diff options
Diffstat (limited to 'src/lib')
48 files changed, 11308 insertions, 2324 deletions
diff --git a/src/lib/CMakeLists.txt b/src/lib/CMakeLists.txt index 79263924..3abf39d0 100644 --- a/src/lib/CMakeLists.txt +++ b/src/lib/CMakeLists.txt @@ -17,8 +17,12 @@ protobuf_generate_c(IPCP_PROTO_SRCS IPCP_PROTO_HDRS set(SOURCE_FILES_COMMON bitmap.c btree.c - crc32.c + crc/crc8.c + crc/crc16.c + crc/crc32.c + crc/crc64.c crypt.c + crypt/keyrot.c hash.c lockfile.c logs.c @@ -36,6 +40,7 @@ set(SOURCE_FILES_COMMON ssm/pool.c sockets.c tpm.c + tw.c utils.c ) @@ -88,6 +93,13 @@ if(HAVE_FUSE) target_link_libraries(ouroboros-common PRIVATE Fuse::Fuse) endif() +if(HAVE_LIBURCU) + target_link_libraries(ouroboros-common PRIVATE Urcu::Urcu) + # urcu headers require C99; override the global -std=c89 for this TU only. + set_source_files_properties(crypt/keyrot.c PROPERTIES + COMPILE_OPTIONS "-std=gnu99") +endif() + install(TARGETS ouroboros-common EXPORT OuroborosTargets LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}) @@ -155,5 +167,6 @@ configure_file("${CMAKE_CURRENT_SOURCE_DIR}/ssm/ssm.h.in" if(BUILD_TESTS) add_subdirectory(tests) + add_subdirectory(crc/tests) add_subdirectory(ssm/tests) endif() diff --git a/src/lib/config.h.in b/src/lib/config.h.in index 08e9baf6..4ebaa194 100644 --- a/src/lib/config.h.in +++ b/src/lib/config.h.in @@ -20,16 +20,23 @@ * Foundation, Inc., http://www.fsf.org/about/contact/. */ +#ifndef MILLION +#define MILLION 1000000LL +#endif + +#ifndef BILLION +#define BILLION 1000000000LL +#endif + #cmakedefine HAVE_SYS_RANDOM #cmakedefine HAVE_EXPLICIT_BZERO #cmakedefine HAVE_LIBGCRYPT #cmakedefine HAVE_OPENSSL #ifdef HAVE_OPENSSL -#cmakedefine HAVE_OPENSSL_ML_KEM -#cmakedefine HAVE_OPENSSL_ML_DSA -#cmakedefine HAVE_OPENSSL_SLH_DSA +#cmakedefine HAVE_ML +#cmakedefine HAVE_SLH #define HAVE_ENCRYPTION -#define SECMEM_GUARD @SECMEM_GUARD@ +#define SECMEM_MINSIZE @SECMEM_MINSIZE@ #endif #define PROC_SECMEM_MAX @PROC_SECMEM_MAX@ @@ -37,6 +44,8 @@ #cmakedefine QOS_DISABLE_CRC #cmakedefine HAVE_OPENSSL_RNG +#cmakedefine HAVE_PCLMUL +#cmakedefine HAVE_PMULL #define SHM_LOCKFILE_NAME "@SHM_LOCKFILE_NAME@" #define FLOW_ALLOC_TIMEOUT @FLOW_ALLOC_TIMEOUT@ @@ -60,16 +69,20 @@ #cmakedefine PROC_FLOW_STATS #endif +#cmakedefine HAVE_LIBURCU + +#cmakedefine FRCT_DEBUG_STDOUT + #define PTHREAD_COND_CLOCK @PTHREAD_COND_CLOCK@ -#define PROG_MAX_FLOWS @PROG_MAX_FLOWS@ -#define PROG_RES_FDS @PROG_RES_FDS@ -#define PROG_MAX_FQUEUES @PROG_MAX_FQUEUES@ +#define PROC_MAX_FLOWS @PROC_MAX_FLOWS@ +#define PROC_RES_FDS @PROC_RES_FDS@ +#define PROC_MAX_FQUEUES @PROC_MAX_FQUEUES@ /* Default Delta-t parameters */ #cmakedefine FRCT_LINUX_RTT_ESTIMATOR -#define DELT_A (@DELTA_T_ACK@) /* ns */ -#define DELT_R (@DELTA_T_RTX@) /* ns */ +#define DELT_A (@DELTA_T_ACK@) /* ms */ +#define DELT_R (@DELTA_T_RTX@) /* ms */ #define RQ_SIZE (@FRCT_REORDER_QUEUE_SIZE@) #define START_WINDOW (@FRCT_START_WINDOW@) @@ -80,9 +93,6 @@ #define TICTIME (@FRCT_TICK_TIME@ * 1000) /* ns */ /* Retransmission tuning */ -#cmakedefine RXM_BUFFER_ON_HEAP -#cmakedefine RXM_BLOCKING - #define RXMQ_RES (@RXM_MIN_RESOLUTION@) /* 2^N ns */ #define RXMQ_BUMP (@RXM_WHEEL_MULTIPLIER@) #define RXMQ_LVLS (@RXM_WHEEL_LEVELS@) @@ -91,4 +101,9 @@ #define ACKQ_SLOTS (@ACK_WHEEL_SLOTS@) #define ACKQ_RES (@ACK_WHEEL_RESOLUTION@) /* 2^N ns */ -#define KEY_ROTATION_BIT (@KEY_ROTATION_BIT@) /* Bit for key rotation */ +#define KEY_LEAF_BITS (@KEY_LEAF_BITS@) /* pkts/leaf-key = 2^n */ +#define KEY_NODE_BITS (@KEY_NODE_BITS@) /* leaf-keys/node = 2^n */ +#define KEY_NODE_COUNT (@KEY_NODE_COUNT@) /* node keys/batch N */ +#define KEY_REKEY_WATERMARK (@KEY_REKEY_WATERMARK@) /* node-keys-left trig */ +#define KEY_REPLAY_WINDOW (@KEY_REPLAY_WINDOW@) /* rx replay win pkts */ +#define FLOW_WM_CHECK (1u << @KEY_REKEY_WM_CHECK_BITS@) /* wm chk/n wr */ diff --git a/src/lib/crc/crc16.c b/src/lib/crc/crc16.c new file mode 100644 index 00000000..9dc59429 --- /dev/null +++ b/src/lib/crc/crc16.c @@ -0,0 +1,61 @@ +/* + * Ouroboros - Copyright (C) 2016 - 2026 + * + * 16-bit Cyclic Redundancy Check (CCITT-FALSE variant) + * + * Dimitri Staessens <dimitri@ouroboros.rocks> + * Sander Vrijders <sander@ouroboros.rocks> + * + * This library is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public License + * version 2.1 as published by the Free Software Foundation. + * + * This library is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with this library; if not, write to the Free Software + * Foundation, Inc., http://www.fsf.org/about/contact/. + */ + +/* + * CRC-16/CCITT-FALSE (reveng catalog, alias CRC-16/IBM-3740): + * poly = 0x1021 + * init = 0xffff + * refin = false + * refout = false + * xorout = 0x0000 + * check = crc16_ccitt_false("123456789") == 0x29b1 + */ + +#include "config.h" + +#include <ouroboros/crc16.h> + +/* Bit-by-bit MSB-first CRC. */ +void crc16_ccitt_false(uint16_t * crc, + const void * buf, + size_t len) +{ + const uint8_t * p; + uint16_t c; + size_t n; + int i; + + p = (const uint8_t *) buf; + c = *crc ^ 0xffff; + + for (n = 0; n < len; n++) { + c ^= ((uint16_t) p[n]) << 8; + for (i = 0; i < 8; i++) { + if (c & 0x8000) + c = (uint16_t) ((c << 1) ^ 0x1021); + else + c = (uint16_t) (c << 1); + } + } + + *crc = c; +} diff --git a/src/lib/crc32.c b/src/lib/crc/crc32.c index 0fdb62b1..0fdb62b1 100644 --- a/src/lib/crc32.c +++ b/src/lib/crc/crc32.c diff --git a/src/lib/crc/crc64.c b/src/lib/crc/crc64.c new file mode 100644 index 00000000..1b6fb5f6 --- /dev/null +++ b/src/lib/crc/crc64.c @@ -0,0 +1,363 @@ +/* + * Ouroboros - Copyright (C) 2016 - 2026 + * + * 64-bit Cyclic Redundancy Check (NVMe variant) + * + * Dimitri Staessens <dimitri@ouroboros.rocks> + * Sander Vrijders <sander@ouroboros.rocks> + * + * This library is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public License + * version 2.1 as published by the Free Software Foundation. + * + * This library is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with this library; if not, write to the Free Software + * Foundation, Inc., http://www.fsf.org/about/contact/. + */ + +/* + * CRC-64/NVMe (reveng catalog): + * poly = 0xad93d23594c93659 + * init = 0xffffffffffffffff + * refin = true + * refout = true + * xorout = 0xffffffffffffffff + * check = crc64_nvme("123456789") == 0xae8b14860a799888 + */ + +#include "config.h" + +#include <ouroboros/crc64.h> + +/* + * Reflected CRC-64/NVMe table. Polynomial in reflected form: + * 0x9a6c9329ac4bc9b5 (bitrev of 0xad93d23594c93659). + */ +static const uint64_t crc64_nvme_tab[256] = { + 0x0000000000000000ULL, 0x7f6ef0c830358979ULL, + 0xfedde190606b12f2ULL, 0x81b31158505e9b8bULL, + 0xc962e5739841b68fULL, 0xb60c15bba8743ff6ULL, + 0x37bf04e3f82aa47dULL, 0x48d1f42bc81f2d04ULL, + 0xa61cecb46814fe75ULL, 0xd9721c7c5821770cULL, + 0x58c10d24087fec87ULL, 0x27affdec384a65feULL, + 0x6f7e09c7f05548faULL, 0x1010f90fc060c183ULL, + 0x91a3e857903e5a08ULL, 0xeecd189fa00bd371ULL, + 0x78e0ff3b88be6f81ULL, 0x078e0ff3b88be6f8ULL, + 0x863d1eabe8d57d73ULL, 0xf953ee63d8e0f40aULL, + 0xb1821a4810ffd90eULL, 0xceecea8020ca5077ULL, + 0x4f5ffbd87094cbfcULL, 0x30310b1040a14285ULL, + 0xdefc138fe0aa91f4ULL, 0xa192e347d09f188dULL, + 0x2021f21f80c18306ULL, 0x5f4f02d7b0f40a7fULL, + 0x179ef6fc78eb277bULL, 0x68f0063448deae02ULL, + 0xe943176c18803589ULL, 0x962de7a428b5bcf0ULL, + 0xf1c1fe77117cdf02ULL, 0x8eaf0ebf2149567bULL, + 0x0f1c1fe77117cdf0ULL, 0x7072ef2f41224489ULL, + 0x38a31b04893d698dULL, 0x47cdebccb908e0f4ULL, + 0xc67efa94e9567b7fULL, 0xb9100a5cd963f206ULL, + 0x57dd12c379682177ULL, 0x28b3e20b495da80eULL, + 0xa900f35319033385ULL, 0xd66e039b2936bafcULL, + 0x9ebff7b0e12997f8ULL, 0xe1d10778d11c1e81ULL, + 0x606216208142850aULL, 0x1f0ce6e8b1770c73ULL, + 0x8921014c99c2b083ULL, 0xf64ff184a9f739faULL, + 0x77fce0dcf9a9a271ULL, 0x08921014c99c2b08ULL, + 0x4043e43f0183060cULL, 0x3f2d14f731b68f75ULL, + 0xbe9e05af61e814feULL, 0xc1f0f56751dd9d87ULL, + 0x2f3dedf8f1d64ef6ULL, 0x50531d30c1e3c78fULL, + 0xd1e00c6891bd5c04ULL, 0xae8efca0a188d57dULL, + 0xe65f088b6997f879ULL, 0x9931f84359a27100ULL, + 0x1882e91b09fcea8bULL, 0x67ec19d339c963f2ULL, + 0xd75adabd7a6e2d6fULL, 0xa8342a754a5ba416ULL, + 0x29873b2d1a053f9dULL, 0x56e9cbe52a30b6e4ULL, + 0x1e383fcee22f9be0ULL, 0x6156cf06d21a1299ULL, + 0xe0e5de5e82448912ULL, 0x9f8b2e96b271006bULL, + 0x71463609127ad31aULL, 0x0e28c6c1224f5a63ULL, + 0x8f9bd7997211c1e8ULL, 0xf0f5275142244891ULL, + 0xb824d37a8a3b6595ULL, 0xc74a23b2ba0eececULL, + 0x46f932eaea507767ULL, 0x3997c222da65fe1eULL, + 0xafba2586f2d042eeULL, 0xd0d4d54ec2e5cb97ULL, + 0x5167c41692bb501cULL, 0x2e0934dea28ed965ULL, + 0x66d8c0f56a91f461ULL, 0x19b6303d5aa47d18ULL, + 0x980521650afae693ULL, 0xe76bd1ad3acf6feaULL, + 0x09a6c9329ac4bc9bULL, 0x76c839faaaf135e2ULL, + 0xf77b28a2faafae69ULL, 0x8815d86aca9a2710ULL, + 0xc0c42c4102850a14ULL, 0xbfaadc8932b0836dULL, + 0x3e19cdd162ee18e6ULL, 0x41773d1952db919fULL, + 0x269b24ca6b12f26dULL, 0x59f5d4025b277b14ULL, + 0xd846c55a0b79e09fULL, 0xa72835923b4c69e6ULL, + 0xeff9c1b9f35344e2ULL, 0x90973171c366cd9bULL, + 0x1124202993385610ULL, 0x6e4ad0e1a30ddf69ULL, + 0x8087c87e03060c18ULL, 0xffe938b633338561ULL, + 0x7e5a29ee636d1eeaULL, 0x0134d92653589793ULL, + 0x49e52d0d9b47ba97ULL, 0x368bddc5ab7233eeULL, + 0xb738cc9dfb2ca865ULL, 0xc8563c55cb19211cULL, + 0x5e7bdbf1e3ac9decULL, 0x21152b39d3991495ULL, + 0xa0a63a6183c78f1eULL, 0xdfc8caa9b3f20667ULL, + 0x97193e827bed2b63ULL, 0xe877ce4a4bd8a21aULL, + 0x69c4df121b863991ULL, 0x16aa2fda2bb3b0e8ULL, + 0xf86737458bb86399ULL, 0x8709c78dbb8deae0ULL, + 0x06bad6d5ebd3716bULL, 0x79d4261ddbe6f812ULL, + 0x3105d23613f9d516ULL, 0x4e6b22fe23cc5c6fULL, + 0xcfd833a67392c7e4ULL, 0xb0b6c36e43a74e9dULL, + 0x9a6c9329ac4bc9b5ULL, 0xe50263e19c7e40ccULL, + 0x64b172b9cc20db47ULL, 0x1bdf8271fc15523eULL, + 0x530e765a340a7f3aULL, 0x2c608692043ff643ULL, + 0xadd397ca54616dc8ULL, 0xd2bd67026454e4b1ULL, + 0x3c707f9dc45f37c0ULL, 0x431e8f55f46abeb9ULL, + 0xc2ad9e0da4342532ULL, 0xbdc36ec59401ac4bULL, + 0xf5129aee5c1e814fULL, 0x8a7c6a266c2b0836ULL, + 0x0bcf7b7e3c7593bdULL, 0x74a18bb60c401ac4ULL, + 0xe28c6c1224f5a634ULL, 0x9de29cda14c02f4dULL, + 0x1c518d82449eb4c6ULL, 0x633f7d4a74ab3dbfULL, + 0x2bee8961bcb410bbULL, 0x548079a98c8199c2ULL, + 0xd53368f1dcdf0249ULL, 0xaa5d9839ecea8b30ULL, + 0x449080a64ce15841ULL, 0x3bfe706e7cd4d138ULL, + 0xba4d61362c8a4ab3ULL, 0xc52391fe1cbfc3caULL, + 0x8df265d5d4a0eeceULL, 0xf29c951de49567b7ULL, + 0x732f8445b4cbfc3cULL, 0x0c41748d84fe7545ULL, + 0x6bad6d5ebd3716b7ULL, 0x14c39d968d029fceULL, + 0x95708ccedd5c0445ULL, 0xea1e7c06ed698d3cULL, + 0xa2cf882d2576a038ULL, 0xdda178e515432941ULL, + 0x5c1269bd451db2caULL, 0x237c997575283bb3ULL, + 0xcdb181ead523e8c2ULL, 0xb2df7122e51661bbULL, + 0x336c607ab548fa30ULL, 0x4c0290b2857d7349ULL, + 0x04d364994d625e4dULL, 0x7bbd94517d57d734ULL, + 0xfa0e85092d094cbfULL, 0x856075c11d3cc5c6ULL, + 0x134d926535897936ULL, 0x6c2362ad05bcf04fULL, + 0xed9073f555e26bc4ULL, 0x92fe833d65d7e2bdULL, + 0xda2f7716adc8cfb9ULL, 0xa54187de9dfd46c0ULL, + 0x24f29686cda3dd4bULL, 0x5b9c664efd965432ULL, + 0xb5517ed15d9d8743ULL, 0xca3f8e196da80e3aULL, + 0x4b8c9f413df695b1ULL, 0x34e26f890dc31cc8ULL, + 0x7c339ba2c5dc31ccULL, 0x035d6b6af5e9b8b5ULL, + 0x82ee7a32a5b7233eULL, 0xfd808afa9582aa47ULL, + 0x4d364994d625e4daULL, 0x3258b95ce6106da3ULL, + 0xb3eba804b64ef628ULL, 0xcc8558cc867b7f51ULL, + 0x8454ace74e645255ULL, 0xfb3a5c2f7e51db2cULL, + 0x7a894d772e0f40a7ULL, 0x05e7bdbf1e3ac9deULL, + 0xeb2aa520be311aafULL, 0x944455e88e0493d6ULL, + 0x15f744b0de5a085dULL, 0x6a99b478ee6f8124ULL, + 0x224840532670ac20ULL, 0x5d26b09b16452559ULL, + 0xdc95a1c3461bbed2ULL, 0xa3fb510b762e37abULL, + 0x35d6b6af5e9b8b5bULL, 0x4ab846676eae0222ULL, + 0xcb0b573f3ef099a9ULL, 0xb465a7f70ec510d0ULL, + 0xfcb453dcc6da3dd4ULL, 0x83daa314f6efb4adULL, + 0x0269b24ca6b12f26ULL, 0x7d0742849684a65fULL, + 0x93ca5a1b368f752eULL, 0xeca4aad306bafc57ULL, + 0x6d17bb8b56e467dcULL, 0x12794b4366d1eea5ULL, + 0x5aa8bf68aecec3a1ULL, 0x25c64fa09efb4ad8ULL, + 0xa4755ef8cea5d153ULL, 0xdb1bae30fe90582aULL, + 0xbcf7b7e3c7593bd8ULL, 0xc399472bf76cb2a1ULL, + 0x422a5673a732292aULL, 0x3d44a6bb9707a053ULL, + 0x759552905f188d57ULL, 0x0afba2586f2d042eULL, + 0x8b48b3003f739fa5ULL, 0xf42643c80f4616dcULL, + 0x1aeb5b57af4dc5adULL, 0x6585ab9f9f784cd4ULL, + 0xe436bac7cf26d75fULL, 0x9b584a0fff135e26ULL, + 0xd389be24370c7322ULL, 0xace74eec0739fa5bULL, + 0x2d545fb4576761d0ULL, 0x523aaf7c6752e8a9ULL, + 0xc41748d84fe75459ULL, 0xbb79b8107fd2dd20ULL, + 0x3acaa9482f8c46abULL, 0x45a459801fb9cfd2ULL, + 0x0d75adabd7a6e2d6ULL, 0x721b5d63e7936bafULL, + 0xf3a84c3bb7cdf024ULL, 0x8cc6bcf387f8795dULL, + 0x620ba46c27f3aa2cULL, 0x1d6554a417c62355ULL, + 0x9cd645fc4798b8deULL, 0xe3b8b53477ad31a7ULL, + 0xab69411fbfb21ca3ULL, 0xd407b1d78f8795daULL, + 0x55b4a08fdfd90e51ULL, 0x2ada5047efec8728ULL +}; + +static __inline__ uint64_t crc64_nvme_step(uint64_t c, + const uint8_t * p, + size_t len) +{ + size_t n; + + for (n = 0; n < len; n++) + c = crc64_nvme_tab[(c ^ p[n]) & 0xff] ^ (c >> 8); + + return c; +} + +void crc64_nvme_table(uint64_t * crc, + const void * buf, + size_t len) +{ + uint64_t c; + + c = crc64_nvme_step(*crc ^ UINT64_MAX, + (const uint8_t *) buf, len); + + *crc = c ^ UINT64_MAX; +} + +#ifdef HAVE_PCLMUL + +#include <smmintrin.h> +#include <wmmintrin.h> + +/* + * Fold-by-16 constants for reflected CRC-64/NVMe. Properties of the + * polynomial; identical between the PCLMUL and PMULL backends. + * k3 = bitrev64(x^(128+64) mod P) << 1 + * k4 = bitrev64(x^(128+0) mod P) << 1 + */ +static const uint64_t k3_clmul = 0xeadc41fd2ba3d420ULL; +static const uint64_t k4_clmul = 0x21e9761e252621acULL; + +__attribute__((target("pclmul,sse4.1"))) +static __m128i fold16(__m128i x, + __m128i k) +{ + __m128i lo; + __m128i hi; + + lo = _mm_clmulepi64_si128(x, k, 0x00); + hi = _mm_clmulepi64_si128(x, k, 0x11); + return _mm_xor_si128(lo, hi); +} + +/* + * Fold-by-16 over 16-byte chunks; the 128-bit folded state is then + * emitted as 16 little-endian bytes and run through the byte-table + * loop together with any tail (<=15 bytes). The 16-byte minimum on + * the bulk loop is why the short-input path uses the table directly. + */ +__attribute__((target("pclmul,sse4.1"))) +static void crc64_nvme_clmul(uint64_t * crc, + const void * buf, + size_t len) +{ + const uint8_t * p; + uint64_t seed; + uint64_t c; + size_t off; + __m128i x; + __m128i k; + uint8_t post[16]; + + p = (const uint8_t *) buf; + seed = *crc; + + if (len < 16) { + c = crc64_nvme_step(seed ^ UINT64_MAX, p, len); + *crc = c ^ UINT64_MAX; + return; + } + + x = _mm_loadu_si128((const __m128i *) p); + x = _mm_xor_si128(x, _mm_cvtsi64_si128((int64_t) + (seed ^ UINT64_MAX))); + + k = _mm_set_epi64x((int64_t) k4_clmul, (int64_t) k3_clmul); + + off = 16; + while (off + 16 <= len) { + __m128i d; + + d = _mm_loadu_si128((const __m128i *) (p + off)); + x = _mm_xor_si128(fold16(x, k), d); + off += 16; + } + + _mm_storeu_si128((__m128i *) post, x); + + c = crc64_nvme_step(0, post, 16); + c = crc64_nvme_step(c, p + off, len - off); + + *crc = c ^ UINT64_MAX; +} + +#endif /* HAVE_PCLMUL */ + +#ifdef HAVE_PMULL + +#include <arm_neon.h> + +/* Same fold-by-16 constants as the PCLMUL path (poly properties). */ +static const uint64_t k3_pmull = 0xeadc41fd2ba3d420ULL; +static const uint64_t k4_pmull = 0x21e9761e252621acULL; + +__attribute__((target("+crypto"))) +static uint64x2_t fold16_pmull(uint64x2_t x, + uint64x2_t k) +{ + poly64x2_t xp; + poly64x2_t kp; + uint64x2_t lo; + uint64x2_t hi; + + xp = vreinterpretq_p64_u64(x); + kp = vreinterpretq_p64_u64(k); + lo = vreinterpretq_u64_p128( + vmull_p64((poly64_t) vgetq_lane_u64(x, 0), + (poly64_t) vgetq_lane_u64(k, 0))); + hi = vreinterpretq_u64_p128(vmull_high_p64(xp, kp)); + return veorq_u64(lo, hi); +} + +__attribute__((target("+crypto"))) +static void crc64_nvme_pmull(uint64_t * crc, + const void * buf, + size_t len) +{ + const uint8_t * p; + uint64_t seed; + uint64_t c; + size_t off; + uint64x2_t x; + uint64x2_t k; + uint64_t seed_lane[2]; + uint64_t k_lanes[2]; + uint8_t post[16]; + + p = (const uint8_t *) buf; + seed = *crc; + + if (len < 16) { + c = crc64_nvme_step(seed ^ UINT64_MAX, p, len); + *crc = c ^ UINT64_MAX; + return; + } + + x = vld1q_u64((const uint64_t *) p); + seed_lane[0] = seed ^ UINT64_MAX; + seed_lane[1] = 0; + x = veorq_u64(x, vld1q_u64(seed_lane)); + + k_lanes[0] = k3_pmull; + k_lanes[1] = k4_pmull; + k = vld1q_u64(k_lanes); + + off = 16; + while (off + 16 <= len) { + uint64x2_t d; + + d = vld1q_u64((const uint64_t *) (p + off)); + x = veorq_u64(fold16_pmull(x, k), d); + off += 16; + } + + vst1q_u8(post, vreinterpretq_u8_u64(x)); + + c = crc64_nvme_step(0, post, 16); + c = crc64_nvme_step(c, p + off, len - off); + + *crc = c ^ UINT64_MAX; +} +#endif /* HAVE_PMULL */ + +void crc64_nvme(uint64_t * crc, + const void * buf, + size_t len) +{ +#ifdef HAVE_PCLMUL + crc64_nvme_clmul(crc, buf, len); +#elif defined(HAVE_PMULL) + crc64_nvme_pmull(crc, buf, len); +#else + crc64_nvme_table(crc, buf, len); +#endif +} diff --git a/src/lib/crc/crc8.c b/src/lib/crc/crc8.c new file mode 100644 index 00000000..20976b29 --- /dev/null +++ b/src/lib/crc/crc8.c @@ -0,0 +1,62 @@ +/* + * Ouroboros - Copyright (C) 2016 - 2026 + * + * 8-bit Cyclic Redundancy Check (AUTOSAR variant) + * + * Dimitri Staessens <dimitri@ouroboros.rocks> + * Sander Vrijders <sander@ouroboros.rocks> + * + * This library is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public License + * version 2.1 as published by the Free Software Foundation. + * + * This library is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with this library; if not, write to the Free Software + * Foundation, Inc., http://www.fsf.org/about/contact/. + */ + +/* + * CRC-8/AUTOSAR (reveng catalog): + * poly = 0x2f + * init = 0xff + * refin = false + * refout = false + * xorout = 0xff + * check = crc8_autosar("123456789") == 0xdf + */ + +#include "config.h" + +#include <ouroboros/crc8.h> + + + /* Bit-by-bit MSB-first CRC. */ +void crc8_autosar(uint8_t * crc, + const void * buf, + size_t len) +{ + const uint8_t * p; + uint8_t c; + size_t n; + int i; + + p = (const uint8_t *) buf; + c = *crc ^ 0xff; + + for (n = 0; n < len; n++) { + c ^= p[n]; + for (i = 0; i < 8; i++) { + if (c & 0x80) + c = (uint8_t) ((c << 1) ^ 0x2f); + else + c = (uint8_t) (c << 1); + } + } + + *crc = c ^ 0xff; +} diff --git a/src/lib/crc/tests/CMakeLists.txt b/src/lib/crc/tests/CMakeLists.txt new file mode 100644 index 00000000..11daca5a --- /dev/null +++ b/src/lib/crc/tests/CMakeLists.txt @@ -0,0 +1,21 @@ +get_filename_component(PARENT_PATH ${CMAKE_CURRENT_SOURCE_DIR} DIRECTORY) +get_filename_component(PARENT_DIR ${PARENT_PATH} NAME) + +compute_test_prefix() + +create_test_sourcelist(${PARENT_DIR}_tests test_suite.c + # Add new tests here + crc8_test.c + crc16_test.c + crc32_test.c + crc64_test.c + ) + +add_executable(${PARENT_DIR}_test ${${PARENT_DIR}_tests}) + +disable_test_logging_for_target(${PARENT_DIR}_test) +target_link_libraries(${PARENT_DIR}_test ouroboros-common) + +add_dependencies(build_tests ${PARENT_DIR}_test) + +ouroboros_register_tests(TARGET ${PARENT_DIR}_test TESTS ${${PARENT_DIR}_tests}) diff --git a/src/lib/crc/tests/crc16_test.c b/src/lib/crc/tests/crc16_test.c new file mode 100644 index 00000000..03a5b504 --- /dev/null +++ b/src/lib/crc/tests/crc16_test.c @@ -0,0 +1,67 @@ +/* + * Ouroboros - Copyright (C) 2016 - 2026 + * + * Test of the CRC-16/CCITT-FALSE function + * + * Dimitri Staessens <dimitri@ouroboros.rocks> + * Sander Vrijders <sander@ouroboros.rocks> + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 as + * published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., http://www.fsf.org/about/contact/. + */ + +#include "config.h" + +#include <ouroboros/crc16.h> + +#include <test/test.h> + +#include <stddef.h> +#include <stdint.h> +#include <stdio.h> + +/* reveng-catalog smoke vectors. */ +static int test_crc16_ccitt_false_basic(void) +{ + uint16_t crc; + + TEST_START(); + + crc = 0; + crc16_ccitt_false(&crc, "", 0); + if (crc != 0xffff) + goto fail; + + crc = 0; + crc16_ccitt_false(&crc, "123456789", 9); + if (crc != 0x29b1) + goto fail; + + TEST_SUCCESS(); + return TEST_RC_SUCCESS; + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +int crc16_test(int argc, + char ** argv) +{ + int ret = 0; + + (void) argc; + (void) argv; + + ret |= test_crc16_ccitt_false_basic(); + return ret; +} diff --git a/src/lib/tests/crc32_test.c b/src/lib/crc/tests/crc32_test.c index 5a1ddd87..5a1ddd87 100644 --- a/src/lib/tests/crc32_test.c +++ b/src/lib/crc/tests/crc32_test.c diff --git a/src/lib/crc/tests/crc64_test.c b/src/lib/crc/tests/crc64_test.c new file mode 100644 index 00000000..cf3f5ca3 --- /dev/null +++ b/src/lib/crc/tests/crc64_test.c @@ -0,0 +1,126 @@ +/* + * Ouroboros - Copyright (C) 2016 - 2026 + * + * Test of the CRC-64/NVMe function + * + * Dimitri Staessens <dimitri@ouroboros.rocks> + * Sander Vrijders <sander@ouroboros.rocks> + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 as + * published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., http://www.fsf.org/about/contact/. + */ + +#include "config.h" + +#include <ouroboros/crc64.h> +#include <ouroboros/random.h> + +#include <test/test.h> + +#include <stddef.h> +#include <stdint.h> +#include <stdio.h> + +/* Reference impl, internal to libouroboros-common. */ +extern void crc64_nvme_table(uint64_t * crc, + const void * buf, + size_t len); + +/* reveng-catalog smoke vectors plus a 16-byte fold-boundary check. */ +static int test_crc64_nvme_basic(void) +{ + uint64_t crc; + + TEST_START(); + + crc = 0; + crc64_nvme(&crc, "", 0); + if (crc != 0x0000000000000000ULL) + goto fail; + + crc = 0; + crc64_nvme(&crc, "123456789", 9); + if (crc != 0xae8b14860a799888ULL) + goto fail; + + crc = 0; + crc64_nvme(&crc, "0123456789abcdef", 16); + if (crc != 0x091485ca7018730eULL) + goto fail; + + TEST_SUCCESS(); + return TEST_RC_SUCCESS; + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +#if defined(HAVE_PCLMUL) || defined(HAVE_PMULL) +/* Cross-check the accelerated dispatcher path against the byte-table. */ +static int test_crc64_nvme_random(void) +{ + static const size_t lens[] = { + 0, 1, 7, 8, 15, 16, 17, 31, 32, 33, 63, 64, 65, 127, 128, + 129, 255, 256, 257, 1023, 1024, 1025, 4096 + }; + uint8_t buf[4096]; + size_t i; + uint64_t ref; + uint64_t got; + + TEST_START(); + + if (random_buffer(buf, sizeof(buf)) < 0) { + printf("Failed to generate random data.\n"); + goto fail; + } + + for (i = 0; i < sizeof(lens) / sizeof(lens[0]); i++) { + ref = 0; + crc64_nvme_table(&ref, buf, lens[i]); + + got = 0; + crc64_nvme(&got, buf, lens[i]); + + if (ref == got) + continue; + + printf("Mismatch at len=%zu: table=0x%016lx disp=0x%016lx\n", + lens[i], + (unsigned long) ref, + (unsigned long) got); + goto fail; + } + + TEST_SUCCESS(); + return TEST_RC_SUCCESS; + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +#endif +} + +int crc64_test(int argc, + char ** argv) +{ + int ret = 0; + + (void) argc; + (void) argv; + + ret |= test_crc64_nvme_basic(); +#if defined(HAVE_PCLMUL) || defined(HAVE_PMULL) + ret |= test_crc64_nvme_random(); +#endif + return ret; +} diff --git a/src/lib/crc/tests/crc8_test.c b/src/lib/crc/tests/crc8_test.c new file mode 100644 index 00000000..f7bb33b8 --- /dev/null +++ b/src/lib/crc/tests/crc8_test.c @@ -0,0 +1,67 @@ +/* + * Ouroboros - Copyright (C) 2016 - 2026 + * + * Test of the CRC-8/AUTOSAR function + * + * Dimitri Staessens <dimitri@ouroboros.rocks> + * Sander Vrijders <sander@ouroboros.rocks> + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 as + * published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., http://www.fsf.org/about/contact/. + */ + +#include "config.h" + +#include <ouroboros/crc8.h> + +#include <test/test.h> + +#include <stddef.h> +#include <stdint.h> +#include <stdio.h> + +/* reveng-catalog smoke vectors. */ +static int test_crc8_autosar_basic(void) +{ + uint8_t crc; + + TEST_START(); + + crc = 0; + crc8_autosar(&crc, "", 0); + if (crc != 0x00) + goto fail; + + crc = 0; + crc8_autosar(&crc, "123456789", 9); + if (crc != 0xdf) + goto fail; + + TEST_SUCCESS(); + return TEST_RC_SUCCESS; + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +int crc8_test(int argc, + char ** argv) +{ + int ret = 0; + + (void) argc; + (void) argv; + + ret |= test_crc8_autosar_basic(); + return ret; +} diff --git a/src/lib/crypt.c b/src/lib/crypt.c index cd3421dd..cbbb9bc2 100644 --- a/src/lib/crypt.c +++ b/src/lib/crypt.c @@ -27,10 +27,14 @@ #include <config.h> #include <ouroboros/errno.h> +#include <ouroboros/pthread.h> #include <ouroboros/random.h> #include <ouroboros/crypt.h> +#include "crypt/keyrot.h" + #ifdef HAVE_OPENSSL +#include <openssl/crypto.h> #include <openssl/evp.h> #include "crypt/openssl.h" #endif @@ -50,18 +54,12 @@ static const struct nid_map cipher_nid_map[] = { {NID_aes_192_gcm, "aes-192-gcm"}, {NID_aes_256_gcm, "aes-256-gcm"}, {NID_chacha20_poly1305, "chacha20-poly1305"}, - {NID_aes_128_ctr, "aes-128-ctr"}, - {NID_aes_192_ctr, "aes-192-ctr"}, - {NID_aes_256_ctr, "aes-256-ctr"}, {NID_undef, NULL} }; /* Ordered in strength preference, lowest first */ const uint16_t crypt_supported_nids[] = { #ifdef HAVE_OPENSSL - NID_aes_128_ctr, - NID_aes_192_ctr, - NID_aes_256_ctr, NID_aes_128_gcm, NID_aes_192_gcm, NID_aes_256_gcm, @@ -87,23 +85,23 @@ static const struct nid_map kex_nid_map[] = { {NID_undef, NULL} }; -/* Ordered in strength preference, lowest first */ +/* Ordered in strength preference, lowest first (NIST SP 800-57 levels) */ const uint16_t kex_supported_nids[] = { #ifdef HAVE_OPENSSL - NID_ffdhe2048, - NID_X9_62_prime256v1, - NID_X25519, - NID_ffdhe3072, - NID_secp384r1, - NID_ffdhe4096, - NID_X448, - NID_secp521r1, -#ifdef HAVE_OPENSSL_ML_KEM - NID_MLKEM512, - NID_MLKEM768, - NID_MLKEM1024, - NID_X25519MLKEM768, - NID_X448MLKEM1024, + NID_ffdhe2048, /* FFDHE-2048, ~112-bit */ + NID_X9_62_prime256v1, /* ECDH P-256, 128-bit */ + NID_X25519, /* ECDH X25519, 128-bit */ + NID_ffdhe3072, /* FFDHE-3072, ~128-bit */ + NID_ffdhe4096, /* FFDHE-4096, ~152-bit */ + NID_secp384r1, /* ECDH P-384, 192-bit */ + NID_X448, /* ECDH X448, 224-bit */ + NID_secp521r1, /* ECDH P-521, 256-bit */ +#ifdef HAVE_ML + NID_MLKEM512, /* ML-KEM-512, PQC L1 (~AES-128) */ + NID_MLKEM768, /* ML-KEM-768, PQC L3 (~AES-192) */ + NID_MLKEM1024, /* ML-KEM-1024, PQC L5 (~AES-256) */ + NID_X25519MLKEM768, /* X25519 + ML-KEM-768, PQC L3 */ + NID_X448MLKEM1024, /* X448 + ML-KEM-1024, PQC L5 */ #endif #endif NID_undef @@ -137,11 +135,13 @@ const uint16_t md_supported_nids[] = { }; struct crypt_ctx { - void * ctx; /* Encryption context */ + struct keyrot * kr; /* backend-independent key rotation */ + void * cipher; /* backend AEAD cipher context */ }; struct auth_ctx { - void * store; + void * store; /* trusted anchors */ + void * chain; /* untrusted build-only interm */ }; static int parse_kex_value(const char * value, @@ -162,6 +162,7 @@ int parse_sec_config(struct sec_config * cfg, char * equals; char * key; char * value; + bool no_enc = false; assert(cfg != NULL); assert(fp != NULL); @@ -172,6 +173,7 @@ int parse_sec_config(struct sec_config * cfg, SET_KEX_KDF_NID(cfg, NID_sha256); SET_KEX_CIPHER_NID(cfg, NID_aes_256_gcm); SET_KEX_DIGEST_NID(cfg, NID_sha256); + /* a.req is seeded per-role by the caller; only auth= overrides it */ while (fgets(line, sizeof(line), fp) != NULL) { char * trimmed; @@ -180,12 +182,10 @@ int parse_sec_config(struct sec_config * cfg, if (line[0] == '#' || line[0] == '\n') continue; - /* Check for 'none' keyword */ + /* Bare 'none' keyword replaced by encryption=none */ trimmed = trim_whitespace(line); - if (strcmp(trimmed, "none") == 0) { - memset(cfg, 0, sizeof(*cfg)); - return 0; - } + if (strcmp(trimmed, "none") == 0) + return -EINVAL; /* Find the = separator */ equals = strchr(line, '='); @@ -221,9 +221,34 @@ int parse_sec_config(struct sec_config * cfg, } else { return -EINVAL; } + } else if (strcmp(key, "auth") == 0) { + if (strcmp(value, "required") == 0) { + cfg->a.req = true; + } else if (strcmp(value, "optional") == 0) { + cfg->a.req = false; + } else { + return -EINVAL; + } + } else if (strcmp(key, "cacert") == 0) { + if (strlen(value) >= sizeof(cfg->a.cacert)) + return -EINVAL; + strcpy(cfg->a.cacert, value); + } else if (strcmp(key, "encryption") == 0) { + if (strcmp(value, "none") != 0) + return -EINVAL; + no_enc = true; + } else { + return -EINVAL; } } + if (no_enc) { + /* Digest stays: it belongs to the auth axis */ + CLEAR_KEX_ALGO(cfg); + CLEAR_KEX_KDF(cfg); + CLEAR_KEX_CIPHER(cfg); + } + return 0; } @@ -239,12 +264,17 @@ int load_sec_config_file(struct sec_config * cfg, fp = fopen(path, "r"); if (fp == NULL) { - /* File doesn't exist - disable encryption */ - CLEAR_KEX_ALGO(cfg); - return 0; + /* Absent config disables encryption; other errors fail */ + if (errno == ENOENT) { + CLEAR_KEX_ALGO(cfg); + return 0; + } + return -errno; } + pthread_cleanup_push(__cleanup_fclose, fp); ret = parse_sec_config(cfg, fp); + pthread_cleanup_pop(0); fclose(fp); @@ -592,19 +622,71 @@ int crypt_kex_rank(int nid) return -1; } -/* Hash length now returned by md_digest() */ +/* AEAD primitive: 1:1 backend wrappers used by the data path below. */ +static int crypt_seal(void * cipher, + const uint8_t * key, + const uint8_t * nonce, + buffer_t aad, + buffer_t in, + uint8_t * out, + uint8_t * tag) +{ +#ifdef HAVE_OPENSSL + return openssl_seal(cipher, key, nonce, aad, in, out, tag); +#else + (void) cipher; + (void) key; + (void) nonce; + (void) aad; + (void) in; + (void) out; + (void) tag; -int crypt_encrypt(struct crypt_ctx * ctx, - buffer_t in, - buffer_t * out) + return -ECRYPT; +#endif +} + +static int crypt_open(void * cipher, + const uint8_t * key, + const uint8_t * nonce, + buffer_t aad, + buffer_t in, + const uint8_t * tag, + buffer_t * out) { - assert(ctx != NULL); - assert(ctx->ctx != NULL); +#ifdef HAVE_OPENSSL + return openssl_open(cipher, key, nonce, aad, in, tag, out); +#else + (void) cipher; + (void) key; + (void) nonce; + (void) aad; + (void) in; + (void) tag; + (void) out; + + return -ECRYPT; +#endif +} + +int crypt_oneshot_seal(int nid, + const uint8_t * key, + const uint8_t * nonce, + buffer_t aad, + buffer_t in, + buffer_t * out) +{ + assert(key != NULL); + assert(nonce != NULL); + assert(out != NULL); #ifdef HAVE_OPENSSL - return openssl_encrypt(ctx->ctx, in, out); + return openssl_oneshot_seal(nid, key, nonce, aad, in, out); #else - (void) ctx; + (void) nid; + (void) key; + (void) nonce; + (void) aad; (void) in; (void) out; @@ -612,17 +694,24 @@ int crypt_encrypt(struct crypt_ctx * ctx, #endif } -int crypt_decrypt(struct crypt_ctx * ctx, - buffer_t in, - buffer_t * out) +int crypt_oneshot_open(int nid, + const uint8_t * key, + const uint8_t * nonce, + buffer_t aad, + buffer_t in, + buffer_t * out) { - assert(ctx != NULL); - assert(ctx->ctx != NULL); + assert(key != NULL); + assert(nonce != NULL); + assert(out != NULL); #ifdef HAVE_OPENSSL - return openssl_decrypt(ctx->ctx, in, out); + return openssl_oneshot_open(nid, key, nonce, aad, in, out); #else - (void) ctx; + (void) nid; + (void) key; + (void) nonce; + (void) aad; (void) in; (void) out; @@ -630,6 +719,115 @@ int crypt_decrypt(struct crypt_ctx * ctx, #endif } +/* + * Data-path encrypt: rotate the key, frame selector ‖ ct ‖ tag, seal. + * Backend-agnostic: composed from keyrot_*, crypt_seal and crypt_get_tagsz. + */ +int crypt_encrypt(struct crypt_ctx * ctx, + buffer_t in, + buffer_t * out) +{ + uint8_t nonce[KR_NONCE_LEN]; + const uint8_t * key; + uint8_t * ct; + buffer_t aad; + int tagsz; + int out_sz; + + assert(ctx != NULL); + assert(ctx->kr != NULL); + + tagsz = crypt_get_tagsz(ctx); + if (tagsz < 0) + return -ECRYPT; + + out->data = malloc(KR_SELECTOR_LEN + in.len + (size_t) tagsz); + if (out->data == NULL) + goto fail_malloc; + + ct = out->data + KR_SELECTOR_LEN; + + /* keyrot writes the selector into the wire header (== AAD). */ + if (keyrot_tx_next(ctx->kr, out->data, &key, nonce) != 0) + goto fail_encrypt; + + aad.data = out->data; + aad.len = KR_SELECTOR_LEN; + + out_sz = crypt_seal(ctx->cipher, key, nonce, aad, in, ct, ct + in.len); + if (out_sz < 0) + goto fail_encrypt; + + out->len = KR_SELECTOR_LEN + (size_t) out_sz + (size_t) tagsz; + + return 0; + fail_encrypt: + free(out->data); + fail_malloc: + clrbuf(*out); + return -ECRYPT; +} + +/* + * Data-path decrypt: look up the rotated key from the selector, open, and + * commit the replay window only after the tag verifies. + */ +int crypt_decrypt(struct crypt_ctx * ctx, + buffer_t in, + buffer_t * out) +{ + uint8_t nonce[KR_NONCE_LEN]; + const uint8_t * key; + const uint8_t * tag; + struct kr_rx rx; + buffer_t aad; + buffer_t ct; + int tagsz; + int in_sz; + + assert(ctx != NULL); + assert(ctx->kr != NULL); + + tagsz = crypt_get_tagsz(ctx); + if (tagsz < 0) + return -ECRYPT; + + if (in.len < (size_t) (KR_SELECTOR_LEN + tagsz)) + return -ECRYPT; + + if (keyrot_rx_lookup(ctx->kr, in.data, &key, nonce, &rx) != 0) + return -ECRYPT; + + in_sz = (int) in.len - KR_SELECTOR_LEN - tagsz; + + /* +1 keeps malloc(0) defined for an empty (zero-length) frame. */ + out->data = malloc((size_t) in_sz + 1); + if (out->data == NULL) + goto fail_malloc; + + aad.data = in.data; + aad.len = KR_SELECTOR_LEN; + + ct.data = in.data + KR_SELECTOR_LEN; + ct.len = (size_t) in_sz; + + tag = in.data + KR_SELECTOR_LEN + in_sz; + + if (crypt_open(ctx->cipher, key, nonce, aad, ct, tag, out) < 0) + goto fail_decrypt; + + /* Commit replay state only after the tag verifies. */ + if (keyrot_rx_commit(ctx->kr, &rx) != 0) + goto fail_decrypt; + + return 0; + fail_decrypt: + free(out->data); + fail_malloc: + clrbuf(*out); + return -ECRYPT; +} + struct crypt_ctx * crypt_create_ctx(struct crypt_sk * sk) { struct crypt_ctx * crypt; @@ -643,16 +841,23 @@ struct crypt_ctx * crypt_create_ctx(struct crypt_sk * sk) memset(crypt, 0, sizeof(*crypt)); + crypt->kr = keyrot_create(sk->key, sk->epoch, sk->role); + if (crypt->kr == NULL) + goto fail_kr; + #ifdef HAVE_OPENSSL - crypt->ctx = openssl_crypt_create_ctx(sk); - if (crypt->ctx == NULL) - goto fail_ctx; + crypt->cipher = openssl_crypt_create_ctx(sk); + if (crypt->cipher == NULL) + goto fail_cipher; #endif return crypt; + #ifdef HAVE_OPENSSL - fail_ctx: - free(crypt); + fail_cipher: + keyrot_destroy(crypt->kr); #endif + fail_kr: + free(crypt); fail_crypt: return NULL; } @@ -662,43 +867,76 @@ void crypt_destroy_ctx(struct crypt_ctx * crypt) if (crypt == NULL) return; + keyrot_destroy(crypt->kr); #ifdef HAVE_OPENSSL - assert(crypt->ctx != NULL); - openssl_crypt_destroy_ctx(crypt->ctx); -#else - assert(crypt->ctx == NULL); + openssl_crypt_destroy_ctx(crypt->cipher); #endif free(crypt); } -int crypt_get_ivsz(struct crypt_ctx * ctx) +int crypt_get_headsz(struct crypt_ctx * ctx) { - if (ctx == NULL) - return -EINVAL; + assert(ctx != NULL); + assert(ctx->kr != NULL); -#ifdef HAVE_OPENSSL - assert(ctx->ctx != NULL); - return openssl_crypt_get_ivsz(ctx->ctx); -#else - assert(ctx->ctx == NULL); - return -ENOTSUP; -#endif + (void) ctx; /* validated only; header size is a constant */ + + return KR_SELECTOR_LEN; +} + +int crypt_rekey(struct crypt_ctx * ctx, + struct crypt_sk * sk) +{ + int ret; + + assert(ctx != NULL); + assert(sk != NULL); + assert(ctx->kr != NULL); + + ret = keyrot_rekey(ctx->kr, sk->key, sk->epoch); + if (ret == -EREPLAY) + return -EREPLAY; + + return ret == 0 ? 0 : -ECRYPT; } int crypt_get_tagsz(struct crypt_ctx * ctx) { - if (ctx == NULL) - return -EINVAL; + assert(ctx != NULL); + assert(ctx->cipher != NULL); #ifdef HAVE_OPENSSL - assert(ctx->ctx != NULL); - return openssl_crypt_get_tagsz(ctx->ctx); + return openssl_crypt_get_tagsz(ctx->cipher); #else - assert(ctx->ctx == NULL); + (void) ctx; return -ENOTSUP; #endif } +int crypt_nodes_left(struct crypt_ctx * ctx) +{ + assert(ctx != NULL); + assert(ctx->kr != NULL); + + return (int) keyrot_tx_nodes_left(ctx->kr); +} + +int crypt_peer_synced(struct crypt_ctx * ctx) +{ + assert(ctx != NULL); + assert(ctx->kr != NULL); + + return keyrot_peer_switched(ctx->kr) ? 1 : 0; +} + +void crypt_tx_promote(struct crypt_ctx * ctx) +{ + assert(ctx != NULL); + assert(ctx->kr != NULL); + + keyrot_tx_promote(ctx->kr); +} + int crypt_load_privkey_file(const char * path, void ** key) { @@ -801,6 +1039,25 @@ int crypt_load_privkey_raw_file(const char * path, #endif } +int crypt_ct_cmp(const void * a, + const void * b, + size_t len) +{ +#ifdef HAVE_OPENSSL + return CRYPTO_memcmp(a, b, len); +#else + const volatile uint8_t * pa = a; + const volatile uint8_t * pb = b; + uint8_t d = 0; + size_t i; + + for (i = 0; i < len; i++) + d |= pa[i] ^ pb[i]; + + return d != 0; +#endif +} + int crypt_cmp_key(const void * key1, const void * key2) { @@ -967,9 +1224,15 @@ struct auth_ctx * auth_create_ctx(void) ctx->store = openssl_auth_create_store(); if (ctx->store == NULL) goto fail_store; + + ctx->chain = openssl_auth_create_chain(); + if (ctx->chain == NULL) + goto fail_chain; #endif return ctx; #ifdef HAVE_OPENSSL + fail_chain: + openssl_auth_destroy_store(ctx->store); fail_store: free(ctx); #endif @@ -982,6 +1245,7 @@ void auth_destroy_ctx(struct auth_ctx * ctx) if (ctx == NULL) return; #ifdef HAVE_OPENSSL + openssl_auth_destroy_chain(ctx->chain); openssl_auth_destroy_store(ctx->store); #endif free(ctx); @@ -1003,19 +1267,61 @@ int auth_add_crt_to_store(struct auth_ctx * ctx, #endif } +int auth_add_crt_to_chain(struct auth_ctx * ctx, + void * crt) +{ + assert(ctx != NULL); + assert(crt != NULL); + +#ifdef HAVE_OPENSSL + return openssl_auth_add_crt_to_chain(ctx->chain, crt); +#else + (void) ctx; + (void) crt; + + return 0; +#endif +} + int auth_verify_crt(struct auth_ctx * ctx, void * crt) { #ifdef HAVE_OPENSSL - return openssl_verify_crt(ctx->store, crt); + return openssl_verify_crt(ctx->store, ctx->chain, crt); +#else + (void) ctx; + (void) crt; + + return 0; +#endif +} + +int auth_verify_crt_pin(struct auth_ctx * ctx, + void * crt, + void * pin) +{ +#ifdef HAVE_OPENSSL + return openssl_verify_crt_pin(ctx->store, ctx->chain, crt, pin); #else (void) ctx; (void) crt; + (void) pin; return 0; #endif } +bool crypt_pk_requires_md(const void * pk) +{ +#ifdef HAVE_OPENSSL + return openssl_pk_requires_md((const EVP_PKEY *) pk); +#else + (void) pk; + + return false; +#endif +} + int auth_sign(void * pkp, int md_nid, buffer_t msg, @@ -1077,10 +1383,25 @@ ssize_t md_len(int md_nid) #endif } +int crypt_hkdf_expand(buffer_t key, + buffer_t info, + buffer_t out) +{ +#ifdef HAVE_OPENSSL + return openssl_hkdf_expand(key, info, out) == 0 ? 0 : -ECRYPT; +#else + (void) key; + (void) info; + (void) out; + + return -ECRYPT; +#endif +} + int crypt_secure_malloc_init(size_t max) { #ifdef HAVE_OPENSSL - return openssl_secure_malloc_init(max, SECMEM_GUARD); + return openssl_secure_malloc_init(max, SECMEM_MINSIZE); #else (void) max; return 0; @@ -1094,6 +1415,13 @@ void crypt_secure_malloc_fini(void) #endif } +void crypt_cleanup(void) +{ +#ifdef HAVE_OPENSSL + openssl_cleanup(); +#endif +} + void * crypt_secure_malloc(size_t size) { #ifdef HAVE_OPENSSL diff --git a/src/lib/crypt/keyrot.c b/src/lib/crypt/keyrot.c new file mode 100644 index 00000000..e98df356 --- /dev/null +++ b/src/lib/crypt/keyrot.c @@ -0,0 +1,775 @@ +/* + * Ouroboros - Copyright (C) 2016 - 2026 + * + * Data-plane key-rotation schedule (node/leaf keys, selector) + * + * Dimitri Staessens <dimitri@ouroboros.rocks> + * Sander Vrijders <sander@ouroboros.rocks> + * + * This library is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public License + * version 2.1 as published by the Free Software Foundation. + * + * This library is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with this library; if not, write to the Free Software + * Foundation, Inc., http://www.fsf.org/about/contact/. + */ + +#define _POSIX_C_SOURCE 200809L + +#include <config.h> + +#include <ouroboros/atomics.h> +#include <ouroboros/crypt.h> +#include <ouroboros/errno.h> +#include <ouroboros/pthread.h> +#include <ouroboros/rcu.h> + +#include "crypt/keyrot.h" + +#include <assert.h> +#include <stdbool.h> +#include <stdlib.h> +#include <string.h> + +/* + * Per-flow keys are addressed by (epoch, node, leaf) and derived as: + * root = per-batch HKDF PRK from the OAP exchange, wiped once expanded + * nodes = HKDF-Expand(root, "o7s-keyrot-node") -> KEY_NODE_COUNT keys + * leaf = HKDF-Expand(node, "o7s-keyrot-leaf"|dir|leaf) -> AEAD key + * The epoch is a small wrapping counter, carried in the selector, that picks + * the live batch; a Tier-2 OAP re-key advances it. The "dir" byte forks the + * leaf keys per direction. + * + * Concurrency: cur/prev batch pointers are published by a re-key and read on + * the data path under an rcu_guard (lock-free RCU with liburcu, else a per- + * keyrot rwlock). The per-batch TX counter is atomic, so the (epoch, counter) + * nonce is unique without serialising TX. Leaf caches are THREAD-LOCAL (an app + * writer and the FRCT retransmit timer never share cache state), keyed on a + * global batch id and direct-mapped. + */ + +#define KR_WITHIN_BITS (KEY_LEAF_BITS + KEY_NODE_BITS) +#define KR_WITHIN_MASK (((uint64_t) 1 << KR_WITHIN_BITS) - 1) +#define KR_N (KEY_NODE_COUNT) +#define KR_LEAVES (1u << KEY_NODE_BITS) +#define KR_BATCH_MAX ((uint64_t) KR_N << KR_WITHIN_BITS) +#define KR_NODES_SZ ((size_t) KR_N * SYMMKEYSZ) +#define KR_TCACHE_WAYS 16 /* per-thread cache slots per direction (pow2) */ +#define KR_EPOCHS 16 /* 4-bit wire epoch: gens before wrap */ + +#define KR_RP_WORDS (KEY_REPLAY_WINDOW / 64) /* pow2; RFC 6479 bitmap */ +#define KR_RP_SHIFT 6 +#define KR_RP_MASK 63 +#define KR_RP_WINDOW (KEY_REPLAY_WINDOW - 64) /* reserve 1 slack word */ + +static const char kr_node_label[] = "o7s-keyrot-node"; +static const char kr_leaf_label[] = "o7s-keyrot-leaf"; + +struct kr_batch { + uint64_t id; /* process-global, unique; cache key (no ABA) */ + uint8_t epoch; /* 4-bit wire selector */ + uint8_t * nodes; /* KR_NODES_SZ in secure heap; NULL if empty */ + uint64_t tx_ctr; /* atomic; per-batch so nonces never collide */ + + struct { /* RFC 6479-like anti-replay window */ + uint64_t last; /* highest accepted ctr + 1 */ + uint64_t bits[KR_RP_WORDS]; + pthread_mutex_t mtx; + } rp; +}; + +struct kr_keycache { + uint8_t * key; /* SYMMKEYSZ, points into the per-thread slab */ + uint64_t id; /* batch the cached key belongs to */ + uint16_t node; + uint8_t leaf; + uint8_t dir; + bool valid; +}; + +struct keyrot { + struct kr_batch * cur; /* published; read on data path */ + struct kr_batch * prev; /* NULL = none */ + struct rcu_guard guard; /* re-key vs readers */ + uint8_t role; + uint8_t tx_epoch; /* epoch TX currently stamps */ + bool peer_switched; /* peer is on the cur epoch */ +}; + +/* Per-thread leaf-key caches, freed by the thread-exit destructor. */ +struct kr_tcache { + struct kr_keycache tx[KR_TCACHE_WAYS]; + struct kr_keycache rx[KR_TCACHE_WAYS]; + uint8_t * slab; /* 2*KR_TCACHE_WAYS*SYMMKEYSZ secure heap */ +}; + +static struct { + uint64_t next_id; /* batch-id allocator (atomic) */ + pthread_key_t tcache_key; /* per-thread leaf-key caches */ + pthread_once_t tcache_once; +} kr_g = { 0, 0, PTHREAD_ONCE_INIT }; + +static void kr_tcache_free(void * p) +{ + struct kr_tcache * t = p; + + if (t == NULL) + return; + + crypt_secure_free(t->slab, 2 * KR_TCACHE_WAYS * SYMMKEYSZ); + free(t); +} + +static void kr_tcache_init(void) +{ + pthread_key_create(&kr_g.tcache_key, kr_tcache_free); +} + +static struct kr_tcache * kr_tcache_get(void) +{ + struct kr_tcache * t; + size_t i; + + pthread_once(&kr_g.tcache_once, kr_tcache_init); + + t = pthread_getspecific(kr_g.tcache_key); + if (t != NULL) + return t; + + t = malloc(sizeof(*t)); + if (t == NULL) + goto fail_alloc; + + memset(t, 0, sizeof(*t)); + + t->slab = crypt_secure_malloc(2 * KR_TCACHE_WAYS * SYMMKEYSZ); + if (t->slab == NULL) + goto fail_slab; + + for (i = 0; i < KR_TCACHE_WAYS; i++) { + t->tx[i].key = t->slab + i * SYMMKEYSZ; + t->rx[i].key = t->slab + (KR_TCACHE_WAYS + i) * SYMMKEYSZ; + } + + if (pthread_setspecific(kr_g.tcache_key, t) != 0) + goto fail_set; + + return t; + + fail_set: + crypt_secure_free(t->slab, 2 * KR_TCACHE_WAYS * SYMMKEYSZ); + fail_slab: + free(t); + fail_alloc: + return NULL; +} + +static uint8_t * kr_expand_nodes(const uint8_t * root) +{ + uint8_t * nodes; + buffer_t prk; + buffer_t info; + buffer_t okm; + + nodes = crypt_secure_malloc(KR_NODES_SZ); + if (nodes == NULL) + return NULL; + + prk.len = SYMMKEYSZ; + prk.data = (uint8_t *) root; + info.len = sizeof(kr_node_label) - 1; + info.data = (uint8_t *) kr_node_label; + okm.len = KR_NODES_SZ; + okm.data = nodes; + + if (crypt_hkdf_expand(prk, info, okm) != 0) + goto fail_expand; + + return nodes; + + fail_expand: + crypt_secure_free(nodes, KR_NODES_SZ); + return NULL; +} + +static int kr_leaf_key(const uint8_t * node, + uint8_t leaf, + uint8_t dir, + uint8_t * out) +{ + uint8_t info_buf[sizeof(kr_leaf_label) - 1 + 2]; + buffer_t prk; + buffer_t info; + buffer_t okm; + size_t n = sizeof(kr_leaf_label) - 1; + + memcpy(info_buf, kr_leaf_label, n); + info_buf[n] = dir; + info_buf[n + 1] = leaf; + + prk.len = SYMMKEYSZ; + prk.data = (uint8_t *) node; + info.len = n + 2; + info.data = info_buf; + okm.len = SYMMKEYSZ; + okm.data = out; + + return crypt_hkdf_expand(prk, info, okm); +} + +static __inline__ bool kr_kc_hit(const struct kr_keycache * kc, + const struct kr_batch * b, + uint16_t node, + uint8_t leaf, + uint8_t dir) +{ + if (!kc->valid) + return false; + + if (kc->id != b->id) + return false; + + if (kc->node != node) + return false; + + if (kc->leaf != leaf) + return false; + + return kc->dir == dir; +} + +/* Fetch the leaf key; derive into the (direct-mapped) slot on a miss. */ +static const uint8_t * kr_kc_get(struct kr_keycache * cache, + const struct kr_batch * b, + uint16_t node, + uint8_t leaf, + uint8_t dir) +{ + struct kr_keycache * kc; + uint8_t * nkey; + + kc = &cache[b->id & (KR_TCACHE_WAYS - 1)]; + + if (kr_kc_hit(kc, b, node, leaf, dir)) + return kc->key; + + nkey = b->nodes + (size_t) node * SYMMKEYSZ; + if (kr_leaf_key(nkey, leaf, dir, kc->key) != 0) + return NULL; + + kc->valid = true; + kc->id = b->id; + kc->node = node; + kc->leaf = leaf; + kc->dir = dir; + + return kc->key; +} + +static void kr_sel_enc(uint8_t epoch, + uint16_t node, + uint32_t seq, + uint8_t sel[KR_SELECTOR_LEN]) +{ + sel[0] = (uint8_t) ((epoch << 4) | ((node >> 8) & 0x0F)); + sel[1] = (uint8_t) (node & 0xFF); + sel[2] = (uint8_t) (seq >> 24); + sel[3] = (uint8_t) (seq >> 16); + sel[4] = (uint8_t) (seq >> 8); + sel[5] = (uint8_t) (seq); +} + +static void kr_sel_dec(const uint8_t sel[KR_SELECTOR_LEN], + uint8_t * epoch, + uint16_t * node, + uint32_t * seq) +{ + *epoch = (uint8_t) (sel[0] >> 4); + *node = (uint16_t) (((sel[0] & 0x0F) << 8) | sel[1]); + *seq = ((uint32_t) sel[2] << 24) | ((uint32_t) sel[3] << 16) | + ((uint32_t) sel[4] << 8) | (uint32_t) sel[5]; +} + +static uint64_t kr_ctr(uint16_t node, + uint32_t seq) +{ + return ((uint64_t) node << KR_WITHIN_BITS) | + ((uint64_t) seq & KR_WITHIN_MASK); +} + +static void kr_nonce(uint64_t ctr, + uint8_t * nonce) +{ + size_t i; + + memset(nonce, 0, KR_NONCE_LEN); + + /* ctr big-endian in the low 8 bytes; high bytes stay zero */ + for (i = 0; i < 8; i++) + nonce[i] = (uint8_t) (ctr >> (56 - 8 * i)); +} + +static struct kr_batch * kr_batch_create(uint8_t epoch, + const uint8_t * root) +{ + struct kr_batch * b; + + b = malloc(sizeof(*b)); + if (b == NULL) + goto fail_alloc; + + b->nodes = kr_expand_nodes(root); + if (b->nodes == NULL) + goto fail_nodes; + + b->id = FETCH_ADD_RELAXED(&kr_g.next_id, 1); + b->epoch = epoch; + b->tx_ctr = 0; + if (pthread_mutex_init(&b->rp.mtx, NULL) != 0) + goto fail_lock; + + b->rp.last = 0; + memset(b->rp.bits, 0, sizeof(b->rp.bits)); + + return b; + + fail_lock: + crypt_secure_free(b->nodes, KR_NODES_SZ); + free(b); + return NULL; + fail_nodes: + free(b); + fail_alloc: + return NULL; +} + +static void kr_batch_free(struct kr_batch * b) +{ + if (b == NULL) + return; + + pthread_mutex_destroy(&b->rp.mtx); + crypt_secure_free(b->nodes, KR_NODES_SZ); + free(b); +} + +/* + * RFC 6479 anti-replay window keyed on the per-batch counter, with + * seq = ctr + 1 so 0 means "nothing accepted yet". Returns 0 if the + * packet is fresh (and records it), -1 on a replay or a too-old ctr. + */ +static int kr_rp_commit(struct kr_batch * b, + uint64_t ctr) +{ + uint64_t seq; + uint64_t idx; + uint64_t cur; + uint64_t diff; + + seq = ctr + 1; + + pthread_mutex_lock(&b->rp.mtx); + + if (seq > b->rp.last) { + idx = seq >> KR_RP_SHIFT; + cur = b->rp.last >> KR_RP_SHIFT; + diff = idx - cur; + if (diff > KR_RP_WORDS) + diff = KR_RP_WORDS; + + while (diff-- > 0) { + cur++; + b->rp.bits[cur & (KR_RP_WORDS - 1)] = 0; + } + + b->rp.bits[idx & (KR_RP_WORDS - 1)] |= + (uint64_t) 1 << (seq & KR_RP_MASK); + b->rp.last = seq; + goto finish; + } + + if (b->rp.last - seq >= KR_RP_WINDOW) + goto fail; + + idx = seq >> KR_RP_SHIFT; + if (b->rp.bits[idx & (KR_RP_WORDS - 1)] + & ((uint64_t) 1 << (seq & KR_RP_MASK))) + goto fail; + + b->rp.bits[idx & (KR_RP_WORDS - 1)] |= + (uint64_t) 1 << (seq & KR_RP_MASK); + finish: + pthread_mutex_unlock(&b->rp.mtx); + + return 0; + fail: + pthread_mutex_unlock(&b->rp.mtx); + + return -1; +} + +struct keyrot * keyrot_create(const uint8_t * root, + uint8_t epoch, + uint8_t role) +{ + struct keyrot * kr; + + assert(root != NULL); + assert(role <= 1); + + if (epoch >= KR_EPOCHS) + goto fail_kr; + + kr = malloc(sizeof(*kr)); + if (kr == NULL) + goto fail_kr; + + memset(kr, 0, sizeof(*kr)); + + kr->role = role; + kr->tx_epoch = epoch; + kr->peer_switched = true; + kr->prev = NULL; + + kr->cur = kr_batch_create(epoch, root); + if (kr->cur == NULL) + goto fail_cur; + + if (rcu_guard_init(&kr->guard)) + goto fail_guard; + + return kr; + + fail_guard: + kr_batch_free(kr->cur); + fail_cur: + free(kr); + fail_kr: + return NULL; +} + +void keyrot_destroy(struct keyrot * kr) +{ + if (kr == NULL) + return; + + /* Wait out any in-flight reader before freeing batches. */ + rcu_drain(&kr->guard); + + kr_batch_free(kr->cur); + kr_batch_free(kr->prev); + + rcu_guard_fini(&kr->guard); + + free(kr); +} + +/* A dup live epoch shadows straggler RX; epoch is peer-driven. */ +static struct kr_batch * kr_live_batch(struct kr_batch * cur, + struct kr_batch * prev, + uint8_t epoch) +{ + if (epoch == cur->epoch) + return cur; + + if (prev != NULL && epoch == prev->epoch) + return prev; + + return NULL; +} + +int keyrot_rekey(struct keyrot * kr, + const uint8_t * root, + uint8_t epoch) +{ + struct kr_batch * nb; + struct kr_batch * old_prev; + struct kr_batch * cur; + struct kr_batch * prev; + struct kr_batch * live; + int ret; + + assert(kr != NULL); + assert(root != NULL); + + if (epoch >= KR_EPOCHS) + return -1; + + nb = kr_batch_create(epoch, root); + if (nb == NULL) + return -1; + + rcu_wrlock(&kr->guard); + + cur = rcu_deref(kr->cur); + prev = rcu_deref(kr->prev); + + live = kr_live_batch(cur, prev, epoch); + if (live != NULL) { + /* The first node key identifies the root. */ + if (crypt_ct_cmp(live->nodes, nb->nodes, SYMMKEYSZ) == 0) + ret = -EREPLAY; + else + ret = -1; + rcu_wrunlock(&kr->guard); + kr_batch_free(nb); + return ret; + } + + old_prev = kr->prev; + rcu_assign(kr->prev, kr->cur); + rcu_publish(nb); + rcu_assign(kr->cur, nb); + + /* TX keeps the old epoch until the peer is seen on the new one. */ + STORE_RELEASE(&kr->peer_switched, false); + + rcu_wrunlock(&kr->guard); + + /* old_prev is unreachable now; reclaim past any live reader. */ + rcu_reclaim(&kr->guard); + kr_batch_free(old_prev); + + return 0; +} + +void keyrot_tx_promote(struct keyrot * kr) +{ + assert(kr != NULL); + + /* Serialise with keyrot_rekey so tx_epoch tracks a consistent cur. */ + rcu_wrlock(&kr->guard); + STORE_RELAXED(&kr->tx_epoch, rcu_deref(kr->cur)->epoch); + rcu_wrunlock(&kr->guard); +} + +int keyrot_tx_next(struct keyrot * kr, + uint8_t sel[KR_SELECTOR_LEN], + const uint8_t ** key, + uint8_t nonce[KR_NONCE_LEN]) +{ + struct kr_tcache * tc; + struct kr_batch * cur; + struct kr_batch * prev; + struct kr_batch * b; + uint64_t ctr; + uint16_t node; + uint8_t leaf; + uint8_t txe; + uint8_t epoch; + uint32_t seq; + const uint8_t * k; + + assert(kr != NULL); + assert(key != NULL); + + tc = kr_tcache_get(); + if (tc == NULL) + return -1; + + rcu_rdlock(&kr->guard); + + cur = rcu_deref(kr->cur); + prev = rcu_deref(kr->prev); + rcu_consume(cur); + rcu_consume(prev); + txe = LOAD_RELAXED(&kr->tx_epoch); + + if (cur->epoch == txe) + b = cur; + else if (prev != NULL && prev->epoch == txe) + b = prev; + else + b = NULL; + + if (b == NULL) { + rcu_rdunlock(&kr->guard); + return -1; /* tx_epoch batch gone; next promote resyncs */ + } + + /* Slot reserved even if exhausted; tx_nodes_left clamps the count. */ + ctr = FETCH_ADD_RELAXED(&b->tx_ctr, 1); + if (ctr >= KR_BATCH_MAX) { + rcu_rdunlock(&kr->guard); + return -1; /* batch exhausted */ + } + + node = (uint16_t) (ctr >> KR_WITHIN_BITS); + leaf = (uint8_t) ((ctr >> KEY_LEAF_BITS) & (KR_LEAVES - 1)); + seq = (uint32_t) (ctr & KR_WITHIN_MASK); + epoch = b->epoch; + + k = kr_kc_get(tc->tx, b, node, leaf, kr->role); + + rcu_rdunlock(&kr->guard); + + if (k == NULL) + return -1; + + kr_sel_enc(epoch, node, seq, sel); + kr_nonce(ctr, nonce); + + *key = k; + + return 0; +} + +int keyrot_rx_lookup(struct keyrot * kr, + const uint8_t sel[KR_SELECTOR_LEN], + const uint8_t ** key, + uint8_t nonce[KR_NONCE_LEN], + struct kr_rx * rx) +{ + struct kr_tcache * tc; + struct kr_batch * cur; + struct kr_batch * prev; + struct kr_batch * b; + uint8_t epoch; + uint16_t node; + uint32_t seq; + uint64_t ctr; + uint8_t leaf; + const uint8_t * k; + + assert(kr != NULL); + assert(key != NULL); + + kr_sel_dec(sel, &epoch, &node, &seq); + + if (node >= KR_N) + return -1; + + tc = kr_tcache_get(); + if (tc == NULL) + return -1; + + rcu_rdlock(&kr->guard); + + cur = rcu_deref(kr->cur); + prev = rcu_deref(kr->prev); + rcu_consume(cur); + rcu_consume(prev); + + if (epoch == cur->epoch) { + b = cur; + } else if (prev != NULL && epoch == prev->epoch) { + b = prev; + } else { + rcu_rdunlock(&kr->guard); + return -1; /* unknown epoch */ + } + + ctr = kr_ctr(node, seq); + leaf = (uint8_t) ((ctr >> KEY_LEAF_BITS) & (KR_LEAVES - 1)); + + /* peer's tx direction */ + k = kr_kc_get(tc->rx, b, node, leaf, (uint8_t) (kr->role ^ 1)); + + rx->id = b->id; + rx->ctr = ctr; + + rcu_rdunlock(&kr->guard); + + if (k == NULL) + return -1; + + kr_nonce(ctr, nonce); + + *key = k; + + return 0; +} + +/* + * Commit a packet that authenticated under the batch keyrot_rx_lookup + * selected. Re-finds that batch by id (epoch may have advanced) and, + * if still resident, advances the replay window and records that the + * peer is on the current batch. Runs only post-AEAD so a forged or + * replayed packet can mutate no receiver state. Returns -1 on replay. + */ +int keyrot_rx_commit(struct keyrot * kr, + const struct kr_rx * rx) +{ + struct kr_batch * cur; + struct kr_batch * prev; + struct kr_batch * b; + int rc; + + assert(kr != NULL); + assert(rx != NULL); + + rcu_rdlock(&kr->guard); + + cur = rcu_deref(kr->cur); + prev = rcu_deref(kr->prev); + rcu_consume(cur); + rcu_consume(prev); + + if (cur->id == rx->id) + b = cur; + else if (prev != NULL && prev->id == rx->id) + b = prev; + else + b = NULL; + + if (b == NULL) { + rcu_rdunlock(&kr->guard); + return 0; /* batch evicted post-auth; nothing to protect */ + } + + rc = kr_rp_commit(b, rx->ctr); + if (rc == 0 && b == cur) + STORE_RELEASE(&kr->peer_switched, true); + + rcu_rdunlock(&kr->guard); + + return rc; +} + +bool keyrot_peer_switched(const struct keyrot * kr) +{ + assert(kr != NULL); + + return LOAD_ACQUIRE(&kr->peer_switched); +} + +unsigned keyrot_tx_nodes_left(struct keyrot * kr) +{ + struct kr_batch * cur; + struct kr_batch * prev; + struct kr_batch * b; + uint64_t ctr; + unsigned used; + uint8_t txe; + + assert(kr != NULL); + + rcu_rdlock(&kr->guard); + cur = rcu_deref(kr->cur); + prev = rcu_deref(kr->prev); + rcu_consume(cur); + rcu_consume(prev); + txe = LOAD_RELAXED(&kr->tx_epoch); + + if (cur->epoch == txe) + b = cur; + else if (prev != NULL && prev->epoch == txe) + b = prev; + else + b = NULL; + + ctr = b != NULL ? LOAD_RELAXED(&b->tx_ctr) : KR_BATCH_MAX; + rcu_rdunlock(&kr->guard); + + used = (unsigned) (ctr >> KR_WITHIN_BITS); + if (used >= KR_N) + return 0; + + return KR_N - used; +} diff --git a/src/lib/crypt/keyrot.h b/src/lib/crypt/keyrot.h new file mode 100644 index 00000000..6a598f76 --- /dev/null +++ b/src/lib/crypt/keyrot.h @@ -0,0 +1,74 @@ +/* + * Ouroboros - Copyright (C) 2016 - 2026 + * + * Data-plane key-rotation schedule (node/leaf keys, selector) + * + * Dimitri Staessens <dimitri@ouroboros.rocks> + * Sander Vrijders <sander@ouroboros.rocks> + * + * This library is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public License + * version 2.1 as published by the Free Software Foundation. + * + * This library is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with this library; if not, write to the Free Software + * Foundation, Inc., http://www.fsf.org/about/contact/. + */ + +#ifndef OUROBOROS_LIB_CRYPT_KEYROT_H +#define OUROBOROS_LIB_CRYPT_KEYROT_H + +#include <ouroboros/crypt.h> /* SYMMKEYSZ, NONCESZ */ + +#include <stdbool.h> +#include <stdint.h> + +#define KR_SELECTOR_LEN 6 +#define KR_NONCE_LEN NONCESZ + +struct keyrot; + +struct kr_rx { + uint64_t id; /* batch id of the matched epoch */ + uint64_t ctr; /* packet counter for replay check */ +}; + +struct keyrot * keyrot_create(const uint8_t * root, + uint8_t epoch, + uint8_t role); + +void keyrot_destroy(struct keyrot * kr); + +int keyrot_rekey(struct keyrot * kr, + const uint8_t * root, + uint8_t epoch); + +/* Promote TX to the installed (new) batch once the peer is on it. */ +void keyrot_tx_promote(struct keyrot * kr); + +int keyrot_tx_next(struct keyrot * kr, + uint8_t sel[KR_SELECTOR_LEN], + const uint8_t ** key, + uint8_t nonce[KR_NONCE_LEN]); + +int keyrot_rx_lookup(struct keyrot * kr, + const uint8_t sel[KR_SELECTOR_LEN], + const uint8_t ** key, + uint8_t nonce[KR_NONCE_LEN], + struct kr_rx * rx); + +/* Commit an authenticated packet: replay window + peer-switched. */ +int keyrot_rx_commit(struct keyrot * kr, + const struct kr_rx * rx); + +/* True once an RX packet under the current batch has been observed. */ +bool keyrot_peer_switched(const struct keyrot * kr); + +unsigned keyrot_tx_nodes_left(struct keyrot * kr); + +#endif /* OUROBOROS_LIB_CRYPT_KEYROT_H */ diff --git a/src/lib/crypt/openssl.c b/src/lib/crypt/openssl.c index 573bc0b3..d5d9adf5 100644 --- a/src/lib/crypt/openssl.c +++ b/src/lib/crypt/openssl.c @@ -30,6 +30,8 @@ #include <ouroboros/errno.h> #include <ouroboros/crypt.h> #include <ouroboros/hash.h> +#include <ouroboros/name.h> +#include <ouroboros/pthread.h> #include <ouroboros/random.h> #include <ouroboros/utils.h> @@ -52,27 +54,14 @@ #define HKDF_INFO_DHE "o7s-ossl-dhe" #define HKDF_INFO_ENCAP "o7s-ossl-encap" -#define HKDF_INFO_ROTATION "o7s-key-rotation" #define HKDF_SALT_LEN 32 /* SHA-256 output size */ +#define AEAD_NONCE_LEN 12 /* 96-bit deterministic IV (SP 800-38D) */ +#define AEAD_TAG_LEN 16 /* 128-bit AEAD authentication tag */ struct ossl_crypt_ctx { EVP_CIPHER_CTX * evp_ctx; const EVP_CIPHER * cipher; - int ivsz; int tagsz; - - struct { - uint8_t * cur; /* current key */ - uint8_t * prv; /* rotated key */ - } keys; - - struct { - uint32_t cntr; /* counter */ - uint32_t mask; /* phase mask */ - uint32_t age; /* counter within epoch */ - uint8_t phase; /* current key phase */ - uint8_t salt[HKDF_SALT_LEN]; - } rot; /* rotation logic */ }; struct kdf_info { @@ -83,17 +72,6 @@ struct kdf_info { buffer_t key; }; -/* Key rotation macros */ -#define HAS_PHASE_BIT_TOGGLED(ctx) \ - (((ctx)->rot.cntr & (ctx)->rot.mask) != \ - (((ctx)->rot.cntr - 1) & (ctx)->rot.mask)) - -#define HAS_GRACE_EXPIRED(ctx) \ - ((ctx)->rot.age >= ((ctx)->rot.mask >> 1)) - -#define ROTATION_TOO_RECENT(ctx) \ - ((ctx)->rot.age < ((ctx)->rot.mask - ((ctx)->rot.mask >> 2))) - /* Convert hash NID to OpenSSL digest name string for HKDF */ static const char * hash_nid_to_digest_name(int nid) { @@ -102,11 +80,11 @@ static const char * hash_nid_to_digest_name(int nid) md = EVP_get_digestbynid(nid); if (md == NULL) - return "SHA256"; /* fallback to SHA-256 */ + return NULL; name = EVP_MD_get0_name(md); if (name == NULL) - return "SHA256"; /* fallback to SHA-256 */ + return NULL; return name; } @@ -144,21 +122,20 @@ static int get_pk_bytes_from_key(EVP_PKEY * key, } /* Derive salt from public key bytes by hashing them */ -static int derive_salt_from_pk_bytes(buffer_t pk, - uint8_t * salt, - size_t salt_len) +static int derive_salt_from_pk_bytes(buffer_t pk, + buffer_t salt) { uint8_t hash[EVP_MAX_MD_SIZE]; unsigned hash_len; assert(pk.data != NULL); - assert(salt != NULL); + assert(salt.data != NULL); if (EVP_Digest(pk.data, pk.len, hash, &hash_len, EVP_sha256(), NULL) != 1) goto fail_digest; - memcpy(salt, hash, salt_len < hash_len ? salt_len : hash_len); + memcpy(salt.data, hash, salt.len < hash_len ? salt.len : hash_len); return 0; fail_digest: @@ -166,10 +143,9 @@ static int derive_salt_from_pk_bytes(buffer_t pk, } /* Derive salt from two public key byte buffers (DHE) in canonical order */ -static int derive_salt_from_pk_bytes_dhe(buffer_t local, - buffer_t remote, - uint8_t * salt, - size_t salt_len) +static int derive_salt_from_pk_bytes_dhe(buffer_t local, + buffer_t remote, + buffer_t salt) { uint8_t * concat; size_t concat_len; @@ -180,7 +156,7 @@ static int derive_salt_from_pk_bytes_dhe(buffer_t local, assert(local.data != NULL); assert(remote.data != NULL); - assert(salt != NULL); + assert(salt.data != NULL); concat_len = local.len + remote.len; concat = OPENSSL_malloc(concat_len); @@ -204,7 +180,7 @@ static int derive_salt_from_pk_bytes_dhe(buffer_t local, OPENSSL_free(concat); - memcpy(salt, hash, salt_len < hash_len ? salt_len : hash_len); + memcpy(salt.data, hash, salt.len < hash_len ? salt.len : hash_len); return 0; fail_digest: @@ -225,6 +201,8 @@ static int derive_key_hkdf(struct kdf_info * ki) int idx; digest = hash_nid_to_digest_name(ki->nid); + if (digest == NULL) + goto fail_fetch; kdf = EVP_KDF_fetch(NULL, "HKDF", NULL); if (kdf == NULL) @@ -258,117 +236,144 @@ static int derive_key_hkdf(struct kdf_info * ki) return -ECRYPT; } -/* Key rotation helper functions implementation */ -static int should_rotate_key_rx(struct ossl_crypt_ctx * ctx, - uint8_t rx_phase) +int openssl_hkdf_expand(buffer_t key, + buffer_t info, + buffer_t out) { - assert(ctx != NULL); + EVP_KDF * kdf; + EVP_KDF_CTX * kctx; + OSSL_PARAM params[5]; + int mode = EVP_KDF_HKDF_MODE_EXPAND_ONLY; + int idx = 0; + int ret = -1; - /* Phase must have changed */ - if (rx_phase == ctx->rot.phase) - return 0; + kdf = EVP_KDF_fetch(NULL, "HKDF", NULL); + if (kdf == NULL) + goto fail_fetch; - if (ROTATION_TOO_RECENT(ctx)) - return 0; + kctx = EVP_KDF_CTX_new(kdf); + if (kctx == NULL) + goto fail_ctx; - return 1; + params[idx++] = OSSL_PARAM_construct_utf8_string( + "digest", (char *) "SHA256", 0); + params[idx++] = OSSL_PARAM_construct_int("mode", &mode); + params[idx++] = OSSL_PARAM_construct_octet_string( + "key", key.data, key.len); + params[idx++] = OSSL_PARAM_construct_octet_string( + "info", info.data, info.len); + params[idx] = OSSL_PARAM_construct_end(); + + if (EVP_KDF_derive(kctx, out.data, out.len, params) == 1) + ret = 0; + + EVP_KDF_CTX_free(kctx); + fail_ctx: + EVP_KDF_free(kdf); + fail_fetch: + return ret; } -static int rotate_key(struct ossl_crypt_ctx * ctx) +/* AEAD seal: encrypt in with key/nonce, bind aad, append tag */ +int openssl_seal(struct ossl_crypt_ctx * ctx, + const uint8_t * key, + const uint8_t * nonce, + buffer_t aad, + buffer_t in, + uint8_t * out, + uint8_t * tag) { - struct kdf_info ki; - uint8_t * tmp; + int out_sz; + int tmp_sz; assert(ctx != NULL); + assert(ctx->tagsz > 0); /* AEAD mandated at ctx creation */ - /* Swap keys - move current to prev */ - tmp = ctx->keys.prv; - ctx->keys.prv = ctx->keys.cur; + EVP_CIPHER_CTX_reset(ctx->evp_ctx); - if (tmp != NULL) { - /* Reuse old prev_key memory for new key */ - ctx->keys.cur = tmp; - } else { - /* First rotation - allocate new memory */ - ctx->keys.cur = OPENSSL_secure_malloc(SYMMKEYSZ); - if (ctx->keys.cur == NULL) - return -ECRYPT; - } + if (EVP_EncryptInit_ex(ctx->evp_ctx, ctx->cipher, NULL, + NULL, NULL) != 1) + return -1; - /* Derive new key from previous key using HKDF */ - ki.secret.data = ctx->keys.prv; - ki.secret.len = SYMMKEYSZ; - ki.nid = NID_sha256; - ki.salt.data = ctx->rot.salt; - ki.salt.len = HKDF_SALT_LEN; - ki.info.data = (uint8_t *) HKDF_INFO_ROTATION; - ki.info.len = strlen(HKDF_INFO_ROTATION); - ki.key.data = ctx->keys.cur; - ki.key.len = SYMMKEYSZ; + /* Pin the AEAD nonce to 96 bits (SP 800-38D deterministic IV). */ + if (EVP_CIPHER_CTX_ctrl(ctx->evp_ctx, EVP_CTRL_AEAD_SET_IVLEN, + AEAD_NONCE_LEN, NULL) != 1) + return -1; - if (derive_key_hkdf(&ki) != 0) - return -ECRYPT; + if (EVP_EncryptInit_ex(ctx->evp_ctx, NULL, NULL, + key, nonce) != 1) + return -1; - ctx->rot.age = 0; - ctx->rot.phase = !ctx->rot.phase; + if (EVP_EncryptUpdate(ctx->evp_ctx, NULL, &tmp_sz, + aad.data, (int) aad.len) != 1) + return -1; - return 0; -} + if (EVP_EncryptUpdate(ctx->evp_ctx, out, &out_sz, + in.data, (int) in.len) != 1) + return -1; -static void cleanup_old_key(struct ossl_crypt_ctx * ctx) -{ - assert(ctx != NULL); + if (EVP_EncryptFinal_ex(ctx->evp_ctx, out + out_sz, &tmp_sz) != 1) + return -1; - if (ctx->keys.prv == NULL) - return; + out_sz += tmp_sz; - if (!HAS_GRACE_EXPIRED(ctx)) - return; + if (EVP_CIPHER_CTX_ctrl(ctx->evp_ctx, EVP_CTRL_AEAD_GET_TAG, + ctx->tagsz, tag) != 1) + return -1; - OPENSSL_secure_clear_free(ctx->keys.prv, SYMMKEYSZ); - ctx->keys.prv = NULL; + return out_sz; } -static int try_decrypt(struct ossl_crypt_ctx * ctx, - uint8_t * key, - uint8_t * iv, - uint8_t * input, - int in_sz, - uint8_t * out, - int * out_sz) +/* AEAD open: decrypt in with key/nonce, verify aad and tag */ +int openssl_open(struct ossl_crypt_ctx * ctx, + const uint8_t * key, + const uint8_t * nonce, + buffer_t aad, + buffer_t in, + const uint8_t * tag, + buffer_t * out) { - uint8_t * tag; - int tmp_sz; - int ret; + int out_sz; + int tmp_sz; - tag = input + in_sz; + assert(ctx != NULL); + assert(ctx->tagsz > 0); /* AEAD mandated at ctx creation */ EVP_CIPHER_CTX_reset(ctx->evp_ctx); - ret = EVP_DecryptInit_ex(ctx->evp_ctx, ctx->cipher, NULL, key, iv); - if (ret != 1) + if (EVP_DecryptInit_ex(ctx->evp_ctx, ctx->cipher, NULL, + NULL, NULL) != 1) return -1; - if (ctx->tagsz > 0) { - ret = EVP_CIPHER_CTX_ctrl(ctx->evp_ctx, EVP_CTRL_AEAD_SET_TAG, - ctx->tagsz, tag); - if (ret != 1) - return -1; - } + /* Pin the AEAD nonce to 96 bits (SP 800-38D deterministic IV). */ + if (EVP_CIPHER_CTX_ctrl(ctx->evp_ctx, EVP_CTRL_AEAD_SET_IVLEN, + AEAD_NONCE_LEN, NULL) != 1) + return -1; - ret = EVP_DecryptUpdate(ctx->evp_ctx, out, &tmp_sz, input, in_sz); - if (ret != 1) + if (EVP_DecryptInit_ex(ctx->evp_ctx, NULL, NULL, key, nonce) != 1) return -1; - *out_sz = tmp_sz; + if (EVP_CIPHER_CTX_ctrl(ctx->evp_ctx, EVP_CTRL_AEAD_SET_TAG, + ctx->tagsz, (void *) tag) != 1) + return -1; - ret = EVP_DecryptFinal_ex(ctx->evp_ctx, out + tmp_sz, &tmp_sz); - if (ret != 1) + if (EVP_DecryptUpdate(ctx->evp_ctx, NULL, &tmp_sz, + aad.data, (int) aad.len) != 1) return -1; - *out_sz += tmp_sz; + if (EVP_DecryptUpdate(ctx->evp_ctx, out->data, &out_sz, + in.data, (int) in.len) != 1) + return -1; - return 0; + if (EVP_DecryptFinal_ex(ctx->evp_ctx, out->data + out_sz, + &tmp_sz) != 1) + return -1; + + out_sz += tmp_sz; + + out->len = (size_t) out_sz; + + return out_sz; } /* @@ -396,11 +401,14 @@ static int __openssl_dhe_derive(EVP_PKEY * pkp, ret = i2d_PUBKEY(pkp, &local_pk.data); if (ret <= 0) goto fail_local; + local_pk.len = (size_t) ret; + ki.salt.len = HKDF_SALT_LEN; + ki.salt.data = salt_buf; + /* Derive salt from both public keys */ - if (derive_salt_from_pk_bytes_dhe(local_pk, remote_pk, salt_buf, - HKDF_SALT_LEN) < 0) + if (derive_salt_from_pk_bytes_dhe(local_pk, remote_pk, ki.salt) < 0) goto fail_salt; ctx = EVP_PKEY_CTX_new(pkp, NULL); @@ -437,13 +445,11 @@ static int __openssl_dhe_derive(EVP_PKEY * pkp, ki.info.data = (uint8_t *) HKDF_INFO_DHE; ki.key.len = SYMMKEYSZ; ki.key.data = s; - ki.salt.len = HKDF_SALT_LEN; - ki.salt.data = salt_buf; /* Derive symmetric key from shared secret using HKDF */ ret = derive_key_hkdf(&ki); - OPENSSL_free(secret); + OPENSSL_clear_free(secret, secret_len); EVP_PKEY_CTX_free(ctx); OPENSSL_free(local_pk.data); @@ -452,7 +458,7 @@ static int __openssl_dhe_derive(EVP_PKEY * pkp, return 0; fail_derive: - OPENSSL_free(secret); + OPENSSL_clear_free(secret, secret_len); fail_ctx: EVP_PKEY_CTX_free(ctx); fail_salt: @@ -624,14 +630,22 @@ ssize_t openssl_pkp_create(const char * algo, if (raw.len == 0) goto fail_pubkey; + if (raw.len > CRYPT_KEY_BUFSZ) { + OPENSSL_free(raw.data); + goto fail_pubkey; + } + memcpy(pk, raw.data, raw.len); OPENSSL_free(raw.data); return (ssize_t) raw.len; } else { /* DER encode standard algorithms */ - pos = pk; /* i2d_PUBKEY increments the pointer, don't use pk! */ - len = i2d_PUBKEY(*pkp, &pos); - if (len < 0) + len = i2d_PUBKEY(*pkp, NULL); /* pre-flight length */ + if (len < 0 || len > CRYPT_KEY_BUFSZ) + goto fail_pubkey; + + pos = pk; /* i2d_PUBKEY increments the ptr, don't use pk! */ + if (i2d_PUBKEY(*pkp, &pos) < 0) goto fail_pubkey; return len; @@ -666,7 +680,7 @@ static ssize_t __openssl_kem_encap(EVP_PKEY * pub, /* Get required lengths */ ret = EVP_PKEY_encapsulate(ctx, NULL, &ct_len, NULL, &secret_len); - if (ret != 1 || ct_len > MSGBUFSZ) + if (ret != 1 || ct_len > CRYPT_KEY_BUFSZ) goto fail_encap; /* Allocate buffer for secret */ @@ -692,7 +706,7 @@ static ssize_t __openssl_kem_encap(EVP_PKEY * pub, /* Derive symmetric key from shared secret using HKDF */ ret = derive_key_hkdf(&ki); - OPENSSL_free(secret); + OPENSSL_clear_free(secret, secret_len); EVP_PKEY_CTX_free(ctx); if (ret != 0) @@ -701,7 +715,7 @@ static ssize_t __openssl_kem_encap(EVP_PKEY * pub, return (ssize_t) ct_len; fail_secret: - OPENSSL_free(secret); + OPENSSL_clear_free(secret, secret_len); fail_encap: EVP_PKEY_CTX_free(ctx); fail_ctx: @@ -717,13 +731,17 @@ ssize_t openssl_kem_encap(buffer_t pk, EVP_PKEY * pub; uint8_t * pos; uint8_t salt[HKDF_SALT_LEN]; + buffer_t salt_b; ssize_t ret; assert(pk.data != NULL); assert(ct != NULL); assert(s != NULL); - if (derive_salt_from_pk_bytes(pk, salt, HKDF_SALT_LEN) < 0) + salt_b.len = HKDF_SALT_LEN; + salt_b.data = salt; + + if (derive_salt_from_pk_bytes(pk, salt_b) < 0) goto fail_salt; pos = pk.data; @@ -749,13 +767,17 @@ ssize_t openssl_kem_encap_raw(buffer_t pk, EVP_PKEY * pub; const char * algo; uint8_t salt[HKDF_SALT_LEN]; + buffer_t salt_b; ssize_t ret; assert(pk.data != NULL); assert(ct != NULL); assert(s != NULL); - if (derive_salt_from_pk_bytes(pk, salt, HKDF_SALT_LEN) < 0) + salt_b.len = HKDF_SALT_LEN; + salt_b.data = salt; + + if (derive_salt_from_pk_bytes(pk, salt_b) < 0) goto fail_salt; algo = __openssl_hybrid_algo_from_len(pk.len); @@ -789,12 +811,16 @@ int openssl_kem_decap(EVP_PKEY * priv, size_t secret_len; int ret; uint8_t salt[HKDF_SALT_LEN]; + buffer_t salt_b; /* Extract public key bytes from private key */ if (get_pk_bytes_from_key(priv, &pk) < 0) goto fail_pk; - if (derive_salt_from_pk_bytes(pk, salt, HKDF_SALT_LEN) < 0) + salt_b.len = HKDF_SALT_LEN; + salt_b.data = salt; + + if (derive_salt_from_pk_bytes(pk, salt_b) < 0) goto fail_salt; ctx = EVP_PKEY_CTX_new(priv, NULL); @@ -833,7 +859,7 @@ int openssl_kem_decap(EVP_PKEY * priv, /* Derive symmetric key from shared secret using HKDF */ ret = derive_key_hkdf(&ki); - OPENSSL_free(secret); + OPENSSL_clear_free(secret, secret_len); EVP_PKEY_CTX_free(ctx); OPENSSL_free(pk.data); @@ -843,7 +869,7 @@ int openssl_kem_decap(EVP_PKEY * priv, return 0; fail_secret: - OPENSSL_free(secret); + OPENSSL_clear_free(secret, secret_len); fail_ctx: EVP_PKEY_CTX_free(ctx); fail_salt: @@ -857,13 +883,14 @@ void openssl_pkp_destroy(EVP_PKEY * pkp) EVP_PKEY_free(pkp); } -int __openssl_get_curve(EVP_PKEY * pub, - char * algo) +static int openssl_get_curve(EVP_PKEY * pub, + char * algo) { int ret; size_t len = KEX_ALGO_BUFSZ; ret = EVP_PKEY_get_utf8_string_param(pub, "group", algo, len, &len); + return ret == 1 ? 0 : -ECRYPT; } @@ -888,9 +915,10 @@ int openssl_get_algo_from_pk_der(buffer_t pk, strcpy(algo, type_str); - if ((IS_EC_GROUP(algo) || IS_DH_GROUP(algo)) && - __openssl_get_curve(pub, algo) < 0) - goto fail_pub; + if (IS_EC_GROUP(algo) || IS_DH_GROUP(algo)) { + if (openssl_get_curve(pub, algo) < 0) + goto fail_pub; + } EVP_PKEY_free(pub); return 0; @@ -948,141 +976,122 @@ int openssl_dhe_derive(EVP_PKEY * pkp, return -ECRYPT; } -int openssl_encrypt(struct ossl_crypt_ctx * ctx, - buffer_t in, - buffer_t * out) +/* Set up a fresh AEAD cipher ctx for nid: reject non-AEAD / oversized IV. */ +static int ossl_cipher_ctx_init(struct ossl_crypt_ctx * ctx, + int nid) { - uint8_t * ptr; - uint8_t * iv; - int in_sz; - int out_sz; - int tmp_sz; - int ret; - - assert(ctx != NULL); - - in_sz = (int) in.len; - - out->data = malloc(in.len + EVP_MAX_BLOCK_LENGTH + \ - ctx->ivsz + ctx->tagsz); - if (out->data == NULL) - goto fail_malloc; - - iv = out->data; - ptr = out->data + ctx->ivsz; + ctx->cipher = EVP_get_cipherbynid(nid); + if (ctx->cipher == NULL) + return -1; - if (random_buffer(iv, ctx->ivsz) < 0) - goto fail_encrypt; + /* IV must fit the NONCESZ nonce buffer. */ + if (EVP_CIPHER_get_iv_length(ctx->cipher) > NONCESZ) + return -1; - /* Set IV bit 7 to current key phase (KEY_ROTATION_BIT of counter) */ - if (ctx->rot.cntr & ctx->rot.mask) - iv[0] |= 0x80; - else - iv[0] &= 0x7F; + /* Authenticated encryption is mandatory; reject non-AEAD ciphers. */ + if ((EVP_CIPHER_flags(ctx->cipher) & EVP_CIPH_FLAG_AEAD_CIPHER) == 0) + return -1; - EVP_CIPHER_CTX_reset(ctx->evp_ctx); + ctx->tagsz = AEAD_TAG_LEN; - ret = EVP_EncryptInit_ex(ctx->evp_ctx, ctx->cipher, NULL, - ctx->keys.cur, iv); - if (ret != 1) - goto fail_encrypt; + ctx->evp_ctx = EVP_CIPHER_CTX_new(); + if (ctx->evp_ctx == NULL) + return -1; - ret = EVP_EncryptUpdate(ctx->evp_ctx, ptr, &tmp_sz, in.data, in_sz); - if (ret != 1) - goto fail_encrypt; + return 0; +} - out_sz = tmp_sz; - ret = EVP_EncryptFinal_ex(ctx->evp_ctx, ptr + tmp_sz, &tmp_sz); - if (ret != 1) - goto fail_encrypt; +/* One-shot AEAD seal over an explicit key/nonce (no keyrot). out = ct ‖ tag. */ +int openssl_oneshot_seal(int nid, + const uint8_t * key, + const uint8_t * nonce, + buffer_t aad, + buffer_t in, + buffer_t * out) +{ + struct ossl_crypt_ctx ctx; + int out_sz; - out_sz += tmp_sz; + assert(key != NULL); + assert(nonce != NULL); + assert(out != NULL); - /* For AEAD ciphers, get and append the authentication tag */ - if (ctx->tagsz > 0) { - ret = EVP_CIPHER_CTX_ctrl(ctx->evp_ctx, EVP_CTRL_AEAD_GET_TAG, - ctx->tagsz, ptr + out_sz); - if (ret != 1) - goto fail_encrypt; - out_sz += ctx->tagsz; - } + memset(&ctx, 0, sizeof(ctx)); - assert(out_sz >= in_sz); + if (ossl_cipher_ctx_init(&ctx, nid) < 0) + goto fail_cipher; - out->len = (size_t) out_sz + ctx->ivsz; + out->data = malloc(in.len + EVP_MAX_BLOCK_LENGTH + ctx.tagsz); + if (out->data == NULL) + goto fail_ctx; - /* Increment packet counter and check for key rotation */ - ctx->rot.cntr++; - ctx->rot.age++; + out_sz = openssl_seal(&ctx, key, nonce, aad, in, + out->data, out->data + in.len); + if (out_sz < 0) + goto fail_seal; - if (HAS_PHASE_BIT_TOGGLED(ctx)) { - if (rotate_key(ctx) != 0) - goto fail_encrypt; - } + out->len = (size_t) out_sz + ctx.tagsz; - cleanup_old_key(ctx); + EVP_CIPHER_CTX_free(ctx.evp_ctx); return 0; - fail_encrypt: + + fail_seal: free(out->data); - fail_malloc: + fail_ctx: + EVP_CIPHER_CTX_free(ctx.evp_ctx); + fail_cipher: clrbuf(*out); return -ECRYPT; } -int openssl_decrypt(struct ossl_crypt_ctx * ctx, - buffer_t in, - buffer_t * out) +/* One-shot AEAD open; in = ct ‖ tag, verifies aad and tag. */ +int openssl_oneshot_open(int nid, + const uint8_t * key, + const uint8_t * nonce, + buffer_t aad, + buffer_t in, + buffer_t * out) { - uint8_t * iv; - uint8_t * input; - uint8_t rx_phase; - int out_sz; - int in_sz; - - assert(ctx != NULL); + struct ossl_crypt_ctx ctx; + buffer_t ct; + const uint8_t * tag; + int in_sz; - in_sz = (int) in.len - ctx->ivsz; - if (in_sz < ctx->tagsz) - return -ECRYPT; - - in_sz -= ctx->tagsz; + assert(key != NULL); + assert(nonce != NULL); + assert(out != NULL); - out->data = malloc(in_sz + EVP_MAX_BLOCK_LENGTH); - if (out->data == NULL) - goto fail_malloc; + memset(&ctx, 0, sizeof(ctx)); - iv = in.data; - input = in.data + ctx->ivsz; + if (ossl_cipher_ctx_init(&ctx, nid) < 0) + goto fail_cipher; - /* Extract phase from IV bit 7 and check for key rotation */ - rx_phase = (iv[0] & 0x80) ? 1 : 0; + if (in.len < (size_t) ctx.tagsz) + goto fail_ctx; - if (should_rotate_key_rx(ctx, rx_phase)) { - if (rotate_key(ctx) != 0) - goto fail_decrypt; - } + in_sz = (int) in.len - ctx.tagsz; - ctx->rot.cntr++; - ctx->rot.age++; + out->data = malloc((size_t) in_sz + EVP_MAX_BLOCK_LENGTH); + if (out->data == NULL) + goto fail_ctx; - if (try_decrypt(ctx, ctx->keys.cur, iv, input, in_sz, out->data, - &out_sz) != 0) { - if (ctx->keys.prv == NULL) - goto fail_decrypt; - if (try_decrypt(ctx, ctx->keys.prv, iv, input, in_sz, - out->data, &out_sz) != 0) - goto fail_decrypt; - } + ct.data = in.data; + ct.len = (size_t) in_sz; + tag = in.data + in_sz; - assert(out_sz <= in_sz); + if (openssl_open(&ctx, key, nonce, aad, ct, tag, out) < 0) + goto fail_open; - out->len = (size_t) out_sz; + EVP_CIPHER_CTX_free(ctx.evp_ctx); return 0; - fail_decrypt: + + fail_open: free(out->data); - fail_malloc: + fail_ctx: + EVP_CIPHER_CTX_free(ctx.evp_ctx); + fail_cipher: clrbuf(*out); return -ECRYPT; } @@ -1093,51 +1102,19 @@ struct ossl_crypt_ctx * openssl_crypt_create_ctx(struct crypt_sk * sk) assert(sk != NULL); assert(sk->key != NULL); - assert(sk->rot_bit > 0 && sk->rot_bit < 32); ctx = malloc(sizeof(*ctx)); if (ctx == NULL) - goto fail_malloc; + goto fail_malloc; memset(ctx, 0, sizeof(*ctx)); - ctx->keys.cur = OPENSSL_secure_malloc(SYMMKEYSZ); - if (ctx->keys.cur == NULL) - goto fail_key; - - memcpy(ctx->keys.cur, sk->key, SYMMKEYSZ); - - ctx->keys.prv = NULL; - - /* Derive rotation salt from initial shared secret */ - if (EVP_Digest(sk->key, SYMMKEYSZ, ctx->rot.salt, NULL, - EVP_sha256(), NULL) != 1) - goto fail_cipher; - - ctx->cipher = EVP_get_cipherbynid(sk->nid); - if (ctx->cipher == NULL) - goto fail_cipher; - - ctx->ivsz = EVP_CIPHER_iv_length(ctx->cipher); - - /* Set tag size for AEAD ciphers (GCM, CCM, OCB, ChaCha20-Poly1305) */ - if (EVP_CIPHER_flags(ctx->cipher) & EVP_CIPH_FLAG_AEAD_CIPHER) - ctx->tagsz = 16; /* Standard AEAD tag length (128 bits) */ - - ctx->rot.cntr = 0; - ctx->rot.mask = (1U << sk->rot_bit); - ctx->rot.age = 0; - ctx->rot.phase = 0; - - ctx->evp_ctx = EVP_CIPHER_CTX_new(); - if (ctx->evp_ctx == NULL) + if (ossl_cipher_ctx_init(ctx, sk->nid) < 0) goto fail_cipher; return ctx; fail_cipher: - OPENSSL_secure_clear_free(ctx->keys.cur, SYMMKEYSZ); - fail_key: free(ctx); fail_malloc: return NULL; @@ -1148,23 +1125,10 @@ void openssl_crypt_destroy_ctx(struct ossl_crypt_ctx * ctx) if (ctx == NULL) return; - if (ctx->keys.cur != NULL) - OPENSSL_secure_clear_free(ctx->keys.cur, SYMMKEYSZ); - - if (ctx->keys.prv != NULL) - OPENSSL_secure_clear_free(ctx->keys.prv, SYMMKEYSZ); - EVP_CIPHER_CTX_free(ctx->evp_ctx); free(ctx); } -int openssl_crypt_get_ivsz(struct ossl_crypt_ctx * ctx) -{ - assert(ctx != NULL); - - return ctx->ivsz; -} - int openssl_crypt_get_tagsz(struct ossl_crypt_ctx * ctx) { assert(ctx != NULL); @@ -1184,7 +1148,12 @@ int openssl_load_crt_file(const char * path, if (fp == NULL) goto fail_file; + pthread_cleanup_push(__cleanup_fclose, fp); + xcrt = PEM_read_X509(fp, NULL, NULL, NULL); + + pthread_cleanup_pop(false); + if (xcrt == NULL) goto fail_crt; @@ -1200,35 +1169,58 @@ int openssl_load_crt_file(const char * path, return -1; } -int openssl_load_crt_str(const char * str, - void ** crt) +static void * rd_crt_bio(BIO * bio) +{ + return PEM_read_bio_X509(bio, NULL, NULL, NULL); +} + +static void * rd_privkey_bio(BIO * bio) +{ + return PEM_read_bio_PrivateKey(bio, NULL, NULL, ""); +} + +static void * rd_pubkey_bio(BIO * bio) +{ + return PEM_read_bio_PUBKEY(bio, NULL, NULL, NULL); +} + +/* Decode a PEM object from an in-memory string via rd. */ +static int load_pem_str(const char * str, + void * (* rd)(BIO *), + void ** out) { BIO * bio; - X509 * xcrt; + void * obj; bio = BIO_new(BIO_s_mem()); if (bio == NULL) goto fail_bio; if (BIO_write(bio, str, strlen(str)) < 0) - goto fail_crt; + goto fail_obj; - xcrt = PEM_read_bio_X509(bio, NULL, NULL, NULL); - if (xcrt == NULL) - goto fail_crt; + obj = rd(bio); + if (obj == NULL) + goto fail_obj; BIO_free(bio); - *crt = (void *) xcrt; + *out = obj; return 0; - fail_crt: + fail_obj: BIO_free(bio); fail_bio: - *crt = NULL; + *out = NULL; return -1; } +int openssl_load_crt_str(const char * str, + void ** crt) +{ + return load_pem_str(str, rd_crt_bio, crt); +} + int openssl_load_crt_der(buffer_t buf, void ** crt) { @@ -1283,24 +1275,19 @@ int openssl_load_privkey_file(const char * path, { FILE * fp; EVP_PKEY * pkey; - unsigned long err; - char errbuf[256]; fp = fopen(path, "r"); - if (fp == NULL) { - fprintf(stderr, "Failed to open %s\n", path); + if (fp == NULL) goto fail_file; - } + + pthread_cleanup_push(__cleanup_fclose, fp); pkey = PEM_read_PrivateKey(fp, NULL, NULL, ""); - if (pkey == NULL) { - err = ERR_get_error(); - ERR_error_string_n(err, errbuf, sizeof(errbuf)); - fprintf(stderr, - "OpenSSL error loading privkey from %s: %s\n", - path, errbuf); + + pthread_cleanup_pop(false); + + if (pkey == NULL) goto fail_key; - } fclose(fp); @@ -1317,30 +1304,7 @@ int openssl_load_privkey_file(const char * path, int openssl_load_privkey_str(const char * str, void ** key) { - BIO * bio; - EVP_PKEY * pkey; - - bio = BIO_new(BIO_s_mem()); - if (bio == NULL) - goto fail_bio; - - if (BIO_write(bio, str, strlen(str)) < 0) - goto fail_key; - - pkey = PEM_read_bio_PrivateKey(bio, NULL, NULL, NULL); - if (pkey == NULL) - goto fail_key; - - BIO_free(bio); - - *key = (void *) pkey; - - return 0; - fail_key: - BIO_free(bio); - fail_bio: - *key = NULL; - return -1; + return load_pem_str(str, rd_privkey_bio, key); } int openssl_load_pubkey_file(const char * path, @@ -1353,7 +1317,12 @@ int openssl_load_pubkey_file(const char * path, if (fp == NULL) goto fail_file; + pthread_cleanup_push(__cleanup_fclose, fp); + pkey = PEM_read_PUBKEY(fp, NULL, NULL, NULL); + + pthread_cleanup_pop(false); + if (pkey == NULL) goto fail_key; @@ -1385,7 +1354,12 @@ int openssl_load_pubkey_file_to_der(const char * path, if (fp == NULL) goto fail_file; + pthread_cleanup_push(__cleanup_fclose, fp); + pkey = PEM_read_PUBKEY(fp, NULL, NULL, NULL); + + pthread_cleanup_pop(false); + if (pkey == NULL) goto fail_key; @@ -1412,37 +1386,14 @@ int openssl_load_pubkey_file_to_der(const char * path, int openssl_load_pubkey_str(const char * str, void ** key) { - BIO * bio; - EVP_PKEY * pkey; - - bio = BIO_new(BIO_s_mem()); - if (bio == NULL) - goto fail_bio; - - if (BIO_write(bio, str, strlen(str)) < 0) - goto fail_key; - - pkey = PEM_read_bio_PUBKEY(bio, NULL, NULL, NULL); - if (pkey == NULL) - goto fail_key; - - BIO_free(bio); - - *key = (void *) pkey; - - return 0; - fail_key: - BIO_free(bio); - fail_bio: - *key = NULL; - return -1; + return load_pem_str(str, rd_pubkey_bio, key); } int openssl_load_pubkey_raw_file(const char * path, buffer_t * buf) { FILE * fp; - uint8_t tmp_buf[MSGBUFSZ]; + uint8_t tmp_buf[CRYPT_KEY_BUFSZ]; size_t bytes_read; const char * algo; @@ -1453,7 +1404,12 @@ int openssl_load_pubkey_raw_file(const char * path, if (fp == NULL) goto fail_file; - bytes_read = fread(tmp_buf, 1, MSGBUFSZ, fp); + pthread_cleanup_push(__cleanup_fclose, fp); + + bytes_read = fread(tmp_buf, 1, CRYPT_KEY_BUFSZ, fp); + + pthread_cleanup_pop(false); + if (bytes_read == 0) goto fail_read; @@ -1495,11 +1451,17 @@ static const char * __openssl_hybrid_algo_from_sk_len(size_t len) return NULL; } +/* Wipe the raw-key staging buffer if a cancel aborts the read. */ +static void __cleanse_key_buf(void * o) +{ + OPENSSL_cleanse(o, CRYPT_KEY_BUFSZ); +} + int openssl_load_privkey_raw_file(const char * path, void ** key) { FILE * fp; - uint8_t tmp_buf[4096]; + uint8_t tmp_buf[CRYPT_KEY_BUFSZ]; size_t bytes_read; const char * algo; EVP_PKEY * pkey; @@ -1511,7 +1473,14 @@ int openssl_load_privkey_raw_file(const char * path, if (fp == NULL) goto fail_file; + pthread_cleanup_push(__cleanup_fclose, fp); + pthread_cleanup_push(__cleanse_key_buf, tmp_buf); + bytes_read = fread(tmp_buf, 1, sizeof(tmp_buf), fp); + + pthread_cleanup_pop(false); + pthread_cleanup_pop(false); + if (bytes_read == 0) goto fail_read; @@ -1562,65 +1531,71 @@ void openssl_free_key(EVP_PKEY * key) int openssl_check_crt_name(void * crt, const char * name) { - char * subj; - char * cn; - X509 * xcrt; + const unsigned char * cn; + ASN1_STRING * val; + X509_NAME * nm; + int idx; + int len; + + nm = X509_get_subject_name((X509 *) crt); + if (nm == NULL) + return -1; - xcrt = (X509 *) crt; + idx = X509_NAME_get_index_by_NID(nm, NID_commonName, -1); + if (idx < 0) + return -1; - subj = X509_NAME_oneline(X509_get_subject_name(xcrt), NULL, 0); - if (subj == NULL) - goto fail_subj; + val = X509_NAME_ENTRY_get_data(X509_NAME_get_entry(nm, idx)); + cn = ASN1_STRING_get0_data(val); + len = ASN1_STRING_length(val); - cn = strstr(subj, "CN="); - if (cn == NULL) - goto fail_cn; + if (len < 0 || (size_t) len != strlen(name)) + return -1; - if (strcmp(cn + 3, name) != 0) - goto fail_cn; + if (memchr(cn, '\0', (size_t) len) != NULL) + return -1; - free(subj); + if (memcmp(cn, name, (size_t) len) != 0) + return -1; return 0; - fail_cn: - free(subj); - fail_subj: - return -1; } int openssl_get_crt_name(void * crt, char * name) { - char * subj; - char * cn; - char * end; - X509 * xcrt; + const unsigned char * cn; + ASN1_STRING * val; + X509_NAME * nm; + int idx; + int len; + + nm = X509_get_subject_name((X509 *) crt); + if (nm == NULL) + return -1; - xcrt = (X509 *) crt; + idx = X509_NAME_get_index_by_NID(nm, NID_commonName, -1); + if (idx < 0) + return -1; - subj = X509_NAME_oneline(X509_get_subject_name(xcrt), NULL, 0); - if (subj == NULL) - goto fail_subj; + val = X509_NAME_ENTRY_get_data(X509_NAME_get_entry(nm, idx)); + cn = ASN1_STRING_get0_data(val); + len = ASN1_STRING_length(val); - cn = strstr(subj, "CN="); - if (cn == NULL) - goto fail_cn; + if (len < 0) + return -1; - cn += 3; /* Skip "CN=" */ + if ((size_t) len > NAME_SIZE) + return -ENAME; - /* Find end of CN (comma or slash for next field) */ - end = strpbrk(cn, ",/"); - if (end != NULL) - *end = '\0'; + /* Reject an embedded NUL that would truncate the parsed name. */ + if (memchr(cn, '\0', (size_t) len) != NULL) + return -1; - strcpy(name, cn); - free(subj); + memcpy(name, cn, (size_t) len); + name[len] = '\0'; return 0; - fail_cn: - free(subj); - fail_subj: - return -1; } int openssl_crt_str(const void * crt, @@ -1658,25 +1633,33 @@ int openssl_crt_str(const void * crt, int openssl_crt_der(const void * crt, buffer_t * buf) { - int len; + uint8_t * p; + int len; assert(crt != NULL); assert(buf != NULL); - len = i2d_X509((X509 *) crt, &buf->data); + /* Get the size by encoding to NULL */ + len = i2d_X509((X509 *) crt, NULL); if (len < 0) - goto fail_der; + goto fail_len; + + buf->data = malloc((size_t) len); + if (buf->data == NULL) + goto fail_malloc; + p = buf->data; /* i2d_X509 increments p */ + i2d_X509((X509 *) crt, &p); buf->len = (size_t) len; return 0; - fail_der: + fail_malloc: + fail_len: clrbuf(*buf); return -1; } - void * openssl_auth_create_store(void) { return X509_STORE_new(); @@ -1697,12 +1680,43 @@ int openssl_auth_add_crt_to_store(void * store, return ret == 1 ? 0 : -1; } -int openssl_verify_crt(void * store, - void * crt) +void * openssl_auth_create_chain(void) +{ + return sk_X509_new_null(); +} + +void openssl_auth_destroy_chain(void * chain) +{ + sk_X509_pop_free((STACK_OF(X509) *) chain, X509_free); +} + +int openssl_auth_add_crt_to_chain(void * chain, + void * crt) +{ + if (X509_up_ref((X509 *) crt) != 1) + goto fail_ref; + + if (sk_X509_push((STACK_OF(X509) *) chain, (X509 *) crt) == 0) + goto fail_push; + + return 0; + fail_push: + X509_free((X509 *) crt); + fail_ref: + return -1; +} + +int openssl_verify_crt_pin(void * store, + void * untrusted, + void * crt, + void * pin) { X509_STORE_CTX * ctx; X509_STORE * _store; X509* _crt; + STACK_OF(X509) * chain; + int i; + int n; int ret; _store = (X509_STORE *) store; @@ -1712,7 +1726,8 @@ int openssl_verify_crt(void * store, if (ctx == NULL) goto fail_store_ctx; - ret = X509_STORE_CTX_init(ctx, _store, _crt, NULL); + ret = X509_STORE_CTX_init(ctx, _store, _crt, + (STACK_OF(X509) *) untrusted); if (ret != 1) goto fail_ca; @@ -1720,13 +1735,39 @@ int openssl_verify_crt(void * store, if (ret != 1) goto fail_ca; + /* Peer cert only verifies a signature; gate on sig KU, not role. */ + if ((X509_get_key_usage(_crt) & KU_DIGITAL_SIGNATURE) == 0) + goto fail_ca; + + if (pin != NULL) { + chain = X509_STORE_CTX_get0_chain(ctx); + if (chain == NULL) + goto fail_ca; + n = sk_X509_num(chain); + for (i = 1; i < n; i++) /* Skip the leaf */ + if (X509_cmp(sk_X509_value(chain, i), pin) == 0) + break; + if (i == n) + goto fail_pin; + } + X509_STORE_CTX_free(ctx); return 0; + fail_pin: + X509_STORE_CTX_free(ctx); + return -ENOENT; fail_ca: X509_STORE_CTX_free(ctx); fail_store_ctx: - return -1; + return -EAUTH; +} + +int openssl_verify_crt(void * store, + void * untrusted, + void * crt) +{ + return openssl_verify_crt_pin(store, untrusted, crt, NULL); } static const EVP_MD * select_md(EVP_PKEY * pkey, @@ -1741,6 +1782,12 @@ static const EVP_MD * select_md(EVP_PKEY * pkey, return EVP_get_digestbynid(nid); } +bool openssl_pk_requires_md(const EVP_PKEY * pk) +{ + /* Provider-based (PQC) signatures have an intrinsic digest */ + return EVP_PKEY_get_id(pk) >= 0; +} + int openssl_sign(EVP_PKEY * pkp, int nid, buffer_t msg, @@ -1868,9 +1915,10 @@ void * openssl_secure_malloc(size_t size) return OPENSSL_secure_malloc(size); } -void openssl_secure_free(void * ptr) +void openssl_secure_free(void * ptr, + size_t size) { - OPENSSL_secure_free(ptr); + OPENSSL_secure_clear_free(ptr, size); } void openssl_secure_clear(void * ptr, @@ -1878,3 +1926,8 @@ void openssl_secure_clear(void * ptr, { OPENSSL_cleanse(ptr, size); } + +void openssl_cleanup(void) +{ + OPENSSL_cleanup(); +} diff --git a/src/lib/crypt/openssl.h b/src/lib/crypt/openssl.h index b95d1b0b..e5cc35f7 100644 --- a/src/lib/crypt/openssl.h +++ b/src/lib/crypt/openssl.h @@ -61,20 +61,44 @@ int openssl_get_algo_from_pk_der(buffer_t pk, int openssl_get_algo_from_pk_raw(buffer_t pk, char * algo); -int openssl_encrypt(struct ossl_crypt_ctx * ctx, - buffer_t in, - buffer_t * out); - -int openssl_decrypt(struct ossl_crypt_ctx * ctx, - buffer_t in, - buffer_t * out); +int openssl_seal(struct ossl_crypt_ctx * ctx, + const uint8_t * key, + const uint8_t * nonce, + buffer_t aad, + buffer_t in, + uint8_t * out, + uint8_t * tag); + +int openssl_open(struct ossl_crypt_ctx * ctx, + const uint8_t * key, + const uint8_t * nonce, + buffer_t aad, + buffer_t in, + const uint8_t * tag, + buffer_t * out); + +int openssl_oneshot_seal(int nid, + const uint8_t * key, + const uint8_t * nonce, + buffer_t aad, + buffer_t in, + buffer_t * out); + +int openssl_oneshot_open(int nid, + const uint8_t * key, + const uint8_t * nonce, + buffer_t aad, + buffer_t in, + buffer_t * out); + +int openssl_hkdf_expand(buffer_t key, + buffer_t info, + buffer_t out); struct ossl_crypt_ctx * openssl_crypt_create_ctx(struct crypt_sk * sk); void openssl_crypt_destroy_ctx(struct ossl_crypt_ctx * ctx); -int openssl_crypt_get_ivsz(struct ossl_crypt_ctx * ctx); - int openssl_crypt_get_tagsz(struct ossl_crypt_ctx * ctx); /* AUTHENTICATION */ @@ -136,9 +160,24 @@ void openssl_auth_destroy_store(void * store); int openssl_auth_add_crt_to_store(void * store, void * crt); +void * openssl_auth_create_chain(void); + +void openssl_auth_destroy_chain(void * chain); + +int openssl_auth_add_crt_to_chain(void * chain, + void * crt); + int openssl_verify_crt(void * store, + void * untrusted, void * crt); +int openssl_verify_crt_pin(void * store, + void * untrusted, + void * crt, + void * pin); + +bool openssl_pk_requires_md(const EVP_PKEY * pk); + int openssl_sign(EVP_PKEY * pkp, int md_nid, buffer_t msg, @@ -169,4 +208,6 @@ void openssl_secure_free(void * ptr, void openssl_secure_clear(void * ptr, size_t size); +void openssl_cleanup(void); + #endif /* OUROBOROS_LIB_CRYPT_OPENSSL_H */ diff --git a/src/lib/dev.c b/src/lib/dev.c index 9cfc24ee..166aba5c 100644 --- a/src/lib/dev.c +++ b/src/lib/dev.c @@ -29,10 +29,13 @@ #include "config.h" #include "ssm.h" +#include <ouroboros/atomics.h> #include <ouroboros/bitmap.h> #include <ouroboros/cep.h> +#include <ouroboros/crc16.h> #include <ouroboros/crypt.h> #include <ouroboros/dev.h> +#include <ouroboros/endian.h> #include <ouroboros/errno.h> #include <ouroboros/fccntl.h> #include <ouroboros/flow.h> @@ -45,32 +48,33 @@ #include <ouroboros/np1_flow.h> #include <ouroboros/pthread.h> #include <ouroboros/random.h> +#ifdef PROC_FLOW_STATS +#include <ouroboros/rib.h> +#endif #include <ouroboros/serdes-irm.h> +#include <ouroboros/sockets.h> #include <ouroboros/ssm_flow_set.h> #include <ouroboros/ssm_pool.h> #include <ouroboros/ssm_rbuff.h> -#include <ouroboros/sockets.h> +#include <ouroboros/tw.h> #include <ouroboros/utils.h> -#ifdef PROC_FLOW_STATS -#include <ouroboros/rib.h> -#endif +#include <assert.h> #ifdef HAVE_LIBGCRYPT #include <gcrypt.h> #endif -#include <assert.h> -#include <stdlib.h> -#include <string.h> -#include <stdio.h> #include <stdarg.h> #include <stdbool.h> +#include <inttypes.h> +#include <stdio.h> +#include <stdlib.h> +#include <string.h> #include <sys/types.h> #ifndef CLOCK_REALTIME_COARSE #define CLOCK_REALTIME_COARSE CLOCK_REALTIME #endif -/* Partial read information. */ #define NO_PART -1 #define DONE_PART -2 @@ -78,19 +82,12 @@ #define SECMEMSZ 16384 #define MSGBUFSZ 2048 -/* map flow_ids to flow descriptors; track state of the flow */ struct fmap { int fd; - /* TODO: use actual flow state */ enum flow_state state; }; -#define frcti_to_flow(frcti) \ - ((struct flow *)((uint8_t *) frcti - offsetof(struct flow, frcti))) - struct flow { - struct list_head next; - struct flow_info info; struct ssm_rbuff * rx_rb; @@ -101,8 +98,14 @@ struct flow { ssize_t part_idx; struct crypt_ctx * crypt; - int headsz; /* IV */ - int tailsz; /* Tag + CRC */ + int headsz; /* Selector */ + int tailsz; /* Tag + CRC */ + + struct timespec rk_grace; /* TX-promote deadline (0 = none) */ + struct timespec rk_attempt; /* Last re-key attempt (backoff) */ + bool rk_wm_inflight; /* Re-key trigger in flight */ + uint32_t rk_wm_ctr; /* Throttles the consult */ + bool rk_initiator; /* OAP initiator this re-key */ struct timespec snd_act; struct timespec rcv_act; @@ -135,16 +138,10 @@ struct { struct flow * flows; struct fmap * id_to_fd; - struct list_head flow_list; pthread_mutex_t mtx; pthread_cond_t cond; - pthread_t tx; - pthread_t rx; - size_t n_frcti; - fset_t * frct_set; - pthread_rwlock_t lock; } proc; @@ -243,7 +240,7 @@ static int proc_announce(const struct proc_info * proc) return irm__irm_result_des(&msg); } -/* IRMd will clean up the mess if this fails */ +/* IRMd cleans up on failure. */ static void proc_exit(void) { uint8_t buf[SOCK_BUF_SIZE]; @@ -264,7 +261,7 @@ static int spb_encrypt(struct flow * flow, uint8_t * tail; if (flow->crypt == NULL) - return 0; /* No encryption */ + return 0; in.data = ssm_pk_buff_head(spb); in.len = ssm_pk_buff_len(spb); @@ -272,11 +269,11 @@ static int spb_encrypt(struct flow * flow, if (crypt_encrypt(flow->crypt, in, &out) < 0) goto fail_encrypt; - head = ssm_pk_buff_head_alloc(spb, flow->headsz); + head = ssm_pk_buff_push(spb, flow->headsz); if (head == NULL) goto fail_alloc; - tail = ssm_pk_buff_tail_alloc(spb, flow->tailsz); + tail = ssm_pk_buff_push_tail(spb, flow->tailsz); if (tail == NULL) goto fail_alloc; @@ -299,17 +296,16 @@ static int spb_decrypt(struct flow * flow, uint8_t * head; if (flow->crypt == NULL) - return 0; /* No decryption */ + return 0; in.data = ssm_pk_buff_head(spb); in.len = ssm_pk_buff_len(spb); if (crypt_decrypt(flow->crypt, in, &out) < 0) - return -ENOMEM; - + return -ECRYPT; - head = ssm_pk_buff_head_release(spb, flow->headsz) + flow->headsz; - ssm_pk_buff_tail_release(spb, flow->tailsz); + head = ssm_pk_buff_pop(spb, flow->headsz) + flow->headsz; + ssm_pk_buff_pop_tail(spb, flow->tailsz); memcpy(head, out.data, out.len); @@ -318,130 +314,362 @@ static int spb_decrypt(struct flow * flow, return 0; } -#include "frct.c" +/* tw_move under proc.lock rdlock; gates teardown vs in-flight fires. */ +static void tw_move_safe(void) +{ + pthread_rwlock_rdlock(&proc.lock); + + pthread_cleanup_push(__cleanup_rwlock_unlock, &proc.lock); + + tw_move(); + + pthread_cleanup_pop(1); +} -void * flow_tx(void * o) +static int crc_add(struct ssm_pk_buff * spb, + size_t head_skip) { - struct timespec tic = TIMESPEC_INIT_NS(TICTIME); + uint8_t * head; + uint8_t * tail; - (void) o; + tail = ssm_pk_buff_push_tail(spb, CRCLEN); + if (tail == NULL) + return -ENOMEM; - while (true) { - timerwheel_move(); + head = ssm_pk_buff_head(spb) + head_skip; - nanosleep(&tic, NULL); - } + mem_hash(HASH_CRC32, tail, head, tail - head); - return (void *) 0; + return 0; } -static void flow_send_keepalive(struct flow * flow, - struct timespec now) +static int crc_check(struct ssm_pk_buff * spb, + size_t head_skip) { - struct ssm_pk_buff * spb; - ssize_t idx; - uint8_t * ptr; + uint32_t crc; + uint8_t * head; + uint8_t * tail; - idx = ssm_pool_alloc(proc.pool, 0, &ptr, &spb); - if (idx < 0) - return; + if (ssm_pk_buff_len(spb) < head_skip + CRCLEN) + return 1; - pthread_rwlock_wrlock(&proc.lock); + head = ssm_pk_buff_head(spb) + head_skip; + tail = ssm_pk_buff_pop_tail(spb, CRCLEN); - flow->snd_act = now; + mem_hash(HASH_CRC32, &crc, head, tail - head); - if (ssm_rbuff_write(flow->tx_rb, idx)) - ssm_pool_remove(proc.pool, idx); + return !(crc == *((uint32_t *) tail)); +} + +/* FRCT included here so it can use proc and dev.c statics directly. */ +#include "frct.c" + +/* Decrypt before any check so the plaintext is authoritative. */ +static bool invalid_pkt(struct flow * flow, + struct ssm_pk_buff * spb) +{ + const struct frct_pci * pci; + uint16_t flags; + size_t pci_total; + + if (spb == NULL || ssm_pk_buff_len(spb) == 0) + return true; + + if (spb_decrypt(flow, spb) < 0) + return true; + + if (flow->frcti == NULL) { + if (flow->info.qs.ber == 0 && crc_check(spb, 0) != 0) + return true; + return false; + } + + if (ssm_pk_buff_len(spb) < FRCT_PCILEN) + return true; + + pci = (const struct frct_pci *) ssm_pk_buff_head(spb); + flags = ntoh16(pci->flags); + + /* Untrusted flag read; mismatch on HCS will drop on corrupt. */ + if (flags & FRCT_DATA) + pci_total = frcti_data_hdr_len(flow->frcti); else - ssm_flow_set_notify(flow->set, flow->info.id, FLOW_PKT); + pci_total = frcti_ctrl_hdr_len(flow->frcti); - pthread_rwlock_unlock(&proc.lock); + if (ssm_pk_buff_len(spb) < pci_total) + return true; + + if (frct_hcs_check(pci, flow->frcti) != 0) + return true; + + /* HCS valid: CRC32 on SACK; or on DATA if ber = 0. */ + if (flags & FRCT_SACK) { + if (crc_check(spb, pci_total) != 0) + return true; + + } else if ((flags & FRCT_DATA) && flow->info.qs.ber == 0) { + if (crc_check(spb, pci_total) != 0) + return true; + } + + return false; } -/* Needs rdlock on proc. */ -static void _flow_keepalive(struct flow * flow) +static bool deadline_passed(const struct timespec * abs) { - struct timespec now; - struct timespec s_act; - struct timespec r_act; - int flow_id; - time_t timeo; - uint32_t acl; + struct timespec now; - s_act = flow->snd_act; - r_act = flow->rcv_act; + if (abs == NULL) + return false; - flow_id = flow->info.id; - timeo = flow->info.qs.timeout; + clock_gettime(PTHREAD_COND_CLOCK, &now); - acl = ssm_rbuff_get_acl(flow->rx_rb); - if (timeo == 0 || acl & (ACL_FLOWPEER | ACL_FLOWDOWN)) - return; + return ts_diff_ns(&now, abs) >= 0; +} + +/* Clamp the wait by min(dl, next tw expiry, now + TICTIME). */ +static void compute_wait_deadline(const struct timespec * dl, + struct timespec * out) +{ + struct timespec now; + struct timespec cap; + struct timespec expiry; + struct timespec tic = TIMESPEC_INIT_NS(TICTIME); clock_gettime(PTHREAD_COND_CLOCK, &now); + ts_add(&now, &tic, &cap); - if (ts_diff_ns(&now, &r_act) > (int64_t) timeo * MILLION) { - ssm_rbuff_set_acl(flow->rx_rb, ACL_FLOWPEER); - ssm_flow_set_notify(proc.fqset, flow_id, FLOW_PEER); - return; - } + tw_next_expiry(&expiry); - if (ts_diff_ns(&now, &s_act) > (int64_t) timeo * (MILLION >> 2)) { - pthread_rwlock_unlock(&proc.lock); + *out = (ts_diff_ns(&cap, &expiry) < 0) ? expiry : cap; + if (dl != NULL && ts_diff_ns(out, dl) > 0) + *out = *dl; +} + +static void flow_drain_rx_nb(struct flow * flow) +{ + ssize_t idx; + struct ssm_pk_buff * spb; + struct ssm_rbuff * rx_rb; + struct frcti * frcti; +#ifdef PROC_FLOW_STATS + struct timespec t_a; + struct timespec t_b; +#endif - flow_send_keepalive(flow, now); + if (flow->frcti != NULL) + STAT_BUMP(flow->frcti, drain_calls); + while (true) { pthread_rwlock_rdlock(&proc.lock); + + rx_rb = flow->rx_rb; + if (rx_rb == NULL) { + pthread_rwlock_unlock(&proc.lock); + return; + } + + idx = ssm_rbuff_read(rx_rb); + if (idx < 0) { + pthread_rwlock_unlock(&proc.lock); + return; + } + + spb = ssm_pool_get(proc.pool, idx); + if (invalid_pkt(flow, spb)) { + ssm_pool_remove(proc.pool, idx); + pthread_rwlock_unlock(&proc.lock); + continue; + } + + frcti = flow->frcti; + if (frcti != NULL) { +#ifdef PROC_FLOW_STATS + clock_gettime(CLOCK_MONOTONIC, &t_a); + FRCTI_RCV(frcti, spb); + clock_gettime(CLOCK_MONOTONIC, &t_b); + STAT_ADD(frcti, rcv_proc_ns, + (size_t) ts_diff_ns(&t_b, &t_a)); +#else + FRCTI_RCV(frcti, spb); +#endif + } else { + ssm_pool_remove(proc.pool, idx); + } + + pthread_rwlock_unlock(&proc.lock); + + /* Per-packet so the delayed-ACK fires on time in a burst. */ +#ifdef PROC_FLOW_STATS + clock_gettime(CLOCK_MONOTONIC, &t_a); + tw_move_safe(); + clock_gettime(CLOCK_MONOTONIC, &t_b); + if (frcti != NULL) + STAT_ADD(frcti, tw_move_ns, + (size_t) ts_diff_ns(&t_b, &t_a)); +#else + tw_move_safe(); +#endif } } -static void handle_keepalives(void) +/* TX-promotion grace when the peer's install latency is unknown (raw). */ +#define REKEY_GRACE_MS 1000 + +/* Last-resort promote within N node-keys of exhaustion (< watermark). */ +#define REKEY_PROMOTE_FLOOR 1 + +/* Throttle re-key retries so a failed attempt can't storm the IRMd. */ +#define REKEY_BACKOFF_NS (250 * MILLION) + +/* proc.lock (rd) only guards teardown; crypt_rekey self-synchronises. */ +static void flow_rekey(struct flow * flow) { - struct list_head * p; - struct list_head * h; + struct flow_info info; + struct crypt_sk sk; + struct timespec now; + struct timespec intv; + time_t ms; + uint8_t key[SYMMKEYSZ]; + uint8_t buf[SOCK_BUF_SIZE]; + buffer_t msg = {SOCK_BUF_SIZE, buf}; + bool has_key; + bool initiator = false; pthread_rwlock_rdlock(&proc.lock); + if (flow->info.id < 0 || flow->crypt == NULL) { + pthread_rwlock_unlock(&proc.lock); + return; + } - list_for_each_safe(p, h, &proc.flow_list) { - struct flow * flow; - flow = list_entry(p, struct flow, next); - _flow_keepalive(flow); + /* Back off so a failed attempt can't storm the IRMd per syscall. */ + clock_gettime(PTHREAD_COND_CLOCK, &now); + if (ts_diff_ns(&now, &flow->rk_attempt) < REKEY_BACKOFF_NS) { + pthread_rwlock_unlock(&proc.lock); + return; } + flow->rk_attempt = now; + info = flow->info; pthread_rwlock_unlock(&proc.lock); -} -static void __cleanup_fqueue_destroy(void * fq) -{ - fqueue_destroy((fqueue_t *) fq); + if (flow_update__irm_req_ser(&msg, &info, false) < 0) + return; + + if (send_recv_msg(&msg) < 0) + return; + + sk.key = key; + if (flow_rekey__irm_result_des(&msg, &sk, &has_key, &initiator) < 0) + return; + + if (!has_key) + return; + + pthread_rwlock_rdlock(&proc.lock); + if (flow->info.id == info.id && flow->crypt != NULL) { + if (crypt_rekey(flow->crypt, &sk) == 0) { + flow->rk_initiator = initiator; + /* Hold TX on the old epoch until the peer installs. */ + ms = flow->info.mpl > 0 ? flow->info.mpl * 3 + : REKEY_GRACE_MS; + intv.tv_sec = ms / 1000; + intv.tv_nsec = (ms % 1000) * MILLION; + clock_gettime(PTHREAD_COND_CLOCK, &now); + ts_add(&now, &intv, &flow->rk_grace); + } + /* Re-arm the watermark even if the install was a no-op. */ + STORE_RELAXED(&flow->rk_wm_inflight, false); + } + pthread_rwlock_unlock(&proc.lock); + + crypt_secure_clear(key, SYMMKEYSZ); } -void * flow_rx(void * o) +/* A clamp-timeout means tw work is due, not the caller deadline. */ +static int flow_rx_one(struct flow * flow, + struct timespec * abs) { - struct timespec tic = TIMESPEC_INIT_NS(TICTIME); - int ret; - struct fqueue * fq; + struct timespec wait_abs; + struct ssm_pk_buff * spb; + struct ssm_rbuff * rx_rb; + ssize_t idx; - (void) o; + while (true) { + compute_wait_deadline(abs, &wait_abs); - fq = fqueue_create(); + /* rdlock gates flow_fini; FLOWDOWN preempts the block. */ + pthread_rwlock_rdlock(&proc.lock); - pthread_cleanup_push(__cleanup_fqueue_destroy, fq); + rx_rb = flow->rx_rb; + if (rx_rb == NULL) { + pthread_rwlock_unlock(&proc.lock); + return -EFLOWDOWN; + } - /* fevent will filter all FRCT packets for us */ - while ((ret = fevent(proc.frct_set, fq, &tic)) != 0) { - if (ret == -ETIMEDOUT) { - handle_keepalives(); + /* Pull a parked re-key before re-blocking (idle reader). */ + if (flow->crypt != NULL + && (ssm_rbuff_get_flags(rx_rb) & RB_REKEY)) { + pthread_rwlock_unlock(&proc.lock); + flow_rekey(flow); continue; } - while (fqueue_next(fq) >= 0) - ; /* no need to act */ + pthread_cleanup_push(__cleanup_rwlock_unlock, &proc.lock); + + idx = ssm_rbuff_read_b(rx_rb, &wait_abs); + + pthread_cleanup_pop(false); + + if (idx == -ETIMEDOUT) { + pthread_rwlock_unlock(&proc.lock); + if (deadline_passed(abs)) + return -ETIMEDOUT; + tw_move_safe(); + continue; + } + if (idx < 0) { + pthread_rwlock_unlock(&proc.lock); + return idx; + } + + spb = ssm_pool_get(proc.pool, idx); + if (invalid_pkt(flow, spb)) { + ssm_pool_remove(proc.pool, idx); + pthread_rwlock_unlock(&proc.lock); + continue; + } + + if (flow->frcti != NULL) + FRCTI_RCV(flow->frcti, spb); + else + ssm_pool_remove(proc.pool, idx); + + pthread_rwlock_unlock(&proc.lock); + + tw_move_safe(); + return 0; } +} - pthread_cleanup_pop(true); +/* 0 = window open; -EAGAIN = !block and would block; else flow_rx_one rc. */ +static __inline__ int flow_wait_window(struct flow * flow, + size_t n, + bool block, + struct timespec * dl) +{ + int rc; - return (void *) 0; + while (true) { + flow_drain_rx_nb(flow); + if (FRCTI_IS_WINDOW_OPEN_N(flow->frcti, n)) + return 0; + if (!block) + return -EAGAIN; + rc = flow_rx_one(flow, dl); + if (rc < 0) + return rc; + } } static void flow_clear(int fd) @@ -451,36 +679,36 @@ static void flow_clear(int fd) proc.flows[fd].info.id = -1; } -static void __flow_fini(int fd) +/* Order before flow_fini's wrlock, which blocks on rdlock holders. */ +static void flow_quiesce(int fd) { - assert(fd >= 0 && fd < SYS_MAX_FLOWS); + struct ssm_rbuff * rx_rb = proc.flows[fd].rx_rb; + struct ssm_rbuff * tx_rb = proc.flows[fd].tx_rb; - if (proc.flows[fd].frcti != NULL) { - proc.n_frcti--; - if (proc.n_frcti == 0) { - pthread_cancel(proc.tx); - pthread_join(proc.tx, NULL); - } + if (rx_rb != NULL) + ssm_rbuff_set_bits(rx_rb, RB_FLOWDOWN); - ssm_flow_set_del(proc.fqset, 0, proc.flows[fd].info.id); + if (tx_rb != NULL) + ssm_rbuff_set_bits(tx_rb, RB_FLOWDOWN); +} +static void do_flow_fini(int fd) +{ + assert(fd >= 0 && fd < PROC_MAX_FLOWS); + + if (proc.flows[fd].frcti != NULL) frcti_destroy(proc.flows[fd].frcti); - } if (proc.flows[fd].info.id != -1) { flow_destroy(&proc.id_to_fd[proc.flows[fd].info.id]); bmp_release(proc.fds, fd); } - if (proc.flows[fd].rx_rb != NULL) { - ssm_rbuff_set_acl(proc.flows[fd].rx_rb, ACL_FLOWDOWN); + if (proc.flows[fd].rx_rb != NULL) ssm_rbuff_close(proc.flows[fd].rx_rb); - } - if (proc.flows[fd].tx_rb != NULL) { - ssm_rbuff_set_acl(proc.flows[fd].tx_rb, ACL_FLOWDOWN); + if (proc.flows[fd].tx_rb != NULL) ssm_rbuff_close(proc.flows[fd].tx_rb); - } if (proc.flows[fd].set != NULL) { ssm_flow_set_notify(proc.flows[fd].set, @@ -491,24 +719,40 @@ static void __flow_fini(int fd) crypt_destroy_ctx(proc.flows[fd].crypt); - list_del(&proc.flows[fd].next); - flow_clear(fd); } static void flow_fini(int fd) { + flow_quiesce(fd); + pthread_rwlock_wrlock(&proc.lock); - __flow_fini(fd); + do_flow_fini(fd); pthread_rwlock_unlock(&proc.lock); } #define IS_ENCRYPTED(crypt) ((crypt)->nid != NID_undef) -#define IS_ORDERED(flow) (flow.qs.in_order != 0) +#define IS_ORDERED(info) ((info)->qs.service != SVC_RAW) +#define IS_STREAM(info) ((info)->qs.service == SVC_STREAM) + +/* Raw MTU minus the wrapping (IV/Tag + optional CRC) dev.c adds. */ +static __inline__ size_t flow_user_mtu(const struct flow * flow, + size_t raw) +{ + size_t hdr; + + hdr = flow->headsz + flow->tailsz; + if (flow->info.qs.ber == 0 && flow->crypt == NULL) + hdr += CRCLEN; + + return raw > hdr ? raw - hdr : 0; +} + static int flow_init(struct flow_info * info, - struct crypt_sk * sk) + struct crypt_sk * sk, + time_t rtt_hint) { struct timespec now; struct flow * flow; @@ -550,33 +794,25 @@ static int flow_init(struct flow_info * info, flow->tailsz = 0; if (IS_ENCRYPTED(sk)) { - /* Set to lower value in tests, should we make configurable? */ - sk->rot_bit = KEY_ROTATION_BIT; flow->crypt = crypt_create_ctx(sk); if (flow->crypt == NULL) goto fail_crypt; - flow->headsz = crypt_get_ivsz(flow->crypt); + flow->headsz = crypt_get_headsz(flow->crypt); flow->tailsz = crypt_get_tagsz(flow->crypt); } assert(flow->frcti == NULL); - if (IS_ORDERED(flow->info)) { - flow->frcti = frcti_create(fd, DELT_A, DELT_R, info->mpl); + if (IS_ORDERED(&flow->info)) { + uint32_t frct_mtu = flow_user_mtu(flow, info->mtu); + + flow->frcti = frcti_create(fd, DELT_A, DELT_R, + info->mpl, rtt_hint, + info->qs, frct_mtu); if (flow->frcti == NULL) goto fail_frcti; - - if (ssm_flow_set_add(proc.fqset, 0, info->id)) - goto fail_flow_set_add; - - ++proc.n_frcti; - if (proc.n_frcti == 1 && - pthread_create(&proc.tx, NULL, flow_tx, NULL) < 0) - goto fail_tx_thread; } - list_add_tail(&flow->next, &proc.flow_list); - proc.id_to_fd[info->id].fd = fd; flow_set_state(&proc.id_to_fd[info->id], FLOW_ALLOCATED); @@ -585,10 +821,6 @@ static int flow_init(struct flow_info * info, return fd; - fail_tx_thread: - ssm_flow_set_del(proc.fqset, 0, info->id); - fail_flow_set_add: - frcti_destroy(flow->frcti); fail_frcti: crypt_destroy_ctx(flow->crypt); fail_crypt: @@ -618,9 +850,10 @@ static void init(int argc, char ** argv, char ** envp) { - struct proc_info info; - char * prog = argv[0]; - int i; + struct proc_info info; + char * prog = argv[0]; + int i; + pthread_rwlockattr_t attr; #ifdef PROC_FLOW_STATS char procstr[32]; #endif @@ -655,13 +888,13 @@ static void init(int argc, gcry_control(GCRYCTL_INITIALIZATION_FINISHED, 0); } #endif - proc.fds = bmp_create(PROG_MAX_FLOWS - PROG_RES_FDS, PROG_RES_FDS); + proc.fds = bmp_create(PROC_MAX_FLOWS - PROC_RES_FDS, PROC_RES_FDS); if (proc.fds == NULL) { fprintf(stderr, "FATAL: Could not create fd bitmap.\n"); goto fail_fds; } - proc.fqueues = bmp_create(PROG_MAX_FQUEUES, 0); + proc.fqueues = bmp_create(PROC_MAX_FQUEUES, 0); if (proc.fqueues == NULL) { fprintf(stderr, "FATAL: Could not create fqueue bitmap.\n"); goto fail_fqueues; @@ -677,13 +910,13 @@ static void init(int argc, goto fail_rdrb; } - proc.flows = malloc(sizeof(*proc.flows) * PROG_MAX_FLOWS); + proc.flows = malloc(sizeof(*proc.flows) * PROC_MAX_FLOWS); if (proc.flows == NULL) { fprintf(stderr, "FATAL: Could not malloc flows.\n"); goto fail_flows; } - for (i = 0; i < PROG_MAX_FLOWS; ++i) + for (i = 0; i < PROC_MAX_FLOWS; ++i) flow_clear(i); proc.id_to_fd = malloc(sizeof(*proc.id_to_fd) * SYS_MAX_FLOWS); @@ -705,7 +938,16 @@ static void init(int argc, goto fail_cond; } - if (pthread_rwlock_init(&proc.lock, NULL) < 0) { + /* Writer-preferred: FRCT readers must not starve flow accept. */ + if (pthread_rwlockattr_init(&attr) != 0) { + fprintf(stderr, "FATAL: Could not init rwlock attributes.\n"); + goto fail_rwlock_attr; + } +#if defined(__GLIBC__) + pthread_rwlockattr_setkind_np( + &attr, PTHREAD_RWLOCK_PREFER_WRITER_NONRECURSIVE_NP); +#endif + if (pthread_rwlock_init(&proc.lock, &attr) != 0) { fprintf(stderr, "FATAL: Could not initialize flow lock.\n"); goto fail_flow_lock; } @@ -716,20 +958,14 @@ static void init(int argc, goto fail_fqset; } - proc.frct_set = fset_create(); - if (proc.frct_set == NULL || proc.frct_set->idx != 0) { - fprintf(stderr, "FATAL: Could not create FRCT set.\n"); - goto fail_frct_set; - } - - if (timerwheel_init() < 0) { + if (tw_init() < 0) { fprintf(stderr, "FATAL: Could not initialize timerwheel.\n"); goto fail_timerwheel; } if (crypt_secure_malloc_init(PROC_SECMEM_MAX) < 0) { fprintf(stderr, "FATAL: Could not init secure malloc.\n"); - goto fail_timerwheel; + goto fail_secmem; } #if defined PROC_FLOW_STATS @@ -741,28 +977,23 @@ static void init(int argc, } } #endif - if (pthread_create(&proc.rx, NULL, flow_rx, NULL) < 0) { - fprintf(stderr, "FATAL: Could not start monitor thread.\n"); - goto fail_monitor; - } - - list_head_init(&proc.flow_list); + pthread_rwlockattr_destroy(&attr); return; - fail_monitor: #if defined PROC_FLOW_STATS - rib_fini(); fail_rib_init: + crypt_secure_malloc_fini(); #endif - timerwheel_fini(); + fail_secmem: + tw_fini(); fail_timerwheel: - fset_destroy(proc.frct_set); - fail_frct_set: ssm_flow_set_close(proc.fqset); fail_fqset: pthread_rwlock_destroy(&proc.lock); fail_flow_lock: + pthread_rwlockattr_destroy(&attr); + fail_rwlock_attr: pthread_cond_destroy(&proc.cond); fail_cond: pthread_mutex_destroy(&proc.mtx); @@ -789,19 +1020,20 @@ static void fini(void) if (proc.fds == NULL) return; - pthread_cancel(proc.rx); - pthread_join(proc.rx, NULL); + /* Wake all in-flight readers/writers BEFORE wrlock acquire. */ + for (i = 0; i < PROC_MAX_FLOWS; ++i) + if (proc.flows[i].info.id != -1) + flow_quiesce(i); pthread_rwlock_wrlock(&proc.lock); - for (i = 0; i < PROG_MAX_FLOWS; ++i) { + for (i = 0; i < PROC_MAX_FLOWS; ++i) { struct flow * flow = &proc.flows[i]; if (flow->info.id != -1) { ssize_t idx; - ssm_rbuff_set_acl(flow->rx_rb, ACL_FLOWDOWN); while ((idx = ssm_rbuff_read(flow->rx_rb)) >= 0) ssm_pool_remove(proc.pool, idx); - __flow_fini(i); + do_flow_fini(i); } } @@ -813,9 +1045,9 @@ static void fini(void) #ifdef PROC_FLOW_STATS rib_fini(); #endif - timerwheel_fini(); + crypt_secure_malloc_fini(); - fset_destroy(proc.frct_set); + tw_fini(); ssm_flow_set_close(proc.fqset); @@ -860,6 +1092,10 @@ int flow_accept(qosspec_t * qs, if (qs != NULL) qs->ber = 1; #endif + /* STREAM cannot tolerate loss: drops create silent gaps. */ + if (qs != NULL && qs->service == SVC_STREAM && qs->loss != 0) + return -EINVAL; + memset(&flow, 0, sizeof(flow)); flow.n_pid = getpid(); @@ -872,13 +1108,16 @@ int flow_accept(qosspec_t * qs, if (err < 0) return err; - crypt.key = key; + crypt.key = key; + crypt.epoch = 0; + crypt.role = CRYPT_ROLE_RESP; err = flow__irm_result_des(&msg, &flow, &crypt); if (err < 0) return err; - fd = flow_init(&flow, &crypt); + /* No RTT in accept; rtt_hint=0 bootstraps from first ACK. */ + fd = flow_init(&flow, &crypt, 0); crypt_secure_clear(key, SYMMKEYSZ); @@ -899,11 +1138,16 @@ int flow_alloc(const char * dst, uint8_t key[SYMMKEYSZ]; int fd; int err; + struct timespec t0; + struct timespec t1; #ifdef QOS_DISABLE_CRC if (qs != NULL) qs->ber = 1; #endif + /* STREAM cannot tolerate loss: drops create silent gaps. */ + if (qs != NULL && qs->service == SVC_STREAM && qs->loss != 0) + return -EINVAL; memset(&flow, 0, sizeof(flow)); @@ -913,19 +1157,23 @@ int flow_alloc(const char * dst, if (flow_alloc__irm_req_ser(&msg, &flow, dst, timeo)) return -ENOMEM; + clock_gettime(PTHREAD_COND_CLOCK, &t0); + err = send_recv_msg(&msg); - if (err < 0) { - printf("send_recv_msg error %d\n", err); + if (err < 0) return err; - } - crypt.key = key; + clock_gettime(PTHREAD_COND_CLOCK, &t1); + + crypt.key = key; + crypt.epoch = 0; + crypt.role = CRYPT_ROLE_INIT; err = flow__irm_result_des(&msg, &flow, &crypt); if (err < 0) return err; - fd = flow_init(&flow, &crypt); + fd = flow_init(&flow, &crypt, ts_diff_ns(&t1, &t0)); crypt_secure_clear(key, SYMMKEYSZ); @@ -958,13 +1206,15 @@ int flow_join(const char * dst, if (err < 0) return err; - crypt.key = key; + crypt.key = key; + crypt.epoch = 0; + crypt.role = CRYPT_ROLE_INIT; err = flow__irm_result_des(&msg, &flow, &crypt); if (err < 0) return err; - fd = flow_init(&flow, &crypt); + fd = flow_init(&flow, &crypt, 0); crypt_secure_clear(key, SYMMKEYSZ); @@ -983,10 +1233,10 @@ int flow_dealloc(int fd) struct flow * flow; int err; - if (fd < 0 || fd >= SYS_MAX_FLOWS ) + if (fd < 0 || fd >= PROC_MAX_FLOWS ) return -EINVAL; - memset(&info, 0, sizeof(flow)); + memset(&info, 0, sizeof(info)); flow = &proc.flows[fd]; @@ -1008,9 +1258,8 @@ int flow_dealloc(int fd) pthread_rwlock_rdlock(&proc.lock); - timeo.tv_sec = frcti_dealloc(flow->frcti); - while (timeo.tv_sec < 0) { /* keep the flow active for rtx */ - ssize_t ret; + while (FRCTI_LINGERING(flow->frcti)) { + ssize_t ret; pthread_rwlock_unlock(&proc.lock); @@ -1018,12 +1267,12 @@ int flow_dealloc(int fd) pthread_rwlock_rdlock(&proc.lock); - timeo.tv_sec = frcti_dealloc(flow->frcti); - - if (ret == -EFLOWDOWN && timeo.tv_sec < 0) - timeo.tv_sec = -timeo.tv_sec; + if (ret == -EFLOWDOWN) + break; } + timeo.tv_sec = FRCTI_DEALLOC(flow->frcti); + pthread_cleanup_push(__cleanup_rwlock_unlock, &proc.lock); ssm_rbuff_fini(flow->tx_rb); @@ -1033,15 +1282,18 @@ int flow_dealloc(int fd) info.id = flow->info.id; info.n_pid = getpid(); - if (flow_dealloc__irm_req_ser(&msg, &info, &timeo) < 0) - return -ENOMEM; + if (flow_dealloc__irm_req_ser(&msg, &info, &timeo) < 0) { + err = -ENOMEM; + goto out; + } err = send_recv_msg(&msg); if (err < 0) - return err; + goto out; err = irm__irm_result_des(&msg); + out: flow_fini(fd); return err; @@ -1055,12 +1307,12 @@ int ipcp_flow_dealloc(int fd) struct flow * flow; int err; - if (fd < 0 || fd >= SYS_MAX_FLOWS ) + if (fd < 0 || fd >= PROC_MAX_FLOWS ) return -EINVAL; flow = &proc.flows[fd]; - memset(&info, 0, sizeof(flow)); + memset(&info, 0, sizeof(info)); pthread_rwlock_rdlock(&proc.lock); @@ -1074,15 +1326,18 @@ int ipcp_flow_dealloc(int fd) pthread_rwlock_unlock(&proc.lock); - if (ipcp_flow_dealloc__irm_req_ser(&msg, &info) < 0) - return -ENOMEM; + if (ipcp_flow_dealloc__irm_req_ser(&msg, &info) < 0) { + err = -ENOMEM; + goto out; + } err = send_recv_msg(&msg); if (err < 0) - return err; + goto out; err = irm__irm_result_des(&msg); + out: flow_fini(fd); return err; @@ -1098,12 +1353,20 @@ int fccntl(int fd, va_list l; struct timespec * timeo; qosspec_t * qs; - uint32_t rx_acl; - uint32_t tx_acl; size_t * qlen; struct flow * flow; - - if (fd < 0 || fd >= SYS_MAX_FLOWS) + uint16_t old_acc; + uint16_t new_acc; + size_t max; + size_t * maxp; + size_t rsz; + size_t * rszp; + time_t rto; + time_t * rtop; + int rc; + bool emit_eos = false; + + if (fd < 0 || fd >= PROC_MAX_FLOWS) return -EBADF; flow = &proc.flows[fd]; @@ -1167,36 +1430,44 @@ int fccntl(int fd, qlen = va_arg(l, size_t *); *qlen = ssm_rbuff_queued(flow->tx_rb); break; + case FLOWGMTU: + maxp = va_arg(l, size_t *); + if (maxp == NULL) + goto einval; + *maxp = flow_user_mtu(flow, flow->info.mtu); + break; case FLOWSFLAGS: + old_acc = flow->oflags & FLOWFACCMODE; flow->oflags = va_arg(l, uint32_t); - rx_acl = ssm_rbuff_get_acl(flow->rx_rb); - tx_acl = ssm_rbuff_get_acl(flow->rx_rb); - /* - * Making our own flow write only means making the - * the other side of the flow read only. - */ + new_acc = flow->oflags & FLOWFACCMODE; + + /* Defer EOS emit until after proc.lock is dropped: */ + /* frcti_fin_snd may block on shm-pool/tx-rb. */ + if (new_acc == FLOWFRDONLY + && old_acc != FLOWFRDONLY + && flow->frcti != NULL) + emit_eos = true; + + /* Our flow write-only -> peer's read-only; restore on RDWR. */ if (flow->oflags & FLOWFWRONLY) - rx_acl |= ACL_RDONLY; - if (flow->oflags & FLOWFRDWR) - rx_acl |= ACL_RDWR; + ssm_rbuff_clr_bits(flow->rx_rb, RB_WR); + else + ssm_rbuff_set_bits(flow->rx_rb, RB_WR); if (flow->oflags & FLOWFDOWN) { - rx_acl |= ACL_FLOWDOWN; - tx_acl |= ACL_FLOWDOWN; + ssm_rbuff_set_bits(flow->rx_rb, RB_FLOWDOWN); + ssm_rbuff_set_bits(flow->tx_rb, RB_FLOWDOWN); ssm_flow_set_notify(flow->set, flow->info.id, FLOW_DOWN); } else { - rx_acl &= ~ACL_FLOWDOWN; - tx_acl &= ~ACL_FLOWDOWN; + ssm_rbuff_clr_bits(flow->rx_rb, RB_FLOWDOWN); + ssm_rbuff_clr_bits(flow->tx_rb, RB_FLOWDOWN); ssm_flow_set_notify(flow->set, flow->info.id, FLOW_UP); } - ssm_rbuff_set_acl(flow->rx_rb, rx_acl); - ssm_rbuff_set_acl(flow->tx_rb, tx_acl); - break; case FLOWGFLAGS: fflags = va_arg(l, uint32_t *); @@ -1218,6 +1489,59 @@ int fccntl(int fd, goto eperm; *cflags = frcti_getflags(flow->frcti); break; + case FRCTSMAXSDU: + max = va_arg(l, size_t); + if (flow->frcti == NULL) + goto eperm; + if (frcti_set_max_rcv_sdu(flow->frcti, max) < 0) + goto einval; + break; + case FRCTGMAXSDU: + maxp = va_arg(l, size_t *); + if (maxp == NULL) + goto einval; + if (flow->frcti == NULL) + goto eperm; + *maxp = frcti_get_max_rcv_sdu(flow->frcti); + break; + case FRCTSRRINGSZ: + rsz = va_arg(l, size_t); + if (flow->frcti == NULL) + goto eperm; + rc = frcti_set_rcv_ring_sz(flow->frcti, rsz); + if (rc < 0) { + pthread_rwlock_unlock(&proc.lock); + va_end(l); + return rc; + } + break; + case FRCTGRRINGSZ: + rszp = va_arg(l, size_t *); + if (rszp == NULL) + goto einval; + if (flow->frcti == NULL) + goto eperm; + *rszp = frcti_get_rcv_ring_sz(flow->frcti); + break; + case FRCTSRTOMIN: + if (flow->frcti == NULL) + goto eperm; + rto = va_arg(l, time_t); + rc = frcti_set_rto_min(flow->frcti, rto); + if (rc < 0) { + pthread_rwlock_unlock(&proc.lock); + va_end(l); + return rc; + } + break; + case FRCTGRTOMIN: + if (flow->frcti == NULL) + goto eperm; + rtop = va_arg(l, time_t *); + if (rtop == NULL) + goto einval; + *rtop = frcti_get_rto_min(flow->frcti); + break; default: pthread_rwlock_unlock(&proc.lock); va_end(l); @@ -1227,6 +1551,9 @@ int fccntl(int fd, pthread_rwlock_unlock(&proc.lock); + if (emit_eos) + frcti_fin_snd(flow->frcti); + va_end(l); return 0; @@ -1241,86 +1568,275 @@ int fccntl(int fd, return -EPERM; } -static int chk_crc(struct ssm_pk_buff * spb) -{ - uint32_t crc; - uint8_t * head = ssm_pk_buff_head(spb); - uint8_t * tail = ssm_pk_buff_tail_release(spb, CRCLEN); - - mem_hash(HASH_CRC32, &crc, head, tail - head); - - return !(crc == *((uint32_t *) tail)); -} - -static int add_crc(struct ssm_pk_buff * spb) -{ - uint8_t * head; - uint8_t * tail; - - tail = ssm_pk_buff_tail_alloc(spb, CRCLEN); - if (tail == NULL) - return -ENOMEM; - - head = ssm_pk_buff_head(spb); - mem_hash(HASH_CRC32, tail, head, tail - head); - - return 0; -} - static int flow_tx_spb(struct flow * flow, struct ssm_pk_buff * spb, + uint16_t flags, bool block, struct timespec * abstime) { struct timespec now; ssize_t idx; + size_t pci_total; int ret; clock_gettime(PTHREAD_COND_CLOCK, &now); - - pthread_rwlock_wrlock(&proc.lock); - flow->snd_act = now; - pthread_rwlock_unlock(&proc.lock); - - idx = ssm_pk_buff_get_idx(spb); - - pthread_rwlock_rdlock(&proc.lock); + idx = ssm_pk_buff_get_off(spb); if (ssm_pk_buff_len(spb) > 0) { - if (frcti_snd(flow->frcti, spb) < 0) + if (FRCTI_SND(flow->frcti, spb, flags) < 0) goto enomem; - if (spb_encrypt(flow, spb) < 0) - goto enomem; + if (flow->info.qs.ber == 0) { + pci_total = flow->frcti != NULL + ? frcti_data_hdr_len(flow->frcti) : 0; + if (crc_add(spb, pci_total) != 0) + goto enomem; + } - if (flow->info.qs.ber == 0 && add_crc(spb) != 0) + if (spb_encrypt(flow, spb) < 0) goto enomem; } - pthread_cleanup_push(__cleanup_rwlock_unlock, &proc.lock); - if (!block) ret = ssm_rbuff_write(flow->tx_rb, idx); else ret = ssm_rbuff_write_b(flow->tx_rb, idx, abstime); if (ret < 0) - ssm_pool_remove(proc.pool, idx); - else - ssm_flow_set_notify(flow->set, flow->info.id, FLOW_PKT); - - pthread_cleanup_pop(true); + return ret; + ssm_flow_set_notify(flow->set, flow->info.id, FLOW_PKT); return 0; -enomem: - pthread_rwlock_unlock(&proc.lock); - ssm_pool_remove(proc.pool, idx); + enomem: return -ENOMEM; } +/* Per-fragment role for fragment i out of n; n == 1 yields SOLE. */ +static __inline__ uint16_t flow_frag_role(size_t i, size_t n) +{ + if (n == 1) + return FRCT_FR_SOLE; + + if (i == 0) + return FRCT_FR_FIRST; + + if (i + 1 == n) + return FRCT_FR_LAST; + + return FRCT_FR_MID; +} + +static ssize_t flow_write_stream(struct flow * flow, + const void * buf, + size_t count, + int oflags, + struct timespec * dl) +{ + const uint8_t * src = buf; + size_t payload; + size_t off = 0; + bool block = !(oflags & FLOWFWNOBLOCK); + + if (!FRCTI_IS_FRTX(flow->frcti)) + return -EMSGSIZE; + + payload = FRCTI_PAYLOAD_CAP(flow->frcti); + + while (off < count) { + struct ssm_pk_buff * spb; + uint8_t * ptr; + ssize_t idx; + size_t clen; + int ret; + + ret = flow_wait_window(flow, 1, block, dl); + if (ret < 0) + return off > 0 ? (ssize_t) off : (ssize_t) ret; + + clen = MIN(count - off, payload); + + if (block) + idx = ssm_pool_alloc_b(proc.pool, clen, &ptr, + &spb, dl); + else + idx = ssm_pool_alloc(proc.pool, clen, &ptr, &spb); + if (idx < 0) + return off > 0 ? (ssize_t) off : idx; + + memcpy(ptr, src + off, clen); + + ret = flow_tx_spb(flow, spb, 0, block, dl); + if (ret < 0) { + ssm_pool_remove(proc.pool, idx); + return off > 0 ? (ssize_t) off : (ssize_t) ret; + } + + off += clen; + } + + return (ssize_t) count; +} + +/* Per-fragment flow_tx_spb loop. Raw flows refuse; FRCT splits the SDU. */ +static ssize_t flow_write_frag(struct flow * flow, + const void * buf, + size_t count, + int oflags, + struct timespec * dl) +{ + const uint8_t * src = buf; + size_t frag_payload; + size_t n; + size_t off = 0; + size_t i; + int ret; + bool block = !(oflags & FLOWFWNOBLOCK); + + /* Raw flows carry no PCI; cannot fragment. */ + if (flow->frcti == NULL) + return -EMSGSIZE; + + frag_payload = FRCTI_PAYLOAD_CAP(flow->frcti); + + /* Guard the ceil-divide against size_t overflow. */ + if (count > SIZE_MAX - frag_payload + 1) + return -EMSGSIZE; + + n = (count + frag_payload - 1) / frag_payload; + + /* SDU larger than the FC window can ever offer would deadlock. */ + if (n > RQ_SIZE) + return -EMSGSIZE; + + /* SDU-atomic FC: wait for n seqnos to avoid overshoot mid-SDU. */ + ret = flow_wait_window(flow, n, block, dl); + if (ret < 0) + return (ssize_t) ret; + + STAT_BUMP(flow->frcti, sdu_snd_frag); + + for (i = 0; i < n; ++i) { + struct ssm_pk_buff * spb; + uint8_t * ptr; + ssize_t idx; + size_t clen; + + clen = (i + 1 == n) ? (count - off) : frag_payload; + + if (block) + idx = ssm_pool_alloc_b(proc.pool, clen, &ptr, + &spb, dl); + else + idx = ssm_pool_alloc(proc.pool, clen, &ptr, &spb); + if (idx < 0) { + if (off > 0) + STAT_BUMP(flow->frcti, sdu_snd_alloc); + return off > 0 ? (ssize_t) off : idx; + } + + memcpy(ptr, src + off, clen); + + ret = flow_tx_spb(flow, spb, flow_frag_role(i, n), + block, dl); + if (ret < 0) { + ssm_pool_remove(proc.pool, idx); + if (off > 0) + STAT_BUMP(flow->frcti, sdu_snd_tx); + return off > 0 ? (ssize_t) off : (ssize_t) ret; + } + + off += clen; + } + + return (ssize_t) count; +} + +/* + * Initiator promotes on the install grace (it holds the key-confirm + * tag); responder waits for peer_synced, with a near-exhaustion floor. + */ +static void flow_tx_promote(struct flow * flow) +{ + struct timespec now; + int nodes_left; + bool promote; + + if (flow->crypt == NULL) + return; + + if (flow->rk_grace.tv_sec == 0 && flow->rk_grace.tv_nsec == 0) + return; + + promote = crypt_peer_synced(flow->crypt); + + if (!promote && flow->rk_initiator) { + clock_gettime(PTHREAD_COND_CLOCK, &now); + promote = ts_diff_ns(&now, &flow->rk_grace) >= 0; + } + + if (!promote && !flow->rk_initiator) { + nodes_left = crypt_nodes_left(flow->crypt); + promote = nodes_left >= 0 && nodes_left <= REKEY_PROMOTE_FLOOR; + } + + if (!promote) + return; + + crypt_tx_promote(flow->crypt); + flow->rk_grace.tv_sec = 0; + flow->rk_grace.tv_nsec = 0; +} + +/* The reply carries no key; the seed arrives later over RB_REKEY. */ +static int flow_rekey_trigger(struct flow * flow) +{ + struct flow_info info; + uint8_t buf[SOCK_BUF_SIZE]; + buffer_t msg = {SOCK_BUF_SIZE, buf}; + + pthread_rwlock_rdlock(&proc.lock); + if (flow->info.id < 0 || flow->crypt == NULL) { + pthread_rwlock_unlock(&proc.lock); + return -1; + } + info = flow->info; + pthread_rwlock_unlock(&proc.lock); + + if (flow_update__irm_req_ser(&msg, &info, true) < 0) + return -1; + + if (send_recv_msg(&msg) < 0) + return -1; + + return 0; +} + +static bool flow_wm_due(struct flow * flow) +{ + uint32_t tick; + + if (KEY_REKEY_WATERMARK == 0) + return false; + + if (flow->crypt == NULL) + return false; + + if (LOAD_RELAXED(&flow->rk_wm_inflight)) + return false; + + tick = FETCH_ADD_RELAXED(&flow->rk_wm_ctr, 1); + if ((tick & (FLOW_WM_CHECK - 1)) != 0) + return false; + + if (ssm_rbuff_get_flags(flow->rx_rb) & RB_REKEY) + return false; + + return crypt_nodes_left(flow->crypt) <= KEY_REKEY_WATERMARK; +} + ssize_t flow_write(int fd, const void * buf, size_t count) @@ -1330,74 +1846,90 @@ ssize_t flow_write(int fd, int ret; int flags; struct timespec abs; - struct timespec * abstime = NULL; + struct timespec now; + struct timespec * dl = NULL; struct ssm_pk_buff * spb; uint8_t * ptr; if (buf == NULL && count != 0) return -EINVAL; - if (fd < 0 || fd >= PROG_MAX_FLOWS) + if (fd < 0 || fd >= PROC_MAX_FLOWS) return -EBADF; flow = &proc.flows[fd]; - clock_gettime(PTHREAD_COND_CLOCK, &abs); - - pthread_rwlock_wrlock(&proc.lock); + pthread_rwlock_rdlock(&proc.lock); if (flow->info.id < 0) { pthread_rwlock_unlock(&proc.lock); return -ENOTALLOC; } + flags = flow->oflags; + + clock_gettime(PTHREAD_COND_CLOCK, &now); + if (flow->snd_timesout) { - ts_add(&abs, &flow->snd_timeo, &abs); - abstime = &abs; + ts_add(&now, &flow->snd_timeo, &abs); + dl = &abs; } - flags = flow->oflags; - pthread_rwlock_unlock(&proc.lock); if ((flags & FLOWFACCMODE) == FLOWFRDONLY) return -EPERM; - if (flags & FLOWFWNOBLOCK) { - if (!frcti_is_window_open(flow->frcti)) - return -EAGAIN; - idx = ssm_pool_alloc(proc.pool, count, &ptr, &spb); - } else { - ret = frcti_window_wait(flow->frcti, abstime); + if (flow->crypt != NULL + && (ssm_rbuff_get_flags(flow->rx_rb) & RB_REKEY)) + flow_rekey(flow); + + flow_tx_promote(flow); + + /* Pre-empt TX key exhaustion; the timer is the backstop. */ + if (flow_wm_due(flow)) { + STORE_RELAXED(&flow->rk_wm_inflight, true); + if (flow_rekey_trigger(flow) < 0) + STORE_RELAXED(&flow->rk_wm_inflight, false); + } + + tw_move_safe(); + + if (flow->frcti != NULL) { + /* Pump rx_rb so a pure-writer processes ACKs. */ + ret = flow_wait_window(flow, 1, !(flags & FLOWFWNOBLOCK), dl); if (ret < 0) return ret; - idx = ssm_pool_alloc_b(proc.pool, count, &ptr, &spb, abstime); + + if (count > 0 && FRCTI_IS_STREAM(flow->frcti)) + return flow_write_stream(flow, buf, count, flags, dl); + + if (FRCTI_NEEDS_FRAG(flow->frcti, count)) + return flow_write_frag(flow, buf, count, flags, dl); + } else if (flow->info.mtu > 0 + && count > flow_user_mtu(flow, flow->info.mtu)) { + /* Raw flows carry no PCI; refuse anything > one n-1 frame. */ + return -EMSGSIZE; } + if (flags & FLOWFWNOBLOCK) + idx = ssm_pool_alloc(proc.pool, count, &ptr, &spb); + else + idx = ssm_pool_alloc_b(proc.pool, count, &ptr, &spb, dl); if (idx < 0) return idx; if (count > 0) memcpy(ptr, buf, count); - ret = flow_tx_spb(flow, spb, !(flags & FLOWFWNOBLOCK), abstime); - - return ret < 0 ? (ssize_t) ret : (ssize_t) count; -} - -static bool invalid_pkt(struct flow * flow, - struct ssm_pk_buff * spb) -{ - if (spb == NULL || ssm_pk_buff_len(spb) == 0) - return true; - - if (flow->info.qs.ber == 0 && chk_crc(spb) != 0) - return true; - - if (spb_decrypt(flow, spb) < 0) - return true; + ret = flow_tx_spb(flow, spb, FRCT_FR_SOLE, + !(flags & FLOWFWNOBLOCK), dl); + if (ret < 0) { + ssm_pool_remove(proc.pool, idx); + return (ssize_t) ret; + } - return false; + return (ssize_t) count; } static ssize_t flow_rx_spb(struct flow * flow, @@ -1408,19 +1940,14 @@ static ssize_t flow_rx_spb(struct flow * flow, ssize_t idx; struct timespec now; - idx = block ? ssm_rbuff_read_b(flow->rx_rb, abstime) : - ssm_rbuff_read(flow->rx_rb); + idx = block ? ssm_rbuff_read_b(flow->rx_rb, abstime) + : ssm_rbuff_read(flow->rx_rb); if (idx < 0) return idx; clock_gettime(PTHREAD_COND_CLOCK, &now); - - pthread_rwlock_wrlock(&proc.lock); - flow->rcv_act = now; - pthread_rwlock_unlock(&proc.lock); - *spb = ssm_pool_get(proc.pool, idx); if (invalid_pkt(flow, *spb)) { @@ -1431,28 +1958,128 @@ static ssize_t flow_rx_spb(struct flow * flow, return idx; } +static ssize_t raw_flow_read_pkt(struct flow * flow, + bool block, + struct timespec * dl) +{ + struct ssm_pk_buff * spb; + struct timespec wait_abs; + ssize_t idx; + + while (true) { + if (flow->crypt != NULL + && (ssm_rbuff_get_flags(flow->rx_rb) & RB_REKEY)) + flow_rekey(flow); + + if (!block) { + idx = ssm_rbuff_read(flow->rx_rb); + if (idx < 0) + return -EAGAIN; + } else { + compute_wait_deadline(dl, &wait_abs); + idx = ssm_rbuff_read_b(flow->rx_rb, &wait_abs); + if (idx == -ETIMEDOUT) { + if (deadline_passed(dl)) + return -ETIMEDOUT; + continue; + } + if (idx < 0) + return idx; + } + + spb = ssm_pool_get(proc.pool, idx); + if (!invalid_pkt(flow, spb)) + return idx; + + ssm_pool_remove(proc.pool, idx); + if (!block) + return -EAGAIN; + } +} + +static ssize_t deliver_pkt(struct flow * flow, + struct ssm_pk_buff * spb, + ssize_t idx, + void * buf, + size_t count, + bool partrd) +{ + uint8_t * packet = ssm_pk_buff_head(spb); + ssize_t n = ssm_pk_buff_len(spb); + + assert(n >= 0); + + if (n <= (ssize_t) count) { + memcpy(buf, packet, n); + ipcp_spb_release(spb); + if (partrd && n == (ssize_t) count) + flow->part_idx = DONE_PART; + else + flow->part_idx = NO_PART; + + return n; + } + + if (partrd) { + memcpy(buf, packet, count); + ssm_pk_buff_pop(spb, n); + flow->part_idx = idx; + return count; + } + + ipcp_spb_release(spb); + return -EMSGSIZE; +} + +/* Drive frcti_consume until it delivers or errors. */ +static ssize_t flow_read_frcti(struct flow * flow, + void * buf, + size_t count, + bool block, + struct timespec * dl) +{ + struct timespec now; + ssize_t bytes; + int rc; + + while (true) { + flow_drain_rx_nb(flow); + bytes = FRCTI_CONSUME(flow->frcti, buf, count); + if (bytes >= 0) + break; + if (bytes != -EAGAIN) + return bytes; + if (!block) + return -EAGAIN; + rc = flow_rx_one(flow, dl); + if (rc < 0) + return rc; + } + + clock_gettime(PTHREAD_COND_CLOCK, &now); + flow->rcv_act = now; + + return bytes; +} + ssize_t flow_read(int fd, void * buf, size_t count) { - ssize_t idx; - ssize_t n; - uint8_t * packet; + struct flow * flow; struct ssm_pk_buff * spb; struct timespec abs; struct timespec now; - struct timespec * abstime = NULL; - struct flow * flow; + struct timespec * dl = NULL; + ssize_t idx; bool block; bool partrd; - if (fd < 0 || fd >= PROG_MAX_FLOWS) + if (fd < 0 || fd >= PROC_MAX_FLOWS) return -EBADF; flow = &proc.flows[fd]; - clock_gettime(PTHREAD_COND_CLOCK, &now); - pthread_rwlock_rdlock(&proc.lock); if (flow->info.id < 0) { @@ -1461,8 +2088,8 @@ ssize_t flow_read(int fd, } if (flow->part_idx == DONE_PART) { - pthread_rwlock_unlock(&proc.lock); flow->part_idx = NO_PART; + pthread_rwlock_unlock(&proc.lock); return 0; } @@ -1470,75 +2097,40 @@ ssize_t flow_read(int fd, partrd = !(flow->oflags & FLOWFRNOPART); if (flow->rcv_timesout) { + clock_gettime(PTHREAD_COND_CLOCK, &now); ts_add(&now, &flow->rcv_timeo, &abs); - abstime = &abs; - } - - idx = flow->part_idx; - if (idx < 0) { - while ((idx = frcti_queued_pdu(flow->frcti)) < 0) { - pthread_rwlock_unlock(&proc.lock); - - idx = flow_rx_spb(flow, &spb, block, abstime); - if (idx < 0) { - if (block && idx != -EAGAIN) - return idx; - if (!block) - return idx; - - pthread_rwlock_rdlock(&proc.lock); - continue; - } - - pthread_rwlock_rdlock(&proc.lock); - - frcti_rcv(flow->frcti, spb); - } + dl = &abs; } - spb = ssm_pool_get(proc.pool, idx); - pthread_rwlock_unlock(&proc.lock); - packet = ssm_pk_buff_head(spb); + if (flow->crypt != NULL + && (ssm_rbuff_get_flags(flow->rx_rb) & RB_REKEY)) + flow_rekey(flow); - n = ssm_pk_buff_len(spb); + /* Advance TX off a stale epoch even on recv-mostly (ACK-only) flows. */ + flow_tx_promote(flow); - assert(n >= 0); + tw_move_safe(); - if (n <= (ssize_t) count) { - memcpy(buf, packet, n); - ipcp_spb_release(spb); - - pthread_rwlock_wrlock(&proc.lock); - - flow->part_idx = (partrd && n == (ssize_t) count) ? - DONE_PART : NO_PART; + idx = flow->part_idx; + if (idx < 0 && flow->frcti != NULL) + return flow_read_frcti(flow, buf, count, block, dl); - flow->rcv_act = now; + if (idx < 0) { + idx = raw_flow_read_pkt(flow, block, dl); + if (idx < 0) + return idx; + } - pthread_rwlock_unlock(&proc.lock); - return n; - } else { - if (partrd) { - memcpy(buf, packet, count); - ssm_pk_buff_head_release(spb, n); - pthread_rwlock_wrlock(&proc.lock); - flow->part_idx = idx; + spb = ssm_pool_get(proc.pool, idx); - flow->rcv_act = now; + clock_gettime(PTHREAD_COND_CLOCK, &now); + flow->rcv_act = now; - pthread_rwlock_unlock(&proc.lock); - return count; - } else { - ipcp_spb_release(spb); - return -EMSGSIZE; - } - } + return deliver_pkt(flow, spb, idx, buf, count, partrd); } -/* fqueue functions. */ - struct flow_set * fset_create(void) { struct flow_set * set; @@ -1614,7 +2206,7 @@ int fset_add(struct flow_set * set, struct flow * flow; int ret; - if (set == NULL || fd < 0 || fd >= SYS_MAX_FLOWS) + if (set == NULL || fd < 0 || fd >= PROC_MAX_FLOWS) return -EINVAL; flow = &proc.flows[fd]; @@ -1650,7 +2242,7 @@ void fset_del(struct flow_set * set, { struct flow * flow; - if (set == NULL || fd < 0 || fd >= SYS_MAX_FLOWS) + if (set == NULL || fd < 0 || fd >= PROC_MAX_FLOWS) return; flow = &proc.flows[fd]; @@ -1661,7 +2253,7 @@ void fset_del(struct flow_set * set, ssm_flow_set_del(proc.fqset, set->idx, flow->info.id); if (flow->frcti != NULL) - ssm_flow_set_add(proc.fqset, 0, proc.flows[fd].info.id); + ssm_flow_set_add(proc.fqset, 0, flow->info.id); pthread_rwlock_unlock(&proc.lock); } @@ -1672,7 +2264,7 @@ bool fset_has(const struct flow_set * set, struct flow * flow; bool ret; - if (set == NULL || fd < 0 || fd >= SYS_MAX_FLOWS) + if (set == NULL || fd < 0 || fd >= PROC_MAX_FLOWS) return false; flow = &proc.flows[fd]; @@ -1691,61 +2283,71 @@ bool fset_has(const struct flow_set * set, return ret; } -/* Filter fqueue events for non-data packets */ static int fqueue_filter(struct fqueue * fq) { struct ssm_pk_buff * spb; int fd; ssize_t idx; struct frcti * frcti; + int ret = 0; + + /* proc.lock rdlock gates frcti_destroy via flow_fini wrlock. */ + pthread_rwlock_rdlock(&proc.lock); while (fq->next < fq->fqsize) { - if (fq->fqueue[fq->next].event != FLOW_PKT) - return 1; + if (fq->fqueue[fq->next].event == FLOW_UPD) { + /* Re-key doorbell: pull internally, never surface. */ + fd = proc.id_to_fd[fq->fqueue[fq->next].flow_id].fd; + ++fq->next; + if (fd >= 0) { + pthread_rwlock_unlock(&proc.lock); + flow_rekey(&proc.flows[fd]); + pthread_rwlock_rdlock(&proc.lock); + } + continue; + } - pthread_rwlock_rdlock(&proc.lock); + if (fq->fqueue[fq->next].event != FLOW_PKT) { + ret = 1; + goto out; + } fd = proc.id_to_fd[fq->fqueue[fq->next].flow_id].fd; if (fd < 0) { ++fq->next; - pthread_rwlock_unlock(&proc.lock); continue; } frcti = proc.flows[fd].frcti; if (frcti == NULL) { - pthread_rwlock_unlock(&proc.lock); - return 1; + ret = 1; + goto out; } - if (__frcti_pdu_ready(frcti) >= 0) { - pthread_rwlock_unlock(&proc.lock); - return 1; + if (FRCTI_PDU_READY(frcti)) { + ret = 1; + goto out; } - pthread_rwlock_unlock(&proc.lock); - idx = flow_rx_spb(&proc.flows[fd], &spb, false, NULL); if (idx < 0) - return 0; - - pthread_rwlock_rdlock(&proc.lock); + goto out; spb = ssm_pool_get(proc.pool, idx); - __frcti_rcv(frcti, spb); + FRCTI_RCV(frcti, spb); - if (__frcti_pdu_ready(frcti) >= 0) { - pthread_rwlock_unlock(&proc.lock); - return 1; + if (FRCTI_PDU_READY(frcti)) { + ret = 1; + goto out; } - pthread_rwlock_unlock(&proc.lock); - ++fq->next; } - return 0; + out: + pthread_rwlock_unlock(&proc.lock); + return ret; } int fqueue_next(struct fqueue * fq) @@ -1792,7 +2394,8 @@ ssize_t fevent(struct flow_set * set, { ssize_t ret = 0; struct timespec abs; - struct timespec * t = NULL; + struct timespec * dl = NULL; + struct timespec wait_abs; if (set == NULL || fq == NULL) return -EINVAL; @@ -1800,17 +2403,26 @@ ssize_t fevent(struct flow_set * set, if (fq->fqsize > 0 && fq->next != fq->fqsize) return 1; - clock_gettime(PTHREAD_COND_CLOCK, &abs); - if (timeo != NULL) { - ts_add(&abs, timeo, &abs); - t = &abs; + struct timespec now; + clock_gettime(PTHREAD_COND_CLOCK, &now); + ts_add(&now, timeo, &abs); + dl = &abs; } while (ret == 0) { - ret = ssm_flow_set_wait(proc.fqset, set->idx, fq->fqueue, t); - if (ret == -ETIMEDOUT) - return -ETIMEDOUT; + tw_move_safe(); + + compute_wait_deadline(dl, &wait_abs); + + ret = ssm_flow_set_wait(proc.fqset, set->idx, + fq->fqueue, &wait_abs); + if (ret == -ETIMEDOUT) { + if (deadline_passed(dl)) + return -ETIMEDOUT; + ret = 0; + continue; + } fq->fqsize = ret; fq->next = 0; @@ -1823,13 +2435,12 @@ ssize_t fevent(struct flow_set * set, return 1; } -/* ipcp-dev functions. */ - int np1_flow_alloc(pid_t n_pid, int flow_id) { struct flow_info flow; - struct crypt_sk crypt = { .nid = NID_undef, .key = NULL }; + struct crypt_sk crypt = { .nid = NID_undef, .key = NULL, + .epoch = 0, .role = CRYPT_ROLE_INIT }; memset(&flow, 0, sizeof(flow)); @@ -1837,9 +2448,10 @@ int np1_flow_alloc(pid_t n_pid, flow.n_pid = getpid(); flow.qs = qos_np1; flow.mpl = 0; - flow.n_1_pid = n_pid; /* This "flow" is upside-down! */ + /* np1 flow: n_1_pid is the upper. */ + flow.n_1_pid = n_pid; - return flow_init(&flow, &crypt); + return flow_init(&flow, &crypt, 0); } int np1_flow_dealloc(int flow_id, @@ -1847,12 +2459,7 @@ int np1_flow_dealloc(int flow_id, { int fd; - /* - * TODO: Don't pass timeo to the IPCP but wait in IRMd. - * This will need async ops, waiting until we bootstrap - * the IRMd over ouroboros. - */ - + /* TODO: wait in IRMd, not here; needs async ops. */ sleep(timeo); pthread_rwlock_rdlock(&proc.lock); @@ -1881,6 +2488,38 @@ int np1_flow_resp(int flow_id, return fd; } +int np1_flow_fd(int flow_id) +{ + int fd; + + if (flow_id < 0 || flow_id >= SYS_MAX_FLOWS) + return -1; + + pthread_rwlock_rdlock(&proc.lock); + + fd = proc.id_to_fd[flow_id].fd; + + pthread_rwlock_unlock(&proc.lock); + + return fd; +} + +int np1_flow_id(int fd) +{ + int flow_id; + + if (fd < 0 || fd >= PROC_MAX_FLOWS) + return -1; + + pthread_rwlock_rdlock(&proc.lock); + + flow_id = proc.flows[fd].info.id; + + pthread_rwlock_unlock(&proc.lock); + + return flow_id; +} + int ipcp_create_r(const struct ipcp_info * info) { uint8_t buf[SOCK_BUF_SIZE]; @@ -1900,6 +2539,7 @@ int ipcp_create_r(const struct ipcp_info * info) int ipcp_flow_req_arr(const buffer_t * dst, qosspec_t qs, time_t mpl, + uint32_t mtu, const buffer_t * data) { struct flow_info flow; @@ -1916,6 +2556,7 @@ int ipcp_flow_req_arr(const buffer_t * dst, flow.n_1_pid = getpid(); flow.qs = qs; flow.mpl = mpl; + flow.mtu = mtu; if (ipcp_flow_req_arr__irm_req_ser(&msg, dst, &flow, data) < 0) return -ENOMEM; @@ -1924,28 +2565,56 @@ int ipcp_flow_req_arr(const buffer_t * dst, if (err < 0) return err; - crypt.key = key; + crypt.key = key; + crypt.epoch = 0; + crypt.role = CRYPT_ROLE_INIT; err = flow__irm_result_des(&msg, &flow, &crypt); if (err < 0) return err; - assert(crypt.nid == NID_undef); /* np1 flows are not encrypted */ + /* np1 flows are not encrypted. */ + assert(crypt.nid == NID_undef); - /* inverted for np1_flow */ + /* Inverted for np1_flow. */ flow.n_1_pid = flow.n_pid; flow.n_pid = getpid(); flow.mpl = 0; + flow.mtu = 0; flow.qs = qos_np1; crypt.nid = NID_undef; - return flow_init(&flow, &crypt); + return flow_init(&flow, &crypt, 0); +} + +int ipcp_flow_update_arr(int flow_id, + const buffer_t * data) +{ + struct flow_info flow; + uint8_t buf[SOCK_BUF_SIZE]; + buffer_t msg = {SOCK_BUF_SIZE, buf}; + int err; + + memset(&flow, 0, sizeof(flow)); + + flow.id = flow_id; + flow.n_1_pid = getpid(); + + if (ipcp_flow_update_arr__irm_req_ser(&msg, &flow, data) < 0) + return -ENOMEM; + + err = send_recv_msg(&msg); + if (err < 0) + return err; + + return irm__irm_result_des(&msg); } int ipcp_flow_alloc_reply(int fd, int response, time_t mpl, + uint32_t mtu, const buffer_t * data) { struct flow_info flow; @@ -1953,7 +2622,7 @@ int ipcp_flow_alloc_reply(int fd, buffer_t msg = {SOCK_BUF_SIZE, buf}; int err; - assert(fd >= 0 && fd < SYS_MAX_FLOWS); + assert(fd >= 0 && fd < PROC_MAX_FLOWS); pthread_rwlock_rdlock(&proc.lock); @@ -1962,6 +2631,7 @@ int ipcp_flow_alloc_reply(int fd, pthread_rwlock_unlock(&proc.lock); flow.mpl = mpl; + flow.mtu = mtu; if (ipcp_flow_alloc_reply__irm_msg_ser(&msg, &flow, response, data) < 0) return -ENOMEM; @@ -1979,25 +2649,33 @@ int ipcp_flow_read(int fd, struct flow * flow; ssize_t idx = -1; - assert(fd >= 0 && fd < SYS_MAX_FLOWS); + assert(fd >= 0 && fd < PROC_MAX_FLOWS); assert(spb); flow = &proc.flows[fd]; pthread_rwlock_rdlock(&proc.lock); - assert(flow->info.id >= 0); + if (flow->info.id < 0) { + pthread_rwlock_unlock(&proc.lock); + return -ENOTALLOC; + } - while (frcti_queued_pdu(flow->frcti) < 0) { + /* Raw flow: deliver the popped pkt directly (no FRCT rq). */ + if (flow->frcti == NULL) { + idx = flow_rx_spb(flow, spb, false, NULL); pthread_rwlock_unlock(&proc.lock); + return idx < 0 ? (int) idx : 0; + } + while (!FRCTI_PDU_READY(flow->frcti)) { idx = flow_rx_spb(flow, spb, false, NULL); - if (idx < 0) + if (idx < 0) { + pthread_rwlock_unlock(&proc.lock); return idx; + } - pthread_rwlock_rdlock(&proc.lock); - - frcti_rcv(flow->frcti, *spb); + FRCTI_RCV(flow->frcti, *spb); } pthread_rwlock_unlock(&proc.lock); @@ -2011,12 +2689,12 @@ int ipcp_flow_write(int fd, struct flow * flow; int ret; - assert(fd >= 0 && fd < SYS_MAX_FLOWS); + assert(fd >= 0 && fd < PROC_MAX_FLOWS); assert(spb); flow = &proc.flows[fd]; - pthread_rwlock_wrlock(&proc.lock); + pthread_rwlock_rdlock(&proc.lock); if (flow->info.id < 0) { pthread_rwlock_unlock(&proc.lock); @@ -2030,30 +2708,28 @@ int ipcp_flow_write(int fd, pthread_rwlock_unlock(&proc.lock); - ret = flow_tx_spb(flow, spb, true, NULL); + ret = flow_tx_spb(flow, spb, FRCT_FR_SOLE, true, NULL); return ret; } -static int pool_copy_spb(struct ssm_pool * src_pool, - ssize_t src_idx, - struct ssm_pool * dst_pool, - struct ssm_pk_buff ** dst_spb) +/* Copy src into dst_pool without consuming src. Caller owns both halves. */ +static int pool_dup_spb(struct ssm_pool * src_pool, + size_t src_off, + struct ssm_pool * dst_pool, + struct ssm_pk_buff ** dst_spb) { struct ssm_pk_buff * src; uint8_t * ptr; size_t len; - src = ssm_pool_get(src_pool, src_idx); + src = ssm_pool_get(src_pool, src_off); len = ssm_pk_buff_len(src); - if (ssm_pool_alloc(dst_pool, len, &ptr, dst_spb) < 0) { - ssm_pool_remove(src_pool, src_idx); + if (ssm_pool_alloc(dst_pool, len, &ptr, dst_spb) < 0) return -ENOMEM; - } memcpy(ptr, ssm_pk_buff_head(src), len); - ssm_pool_remove(src_pool, src_idx); return 0; } @@ -2063,31 +2739,37 @@ int np1_flow_read(int fd, struct ssm_pool * pool) { struct flow * flow; - ssize_t idx = -1; + ssize_t off; - assert(fd >= 0 && fd < SYS_MAX_FLOWS); + assert(fd >= 0 && fd < PROC_MAX_FLOWS); assert(spb); flow = &proc.flows[fd]; - assert(flow->info.id >= 0); - pthread_rwlock_rdlock(&proc.lock); - idx = ssm_rbuff_read(flow->rx_rb); - if (idx < 0) { + if (flow->info.id < 0) { pthread_rwlock_unlock(&proc.lock); - return idx; + return -ENOTALLOC; + } + + off = ssm_rbuff_read(flow->rx_rb); + if (off < 0) { + pthread_rwlock_unlock(&proc.lock); + return off; } pthread_rwlock_unlock(&proc.lock); if (pool == NULL) { - *spb = ssm_pool_get(proc.pool, idx); + *spb = ssm_pool_get(proc.pool, off); } else { /* Cross-pool copy: PUP -> GSPP */ - if (pool_copy_spb(pool, idx, proc.pool, spb) < 0) + if (pool_dup_spb(pool, off, proc.pool, spb) < 0) { + ssm_pool_remove(pool, off); return -ENOMEM; + } + ssm_pool_remove(pool, off); } return 0; @@ -2100,9 +2782,10 @@ int np1_flow_write(int fd, struct flow * flow; struct ssm_pk_buff * dst; int ret; - ssize_t idx; + size_t off; + size_t dst_off; - assert(fd >= 0 && fd < SYS_MAX_FLOWS); + assert(fd >= 0 && fd < PROC_MAX_FLOWS); assert(spb); flow = &proc.flows[fd]; @@ -2121,45 +2804,47 @@ int np1_flow_write(int fd, pthread_rwlock_unlock(&proc.lock); - idx = ssm_pk_buff_get_idx(spb); + off = ssm_pk_buff_get_off(spb); if (pool == NULL) { - ret = ssm_rbuff_write_b(flow->tx_rb, idx, NULL); + ret = ssm_rbuff_write_b(flow->tx_rb, off, NULL); if (ret < 0) - ssm_pool_remove(proc.pool, idx); - else - ssm_flow_set_notify(flow->set, flow->info.id, FLOW_PKT); + return ret; + ssm_flow_set_notify(flow->set, flow->info.id, FLOW_PKT); } else { - /* Cross-pool copy: GSPP -> PUP */ - if (pool_copy_spb(proc.pool, idx, pool, &dst) < 0) + /* Cross-pool copy: GSPP -> PUP. Src kept on error. */ + if (pool_dup_spb(proc.pool, off, pool, &dst) < 0) return -ENOMEM; - idx = ssm_pk_buff_get_idx(dst); - ret = ssm_rbuff_write_b(flow->tx_rb, idx, NULL); - if (ret < 0) - ssm_pool_remove(pool, idx); - else - ssm_flow_set_notify(flow->set, flow->info.id, FLOW_PKT); + dst_off = ssm_pk_buff_get_off(dst); + ret = ssm_rbuff_write_b(flow->tx_rb, dst_off, NULL); + if (ret < 0) { + ssm_pool_remove(pool, dst_off); + return ret; + } + ssm_flow_set_notify(flow->set, flow->info.id, FLOW_PKT); + ssm_pool_remove(proc.pool, off); } - return ret; + return 0; } int ipcp_spb_reserve(struct ssm_pk_buff ** spb, size_t len) { - return ssm_pool_alloc_b(proc.pool, len, NULL, spb, NULL) < 0 ? -1 : 0; + return ssm_pool_alloc_b(proc.pool, len, NULL, spb, NULL) < 0 + ? -1 : 0; } void ipcp_spb_release(struct ssm_pk_buff * spb) { - ssm_pool_remove(proc.pool, ssm_pk_buff_get_idx(spb)); + ssm_pool_remove(proc.pool, ssm_pk_buff_get_off(spb)); } int ipcp_flow_fini(int fd) { struct ssm_rbuff * rx_rb; - assert(fd >= 0 && fd < SYS_MAX_FLOWS); + assert(fd >= 0 && fd < PROC_MAX_FLOWS); pthread_rwlock_rdlock(&proc.lock); @@ -2168,8 +2853,8 @@ int ipcp_flow_fini(int fd) return -1; } - ssm_rbuff_set_acl(proc.flows[fd].rx_rb, ACL_FLOWDOWN); - ssm_rbuff_set_acl(proc.flows[fd].tx_rb, ACL_FLOWDOWN); + ssm_rbuff_set_bits(proc.flows[fd].rx_rb, RB_FLOWDOWN); + ssm_rbuff_set_bits(proc.flows[fd].tx_rb, RB_FLOWDOWN); ssm_flow_set_notify(proc.flows[fd].set, proc.flows[fd].info.id, @@ -2188,7 +2873,7 @@ int ipcp_flow_fini(int fd) int ipcp_flow_get_qoscube(int fd, qoscube_t * cube) { - assert(fd >= 0 && fd < SYS_MAX_FLOWS); + assert(fd >= 0 && fd < PROC_MAX_FLOWS); assert(cube); pthread_rwlock_rdlock(&proc.lock); @@ -2227,7 +2912,7 @@ int local_flow_transfer(int src_fd, struct ssm_pk_buff * dst_spb; struct ssm_pool * sp; struct ssm_pool * dp; - ssize_t idx; + ssize_t off; int ret; assert(src_fd >= 0); @@ -2241,15 +2926,15 @@ int local_flow_transfer(int src_fd, pthread_rwlock_rdlock(&proc.lock); - idx = ssm_rbuff_read(src_flow->rx_rb); - if (idx < 0) { + off = ssm_rbuff_read(src_flow->rx_rb); + if (off < 0) { pthread_rwlock_unlock(&proc.lock); - return idx; + return off; } if (dst_flow->info.id < 0) { pthread_rwlock_unlock(&proc.lock); - ssm_pool_remove(sp, idx); + ssm_pool_remove(sp, off); return -ENOTALLOC; } @@ -2257,21 +2942,24 @@ int local_flow_transfer(int src_fd, if (sp == dp) { /* Same pool: zero-copy */ - ret = ssm_rbuff_write_b(dst_flow->tx_rb, idx, NULL); + ret = ssm_rbuff_write_b(dst_flow->tx_rb, off, NULL); if (ret < 0) - ssm_pool_remove(sp, idx); + ssm_pool_remove(sp, off); else ssm_flow_set_notify(dst_flow->set, dst_flow->info.id, FLOW_PKT); } else { /* Different pools: single copy */ - if (pool_copy_spb(sp, idx, dp, &dst_spb) < 0) + if (pool_dup_spb(sp, off, dp, &dst_spb) < 0) { + ssm_pool_remove(sp, off); return -ENOMEM; + } - idx = ssm_pk_buff_get_idx(dst_spb); - ret = ssm_rbuff_write_b(dst_flow->tx_rb, idx, NULL); + ssm_pool_remove(sp, off); + off = ssm_pk_buff_get_off(dst_spb); + ret = ssm_rbuff_write_b(dst_flow->tx_rb, off, NULL); if (ret < 0) - ssm_pool_remove(dp, idx); + ssm_pool_remove(dp, off); else ssm_flow_set_notify(dst_flow->set, dst_flow->info.id, FLOW_PKT); diff --git a/src/lib/frct.c b/src/lib/frct.c index fad2cf69..c055433d 100644 --- a/src/lib/frct.c +++ b/src/lib/frct.c @@ -1,7 +1,7 @@ /* * Ouroboros - Copyright (C) 2016 - 2026 * - * Flow and Retransmission Control + * Flow and Retransmission Control Task (FRCT) * * Dimitri Staessens <dimitri@ouroboros.rocks> * Sander Vrijders <sander@ouroboros.rocks> @@ -20,97 +20,416 @@ * Foundation, Inc., http://www.fsf.org/about/contact/. */ -#include <ouroboros/endian.h> +/* Included by dev.c; uses dev.c statics (proc, spb_encrypt, ...). */ #define DELT_RDV (100 * MILLION) /* ns */ -#define MAX_RDV (1 * BILLION) /* ns */ +#define MAX_RDV (1 * BILLION) /* ns */ + +#define MAX_RTO_MUL 8 /* caps the RTO backoff shift */ +#define MAX_TLP_PER_EP 2 /* RFC 8985 §7.3: up to 2 TLPs */ +#define INITIAL_RTO (1 * BILLION) /* RFC 6298 §2.1: 1 s default */ +#define RTT_BOOT_NS (10 * MILLION) /* rtt_hint floor + initial mdev */ +#define SRTT_FLOOR_NS 1000L /* 1 us; smoothed RTT floor */ +#define MDEV_FLOOR_NS 100L /* 100 ns; mdev sanity floor */ +#define RTT_CLAMP_MUL 16 /* probe sample cap = N * srtt */ +#define MIN_RTT_WIN_NS (300ULL * BILLION) /* 5 min, Linux tcp default */ +#define NACK_COOLDOWN_NS (100 * MILLION) /* pre-DRF NACK cooldown */ +#define FRCT_TX_TIMEO_NS (250 * 1000) /* tx ring write deadline */ +#define ACK_DELAY_NS (2ULL * TICTIME) /* delayed-ACK fire delay */ #define FRCT "frct" #define FRCT_PCILEN (sizeof(struct frct_pci)) #define FRCT_NAME_STRLEN 32 -struct frct_cr { - uint32_t lwe; /* Left window edge */ - uint32_t rwe; /* Right window edge */ +/* Wire-protocol cap on SACK blocks per packet; binds both peers. */ +#define SACK_MAX_BLOCKS 2048 +#define SACK_BLOCK_SIZE (2 * sizeof(uint32_t)) +/* 2B count + 2B pad to 4-byte align the block list. */ +#define SACK_HDR_SIZE (sizeof(uint32_t)) +#define SACK_MIN_GAP_NS (250u * 1000u) /* 250 us SACK gap */ +#define MIN_REORDER_NS (250u * 1000u) /* 250 us RACK floor */ +#define SACK_RXM_MAX 32 /* Cap on retransmits staged from single SACK.*/ +#define DUP_THRESH 3 /* RFC 8985 §6.2 step 2.2 SACK count gate. */ + +/* RFC 8985 §7.2 RACK reorder-window scaling cap. */ +#define REO_WND_MULT_MAX 20 +/* RFC 8985 §7.2 step 5: round trips of no DSACK before halving. */ +#define REO_DECAY_PKTS 16 +/* DSACK seqno sanity: reject reports older/farther than one rcv window. */ +#define MAX_DSACK_LAG RQ_SIZE + +/* Signed ns elapsed; negative under concurrent update (no underflow). */ +static __inline__ int64_t ts_age_ns(uint64_t now_ns, + uint64_t then_ns) +{ + return (int64_t)(now_ns - then_ns); +} - uint8_t cflags; - uint32_t seqno; /* SEQ to send, or last SEQ Ack'd */ +/* True iff strictly more than thr_ns elapsed since then_ns. */ +static __inline__ bool ts_aged_ns(uint64_t now_ns, + uint64_t then_ns, + uint64_t thr_ns) +{ + return ts_age_ns(now_ns, then_ns) > (int64_t) thr_ns; +} - struct timespec act; /* Last seen activity */ - time_t inact; /* Inactivity (s) */ -}; +/* FRCT r-timer: do not retransmit packet older than t_r (from first send). */ +#define RXM_AGED_OUT(t0, now_ns, t_r) \ + ts_aged_ns((now_ns), (t0), (uint64_t)(t_r)) -struct frcti { - int fd; +/* FRCT a-timer: do not (re)transmit ACK after t_a from last data receive. */ +#define ACK_AGED_OUT(act, now_ns, t_a) \ + ts_aged_ns((now_ns), (act), (uint64_t)(t_a)) - time_t mpl; - time_t a; - time_t r; - time_t rdv; - - time_t srtt; /* Smoothed rtt */ - time_t mdev; /* Deviation */ - time_t rto; /* Retransmission timeout */ - uint32_t rttseq; - struct timespec t_probe; /* Probe time */ - bool probe; /* Probe active */ -#ifdef PROC_FLOW_STATS - size_t n_rtx; /* Number of rxm packets */ - size_t n_prb; /* Number of rtt probes */ - size_t n_rtt; /* Number of estimates */ - size_t n_dup; /* Duplicates received */ - size_t n_dak; /* Delayed ACKs received */ - size_t n_rdv; /* Number of rdv packets */ - size_t n_out; /* Packets out of window */ - size_t n_rqo; /* Packets out of rqueue */ -#endif - struct frct_cr snd_cr; - struct frct_cr rcv_cr; +struct sack_args { + uint16_t n; + bool dsack; /* RFC 2883: block[0] is a DSACK report */ + uint32_t ack; + uint32_t rwe; + uint32_t blocks[][2]; /* flexible — sized at alloc time */ +}; +/* NewReno-careful (RFC 6582) exit pad; gates RTT samples post-signal. */ +#define RTT_QUARANTINE 32 +#define RTTP_NONCE_LEN 16 - ssize_t rq[RQ_SIZE]; - pthread_rwlock_t lock; +/* RTT-probe wire payload (after the FRCT PCI). */ +struct frct_rttp { + uint32_t probe_id; /* sender counter; 0 on reply */ + uint32_t echo_id; /* peer's probe_id; 0 outbound */ + uint8_t nonce[RTTP_NONCE_LEN]; /* random; echoed verbatim */ +} __attribute__((packed)); - bool open; /* Window open/closed */ - struct timespec t_wnd; /* Window closed time */ - struct timespec t_rdvs; /* Last rendez-vous sent */ - pthread_cond_t cond; - pthread_mutex_t mtx; -}; +#define RTTP_PAYLOAD sizeof(struct frct_rttp) +#define RTTP_POS(id) ((id) & (RTTP_RING - 1)) +/* + * Flag values are assigned MSB-first on the wire (RFC convention): + * bit 0 = 0x8000 occupies wire-position 0 of the 16-bit flags + * field, bit 12 = 0x0008 is the last assigned bit, and the three + * LSBs (0x0007) are reserved. + */ enum frct_flags { - FRCT_DATA = 0x01, /* PDU carries data */ - FRCT_DRF = 0x02, /* Data run flag */ - FRCT_ACK = 0x04, /* ACK field valid */ - FRCT_FC = 0x08, /* FC window valid */ - FRCT_RDVS = 0x10, /* Rendez-vous */ - FRCT_FFGM = 0x20, /* First Fragment */ - FRCT_MFGM = 0x40, /* More fragments */ + FRCT_DATA = 0x8000, /* PDU carries data */ + FRCT_DRF = 0x4000, /* Data run flag */ + FRCT_ACK = 0x2000, /* ACK field valid */ + FRCT_NACK = 0x1000, /* Neg-ACK: pci->seqno is arrival_seqno - 1 */ + FRCT_FC = 0x0800, /* FC window valid */ + FRCT_RDVS = 0x0400, /* Rendez-vous */ + FRCT_FFGM = 0x0200, /* First fragment (begin) */ + FRCT_LFGM = 0x0100, /* Last fragment (end) */ + FRCT_RXM = 0x0080, /* Retransmission */ + FRCT_SACK = 0x0040, /* SACK block list follows */ + FRCT_RTTP = 0x0020, /* RTT probe / echo */ + FRCT_KA = 0x0010, /* Keepalive */ + FRCT_FIN = 0x0008, /* End of stream */ }; -struct frct_pci { - uint8_t flags; +/* + * DATA-packet fragment role (FFGM = begin, LFGM = end), SCTP-style: + * 1 1 = sole / un-fragmented SDU (begin AND end) + * 1 0 = first fragment of a multi-fragment SDU + * 0 0 = middle fragment + * 0 1 = last fragment + */ +#define FRCT_FR_MASK (FRCT_FFGM | FRCT_LFGM) +#define FRCT_FR_SOLE (FRCT_FFGM | FRCT_LFGM) +#define FRCT_FR_FIRST (FRCT_FFGM) +#define FRCT_FR_MID (0) +#define FRCT_FR_LAST (FRCT_LFGM) + +/* Default cap on a single reassembled SDU. App can raise via FRCTSMAXSDU */ +#define FRCT_MAX_SDU (1U << 20) + +/* Stream-mode PCI extension: [start, end) byte range on every DATA pkt. */ +struct frct_pci_stream { + uint32_t start; + uint32_t end; +} __attribute__((packed)); + +#define FRCT_PCI_STREAM_LEN (sizeof(struct frct_pci_stream)) - uint8_t pad; /* 24 bit window! */ - uint16_t window; +/* Bytes following PCI: SACK list / RTTP nonce / control payload. */ +#define FRCT_BODY(pci) ((uint8_t *) (pci) + FRCT_PCILEN) +/* Typed access to the stream PCI extension on stream DATA packets. */ +#define FRCT_SPCI(pci) \ + ((struct frct_pci_stream *) ((uint8_t *) (pci) + FRCT_PCILEN)) +/* Push the FRCT header onto spb's head. */ +#define FRCT_HDR_PUSH(spb, frcti) \ + ((struct frct_pci *) ssm_pk_buff_push((spb), \ + frcti_data_hdr_len(frcti))) + +/* Pop a fixed-size header off spb's head; cast to type *. */ +#define FRCT_HDR_POP(spb, type) \ + ((struct type *) ssm_pk_buff_pop((spb), sizeof(struct type))) + +/* Default / max per-flow stream rx ring (pow2); min N * per_pkt. */ +#define FRCT_STREAM_RING_MIN_PKTS 4 +#define FRCT_STREAM_RING_SZ (1U << 20) /* 1 MiB default */ +#define FRCT_STREAM_RING_SZ_MAX (1U << 27) /* 128 MiB */ + +struct frct_pci { + uint16_t flags; + uint16_t hcs; + + uint32_t window; uint32_t seqno; uint32_t ackno; } __attribute__((packed)); +/* Stat counters; fold to no-ops without PROC_FLOW_STATS. */ +#ifdef PROC_FLOW_STATS +struct frcti_stat { + size_t rxm_rto; /* RTO-timer driven retransmits */ + size_t rxm_rcv; /* RXM packets received (all) */ + size_t rxm_dup_rcv; /* RXM dups (peer already had it) */ + size_t rxm_sack; /* SACK-mechanism retransmits */ + size_t rxm_rack; /* RACK-driven retransmits */ + size_t rxm_dupthresh; /* DupThresh-driven retransmits */ + size_t rxm_nack; /* NACK-pulled retransmits */ + size_t rxm_due_count; /* rxm_due entries (pre-bail) */ + size_t rxm_due_acked; /* bail: seqno < snd_lwe */ + size_t rxm_due_unowned; /* bail: slot.rxm replaced */ + size_t rxm_due_aged; /* bail: r->t0 + t_r < now */ + size_t rxm_due_defer; /* bail: non-HoL, deferred to HoL */ + size_t rxm_arm_fail; /* rxm_arm: malloc failed */ + size_t rxm_cancel; /* entries cancelled at teardown */ + size_t rxm_tx_dead; /* RXM tx into terminal flow */ + size_t tx_drop; /* frct_tx fail (any cause) */ + size_t tx_drop_ack; /* bare ACK dropped */ + size_t tx_drop_sack; /* SACK dropped */ + size_t tx_drop_ka; /* keepalive dropped */ + size_t tx_drop_rttp; /* RTT probe/echo dropped */ + size_t tx_drop_nack; /* pre-DRF NACK dropped */ + size_t tx_drop_rdv; /* rendez-vous dropped */ + size_t tx_drop_other; /* anything not matched above */ + size_t ack_snd; /* ACK packets sent (bare + SACK) */ + size_t ack_fire; /* delayed-ACK timer fires */ + size_t ack_supp_seqno; /* fire suppressed: seqno */ + size_t ack_supp_inact; /* fire suppressed: inact */ + size_t ack_supp_rate; /* fire suppressed: rate */ + size_t ack_rcv; /* ACK packets received */ + size_t ack_rtt; /* ACKs that fed RTT estimator */ + size_t ack_dup_rcv; /* ACK packet wire dups dropped */ + size_t dup_rcv; /* duplicates received */ + size_t out_rcv; /* pkts out of window */ + size_t rqo_rcv; /* pkts out of rqueue */ + size_t ooo_rcv; /* OOO arrivals */ + size_t sack_snd; /* SACK packets sent */ + size_t sack_rcv; /* SACK packets received */ + size_t dsack_snd; /* SACK pkts carrying a DSACK */ + size_t dsack_rcv; /* DSACK blocks parsed */ + size_t dsack_drop; /* DSACK blocks past MAX_DSACK_LAG */ + size_t nack_snd; /* pre-DRF NACKs sent */ + size_t nack_rcv; /* pre-DRF NACKs received */ + size_t tlp_snd; /* tail loss probes sent */ + size_t inact_drop; /* inactivity drop (NACK on cd) */ + size_t drf_rebase; /* DRF-triggered window rebase */ + size_t rq_released; /* slots cleared by release_rq */ + size_t rttp_snd; /* RTT probes sent */ + size_t rttp_rcv; /* RTT probe replies rcvd */ + size_t rtt_smpl; /* RTT estimator samples */ + size_t rdv_snd; /* rendez-vous packets sent */ + size_t rdv_rcv; /* rendez-vous packets rcvd */ + size_t ka_snd; /* keepalives sent */ + size_t ka_rcv; /* keepalives received */ + size_t sdu_snd_frag; /* writes that fragmented */ + size_t sdu_snd_alloc; /* alloc fail truncated SDU send */ + size_t sdu_snd_tx; /* tx fail truncated SDU send */ + size_t frag_snd; /* fragments sent: FIRST/MID/LAST */ + size_t frag_rcv; /* fragments stashed in rq[] */ + size_t sdu_reasm; /* SDUs delivered reassembled */ + size_t sdu_sole; /* SOLE SDUs delivered (n==1) */ + size_t frag_drop; /* dropped at malformed run */ + size_t strm_snd_byte; /* bytes sent on stream */ + size_t strm_rcv_byte; /* bytes copied to ring */ + size_t strm_dlv_byte; /* bytes delivered to reader */ + size_t strm_drop; /* stream rcvs dropped */ + size_t strm_fin_drop; /* stream FIN packets rejected */ + /* Profiling instrumentation. */ + size_t rcv_proc_ns; /* time inside FRCTI_RCV (ns) */ + size_t tw_move_ns; /* time inside tw_move (ns) */ + size_t drain_calls; /* flow_drain_rx_nb invocations */ +}; + +#define STAT_BUMP(frcti, field) FETCH_ADD_RELAXED(&(frcti)->stat.field, 1) +#define STAT_ADD(frcti, field, v) FETCH_ADD_RELAXED(&(frcti)->stat.field, (v)) +#define STAT_LOAD(frcti, field) LOAD_RELAXED(&(frcti)->stat.field) +#else +#define STAT_BUMP(frcti, field) ((void) (frcti)) +#define STAT_ADD(frcti, field, v) ((void) (frcti)) +#define STAT_LOAD(frcti, field) ((void) (frcti), (size_t) 0) +#endif + +#define frcti_to_flow(f) (&proc.flows[(f)->fd]) + +#define RTTP_RING 8 +#define RTTP_COLD_NS (100 * MILLION) /* cold-probe cadence */ +#define RQ_SLOT(seqno) ((seqno) & (RQ_SIZE - 1)) + +struct rxm_entry; + +enum snd_slot_flags { + SND_RTX = 0x01, /* Any retransmit; Karn skips next RTT sample. */ + SND_FAST_RXM = 0x02, /* Fast-retx one-shot gate per loss event. */ + SND_TLP = 0x04, /* Tail loss probe; ACK resets rto_mul. */ +}; + +struct snd_slot { + struct rxm_entry * rxm; /* RXM entry, NULL if none. */ + uint64_t time; /* ts_to_ns of last send (any kind). */ + uint8_t flags; /* SND_* bits above. */ +}; + +/* Per-seqno reorder slot (FRTX) and stream-mode byte/FIN metadata. */ +struct rcv_slot { + ssize_t idx; /* spb idx; -1 = empty */ + uint32_t start; /* stream byte start */ + uint32_t end; /* stream byte end */ + uint8_t fin; /* stream FIN bit */ +}; + +struct frct_cr { + uint32_t lwe; /* Left window edge */ + uint32_t rwe; /* Right window edge */ + + uint8_t cflags; + uint32_t seqno; /* SEQ to send, or last SEQ Ack'd */ + uint32_t ackno; /* snd: ACK-pkt seqno; rcv: dedup */ + + uint64_t act; /* ts_to_ns of last activity */ + uint64_t inact; /* Inactivity threshold (ns) */ +}; + +struct frcti { + /* IMM: set once in frcti_create; read-only thereafter. */ + int fd; + uint64_t t_mpl; /* MPL (ns) */ + uint64_t t_a; /* a-timer (ns) */ + uint64_t t_r; /* r-timer (ns) */ + uint64_t t_rdv; /* RDV cooldown (ns) */ + time_t ber; /* cached qs.ber */ + bool lossy; /* qs.loss != 0 */ + time_t qs_timeout; /* cached qs.timeout (ms) */ + size_t frag_mtu; /* max FRCT pkt: PCI + payload */ + uint16_t sack_n_max; /* SACK blocks that fit MTU */ + bool stream; + + /* All fields below are protected by lock (rwlock/LOAD_ACQUIRE). */ + struct { + struct frct_cr snd_cr; + struct frct_cr rcv_cr; + + /* RTT/RACK estimator */ + time_t srtt; /* smoothed RTT */ + time_t mdev; /* mean deviation */ + time_t min_rtt; /* RACK base, ns */ + uint64_t t_min_rtt; /* min_rtt last set */ + time_t rto; /* retransmit TO */ + time_t rto_min; /* RTO floor (ns) */ + uint8_t rto_mul; /* RTO backoff bits */ + uint32_t rtt_lwe; /* RTT-sample fence */ + uint64_t t_rcv_rtt; /* last RTT feed */ + uint64_t t_snd_probe; /* last probe sent */ + uint64_t t_latest_ack; /* RACK.fack snd-ts */ + uint32_t probe_id_next; + struct { + uint32_t id; + uint64_t ts; /* ts_to_ns send */ + uint8_t nonce[RTTP_NONCE_LEN]; /* echoed back */ + } probes[RTTP_RING]; + + /* rcv reassembly */ + size_t max_rcv_sdu; /* max reasm bytes */ + uint8_t * rcv_ring; /* lazy alloc */ + size_t rcv_ring_sz; /* power of 2 */ + uint32_t ring_seq_cap; /* ring/per_pkt */ + + uint32_t snd_byte_next; + bool snd_fin_sent; + uint32_t snd_fin_seqno; + uint32_t rcv_byte_next; + uint32_t rcv_byte_high; /* contiguous high */ + uint32_t rcv_byte_fin; /* set when FIN */ + bool rcv_fin_seen; + + struct rcv_slot rcv_slots[RQ_SIZE]; + struct snd_slot snd_slots[RQ_SIZE]; /* .rxm is ATOM */ + + /* rcv SACK dedup */ + uint64_t t_snd_sack; + uint32_t sack_lwe; /* rcv lwe at SACK */ + uint16_t sack_n; /* SACK block count */ + + /* RFC 2883 D-SACK: pending report (single-slot, latest). */ + uint32_t dsack_seqno; + bool dsack_valid; + + /* RFC 8985 §7.2 RACK reorder-window scaling. */ + uint8_t reo_wnd_mult; /* REO_WND_MULT_MAX */ + uint32_t dsack_lwe_snap; /* lwe @ last DSACK */ + uint64_t t_last_reo_widen; /* once-per-RTT */ + + uint32_t dup_thresh; /* RFC 8985 */ + uint32_t tlp_high_seq; /* §7.3: 0 = none */ + uint8_t tlp_count; /* §7.3 per-episode */ + uint64_t t_nack; + bool open; /* FC window state */ + bool in_recovery; + uint32_t recovery_high; /* seqno @ entry */ + uint32_t rack_fired_lwe; /* lwe @ last RACK */ + struct timespec t_wnd; /* window-closed ts */ + struct timespec t_last_rdv; /* last RDV sent */ + struct list_head rxm_list; /* live rxm entries */ + + pthread_rwlock_t lock; + }; + + /* Read/written via __atomic without holding lock. */ + uint64_t t_ka_rcv; /* ts_to_ns of last KA rx */ + uint8_t ack_pending; /* delayed-ACK dedup */ + uint8_t tlp_pending; /* TLP arm dedup (lazy) */ + + /* Timer entries; ownership belongs to the tw module. */ + struct tw_entry ack_tw; /* delayed-ACK timer */ + struct tw_entry ka_tw; /* keepalive timer */ + struct tw_entry tlp_tw; /* tail-loss probe timer */ + +#ifdef PROC_FLOW_STATS + /* STAT: lock-free relaxed atomic counters. */ + struct frcti_stat stat; +#endif +}; + #ifdef PROC_FLOW_STATS +__attribute__((cold)) static int frct_rib_read(const char * path, char * buf, size_t len) { + struct frcti * frcti; struct timespec now; + uint64_t now_ns; char * entry; - struct flow * flow; - struct frcti * frcti; int fd; - - (void) len; + int written; + /* Snapshot under the locks; format outside (pure userspace). */ + struct { + uint64_t t_mpl; + uint64_t t_a; + uint64_t t_r; + time_t srtt; + time_t mdev; + time_t rto; + time_t min_rtt; + struct frct_cr snd_cr; + struct frct_cr rcv_cr; + size_t rx_q_now; + size_t tx_q_now; + struct frcti_stat stat; + } s; entry = strstr(path, RIB_SEPARATOR); assert(entry); @@ -118,23 +437,50 @@ static int frct_rib_read(const char * path, fd = atoi(path); - flow = &proc.flows[fd]; - clock_gettime(PTHREAD_COND_CLOCK, &now); + now_ns = TS_TO_UINT64(now); + + if (fd < 0 || fd >= PROC_MAX_FLOWS) + return 0; pthread_rwlock_rdlock(&proc.lock); - frcti = flow->frcti; + frcti = proc.flows[fd].frcti; + if (frcti == NULL) { + pthread_rwlock_unlock(&proc.lock); + return 0; + } + + s.t_mpl = frcti->t_mpl; + s.t_a = frcti->t_a; + s.t_r = frcti->t_r; + + s.rx_q_now = proc.flows[fd].rx_rb != NULL + ? ssm_rbuff_queued(proc.flows[fd].rx_rb) : 0; + s.tx_q_now = proc.flows[fd].tx_rb != NULL + ? ssm_rbuff_queued(proc.flows[fd].tx_rb) : 0; pthread_rwlock_rdlock(&frcti->lock); - sprintf(buf, - "Maximum packet lifetime (ns): %20ld\n" - "Max time to Ack (ns): %20ld\n" - "Max time to Retransmit (ns): %20ld\n" + s.srtt = frcti->srtt; + s.mdev = frcti->mdev; + s.rto = frcti->rto; + s.min_rtt = frcti->min_rtt; + s.snd_cr = frcti->snd_cr; + s.rcv_cr = frcti->rcv_cr; + s.stat = frcti->stat; + + pthread_rwlock_unlock(&frcti->lock); + pthread_rwlock_unlock(&proc.lock); + + written = snprintf(buf, len, + "Maximum packet lifetime (ns): %20" PRIu64 "\n" + "Max time to Ack (ns): %20" PRIu64 "\n" + "Max time to Retransmit (ns): %20" PRIu64 "\n" "Smoothed rtt (ns): %20ld\n" "RTT standard deviation (ns): %20ld\n" "Retransmit timeout RTO (ns): %20ld\n" + "Minimum RTT (RACK base, ns): %20ld\n" "Sender left window edge: %20u\n" "Sender right window edge: %20u\n" "Sender inactive (ns): %20lld\n" @@ -143,44 +489,132 @@ static int frct_rib_read(const char * path, "Receiver right window edge: %20u\n" "Receiver inactive (ns): %20lld\n" "Receiver last ack: %20u\n" - "Number of pkt retransmissions: %20zu\n" - "Number of rtt probes: %20zu\n" - "Number of rtt estimates: %20zu\n" - "Number of duplicates received: %20zu\n" - "Number of delayed acks received: %20zu\n" - "Number of rendez-vous sent: %20zu\n" - "Number of packets out of window: %20zu\n" - "Number of packets out of rqueue: %20zu\n", - frcti->mpl, - frcti->a, - frcti->r, - frcti->srtt, - frcti->mdev, - frcti->rto, - frcti->snd_cr.lwe, - frcti->snd_cr.rwe, - ts_diff_ns(&now, &frcti->snd_cr.act), - frcti->snd_cr.seqno, - frcti->rcv_cr.lwe, - frcti->rcv_cr.rwe, - ts_diff_ns(&now, &frcti->rcv_cr.act), - frcti->rcv_cr.seqno, - frcti->n_rtx, - frcti->n_prb, - frcti->n_rtt, - frcti->n_dup, - frcti->n_dak, - frcti->n_rdv, - frcti->n_out, - frcti->n_rqo); - - pthread_rwlock_unlock(&flow->frcti->lock); + "RXM (RTO-driven) sent: %20zu\n" + "RXM packets received: %20zu\n" + " duplicates received: %20zu\n" + "RXM (SACK mechanism) sent: %20zu\n" + "RXM (RACK-driven) sent: %20zu\n" + "RXM (DupThresh-driven) sent: %20zu\n" + "RXM (NACK-driven) sent: %20zu\n" + "ACK packets sent: %20zu\n" + "Delayed-ACK timer fires: %20zu\n" + " suppressed (seqno): %20zu\n" + " suppressed (inact): %20zu\n" + " suppressed (rate): %20zu\n" + "ACK packets received: %20zu\n" + " fed RTT estimator: %20zu\n" + " wire dups dropped: %20zu\n" + "Duplicates received: %20zu\n" + "Out-of-window pkts received: %20zu\n" + "Out-of-rqueue pkts received: %20zu\n" + "OOO arrivals: %20zu\n" + "SACKs sent: %20zu\n" + "SACKs received: %20zu\n" + "D-SACKs sent: %20zu\n" + "D-SACKs received: %20zu\n" + "D-SACK out-of-range dropped: %20zu\n" + "Pre-DRF NACKs sent: %20zu\n" + "Pre-DRF NACKs received: %20zu\n" + "Tail loss probes sent: %20zu\n" + "Inactivity drops (silent): %20zu\n" + "DRF window rebases: %20zu\n" + "rq slots cleared by release_rq: %20zu\n" + "RTT probes sent: %20zu\n" + "RTT probe replies received: %20zu\n" + "RTT estimator samples: %20zu\n" + "Rendez-vous packets sent: %20zu\n" + "Rendez-vous packets received: %20zu\n" + "Keepalives sent: %20zu\n" + "Keepalives received: %20zu\n" + "SDU writes fragmented: %20zu\n" + " alloc fail mid-SDU: %20zu\n" + " tx fail mid-SDU: %20zu\n" + "Fragments sent: %20zu\n" + "Fragments received: %20zu\n" + "SDUs delivered reassembled: %20zu\n" + "SDUs delivered (SOLE): %20zu\n" + "Fragments dropped (malformed): %20zu\n" + "Stream bytes sent: %20zu\n" + "Stream bytes received: %20zu\n" + "Stream bytes delivered: %20zu\n" + "Stream packets dropped: %20zu\n" + "Stream FINs dropped: %20zu\n" + "FRCTI_RCV time (ns): %20zu\n" + "tw_move time (ns): %20zu\n" + "drain_rx_nb calls: %20zu\n" + "RX rbuff queued: %20zu\n" + "TX rbuff queued: %20zu\n" + "RXM-due entries: %20zu\n" + " bail (acked): %20zu\n" + " bail (unowned): %20zu\n" + " bail (aged): %20zu\n" + " bail (defer): %20zu\n" + "RXM-arm malloc failures: %20zu\n" + "RXM cancels (teardown): %20zu\n" + "RXM tx into dead flow: %20zu\n" + "Tx ring drops (any cause): %20zu\n" + " ack: %20zu\n" + " sack: %20zu\n" + " ka: %20zu\n" + " rttp: %20zu\n" + " nack: %20zu\n" + " rdv: %20zu\n" + " other: %20zu\n", + /* Check getattr size below when adding stats. */ + s.t_mpl, s.t_a, s.t_r, + s.srtt, s.mdev, s.rto, s.min_rtt, + s.snd_cr.lwe, s.snd_cr.rwe, + (long long)(now_ns - s.snd_cr.act), + s.snd_cr.seqno, + s.rcv_cr.lwe, s.rcv_cr.rwe, + (long long)(now_ns - s.rcv_cr.act), + s.rcv_cr.seqno, + s.stat.rxm_rto, s.stat.rxm_rcv, s.stat.rxm_dup_rcv, + s.stat.rxm_sack, s.stat.rxm_rack, s.stat.rxm_dupthresh, + s.stat.rxm_nack, + s.stat.ack_snd, s.stat.ack_fire, + s.stat.ack_supp_seqno, s.stat.ack_supp_inact, + s.stat.ack_supp_rate, + s.stat.ack_rcv, s.stat.ack_rtt, s.stat.ack_dup_rcv, + s.stat.dup_rcv, s.stat.out_rcv, s.stat.rqo_rcv, + s.stat.ooo_rcv, + s.stat.sack_snd, s.stat.sack_rcv, + s.stat.dsack_snd, s.stat.dsack_rcv, s.stat.dsack_drop, + s.stat.nack_snd, s.stat.nack_rcv, s.stat.tlp_snd, + s.stat.inact_drop, s.stat.drf_rebase, s.stat.rq_released, + s.stat.rttp_snd, s.stat.rttp_rcv, s.stat.rtt_smpl, + s.stat.rdv_snd, s.stat.rdv_rcv, + s.stat.ka_snd, s.stat.ka_rcv, + s.stat.sdu_snd_frag, s.stat.sdu_snd_alloc, s.stat.sdu_snd_tx, + s.stat.frag_snd, s.stat.frag_rcv, + s.stat.sdu_reasm, s.stat.sdu_sole, s.stat.frag_drop, + s.stat.strm_snd_byte, s.stat.strm_rcv_byte, + s.stat.strm_dlv_byte, + s.stat.strm_drop, s.stat.strm_fin_drop, + s.stat.rcv_proc_ns, s.stat.tw_move_ns, + s.stat.drain_calls, + s.rx_q_now, s.tx_q_now, + s.stat.rxm_due_count, + s.stat.rxm_due_acked, s.stat.rxm_due_unowned, + s.stat.rxm_due_aged, s.stat.rxm_due_defer, + s.stat.rxm_arm_fail, + s.stat.rxm_cancel, + s.stat.rxm_tx_dead, s.stat.tx_drop, + s.stat.tx_drop_ack, s.stat.tx_drop_sack, + s.stat.tx_drop_ka, s.stat.tx_drop_rttp, + s.stat.tx_drop_nack, s.stat.tx_drop_rdv, + s.stat.tx_drop_other); + + if (written < 0) + return 0; - pthread_rwlock_unlock(&proc.lock); + if ((size_t) written >= len) + return (int) (len - 1); - return strlen(buf); + return written; } +__attribute__((cold)) static int frct_rib_readdir(char *** buf) { *buf = malloc(sizeof(**buf)); @@ -199,13 +633,14 @@ static int frct_rib_readdir(char *** buf) return -ENOMEM; } +__attribute__((cold)) static int frct_rib_getattr(const char * path, struct rib_attr * attr) { (void) path; - (void) attr; - attr->size = 1189; + /* Must be >= the sprintf output in frct_rib_read. */ + attr->size = 8192; attr->mtime = 0; return 0; @@ -220,128 +655,1168 @@ static struct rib_ops r_ops = { #endif /* PROC_FLOW_STATS */ -static bool before(uint32_t seq1, - uint32_t seq2) +static __inline__ bool before(uint32_t s1, uint32_t s2) { - return (int32_t)(seq1 - seq2) < 0; + return (int32_t)(s1 - s2) < 0; } -static bool after(uint32_t seq1, - uint32_t seq2) +static __inline__ bool after(uint32_t s1, uint32_t s2) { - return (int32_t)(seq2 - seq1) < 0; + return (int32_t)(s2 - s1) < 0; } -static void __send_frct_pkt(int fd, - uint8_t flags, - uint32_t ackno, - uint32_t rwe) +static __inline__ bool within(uint32_t seq, uint32_t lo, uint32_t hi) { - struct ssm_pk_buff * spb; - struct frct_pci * pci; - ssize_t idx; - struct flow * f; + return after(seq, lo) && !after(seq, hi); +} - /* Raw calls needed to bypass frcti. */ -#ifdef RXM_BLOCKING - idx = ssm_pool_alloc_b(proc.pool, sizeof(*pci), NULL, &spb, NULL); -#else - idx = ssm_pool_alloc(proc.pool, sizeof(*pci), NULL, &spb); -#endif - if (idx < 0) +static __inline__ bool in_window(uint32_t seq, const struct frct_cr * cr) +{ + return !before(seq, cr->lwe) && before(seq, cr->rwe); +} + +/* DRF arrival that stays within the current receive epoch. */ +static __inline__ bool same_epoch_drf(uint32_t seq, + uint16_t flags, + const struct frct_cr * cr) +{ + if (cr->lwe == cr->rwe) + return false; + + return (flags & FRCT_RXM) || in_window(seq, cr); +} + +/* + * RACK reorder window R (RFC 8985 §6.2): + * R = MIN(reo_wnd_mult * RACK.min_RTT / 4, SRTT) + * reo_wnd_mult scales on D-SACK evidence of under-tolerance (§7.2). + * Fall back to srtt when no min_rtt sample exists yet; MIN_REORDER_NS + * floor guards collapse below the timer-tick resolution. + */ +static __inline__ uint64_t rack_reorder_window(struct frcti * frcti) +{ + uint64_t mult = frcti->reo_wnd_mult > 0 ? frcti->reo_wnd_mult : 1; + uint64_t base = frcti->min_rtt > 0 ? (uint64_t) frcti->min_rtt + : (uint64_t) frcti->srtt; + uint64_t R = mult * (base / 4); + + R = MAX(R, (uint64_t) MIN_REORDER_NS); + R = MIN(R, (uint64_t) frcti->srtt); + + return R; +} + +static __inline__ int frct_spb_reserve(size_t len, + struct ssm_pk_buff ** spb) +{ + ssize_t idx = ssm_pool_alloc_b(proc.pool, len, NULL, spb, NULL); + + return idx < 0 ? (int) idx : 0; +} + +static __inline__ void frct_spb_release(struct ssm_pk_buff * spb) +{ + ssm_pool_remove(proc.pool, ssm_pk_buff_get_off(spb)); +} + +static __inline__ void frct_spb_release_idx(size_t idx) +{ + ssm_pool_remove(proc.pool, idx); +} + +/* Fetch the spb stashed at the rq slot for seqno. */ +static __inline__ struct ssm_pk_buff * rq_frag(const struct frcti * frcti, + uint32_t seqno) +{ + return ssm_pool_get(proc.pool, frcti->rcv_slots[RQ_SLOT(seqno)].idx); +} + +static __inline__ size_t frcti_data_hdr_len(const struct frcti * frcti) +{ + return FRCT_PCILEN + (frcti->stream ? FRCT_PCI_STREAM_LEN : 0); +} + +static __inline__ size_t frcti_ctrl_hdr_len(const struct frcti * frcti) +{ + (void) frcti; + + return FRCT_PCILEN; +} + +/* + * HCS at offset 2 inside PCI. Covers flags (bytes 0..1) and + * window/seqno/ackno (bytes 4..15), plus SPCI for stream DATA. + */ +static void frct_hcs_set(struct frct_pci * pci, + bool stream) +{ + uint16_t hcs = 0; + size_t tail; + + tail = sizeof(*pci) - sizeof(pci->flags) - sizeof(pci->hcs); + if (stream) + tail += FRCT_PCI_STREAM_LEN; + + crc16_ccitt_false(&hcs, pci, sizeof(pci->flags)); + crc16_ccitt_false(&hcs, &pci->window, tail); + + pci->hcs = hton16(hcs); +} + +static int frct_hcs_check(const struct frct_pci * pci, + const struct frcti * frcti) +{ + uint16_t hcs = 0; + uint16_t flags; + size_t tail; + + /* Untrusted flag read; mismatch on HCS will drop on corrupt. */ + flags = ntoh16(pci->flags); + + tail = sizeof(*pci) - sizeof(pci->flags) - sizeof(pci->hcs); + if (frcti->stream && (flags & FRCT_DATA)) + tail += FRCT_PCI_STREAM_LEN; + + crc16_ccitt_false(&hcs, pci, sizeof(pci->flags)); + crc16_ccitt_false(&hcs, &pci->window, tail); + + return hcs != ntoh16(pci->hcs); +} + +/* Bump tx_drop plus the per-frame-type counter matching `flags`. */ +static void frct_tx_drop_bump(struct frcti * frcti, + uint16_t flags) +{ + STAT_BUMP(frcti, tx_drop); + + if (flags & FRCT_SACK) { + STAT_BUMP(frcti, tx_drop_sack); return; + } - pci = (struct frct_pci *) ssm_pk_buff_head(spb); - memset(pci, 0, sizeof(*pci)); + if (flags & FRCT_KA) { + STAT_BUMP(frcti, tx_drop_ka); + return; + } - *((uint32_t *) pci) = hton32(rwe); + if (flags & FRCT_RTTP) { + STAT_BUMP(frcti, tx_drop_rttp); + return; + } - pci->flags = flags; - pci->ackno = hton32(ackno); + if (flags & FRCT_NACK) { + STAT_BUMP(frcti, tx_drop_nack); + return; + } + + if (flags & FRCT_RDVS) { + STAT_BUMP(frcti, tx_drop_rdv); + return; + } + + if (flags & FRCT_ACK) { + STAT_BUMP(frcti, tx_drop_ack); + return; + } + + STAT_BUMP(frcti, tx_drop_other); +} - f = &proc.flows[fd]; +static int frct_tx(struct frcti * frcti, struct ssm_pk_buff * spb) +{ + struct flow * f = frcti_to_flow(frcti); + const struct frct_pci * pci; + const struct timespec * dl = NULL; + struct timespec now; + struct timespec intv = TIMESPEC_INIT_NS(FRCT_TX_TIMEO_NS); + struct timespec deadline; + uint16_t flags; + ssize_t idx; + int ret = -ENOMEM; + + pci = (const struct frct_pci *) ssm_pk_buff_head(spb); + flags = ntoh16(pci->flags); + + /* CRC32 covers plaintext body; PCI is in HCS. Pre-encrypt. */ + if (flags & FRCT_SACK) { + if (crc_add(spb, frcti_ctrl_hdr_len(frcti)) != 0) + goto fail; + } else if ((flags & FRCT_DATA) && f->info.qs.ber == 0) { + if (crc_add(spb, frcti_data_hdr_len(frcti)) != 0) + goto fail; + } if (spb_encrypt(f, spb) < 0) goto fail; -#ifdef RXM_BLOCKING - if (ssm_rbuff_write_b(f->tx_rb, idx, NULL)) -#else - if (ssm_rbuff_write(f->tx_rb, idx)) -#endif + idx = ssm_pk_buff_get_off(spb); + + /* DATA blocks; control times out so a full ring can't stall wheel. */ + if (!(flags & FRCT_DATA)) { + clock_gettime(PTHREAD_COND_CLOCK, &now); + ts_add(&now, &intv, &deadline); + dl = &deadline; + } + + ret = ssm_rbuff_write_b(f->tx_rb, idx, dl); + if (ret < 0) goto fail; ssm_flow_set_notify(f->set, f->info.id, FLOW_PKT); - return; + return 0; fail: - ipcp_spb_release(spb); - return; + frct_tx_drop_bump(frcti, flags); + ssm_pool_remove(proc.pool, ssm_pk_buff_get_off(spb)); + return ret; +} + +__attribute__((cold)) +static void frct_mark_flow_down(struct frcti * frcti) +{ + struct flow * f = frcti_to_flow(frcti); + + if (f->rx_rb != NULL) + ssm_rbuff_set_bits(f->rx_rb, RB_FLOWDOWN); + + if (f->tx_rb != NULL) + ssm_rbuff_set_bits(f->tx_rb, RB_FLOWDOWN); +} + +__attribute__((cold)) +static void frct_mark_peer_dead(struct frcti * frcti) +{ + struct flow * f = frcti_to_flow(frcti); + + if (f->rx_rb != NULL) + ssm_rbuff_set_bits(f->rx_rb, RB_FLOWPEER); + + if (proc.fqset != NULL) + ssm_flow_set_notify(proc.fqset, f->info.id, FLOW_PEER); +} + +static __inline__ int frct_ctrl_alloc(struct ssm_pk_buff ** spb, + struct frct_pci ** pci, + size_t payload_len) +{ + if (frct_spb_reserve(FRCT_PCILEN + payload_len, spb) < 0) + return -1; + + *pci = (struct frct_pci *) ssm_pk_buff_head(*spb); + memset(*pci, 0, FRCT_PCILEN); + + return 0; +} + +/* + * Advertised rwe. Stream mode clamps to lwe + ring_seq_cap so the + * byte-equivalent fits the rx ring. Caller holds at least the rdlock. + */ +static __inline__ uint32_t frcti_advert_rwe(struct frcti * frcti) +{ + uint32_t rwe; + uint32_t cap; + + rwe = frcti->rcv_cr.rwe; + + if (!frcti->stream) + return rwe; + + cap = frcti->rcv_cr.lwe + frcti->ring_seq_cap; + + return before(cap, rwe) ? cap : rwe; +} + +static void frcti_pkt_snd(struct frcti * frcti, + uint16_t flags, + uint32_t ackno, + uint32_t rwe) +{ + struct ssm_pk_buff * spb; + struct frct_pci * pci; + + if (frct_ctrl_alloc(&spb, &pci, 0) < 0) + return; + + pci->flags = hton16(flags); + pci->window = hton32(rwe); + pci->ackno = hton32(ackno); + if (flags & FRCT_ACK) { + /* reuse ackno for the sequence number of delayed ACK */ + ackno = FETCH_ADD_RELAXED(&frcti->snd_cr.ackno, 1); + pci->seqno = hton32(ackno + 1); + } + + frct_hcs_set(pci, false); + + frct_tx(frcti, spb); +} + +/* RTO floor scales with srtt; hard floor rto_min guards sub-ms RTT. */ +static void rtt_init(struct frcti * frcti, + time_t rtt_hint) +{ + time_t floor; + + if (rtt_hint > 0) { + rtt_hint = MAX(rtt_hint, (time_t) RTT_BOOT_NS); + frcti->srtt = rtt_hint; + frcti->mdev = rtt_hint >> 3; + floor = MAX(frcti->rto_min, 2 * frcti->srtt); + frcti->rto = MAX(floor, rtt_hint + (frcti->mdev << MDEV_MUL)); + frcti->min_rtt = rtt_hint; + } else { + /* Boot from first ACK. */ + frcti->srtt = 0; + frcti->mdev = RTT_BOOT_NS; + frcti->rto = MAX((time_t) INITIAL_RTO, frcti->rto_min); + frcti->min_rtt = 0; + } + + frcti->rto_mul = 0; +} + +/* RFC 8985 §6.2: replace min_RTT on unset, smaller sample, or expiry. */ +static __inline__ bool min_rtt_stale(struct frcti * frcti, + time_t mrtt, + uint64_t now_ns) +{ + if (frcti->min_rtt == 0) + return true; + + if (mrtt < frcti->min_rtt) + return true; + + return ts_aged_ns(now_ns, frcti->t_min_rtt, MIN_RTT_WIN_NS); +} + +/* Linux-style windowed-min refresh of RACK.min_RTT. */ +static __inline__ void min_rtt_update(struct frcti * frcti, + time_t mrtt, + uint64_t now_ns) +{ + if (!min_rtt_stale(frcti, mrtt, now_ns)) + return; + + frcti->min_rtt = mrtt; + frcti->t_min_rtt = now_ns; +} + +static void rtt_update(struct frcti * frcti, + time_t mrtt, + uint64_t now_ns) +{ + time_t srtt = frcti->srtt; + time_t rttvar = frcti->mdev; + time_t floor; + time_t rto; + + if (srtt == 0) { + srtt = mrtt; + rttvar = mrtt >> 1; + } else { + /* RFC 6298 symmetric EWMA. */ + time_t delta = mrtt - srtt; + srtt += (delta >> 3); + delta = (ABS(delta) - rttvar) >> 2; +#ifdef FRCT_LINUX_RTT_ESTIMATOR + if (delta < 0) + delta >>= 3; +#endif + rttvar += delta; + } + STAT_BUMP(frcti, rtt_smpl); + frcti->srtt = MAX(SRTT_FLOOR_NS, srtt); + frcti->mdev = MAX(MDEV_FLOOR_NS, rttvar); + + min_rtt_update(frcti, mrtt, now_ns); + + floor = MAX(frcti->rto_min, 2 * frcti->srtt); + rto = MAX(floor, frcti->srtt + (frcti->mdev << MDEV_MUL)); + + STORE_RELEASE(&frcti->rto, rto); + STORE_RELEASE(&frcti->rto_mul, 0); +} + +/* Fill probes[pos], return new probe_id; 0 on entropy failure. Wrlock. */ +static uint32_t rttp_alloc_probe(struct frcti * frcti, + uint64_t now_ns, + uint8_t nonce[RTTP_NONCE_LEN]) +{ + uint32_t probe_id; + size_t pos; + + if (random_buffer(nonce, RTTP_NONCE_LEN) < 0) + return 0; + + probe_id = frcti->probe_id_next++; + if (probe_id == 0) + probe_id = frcti->probe_id_next++; + + pos = RTTP_POS(probe_id); + frcti->probes[pos].id = probe_id; + frcti->probes[pos].ts = now_ns; + memcpy(frcti->probes[pos].nonce, nonce, RTTP_NONCE_LEN); + frcti->t_snd_probe = now_ns; + + STAT_BUMP(frcti, rttp_snd); + + return probe_id; +} + +/* Caller wrlock; out args valid on true (caller emits post-unlock). */ +static bool rtt_probe_arm(struct frcti * frcti, + uint64_t now_ns, + uint32_t * probe_id, + uint8_t nonce[RTTP_NONCE_LEN]) +{ + if (frcti->srtt == 0) + return false; + + if (!after(frcti->snd_cr.seqno, frcti->snd_cr.lwe)) + return false; + + if (!ts_aged_ns(now_ns, frcti->t_rcv_rtt, + 2u * (uint64_t) frcti->srtt)) + return false; + + if (!ts_aged_ns(now_ns, frcti->t_snd_probe, + (uint64_t) frcti->srtt)) + return false; + + *probe_id = rttp_alloc_probe(frcti, now_ns, nonce); + + return *probe_id != 0; +} + +static void frcti_rttp_snd(struct frcti * frcti, + uint32_t probe_id, + uint32_t echo_id, + const uint8_t * nonce) +{ + struct ssm_pk_buff * spb; + struct frct_pci * pci; + struct frct_rttp * rttp; + + if (frct_ctrl_alloc(&spb, &pci, RTTP_PAYLOAD) < 0) + return; + + pci->flags = hton16(FRCT_RTTP); + + frct_hcs_set(pci, false); + + rttp = (struct frct_rttp *) FRCT_BODY(pci); + rttp->probe_id = hton32(probe_id); + rttp->echo_id = hton32(echo_id); + memcpy(rttp->nonce, nonce, sizeof(rttp->nonce)); + + frct_tx(frcti, spb); +} + +struct rxm_entry { + struct tw_entry tw; + struct list_head next; /* in frcti->rxm_list */ + struct frcti * frcti; + uint32_t seqno; + uint64_t t0; + size_t len; + uint8_t pkt[]; /* flexible — sized at alloc time */ +}; + +static void rxm_entry_destroy(struct rxm_entry * r) +{ + free(r); +} + +static bool rxm_still_owned(struct frcti * frcti, + size_t pos, + struct rxm_entry * r) +{ + return LOAD_ACQUIRE(&frcti->snd_slots[pos].rxm) == r; +} + +/* + * All in-flight slots share the HoL backoff; otherwise non-HoL timers + * cycle at base RTO and storm the wire while HoL is still backing off. + */ +static uint64_t rxm_next_deadline(struct frcti * frcti, + uint64_t now_ns) +{ + time_t rto = LOAD_RELAXED(&frcti->rto); + uint8_t rto_mul = LOAD_RELAXED(&frcti->rto_mul); + + return now_ns + ((uint64_t) rto << rto_mul); +} + +/* Copy pkt, set FRCT_RXM, refresh ackno, re-seal HCS. */ +static struct ssm_pk_buff * rxm_pkt_prepare(const void * pkt, + size_t len, + uint32_t rcv_lwe, + bool stream) +{ + struct ssm_pk_buff * spb; + struct frct_pci * pci; + uint16_t flags; + + if (frct_spb_reserve(len, &spb) < 0) + return NULL; + + pci = (struct frct_pci *) ssm_pk_buff_head(spb); + memcpy(pci, pkt, len); + + flags = ntoh16(pci->flags) | FRCT_RXM; + pci->flags = hton16(flags); + pci->ackno = hton32(rcv_lwe); + + frct_hcs_set(pci, stream); + + return spb; +} + +/* Caller must NOT hold frcti->lock. */ +static void rxm_snd(struct frcti * frcti, + uint32_t seqno, + const void * pkt, + size_t len) +{ + struct ssm_pk_buff * spb; + struct timespec now; + struct snd_slot * slot; + uint32_t snd_lwe; + uint32_t rcv_lwe; + size_t pos; + int ret; + + snd_lwe = LOAD_RELAXED(&frcti->snd_cr.lwe); + rcv_lwe = LOAD_RELAXED(&frcti->rcv_cr.lwe); + + clock_gettime(PTHREAD_COND_CLOCK, &now); + + pthread_rwlock_wrlock(&frcti->lock); + + pos = RQ_SLOT(seqno); + slot = &frcti->snd_slots[pos]; + + slot->time = TS_TO_UINT64(now); + /* RTO supersedes any pending TLP/fast-rxm on this slot. */ + slot->flags = (slot->flags & ~(SND_FAST_RXM | SND_TLP)) | SND_RTX; + /* §7.3: RTO supersedes TLP probes and ends the probe episode. */ + frcti->tlp_high_seq = 0; + frcti->tlp_count = 0; + + frcti->rtt_lwe = seqno + 1; + + /* Only the HoL retransmit bumps the global RTO backoff. */ + if (seqno == snd_lwe && frcti->rto_mul < MAX_RTO_MUL) + STORE_RELEASE(&frcti->rto_mul, frcti->rto_mul + 1); + + /* RFC 8985 §7.2 step 4: RTO on HoL resets RACK reo scaling. */ + if (seqno == snd_lwe) + frcti->reo_wnd_mult = 1; + + pthread_rwlock_unlock(&frcti->lock); + + STAT_BUMP(frcti, rxm_rto); + + spb = rxm_pkt_prepare(pkt, len, rcv_lwe, frcti->stream); + if (spb == NULL) + return; + + /* ETIMEDOUT/ENOMEM: let r-timer drive teardown. */ + ret = frct_tx(frcti, spb); + if (ret == -EFLOWDOWN || ret == -ENOTALLOC) + STAT_BUMP(frcti, rxm_tx_dead); +} + +static void rxm_due(void * arg) +{ + struct rxm_entry * r = arg; + struct frcti * frcti = r->frcti; + struct timespec now; + uint64_t now_ns; + uint32_t snd_lwe; + size_t pos = RQ_SLOT(r->seqno); + + STAT_BUMP(frcti, rxm_due_count); + + snd_lwe = LOAD_RELAXED(&frcti->snd_cr.lwe); + + /* Already ACK'd: expected for the steady-state majority. */ + if (before(r->seqno, snd_lwe)) { + STAT_BUMP(frcti, rxm_due_acked); + goto cleanup; + } + + /* SACK/RACK-cleared the slot (caller NULL'd snd_slots[pos].rxm). */ + if (!rxm_still_owned(frcti, pos, r)) { + STAT_BUMP(frcti, rxm_due_unowned); + goto cleanup; + } + + clock_gettime(PTHREAD_COND_CLOCK, &now); + now_ns = TS_TO_UINT64(now); + + /* R-timer expired: peer unreachable. */ + if (RXM_AGED_OUT(r->t0, now_ns, frcti->t_r)) { + STAT_BUMP(frcti, rxm_due_aged); + frct_mark_flow_down(frcti); + goto cleanup; + } + + /* HoL-only retx; defer at base rto so HoL transitions react. */ + if (r->seqno != snd_lwe) { + STAT_BUMP(frcti, rxm_due_defer); + tw_post(&r->tw, now_ns + LOAD_RELAXED(&frcti->rto), + rxm_due, r); + return; + } + + rxm_snd(frcti, r->seqno, r->pkt, r->len); + + /* Re-check ownership: fire path may have replaced our entry. */ + if (rxm_still_owned(frcti, pos, r)) { + uint64_t anchor; + + /* Per-slot anchor breaks co-fire re-bin. */ + anchor = frcti->snd_slots[pos].time; + tw_post(&r->tw, rxm_next_deadline(frcti, anchor), rxm_due, r); + return; + } + + cleanup: + pthread_rwlock_wrlock(&frcti->lock); + + if (rxm_still_owned(frcti, pos, r)) + STORE_RELEASE(&frcti->snd_slots[pos].rxm, NULL); + + list_del(&r->next); + + pthread_rwlock_unlock(&frcti->lock); + + rxm_entry_destroy(r); +} + +/* Pre-allocate rxm entry so frcti_snd can fail before committing seqno. */ +static struct rxm_entry * rxm_alloc(struct frcti * frcti, + size_t pkt_len) +{ + struct rxm_entry * r; + + r = malloc(sizeof(*r) + pkt_len); + if (r == NULL) { + STAT_BUMP(frcti, rxm_arm_fail); + return NULL; + } + + r->frcti = frcti; + tw_init_entry(&r->tw); + + return r; +} + +static void rxm_arm(struct frcti * frcti, + uint32_t seqno, + struct rxm_entry * r, + const struct ssm_pk_buff * spb) +{ + struct timespec now; + time_t rto; + uint8_t rto_mul; + uint64_t deadline; + size_t len = ssm_pk_buff_len(spb); + + memcpy(r->pkt, ssm_pk_buff_head(spb), len); + r->len = len; + r->seqno = seqno; + + clock_gettime(PTHREAD_COND_CLOCK, &now); + r->t0 = TS_TO_UINT64(now); + + rto = LOAD_RELAXED(&frcti->rto); + rto_mul = LOAD_RELAXED(&frcti->rto_mul); + deadline = r->t0 + ((uint64_t) rto << rto_mul); + + pthread_rwlock_wrlock(&frcti->lock); + + assert(before(seqno, frcti->snd_cr.lwe + RQ_SIZE)); + + list_add_tail(&r->next, &frcti->rxm_list); + STORE_RELEASE(&frcti->snd_slots[RQ_SLOT(seqno)].rxm, r); + + pthread_rwlock_unlock(&frcti->lock); + + tw_post(&r->tw, deadline, rxm_due, r); +} + +static void rxm_cancel_all(struct frcti * frcti) +{ + struct list_head * p; + struct list_head * t; + + list_for_each_safe(p, t, &frcti->rxm_list) { + struct rxm_entry * r = list_entry(p, struct rxm_entry, next); + list_del(&r->next); + tw_cancel(&r->tw); + rxm_entry_destroy(r); + STAT_BUMP(frcti, rxm_cancel); + } +} + +static __inline__ void sack_block_put(uint8_t * payload, + uint16_t i, + uint32_t s, + uint32_t e) +{ + uint32_t * blk = (uint32_t *) + (payload + SACK_HDR_SIZE + i * SACK_BLOCK_SIZE); + + blk[0] = hton32(s); + blk[1] = hton32(e); +} + +static __inline__ void sack_block_get(const uint8_t * payload, + uint16_t i, + uint32_t * s, + uint32_t * e) +{ + const uint32_t * blk = (const uint32_t *) + (payload + SACK_HDR_SIZE + i * SACK_BLOCK_SIZE); + + *s = ntoh32(blk[0]); + *e = ntoh32(blk[1]); } -static void send_frct_pkt(struct frcti * frcti) +/* + * Build SACK blocks for ranges *above* rcv_cr.lwe. Wire invariant + * (see doc/frct.txt §1.3): every block produced here satisfies + * blocks[i].start > rcv_cr.lwe = ackno, which makes the "first block + * below ackno" convention used to mark a D-SACK (RFC 2883 §4 case 1) + * unambiguous. Caller holds frcti->lock. + */ +static uint16_t sack_blocks_build(struct frcti * frcti, + uint32_t blocks[][2], + uint16_t max_n) +{ + const struct rcv_slot * slots = frcti->rcv_slots; + uint32_t s; + uint32_t end; + uint16_t n = 0; + + s = frcti->rcv_cr.lwe + 1; + end = frcti->rcv_cr.lwe + RQ_SIZE; + if (after(end, frcti->rcv_cr.rwe)) + end = frcti->rcv_cr.rwe; + + while (before(s, end) && n < max_n) { + while (before(s, end) && slots[RQ_SLOT(s)].idx == -1) + ++s; + + if (!before(s, end)) + break; + + blocks[n][0] = s; + while (before(s, end) && slots[RQ_SLOT(s)].idx != -1) + ++s; + blocks[n][1] = s; + ++n; + } + + return n; +} + +/* + * Prepend the pending D-SACK report (if any) as block[0]; clear flag. + * Returns the number of slots consumed at the head (0 or 1). Caller + * holds wrlock. + */ +static __inline__ uint16_t dsack_consume(struct frcti * frcti, + uint32_t blocks[][2]) +{ + if (!frcti->dsack_valid || frcti->sack_n_max == 0) + return 0; + + blocks[0][0] = frcti->dsack_seqno; + blocks[0][1] = frcti->dsack_seqno + 1; + frcti->dsack_valid = false; + return 1; +} + +/* Caller must NOT hold frcti->lock. */ +static void frcti_sack_snd(struct frcti * frcti, + const struct sack_args * sa) +{ + struct ssm_pk_buff * spb; + struct frct_pci * pci; + buffer_t buf; + uint16_t i; + + assert(sa->n <= SACK_MAX_BLOCKS); + + buf.len = SACK_HDR_SIZE + sa->n * SACK_BLOCK_SIZE; + + if (frct_ctrl_alloc(&spb, &pci, buf.len) < 0) + return; + + pci->flags = hton16(FRCT_ACK | FRCT_FC | FRCT_SACK); + pci->window = hton32(sa->rwe); + pci->ackno = hton32(sa->ack); + pci->seqno = hton32(FETCH_ADD_RELAXED(&frcti->snd_cr.ackno, 1) + 1); + + frct_hcs_set(pci, false); + + buf.data = FRCT_BODY(pci); + memset(buf.data, 0, SACK_HDR_SIZE); + *(uint16_t *) buf.data = hton16(sa->n); + for (i = 0; i < sa->n; ++i) + sack_block_put(buf.data, i, sa->blocks[i][0], sa->blocks[i][1]); + + frct_tx(frcti, spb); +} + +static void ack_snd(struct frcti * frcti, + bool with_sack) { struct timespec now; + uint64_t now_ns; time_t diff; uint32_t ackno; uint32_t rwe; - int fd; + struct sack_args * sa = NULL; + size_t sa_sz; + bool sacking = false; assert(frcti); + STAT_BUMP(frcti, ack_fire); + clock_gettime(PTHREAD_COND_CLOCK, &now); + now_ns = TS_TO_UINT64(now); + + if (with_sack && frcti->sack_n_max > 0) { + sa_sz = sizeof(*sa) + frcti->sack_n_max * sizeof(sa->blocks[0]); + sa = malloc(sa_sz); + /* If alloc fails, fall through and send a bare cum-ACK. */ + } pthread_rwlock_wrlock(&frcti->lock); - if (!after(frcti->rcv_cr.lwe, frcti->rcv_cr.seqno)) { + /* D-SACK rides through cum-ACK freshness; signal is the duplicate. */ + if (!after(frcti->rcv_cr.lwe, frcti->rcv_cr.seqno) + && !frcti->dsack_valid) { pthread_rwlock_unlock(&frcti->lock); - return; + STAT_BUMP(frcti, ack_supp_seqno); + goto out; } - fd = frcti->fd; ackno = frcti->rcv_cr.lwe; - rwe = frcti->rcv_cr.rwe; + rwe = frcti_advert_rwe(frcti); - diff = ts_diff_ns(&now, &frcti->rcv_cr.act); - if (diff > frcti->a) { + if (ACK_AGED_OUT(frcti->rcv_cr.act, now_ns, frcti->t_a)) { pthread_rwlock_unlock(&frcti->lock); - return; + STAT_BUMP(frcti, ack_supp_inact); + goto out; } - diff = ts_diff_ns(&now, &frcti->snd_cr.act); - if (diff < TICTIME) { + diff = (time_t) ts_age_ns(now_ns, frcti->snd_cr.act); + if (diff < TICTIME && !frcti->dsack_valid) { pthread_rwlock_unlock(&frcti->lock); - return; + STAT_BUMP(frcti, ack_supp_rate); + goto out; } + /* RFC 2018: piggyback SACK on timer ACK; dedup unchanged board. */ + if (sa == NULL || (frcti->sack_n == 0 && !frcti->dsack_valid)) + goto no_sack; + + sa->dsack = false; + sa->n = dsack_consume(frcti, sa->blocks); + if (sa->n == 1) + sa->dsack = true; + + sa->n += sack_blocks_build(frcti, sa->blocks + sa->n, + frcti->sack_n_max - sa->n); + if (sa->n == 0) + goto no_sack; + + if (!sa->dsack && ackno == frcti->sack_lwe && sa->n == frcti->sack_n) + goto no_sack; + + sa->ack = ackno; + sa->rwe = rwe; + frcti->sack_lwe = ackno; + frcti->sack_n = sa->n; + frcti->t_snd_sack = now_ns; + sacking = true; + + no_sack: frcti->rcv_cr.seqno = frcti->rcv_cr.lwe; pthread_rwlock_unlock(&frcti->lock); - __send_frct_pkt(fd, FRCT_ACK | FRCT_FC, ackno, rwe); + STAT_BUMP(frcti, ack_snd); + + if (sacking) { + STAT_BUMP(frcti, sack_snd); + if (sa->dsack) + STAT_BUMP(frcti, dsack_snd); + frcti_sack_snd(frcti, sa); + } else { + frcti_pkt_snd(frcti, FRCT_ACK | FRCT_FC, ackno, rwe); + } + + out: + free(sa); } -static void __send_rdv(int fd) +/* Delayed-ACK timer: per-flow, dedup'd via atomic test-and-set. */ +static void ack_due(void * arg) { - __send_frct_pkt(fd, FRCT_RDVS, 0, 0); + struct frcti * frcti = arg; + + __atomic_clear(&frcti->ack_pending, __ATOMIC_RELAXED); + + ack_snd(frcti, true); } -static struct frcti * frcti_create(int fd, - time_t a, - time_t r, - time_t mpl) +static int ack_arm(struct frcti * frcti) { - struct frcti * frcti; - ssize_t idx; - struct timespec now; - pthread_condattr_t cattr; + struct timespec now; + uint64_t deadline; + + if (__atomic_test_and_set(&frcti->ack_pending, __ATOMIC_RELAXED)) + return 0; + + clock_gettime(PTHREAD_COND_CLOCK, &now); + deadline = TS_TO_UINT64(now) + ACK_DELAY_NS; + + tw_post(&frcti->ack_tw, deadline, ack_due, frcti); + + return 0; +} + +/* Forward decl breaks the keepalive cycle: ka_arm <-> ka_due. */ +static void ka_due(void * arg); + +static int ka_arm(struct frcti * frcti) +{ + struct timespec now; + uint64_t now_ns; + uint64_t timeo_ns; + uint64_t snd_ns; + uint64_t rcv_ns; + uint64_t deadline; + + timeo_ns = (uint64_t) frcti->qs_timeout * MILLION; /* IMM */ + snd_ns = LOAD_RELAXED(&frcti->snd_cr.act) + timeo_ns / 4; + rcv_ns = LOAD_RELAXED(&frcti->rcv_cr.act) + timeo_ns; + + clock_gettime(PTHREAD_COND_CLOCK, &now); + now_ns = TS_TO_UINT64(now); + deadline = MIN(snd_ns, rcv_ns); + if (deadline <= now_ns) + deadline = now_ns + timeo_ns / 4; + + tw_post(&frcti->ka_tw, deadline, ka_due, frcti); + + return 0; +} + +__attribute__((cold)) +static void ka_snd(struct frcti * frcti) +{ + struct ssm_pk_buff * spb; + struct frct_pci * pci; + struct timespec now; + uint64_t now_ns; + time_t timeo_ns; + uint64_t rcv_act; + uint64_t ka_rcv; + int64_t rcv_idle; + int64_t snd_idle; + uint32_t ackno; + + assert(frcti); + + clock_gettime(PTHREAD_COND_CLOCK, &now); + now_ns = TS_TO_UINT64(now); + + timeo_ns = (time_t)(frcti->qs_timeout) * MILLION; /* IMM */ + rcv_act = LOAD_RELAXED(&frcti->rcv_cr.act); + ka_rcv = LOAD_RELAXED(&frcti->t_ka_rcv); + rcv_idle = ts_age_ns(now_ns, rcv_act > ka_rcv ? rcv_act : ka_rcv); + snd_idle = ts_age_ns(now_ns, LOAD_RELAXED(&frcti->snd_cr.act)); + + if (rcv_idle > timeo_ns) { + frct_mark_peer_dead(frcti); + return; + } + + if (snd_idle <= timeo_ns / 4) { + ka_arm(frcti); + return; + } + + if (frct_ctrl_alloc(&spb, &pci, 0) < 0) { + ka_arm(frcti); + return; + } + + ackno = LOAD_RELAXED(&frcti->rcv_cr.lwe); + + pci->flags = hton16(FRCT_KA | FRCT_ACK); + pci->ackno = hton32(ackno); + + frct_hcs_set(pci, false); + + STAT_BUMP(frcti, ka_snd); + frct_tx(frcti, spb); + + ka_arm(frcti); +} + +/* Keepalive timer: re-posted by the fire callback itself. */ +static void ka_due(void * arg) +{ + ka_snd((struct frcti *) arg); +} + +static void frcti_rdv_snd(struct frcti * frcti) +{ + frcti_pkt_snd(frcti, FRCT_RDVS, 0, 0); +} + +#define HAS_RESCNTL(cr) ((cr)->cflags & FRCTFRESCNTL) +static bool frcti_is_window_open(struct frcti * frcti) +{ + struct frct_cr * snd_cr = &frcti->snd_cr; + struct timespec now; + time_t diff; + bool ret = false; + + if (!HAS_RESCNTL(snd_cr)) + return true; + + if (before(snd_cr->seqno, LOAD_RELAXED(&snd_cr->rwe))) + return true; + + /* Window may be closed; wrlock for RDV state mutations. */ + pthread_rwlock_wrlock(&frcti->lock); + + if (before(snd_cr->seqno, snd_cr->rwe)) { + ret = true; + goto unlock; + } + + clock_gettime(PTHREAD_COND_CLOCK, &now); + + if (frcti->open) { + frcti->open = false; + frcti->t_wnd = now; + frcti->t_last_rdv = now; + goto unlock; + } + + diff = ts_diff_ns(&now, &frcti->t_wnd); + if (diff > MAX_RDV) + goto unlock; + + diff = ts_diff_ns(&now, &frcti->t_last_rdv); + if (diff > (time_t) frcti->t_rdv) { + frcti->t_last_rdv = now; + frcti_rdv_snd(frcti); + STAT_BUMP(frcti, rdv_snd); + } + unlock: + pthread_rwlock_unlock(&frcti->lock); + + return ret; +} + +/* n contiguous seqnos free? No RDV: the n=1 path drives it. */ +static bool frcti_is_window_open_n(struct frcti * frcti, + size_t n) +{ + struct frct_cr * snd_cr = &frcti->snd_cr; + + if (!HAS_RESCNTL(snd_cr)) + return true; + + if (n <= 1) + return frcti_is_window_open(frcti); + + return before(snd_cr->seqno + (uint32_t)(n - 1), + LOAD_RELAXED(&snd_cr->rwe)); +} + +static void release_rq(struct frcti * frcti) +{ + size_t i; + + for (i = 0; i < RQ_SIZE; ++i) { + if (frcti->rcv_slots[i].idx == -1) + continue; + + /* Stream rq entries are sentinels (no spb owned). */ + if (!frcti->stream) + frct_spb_release_idx(frcti->rcv_slots[i].idx); + + frcti->rcv_slots[i].idx = -1; + STAT_BUMP(frcti, rq_released); + } +} + +static __inline__ bool stream_ring_sz_ok(struct frcti * frcti, + size_t n) +{ + size_t per_pkt; + + if (n > FRCT_STREAM_RING_SZ_MAX) + return false; + + if ((n & (n - 1)) != 0) + return false; + + per_pkt = frcti->frag_mtu - frcti_data_hdr_len(frcti); + + return n >= FRCT_STREAM_RING_MIN_PKTS * per_pkt; +} + +/* Default ring sized for full RQ_SIZE seqno window; pow2, capped. */ +static size_t default_stream_ring_sz(size_t per_pkt) +{ + size_t need; + size_t sz; + + need = (size_t) RQ_SIZE * per_pkt; + sz = FRCT_STREAM_RING_SZ; + + while (sz < need && sz < FRCT_STREAM_RING_SZ_MAX) + sz <<= 1; + + return sz; +} + +struct frcti * frcti_create(int fd, + uint64_t a, + uint64_t r, + uint64_t mpl, + time_t rtt_hint, + qosspec_t qs, + uint32_t mtu) +{ + struct frcti * frcti; + ssize_t idx; + struct timespec now; + uint64_t now_ns; + size_t bb; + size_t per_pkt; #ifdef PROC_FLOW_STATS - char frctstr[FRCT_NAME_STRLEN + 1]; + char frctstr[FRCT_NAME_STRLEN + 1]; #endif - mpl *= MILLION; - a *= BILLION; - r *= BILLION; + mpl *= MILLION; /* ms -> ns */ + a *= MILLION; /* ms -> ns */ + r *= MILLION; /* ms -> ns */ frcti = malloc(sizeof(*frcti)); if (frcti == NULL) @@ -349,56 +1824,76 @@ static struct frcti * frcti_create(int fd, memset(frcti, 0, sizeof(*frcti)); + list_head_init(&frcti->rxm_list); + if (pthread_rwlock_init(&frcti->lock, NULL)) goto fail_lock; - if (pthread_mutex_init(&frcti->mtx, NULL)) - goto fail_mutex; - - if (pthread_condattr_init(&cattr)) - goto fail_cattr; -#ifndef __APPLE__ - pthread_condattr_setclock(&cattr, PTHREAD_COND_CLOCK); -#endif - if (pthread_cond_init(&frcti->cond, &cattr)) - goto fail_cond; - #ifdef PROC_FLOW_STATS sprintf(frctstr, "%d", fd); if (rib_reg(frctstr, &r_ops)) goto fail_rib_reg; #endif - pthread_condattr_destroy(&cattr); for (idx = 0; idx < RQ_SIZE; ++idx) - frcti->rq[idx] = -1; + frcti->rcv_slots[idx].idx = -1; clock_gettime(PTHREAD_COND_CLOCK, &now); + now_ns = TS_TO_UINT64(now); + + frcti->t_mpl = mpl; + frcti->t_a = a; + frcti->t_r = r; + frcti->t_rdv = DELT_RDV; + frcti->fd = fd; + frcti->ber = (time_t) qs.ber; + frcti->lossy = (qs.loss != 0); + frcti->qs_timeout = (time_t) qs.timeout; + + frcti->frag_mtu = (size_t) mtu; + + /* Cap blocks per SACK at what fits in the per-flow frag_mtu. */ + bb = (frcti->frag_mtu - FRCT_PCILEN - SACK_HDR_SIZE) + / SACK_BLOCK_SIZE; + if (bb > SACK_MAX_BLOCKS) + bb = SACK_MAX_BLOCKS; + frcti->sack_n_max = (uint16_t) bb; + + frcti->max_rcv_sdu = FRCT_MAX_SDU; + + frcti->stream = (qs.service == SVC_STREAM); + if (frcti->stream) { + per_pkt = frcti->frag_mtu - frcti_data_hdr_len(frcti); + frcti->rcv_ring_sz = default_stream_ring_sz(per_pkt); + frcti->ring_seq_cap = + (uint32_t) (frcti->rcv_ring_sz / per_pkt); + } - frcti->mpl = mpl; - frcti->a = a; - frcti->r = r; - frcti->rdv = DELT_RDV; - frcti->fd = fd; - - - frcti->rttseq = 0; - frcti->probe = false; - - frcti->srtt = 0; /* Updated on first ACK */ - frcti->mdev = 10 * MILLION; /* Updated on first ACK */ - frcti->rto = BILLION; /* Initial rxm will be after 1 s */ -#ifdef PROC_FLOW_STATS - frcti->n_rtx = 0; - frcti->n_prb = 0; - frcti->n_rtt = 0; - frcti->n_dup = 0; - frcti->n_dak = 0; - frcti->n_rdv = 0; - frcti->n_out = 0; - frcti->n_rqo = 0; -#endif - if (proc.flows[fd].info.qs.loss == 0) { + frcti->rto_min = (time_t) MAX(RTO_MIN, 1ULL << RXMQ_RES); + rtt_init(frcti, rtt_hint); + frcti->t_min_rtt = now_ns; + frcti->probe_id_next = 1; + frcti->t_rcv_rtt = now_ns; + frcti->t_snd_probe = now_ns; + frcti->t_snd_sack = 0; + frcti->sack_lwe = 0; + frcti->sack_n = 0; + frcti->dsack_seqno = 0; + frcti->dsack_valid = false; + frcti->reo_wnd_mult = 1; + frcti->dsack_lwe_snap = 0; + frcti->t_last_reo_widen = 0; + /* So the first pre-DRF NACK fires without waiting cooldown. */ + frcti->t_nack = now_ns - BILLION; + frcti->in_recovery = false; + frcti->recovery_high = 0; + frcti->rack_fired_lwe = 0; + + tw_init_entry(&frcti->ack_tw); + tw_init_entry(&frcti->ka_tw); + tw_init_entry(&frcti->tlp_tw); + + if (!frcti->lossy) { frcti->snd_cr.cflags |= FRCTFRTX | FRCTFLINGER; frcti->rcv_cr.cflags |= FRCTFRTX; } @@ -406,24 +1901,31 @@ static struct frcti * frcti_create(int fd, frcti->snd_cr.cflags |= FRCTFRESCNTL; frcti->snd_cr.rwe = START_WINDOW; + if (frcti->lossy) + frcti->snd_cr.rwe = RQ_SIZE; + + frcti->snd_cr.inact = 3 * mpl + a + r + BILLION; /* ns */ + frcti->snd_cr.act = now_ns - frcti->snd_cr.inact - BILLION; - frcti->snd_cr.inact = (3 * mpl + a + r) / BILLION + 1; /* s */ - frcti->snd_cr.act.tv_sec = now.tv_sec - (frcti->snd_cr.inact + 1); + frcti->rcv_cr.inact = 2 * mpl + a + r + BILLION; /* ns */ + frcti->rcv_cr.act = now_ns - frcti->rcv_cr.inact - BILLION; - frcti->rcv_cr.inact = (2 * mpl + a + r) / BILLION + 1; /* s */ - frcti->rcv_cr.act.tv_sec = now.tv_sec - (frcti->rcv_cr.inact + 1); + frcti->t_ka_rcv = now_ns; + + /* qs_timeout == 0: no KA, silent peer crash goes undetected. */ + if (frcti->qs_timeout > 0) { + if (ka_arm(frcti) < 0) + goto fail_ka_arm; + } return frcti; + fail_ka_arm: #ifdef PROC_FLOW_STATS + sprintf(frctstr, "%d", fd); + rib_unreg(frctstr); fail_rib_reg: - pthread_cond_destroy(&frcti->cond); #endif - fail_cond: - pthread_condattr_destroy(&cattr); - fail_cattr: - pthread_mutex_destroy(&frcti->mtx); - fail_mutex: pthread_rwlock_destroy(&frcti->lock); fail_lock: free(frcti); @@ -431,21 +1933,55 @@ static struct frcti * frcti_create(int fd, return NULL; } -static void frcti_destroy(struct frcti * frcti) +void frcti_destroy(struct frcti * frcti) { #ifdef PROC_FLOW_STATS char frctstr[FRCT_NAME_STRLEN + 1]; +#endif + /* Drop every wheel entry referencing frcti before freeing it. */ + rxm_cancel_all(frcti); + tw_cancel(&frcti->ack_tw); + tw_cancel(&frcti->ka_tw); + tw_cancel(&frcti->tlp_tw); + +#if defined(PROC_FLOW_STATS) && defined(FRCT_DEBUG_STDOUT) + printf("[FRCT teardown] pid=%d fd=%d " + "sdu_snd=%zu sdu_reasm=%zu sdu_sole=%zu " + "frag_snd=%zu frag_rcv=%zu frag_drop=%zu " + "rxm_rto=%zu rxm_sack=%zu rxm_dup=%zu " + "rxm_due=%zu acked=%zu unowned=%zu aged=%zu defer=%zu " + "cancel=%zu arm_fail=%zu inflight=%u " + "nack_snd=%zu nack_rcv=%zu inact_drop=%zu " + "drf_rebase=%zu rq_released=%zu\n", + (int) getpid(), frcti->fd, + frcti->stat.sdu_snd_frag, frcti->stat.sdu_reasm, + frcti->stat.sdu_sole, + frcti->stat.frag_snd, frcti->stat.frag_rcv, + frcti->stat.frag_drop, + frcti->stat.rxm_rto, frcti->stat.rxm_sack, + frcti->stat.rxm_dupthresh, + frcti->stat.rxm_due_count, frcti->stat.rxm_due_acked, + frcti->stat.rxm_due_unowned, frcti->stat.rxm_due_aged, + frcti->stat.rxm_due_defer, + frcti->stat.rxm_cancel, frcti->stat.rxm_arm_fail, + frcti->snd_cr.seqno - frcti->snd_cr.lwe, + frcti->stat.nack_snd, frcti->stat.nack_rcv, + frcti->stat.inact_drop, + frcti->stat.drf_rebase, frcti->stat.rq_released); +#endif + + release_rq(frcti); + free(frcti->rcv_ring); +#ifdef PROC_FLOW_STATS sprintf(frctstr, "%d", frcti->fd); rib_unreg(frctstr); #endif - pthread_cond_destroy(&frcti->cond); - pthread_mutex_destroy(&frcti->mtx); pthread_rwlock_destroy(&frcti->lock); free(frcti); } -static uint16_t frcti_getflags(struct frcti * frcti) +uint16_t frcti_getflags(struct frcti * frcti) { uint16_t ret; @@ -453,89 +1989,91 @@ static uint16_t frcti_getflags(struct frcti * frcti) pthread_rwlock_rdlock(&frcti->lock); - ret = frcti->snd_cr.cflags; + ret = frcti->snd_cr.cflags & FRCTFMASK; pthread_rwlock_unlock(&frcti->lock); return ret; } -static void frcti_setflags(struct frcti * frcti, - uint16_t flags) +void frcti_setflags(struct frcti * frcti, + uint16_t flags) { - flags |= FRCTFRTX; /* Should not be set by command */ - assert(frcti); - pthread_rwlock_wrlock(&frcti->lock); + flags &= FRCTFSETMASK; - frcti->snd_cr.cflags &= FRCTFRTX; /* Zero other flags */ + pthread_rwlock_wrlock(&frcti->lock); - frcti->snd_cr.cflags &= flags; + frcti->snd_cr.cflags = (frcti->snd_cr.cflags & ~FRCTFSETMASK) | flags; pthread_rwlock_unlock(&frcti->lock); } -#define frcti_queued_pdu(frcti) \ - (frcti == NULL ? idx : __frcti_queued_pdu(frcti)) +size_t frcti_get_max_rcv_sdu(struct frcti * frcti) +{ + size_t ret; -#define frcti_snd(frcti, spb) \ - (frcti == NULL ? 0 : __frcti_snd(frcti, spb)) + assert(frcti); + + pthread_rwlock_rdlock(&frcti->lock); + ret = frcti->max_rcv_sdu; + pthread_rwlock_unlock(&frcti->lock); -#define frcti_rcv(frcti, spb) \ - (frcti == NULL ? 0 : __frcti_rcv(frcti, spb)) + return ret; +} -#define frcti_dealloc(frcti) \ - (frcti == NULL ? 0 : __frcti_dealloc(frcti)) +int frcti_set_max_rcv_sdu(struct frcti * frcti, + size_t max) +{ + assert(frcti); -#define frcti_is_window_open(frcti) \ - (frcti == NULL ? true : __frcti_is_window_open(frcti)) + if (max == 0) + return -EINVAL; -#define frcti_window_wait(frcti, abstime) \ - (frcti == NULL ? 0 : __frcti_window_wait(frcti, abstime)) + pthread_rwlock_wrlock(&frcti->lock); + frcti->max_rcv_sdu = max; + pthread_rwlock_unlock(&frcti->lock); + return 0; +} -static bool __frcti_is_window_open(struct frcti * frcti) +size_t frcti_get_rcv_ring_sz(struct frcti * frcti) { - struct frct_cr * snd_cr = &frcti->snd_cr; - bool ret = true; + size_t ret; + + assert(frcti); pthread_rwlock_rdlock(&frcti->lock); + ret = frcti->rcv_ring_sz; + pthread_rwlock_unlock(&frcti->lock); - if (snd_cr->cflags & FRCTFRESCNTL) - ret = before(snd_cr->seqno, snd_cr->rwe); + return ret; +} - if (!ret) { - struct timespec now; +/* Set before any stream byte has been delivered; -EBUSY otherwise. */ +int frcti_set_rcv_ring_sz(struct frcti * frcti, + size_t n) +{ + int ret = 0; + size_t per_pkt; - clock_gettime(PTHREAD_COND_CLOCK, &now); + assert(frcti); - pthread_mutex_lock(&frcti->mtx); - if (frcti->open) { - frcti->open = false; - frcti->t_wnd = now; - frcti->t_rdvs = now; - } else { - time_t diff; - diff = ts_diff_ns(&now, &frcti->t_wnd); - if (diff > MAX_RDV) { - pthread_mutex_unlock(&frcti->mtx); - pthread_rwlock_unlock(&frcti->lock); - return false; - } - - diff = ts_diff_ns(&now, &frcti->t_rdvs); - if (diff > frcti->rdv) { - frcti->t_rdvs = now; - __send_rdv(frcti->fd); -#ifdef PROC_FLOW_STATS - frcti->n_rdv++; -#endif + if (!frcti->stream) + return -ENOTSUP; + if (!stream_ring_sz_ok(frcti, n)) + return -EINVAL; - } - } + per_pkt = frcti->frag_mtu - frcti_data_hdr_len(frcti); + + pthread_rwlock_wrlock(&frcti->lock); - pthread_mutex_unlock(&frcti->mtx); + if (frcti->rcv_ring != NULL) { + ret = -EBUSY; + } else { + frcti->rcv_ring_sz = n; + frcti->ring_seq_cap = (uint32_t) (n / per_pkt); } pthread_rwlock_unlock(&frcti->lock); @@ -543,392 +2081,2130 @@ static bool __frcti_is_window_open(struct frcti * frcti) return ret; } -static int __frcti_window_wait(struct frcti * frcti, - struct timespec * abstime) +time_t frcti_get_rto_min(struct frcti * frcti) { - struct frct_cr * snd_cr = &frcti->snd_cr; - int ret = 0; + time_t v; + + assert(frcti); pthread_rwlock_rdlock(&frcti->lock); + v = frcti->rto_min; + pthread_rwlock_unlock(&frcti->lock); - if (!(snd_cr->cflags & FRCTFRESCNTL)) { - pthread_rwlock_unlock(&frcti->lock); + return v; +} + +/* Floor at the timer-wheel resolution; finer granularity is unrepresentable. */ +int frcti_set_rto_min(struct frcti * frcti, + time_t rto_min) +{ + time_t floor = (time_t) (1ULL << RXMQ_RES); + time_t rto_floor; + time_t rto; + + assert(frcti); + + if (rto_min < floor) + return -EINVAL; + + pthread_rwlock_wrlock(&frcti->lock); + + frcti->rto_min = rto_min; + if (frcti->srtt > 0) { + rto_floor = MAX(rto_min, 2 * frcti->srtt); + rto = MAX(rto_floor, + frcti->srtt + (frcti->mdev << MDEV_MUL)); + STORE_RELEASE(&frcti->rto, rto); + } else if (frcti->rto < rto_min) { + STORE_RELEASE(&frcti->rto, rto_min); + } + + pthread_rwlock_unlock(&frcti->lock); + + return 0; +} + +/* Re-arm a fresh rxm so a lost fast-retx still recovers via RTO. */ +static void sack_rxm_snd(struct frcti * frcti, + void * pkt, + size_t len) +{ + struct ssm_pk_buff * spb; + const struct frct_pci * pci; + struct rxm_entry * rxm; + uint32_t rcv_lwe; + uint32_t seqno; + int ret; + + rcv_lwe = LOAD_RELAXED(&frcti->rcv_cr.lwe); + + spb = rxm_pkt_prepare(pkt, len, rcv_lwe, frcti->stream); + if (spb == NULL) + return; + + pci = (const struct frct_pci *) ssm_pk_buff_head(spb); + seqno = ntoh32(pci->seqno); + + rxm = rxm_alloc(frcti, ssm_pk_buff_len(spb)); + if (rxm == NULL) { + frct_spb_release(spb); + return; + } + rxm_arm(frcti, seqno, rxm, spb); + + STAT_BUMP(frcti, rxm_sack); + ret = frct_tx(frcti, spb); + if (ret == -EFLOWDOWN || ret == -ENOTALLOC) + STAT_BUMP(frcti, rxm_tx_dead); +} + +/* Additive HoL emit; original snd_slots[hp].rxm stays armed (NewReno). */ +static int fast_rxm_send(struct frcti * frcti, + void * pkt, + size_t len) +{ + struct ssm_pk_buff * spb; + uint32_t rcv_lwe; + + rcv_lwe = LOAD_RELAXED(&frcti->rcv_cr.lwe); + + spb = rxm_pkt_prepare(pkt, len, rcv_lwe, frcti->stream); + if (spb == NULL) return 0; + + return frct_tx(frcti, spb); +} + +/* PCI bytes survive head_release at receive; just rewind the pointer. */ +static __inline__ uint16_t frag_role_peek(struct ssm_pk_buff * spb) +{ + const struct frct_pci * pci; + + assert(ssm_pk_buff_head(spb) != NULL); + + pci = (const struct frct_pci *) (ssm_pk_buff_head(spb) - FRCT_PCILEN); + + return ntoh16(pci->flags) & FRCT_FR_MASK; +} + +enum frag_state { + FRAG_NOT_READY, /* head missing / FIRST..LAST run incomplete */ + FRAG_DELIVER, /* *count fragments form a deliverable SDU */ + FRAG_DROP, /* *count fragments at lwe are malformed */ +}; + +/* + * On a gap in the run: FRTX waits (NOT_READY); best-effort scans forward + * for the next FIRST/SOLE and returns DROP for the broken prefix. *count + * gets the offset from the trailing edge. NOT_READY if no later run is + * in window. Caller rdlock. + */ +static enum frag_state frag_inspect_gap(struct frcti * frcti, + size_t start, + size_t * count) +{ + const struct rcv_slot * slots = frcti->rcv_slots; + struct ssm_pk_buff * spb; + uint32_t k; + uint16_t role; + size_t m; + + if (frcti->rcv_cr.cflags & FRCTFRTX) + return FRAG_NOT_READY; + + k = frcti->rcv_cr.rwe - RQ_SIZE; + + for (m = start; m < RQ_SIZE; ++m) { + if (slots[RQ_SLOT(k + m)].idx == -1) + continue; + + spb = rq_frag(frcti, k + m); + role = frag_role_peek(spb); + + if (role == FRCT_FR_SOLE || role == FRCT_FR_FIRST) { + if (m == 0) + return FRAG_NOT_READY; + + *count = m; + return FRAG_DROP; + } } - while (snd_cr->seqno == snd_cr->rwe && ret != -ETIMEDOUT) { - struct timespec now; - pthread_rwlock_unlock(&frcti->lock); - pthread_mutex_lock(&frcti->mtx); + return FRAG_NOT_READY; +} + +/* + * Inspect rq[lwe..]; set *count and return DELIVER/DROP/NOT_READY. DROP + * covers broken prefixes (mid/last at HoL, FIRST..[non-LAST]..new-FIRST). + * Non-FRTX flows skip past gaps to the next FIRST/SOLE. Caller rdlock. + */ +static enum frag_state frag_run_inspect(struct frcti * frcti, + size_t * count) +{ + const struct rcv_slot * slots = frcti->rcv_slots; + struct ssm_pk_buff * spb; + uint32_t k = frcti->rcv_cr.rwe - RQ_SIZE; + uint16_t role; + size_t n = 0; + + if (slots[RQ_SLOT(k)].idx == -1) + return frag_inspect_gap(frcti, 0, count); + + spb = rq_frag(frcti, k); + role = frag_role_peek(spb); + + if (role == FRCT_FR_SOLE) { + *count = 1; + return FRAG_DELIVER; + } - if (frcti->open) { - clock_gettime(PTHREAD_COND_CLOCK, &now); + if (role != FRCT_FR_FIRST) { + *count = 1; + return FRAG_DROP; + } + + while (true) { + if (n == RQ_SIZE || slots[RQ_SLOT(k + n)].idx == -1) + return frag_inspect_gap(frcti, n, count); + + spb = rq_frag(frcti, k + n); + role = frag_role_peek(spb); + ++n; + + if (role == FRCT_FR_LAST) { + *count = n; + return FRAG_DELIVER; + } - frcti->t_wnd = now; - frcti->t_rdvs = now; - frcti->open = false; + if (n > 1 && role != FRCT_FR_MID) { + /* SOLE or new FIRST mid-run: drop the prefix. */ + *count = n - 1; + return FRAG_DROP; } + } +} + +/* Caller wrlock. Delivery edge is implicit: rwe - RQ_SIZE. */ +static void frag_drop(struct frcti * frcti, + size_t count) +{ + uint32_t k = frcti->rcv_cr.rwe - RQ_SIZE; + uint32_t edge; + size_t i; + + for (i = 0; i < count; ++i) { + size_t pos = RQ_SLOT(k + i); + + if (frcti->rcv_slots[pos].idx == -1) + continue; + + frct_spb_release_idx(frcti->rcv_slots[pos].idx); + frcti->rcv_slots[pos].idx = -1; + } + + frcti->rcv_cr.rwe += count; + + /* Drop may span a gap; pull lwe up to preserve rwe - RQ_SIZE <= lwe. */ + edge = frcti->rcv_cr.rwe - RQ_SIZE; + if (before(frcti->rcv_cr.lwe, edge)) + STORE_RELEASE(&frcti->rcv_cr.lwe, edge); +} + +/* Copy `count` fragments at rq[lwe..] into buf; release + advance lwe. */ +static size_t frag_gather(struct frcti * frcti, + size_t count, + uint8_t * buf) +{ + struct ssm_pk_buff * frag; + size_t off = 0; + size_t i; + uint32_t k = frcti->rcv_cr.rwe - RQ_SIZE; + + for (i = 0; i < count; ++i) { + size_t pos = RQ_SLOT(k + i); + size_t flen; + + frag = rq_frag(frcti, k + i); + flen = ssm_pk_buff_len(frag); + memcpy(buf + off, ssm_pk_buff_head(frag), flen); + off += flen; + frct_spb_release_idx(frcti->rcv_slots[pos].idx); + frcti->rcv_slots[pos].idx = -1; + } + + frcti->rcv_cr.rwe += count; + + return off; +} + +/* Caller holds lock. */ +static size_t frag_total_len(struct frcti * frcti, + size_t count, + bool * overflow) +{ + struct ssm_pk_buff * frag; + size_t total = 0; + size_t i; + uint32_t k = frcti->rcv_cr.rwe - RQ_SIZE; - pthread_cleanup_push(__cleanup_mutex_unlock, &frcti->mtx); + *overflow = false; - ret = -__timedwait(&frcti->cond, &frcti->mtx, abstime); + for (i = 0; i < count; ++i) { + size_t flen; - pthread_cleanup_pop(false); + frag = rq_frag(frcti, k + i); + flen = ssm_pk_buff_len(frag); + if (total + flen < total) { + *overflow = true; + return 0; + } + total += flen; + } - if (ret == -ETIMEDOUT) { - time_t diff; + return total; +} - clock_gettime(PTHREAD_COND_CLOCK, &now); +/* + * Process a delivered slot at lwe: latch FIN if acceptable, + * advance byte_high (clamped to byte_fin once latched). + */ +static __inline__ void stream_deliver_slot(struct frcti * frcti, + size_t lp) +{ + uint32_t end; - diff = ts_diff_ns(&now, &frcti->t_wnd); - if (diff > MAX_RDV) { - pthread_mutex_unlock(&frcti->mtx); - return -ECONNRESET; /* write fails! */ - } + end = frcti->rcv_slots[lp].end; - diff = ts_diff_ns(&now, &frcti->t_rdvs); - if (diff > frcti->rdv) { - frcti->t_rdvs = now; - __send_rdv(frcti->fd); - } + if (frcti->rcv_slots[lp].fin) { + if (end == frcti->rcv_byte_high && !frcti->rcv_fin_seen) { + frcti->rcv_fin_seen = true; + frcti->rcv_byte_fin = end; + } else { + STAT_BUMP(frcti, strm_fin_drop); } + } + + if (frcti->rcv_fin_seen && after(end, frcti->rcv_byte_fin)) + end = frcti->rcv_byte_fin; + + frcti->rcv_byte_high = end; +} + +/* Two-segment memcpy from buf into the rx ring at byte offset start. */ +static void stream_ring_write(struct frcti * frcti, + uint32_t start, + buffer_t buf) +{ + size_t mask = frcti->rcv_ring_sz - 1; + size_t off = start & mask; - pthread_mutex_unlock(&frcti->mtx); - pthread_rwlock_rdlock(&frcti->lock); + if (off + buf.len <= frcti->rcv_ring_sz) { + memcpy(frcti->rcv_ring + off, buf.data, buf.len); + } else { + size_t first = frcti->rcv_ring_sz - off; + memcpy(frcti->rcv_ring + off, buf.data, first); + memcpy(frcti->rcv_ring, buf.data + first, buf.len - first); } +} +/* Two-segment memcpy from the rx ring at byte offset start into buf. */ +static void stream_ring_read(struct frcti * frcti, + uint32_t start, + buffer_t buf) +{ + size_t mask = frcti->rcv_ring_sz - 1; + size_t off = start & mask; + + if (off + buf.len <= frcti->rcv_ring_sz) { + memcpy(buf.data, frcti->rcv_ring + off, buf.len); + } else { + size_t first = frcti->rcv_ring_sz - off; + memcpy(buf.data, frcti->rcv_ring + off, first); + memcpy(buf.data + first, frcti->rcv_ring, buf.len - first); + } +} + +/* Deliver-or-drop one stashed slot at lwe; advance lwe/rwe. Caller wrlock. */ +static void stream_advance_lwe(struct frcti * frcti) +{ + size_t lp; + + lp = RQ_SLOT(frcti->rcv_cr.lwe); + + if (frcti->rcv_slots[lp].start != frcti->rcv_byte_high) + STAT_BUMP(frcti, strm_drop); + else + stream_deliver_slot(frcti, lp); + + frcti->rcv_slots[lp].fin = 0; + frcti->rcv_slots[lp].idx = -1; + STORE_RELEASE(&frcti->rcv_cr.lwe, frcti->rcv_cr.lwe + 1); + frcti->rcv_cr.rwe++; +} + +/* + * Validate a stream DATA packet before stashing. Returns 0 if the + * packet may be written into rcv_ring + rq[], -1 otherwise. + */ +static __inline__ int stream_stash_check(struct frcti * frcti, + uint32_t start, + uint32_t end, + size_t plen, + uint16_t flags) +{ + if (end - start != (uint32_t) plen) + return -1; + + /* FIN MUST be 0-byte. */ + if ((flags & FRCT_FIN) && plen != 0) + return -1; + + /* Post-EOS: no further FIN once latched. */ + if (frcti->rcv_fin_seen && (flags & FRCT_FIN)) + return -1; + + /* Post-EOS: reject data at or past byte_fin. */ + if (frcti->rcv_fin_seen && !before(start, frcti->rcv_byte_fin)) + return -1; + + /* Stale: peer is behind the delivered edge. */ + if (before(end, frcti->rcv_byte_next)) + return -1; + + /* Exact-edge: only an empty-stream FIN is meaningful. */ + if (end == frcti->rcv_byte_next && !(flags & FRCT_FIN)) + return -1; + + if (end - frcti->rcv_byte_next > frcti->rcv_ring_sz) + return -1; + + return 0; +} + +/* + * Stream-mode DATA receive: validate, stash payload in rcv_ring, mark + * rq[pos], advance lwe through any newly-contiguous run. Returns 0 + * (spb released) or -1 (caller releases). Caller wrlock. + */ +static int frcti_stream_data_rcv(struct frcti * frcti, + struct ssm_pk_buff * spb, + size_t pos, + uint16_t flags) +{ + struct frct_pci_stream * spci; + uint32_t start; + uint32_t end; + buffer_t buf; + size_t skip; + + if (ssm_pk_buff_len(spb) < FRCT_PCI_STREAM_LEN) + return -1; + + if (frcti->rcv_ring == NULL) { + frcti->rcv_ring = calloc(1, frcti->rcv_ring_sz); + if (frcti->rcv_ring == NULL) + return -ENOMEM; + } + + spci = FRCT_HDR_POP(spb, frct_pci_stream); + start = ntoh32(spci->start); + end = ntoh32(spci->end); + + buf.data = ssm_pk_buff_head(spb); + buf.len = ssm_pk_buff_len(spb); + + if (stream_stash_check(frcti, start, end, buf.len, flags) < 0) + return -1; + + /* Trim front-overlap with already-delivered region. */ + if (before(start, frcti->rcv_byte_next)) { + skip = frcti->rcv_byte_next - start; + buf.data += skip; + buf.len -= skip; + start = frcti->rcv_byte_next; + } + + stream_ring_write(frcti, start, buf); + STAT_ADD(frcti, strm_rcv_byte, buf.len); + + frcti->rcv_slots[pos].idx = 1; + frcti->rcv_slots[pos].start = start; + frcti->rcv_slots[pos].end = end; + frcti->rcv_slots[pos].fin = (flags & FRCT_FIN) ? 1 : 0; + + while (frcti->rcv_slots[RQ_SLOT(frcti->rcv_cr.lwe)].idx != -1) + stream_advance_lwe(frcti); + + frct_spb_release(spb); + + return 0; +} + +/* + * DATA receive: stash idx at rq[pos], advance lwe through any + * contiguous run. Caller wrlock. + */ +static void frcti_data_stash(struct frcti * frcti, + ssize_t idx, + size_t pos, + uint16_t flags) +{ + frcti->rcv_slots[pos].idx = idx; + + if ((flags & FRCT_FR_MASK) != FRCT_FR_SOLE) + STAT_BUMP(frcti, frag_rcv); + + /* lwe = cum-ACK edge; advance per fragment through contiguous run. */ + while (before(frcti->rcv_cr.lwe, frcti->rcv_cr.rwe) + && frcti->rcv_slots[RQ_SLOT(frcti->rcv_cr.lwe)].idx != -1) + STORE_RELEASE(&frcti->rcv_cr.lwe, frcti->rcv_cr.lwe + 1); +} + +/* Stream consume: copy up to `count` contiguous bytes from ring into buf. */ +static ssize_t frcti_consume_stream(struct frcti * frcti, + uint8_t * buf, + size_t count) +{ + size_t avail; + size_t copy; + ssize_t ret; + buffer_t dst; + + assert(frcti); + + pthread_rwlock_wrlock(&frcti->lock); + + avail = (size_t) (frcti->rcv_byte_high - frcti->rcv_byte_next); + if (avail == 0) { + /* EOS drained: signal EOF to the reader. */ + if (frcti->rcv_fin_seen + && frcti->rcv_byte_next == frcti->rcv_byte_fin) + ret = 0; + else + ret = -EAGAIN; + goto unlock; + } + + copy = MIN(avail, count); + + dst.data = buf; + dst.len = copy; + stream_ring_read(frcti, frcti->rcv_byte_next, dst); + + frcti->rcv_byte_next += (uint32_t) copy; + STAT_ADD(frcti, strm_dlv_byte, copy); + + ret = (ssize_t) copy; + + unlock: pthread_rwlock_unlock(&frcti->lock); return ret; } -static ssize_t __frcti_queued_pdu(struct frcti * frcti) +/* + * FRTX consume: copy next ready PDU (full SDU or nothing). Returns bytes, + * -EAGAIN (no PDU), or -EMSGSIZE (oversize: run dropped to unblock flow). + */ +static ssize_t frcti_consume(struct frcti * frcti, + uint8_t * buf, + size_t count) { - ssize_t idx; - size_t pos; + size_t n; + size_t total; + bool overflow; + enum frag_state st; + ssize_t ret; assert(frcti); - /* See if we already have the next PDU. */ pthread_rwlock_wrlock(&frcti->lock); - pos = frcti->rcv_cr.lwe & (RQ_SIZE - 1); - - idx = frcti->rq[pos]; - if (idx != -1) { - ++frcti->rcv_cr.lwe; - ++frcti->rcv_cr.rwe; - frcti->rq[pos] = -1; + while (true) { + st = frag_run_inspect(frcti, &n); + if (st == FRAG_NOT_READY) { + ret = -EAGAIN; + goto unlock; + } + if (st == FRAG_DROP) { + STAT_ADD(frcti, frag_drop, n); + frag_drop(frcti, n); + continue; + } + /* FRAG_DELIVER */ + total = frag_total_len(frcti, n, &overflow); + if (overflow || total > frcti->max_rcv_sdu || total > count) { + STAT_ADD(frcti, frag_drop, n); + frag_drop(frcti, n); + ret = -EMSGSIZE; + goto unlock; + } + ret = (ssize_t) frag_gather(frcti, n, buf); + if (n > 1) + STAT_BUMP(frcti, sdu_reasm); + else + STAT_BUMP(frcti, sdu_sole); + goto unlock; } + unlock: pthread_rwlock_unlock(&frcti->lock); - return idx; + return ret; } -static ssize_t __frcti_pdu_ready(struct frcti * frcti) +static bool frcti_pdu_ready(struct frcti * frcti) { - ssize_t idx; - size_t pos; + size_t pos; + size_t count; + bool ready; assert(frcti); - /* See if we already have the next PDU. */ pthread_rwlock_rdlock(&frcti->lock); - pos = frcti->rcv_cr.lwe & (RQ_SIZE - 1); - idx = frcti->rq[pos]; + if (frcti->stream) { + ready = frcti->rcv_byte_high != frcti->rcv_byte_next; + pthread_rwlock_unlock(&frcti->lock); + return ready; + } + + if (frag_run_inspect(frcti, &count) != FRAG_DELIVER) { + /* Drop case: frcti_consume will handle it; not ready. */ + pthread_rwlock_unlock(&frcti->lock); + return false; + } + + pos = RQ_SLOT(frcti->rcv_cr.rwe - RQ_SIZE); + ready = frcti->rcv_slots[pos].idx != -1; + + pthread_rwlock_unlock(&frcti->lock); + + return ready; +} + +/* No srtt yet: probe at the cold-probe cadence to seed it. */ +#define PROBE_DUE_COLD(frcti, now_ns) \ + ((now_ns) - (frcti)->t_snd_probe > (uint64_t) RTTP_COLD_NS) + +/* Have srtt: probe when peer quiet for > 2*srtt and last probe > srtt. */ +#define PROBE_DUE_WARM(frcti, now_ns) \ + ((now_ns) - (frcti)->t_rcv_rtt > 2u * (uint64_t)(frcti)->srtt \ + && (now_ns) - (frcti)->t_snd_probe > (uint64_t)(frcti)->srtt) + +/* Seeds srtt for receive-only sides so they don't fall back to 1 s RTO. */ +__attribute__((cold)) +static void frcti_rcv_probe(struct frcti * frcti, + uint64_t now_ns) +{ + uint32_t probe_id; + uint8_t nonce[RTTP_NONCE_LEN] = { 0 }; + + pthread_rwlock_wrlock(&frcti->lock); + + if (frcti->srtt == 0 && !PROBE_DUE_COLD(frcti, now_ns)) { + pthread_rwlock_unlock(&frcti->lock); + return; + } + + if (frcti->srtt != 0 && !PROBE_DUE_WARM(frcti, now_ns)) { + pthread_rwlock_unlock(&frcti->lock); + return; + } + + probe_id = rttp_alloc_probe(frcti, now_ns, nonce); pthread_rwlock_unlock(&frcti->lock); - return idx; + if (probe_id != 0) + frcti_rttp_snd(frcti, probe_id, 0, nonce); } -#include <timerwheel.c> +/* Echo at slot `pos` matches our probe: id, slot, nonce all intact. */ +static __inline__ bool probe_echo_matches(struct frcti * frcti, + size_t pos, + uint32_t echo_id, + const uint8_t nonce[RTTP_NONCE_LEN]) +{ + if (frcti->probes[pos].id != echo_id) + return false; + + if (frcti->probes[pos].ts == 0) + return false; + + return memcmp(frcti->probes[pos].nonce, nonce, RTTP_NONCE_LEN) == 0; +} /* - * Send a final ACK for everything that has not been ACK'd. - * If the flow should be kept active for retransmission, - * the returned time will be negative. + * RTT probe (echo_id == 0): bounce the nonce back to peer. + * RTT echo (echo_id != 0): verify nonce + feed sample. */ -static time_t __frcti_dealloc(struct frcti * frcti) +static void frcti_rttp_rcv(struct frcti * frcti, + buffer_t pkt, + uint64_t now_ns) { - struct timespec now; - time_t wait; - int ackno; - int fd = -1; + const struct frct_rttp * rttp; + uint32_t probe_id; + uint32_t echo_id; + uint8_t nonce[RTTP_NONCE_LEN]; + size_t ring_pos; + int64_t elapsed; + uint64_t sample; + + if (pkt.len < RTTP_PAYLOAD) + return; + + rttp = (const struct frct_rttp *) pkt.data; + probe_id = ntoh32(rttp->probe_id); + echo_id = ntoh32(rttp->echo_id); + + /* Forged/malformed: bouncing this would loop on echo_id == 0. */ + if (probe_id == 0 && echo_id == 0) + return; + + memcpy(nonce, rttp->nonce, sizeof(nonce)); + + if (echo_id == 0) { + /* Probe: echo back with same nonce so peer can verify. */ + STAT_BUMP(frcti, rttp_rcv); + frcti_rttp_snd(frcti, 0, probe_id, nonce); + return; + } + + ring_pos = RTTP_POS(echo_id); + + pthread_rwlock_wrlock(&frcti->lock); + + if (!probe_echo_matches(frcti, ring_pos, echo_id, nonce)) { + pthread_rwlock_unlock(&frcti->lock); + return; + } + + elapsed = ts_age_ns(now_ns, frcti->probes[ring_pos].ts); + frcti->probes[ring_pos].ts = 0; + frcti->t_rcv_rtt = now_ns; + + if (elapsed <= 0) { + pthread_rwlock_unlock(&frcti->lock); + return; + } + sample = (uint64_t) elapsed; + + /* Clamp probe sample to RTT_CLAMP_MUL * srtt to avoid poisoning. */ + if (frcti->srtt > 0) + sample = MIN(sample, (uint64_t) frcti->srtt * RTT_CLAMP_MUL); + + rtt_update(frcti, sample, now_ns); + + pthread_rwlock_unlock(&frcti->lock); +} + +/* Honours piggybacked ACK on the KA. */ +static void frcti_ka_rcv(struct frcti * frcti, + const struct frct_pci * pci, + uint64_t now_ns, + uint16_t flags) +{ + uint32_t ka_ackno; + + STORE_RELEASE(&frcti->t_ka_rcv, now_ns); + STAT_BUMP(frcti, ka_rcv); + + if (!(flags & FRCT_ACK)) + return; + + ka_ackno = ntoh32(pci->ackno); + + pthread_rwlock_wrlock(&frcti->lock); + + if (within(ka_ackno, frcti->snd_cr.lwe, frcti->snd_cr.seqno)) + STORE_RELEASE(&frcti->snd_cr.lwe, ka_ackno); + + pthread_rwlock_unlock(&frcti->lock); +} + +/* + * Additive HoL re-emit (carries DRF); runs before rcv_cr->act + * refresh so it doesn't pre-empt peer's first DRF. + */ +__attribute__((cold)) +static void frcti_nack_rcv(struct frcti * frcti) +{ + struct timespec now; + uint64_t now_ns; + size_t hp; + struct rxm_entry * rxm; + void * pkt_copy = NULL; + size_t pkt_len = 0; clock_gettime(PTHREAD_COND_CLOCK, &now); + now_ns = TS_TO_UINT64(now); + + pthread_rwlock_wrlock(&frcti->lock); + + STAT_BUMP(frcti, nack_rcv); + + if (frcti->snd_cr.seqno == frcti->snd_cr.lwe) { + pthread_rwlock_unlock(&frcti->lock); + return; + } + + hp = RQ_SLOT(frcti->snd_cr.lwe); + rxm = LOAD_ACQUIRE(&frcti->snd_slots[hp].rxm); + if (rxm == NULL || RXM_AGED_OUT(rxm->t0, now_ns, frcti->t_r)) { + pthread_rwlock_unlock(&frcti->lock); + return; + } + + pkt_copy = malloc(rxm->len); + if (pkt_copy != NULL) { + memcpy(pkt_copy, rxm->pkt, rxm->len); + pkt_len = rxm->len; + /* Karn: suppress RTT sample. NACK supersedes pending TLP. */ + frcti->snd_slots[hp].flags = + (frcti->snd_slots[hp].flags & ~SND_TLP) + | SND_RTX | SND_FAST_RXM; + frcti->rtt_lwe = frcti->snd_cr.lwe + 1; + STAT_BUMP(frcti, rxm_nack); + } + + pthread_rwlock_unlock(&frcti->lock); + + if (pkt_copy != NULL) { + int ret = fast_rxm_send(frcti, pkt_copy, pkt_len); + if (ret == -EFLOWDOWN || ret == -ENOTALLOC) + STAT_BUMP(frcti, rxm_tx_dead); + free(pkt_copy); + } +} + +__attribute__((cold)) +static void frcti_rdv_rcv(struct frcti * frcti) +{ + uint32_t rwe; pthread_rwlock_rdlock(&frcti->lock); - ackno = frcti->rcv_cr.lwe; - if (frcti->rcv_cr.lwe != frcti->rcv_cr.seqno) - fd = frcti->fd; + rwe = frcti_advert_rwe(frcti); + + pthread_rwlock_unlock(&frcti->lock); + + STAT_BUMP(frcti, rdv_rcv); + + frcti_pkt_snd(frcti, FRCT_FC, 0, rwe); +} + +/* §7.2: PTO = 2*SRTT + max delayed-ACK delay; fallback when unseeded. */ +static __inline__ uint64_t tlp_pto(const struct frcti * frcti) +{ + if (frcti->srtt > 0) + return 2ULL * (uint64_t) frcti->srtt + ACK_DELAY_NS; + + return NACK_COOLDOWN_NS; +} + +/* + * RFC 8985 §7: lazy probe. Re-evaluate on fire — if sender was active + * within PTO, re-post; else probe HoL once and hand off to RTO. + */ +__attribute__((cold)) +static void tlp_due(void * arg) +{ + struct frcti * frcti = arg; + struct timespec now; + uint64_t now_ns; + uint64_t pto; + uint64_t rto_at; + size_t hp; + struct rxm_entry * rxm; + void * pkt_copy = NULL; + size_t pkt_len = 0; + bool re_post = false; + uint64_t deadline = 0; + + clock_gettime(PTHREAD_COND_CLOCK, &now); + now_ns = TS_TO_UINT64(now); + + pthread_rwlock_wrlock(&frcti->lock); + + if (frcti->snd_cr.seqno == frcti->snd_cr.lwe) + goto unlock; - wait = MAX(frcti->rcv_cr.inact - now.tv_sec + frcti->rcv_cr.act.tv_sec, - frcti->snd_cr.inact - now.tv_sec + frcti->snd_cr.act.tv_sec); - wait = MAX(wait, 0); + if (!before(frcti->snd_cr.seqno, frcti->snd_cr.rwe)) + goto unlock; /* FC-blocked: RDV handles it. */ - if (frcti->snd_cr.cflags & FRCTFLINGER - && before(frcti->snd_cr.lwe, frcti->snd_cr.seqno)) - wait = -wait; + /* RFC 8985 §7.3: one outstanding probe, MAX_TLP_PER_EP per ep. */ + if (frcti->tlp_high_seq != 0) + goto unlock; + if (frcti->tlp_count >= MAX_TLP_PER_EP) + goto unlock; + + pto = tlp_pto(frcti); + + /* §7.2: anchor PTO on most recent send; defer if still active. */ + if (now_ns < frcti->snd_cr.act + pto) { + deadline = frcti->snd_cr.act + pto; + re_post = true; + goto unlock; + } + + hp = RQ_SLOT(frcti->snd_cr.lwe); + rxm = LOAD_ACQUIRE(&frcti->snd_slots[hp].rxm); + if (rxm == NULL || RXM_AGED_OUT(rxm->t0, now_ns, frcti->t_r)) + goto unlock; + + /* Cap: if HoL RTO is due, let rxm_due fire instead. */ + rto_at = rxm->t0 + ((uint64_t) frcti->rto + << LOAD_RELAXED(&frcti->rto_mul)); + if (rto_at <= now_ns) + goto unlock; + + pkt_copy = malloc(rxm->len); + if (pkt_copy != NULL) { + memcpy(pkt_copy, rxm->pkt, rxm->len); + pkt_len = rxm->len; + frcti->snd_slots[hp].time = now_ns; + frcti->snd_slots[hp].flags |= SND_TLP | SND_FAST_RXM; + frcti->rtt_lwe = frcti->snd_cr.lwe + 1; + /* §7.3 outstanding-probe marker; ack_rcv/rxm_snd clear. */ + frcti->tlp_high_seq = frcti->snd_cr.seqno; + frcti->tlp_count++; + STAT_BUMP(frcti, tlp_snd); + } + + unlock: pthread_rwlock_unlock(&frcti->lock); - if (fd != -1) - __send_frct_pkt(fd, FRCT_ACK, ackno, 0); + if (pkt_copy != NULL) { + fast_rxm_send(frcti, pkt_copy, pkt_len); + free(pkt_copy); + } - return wait; + if (re_post) + tw_post(&frcti->tlp_tw, deadline, tlp_due, frcti); + else + __atomic_clear(&frcti->tlp_pending, __ATOMIC_RELAXED); } -static int __frcti_snd(struct frcti * frcti, - struct ssm_pk_buff * spb) +/* §7.2 lazy: post once per quiet period. tlp_due re-evaluates on fire. */ +static int tlp_arm(struct frcti * frcti) { - struct frct_pci * pci; - struct timespec now; - struct frct_cr * snd_cr; - struct frct_cr * rcv_cr; - uint32_t seqno; - bool rtx; + struct timespec now; + uint64_t now_ns; + uint64_t pto; + uint64_t deadline; + + /* §7.3: one outstanding probe, MAX_TLP_PER_EP per recovery ep. */ + if (LOAD_RELAXED(&frcti->tlp_high_seq) != 0) + return 0; + if (LOAD_RELAXED(&frcti->tlp_count) >= MAX_TLP_PER_EP) + return 0; + if (__atomic_test_and_set(&frcti->tlp_pending, __ATOMIC_RELAXED)) + return 0; + + clock_gettime(PTHREAD_COND_CLOCK, &now); + now_ns = TS_TO_UINT64(now); + + pto = tlp_pto(frcti); + + deadline = LOAD_RELAXED(&frcti->snd_cr.act) + pto; + if (deadline <= now_ns) + deadline = now_ns + pto; + + tw_post(&frcti->tlp_tw, deadline, tlp_due, frcti); + + return 0; +} + +/* + * FC window advert from any flag-bearing packet. Caps at lwe + RQ_SIZE, + * rejects backward shrink (forged/stale FC), marks window open. + * Caller wrlock. + */ +static __inline__ void frcti_fc_rcv(struct frcti * frcti, + const struct frct_pci * pci) +{ + struct frct_cr * snd_cr; + uint32_t rwe; + uint32_t rwe_max; + + snd_cr = &frcti->snd_cr; + rwe = ntoh32(pci->window); + rwe_max = snd_cr->lwe + RQ_SIZE; + + if (after(rwe, rwe_max)) + rwe = rwe_max; + + /* Reject backward shrink (forged/stale FC). */ + if (before(rwe, snd_cr->rwe)) + rwe = snd_cr->rwe; + + STORE_RELAXED(&snd_cr->rwe, rwe); + frcti->open = true; +} + +/* Packet copies captured under frcti->lock; emitted after release. */ +struct pending { + buffer_t fast_rxm; + buffer_t sack_rxm[SACK_RXM_MAX]; + size_t sack_rxm_cnt; +}; + +/* RFC 6582 §3.2: seal recovery_high on entry; do not extend on new gaps. */ +static void recovery_enter(struct frcti * frcti) +{ + if (frcti->in_recovery) + return; + + frcti->in_recovery = true; + frcti->recovery_high = frcti->snd_cr.seqno + RTT_QUARANTINE; +} + +/* True when cum-ACK clears recovery_high or all in-flight ACKed. */ +static bool recovery_exit_reached(struct frcti * frcti, + uint32_t ackno) +{ + if (!frcti->in_recovery) + return false; + + if (!before(ackno, frcti->recovery_high)) + return true; + + return ackno == frcti->snd_cr.seqno; +} + +/* RTT sample gate: Karn + SACK-consume + don't-seed. */ +static bool rtt_sample_eligible(struct frcti * frcti, + size_t p, + uint16_t flags, + uint32_t lwe) +{ + if (flags & FRCT_RXM) + return false; + if (frcti->snd_slots[p].flags & (SND_RTX | SND_TLP)) + return false; + if (LOAD_ACQUIRE(&frcti->snd_slots[p].rxm) == NULL) + return false; + if (before(lwe, frcti->rtt_lwe)) + return false; + /* Don't seed srtt from a cum-ACK; let probes seed. */ + if (frcti->srtt == 0) + return false; + return true; +} + +#define RXM_SLOT_EMPTY(rxm) ((rxm) == NULL) +#define FAST_RXM_STAGED(pending) ((pending)->fast_rxm.data != NULL) +#define RXM_FAST_DONE(flags) (((flags) & SND_FAST_RXM) != 0) + +/* RACK fast retransmit on cum-ACK: HoL aged past R, not yet retransmitted. */ +static void fast_rxm_consider(struct frcti * frcti, + uint64_t now_ns, + struct pending * pending) +{ + struct rxm_entry * rxm; + struct snd_slot * slot; + size_t hp; + uint64_t R; + bool rack_ok; + + hp = RQ_SLOT(frcti->snd_cr.lwe); + slot = &frcti->snd_slots[hp]; + rxm = LOAD_ACQUIRE(&slot->rxm); + R = rack_reorder_window(frcti); + + if (RXM_SLOT_EMPTY(rxm)) + return; + + /* RFC 8985 §6.2: time-based RACK OR DupThresh count. */ + rack_ok = (int64_t)(frcti->t_latest_ack - slot->time) > (int64_t) R; + if (!rack_ok && frcti->dup_thresh < DUP_THRESH) + return; + + /* HoL aged past t_r; let rxm_due tear the flow down. */ + if (RXM_AGED_OUT(rxm->t0, now_ns, frcti->t_r)) + return; + + /* Already on it. */ + if (FAST_RXM_STAGED(pending) || RXM_FAST_DONE(slot->flags)) + return; + + recovery_enter(frcti); + + pending->fast_rxm.data = malloc(rxm->len); + if (pending->fast_rxm.data == NULL) + return; + + pending->fast_rxm.len = rxm->len; + memcpy(pending->fast_rxm.data, rxm->pkt, rxm->len); + slot->flags |= SND_RTX | SND_FAST_RXM; + frcti->rtt_lwe = frcti->snd_cr.lwe + 1; + if (rack_ok) + STAT_BUMP(frcti, rxm_rack); + else + STAT_BUMP(frcti, rxm_dupthresh); +} + +/* Caller holds wrlock; RACK fast retransmit queued in pending. */ +__attribute__((hot)) +static void frcti_ack_rcv(struct frcti * frcti, + const struct frct_pci * pci, + uint16_t flags, + uint64_t now_ns, + struct pending * pending) +{ + uint32_t ackno; + uint32_t lwe; + size_t p; + size_t fresh; + + if (!(flags & FRCT_DATA)) + STAT_BUMP(frcti, ack_rcv); + + ackno = ntoh32(pci->ackno); + if (ackno == frcti->snd_cr.lwe) { + /* RFC 8985 §6.2: only on scoreboard change. */ + if (frcti->snd_cr.lwe != frcti->rack_fired_lwe) { + fast_rxm_consider(frcti, now_ns, pending); + frcti->rack_fired_lwe = frcti->snd_cr.lwe; + } + return; + } + + if (!within(ackno, frcti->snd_cr.lwe, frcti->snd_cr.seqno)) + return; + + lwe = frcti->snd_cr.lwe; + p = RQ_SLOT(lwe); + + STORE_RELEASE(&frcti->snd_cr.lwe, ackno); + + /* §7.3: cum-ACK past the probed seqno resolves the TLP. */ + if (frcti->tlp_high_seq != 0 + && !before(ackno, frcti->tlp_high_seq)) + frcti->tlp_high_seq = 0; + + /* §7.3: end the probe episode once inflight drains. */ + if (ackno == frcti->snd_cr.seqno) + frcti->tlp_count = 0; + + /* RFC 8985 §7.2: halve mult per REO_DECAY_PKTS fresh-ACK'd seqnos. */ + fresh = ackno - frcti->dsack_lwe_snap; + if (frcti->reo_wnd_mult > 1 && fresh >= REO_DECAY_PKTS) { + uint8_t half = frcti->reo_wnd_mult >> 1; + frcti->reo_wnd_mult = half < 1 ? 1 : half; + frcti->dsack_lwe_snap = ackno; + } + + /* RFC 8985: latest cum-ACKed send-time (slot of ackno-1). */ + frcti->t_latest_ack = frcti->snd_slots[RQ_SLOT(ackno - 1)].time; + + /* RFC 8985: SACK-above-lwe count is per-recovery-episode. */ + frcti->dup_thresh = 0; + + /* Karn-skip on retx; TLP ACK clears rto_mul (no CC backoff). */ + if ((frcti->snd_slots[p].flags & SND_RTX) == 0 + || (frcti->snd_slots[p].flags & SND_TLP) != 0) + STORE_RELEASE(&frcti->rto_mul, 0); + + if (recovery_exit_reached(frcti, ackno)) + frcti->in_recovery = false; + + if (rtt_sample_eligible(frcti, p, flags, lwe)) { + int64_t mrtt = ts_age_ns(now_ns, frcti->snd_slots[p].time); + if (mrtt > 0) { + if (!(flags & FRCT_DATA)) + STAT_BUMP(frcti, ack_rtt); + rtt_update(frcti, (time_t) mrtt, now_ns); + frcti->t_rcv_rtt = now_ns; + } + } +} + +/* Skip k == lwe under clamp: NULLing HoL from a stale SACK wedges it. */ +static uint32_t sack_mark_blocks(struct frcti * frcti, + const uint8_t * payload, + uint16_t n, + uint32_t * newly_marked) +{ + uint32_t hi_sacked = frcti->snd_cr.lwe; + uint32_t marked = 0; + uint16_t i; + + for (i = 0; i < n; ++i) { + uint32_t s; + uint32_t e; + uint32_t k; + bool clamped; + + sack_block_get(payload, i, &s, &e); + + if (!before(s, e)) + continue; + + clamped = before(s, frcti->snd_cr.lwe); + if (clamped) + s = frcti->snd_cr.lwe; + if (after(e, frcti->snd_cr.seqno)) + e = frcti->snd_cr.seqno; + + for (k = s; before(k, e); ++k) { + size_t kp = RQ_SLOT(k); + uint64_t t_k; + if (clamped && k == frcti->snd_cr.lwe) + continue; + if (LOAD_ACQUIRE(&frcti->snd_slots[kp].rxm) == NULL) + continue; + STORE_RELEASE(&frcti->snd_slots[kp].rxm, NULL); + frcti->snd_slots[kp].flags = 0; + marked++; + /* RACK.fack: latest SACK-confirmed send-time. */ + t_k = frcti->snd_slots[kp].time; + if (t_k > frcti->t_latest_ack) + frcti->t_latest_ack = t_k; + } + + if (after(e, hi_sacked)) + hi_sacked = e; + } + + *newly_marked = marked; + return hi_sacked; +} + +/* Queue once per loss event (SND_FAST_RXM gates). Emit after unlock. */ +static void sack_queue_rxm(struct frcti * frcti, + uint32_t hi_sacked, + uint64_t now_ns, + struct pending * pending) +{ + uint64_t R = rack_reorder_window(frcti); + uint32_t k; + bool rack_ok; + + for (k = frcti->snd_cr.lwe; before(k, hi_sacked); ++k) { + struct rxm_entry * rxm; + size_t kp = RQ_SLOT(k); + size_t cnt = pending->sack_rxm_cnt; + size_t rack_age; + + rxm = LOAD_ACQUIRE(&frcti->snd_slots[kp].rxm); + + if (cnt >= SACK_RXM_MAX) + break; + + if (rxm == NULL) + continue; + + if (frcti->snd_slots[kp].flags & SND_FAST_RXM) + continue; + + if (RXM_AGED_OUT(rxm->t0, now_ns, frcti->t_r)) + continue; + + rack_age = frcti->t_latest_ack - frcti->snd_slots[kp].time; + /* RFC 8985 §6.2: time-based RACK OR DupThresh count. */ + rack_ok = (int64_t) rack_age > (int64_t) R; + if (!rack_ok && frcti->dup_thresh < DUP_THRESH) + continue; + + if (rack_ok) + STAT_BUMP(frcti, rxm_rack); + else + STAT_BUMP(frcti, rxm_dupthresh); + + pending->sack_rxm[cnt].data = malloc(rxm->len); + if (pending->sack_rxm[cnt].data == NULL) + break; + + pending->sack_rxm[cnt].len = rxm->len; + memcpy(pending->sack_rxm[cnt].data, rxm->pkt, rxm->len); + pending->sack_rxm_cnt++; + /* NULL slot so the original timer self-cleans. */ + STORE_RELEASE(&frcti->snd_slots[kp].rxm, NULL); + frcti->snd_slots[kp].time = now_ns; + frcti->snd_slots[kp].flags |= SND_RTX | SND_FAST_RXM; + frcti->rtt_lwe = k + 1; + } +} + +/* + * RFC 2883 D-SACK detector. Returns true iff block[0] is a D-SACK + * report: + * case 1: blocks[0].start < pkt_ackno (strictly below cum-ACK). + * case 2: blocks[0] is a strict sub-range of some blocks[i>0]. + * MAX_DSACK_LAG bounds case-1 distance to one rcv window (sanity). + */ +static bool sack_is_dsack(struct frcti * frcti, + const uint8_t * payload, + uint16_t n, + uint32_t pkt_ackno) +{ + uint32_t s0; + uint32_t e0; + uint16_t i; + + if (n == 0) + return false; + + sack_block_get(payload, 0, &s0, &e0); + if (!before(s0, e0)) + return false; + + if (before(s0, pkt_ackno)) { + if ((pkt_ackno - s0) <= (uint32_t) MAX_DSACK_LAG) + return true; + STAT_BUMP(frcti, dsack_drop); + return false; + } + + for (i = 1; i < n; ++i) { + uint32_t si; + uint32_t ei; + + sack_block_get(payload, i, &si, &ei); + if (!before(si, ei)) + continue; + if (!before(s0, si) && !after(e0, ei) + && (s0 != si || e0 != ei)) + return true; + } + + return false; +} + +/* RFC 8985 §7.2: grow reo_wnd_mult on DSACK; at most once per RTT. */ +static __inline__ void reo_wnd_on_dsack(struct frcti * frcti, + uint64_t now_ns) +{ + time_t srtt = frcti->srtt; + + /* Snap is unconditional: feeds the per-D-SACK decay clock. */ + frcti->dsack_lwe_snap = frcti->snd_cr.lwe; + + if (srtt > 0 + && now_ns - frcti->t_last_reo_widen <= (uint64_t) srtt) + return; + + if (frcti->reo_wnd_mult < REO_WND_MULT_MAX) + frcti->reo_wnd_mult++; + + frcti->t_last_reo_widen = now_ns; +} + +/* Caller holds wrlock; retransmits queued for post-unlock emission. */ +static void frcti_sack_rcv(struct frcti * frcti, + buffer_t pkt, + uint32_t pkt_ackno, + uint64_t now_ns, + struct pending * pending) +{ + uint32_t hi_sacked; + uint32_t marked; + uint16_t n; + bool dsack; + uint16_t n_real; + + if (pkt.len < SACK_HDR_SIZE) + return; + + n = ntoh16(*(const uint16_t *) pkt.data); + if (n > SACK_MAX_BLOCKS) + return; + + if (pkt.len < SACK_HDR_SIZE + (size_t) n * SACK_BLOCK_SIZE) + return; + + STAT_BUMP(frcti, sack_rcv); + + dsack = sack_is_dsack(frcti, pkt.data, n, pkt_ackno); + n_real = n - (dsack ? 1 : 0); + + if (dsack) { + STAT_BUMP(frcti, dsack_rcv); + reo_wnd_on_dsack(frcti, now_ns); + } + + /* DSACK-only carries no new gap; don't enter recovery. */ + if (n_real > 0) + recovery_enter(frcti); + + marked = 0; + hi_sacked = sack_mark_blocks(frcti, pkt.data, n, &marked); + frcti->dup_thresh += marked; + + if (after(hi_sacked, frcti->snd_cr.lwe)) + sack_queue_rxm(frcti, hi_sacked, now_ns, pending); +} + +/* Emit and free queued packet copies. */ +static void pending_flush(struct frcti * frcti, + struct pending * pending) +{ + size_t i; + + for (i = 0; i < pending->sack_rxm_cnt; ++i) { + sack_rxm_snd(frcti, pending->sack_rxm[i].data, + pending->sack_rxm[i].len); + free(pending->sack_rxm[i].data); + } + + if (pending->fast_rxm.data != NULL) { + int ret = fast_rxm_send(frcti, pending->fast_rxm.data, + pending->fast_rxm.len); + if (ret == -EFLOWDOWN || ret == -ENOTALLOC) + STAT_BUMP(frcti, rxm_tx_dead); + free(pending->fast_rxm.data); + } +} + +/* Pre-DRF NACK: ask peer to retransmit HoL; seqno is informational. */ +static void frcti_nack_snd(struct frcti * frcti, + uint32_t seqno_unseen) +{ + struct ssm_pk_buff * spb; + struct frct_pci * pci; + + if (frct_ctrl_alloc(&spb, &pci, 0) < 0) + return; + + pci->flags = hton16(FRCT_NACK); + pci->seqno = hton32(seqno_unseen); + + frct_hcs_set(pci, false); + + frct_tx(frcti, spb); +} + +enum frct_act { + FRCT_ACTIVE, + FRCT_INACT_NEED_NACK, + FRCT_INACT_DROP, +}; + +/* On rcv inactivity: rebase on DRF, or arm pre-DRF NACK. Caller wrlock. */ +static enum frct_act rcv_inact_check(struct frcti * frcti, + uint16_t flags, + uint32_t seqno, + uint64_t now_ns) +{ + struct frct_cr * rcv_cr = &frcti->rcv_cr; + uint64_t cd; + + if (!ts_aged_ns(now_ns, rcv_cr->act, rcv_cr->inact)) + return FRCT_ACTIVE; + + if (flags & FRCT_DRF) { + if (same_epoch_drf(seqno, flags, rcv_cr)) + return FRCT_ACTIVE; + + /* Bootstrap or fresh epoch: rebase. */ + STAT_BUMP(frcti, drf_rebase); + release_rq(frcti); + STORE_RELEASE(&rcv_cr->lwe, seqno); + rcv_cr->rwe = seqno + RQ_SIZE; + rcv_cr->seqno = seqno; + return FRCT_ACTIVE; + } + + if (!(flags & FRCT_DATA)) + return FRCT_ACTIVE; + + /* Pre-DRF: nudge sender with NACK (rate-limited). */ + cd = frcti->srtt > 0 ? (uint64_t) frcti->srtt : NACK_COOLDOWN_NS; + if (!ts_aged_ns(now_ns, frcti->t_nack, cd)) + return FRCT_INACT_DROP; + + frcti->t_nack = now_ns; + STAT_BUMP(frcti, nack_snd); + + return FRCT_INACT_NEED_NACK; +} + +/* Both modes: bounded accept into rq[seqno]. Caller wrlock. */ +__attribute__((hot)) +static bool rq_accept(struct frcti * frcti, + uint32_t seqno, + size_t pos, + uint16_t flags) +{ + struct frct_cr * rcv_cr = &frcti->rcv_cr; + + if (!before(seqno, rcv_cr->rwe)) { + STAT_BUMP(frcti, out_rcv); + return false; + } + + if (!before(seqno, rcv_cr->lwe + RQ_SIZE)) { + STAT_BUMP(frcti, rqo_rcv); + return false; + } + + if (frcti->rcv_slots[pos].idx != -1) { + if (flags & FRCT_RXM) + STAT_BUMP(frcti, rxm_dup_rcv); + else + STAT_BUMP(frcti, dup_rcv); + /* RFC 2883 §4 case 2: in-window dup; sub-range marker. */ + frcti->dsack_seqno = seqno; + frcti->dsack_valid = true; + return false; + } + + return true; +} + +/* OOO arrival; throttle by min_gap + scoreboard dedup. */ +static bool sack_check(struct frcti * frcti, + uint32_t seqno, + uint64_t now_ns, + struct sack_args * out) +{ + struct frct_cr * rcv_cr = &frcti->rcv_cr; + uint64_t min_gap; + uint16_t n; + + if (!after(seqno, rcv_cr->lwe)) + return false; + + STAT_BUMP(frcti, ooo_rcv); + + /* SACK carries cum-ACK; bound by t_a like any other ACK. */ + if (ACK_AGED_OUT(rcv_cr->act, now_ns, frcti->t_a)) + return false; + + /* srtt/8 gate starved recovery under burst loss; floor to save CPU. */ + min_gap = (uint64_t) SACK_MIN_GAP_NS; + + if (!ts_aged_ns(now_ns, frcti->t_snd_sack, min_gap)) + return false; + + out->dsack = false; + n = dsack_consume(frcti, out->blocks); + if (n == 1) + out->dsack = true; + n += sack_blocks_build(frcti, out->blocks + n, + frcti->sack_n_max - n); + + if (!out->dsack + && rcv_cr->lwe == frcti->sack_lwe && n == frcti->sack_n) + return false; + + out->n = n; + out->ack = rcv_cr->lwe; + out->rwe = frcti_advert_rwe(frcti); + frcti->t_snd_sack = now_ns; + frcti->sack_lwe = rcv_cr->lwe; + frcti->sack_n = n; + + return true; +} + +/* Wire-dup of fresh DATA at an already-ACKed seqno. */ +static __inline__ bool is_dup_data(uint16_t flags, + uint32_t seqno, + uint32_t lwe) +{ + if (!(flags & FRCT_DATA)) + return false; + + if (flags & FRCT_RXM) + return false; + + return before(seqno, lwe); +} + +/* + * Wire-dup ACK packet: same seqno as the previous emission. Updates + * the dedup ackno on a fresh ACK; caller drops on true. + */ +static __inline__ bool is_dup_ack(struct frcti * frcti, + uint16_t flags, + uint32_t seqno) +{ + if (flags & FRCT_DATA) + return false; + + if (!(flags & FRCT_ACK)) + return false; + + if (seqno == frcti->rcv_cr.ackno) + return true; + + frcti->rcv_cr.ackno = seqno; + + return false; +} + +/* Caller wrlock. */ +__attribute__((cold)) +static void seqno_rotate(struct frcti * frcti, + uint64_t now_ns) +{ + struct frct_cr * snd_cr = &frcti->snd_cr; + + if (!ts_aged_ns(now_ns, snd_cr->act, snd_cr->inact)) + return; + /* Idle-on-wire ≠ idle e2e: don't orphan in-flight rxm. */ + if (snd_cr->seqno != snd_cr->lwe) + return; + + /* Avoid colliding with peer's current rcv window. */ + do { + random_buffer(&snd_cr->seqno, sizeof(snd_cr->seqno)); + } while (in_window(snd_cr->seqno, snd_cr)); + STORE_RELEASE(&snd_cr->lwe, snd_cr->seqno); + STORE_RELAXED(&snd_cr->rwe, snd_cr->lwe + START_WINDOW); + frcti->rtt_lwe = snd_cr->seqno; + frcti->in_recovery = false; + frcti->recovery_high = snd_cr->seqno; +} + +__attribute__((hot)) +static int frcti_snd(struct frcti * frcti, + struct ssm_pk_buff * spb, + uint16_t flags) +{ + struct frct_pci * pci; + struct frct_pci_stream * spci = NULL; + struct timespec now; + struct frct_cr * snd_cr; + struct frct_cr * rcv_cr; + struct rxm_entry * rxm = NULL; + uint32_t seqno; + uint16_t pci_flags = 0; + bool rtx; + uint64_t now_ns; + int64_t rcv_idle; + uint32_t probe_id = 0; + uint8_t probe_nonce[RTTP_NONCE_LEN] = { 0 }; + bool probe; + size_t payload_len = 0; assert(frcti); - assert(ssm_pk_buff_len(spb) != 0); + /* Stream mode permits 0-byte sends for the EOS marker. */ + assert(ssm_pk_buff_len(spb) != 0 || frcti->stream); snd_cr = &frcti->snd_cr; rcv_cr = &frcti->rcv_cr; - timerwheel_move(); + tw_move_safe(); + + if (frcti->stream) + payload_len = ssm_pk_buff_len(spb); - pci = (struct frct_pci *) ssm_pk_buff_head_alloc(spb, FRCT_PCILEN); + pci = FRCT_HDR_PUSH(spb, frcti); if (pci == NULL) return -ENOMEM; - memset(pci, 0, sizeof(*pci)); + /* Pre-allocate rxm so alloc fail can't orphan a seqno. */ + if (snd_cr->cflags & FRCTFRTX) { + rxm = rxm_alloc(frcti, ssm_pk_buff_len(spb)); + if (rxm == NULL) { + ssm_pk_buff_pop(spb, frcti_data_hdr_len(frcti)); + return -ENOMEM; + } + } + + memset(pci, 0, FRCT_PCILEN); + + if (frcti->stream) + spci = FRCT_SPCI(pci); clock_gettime(PTHREAD_COND_CLOCK, &now); + now_ns = TS_TO_UINT64(now); pthread_rwlock_wrlock(&frcti->lock); rtx = snd_cr->cflags & FRCTFRTX; - pci->flags |= FRCT_DATA; + pci_flags |= FRCT_DATA; + if (!frcti->stream) + pci_flags |= (flags & FRCT_FR_MASK); - /* Set DRF if there are no unacknowledged packets. */ - if (snd_cr->seqno == snd_cr->lwe) - pci->flags |= FRCT_DRF; + if (!frcti->stream && (flags & FRCT_FR_MASK) != FRCT_FR_SOLE) + STAT_BUMP(frcti, frag_snd); - /* Choose a new sequence number if sender inactivity expired. */ - if (now.tv_sec - snd_cr->act.tv_sec > snd_cr->inact) { - /* There are no unacknowledged packets. */ - assert(snd_cr->seqno == snd_cr->lwe); - random_buffer(&snd_cr->seqno, sizeof(snd_cr->seqno)); - snd_cr->lwe = snd_cr->seqno; - snd_cr->rwe = snd_cr->lwe + START_WINDOW; + if (frcti->stream) { + if (flags & FRCT_FIN) + pci_flags |= FRCT_FIN; + + spci->start = hton32(frcti->snd_byte_next); + frcti->snd_byte_next += (uint32_t) payload_len; + spci->end = hton32(frcti->snd_byte_next); + STAT_ADD(frcti, strm_snd_byte, payload_len); } + if (snd_cr->seqno == snd_cr->lwe) + pci_flags |= FRCT_DRF; + + seqno_rotate(frcti, now_ns); + seqno = snd_cr->seqno; pci->seqno = hton32(seqno); - if (now.tv_sec - rcv_cr->act.tv_sec < rcv_cr->inact) { - pci->flags |= FRCT_FC; - *((uint32_t *) pci) |= hton32(rcv_cr->rwe & 0x00FFFFFF); + rcv_idle = ts_age_ns(now_ns, rcv_cr->act); + + if (rcv_idle < (int64_t) rcv_cr->inact) { + pci_flags |= FRCT_FC; + pci->window = hton32(frcti_advert_rwe(frcti)); } if (!rtx) { - snd_cr->lwe++; + STORE_RELEASE(&snd_cr->lwe, snd_cr->lwe + 1); + STORE_RELEASE(&snd_cr->rwe, snd_cr->lwe + RQ_SIZE); } else { - if (!frcti->probe) { - frcti->rttseq = snd_cr->seqno; - frcti->t_probe = now; - frcti->probe = true; -#ifdef PROC_FLOW_STATS - frcti->n_prb++; -#endif - } - if ((now.tv_sec - rcv_cr->act.tv_sec) * BILLION <= frcti->a) { - pci->flags |= FRCT_ACK; + size_t p = RQ_SLOT(seqno); + frcti->snd_slots[p].time = now_ns; + /* Fresh send clears RTX bits. */ + frcti->snd_slots[p].flags = 0; + if (rcv_idle <= (int64_t) frcti->t_a) { + pci_flags |= FRCT_ACK; pci->ackno = hton32(rcv_cr->lwe); rcv_cr->seqno = rcv_cr->lwe; } } + pci->flags = hton16(pci_flags); + + frct_hcs_set(pci, frcti->stream); + snd_cr->seqno++; - snd_cr->act = now; + STORE_RELEASE(&snd_cr->act, now_ns); + + probe = rtt_probe_arm(frcti, now_ns, &probe_id, probe_nonce); pthread_rwlock_unlock(&frcti->lock); - if (rtx) - timerwheel_rxm(frcti, seqno, spb); + if (probe) + frcti_rttp_snd(frcti, probe_id, 0, probe_nonce); + + if (rtx) { + assert(rxm != NULL); + rxm_arm(frcti, seqno, rxm, spb); + tlp_arm(frcti); + } return 0; } -static void rtt_estimator(struct frcti * frcti, - time_t mrtt) +/* Stream FIN is armed for rxm; needs to be in window. */ +static __inline__ bool stream_fin_blocked(struct frcti * frcti) { - time_t srtt = frcti->srtt; - time_t rttvar = frcti->mdev; + if (!frcti->stream) + return false; - if (srtt == 0) { /* first measurement */ - srtt = mrtt; - rttvar = mrtt >> 1; - } else { - time_t delta = mrtt - srtt; - srtt += (delta >> 3); - delta = (ABS(delta) - rttvar) >> 2; -#ifdef FRCT_LINUX_RTT_ESTIMATOR - if (delta < 0) - delta >>= 3; -#endif - rttvar += delta; + return !before(frcti->snd_cr.seqno, frcti->snd_cr.lwe + RQ_SIZE); +} + +/* + * Stream: 0-byte FRCT_FIN DATA so peer's flow_read returns 0 at this + * byte. Msg: control packet with FRCT_FIN flag, snd_cr.seqno carried + * in pci->ackno (sender packs via frcti_pkt_snd's ackno parameter). + */ +static void frcti_fin_snd(struct frcti * frcti) +{ + struct ssm_pk_buff * spb; + bool already; + uint32_t fin_seqno; + + if (!(frcti->snd_cr.cflags & FRCTFLINGER)) + return; + + pthread_rwlock_wrlock(&frcti->lock); + + already = frcti->snd_fin_sent; + + /* Defer before committing snd_fin_sent; linger loop retries. */ + if (!already && stream_fin_blocked(frcti)) { + pthread_rwlock_unlock(&frcti->lock); + return; } -#ifdef PROC_FLOW_STATS - frcti->n_rtt++; -#endif - frcti->srtt = MAX(1000L, srtt); - frcti->mdev = MAX(100L, rttvar); - frcti->rto = MAX(RTO_MIN, frcti->srtt + (frcti->mdev << MDEV_MUL)); -} - -/* Always queues the next application packet on the RQ. */ -static void __frcti_rcv(struct frcti * frcti, - struct ssm_pk_buff * spb) -{ - ssize_t idx; - size_t pos; - struct frct_pci * pci; - struct timespec now; - struct frct_cr * rcv_cr; - struct frct_cr * snd_cr; - uint32_t seqno; - uint32_t ackno; - uint32_t rwe; - int fd = -1; - assert(frcti); + frcti->snd_fin_sent = true; + fin_seqno = frcti->snd_cr.seqno; + + if (!already && !frcti->stream) + frcti->snd_fin_seqno = fin_seqno; + pthread_rwlock_unlock(&frcti->lock); + + if (already) + return; + + if (!frcti->stream) { + frcti_pkt_snd(frcti, FRCT_FIN, fin_seqno, 0); + return; + } + + if (frct_spb_reserve(frcti_data_hdr_len(frcti), &spb) < 0) + return; + + /* Reset spb to 0-len so frcti_snd's head_alloc populates PCI. */ + ssm_pk_buff_truncate(spb, 0); + + if (frcti_snd(frcti, spb, FRCT_FIN) < 0) { + frct_spb_release(spb); + return; + } + + if (frct_tx(frcti, spb) < 0) + return; + + pthread_rwlock_wrlock(&frcti->lock); + + frcti->snd_fin_seqno = frcti->snd_cr.seqno - 1; + + pthread_rwlock_unlock(&frcti->lock); +} + +static bool final_ack_due(struct frcti * frcti, + struct frct_cr * rcv_cr, + uint64_t now_ns) +{ + if (rcv_cr->lwe == rcv_cr->seqno) + return false; + + if (ACK_AGED_OUT(rcv_cr->act, now_ns, frcti->t_a)) + return false; + + return true; +} + +/* Snd-side has FLINGER cflag and unACK'd data below the FIN/seqno. */ +static __inline__ bool snd_drain_pending(struct frct_cr * snd_cr, + uint32_t edge) +{ + if (!(snd_cr->cflags & FRCTFLINGER)) + return false; + + return before(snd_cr->lwe, edge); +} + +/* Peer is still active and we haven't seen their FIN yet. */ +static __inline__ bool rcv_drain_pending(struct frcti * frcti, + struct frct_cr * rcv_cr, + uint64_t now_ns) +{ + if (frcti->rcv_fin_seen) + return false; + + return !ts_aged_ns(now_ns, rcv_cr->act, rcv_cr->inact); +} + +/* Drain-loop predicate: snd-side unACK'd data OR peer still active. */ +static bool frcti_lingering(struct frcti * frcti) +{ + struct timespec now; + struct frct_cr * snd_cr; + struct frct_cr * rcv_cr; + uint32_t edge; + uint64_t now_ns; + bool snd_linger; + bool rcv_linger; + + /* Idempotent; emits FIN once per side, both stream and msg. */ + frcti_fin_snd(frcti); + + clock_gettime(PTHREAD_COND_CLOCK, &now); + now_ns = TS_TO_UINT64(now); + + pthread_rwlock_rdlock(&frcti->lock); + + snd_cr = &frcti->snd_cr; rcv_cr = &frcti->rcv_cr; + + if (frcti->snd_fin_sent) + edge = frcti->snd_fin_seqno; + else + edge = snd_cr->seqno; + + snd_linger = snd_drain_pending(snd_cr, edge); + rcv_linger = rcv_drain_pending(frcti, rcv_cr, now_ns); + + pthread_rwlock_unlock(&frcti->lock); + + return snd_linger || rcv_linger; +} + +static time_t frcti_dealloc(struct frcti * frcti) +{ + struct timespec now; + struct frct_cr * snd_cr; + struct frct_cr * rcv_cr; + int ackno; + bool due; + int64_t now_ns; + int64_t rcv; + int64_t snd; + snd_cr = &frcti->snd_cr; + rcv_cr = &frcti->rcv_cr; + + /* Idempotent; usually already sent by frcti_lingering. */ + frcti_fin_snd(frcti); clock_gettime(PTHREAD_COND_CLOCK, &now); + now_ns = TS_TO_UINT64(now); - pci = (struct frct_pci *) ssm_pk_buff_head_release(spb, FRCT_PCILEN); + pthread_rwlock_rdlock(&frcti->lock); - idx = ssm_pk_buff_get_idx(spb); - seqno = ntoh32(pci->seqno); - pos = seqno & (RQ_SIZE - 1); + ackno = rcv_cr->lwe; + rcv = (int64_t)(rcv_cr->act + rcv_cr->inact) - now_ns; + snd = (int64_t)(snd_cr->act + snd_cr->inact) - now_ns; + due = final_ack_due(frcti, rcv_cr, now_ns); - pthread_rwlock_wrlock(&frcti->lock); + pthread_rwlock_unlock(&frcti->lock); - if (now.tv_sec - rcv_cr->act.tv_sec > rcv_cr->inact) { - if (pci->flags & FRCT_DRF) { /* New run. */ - rcv_cr->lwe = seqno; - rcv_cr->rwe = seqno + RQ_SIZE; - rcv_cr->seqno = seqno; - } else if (pci->flags & FRCT_DATA) { - goto drop_packet; - } - } + if (due) + frcti_pkt_snd(frcti, FRCT_ACK, ackno, 0); - rcv_cr->act = now; + return (time_t) MAX((MAX(rcv, snd) / BILLION), 0); +} - /* For now, just send an immediate window update. */ - if (pci->flags & FRCT_RDVS) { - fd = frcti->fd; - rwe = rcv_cr->rwe; - pthread_rwlock_unlock(&frcti->lock); +__attribute__((hot)) +static void frcti_rcv(struct frcti * frcti, + struct ssm_pk_buff * spb) +{ + ssize_t idx; + size_t pos; + struct frct_pci * pci; + struct timespec now; + uint64_t now_ns; + struct frct_cr * rcv_cr; + uint32_t seqno; + uint16_t flags; + buffer_t pkt; + struct pending pending = { 0 }; + bool in_order; + struct sack_args * sa = NULL; + bool send_sack = false; + + assert(frcti); - __send_frct_pkt(fd, FRCT_FC, 0, rwe); + rcv_cr = &frcti->rcv_cr; + + clock_gettime(PTHREAD_COND_CLOCK, &now); + now_ns = TS_TO_UINT64(now); - ssm_pool_remove(proc.pool, idx); + if (ssm_pk_buff_len(spb) < FRCT_PCILEN) { + frct_spb_release(spb); return; } - if (pci->flags & FRCT_ACK) { - ackno = ntoh32(pci->ackno); - if (after(ackno, frcti->snd_cr.lwe)) - frcti->snd_cr.lwe = ackno; + pci = FRCT_HDR_POP(spb, frct_pci); - if (frcti->probe && after(ackno, frcti->rttseq)) { -#ifdef PROC_FLOW_STATS - if (!(pci->flags & FRCT_DATA)) - frcti->n_dak++; -#endif - rtt_estimator(frcti, ts_diff_ns(&now, &frcti->t_probe)); - frcti->probe = false; - } + idx = ssm_pk_buff_get_off(spb); + seqno = ntoh32(pci->seqno); + pos = RQ_SLOT(seqno); + + flags = ntoh16(pci->flags); + + pkt.data = ssm_pk_buff_head(spb); + pkt.len = ssm_pk_buff_len(spb); + + if (flags & FRCT_RXM) + STAT_BUMP(frcti, rxm_rcv); + + /* Stateless / lock-free dispatches. spb released via ctrl_done. */ + if (flags & FRCT_KA) { + frcti_ka_rcv(frcti, pci, now_ns, flags); + goto ctrl_done; } - if (pci->flags & FRCT_FC) { - uint32_t rwe; + if (flags & FRCT_RTTP) { + frcti_rttp_rcv(frcti, pkt, now_ns); + goto ctrl_done; + } - rwe = ntoh32(*((uint32_t *)pci) & hton32(0x00FFFFFF)); - rwe |= snd_cr->rwe & 0xFF000000; + if (flags & FRCT_NACK) { + frcti_nack_rcv(frcti); + goto ctrl_done; + } + + if (flags & FRCT_RDVS) { + frcti_rdv_rcv(frcti); + goto ctrl_done; + } - /* Rollover for 24 bit */ - if (before(rwe, snd_cr->rwe) && snd_cr->rwe - rwe > 0x007FFFFF) - rwe += 0x01000000; + /* Msg-mode FIN: control packet, FIN seqno carried in pci->ackno. */ + if ((flags & FRCT_FIN) && !(flags & FRCT_DATA)) { + pthread_rwlock_wrlock(&frcti->lock); + if (!frcti->rcv_fin_seen) { + frcti->rcv_fin_seen = true; + frcti->rcv_byte_fin = ntoh32(pci->ackno); + } + pthread_rwlock_unlock(&frcti->lock); + goto ctrl_done; + } - snd_cr->rwe = rwe; + pthread_rwlock_wrlock(&frcti->lock); - pthread_mutex_lock(&frcti->mtx); - if (!frcti->open) { - frcti->open = true; - pthread_cond_broadcast(&frcti->cond); + /* rcv_inact_check is a no-op for non-DATA non-DRF packets. */ + if (flags & (FRCT_DATA | FRCT_DRF)) { + switch (rcv_inact_check(frcti, flags, seqno, now_ns)) { + case FRCT_INACT_NEED_NACK: + pthread_rwlock_unlock(&frcti->lock); + frcti_nack_snd(frcti, seqno - 1); + frct_spb_release(spb); + return; + case FRCT_INACT_DROP: + STAT_BUMP(frcti, inact_drop); + goto drop_packet; + case FRCT_ACTIVE: + /* FALLTHRU */ + default: + break; } - pthread_mutex_unlock(&frcti->mtx); } - if (!(pci->flags & FRCT_DATA)) + /* DATA-only act refresh: non-DATA would lock out DRF rebase. */ + if (flags & FRCT_DATA) + STORE_RELEASE(&rcv_cr->act, now_ns); + + /* Wire-dup ACK packet: same seqno as the previous emission. */ + if (is_dup_ack(frcti, flags, seqno)) { + STAT_BUMP(frcti, ack_dup_rcv); + goto drop_packet; + } + + /* Wire-dup of DATA: piggybacked ACK info already processed. */ + if (is_dup_data(flags, seqno, rcv_cr->lwe)) { + rcv_cr->seqno = seqno; + STAT_BUMP(frcti, dup_rcv); + /* RFC 2883 §4 case 1: dup below cum-ACK. */ + frcti->dsack_seqno = seqno; + frcti->dsack_valid = true; + goto drop_packet; + } + + if (flags & FRCT_ACK) + frcti_ack_rcv(frcti, pci, flags, now_ns, &pending); + + if (flags & FRCT_SACK) + frcti_sack_rcv(frcti, pkt, ntoh32(pci->ackno), + now_ns, &pending); + + if (flags & FRCT_FC) + frcti_fc_rcv(frcti, pci); + + if (!(flags & FRCT_DATA)) goto drop_packet; if (before(seqno, rcv_cr->lwe)) { - rcv_cr->seqno = seqno; /* Ensures we send a new ACK. */ -#ifdef PROC_FLOW_STATS - frcti->n_dup++; -#endif + /* Bump rcv_cr.seqno to force ack_snd to fire on the dup. */ + rcv_cr->seqno = seqno; + if (flags & FRCT_RXM) + STAT_BUMP(frcti, rxm_dup_rcv); + else + STAT_BUMP(frcti, dup_rcv); + /* RFC 2883 §4 case 1: dup below cum-ACK. */ + frcti->dsack_seqno = seqno; + frcti->dsack_valid = true; goto drop_packet; } - if (rcv_cr->cflags & FRCTFRTX) { + if (!rq_accept(frcti, seqno, pos, flags)) + goto drop_packet; - if (!before(seqno, rcv_cr->rwe)) { /* Out of window. */ -#ifdef PROC_FLOW_STATS - frcti->n_out++; -#endif + if (frcti->stream) { + if (frcti_stream_data_rcv(frcti, spb, pos, flags) < 0) { + STAT_BUMP(frcti, strm_drop); goto drop_packet; } - - if (!before(seqno, rcv_cr->lwe + RQ_SIZE)) { -#ifdef PROC_FLOW_STATS - frcti->n_rqo++; -#endif - goto drop_packet; /* Out of rq. */ - } - if (frcti->rq[pos] != -1) { -#ifdef PROC_FLOW_STATS - frcti->n_dup++; -#endif - goto drop_packet; /* Duplicate in rq. */ - } - fd = frcti->fd; + /* spb consumed by stash; do not release in drop path. */ + spb = NULL; } else { - rcv_cr->lwe = seqno; + frcti_data_stash(frcti, idx, pos, flags); + } + + /* Lazy alloc: only OOO arrivals can trigger a SACK send. */ + if (after(seqno, rcv_cr->lwe) && frcti->sack_n_max > 0) { + size_t sa_sz = sizeof(*sa) + + frcti->sack_n_max * sizeof(sa->blocks[0]); + sa = malloc(sa_sz); + /* If alloc fails, sack_check sees NULL and we skip SACK. */ } - frcti->rq[pos] = idx; + send_sack = sa != NULL && sack_check(frcti, seqno, now_ns, sa); + in_order = !after(seqno, rcv_cr->lwe); pthread_rwlock_unlock(&frcti->lock); - if (fd != -1) - timerwheel_delayed_ack(fd, frcti); + if (send_sack) { + STAT_BUMP(frcti, sack_snd); + if (sa->dsack) + STAT_BUMP(frcti, dsack_snd); + frcti_sack_snd(frcti, sa); + } else if (in_order) { + ack_arm(frcti); + } + + if ((flags & FRCT_ACK) && frcti->snd_cr.seqno != frcti->snd_cr.lwe) + tlp_arm(frcti); + + pending_flush(frcti, &pending); + + frcti_rcv_probe(frcti, now_ns); + free(sa); + return; + + ctrl_done: + frct_spb_release(spb); return; drop_packet: pthread_rwlock_unlock(&frcti->lock); - ssm_pool_remove(proc.pool, idx); - send_frct_pkt(frcti); - return; + frct_spb_release(spb); + /* with_sack=true: ack_snd no-ops if neither dsack nor SACK is due. */ + ack_snd(frcti, true); + + pending_flush(frcti, &pending); + free(sa); } + +/* NULL-shim macros for the no-FRCT case. */ + +#define FRCTI_SND(frcti, spb, flags) \ + ((frcti) == NULL ? 0 : frcti_snd((frcti), (spb), (flags))) + +#define FRCTI_RCV(frcti, spb) \ + do { \ + if ((frcti) != NULL) \ + frcti_rcv((frcti), (spb)); \ + } while (0) + +#define FRCTI_PDU_READY(frcti) \ + ((frcti) != NULL && frcti_pdu_ready(frcti)) + +#define FRCTI_CONSUME(frcti, buf, count) \ + ((frcti) == NULL ? (ssize_t) -EAGAIN \ + : (frcti)->stream \ + ? frcti_consume_stream((frcti), (buf), (count)) \ + : frcti_consume((frcti), (buf), (count))) + +#define FRCTI_IS_FRTX(frcti) \ + ((frcti) != NULL && ((frcti)->rcv_cr.cflags & FRCTFRTX)) + +#define FRCTI_IS_STREAM(frcti) ((frcti) != NULL && (frcti)->stream) + +#define FRCTI_PAYLOAD_CAP(frcti) \ + ((frcti)->frag_mtu - frcti_data_hdr_len(frcti)) + +#define FRCTI_NEEDS_FRAG(frcti, count) \ + ((frcti) != NULL && (count) > FRCTI_PAYLOAD_CAP(frcti)) + +#define FRCTI_IS_WINDOW_OPEN(frcti) \ + ((frcti) == NULL ? true : frcti_is_window_open(frcti)) + +#define FRCTI_IS_WINDOW_OPEN_N(frcti, n) \ + ((frcti) == NULL ? true : frcti_is_window_open_n((frcti), (n))) + +#define FRCTI_LINGERING(frcti) \ + ((frcti) == NULL ? false : frcti_lingering(frcti)) + +#define FRCTI_DEALLOC(frcti) \ + ((frcti) == NULL ? (time_t) 0 : frcti_dealloc(frcti)) + diff --git a/src/lib/hash.c b/src/lib/hash.c index 7adee968..903474df 100644 --- a/src/lib/hash.c +++ b/src/lib/hash.c @@ -39,6 +39,9 @@ #include <ouroboros/md5.h> #include <ouroboros/sha3.h> #endif +#include <ouroboros/crc8.h> +#include <ouroboros/crc16.h> +#include <ouroboros/crc64.h> #include <string.h> #include <assert.h> #include <stdbool.h> @@ -69,6 +72,14 @@ int hash_len_tbl [] = { uint16_t hash_len(enum hash_algo algo) { + if (algo == HASH_CRC8) + return CRC8_HASH_LEN; + + if (algo == HASH_CRC16) + return CRC16_HASH_LEN; + + if (algo == HASH_CRC64) + return CRC64_HASH_LEN; #ifdef HAVE_LIBGCRYPT return (uint16_t) gcry_md_get_algo_dlen(gcry_algo_tbl[algo]); #else @@ -81,12 +92,36 @@ void mem_hash(enum hash_algo algo, const uint8_t * buf, size_t len) { -#ifdef HAVE_LIBGCRYPT - gcry_md_hash_buffer(gcry_algo_tbl[algo], dst, buf, len); -#else +#ifndef HAVE_LIBGCRYPT struct sha3_ctx sha3_ctx; struct md5_ctx md5_ctx; +#endif + if (algo == HASH_CRC8) { + uint8_t crc = 0; + + crc8_autosar(&crc, buf, len); + *(uint8_t *) dst = crc; + return; + } + if (algo == HASH_CRC16) { + uint16_t crc = 0; + + crc16_ccitt_false(&crc, buf, len); + *(uint16_t *) dst = htobe16(crc); + return; + } + + if (algo == HASH_CRC64) { + uint64_t crc = 0; + + crc64_nvme(&crc, buf, len); + *(uint64_t *) dst = htobe64(crc); + return; + } +#ifdef HAVE_LIBGCRYPT + gcry_md_hash_buffer(gcry_algo_tbl[algo], dst, buf, len); +#else switch (algo) { case HASH_CRC32: memset(dst, 0, CRC32_HASH_LEN); @@ -131,3 +166,14 @@ void str_hash(enum hash_algo algo, { return mem_hash(algo, dst, (const uint8_t *) str, strlen(str)); } + +uint64_t hash_mix64(uint64_t key) +{ + key ^= key >> 33; + key *= 0xff51afd7ed558ccdULL; + key ^= key >> 33; + key *= 0xc4ceb9fe1a85ec53ULL; + key ^= key >> 33; + + return key; +} diff --git a/src/lib/irm.c b/src/lib/irm.c index 594014f7..c62701aa 100644 --- a/src/lib/irm.c +++ b/src/lib/irm.c @@ -614,7 +614,7 @@ ssize_t irm_list_names(struct name_info ** names) return 0; } - *names = malloc(nr * sizeof(**names)); + *names = calloc(nr, sizeof(**names)); if (*names == NULL) { irm_msg__free_unpacked(recv_msg, NULL); return -ENOMEM; diff --git a/src/lib/pb/ipcp.proto b/src/lib/pb/ipcp.proto index 9dc402f5..afee4f91 100644 --- a/src/lib/pb/ipcp.proto +++ b/src/lib/pb/ipcp.proto @@ -39,6 +39,7 @@ enum ipcp_msg_code { IPCP_CONNECT = 10; IPCP_DISCONNECT = 11; IPCP_REPLY = 12; + IPCP_FLOW_UPDATE = 13; } message ipcp_msg { @@ -54,7 +55,7 @@ message ipcp_msg { optional int32 response = 10; optional string comp = 11; optional uint32 timeo_sec = 12; - optional sint32 mpl = 13; + optional sint32 mpl = 13; /* MPL in ms. */ optional int32 result = 14; optional uint32 uid = 15; /* 0 = GSPP, >0 = PUP uid */ } diff --git a/src/lib/pb/irm.proto b/src/lib/pb/irm.proto index 9ed0a29b..f54bc9ea 100644 --- a/src/lib/pb/irm.proto +++ b/src/lib/pb/irm.proto @@ -53,6 +53,8 @@ enum irm_msg_code { IPCP_FLOW_REQ_ARR = 25; IPCP_FLOW_ALLOC_REPLY = 26; IRM_REPLY = 27; + IRM_FLOW_UPDATE = 28; + IPCP_FLOW_UPDATE_ARR = 29; } message timespec_msg { @@ -88,12 +90,15 @@ message irm_msg { repeated ipcp_list_msg ipcps = 17; repeated name_info_msg names = 18; optional timespec_msg timeo = 19; - optional sint32 mpl = 20; + optional sint32 mpl = 20; /* MPL in ms. */ optional string comp = 21; optional bytes pk = 22; /* piggyback */ optional uint32 timeo_sec = 23; optional uint32 timeo_nsec = 24; optional sint32 result = 25; - optional bytes sym_key = 26; /* symmetric encryption key */ - optional sint32 cipher_nid = 27; /* cipher NID */ + optional bytes sym_key = 26; /* symmetric encryption key */ + optional sint32 cipher_nid = 27; /* cipher NID */ + optional uint32 generation = 28; /* re-key batch generation */ + optional bool rekey = 29; /* re-key watermark trigger */ + optional bool rk_initiator = 30; /* re-key proof-holder side */ } diff --git a/src/lib/pb/model.proto b/src/lib/pb/model.proto index f1382f3d..4c1564a5 100644 --- a/src/lib/pb/model.proto +++ b/src/lib/pb/model.proto @@ -28,7 +28,7 @@ message qosspec_msg { required uint32 availability = 3; /* Class of 9s. */ required uint32 loss = 4; /* Packet loss. */ required uint32 ber = 5; /* Bit error rate, ppb. */ - required uint32 in_order = 6; /* In-order delivery. */ + required uint32 service = 6; /* enum qos_service. */ required uint32 max_gap = 7; /* In ms. */ required uint32 timeout = 8; /* Timeout in ms. */ } @@ -37,10 +37,11 @@ message flow_info_msg { required uint32 id = 1; required uint32 n_pid = 2; required uint32 n_1_pid = 3; - required uint32 mpl = 4; + required uint32 mpl = 4; /* MPL in ms. */ required uint32 state = 5; required qosspec_msg qos = 6; required uint32 uid = 7; + required uint32 mtu = 8; /* Layer MTU (bytes). */ } message name_info_msg { diff --git a/src/lib/protobuf.c b/src/lib/protobuf.c index d419a9f1..a824d357 100644 --- a/src/lib/protobuf.c +++ b/src/lib/protobuf.c @@ -81,6 +81,7 @@ flow_info_msg_t * flow_info_s_to_msg(const struct flow_info * s) msg->mpl = s->mpl; msg->state = s->state; msg->uid = s->uid; + msg->mtu = s->mtu; msg->qos = qos_spec_s_to_msg(&s->qs); if (msg->qos == NULL) goto fail_msg; @@ -107,6 +108,7 @@ struct flow_info flow_info_msg_to_s(const flow_info_msg_t * msg) s.mpl = msg->mpl; s.state = msg->state; s.uid = msg->uid; + s.mtu = msg->mtu; s.qs = qos_spec_msg_to_s(msg->qos); return s; @@ -137,7 +139,7 @@ name_info_msg_t * name_info_s_to_msg(const struct name_info * info) goto fail_msg; msg->ckey = strdup(info->c.key); - if (msg->skey == NULL) + if (msg->ckey == NULL) goto fail_msg; msg->ccrt = strdup(info->c.crt); @@ -161,6 +163,8 @@ struct name_info name_info_msg_to_s(const name_info_msg_t * msg) assert(msg != NULL); assert(strlen(msg->name) <= NAME_SIZE); + memset(&s, 0, sizeof(s)); + strcpy(s.name, msg->name); strcpy(s.s.key, msg->skey); strcpy(s.s.crt, msg->scrt); @@ -755,7 +759,7 @@ qosspec_msg_t * qos_spec_s_to_msg(const struct qos_spec * s) msg->availability = s->availability; msg->loss = s->loss; msg->ber = s->ber; - msg->in_order = s->in_order; + msg->service = s->service; msg->max_gap = s->max_gap; msg->timeout = s->timeout; @@ -773,7 +777,7 @@ struct qos_spec qos_spec_msg_to_s(const qosspec_msg_t * msg) s.availability = msg->availability; s.loss = msg->loss; s.ber = msg->ber; - s.in_order = msg->in_order; + s.service = msg->service; s.max_gap = msg->max_gap; s.timeout = msg->timeout; diff --git a/src/lib/qoscube.c b/src/lib/qoscube.c index 1eaa0d7c..5d7ae17d 100644 --- a/src/lib/qoscube.c +++ b/src/lib/qoscube.c @@ -29,15 +29,11 @@ qoscube_t qos_spec_to_cube(qosspec_t qs) { - if (qs.delay <= qos_voice.delay && - qs.bandwidth <= qos_voice.bandwidth && - qs.availability >= qos_voice.availability && - qs.max_gap <= qos_voice.max_gap) + if (qs.delay <= 50 && qs.bandwidth <= 100000 + && qs.availability >= 5 && qs.max_gap <= 50) return QOS_CUBE_VOICE; - else if (qs.delay <= qos_video.delay && - qs.bandwidth <= qos_video.bandwidth && - qs.availability >= qos_video.availability && - qs.max_gap <= qos_video.max_gap) + else if (qs.delay <= 100 && qs.availability >= 3 + && qs.max_gap <= 100) return QOS_CUBE_VIDEO; else return QOS_CUBE_BE; diff --git a/src/lib/random.c b/src/lib/random.c index 96315132..a132f470 100644 --- a/src/lib/random.c +++ b/src/lib/random.c @@ -28,6 +28,8 @@ #include <stdlib.h> #elif defined(HAVE_SYS_RANDOM) #include <sys/random.h> +#include <errno.h> +#include <stdint.h> #elif defined(HAVE_LIBGCRYPT) #include <gcrypt.h> #elif defined(HAVE_OPENSSL_RNG) @@ -42,13 +44,28 @@ int random_buffer(void * buf, arc4random_buf(buf, len); return 0; #elif defined(HAVE_SYS_RANDOM) - return getrandom(buf, len, GRND_NONBLOCK); + size_t off = 0; + ssize_t ret; + + while (off < len) { + ret = getrandom((uint8_t *) buf + off, len - off, + GRND_NONBLOCK); + if (ret < 0) { + if (errno == EINTR) + continue; + return -1; + } + off += (size_t) ret; + } + + return 0; #elif defined(HAVE_LIBGCRYPT) gcry_randomize(buf, len, GCRY_STRONG_RANDOM); return 0; #elif defined(HAVE_OPENSSL_RNG) - if (len > 0 && len < INT_MAX) - return RAND_bytes((unsigned char *) buf, (int) len); - return -1; + if (len == 0 || len >= INT_MAX) + return -1; + + return RAND_bytes((unsigned char *) buf, (int) len) == 1 ? 0 : -1; #endif } diff --git a/src/lib/rib.c b/src/lib/rib.c index a8d535c9..6e421397 100644 --- a/src/lib/rib.c +++ b/src/lib/rib.c @@ -112,14 +112,14 @@ static int rib_read(const char * path, (void) info; (void) offset; - pthread_rwlock_wrlock(&rib.lock); + pthread_rwlock_rdlock(&rib.lock); list_for_each(p, &rib.reg_comps) { struct reg_comp * r = list_entry(p, struct reg_comp, next); if (strcmp(comp, r->path) == 0) { - int ret = r->ops->read(path + 1, buf, size); + struct rib_ops * ops = r->ops; pthread_rwlock_unlock(&rib.lock); - return ret; + return ops->read(path + 1, buf, size); } } @@ -160,19 +160,25 @@ static int rib_readdir(const char * path, ssize_t len; ssize_t i; struct reg_comp * c; + struct rib_ops * ops; c = list_entry(p, struct reg_comp, next); if (strcmp(path + 1, c->path) != 0) continue; - assert(c->ops->readdir != NULL); + ops = c->ops; + + assert(ops->readdir != NULL); + + pthread_rwlock_unlock(&rib.lock); - len = c->ops->readdir(&dir_entries); + len = ops->readdir(&dir_entries); if (len < 0) - break; + return 0; for (i = 0; i < len; ++i) filler(buf, dir_entries[i], NULL, 0); freepp(char, dir_entries, len); + return 0; } } diff --git a/src/lib/serdes-irm.c b/src/lib/serdes-irm.c index 65f2c02d..1d9b4dec 100644 --- a/src/lib/serdes-irm.c +++ b/src/lib/serdes-irm.c @@ -174,6 +174,54 @@ int flow__irm_result_des(buffer_t * buf, else memset(sk->key, 0, SYMMKEYSZ); + sk->epoch = msg->has_generation ? (uint8_t) msg->generation : 0; + + if (msg->sym_key.len == SYMMKEYSZ) + crypt_secure_clear(msg->sym_key.data, msg->sym_key.len); + + irm_msg__free_unpacked(msg, NULL); + + return 0; + fail: + irm_msg__free_unpacked(msg, NULL); + fail_msg: + return err; +} + +int flow_rekey__irm_result_des(buffer_t * buf, + struct crypt_sk * sk, + bool * has_key, + bool * initiator) +{ + irm_msg_t * msg; + int err; + + msg = irm_msg__unpack(NULL, buf->len, buf->data); + if (msg == NULL) { + err = -EIRMD; + goto fail_msg; + } + + if (!msg->has_result) { + err = -EIRMD; + goto fail; + } + + if (msg->result < 0) { + err = msg->result; + goto fail; + } + + *has_key = msg->has_sym_key && msg->sym_key.len == SYMMKEYSZ; + if (*has_key) { + memcpy(sk->key, msg->sym_key.data, SYMMKEYSZ); + sk->nid = NID_undef; + sk->epoch = msg->has_generation ? + (uint8_t) msg->generation : 0; + *initiator = msg->has_rk_initiator && msg->rk_initiator; + crypt_secure_clear(msg->sym_key.data, msg->sym_key.len); + } + irm_msg__free_unpacked(msg, NULL); return 0; @@ -222,6 +270,44 @@ int flow_dealloc__irm_req_ser(buffer_t * buf, return -ENOMEM; } +int flow_update__irm_req_ser(buffer_t * buf, + const struct flow_info * flow, + bool rekey) +{ + irm_msg_t * msg; + size_t len; + + msg = malloc(sizeof(*msg)); + if (msg == NULL) + goto fail_malloc; + + irm_msg__init(msg); + + msg->code = IRM_MSG_CODE__IRM_FLOW_UPDATE; + msg->flow_info = flow_info_s_to_msg(flow); + if (msg->flow_info == NULL) + goto fail_msg; + + msg->has_rekey = true; + msg->rekey = rekey; + + len = irm_msg__get_packed_size(msg); + if (len == 0 || len > buf->len) + goto fail_msg; + + buf->len = len; + + irm_msg__pack(msg, buf->data); + irm_msg__free_unpacked(msg, NULL); + + return 0; + + fail_msg: + irm_msg__free_unpacked(msg, NULL); + fail_malloc: + return -ENOMEM; +} + int ipcp_flow_dealloc__irm_req_ser(buffer_t * buf, const struct flow_info * flow) { @@ -398,6 +484,56 @@ int ipcp_flow_req_arr__irm_req_ser(buffer_t * buf, return 0; fail_msg: + /* hash/pk are borrowed from the caller; detach before free. */ + msg->hash.len = 0; + msg->hash.data = NULL; + msg->pk.len = 0; + msg->pk.data = NULL; + irm_msg__free_unpacked(msg, NULL); + fail_malloc: + return -ENOMEM; +} + +int ipcp_flow_update_arr__irm_req_ser(buffer_t * buf, + const struct flow_info * flow, + const buffer_t * data) +{ + irm_msg_t * msg; + size_t len; + + msg = malloc(sizeof(*msg)); + if (msg == NULL) + goto fail_malloc; + + irm_msg__init(msg); + + msg->code = IRM_MSG_CODE__IPCP_FLOW_UPDATE_ARR; + msg->flow_info = flow_info_s_to_msg(flow); + if (msg->flow_info == NULL) + goto fail_msg; + + msg->has_pk = true; + msg->pk.len = data->len; + msg->pk.data = data->data; + + len = irm_msg__get_packed_size(msg); + if (len == 0 || len > buf->len) + goto fail_msg; + + buf->len = len; + + irm_msg__pack(msg, buf->data); + + /* Don't free data! */ + msg->pk.len = 0; + msg->pk.data = NULL; + irm_msg__free_unpacked(msg, NULL); + + return 0; + fail_msg: + /* pk.data is borrowed from the caller; detach before free. */ + msg->pk.len = 0; + msg->pk.data = NULL; irm_msg__free_unpacked(msg, NULL); fail_malloc: return -ENOMEM; diff --git a/src/lib/ssm/flow_set.c b/src/lib/ssm/flow_set.c index 73d0db55..2e33b408 100644 --- a/src/lib/ssm/flow_set.c +++ b/src/lib/ssm/flow_set.c @@ -58,9 +58,9 @@ #define QUEUESIZE ((SSM_RBUFF_SIZE) * sizeof(struct flowevent)) #define SSM_FSET_FILE_SIZE (SYS_MAX_FLOWS * sizeof(ssize_t) \ - + PROG_MAX_FQUEUES * sizeof(size_t) \ - + PROG_MAX_FQUEUES * sizeof(pthread_cond_t) \ - + PROG_MAX_FQUEUES * QUEUESIZE \ + + PROC_MAX_FQUEUES * sizeof(size_t) \ + + PROC_MAX_FQUEUES * sizeof(pthread_cond_t) \ + + PROC_MAX_FQUEUES * QUEUESIZE \ + sizeof(pthread_mutex_t)) #define fqueue_ptr(fs, idx) (fs->fqueues + (SSM_RBUFF_SIZE) * idx) @@ -104,10 +104,10 @@ static struct ssm_flow_set * flow_set_create(pid_t pid, set->mtable = shm_base; set->heads = (size_t *) (set->mtable + SYS_MAX_FLOWS); - set->conds = (pthread_cond_t *)(set->heads + PROG_MAX_FQUEUES); - set->fqueues = (struct flowevent *) (set->conds + PROG_MAX_FQUEUES); + set->conds = (pthread_cond_t *)(set->heads + PROC_MAX_FQUEUES); + set->fqueues = (struct flowevent *) (set->conds + PROC_MAX_FQUEUES); set->lock = (pthread_mutex_t *) - (set->fqueues + PROG_MAX_FQUEUES * (SSM_RBUFF_SIZE)); + (set->fqueues + PROC_MAX_FQUEUES * (SSM_RBUFF_SIZE)); return set; @@ -164,7 +164,7 @@ struct ssm_flow_set * ssm_flow_set_create(pid_t pid) if (pthread_condattr_setclock(&cattr, PTHREAD_COND_CLOCK)) goto fail_condattr_set; #endif - for (i = 0; i < PROG_MAX_FQUEUES; ++i) { + for (i = 0; i < PROC_MAX_FQUEUES; ++i) { set->heads[i] = 0; if (pthread_cond_init(&set->conds[i], &cattr)) goto fail_init; @@ -222,7 +222,7 @@ void ssm_flow_set_zero(struct ssm_flow_set * set, ssize_t i = 0; assert(set); - assert(idx < PROG_MAX_FQUEUES); + assert(idx < PROC_MAX_FQUEUES); pthread_mutex_lock(set->lock); @@ -242,7 +242,7 @@ int ssm_flow_set_add(struct ssm_flow_set * set, { assert(set); assert(!(flow_id < 0) && flow_id < SYS_MAX_FLOWS); - assert(idx < PROG_MAX_FQUEUES); + assert(idx < PROC_MAX_FQUEUES); pthread_mutex_lock(set->lock); @@ -264,7 +264,7 @@ void ssm_flow_set_del(struct ssm_flow_set * set, { assert(set); assert(!(flow_id < 0) && flow_id < SYS_MAX_FLOWS); - assert(idx < PROG_MAX_FQUEUES); + assert(idx < PROC_MAX_FQUEUES); pthread_mutex_lock(set->lock); @@ -282,7 +282,7 @@ int ssm_flow_set_has(struct ssm_flow_set * set, assert(set); assert(!(flow_id < 0) && flow_id < SYS_MAX_FLOWS); - assert(idx < PROG_MAX_FQUEUES); + assert(idx < PROC_MAX_FQUEUES); pthread_mutex_lock(set->lock); @@ -299,26 +299,34 @@ void ssm_flow_set_notify(struct ssm_flow_set * set, int event) { struct flowevent * e; + ssize_t idx; assert(set); assert(!(flow_id < 0) && flow_id < SYS_MAX_FLOWS); pthread_mutex_lock(set->lock); - if (set->mtable[flow_id] == -1) { + idx = set->mtable[flow_id]; + if (idx == -1) { pthread_mutex_unlock(set->lock); return; } - e = fqueue_ptr(set, set->mtable[flow_id]) + - set->heads[set->mtable[flow_id]]; + /* Ring full: drop redundant FLOW_PKT, reserve a slot for ctrl. */ + if (set->heads[idx] >= SSM_RBUFF_SIZE + || (event == FLOW_PKT && set->heads[idx] >= SSM_RBUFF_SIZE - 1)) { + pthread_mutex_unlock(set->lock); + return; + } + + e = fqueue_ptr(set, idx) + set->heads[idx]; e->flow_id = flow_id; e->event = event; - ++set->heads[set->mtable[flow_id]]; + ++set->heads[idx]; - pthread_cond_signal(&set->conds[set->mtable[flow_id]]); + pthread_cond_signal(&set->conds[idx]); pthread_mutex_unlock(set->lock); } @@ -332,7 +340,7 @@ ssize_t ssm_flow_set_wait(const struct ssm_flow_set * set, ssize_t ret = 0; assert(set); - assert(idx < PROG_MAX_FQUEUES); + assert(idx < PROC_MAX_FQUEUES); assert(fqueue); #ifndef HAVE_ROBUST_MUTEX diff --git a/src/lib/ssm/pool.c b/src/lib/ssm/pool.c index 5c98b515..705de147 100644 --- a/src/lib/ssm/pool.c +++ b/src/lib/ssm/pool.c @@ -24,6 +24,7 @@ #include "config.h" +#include <ouroboros/atomics.h> #include <ouroboros/errno.h> #include <ouroboros/pthread.h> #include <ouroboros/ssm_pool.h> @@ -37,10 +38,20 @@ #include <stdio.h> #include <stdlib.h> #include <string.h> +#include <time.h> #include <unistd.h> #include <sys/mman.h> #include <sys/stat.h> +static __inline__ uint64_t pool_now_ns(void) +{ + struct timespec ts; + + clock_gettime(CLOCK_MONOTONIC, &ts); + + return (uint64_t) ts.tv_sec * 1000000000ULL + (uint64_t) ts.tv_nsec; +} + /* Global Shared Packet Pool (GSPP) configuration */ static const struct ssm_size_class_cfg ssm_gspp_cfg[SSM_POOL_MAX_CLASSES] = { { (1 << 8), SSM_GSPP_256_BLOCKS }, @@ -75,26 +86,6 @@ static const struct ssm_size_class_cfg ssm_pup_cfg[SSM_POOL_MAX_CLASSES] = { #define GET_SHARD_FOR_PID(pid) ((int)((pid) % SSM_POOL_SHARDS)) -#define LOAD_RELAXED(ptr) \ - (__atomic_load_n(ptr, __ATOMIC_RELAXED)) - -#define LOAD_ACQUIRE(ptr) \ - (__atomic_load_n(ptr, __ATOMIC_ACQUIRE)) - -#define STORE_RELEASE(ptr, val) \ - (__atomic_store_n(ptr, val, __ATOMIC_RELEASE)) - -#define LOAD(ptr) \ - (__atomic_load_n(ptr, __ATOMIC_SEQ_CST)) - -#define STORE(ptr, val) \ - (__atomic_store_n(ptr, val, __ATOMIC_SEQ_CST)) - -#define FETCH_ADD(ptr, val) \ - (__atomic_fetch_add(ptr, val, __ATOMIC_SEQ_CST)) - -#define FETCH_SUB(ptr, val) \ - (__atomic_fetch_sub(ptr, val, __ATOMIC_SEQ_CST)) #define SSM_FILE_SIZE (SSM_POOL_TOTAL_SIZE + sizeof(struct _ssm_pool_hdr)) #define SSM_GSPP_FILE_SIZE (SSM_GSPP_TOTAL_SIZE + sizeof(struct _ssm_pool_hdr)) @@ -107,6 +98,8 @@ static const struct ssm_size_class_cfg ssm_pup_cfg[SSM_POOL_MAX_CLASSES] = { : SSM_PUP_FILE_SIZE) #define GET_POOL_CFG(uid) (IS_GSPP(uid) ? ssm_gspp_cfg : ssm_pup_cfg) +#define NEEDS_CHOWN(uid, gid) ((uid) != geteuid() || (gid) != getegid()) + struct ssm_pool { uint8_t * shm_base; /* start of blocks */ struct _ssm_pool_hdr * hdr; /* shared memory header */ @@ -163,29 +156,6 @@ static __inline__ void list_add_head(struct _ssm_list_head * head, STORE(&head->count, LOAD(&head->count) + 1); } -static __inline__ int select_size_class(struct ssm_pool * pool, - size_t len) -{ - size_t sz; - int i; - - assert(pool != NULL); - - /* Total space needed: header + headspace + data + tailspace */ - sz = sizeof(struct ssm_pk_buff) + SSM_PK_BUFF_HEADSPACE + len - + SSM_PK_BUFF_TAILSPACE; - - for (i = 0; i < SSM_POOL_MAX_CLASSES; i++) { - struct _ssm_size_class * sc; - - sc = &pool->hdr->size_classes[i]; - if (sc->object_size > 0 && sz <= sc->object_size) - return i; - } - - return -1; -} - static __inline__ int find_size_class_for_offset(struct ssm_pool * pool, size_t offset) { @@ -276,6 +246,7 @@ static void init_size_classes(struct ssm_pool * pool) STORE(&blk->refcount, 0); blk->allocator_pid = 0; + blk->alloc_ts = 0; STORE(&blk->next_offset, 0); list_add_head(&sc->shards[0].free_list, blk, @@ -306,19 +277,31 @@ static size_t reclaim_pid_from_sc(struct _ssm_size_class * sc, size_t i; size_t recovered = 0; struct ssm_pk_buff * blk; + uint64_t now; + uint64_t min_age_ns; - region = (uint8_t *) pool_base + sc->pool_start; + region = (uint8_t *) pool_base + sc->pool_start; + now = pool_now_ns(); + min_age_ns = (uint64_t) SSM_POOL_RECLAIM_AGE_S * 1000000000ULL; for (i = 0; i < sc->object_count; ++i) { blk = (struct ssm_pk_buff *)(region + i * sc->object_size); - if (blk->allocator_pid == pid && LOAD(&blk->refcount) > 0) { - STORE(&blk->refcount, 0); - blk->allocator_pid = 0; - list_add_head(&shard->free_list, blk, pool_base); - FETCH_ADD(&shard->free_count, 1); - recovered++; - } + if (blk->allocator_pid != pid) + continue; + + if (LOAD(&blk->refcount) == 0) + continue; + + /* Recent: a live consumer may still hold the handoff. */ + if (now - blk->alloc_ts < min_age_ns) + continue; + + STORE(&blk->refcount, 0); + blk->allocator_pid = 0; + list_add_head(&shard->free_list, blk, pool_base); + FETCH_ADD(&shard->free_count, 1); + recovered++; } return recovered; @@ -379,6 +362,7 @@ static __inline__ ssize_t init_block(struct ssm_pool * pool, { STORE(&blk->refcount, 1); blk->allocator_pid = getpid(); + blk->alloc_ts = pool_now_ns(); blk->size = (uint32_t) (sc->object_size - sizeof(struct ssm_pk_buff)); blk->pk_head = SSM_PK_BUFF_HEADSPACE; @@ -548,7 +532,7 @@ static struct ssm_pool * __pool_create(const char * name, if (flags & O_CREAT) { if (ftruncate(fd, (off_t) file_size) < 0) goto fail_truncate; - if (uid != geteuid() && fchown(fd, uid, gid) < 0) + if (NEEDS_CHOWN(uid, gid) && fchown(fd, uid, gid) < 0) goto fail_truncate; } @@ -700,7 +684,7 @@ ssize_t ssm_pool_alloc(struct ssm_pool * pool, assert(pool != NULL); assert(spb != NULL); - idx = select_size_class(pool, count); + idx = select_size_class(pool->hdr, count); if (idx >= 0) return alloc_from_sc(pool, idx, count, ptr, spb); @@ -718,7 +702,7 @@ ssize_t ssm_pool_alloc_b(struct ssm_pool * pool, assert(pool != NULL); assert(spb != NULL); - idx = select_size_class(pool, count); + idx = select_size_class(pool->hdr, count); if (idx >= 0) return alloc_from_sc_b(pool, idx, count, ptr, spb, abstime); @@ -744,7 +728,7 @@ ssize_t ssm_pool_read(uint8_t ** dst, } struct ssm_pk_buff * ssm_pool_get(struct ssm_pool * pool, - size_t off) + size_t off) { struct ssm_pk_buff * blk; @@ -823,36 +807,36 @@ int ssm_pool_remove(struct ssm_pool * pool, return 0; } -size_t ssm_pk_buff_get_idx(struct ssm_pk_buff * spb) +size_t ssm_pk_buff_get_off(const struct ssm_pk_buff * spb) { assert(spb != NULL); return spb->off; } -uint8_t * ssm_pk_buff_head(struct ssm_pk_buff * spb) +uint8_t * ssm_pk_buff_head(const struct ssm_pk_buff * spb) { assert(spb != NULL); - return spb->data + spb->pk_head; + return (uint8_t *) spb->data + spb->pk_head; } -uint8_t * ssm_pk_buff_tail(struct ssm_pk_buff * spb) +uint8_t * ssm_pk_buff_tail(const struct ssm_pk_buff * spb) { assert(spb != NULL); - return spb->data + spb->pk_tail; + return (uint8_t *) spb->data + spb->pk_tail; } -size_t ssm_pk_buff_len(struct ssm_pk_buff * spb) +size_t ssm_pk_buff_len(const struct ssm_pk_buff * spb) { assert(spb != NULL); return spb->pk_tail - spb->pk_head; } -uint8_t * ssm_pk_buff_head_alloc(struct ssm_pk_buff * spb, - size_t size) +uint8_t * ssm_pk_buff_push(struct ssm_pk_buff * spb, + size_t size) { assert(spb != NULL); @@ -864,8 +848,8 @@ uint8_t * ssm_pk_buff_head_alloc(struct ssm_pk_buff * spb, return spb->data + spb->pk_head; } -uint8_t * ssm_pk_buff_tail_alloc(struct ssm_pk_buff * spb, - size_t size) +uint8_t * ssm_pk_buff_push_tail(struct ssm_pk_buff * spb, + size_t size) { uint8_t * buf; @@ -881,8 +865,8 @@ uint8_t * ssm_pk_buff_tail_alloc(struct ssm_pk_buff * spb, return buf; } -uint8_t * ssm_pk_buff_head_release(struct ssm_pk_buff * spb, - size_t size) +uint8_t * ssm_pk_buff_pop(struct ssm_pk_buff * spb, + size_t size) { uint8_t * buf; @@ -896,8 +880,8 @@ uint8_t * ssm_pk_buff_head_release(struct ssm_pk_buff * spb, return buf; } -uint8_t * ssm_pk_buff_tail_release(struct ssm_pk_buff * spb, - size_t size) +uint8_t * ssm_pk_buff_pop_tail(struct ssm_pk_buff * spb, + size_t size) { assert(spb != NULL); assert(!(size > spb->pk_tail - spb->pk_head)); diff --git a/src/lib/ssm/rbuff.c b/src/lib/ssm/rbuff.c index e4558c31..7886b5c2 100644 --- a/src/lib/ssm/rbuff.c +++ b/src/lib/ssm/rbuff.c @@ -74,12 +74,13 @@ struct ssm_rbuff { ssize_t * shm_base; /* start of shared memory */ size_t * head; /* start of ringbuffer */ size_t * tail; - size_t * acl; /* access control */ + size_t * flags; /* out-of-band flags (RB_*) */ pthread_mutex_t * mtx; /* lock for cond vars only */ pthread_cond_t * add; /* signal when new data */ pthread_cond_t * del; /* signal when data removed */ pid_t pid; /* pid of the owner */ int flow_id; /* flow_id of the flow */ + size_t n_users; /* in-flight users */ }; #define MM_FLAGS (PROT_READ | PROT_WRITE) @@ -113,12 +114,13 @@ static struct ssm_rbuff * rbuff_create(pid_t pid, rb->shm_base = shm_base; rb->head = (size_t *) (rb->shm_base + (SSM_RBUFF_SIZE)); rb->tail = (size_t *) (rb->head + 1); - rb->acl = (size_t *) (rb->tail + 1); - rb->mtx = (pthread_mutex_t *) (rb->acl + 1); + rb->flags = (size_t *) (rb->tail + 1); + rb->mtx = (pthread_mutex_t *) (rb->flags + 1); rb->add = (pthread_cond_t *) (rb->mtx + 1); rb->del = rb->add + 1; rb->pid = pid; rb->flow_id = flow_id; + rb->n_users = 0; return rb; @@ -179,7 +181,7 @@ struct ssm_rbuff * ssm_rbuff_create(pid_t pid, if (pthread_cond_init(rb->del, &cattr)) goto fail_del; - *rb->acl = ACL_RDWR; + *rb->flags = RB_RDWR; *rb->head = 0; *rb->tail = 0; @@ -228,27 +230,47 @@ void ssm_rbuff_close(struct ssm_rbuff * rb) { assert(rb); + /* + * Caller must set RB_FLOWDOWN first; if a user becomes + * cancellable, push a cleanup that decrements n_users. + */ + while (__atomic_load_n(&rb->n_users, __ATOMIC_SEQ_CST) > 0) { + struct timespec tic = { 0, 100000 }; + nanosleep(&tic, NULL); + } + rbuff_destroy(rb); } +/* Cancel cleanup for a blocked reader: unlock mtx AND drop the n_users ref. */ +static void __cleanup_rbuff_reader(void * o) +{ + struct ssm_rbuff * rb = (struct ssm_rbuff *) o; + + pthread_mutex_unlock(rb->mtx); + __atomic_fetch_sub(&rb->n_users, 1, __ATOMIC_SEQ_CST); +} + int ssm_rbuff_write(struct ssm_rbuff * rb, - size_t idx) + size_t off) { - size_t acl; + size_t flags; bool was_empty; int ret = 0; assert(rb != NULL); - acl = __atomic_load_n(rb->acl, __ATOMIC_SEQ_CST); - if (acl != ACL_RDWR) { - if (acl & ACL_FLOWDOWN) { + __atomic_fetch_add(&rb->n_users, 1, __ATOMIC_SEQ_CST); + + flags = __atomic_load_n(rb->flags, __ATOMIC_SEQ_CST); + if (flags != RB_RDWR) { + if (flags & RB_FLOWDOWN) { ret = -EFLOWDOWN; - goto fail_acl; + goto fail_flags; } - if (acl & ACL_RDONLY) { + if (!(flags & RB_WR)) { ret = -ENOTALLOC; - goto fail_acl; + goto fail_flags; } } @@ -261,7 +283,7 @@ int ssm_rbuff_write(struct ssm_rbuff * rb, was_empty = IS_EMPTY(rb); - HEAD(rb) = (ssize_t) idx; + HEAD(rb) = (ssize_t) off; ADVANCE_HEAD(rb); if (was_empty) @@ -269,43 +291,47 @@ int ssm_rbuff_write(struct ssm_rbuff * rb, pthread_mutex_unlock(rb->mtx); + __atomic_fetch_sub(&rb->n_users, 1, __ATOMIC_SEQ_CST); return 0; fail_mutex: pthread_mutex_unlock(rb->mtx); - fail_acl: + fail_flags: + __atomic_fetch_sub(&rb->n_users, 1, __ATOMIC_SEQ_CST); return ret; } int ssm_rbuff_write_b(struct ssm_rbuff * rb, - size_t idx, + size_t off, const struct timespec * abstime) { - size_t acl; + size_t flags; int ret = 0; bool was_empty; assert(rb != NULL); - acl = __atomic_load_n(rb->acl, __ATOMIC_SEQ_CST); - if (acl != ACL_RDWR) { - if (acl & ACL_FLOWDOWN) { + __atomic_fetch_add(&rb->n_users, 1, __ATOMIC_SEQ_CST); + + flags = __atomic_load_n(rb->flags, __ATOMIC_SEQ_CST); + if (flags != RB_RDWR) { + if (flags & RB_FLOWDOWN) { ret = -EFLOWDOWN; - goto fail_acl; + goto fail_flags; } - if (acl & ACL_RDONLY) { + if (!(flags & RB_WR)) { ret = -ENOTALLOC; - goto fail_acl; + goto fail_flags; } } robust_mutex_lock(rb->mtx); - pthread_cleanup_push(__cleanup_mutex_unlock, rb->mtx); + pthread_cleanup_push(__cleanup_rbuff_reader, rb); while (IS_FULL(rb) && ret != -ETIMEDOUT) { - acl = __atomic_load_n(rb->acl, __ATOMIC_SEQ_CST); - if (acl & ACL_FLOWDOWN) { + flags = __atomic_load_n(rb->flags, __ATOMIC_SEQ_CST); + if (flags & RB_FLOWDOWN) { ret = -EFLOWDOWN; break; } @@ -316,7 +342,7 @@ int ssm_rbuff_write_b(struct ssm_rbuff * rb, if (ret != -ETIMEDOUT && ret != -EFLOWDOWN) { was_empty = IS_EMPTY(rb); - HEAD(rb) = (ssize_t) idx; + HEAD(rb) = (ssize_t) off; ADVANCE_HEAD(rb); if (was_empty) pthread_cond_broadcast(rb->add); @@ -324,24 +350,28 @@ int ssm_rbuff_write_b(struct ssm_rbuff * rb, pthread_mutex_unlock(rb->mtx); - fail_acl: + fail_flags: + __atomic_fetch_sub(&rb->n_users, 1, __ATOMIC_SEQ_CST); return ret; } -static int check_rb_acl(struct ssm_rbuff * rb) +static int check_rb_flags(struct ssm_rbuff * rb) { - size_t acl; + size_t flags; assert(rb != NULL); - acl = __atomic_load_n(rb->acl, __ATOMIC_SEQ_CST); + flags = __atomic_load_n(rb->flags, __ATOMIC_SEQ_CST); - if (acl & ACL_FLOWDOWN) + if (flags & RB_FLOWDOWN) return -EFLOWDOWN; - if (acl & ACL_FLOWPEER) + if (flags & RB_FLOWPEER) return -EFLOWPEER; + if (!(flags & RB_RD)) + return -ENOTALLOC; + return -EAGAIN; } @@ -351,11 +381,21 @@ ssize_t ssm_rbuff_read(struct ssm_rbuff * rb) assert(rb != NULL); - if (IS_EMPTY(rb)) - return check_rb_acl(rb); + __atomic_fetch_add(&rb->n_users, 1, __ATOMIC_SEQ_CST); + + if (IS_EMPTY(rb)) { + ret = check_rb_flags(rb); + goto out; + } robust_mutex_lock(rb->mtx); + if (IS_EMPTY(rb)) { + pthread_mutex_unlock(rb->mtx); + ret = check_rb_flags(rb); + goto out; + } + ret = TAIL(rb); ADVANCE_TAIL(rb); @@ -363,6 +403,8 @@ ssize_t ssm_rbuff_read(struct ssm_rbuff * rb) pthread_mutex_unlock(rb->mtx); + out: + __atomic_fetch_sub(&rb->n_users, 1, __ATOMIC_SEQ_CST); return ret; } @@ -370,21 +412,25 @@ ssize_t ssm_rbuff_read_b(struct ssm_rbuff * rb, const struct timespec * abstime) { ssize_t idx = -1; - size_t acl; + size_t flags; assert(rb != NULL); - acl = __atomic_load_n(rb->acl, __ATOMIC_SEQ_CST); - if (IS_EMPTY(rb) && (acl & ACL_FLOWDOWN)) - return -EFLOWDOWN; + __atomic_fetch_add(&rb->n_users, 1, __ATOMIC_SEQ_CST); + + flags = __atomic_load_n(rb->flags, __ATOMIC_SEQ_CST); + if (IS_EMPTY(rb) && (flags & RB_FLOWDOWN)) { + idx = -EFLOWDOWN; + goto out; + } robust_mutex_lock(rb->mtx); - pthread_cleanup_push(__cleanup_mutex_unlock, rb->mtx); + pthread_cleanup_push(__cleanup_rbuff_reader, rb); while (IS_EMPTY(rb) && idx != -ETIMEDOUT && - check_rb_acl(rb) == -EAGAIN) { + check_rb_flags(rb) == -EAGAIN) { idx = -robust_wait(rb->add, rb->mtx, abstime); } @@ -395,43 +441,67 @@ ssize_t ssm_rbuff_read_b(struct ssm_rbuff * rb, ADVANCE_TAIL(rb); pthread_cond_broadcast(rb->del); } else if (idx != -ETIMEDOUT) { - idx = check_rb_acl(rb); + idx = check_rb_flags(rb); } pthread_mutex_unlock(rb->mtx); assert(idx != -EAGAIN); + out: + __atomic_fetch_sub(&rb->n_users, 1, __ATOMIC_SEQ_CST); return idx; } -void ssm_rbuff_set_acl(struct ssm_rbuff * rb, - uint32_t flags) +void ssm_rbuff_set_bits(struct ssm_rbuff * rb, + uint32_t bits) { assert(rb != NULL); - __atomic_store_n(rb->acl, (size_t) flags, __ATOMIC_SEQ_CST); + robust_mutex_lock(rb->mtx); + __atomic_fetch_or(rb->flags, (size_t) bits, __ATOMIC_SEQ_CST); + pthread_cond_broadcast(rb->add); + pthread_cond_broadcast(rb->del); + pthread_mutex_unlock(rb->mtx); +} + +void ssm_rbuff_clr_bits(struct ssm_rbuff * rb, + uint32_t bits) +{ + assert(rb != NULL); + + robust_mutex_lock(rb->mtx); + __atomic_fetch_and(rb->flags, ~(size_t) bits, __ATOMIC_SEQ_CST); + pthread_cond_broadcast(rb->add); + pthread_cond_broadcast(rb->del); + pthread_mutex_unlock(rb->mtx); } -uint32_t ssm_rbuff_get_acl(struct ssm_rbuff * rb) +uint32_t ssm_rbuff_get_flags(struct ssm_rbuff * rb) { assert(rb != NULL); - return (uint32_t) __atomic_load_n(rb->acl, __ATOMIC_SEQ_CST); + return (uint32_t) __atomic_load_n(rb->flags, __ATOMIC_SEQ_CST); } void ssm_rbuff_fini(struct ssm_rbuff * rb) { assert(rb != NULL); + __atomic_fetch_add(&rb->n_users, 1, __ATOMIC_SEQ_CST); + robust_mutex_lock(rb->mtx); - pthread_cleanup_push(__cleanup_mutex_unlock, rb->mtx); + pthread_cleanup_push(__cleanup_rbuff_reader, rb); while (!IS_EMPTY(rb)) robust_wait(rb->del, rb->mtx, NULL); - pthread_cleanup_pop(true); + pthread_cleanup_pop(false); + + pthread_mutex_unlock(rb->mtx); + + __atomic_fetch_sub(&rb->n_users, 1, __ATOMIC_SEQ_CST); } size_t ssm_rbuff_queued(struct ssm_rbuff * rb) diff --git a/src/lib/ssm/ssm.h.in b/src/lib/ssm/ssm.h.in index b9246c8b..57febae4 100644 --- a/src/lib/ssm/ssm.h.in +++ b/src/lib/ssm/ssm.h.in @@ -38,7 +38,6 @@ #define SSM_RBUFF_PREFIX "@SSM_RBUFF_PREFIX@" #define SSM_FLOW_SET_PREFIX "@SSM_FLOW_SET_PREFIX@" #define SSM_POOL_NAME "@SSM_POOL_NAME@" -#define SSM_POOL_BLOCKS @SSM_POOL_BLOCKS@ #define SSM_RBUFF_SIZE @SSM_RBUFF_SIZE@ /* Packet buffer space reservation */ @@ -84,6 +83,7 @@ /* Size class configuration */ #define SSM_POOL_MAX_CLASSES 9 #define SSM_POOL_SHARDS @SSM_POOL_SHARDS@ +#define SSM_POOL_RECLAIM_AGE_S @SSM_POOL_RECLAIM_AGE_S@ /* Internal structures - exposed for testing */ #ifdef __cplusplus @@ -126,6 +126,7 @@ struct ssm_pk_buff { uint32_t pk_head; /* Head offset into data */ uint32_t pk_tail; /* Tail offset into data */ uint32_t off; /* Block offset in pool */ + uint64_t alloc_ts; /* CLOCK_MONOTONIC ns at alloc */ uint8_t data[]; /* Packet data */ }; @@ -164,6 +165,24 @@ struct _ssm_pool_hdr { struct _ssm_size_class size_classes[SSM_POOL_MAX_CLASSES]; }; +#define SSM_PK_BUFF_TOTALSPACE (SSM_PK_BUFF_HEADSPACE + SSM_PK_BUFF_TAILSPACE) +static __inline__ int select_size_class(struct _ssm_pool_hdr * hdr, + size_t len) +{ + size_t sz; + int i; + + sz = sizeof(struct ssm_pk_buff) + SSM_PK_BUFF_TOTALSPACE + len; + + for (i = 0; i < SSM_POOL_MAX_CLASSES; i++) { + struct _ssm_size_class * sc = &hdr->size_classes[i]; + if (sc->object_size > 0 && sz <= sc->object_size) + return i; + } + + return -1; +} + #ifdef __cplusplus } #endif diff --git a/src/lib/ssm/tests/pool_sharding_test.c b/src/lib/ssm/tests/pool_sharding_test.c index c53105e3..ec464a92 100644 --- a/src/lib/ssm/tests/pool_sharding_test.c +++ b/src/lib/ssm/tests/pool_sharding_test.c @@ -80,19 +80,13 @@ static int test_lazy_distribution(void) goto fail_pool; } - /* Find the first size class with blocks */ - sc_idx = -1; - for (i = 0; i < SSM_POOL_MAX_CLASSES; i++) { - if (hdr->size_classes[i].object_count > 0) { - sc_idx = i; - break; - } - } - + /* Inspect the class that TEST_SIZE allocations will use */ + sc_idx = select_size_class(hdr, TEST_SIZE); if (sc_idx < 0) { - printf("No size classes configured.\n"); + printf("No size class fits TEST_SIZE=%d.\n", TEST_SIZE); for (i = 0; i < SSM_POOL_MAX_CLASSES; i++) { - printf(" Class %d: count=%zu\n", i, + printf(" Class %d: object_size=%zu count=%zu\n", i, + hdr->size_classes[i].object_size, hdr->size_classes[i].object_count); } goto fail_pool; @@ -137,7 +131,6 @@ static int test_shard_migration(void) ssize_t off; int shard_idx; int sc_idx; - int i; TEST_START(); @@ -149,18 +142,11 @@ static int test_shard_migration(void) hdr = get_pool_hdr(pool); - /* Find the first size class with blocks */ - sc_idx = -1; - for (i = 0; i < SSM_POOL_MAX_CLASSES; i++) { - if (hdr->size_classes[i].object_count > 0) { - sc_idx = i; - break; - } - } - + /* Inspect the class that TEST_SIZE allocations will use */ + sc_idx = select_size_class(hdr, TEST_SIZE); if (sc_idx < 0) { - printf("No size classes configured.\n"); - goto fail; + printf("No size class fits TEST_SIZE=%d.\n", TEST_SIZE); + goto fail_pool; } sc = &hdr->size_classes[sc_idx]; @@ -209,7 +195,6 @@ static int test_fallback_stealing(void) size_t total_free; size_t i; int sc_idx; - int c; TEST_START(); @@ -221,18 +206,11 @@ static int test_fallback_stealing(void) hdr = get_pool_hdr(pool); - /* Find the first size class with blocks */ - sc_idx = -1; - for (c = 0; c < SSM_POOL_MAX_CLASSES; c++) { - if (hdr->size_classes[c].object_count > 0) { - sc_idx = c; - break; - } - } - + /* Inspect the class that TEST_SIZE allocations will use */ + sc_idx = select_size_class(hdr, TEST_SIZE); if (sc_idx < 0) { - printf("No size classes configured.\n"); - goto fail; + printf("No size class fits TEST_SIZE=%d.\n", TEST_SIZE); + goto fail_pool; } sc = &hdr->size_classes[sc_idx]; @@ -261,7 +239,7 @@ static int test_fallback_stealing(void) /* Free them all - they go to local_shard */ for (i = 0; i < total_blocks / 2; i++) { - size_t off = ssm_pk_buff_get_idx(spbs[i]); + size_t off = ssm_pk_buff_get_off(spbs[i]); if (ssm_pool_remove(pool, off) != 0) { printf("Remove %zu failed.\n", i); free(spbs); @@ -299,7 +277,7 @@ static int test_fallback_stealing(void) /* Now all allocated blocks are in use again */ /* Cleanup - free all allocated blocks */ for (i = 0; i < total_blocks / 2; i++) { - size_t off = ssm_pk_buff_get_idx(spbs[i]); + size_t off = ssm_pk_buff_get_off(spbs[i]); ssm_pool_remove(pool, off); } @@ -396,20 +374,15 @@ static int test_multiprocess_sharding(void) /* Verify blocks distributed across shards */ hdr = get_pool_hdr(pool); - /* Find the first size class with blocks */ - sc = NULL; - for (i = 0; i < SSM_POOL_MAX_CLASSES; i++) { - if (hdr->size_classes[i].object_count > 0) { - sc = &hdr->size_classes[i]; - break; - } - } - - if (sc == NULL) { - printf("No size classes configured.\n"); + /* Inspect the class that TEST_SIZE allocations used */ + i = select_size_class(hdr, TEST_SIZE); + if (i < 0) { + printf("No size class fits TEST_SIZE=%d.\n", TEST_SIZE); goto fail_pool; } + sc = &hdr->size_classes[i]; + /* After children allocate and free, blocks should be in shards * (though exact distribution depends on PID values) */ diff --git a/src/lib/ssm/tests/pool_test.c b/src/lib/ssm/tests/pool_test.c index 3fc19cd5..f86fbd9e 100644 --- a/src/lib/ssm/tests/pool_test.c +++ b/src/lib/ssm/tests/pool_test.c @@ -741,14 +741,14 @@ static int test_ssm_pk_buff_operations(void) memcpy(head, data, dlen); - tail = ssm_pk_buff_tail_alloc(spb, 32); + tail = ssm_pk_buff_push_tail(spb, 32); if (tail == NULL) { - printf("Tail_alloc failed.\n"); + printf("push_tail failed.\n"); goto fail_ops; } if (ssm_pk_buff_len(spb) != POOL_256 + 32) { - printf("Length after tail_alloc: %zu.\n", + printf("Length after push_tail: %zu.\n", ssm_pk_buff_len(spb)); goto fail_ops; } @@ -758,14 +758,14 @@ static int test_ssm_pk_buff_operations(void) goto fail_ops; } - tail = ssm_pk_buff_tail_release(spb, 32); + tail = ssm_pk_buff_pop_tail(spb, 32); if (tail == NULL) { - printf("Tail_release failed.\n"); + printf("pop_tail failed.\n"); goto fail_ops; } if (ssm_pk_buff_len(spb) != POOL_256) { - printf("Length after tail_release: %zu.\n", + printf("Length after pop_tail: %zu.\n", ssm_pk_buff_len(spb)); goto fail_ops; } @@ -956,6 +956,8 @@ static int test_ssm_pool_reclaim_orphans(void) ssize_t ret3; pid_t my_pid; pid_t fake_pid = 99999; + struct timespec now; + uint64_t old_ts; TEST_START(); @@ -976,9 +978,15 @@ static int test_ssm_pool_reclaim_orphans(void) goto fail_alloc; } - /* Simulate blocks from another process by changing allocator_pid */ + /* Simulate blocks leaked by a dead process: foreign pid, aged out. */ + clock_gettime(CLOCK_MONOTONIC, &now); + old_ts = ((uint64_t) now.tv_sec - (SSM_POOL_RECLAIM_AGE_S + 1)) + * 1000000000ULL + (uint64_t) now.tv_nsec; + spb1->allocator_pid = fake_pid; spb2->allocator_pid = fake_pid; + spb1->alloc_ts = old_ts; + spb2->alloc_ts = old_ts; /* Keep spb3 with our pid */ /* Reclaim orphans from fake_pid */ diff --git a/src/lib/ssm/tests/rbuff_test.c b/src/lib/ssm/tests/rbuff_test.c index 58cb39c3..48e5a714 100644 --- a/src/lib/ssm/tests/rbuff_test.c +++ b/src/lib/ssm/tests/rbuff_test.c @@ -206,10 +206,10 @@ static int test_ssm_rbuff_fill_drain(void) return TEST_RC_FAIL; } -static int test_ssm_rbuff_acl(void) +static int test_ssm_rbuff_flags(void) { struct ssm_rbuff * rb; - uint32_t acl; + uint32_t flags; TEST_START(); @@ -219,16 +219,16 @@ static int test_ssm_rbuff_acl(void) goto fail; } - acl = ssm_rbuff_get_acl(rb); - if (acl != ACL_RDWR) { - printf("Expected ACL_RDWR, got %u.\n", acl); + flags = ssm_rbuff_get_flags(rb); + if (flags != RB_RDWR) { + printf("Expected RB_RDWR, got %u.\n", flags); goto fail_rb; } - ssm_rbuff_set_acl(rb, ACL_RDONLY); - acl = ssm_rbuff_get_acl(rb); - if (acl != ACL_RDONLY) { - printf("Expected ACL_RDONLY, got %u.\n", acl); + ssm_rbuff_clr_bits(rb, RB_WR); + flags = ssm_rbuff_get_flags(rb); + if (flags != RB_RD) { + printf("Expected RB_RD, got %u.\n", flags); goto fail_rb; } @@ -237,7 +237,7 @@ static int test_ssm_rbuff_acl(void) goto fail_rb; } - ssm_rbuff_set_acl(rb, ACL_FLOWDOWN); + ssm_rbuff_set_bits(rb, RB_FLOWDOWN); if (ssm_rbuff_write(rb, 1) != -EFLOWDOWN) { printf("Expected -EFLOWDOWN on FLOWDOWN.\n"); goto fail_rb; @@ -553,7 +553,7 @@ static int test_ssm_rbuff_blocking_flowdown(void) clock_gettime(PTHREAD_COND_CLOCK, &now); ts_add(&now, &interval, &abs_timeout); - ssm_rbuff_set_acl(rb, ACL_FLOWDOWN); + ssm_rbuff_set_bits(rb, RB_FLOWDOWN); ret = ssm_rbuff_read_b(rb, &abs_timeout); if (ret != -EFLOWDOWN) { @@ -561,7 +561,7 @@ static int test_ssm_rbuff_blocking_flowdown(void) goto fail_rb; } - ssm_rbuff_set_acl(rb, ACL_RDWR); + ssm_rbuff_clr_bits(rb, RB_FLOWDOWN); for (i = 0; i < SSM_RBUFF_SIZE - 1; ++i) { if (ssm_rbuff_write(rb, i) < 0) { @@ -573,7 +573,7 @@ static int test_ssm_rbuff_blocking_flowdown(void) clock_gettime(PTHREAD_COND_CLOCK, &now); ts_add(&now, &interval, &abs_timeout); - ssm_rbuff_set_acl(rb, ACL_FLOWDOWN); + ssm_rbuff_set_bits(rb, RB_FLOWDOWN); ret = ssm_rbuff_write_b(rb, 999, &abs_timeout); if (ret != -EFLOWDOWN) { @@ -581,7 +581,7 @@ static int test_ssm_rbuff_blocking_flowdown(void) goto fail_rb; } - ssm_rbuff_set_acl(rb, ACL_RDWR); + ssm_rbuff_clr_bits(rb, RB_FLOWDOWN); while (ssm_rbuff_read(rb) >= 0) ; @@ -664,7 +664,7 @@ int rbuff_test(int argc, ret |= test_ssm_rbuff_write_read(); ret |= test_ssm_rbuff_read_empty(); ret |= test_ssm_rbuff_fill_drain(); - ret |= test_ssm_rbuff_acl(); + ret |= test_ssm_rbuff_flags(); ret |= test_ssm_rbuff_open_close(); ret |= test_ssm_rbuff_threaded(); ret |= test_ssm_rbuff_blocking(); diff --git a/src/lib/tests/CMakeLists.txt b/src/lib/tests/CMakeLists.txt index 5a2f2c52..002d94af 100644 --- a/src/lib/tests/CMakeLists.txt +++ b/src/lib/tests/CMakeLists.txt @@ -10,20 +10,24 @@ create_test_sourcelist(${PARENT_DIR}_tests test_suite.c auth_test_slh_dsa.c bitmap_test.c btree_test.c - crc32_test.c crypt_test.c hash_test.c kex_test.c kex_test_ml_kem.c + keyrot_test.c md5_test.c sha3_test.c sockets_test.c time_test.c tpm_test.c + tw_test.c ) add_executable(${PARENT_DIR}_test ${${PARENT_DIR}_tests}) +target_include_directories(${PARENT_DIR}_test PRIVATE + ${CMAKE_SOURCE_DIR}/src/lib) + disable_test_logging_for_target(${PARENT_DIR}_test) target_link_libraries(${PARENT_DIR}_test ouroboros-common) diff --git a/src/lib/tests/auth_test.c b/src/lib/tests/auth_test.c index 1a5a87af..61f97683 100644 --- a/src/lib/tests/auth_test.c +++ b/src/lib/tests/auth_test.c @@ -24,11 +24,14 @@ #include <test/test.h> #include <ouroboros/crypt.h> +#include <ouroboros/name.h> #include <ouroboros/random.h> #include <ouroboros/utils.h> #include <test/certs/ecdsa.h> +#include <string.h> + #define TEST_MSG_SIZE 1500 static int test_auth_create_destroy_ctx(void) @@ -138,6 +141,47 @@ static int test_check_crt_name(void) return TEST_RC_FAIL; } +static int test_crt_name_confusion(void) +{ + char name[NAME_SIZE + 1]; + void * crt; + + TEST_START(); + + if (crypt_load_crt_str(confused_crt_ec, &crt) < 0) { + printf("Failed to load name-confusion certificate.\n"); + goto fail_load; + } + + /* Must extract the real CN, not the "CN=" decoy in the O field. */ + if (crypt_get_crt_name(crt, name) < 0) { + printf("Failed to extract name from certificate.\n"); + goto fail_check; + } + + if (strcmp(name, "attacker.unittest.o7s") != 0) { + printf("Extracted '%s', expected real CN.\n", name); + goto fail_check; + } + + /* The decoy name in the O field must never authenticate. */ + if (crypt_check_crt_name(crt, "victim.unittest.o7s") == 0) { + printf("Accepted spoofed name from O field.\n"); + goto fail_check; + } + + crypt_free_crt(crt); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_check: + crypt_free_crt(crt); + fail_load: + TEST_FAIL(); + return TEST_RC_FAIL; +} + static int test_load_free_privkey(void) { void * key; @@ -260,7 +304,7 @@ static int test_store_add(void) fail_add: crypt_free_crt(_root_ca_crt); fail_load: - crypt_free_crt(_root_ca_crt); + auth_destroy_ctx(ctx); fail_create: TEST_FAIL(); return TEST_RC_FAIL; @@ -347,6 +391,151 @@ static int test_verify_crt(void) return TEST_RC_FAIL; } +static int test_verify_crt_missing_root_ca(void) +{ + struct auth_ctx * auth; + void * _signed_server_crt; + void * _im_ca_crt; + + TEST_START(); + + auth = auth_create_ctx(); + if (auth == NULL) { + printf("Failed to create auth context.\n"); + goto fail_create_ctx; + } + + if (crypt_load_crt_str(signed_server_crt_ec, &_signed_server_crt) < 0) { + printf("Failed to load signed crt from string.\n"); + goto fail_load_signed; + } + + if (crypt_load_crt_str(im_ca_crt_ec, &_im_ca_crt) < 0) { + printf("Failed to load intermediate crt from string.\n"); + goto fail_load_im_ca; + } + + /* Add only the intermediate CA - root CA is missing */ + if (auth_add_crt_to_store(auth, _im_ca_crt) < 0) { + printf("Failed to add intermediate ca crt to auth store.\n"); + goto fail_add; + } + + if (auth_verify_crt(auth, _signed_server_crt) == 0) { + printf("Verification should fail without root CA.\n"); + goto fail_add; + } + + crypt_free_crt(_im_ca_crt); + crypt_free_crt(_signed_server_crt); + auth_destroy_ctx(auth); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_add: + crypt_free_crt(_im_ca_crt); + fail_load_im_ca: + crypt_free_crt(_signed_server_crt); + fail_load_signed: + auth_destroy_ctx(auth); + fail_create_ctx: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* auth_verify_crt_pin: pin must lie in the verified chain (NULL: any) */ +static int test_verify_crt_pin(void) +{ + struct auth_ctx * auth; + void * _root_ca_crt; + void * _im_ca_crt; + void * _signed_server_crt; + void * _other_ca_crt; + + TEST_START(); + + auth = auth_create_ctx(); + if (auth == NULL) { + printf("Failed to create auth context.\n"); + goto fail_create_ctx; + } + + if (crypt_load_crt_str(root_ca_crt_ec, &_root_ca_crt) < 0) { + printf("Failed to load root crt from string.\n"); + goto fail_load_root_ca; + } + + if (crypt_load_crt_str(im_ca_crt_ec, &_im_ca_crt) < 0) { + printf("Failed to load intermediate crt from string.\n"); + goto fail_load_im_ca; + } + + if (crypt_load_crt_str(signed_server_crt_ec, &_signed_server_crt) < 0) { + printf("Failed to load signed crt from string.\n"); + goto fail_load_signed; + } + + if (crypt_load_crt_str(other_ca_crt_ec, &_other_ca_crt) < 0) { + printf("Failed to load out-of-chain crt from string.\n"); + goto fail_load_other; + } + + if (auth_add_crt_to_store(auth, _root_ca_crt) < 0) { + printf("Failed to add root ca crt to auth store.\n"); + goto fail_verify; + } + + if (auth_add_crt_to_store(auth, _im_ca_crt) < 0) { + printf("Failed to add intermediate ca crt to auth store.\n"); + goto fail_verify; + } + + if (auth_verify_crt_pin(auth, _signed_server_crt, _im_ca_crt) < 0) { + printf("Failed to accept pin on intermediate CA.\n"); + goto fail_verify; + } + + if (auth_verify_crt_pin(auth, _signed_server_crt, _root_ca_crt) < 0) { + printf("Failed to accept pin on root CA.\n"); + goto fail_verify; + } + + if (auth_verify_crt_pin(auth, _signed_server_crt, _other_ca_crt) == 0) { + printf("Failed to reject out-of-chain pin.\n"); + goto fail_verify; + } + + if (auth_verify_crt_pin(auth, _signed_server_crt, NULL) < 0) { + printf("Failed to accept NULL (any) pin.\n"); + goto fail_verify; + } + + crypt_free_crt(_other_ca_crt); + crypt_free_crt(_signed_server_crt); + crypt_free_crt(_im_ca_crt); + crypt_free_crt(_root_ca_crt); + + auth_destroy_ctx(auth); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_verify: + crypt_free_crt(_other_ca_crt); + fail_load_other: + crypt_free_crt(_signed_server_crt); + fail_load_signed: + crypt_free_crt(_im_ca_crt); + fail_load_im_ca: + crypt_free_crt(_root_ca_crt); + fail_load_root_ca: + auth_destroy_ctx(auth); + fail_create_ctx: + TEST_FAIL(); + return TEST_RC_FAIL; +} + int test_auth_sign(void) { uint8_t buf[TEST_MSG_SIZE]; @@ -520,29 +709,36 @@ int auth_test(int argc, #ifdef HAVE_OPENSSL ret |= test_load_free_crt(); ret |= test_check_crt_name(); + ret |= test_crt_name_confusion(); ret |= test_crypt_get_pubkey_crt(); ret |= test_load_free_privkey(); ret |= test_load_free_pubkey(); ret |= test_crypt_check_pubkey_crt(); ret |= test_store_add(); ret |= test_verify_crt(); + ret |= test_verify_crt_missing_root_ca(); + ret |= test_verify_crt_pin(); ret |= test_auth_sign(); ret |= test_auth_bad_signature(); ret |= test_crt_str(); #else (void) test_load_free_crt; (void) test_check_crt_name; + (void) test_crt_name_confusion; (void) test_crypt_get_pubkey_crt; (void) test_load_free_privkey; (void) test_load_free_pubkey; (void) test_crypt_check_pubkey_crt; (void) test_store_add; (void) test_verify_crt; + (void) test_verify_crt_missing_root_ca; + (void) test_verify_crt_pin; (void) test_auth_sign; (void) test_auth_bad_signature; (void) test_crt_str; - ret = TEST_RC_SKIP; + if (ret == 0) + ret = TEST_RC_SKIP; #endif return ret; } diff --git a/src/lib/tests/auth_test_ml_dsa.c b/src/lib/tests/auth_test_ml_dsa.c index cc72e61b..e324c32d 100644 --- a/src/lib/tests/auth_test_ml_dsa.c +++ b/src/lib/tests/auth_test_ml_dsa.c @@ -333,7 +333,7 @@ int auth_test_ml_dsa(int argc, (void) argc; (void) argv; -#ifdef HAVE_OPENSSL_ML_DSA +#ifdef HAVE_ML ret |= test_auth_create_destroy_ctx(); ret |= test_load_free_crt(); ret |= test_load_free_privkey(); diff --git a/src/lib/tests/auth_test_slh_dsa.c b/src/lib/tests/auth_test_slh_dsa.c index 511d20fe..e9af8da8 100644 --- a/src/lib/tests/auth_test_slh_dsa.c +++ b/src/lib/tests/auth_test_slh_dsa.c @@ -344,7 +344,7 @@ int auth_test_slh_dsa(int argc, (void) argc; (void) argv; -#ifdef HAVE_OPENSSL_SLH_DSA +#ifdef HAVE_SLH ret |= test_auth_create_destroy_ctx(); ret |= test_load_free_crt(); ret |= test_load_free_privkey(); diff --git a/src/lib/tests/crypt_test.c b/src/lib/tests/crypt_test.c index 028c4eb5..f00618d8 100644 --- a/src/lib/tests/crypt_test.c +++ b/src/lib/tests/crypt_test.c @@ -30,6 +30,7 @@ #include <stdio.h> #define TEST_PACKET_SIZE 1500 +#define TEST_N_PACKETS 1000 extern const uint16_t crypt_supported_nids[]; extern const uint16_t md_supported_nids[]; @@ -39,9 +40,10 @@ static int test_crypt_create_destroy(void) struct crypt_ctx * ctx; uint8_t key[SYMMKEYSZ]; struct crypt_sk sk = { - .nid = NID_aes_256_gcm, - .key = key, - .rot_bit = KEY_ROTATION_BIT + .nid = NID_aes_256_gcm, + .key = key, + .epoch = 0, + .role = CRYPT_ROLE_INIT }; TEST_START(); @@ -67,18 +69,27 @@ static int test_crypt_create_destroy(void) static int test_crypt_encrypt_decrypt(int nid) { uint8_t pkt[TEST_PACKET_SIZE]; - struct crypt_ctx * ctx; + struct crypt_ctx * tx; + struct crypt_ctx * rx; uint8_t key[SYMMKEYSZ]; - struct crypt_sk sk = { - .nid = NID_aes_256_gcm, - .key = key, - .rot_bit = KEY_ROTATION_BIT + struct crypt_sk sk_tx = { + .key = key, + .epoch = 0, + .role = CRYPT_ROLE_INIT + }; + struct crypt_sk sk_rx = { + .key = key, + .epoch = 0, + .role = CRYPT_ROLE_RESP }; buffer_t in; buffer_t out; buffer_t out2; const char * cipher; + sk_tx.nid = nid; + sk_rx.nid = nid; + cipher = crypt_nid_to_str(nid); TEST_START("(%s)", cipher); @@ -92,53 +103,63 @@ static int test_crypt_encrypt_decrypt(int nid) goto fail_init; } - ctx = crypt_create_ctx(&sk); - if (ctx == NULL) { - printf("Failed to initialize cryptography.\n"); + tx = crypt_create_ctx(&sk_tx); + if (tx == NULL) { + printf("Failed to initialize TX cryptography.\n"); goto fail_init; } + rx = crypt_create_ctx(&sk_rx); + if (rx == NULL) { + printf("Failed to initialize RX cryptography.\n"); + goto fail_tx; + } + in.len = sizeof(pkt); in.data = pkt; - if (crypt_encrypt(ctx, in, &out) < 0) { + if (crypt_encrypt(tx, in, &out) < 0) { printf("Encryption failed.\n"); goto fail_encrypt; } if (out.len < in.len) { printf("Encryption returned too little data.\n"); - goto fail_encrypt; + goto fail_chk; } - if (crypt_decrypt(ctx, out, &out2) < 0) { + if (crypt_decrypt(rx, out, &out2) < 0) { printf("Decryption failed.\n"); goto fail_decrypt; } if (out2.len != in.len) { printf("Decrypted data length does not match original.\n"); - goto fail_chk; + goto fail_chk2; } if (memcmp(in.data, out2.data, in.len) != 0) { printf("Decrypted data does not match original.\n"); - goto fail_chk; + goto fail_chk2; } - crypt_destroy_ctx(ctx); freebuf(out2); freebuf(out); + crypt_destroy_ctx(rx); + crypt_destroy_ctx(tx); TEST_SUCCESS("(%s)", cipher); return TEST_RC_SUCCESS; - fail_chk: + fail_chk2: freebuf(out2); fail_decrypt: + fail_chk: freebuf(out); fail_encrypt: - crypt_destroy_ctx(ctx); + crypt_destroy_ctx(rx); + fail_tx: + crypt_destroy_ctx(tx); fail_init: TEST_FAIL("(%s)", cipher); return TEST_RC_FAIL; @@ -155,6 +176,214 @@ static int test_encrypt_decrypt_all(void) return ret; } +static int test_crypt_multi_packet(int nid) +{ + uint8_t pkt[TEST_PACKET_SIZE]; + struct crypt_ctx * tx; + struct crypt_ctx * rx; + uint8_t key[SYMMKEYSZ]; + struct crypt_sk sk_tx = { + .key = key, + .epoch = 0, + .role = CRYPT_ROLE_INIT + }; + struct crypt_sk sk_rx = { + .key = key, + .epoch = 0, + .role = CRYPT_ROLE_RESP + }; + buffer_t in; + buffer_t enc; + buffer_t dec; + const char * cipher; + int i; + + sk_tx.nid = nid; + sk_rx.nid = nid; + + cipher = crypt_nid_to_str(nid); + TEST_START("(%s)", cipher); + + if (random_buffer(key, sizeof(key)) < 0) { + printf("Failed to generate random key.\n"); + goto fail_init; + } + + if (random_buffer(pkt, sizeof(pkt)) < 0) { + printf("Failed to generate random data.\n"); + goto fail_init; + } + + tx = crypt_create_ctx(&sk_tx); + if (tx == NULL) { + printf("Failed to create TX context.\n"); + goto fail_init; + } + + rx = crypt_create_ctx(&sk_rx); + if (rx == NULL) { + printf("Failed to create RX context.\n"); + goto fail_tx; + } + + in.len = sizeof(pkt); + in.data = pkt; + + for (i = 0; i < TEST_N_PACKETS; i++) { + if (crypt_encrypt(tx, in, &enc) < 0) { + printf("Encryption failed at packet %d.\n", i); + goto fail_rx; + } + + if (crypt_decrypt(rx, enc, &dec) < 0) { + printf("Decryption failed at packet %d.\n", i); + freebuf(enc); + goto fail_rx; + } + + if (dec.len != in.len || + memcmp(in.data, dec.data, in.len) != 0) { + printf("Data mismatch at packet %d.\n", i); + freebuf(dec); + freebuf(enc); + goto fail_rx; + } + + freebuf(dec); + freebuf(enc); + } + + crypt_destroy_ctx(rx); + crypt_destroy_ctx(tx); + + TEST_SUCCESS("(%s)", cipher); + + return TEST_RC_SUCCESS; + fail_rx: + crypt_destroy_ctx(rx); + fail_tx: + crypt_destroy_ctx(tx); + fail_init: + TEST_FAIL("(%s)", cipher); + return TEST_RC_FAIL; +} + +static int test_multi_packet_all(void) +{ + int ret = 0; + int i; + + for (i = 0; crypt_supported_nids[i] != NID_undef; i++) + ret |= test_crypt_multi_packet(crypt_supported_nids[i]); + + return ret; +} + +static int test_crypt_aad_tamper(int nid) +{ + uint8_t pkt[TEST_PACKET_SIZE]; + struct crypt_ctx * tx; + struct crypt_ctx * rx; + uint8_t key[SYMMKEYSZ]; + struct crypt_sk sk_tx = { + .key = key, + .epoch = 0, + .role = CRYPT_ROLE_INIT + }; + struct crypt_sk sk_rx = { + .key = key, + .epoch = 0, + .role = CRYPT_ROLE_RESP + }; + buffer_t in; + buffer_t enc; + buffer_t dec; + const char * cipher; + + sk_tx.nid = nid; + sk_rx.nid = nid; + + cipher = crypt_nid_to_str(nid); + TEST_START("(%s)", cipher); + + if (random_buffer(key, sizeof(key)) < 0) { + printf("Failed to generate random key.\n"); + goto fail_init; + } + + if (random_buffer(pkt, sizeof(pkt)) < 0) { + printf("Failed to generate random data.\n"); + goto fail_init; + } + + tx = crypt_create_ctx(&sk_tx); + if (tx == NULL) { + printf("Failed to create TX context.\n"); + goto fail_init; + } + + rx = crypt_create_ctx(&sk_rx); + if (rx == NULL) { + printf("Failed to create RX context.\n"); + goto fail_tx; + } + + /* Only AEAD ciphers bind the selector as AAD. */ + if (crypt_get_tagsz(tx) == 0) { + crypt_destroy_ctx(rx); + crypt_destroy_ctx(tx); + + TEST_SUCCESS("(%s)", cipher); + + return TEST_RC_SUCCESS; + } + + in.len = sizeof(pkt); + in.data = pkt; + + if (crypt_encrypt(tx, in, &enc) < 0) { + printf("Encryption failed.\n"); + goto fail_rx; + } + + /* Flip a seq byte: epoch/node stay valid so the AEAD tag rejects. */ + enc.data[5] ^= 0x01; + + if (crypt_decrypt(rx, enc, &dec) == 0) { + printf("Decryption accepted a tampered selector.\n"); + freebuf(dec); + freebuf(enc); + goto fail_rx; + } + + freebuf(enc); + + crypt_destroy_ctx(rx); + crypt_destroy_ctx(tx); + + TEST_SUCCESS("(%s)", cipher); + + return TEST_RC_SUCCESS; + fail_rx: + crypt_destroy_ctx(rx); + fail_tx: + crypt_destroy_ctx(tx); + fail_init: + TEST_FAIL("(%s)", cipher); + return TEST_RC_FAIL; +} + +static int test_aad_tamper_all(void) +{ + int ret = 0; + int i; + + for (i = 0; crypt_supported_nids[i] != NID_undef; i++) + ret |= test_crypt_aad_tamper(crypt_supported_nids[i]); + + return ret; +} + #ifdef HAVE_OPENSSL #include <openssl/evp.h> #include <openssl/obj_mac.h> @@ -256,22 +485,17 @@ static int test_md_nid_values(void) } #endif -static int test_key_rotation(void) +static int test_crypt_headsz(void) { - uint8_t pkt[TEST_PACKET_SIZE]; - struct crypt_ctx * tx_ctx; - struct crypt_ctx * rx_ctx; - uint8_t key[SYMMKEYSZ]; - struct crypt_sk sk = { - .nid = NID_aes_256_gcm, - .key = key, - .rot_bit = 7 + struct crypt_ctx * ctx; + uint8_t key[SYMMKEYSZ]; + struct crypt_sk sk = { + .nid = NID_aes_256_gcm, + .key = key, + .epoch = 0, + .role = CRYPT_ROLE_INIT }; - buffer_t in; - buffer_t enc; - buffer_t dec; - uint32_t i; - uint32_t threshold; + int headsz; TEST_START(); @@ -280,155 +504,66 @@ static int test_key_rotation(void) goto fail; } - if (random_buffer(pkt, sizeof(pkt)) < 0) { - printf("Failed to generate random data.\n"); - goto fail; - } - - tx_ctx = crypt_create_ctx(&sk); - if (tx_ctx == NULL) { - printf("Failed to create TX context.\n"); + ctx = crypt_create_ctx(&sk); + if (ctx == NULL) { + printf("Failed to initialize cryptography.\n"); goto fail; } - rx_ctx = crypt_create_ctx(&sk); - if (rx_ctx == NULL) { - printf("Failed to create RX context.\n"); - goto fail_tx; - } - - in.len = sizeof(pkt); - in.data = pkt; - - threshold = (1U << sk.rot_bit); - - /* Encrypt and decrypt across multiple rotations */ - for (i = 0; i < threshold * 3; i++) { - if (crypt_encrypt(tx_ctx, in, &enc) < 0) { - printf("Encryption failed at packet %u.\n", i); - goto fail_rx; - } - - if (crypt_decrypt(rx_ctx, enc, &dec) < 0) { - printf("Decryption failed at packet %u.\n", i); - freebuf(enc); - goto fail_rx; - } - - if (dec.len != in.len || - memcmp(in.data, dec.data, in.len) != 0) { - printf("Data mismatch at packet %u.\n", i); - freebuf(dec); - freebuf(enc); - goto fail_rx; - } - - freebuf(dec); - freebuf(enc); + headsz = crypt_get_headsz(ctx); + if (headsz != 6) { + printf("Unexpected header size: %d (expected 6).\n", headsz); + goto fail_ctx; } - crypt_destroy_ctx(rx_ctx); - crypt_destroy_ctx(tx_ctx); + crypt_destroy_ctx(ctx); TEST_SUCCESS(); return TEST_RC_SUCCESS; - fail_rx: - crypt_destroy_ctx(rx_ctx); - fail_tx: - crypt_destroy_ctx(tx_ctx); + fail_ctx: + crypt_destroy_ctx(ctx); fail: TEST_FAIL(); return TEST_RC_FAIL; } -static int test_key_phase_bit(void) +static int test_crypt_ct_cmp(void) { - uint8_t pkt[TEST_PACKET_SIZE]; - struct crypt_ctx * ctx; - uint8_t key[SYMMKEYSZ]; - struct crypt_sk sk = { - .nid = NID_aes_256_gcm, - .key = key, - .rot_bit = 7 - }; - buffer_t in; - buffer_t out; - uint32_t count; - uint32_t threshold; - uint8_t phase_before; - uint8_t phase_after; - int ivsz; + uint8_t a[64]; + uint8_t b[64]; + size_t i; TEST_START(); - if (random_buffer(key, sizeof(key)) < 0) { - printf("Failed to generate random key.\n"); - goto fail; - } + for (i = 0; i < sizeof(a); i++) + a[i] = (uint8_t) i; - if (random_buffer(pkt, sizeof(pkt)) < 0) { - printf("Failed to generate random data.\n"); - goto fail; - } + memcpy(b, a, sizeof(a)); - ctx = crypt_create_ctx(&sk); - if (ctx == NULL) { - printf("Failed to initialize cryptography.\n"); + if (crypt_ct_cmp(a, b, sizeof(a)) != 0) { + printf("Equal buffers should compare equal.\n"); goto fail; } - ivsz = crypt_get_ivsz(ctx); - if (ivsz <= 0) { - printf("Invalid IV size.\n"); - goto fail_ctx; + if (crypt_ct_cmp(a, b, 0) != 0) { + printf("Zero length should compare equal.\n"); + goto fail; } - in.len = sizeof(pkt); - in.data = pkt; - - /* Encrypt packets up to just before rotation threshold */ - threshold = (1U << sk.rot_bit); - - /* Encrypt threshold - 1 packets (indices 0 to threshold-2) */ - for (count = 0; count < threshold - 1; count++) { - if (crypt_encrypt(ctx, in, &out) < 0) { - printf("Encryption failed at count %u.\n", count); - goto fail_ctx; + for (i = 0; i < sizeof(a); i++) { + b[i] ^= 0x01; + if (crypt_ct_cmp(a, b, sizeof(a)) == 0) { + printf("Difference at byte %zu not detected.\n", i); + goto fail; } - freebuf(out); - } - - /* Packet at index threshold-1: phase should still be initial */ - if (crypt_encrypt(ctx, in, &out) < 0) { - printf("Encryption failed before rotation.\n"); - goto fail_ctx; - } - phase_before = (out.data[0] & 0x80) ? 1 : 0; - freebuf(out); - - /* Packet at index threshold: phase should have toggled */ - if (crypt_encrypt(ctx, in, &out) < 0) { - printf("Encryption failed at rotation threshold.\n"); - goto fail_ctx; - } - phase_after = (out.data[0] & 0x80) ? 1 : 0; - freebuf(out); - /* Phase bit should have toggled */ - if (phase_before == phase_after) { - printf("Phase bit did not toggle: before=%u, after=%u.\n", - phase_before, phase_after); - goto fail_ctx; + b[i] ^= 0x01; } - crypt_destroy_ctx(ctx); - TEST_SUCCESS(); return TEST_RC_SUCCESS; - fail_ctx: - crypt_destroy_ctx(ctx); fail: TEST_FAIL(); return TEST_RC_FAIL; @@ -444,16 +579,20 @@ int crypt_test(int argc, ret |= test_crypt_create_destroy(); ret |= test_encrypt_decrypt_all(); + ret |= test_crypt_ct_cmp(); #ifdef HAVE_OPENSSL ret |= test_cipher_nid_values(); ret |= test_md_nid_values(); - ret |= test_key_rotation(); - ret |= test_key_phase_bit(); + ret |= test_multi_packet_all(); + ret |= test_aad_tamper_all(); + ret |= test_crypt_headsz(); #else - (void) test_key_rotation; - (void) test_key_phase_bit; + (void) test_multi_packet_all; + (void) test_aad_tamper_all; + (void) test_crypt_headsz; - return TEST_RC_SKIP; + if (ret == 0) + ret = TEST_RC_SKIP; #endif return ret; } diff --git a/src/lib/tests/hash_test.c b/src/lib/tests/hash_test.c index e43847e1..a2ba62cc 100644 --- a/src/lib/tests/hash_test.c +++ b/src/lib/tests/hash_test.c @@ -39,6 +39,79 @@ struct vec_entry { char * out; }; +struct mix_entry { + uint64_t in; + uint64_t out; +}; + +static int test_crc8(void) +{ + int ret = 0; + + struct vec_entry vec [] = { + { "", "00" }, + { "123456789", "df" }, + { NULL, NULL } + }; + + struct vec_entry * cur = vec; + + TEST_START(); + + while (cur->in != NULL) { + uint8_t crc; + char res[3]; + + str_hash(HASH_CRC8, &crc, cur->in); + + sprintf(res, "%02x", crc); + if (strcmp(res, cur->out) != 0) { + printf("Hash failed %s != %s.\n", res, cur->out); + ret |= -1; + } + + ++cur; + } + + TEST_END(ret); + + return ret; +} + +static int test_crc16(void) +{ + int ret = 0; + + struct vec_entry vec [] = { + { "", "ffff" }, + { "123456789", "29b1" }, + { NULL, NULL } + }; + + struct vec_entry * cur = vec; + + TEST_START(); + + while (cur->in != NULL) { + uint8_t crc[2]; + char res[5]; + + str_hash(HASH_CRC16, crc, cur->in); + + sprintf(res, "%02x%02x", crc[0], crc[1]); + if (strcmp(res, cur->out) != 0) { + printf("Hash failed %s != %s.\n", res, cur->out); + ret |= -1; + } + + ++cur; + } + + TEST_END(ret); + + return ret; +} + static int test_crc32(void) { int ret = 0; @@ -74,6 +147,42 @@ static int test_crc32(void) return ret; } +static int test_crc64(void) +{ + int ret = 0; + + struct vec_entry vec [] = { + { "", "0000000000000000" }, + { "123456789", "ae8b14860a799888" }, + { "0123456789abcdef", + "091485ca7018730e" }, + { NULL, NULL } + }; + + struct vec_entry * cur = vec; + + TEST_START(); + + while (cur->in != NULL) { + uint8_t crc[8]; + char res[17]; + + str_hash(HASH_CRC64, crc, cur->in); + + sprintf(res, HASH_FMT64, HASH_VAL64(crc)); + if (strcmp(res, cur->out) != 0) { + printf("Hash failed %s != %s.\n", res, cur->out); + ret |= -1; + } + + ++cur; + } + + TEST_END(ret); + + return ret; +} + static int test_md5(void) { int ret = 0; @@ -184,6 +293,36 @@ static int test_sha3(void) return ret; } +static int test_mix64(void) +{ + int ret = 0; + + struct mix_entry vec [] = { + { 0x0000000000000000ULL, 0x0000000000000000ULL }, + { 0x123456789abcdefeULL, 0xb1943cfea4f78f08ULL } + }; + + size_t n = sizeof(vec) / sizeof(vec[0]); + size_t i; + + TEST_START(); + + for (i = 0; i < n; i++) { + uint64_t res = hash_mix64(vec[i].in); + + if (res != vec[i].out) { + printf("Mix failed %016llx != %016llx.\n", + (unsigned long long) res, + (unsigned long long) vec[i].out); + ret |= -1; + } + } + + TEST_END(ret); + + return ret; +} + int hash_test(int argc, char ** argv) { @@ -192,11 +331,19 @@ int hash_test(int argc, (void) argc; (void) argv; + ret |= test_crc8(); + + ret |= test_crc16(); + ret |= test_crc32(); + ret |= test_crc64(); + ret |= test_md5(); ret |= test_sha3(); + ret |= test_mix64(); + return ret; } diff --git a/src/lib/tests/kex_test.c b/src/lib/tests/kex_test.c index ced760fe..5b2ccfc2 100644 --- a/src/lib/tests/kex_test.c +++ b/src/lib/tests/kex_test.c @@ -44,6 +44,9 @@ #define KEX_CONFIG_NONE \ "none\n" +#define KEX_CONFIG_NO_ENC \ + "encryption=none\n" + #define KEX_CONFIG_WHITESPACE \ "# Comment line\n" \ "kex = X448" \ @@ -58,6 +61,31 @@ "kex=X25519\n" \ "digest=sha384\n" +#define KEX_CONFIG_AUTH \ + "auth=required\n" + +#define KEX_CONFIG_AUTH_INVALID \ + "auth=mandatory\n" + +#define KEX_CONFIG_AUTH_OPTIONAL \ + "auth=optional\n" + +#define KEX_CONFIG_AUTH_THEN_NO_ENC \ + "auth=required\n" \ + "digest=sha512\n" \ + "encryption=none\n" + +#define KEX_CONFIG_NO_ENC_THEN_AUTH \ + "encryption=none\n" \ + "auth=required\n" \ + "digest=sha512\n" + +#define KEX_CONFIG_CACERT \ + "cacert=/etc/ouroboros/security/cacert/ca.crt\n" + +#define KEX_CONFIG_UNKNOWN_KEY \ + "autth=required\n" + /* Test key material for key loading tests */ #define X25519_PRIVKEY_PEM \ "-----BEGIN PRIVATE KEY-----\n" \ @@ -106,7 +134,7 @@ static int test_kex_dh_pkp_create_destroy(void) { struct sec_config kex; void * pkp; - uint8_t buf[MSGBUFSZ]; + uint8_t buf[CRYPT_KEY_BUFSZ]; TEST_START(); @@ -134,7 +162,7 @@ static int test_kex_get_algo_from_pk(const char * algo) void * pkp; buffer_t pk; ssize_t len; - uint8_t buf[MSGBUFSZ]; + uint8_t buf[CRYPT_KEY_BUFSZ]; char extracted_algo[256]; TEST_START("(%s)", algo); @@ -204,8 +232,8 @@ static int test_kex_dhe_derive(const char * algo) buffer_t pk1; buffer_t pk2; ssize_t len; - uint8_t buf1[MSGBUFSZ]; - uint8_t buf2[MSGBUFSZ]; + uint8_t buf1[CRYPT_KEY_BUFSZ]; + uint8_t buf2[CRYPT_KEY_BUFSZ]; uint8_t s1[SYMMKEYSZ]; uint8_t s2[SYMMKEYSZ]; @@ -213,6 +241,7 @@ static int test_kex_dhe_derive(const char * algo) memset(&kex, 0, sizeof(kex)); SET_KEX_ALGO(&kex, algo); + SET_KEX_KDF_NID(&kex, NID_sha256); len = kex_pkp_create(&kex, &pkp1, buf1); if (len < 0) { @@ -276,7 +305,7 @@ static int test_kex_validate_algo(void) goto fail; } -#ifdef HAVE_OPENSSL_ML_KEM +#ifdef HAVE_ML if (kex_validate_algo("ML-KEM-768") != 0) { printf("ML-KEM-768 should be valid.\n"); goto fail; @@ -317,13 +346,14 @@ static int test_kex_dhe_corrupted_pubkey(const char * algo) void * pkp; buffer_t pk; ssize_t len; - uint8_t buf[MSGBUFSZ]; + uint8_t buf[CRYPT_KEY_BUFSZ]; uint8_t s[SYMMKEYSZ]; TEST_START("(%s)", algo); memset(&kex, 0, sizeof(kex)); SET_KEX_ALGO(&kex, algo); + SET_KEX_KDF_NID(&kex, NID_sha256); len = kex_pkp_create(&kex, &pkp, buf); if (len < 0) { @@ -363,8 +393,8 @@ static int test_kex_dhe_wrong_algo(void) void * pkp2; buffer_t pk2; ssize_t len; - uint8_t buf1[MSGBUFSZ]; - uint8_t buf2[MSGBUFSZ]; + uint8_t buf1[CRYPT_KEY_BUFSZ]; + uint8_t buf2[CRYPT_KEY_BUFSZ]; uint8_t s[SYMMKEYSZ]; const char * algo1 = "X25519"; const char * algo2 = "X448"; @@ -375,6 +405,8 @@ static int test_kex_dhe_wrong_algo(void) memset(&kex2, 0, sizeof(kex2)); SET_KEX_ALGO(&kex1, algo1); SET_KEX_ALGO(&kex2, algo2); + SET_KEX_KDF_NID(&kex1, NID_sha256); + SET_KEX_KDF_NID(&kex2, NID_sha256); if (kex_pkp_create(&kex1, &pkp1, buf1) < 0) { printf("Failed to create first key pair.\n"); @@ -411,6 +443,57 @@ static int test_kex_dhe_wrong_algo(void) return TEST_RC_FAIL; } +static int test_kex_dhe_no_kdf(void) +{ + struct sec_config kex; + void * pkp1; + void * pkp2; + buffer_t pk2; + ssize_t len; + uint8_t buf1[CRYPT_KEY_BUFSZ]; + uint8_t buf2[CRYPT_KEY_BUFSZ]; + uint8_t s[SYMMKEYSZ]; + + TEST_START(); + + memset(&kex, 0, sizeof(kex)); + SET_KEX_ALGO(&kex, "X25519"); + + if (kex_pkp_create(&kex, &pkp1, buf1) < 0) { + printf("Failed to create first key pair.\n"); + goto fail; + } + + len = kex_pkp_create(&kex, &pkp2, buf2); + if (len < 0) { + printf("Failed to create second key pair.\n"); + goto fail_pkp1; + } + + pk2.len = (size_t) len; + pk2.data = buf2; + + /* No KDF configured: derive must fail, not fall back. */ + if (kex_dhe_derive(&kex, pkp1, pk2, s) == 0) { + printf("Derive succeeded without a KDF.\n"); + goto fail_pkp2; + } + + kex_pkp_destroy(pkp2); + kex_pkp_destroy(pkp1); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_pkp2: + kex_pkp_destroy(pkp2); + fail_pkp1: + kex_pkp_destroy(pkp1); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + static int test_kex_load_dhe_privkey(void) { void * key; @@ -639,7 +722,8 @@ static int test_kex_parse_config_custom(void) return TEST_RC_FAIL; } -static int test_kex_parse_config_none(void) +/* The old bare 'none' keyword must be rejected loudly */ +static int test_kex_parse_config_none_rejected(void) { struct sec_config kex; FILE * fp; @@ -654,14 +738,51 @@ static int test_kex_parse_config_none(void) goto fail; } + if (parse_sec_config(&kex, fp) == 0) { + printf("Bare 'none' keyword should be rejected.\n"); + fclose(fp); + goto fail; + } + + fclose(fp); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_kex_parse_config_no_enc(void) +{ + struct sec_config kex; + FILE * fp; + + TEST_START(); + + memset(&kex, 0, sizeof(kex)); + + fp = FMEMOPEN_STR(KEX_CONFIG_NO_ENC); + if (fp == NULL) { + printf("Failed to open memory stream.\n"); + goto fail; + } + if (parse_sec_config(&kex, fp) < 0) { - printf("Failed to parse 'none' config.\n"); + printf("Failed to parse encryption=none config.\n"); fclose(fp); goto fail; } - if (kex.x.nid != NID_undef) { - printf("'none' keyword should disable encryption.\n"); + if (kex.x.nid != NID_undef || kex.c.nid != NID_undef) { + printf("encryption=none should disable encryption.\n"); + fclose(fp); + goto fail; + } + + if (kex.d.nid != NID_sha256) { + printf("encryption=none should keep the digest.\n"); fclose(fp); goto fail; } @@ -799,6 +920,277 @@ static int test_kex_parse_config_digest(void) return TEST_RC_FAIL; } +static int test_kex_parse_config_auth(void) +{ + struct sec_config kex; + FILE * fp; + + TEST_START(); + + memset(&kex, 0, sizeof(kex)); + + fp = FMEMOPEN_STR(KEX_CONFIG_AUTH); + if (fp == NULL) { + printf("Failed to open memory stream.\n"); + goto fail; + } + + if (parse_sec_config(&kex, fp) < 0) { + printf("Failed to parse auth config.\n"); + fclose(fp); + goto fail; + } + + if (!kex.a.req) { + printf("auth=required not parsed correctly.\n"); + fclose(fp); + goto fail; + } + + fclose(fp); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_kex_parse_config_auth_invalid(void) +{ + struct sec_config kex; + FILE * fp; + + TEST_START(); + + memset(&kex, 0, sizeof(kex)); + + fp = FMEMOPEN_STR(KEX_CONFIG_AUTH_INVALID); + if (fp == NULL) { + printf("Failed to open memory stream.\n"); + goto fail; + } + + if (parse_sec_config(&kex, fp) == 0) { + printf("Invalid auth value should be rejected.\n"); + fclose(fp); + goto fail; + } + + fclose(fp); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* A caller-seeded req_auth survives parsing when no auth= line is set */ +static int test_kex_parse_config_auth_seed(void) +{ + struct sec_config kex; + FILE * fp; + + TEST_START(); + + memset(&kex, 0, sizeof(kex)); + kex.a.req = true; + + fp = FMEMOPEN_STR(KEX_CONFIG_NO_ENC); + if (fp == NULL) { + printf("Failed to open memory stream.\n"); + goto fail; + } + + if (parse_sec_config(&kex, fp) < 0) { + printf("Failed to parse config.\n"); + fclose(fp); + goto fail; + } + + if (!kex.a.req) { + printf("Seeded req_auth should survive parsing.\n"); + fclose(fp); + goto fail; + } + + fclose(fp); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* An explicit auth=optional clears a caller-seeded req_auth */ +static int test_kex_parse_config_auth_optional(void) +{ + struct sec_config kex; + FILE * fp; + + TEST_START(); + + memset(&kex, 0, sizeof(kex)); + kex.a.req = true; + + fp = FMEMOPEN_STR(KEX_CONFIG_AUTH_OPTIONAL); + if (fp == NULL) { + printf("Failed to open memory stream.\n"); + goto fail; + } + + if (parse_sec_config(&kex, fp) < 0) { + printf("Failed to parse auth=optional config.\n"); + fclose(fp); + goto fail; + } + + if (kex.a.req) { + printf("auth=optional should clear req_auth.\n"); + fclose(fp); + goto fail; + } + + fclose(fp); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* encryption=none must not drop auth=required or the digest */ +static int test_kex_parse_config_auth_no_enc(const char * config) +{ + struct sec_config kex; + FILE * fp; + + TEST_START(); + + memset(&kex, 0, sizeof(kex)); + + fp = FMEMOPEN_STR(config); + if (fp == NULL) { + printf("Failed to open memory stream.\n"); + goto fail; + } + + if (parse_sec_config(&kex, fp) < 0) { + printf("Failed to parse auth + encryption=none.\n"); + fclose(fp); + goto fail; + } + + if (!kex.a.req) { + printf("encryption=none should not drop required auth.\n"); + fclose(fp); + goto fail; + } + + if (kex.x.nid != NID_undef) { + printf("encryption=none should disable encryption.\n"); + fclose(fp); + goto fail; + } + + if (kex.d.nid != NID_sha512) { + printf("encryption=none should keep the digest.\n"); + fclose(fp); + goto fail; + } + + fclose(fp); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_kex_parse_config_cacert(void) +{ + struct sec_config kex; + FILE * fp; + + TEST_START(); + + memset(&kex, 0, sizeof(kex)); + + fp = FMEMOPEN_STR(KEX_CONFIG_CACERT); + if (fp == NULL) { + printf("Failed to open memory stream.\n"); + goto fail; + } + + if (parse_sec_config(&kex, fp) < 0) { + printf("Failed to parse cacert config.\n"); + fclose(fp); + goto fail; + } + + if (strcmp(kex.a.cacert, + "/etc/ouroboros/security/cacert/ca.crt") != 0) { + printf("cacert not parsed correctly.\n"); + fclose(fp); + goto fail; + } + + if (kex.a.req) { + printf("cacert must not imply req_auth.\n"); + fclose(fp); + goto fail; + } + + fclose(fp); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_kex_parse_config_unknown_key(void) +{ + struct sec_config kex; + FILE * fp; + + TEST_START(); + + memset(&kex, 0, sizeof(kex)); + + fp = FMEMOPEN_STR(KEX_CONFIG_UNKNOWN_KEY); + if (fp == NULL) { + printf("Failed to open memory stream.\n"); + goto fail; + } + + if (parse_sec_config(&kex, fp) == 0) { + printf("Unknown key should be rejected.\n"); + fclose(fp); + goto fail; + } + + fclose(fp); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + int kex_test(int argc, char ** argv) { @@ -809,7 +1201,16 @@ int kex_test(int argc, ret |= test_kex_create_destroy(); ret |= test_kex_parse_config_empty(); - ret |= test_kex_parse_config_none(); + ret |= test_kex_parse_config_none_rejected(); + ret |= test_kex_parse_config_no_enc(); + ret |= test_kex_parse_config_auth(); + ret |= test_kex_parse_config_auth_invalid(); + ret |= test_kex_parse_config_auth_seed(); + ret |= test_kex_parse_config_auth_optional(); + ret |= test_kex_parse_config_auth_no_enc(KEX_CONFIG_AUTH_THEN_NO_ENC); + ret |= test_kex_parse_config_auth_no_enc(KEX_CONFIG_NO_ENC_THEN_AUTH); + ret |= test_kex_parse_config_cacert(); + ret |= test_kex_parse_config_unknown_key(); #ifdef HAVE_OPENSSL ret |= test_kex_parse_config_custom(); ret |= test_kex_parse_config_whitespace(); @@ -821,6 +1222,7 @@ int kex_test(int argc, ret |= test_kex_validate_algo(); ret |= test_kex_get_algo_from_pk_all(); ret |= test_kex_dhe_wrong_algo(); + ret |= test_kex_dhe_no_kdf(); ret |= test_kex_dhe_corrupted_pubkey_all(); ret |= test_kex_load_dhe_privkey(); ret |= test_kex_load_dhe_pubkey(); @@ -833,12 +1235,14 @@ int kex_test(int argc, (void) test_kex_all; (void) test_kex_validate_algo; (void) test_kex_get_algo_from_pk_all; - (void) test_kex_dhe_wrong_algo(); + (void) test_kex_dhe_wrong_algo; + (void) test_kex_dhe_no_kdf; (void) test_kex_dhe_corrupted_pubkey_all; (void) test_kex_load_dhe_privkey; (void) test_kex_load_dhe_pubkey; - ret = TEST_RC_SKIP; + if (ret == 0) + ret = TEST_RC_SKIP; #endif return ret; } diff --git a/src/lib/tests/kex_test_ml_kem.c b/src/lib/tests/kex_test_ml_kem.c index 3bb9ae7c..1f7d7531 100644 --- a/src/lib/tests/kex_test_ml_kem.c +++ b/src/lib/tests/kex_test_ml_kem.c @@ -197,8 +197,8 @@ static int test_kex_kem(const char * algo) buffer_t ct; ssize_t len; ssize_t ct_len; - uint8_t buf1[MSGBUFSZ]; - uint8_t buf2[MSGBUFSZ]; + uint8_t buf1[CRYPT_KEY_BUFSZ]; + uint8_t buf2[CRYPT_KEY_BUFSZ]; uint8_t s1[SYMMKEYSZ]; uint8_t s2[SYMMKEYSZ]; int kdf; @@ -262,8 +262,8 @@ static int test_kex_kem_corrupted_ciphertext(const char * algo) buffer_t ct; ssize_t len; ssize_t ct_len; - uint8_t buf1[MSGBUFSZ]; - uint8_t buf2[MSGBUFSZ]; + uint8_t buf1[CRYPT_KEY_BUFSZ]; + uint8_t buf2[CRYPT_KEY_BUFSZ]; uint8_t s1[SYMMKEYSZ]; uint8_t s2[SYMMKEYSZ]; int kdf; @@ -334,9 +334,9 @@ static int test_kex_kem_wrong_keypair(const char * algo) buffer_t ct; ssize_t len; ssize_t ct_len; - uint8_t buf1[MSGBUFSZ]; - uint8_t buf2[MSGBUFSZ]; - uint8_t buf3[MSGBUFSZ]; + uint8_t buf1[CRYPT_KEY_BUFSZ]; + uint8_t buf2[CRYPT_KEY_BUFSZ]; + uint8_t buf3[CRYPT_KEY_BUFSZ]; uint8_t s1[SYMMKEYSZ]; uint8_t s2[SYMMKEYSZ]; @@ -402,8 +402,8 @@ static int test_kex_kem_truncated_ciphertext(const char * algo) buffer_t ct; ssize_t len; ssize_t ct_len; - uint8_t buf1[MSGBUFSZ]; - uint8_t buf2[MSGBUFSZ]; + uint8_t buf1[CRYPT_KEY_BUFSZ]; + uint8_t buf2[CRYPT_KEY_BUFSZ]; uint8_t s1[SYMMKEYSZ]; uint8_t s2[SYMMKEYSZ]; @@ -528,7 +528,7 @@ int kex_test_ml_kem(int argc, (void) argc; (void) argv; -#ifdef HAVE_OPENSSL_ML_KEM +#ifdef HAVE_ML ret |= test_kex_load_kem_privkey(); ret |= test_kex_load_kem_pubkey(); ret |= test_kex_kem_all(); diff --git a/src/lib/tests/keyrot_test.c b/src/lib/tests/keyrot_test.c new file mode 100644 index 00000000..efdc718e --- /dev/null +++ b/src/lib/tests/keyrot_test.c @@ -0,0 +1,1234 @@ +/* + * Ouroboros - Copyright (C) 2016 - 2026 + * + * Test of the key-rotation schedule + * + * Dimitri Staessens <dimitri@ouroboros.rocks> + * Sander Vrijders <sander@ouroboros.rocks> + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 as + * published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., http://www.fsf.org/about/contact/. + */ + +#define _POSIX_C_SOURCE 200809L + +#include "config.h" + +#include <test/test.h> + +#ifdef HAVE_OPENSSL +#include <ouroboros/crypt.h> +#include <ouroboros/errno.h> +#include <ouroboros/pthread.h> + +#include "crypt/keyrot.h" + +#include <stdbool.h> +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <time.h> + +static const uint8_t SEED_A[SYMMKEYSZ] = { + 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, + 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, 0x10, + 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, 0x18, + 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f, 0x20 +}; + +static const uint8_t SEED_B[SYMMKEYSZ] = { + 0xa1, 0xa2, 0xa3, 0xa4, 0xa5, 0xa6, 0xa7, 0xa8, + 0xa9, 0xaa, 0xab, 0xac, 0xad, 0xae, 0xaf, 0xb0, + 0xb1, 0xb2, 0xb3, 0xb4, 0xb5, 0xb6, 0xb7, 0xb8, + 0xb9, 0xba, 0xbb, 0xbc, 0xbd, 0xbe, 0xbf, 0xc0 +}; + +static int test_create_destroy(void) +{ + struct keyrot * kr; + + TEST_START(); + + kr = keyrot_create(SEED_A, 0, 0); + if (kr == NULL) + goto fail; + + keyrot_destroy(kr); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_epoch_range(void) +{ + struct keyrot * a; + + TEST_START(); + + /* epoch is a 4-bit wire field; 16 and up must be refused. */ + if (keyrot_create(SEED_A, 16, 0) != NULL) + goto fail; + + a = keyrot_create(SEED_A, 0, 0); + if (a == NULL) + goto fail; + + if (keyrot_rekey(a, SEED_A, 16) == 0) + goto fail_a; + + keyrot_destroy(a); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_a: + keyrot_destroy(a); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* Epochs of the live batches (cur, prev) must stay unique. */ +static int test_rekey_dup_epoch(void) +{ + struct keyrot * a; + + TEST_START(); + + a = keyrot_create(SEED_A, 0, 0); + if (a == NULL) + goto fail; + + if (keyrot_rekey(a, SEED_B, 0) != -1) { + printf("New key to the current epoch not a conflict.\n"); + goto fail_a; + } + + if (keyrot_rekey(a, SEED_B, 1) != 0) { + printf("Re-key to a fresh epoch refused.\n"); + goto fail_a; + } + + if (keyrot_rekey(a, SEED_B, 1) != -EREPLAY) { + printf("Same key to the current epoch not a replay.\n"); + goto fail_a; + } + + if (keyrot_rekey(a, SEED_A, 1) != -1) { + printf("New key to the current epoch not a conflict.\n"); + goto fail_a; + } + + if (keyrot_rekey(a, SEED_A, 0) != -EREPLAY) { + printf("Same key to the previous epoch not a replay.\n"); + goto fail_a; + } + + if (keyrot_rekey(a, SEED_B, 0) != -1) { + printf("New key to the previous epoch not a conflict.\n"); + goto fail_a; + } + + keyrot_destroy(a); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_a: + keyrot_destroy(a); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* The 4-bit wire epoch legitimately wraps 15 -> 0. */ +static int test_rekey_epoch_wrap(void) +{ + struct keyrot * a; + uint8_t sel[KR_SELECTOR_LEN]; + uint8_t n[KR_NONCE_LEN]; + const uint8_t * k; + + TEST_START(); + + a = keyrot_create(SEED_A, 14, 0); + if (a == NULL) + goto fail; + + if (keyrot_rekey(a, SEED_B, 15) != 0) + goto fail_a; + + if (keyrot_rekey(a, SEED_A, 0) != 0) { + printf("Epoch wrap 15 -> 0 refused.\n"); + goto fail_a; + } + + keyrot_tx_promote(a); + + if (keyrot_tx_next(a, sel, &k, n) != 0) { + printf("TX failed after epoch wrap.\n"); + goto fail_a; + } + + keyrot_destroy(a); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_a: + keyrot_destroy(a); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_tx_deterministic(void) +{ + struct keyrot * a; + struct keyrot * b; + uint8_t sela[KR_SELECTOR_LEN]; + uint8_t selb[KR_SELECTOR_LEN]; + uint8_t na[KR_NONCE_LEN]; + uint8_t nb[KR_NONCE_LEN]; + uint8_t ka[SYMMKEYSZ]; + const uint8_t * pa; + const uint8_t * pb; + + TEST_START(); + + a = keyrot_create(SEED_A, 0, 0); + if (a == NULL) + goto fail; + + b = keyrot_create(SEED_A, 0, 0); + if (b == NULL) + goto fail_a; + + if (keyrot_tx_next(a, sela, &pa, na) != 0) + goto fail_b; + + /* Copy out: pa points into the tcache, pb may reuse the slot. */ + memcpy(ka, pa, SYMMKEYSZ); + if (keyrot_tx_next(b, selb, &pb, nb) != 0) + goto fail_b; + + if (memcmp(sela, selb, KR_SELECTOR_LEN) != 0) + goto fail_b; + + if (memcmp(ka, pb, SYMMKEYSZ) != 0) + goto fail_b; + + if (memcmp(na, nb, KR_NONCE_LEN) != 0) + goto fail_b; + + keyrot_destroy(b); + keyrot_destroy(a); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_b: + keyrot_destroy(b); + fail_a: + keyrot_destroy(a); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_selector_layout(void) +{ + struct keyrot * a; + uint8_t sel[KR_SELECTOR_LEN]; + uint8_t nonce[KR_NONCE_LEN]; + const uint8_t * k; + + TEST_START(); + + a = keyrot_create(SEED_A, 3, 0); + if (a == NULL) + goto fail; + + /* First packet: epoch 3, node 0, seq 0 */ + if (keyrot_tx_next(a, sel, &k, nonce) != 0) + goto fail_a; + + if ((sel[0] >> 4) != 3) /* epoch */ + goto fail_a; + + if ((((sel[0] & 0x0F) << 8) | sel[1]) != 0) /* node */ + goto fail_a; + + if (sel[2] != 0 || sel[3] != 0 || sel[4] != 0 || sel[5] != 0) + goto fail_a; + + /* Second packet: seq advances to 1 */ + if (keyrot_tx_next(a, sel, &k, nonce) != 0) + goto fail_a; + + if (sel[5] != 1) + goto fail_a; + + keyrot_destroy(a); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_a: + keyrot_destroy(a); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_nodes_left_initial(void) +{ + struct keyrot * a; + + TEST_START(); + + a = keyrot_create(SEED_A, 0, 0); + if (a == NULL) + goto fail; + + if (keyrot_tx_nodes_left(a) != KEY_NODE_COUNT) + goto fail_a; + + keyrot_destroy(a); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_a: + keyrot_destroy(a); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_roundtrip(void) +{ + struct keyrot * a; /* role 0 */ + struct keyrot * b; /* role 1 */ + uint8_t sel[KR_SELECTOR_LEN]; + uint8_t ntx[KR_NONCE_LEN]; + uint8_t nrx[KR_NONCE_LEN]; + uint8_t ktx[SYMMKEYSZ]; + const uint8_t * ptx; + const uint8_t * prx; + struct kr_rx rx; + int i; + + TEST_START(); + + a = keyrot_create(SEED_A, 0, 0); + if (a == NULL) + goto fail; + + b = keyrot_create(SEED_A, 0, 1); + if (b == NULL) + goto fail_a; + + for (i = 0; i < 256; i++) { + if (keyrot_tx_next(a, sel, &ptx, ntx) != 0) + goto fail_b; + memcpy(ktx, ptx, SYMMKEYSZ); + if (keyrot_rx_lookup(b, sel, &prx, nrx, &rx) != 0) + goto fail_b; + if (keyrot_rx_commit(b, &rx) != 0) + goto fail_b; + if (memcmp(ktx, prx, SYMMKEYSZ) != 0) + goto fail_b; + if (memcmp(ntx, nrx, KR_NONCE_LEN) != 0) + goto fail_b; + } + + keyrot_destroy(b); + keyrot_destroy(a); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_b: + keyrot_destroy(b); + fail_a: + keyrot_destroy(a); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_direction_separation(void) +{ + struct keyrot * a; /* role 0 */ + struct keyrot * b; /* role 1 */ + uint8_t sela[KR_SELECTOR_LEN]; + uint8_t selb[KR_SELECTOR_LEN]; + uint8_t n[KR_NONCE_LEN]; + uint8_t ka[SYMMKEYSZ]; + const uint8_t * pa; + const uint8_t * pb; + + TEST_START(); + + a = keyrot_create(SEED_A, 0, 0); + if (a == NULL) + goto fail; + + b = keyrot_create(SEED_A, 0, 1); + if (b == NULL) + goto fail_a; + + if (keyrot_tx_next(a, sela, &pa, n) != 0) + goto fail_b; + + memcpy(ka, pa, SYMMKEYSZ); + if (keyrot_tx_next(b, selb, &pb, n) != 0) + goto fail_b; + + /* Same position, different role -> different leaf key */ + if (memcmp(ka, pb, SYMMKEYSZ) == 0) + goto fail_b; + + keyrot_destroy(b); + keyrot_destroy(a); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_b: + keyrot_destroy(b); + fail_a: + keyrot_destroy(a); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* Build a selector by hand (test knows the wire format). */ +static void mk_sel(uint8_t epoch, + uint16_t node, + uint32_t seq, + uint8_t sel[KR_SELECTOR_LEN]) +{ + sel[0] = (uint8_t) ((epoch << 4) | ((node >> 8) & 0x0F)); + sel[1] = (uint8_t) (node & 0xFF); + sel[2] = (uint8_t) (seq >> 24); + sel[3] = (uint8_t) (seq >> 16); + sel[4] = (uint8_t) (seq >> 8); + sel[5] = (uint8_t) (seq); +} + +static int test_random_access(void) +{ + struct keyrot * b; + uint8_t s0[KR_SELECTOR_LEN]; + uint8_t s5[KR_SELECTOR_LEN]; + uint8_t n[KR_NONCE_LEN]; + uint8_t k_first[SYMMKEYSZ]; + uint8_t k_node5[SYMMKEYSZ]; + const uint8_t * p; + struct kr_rx rx; + + TEST_START(); + + b = keyrot_create(SEED_A, 0, 1); + if (b == NULL) + goto fail; + + mk_sel(0, 0, 0, s0); + mk_sel(0, 5, 12345, s5); /* a far-ahead node, mid-span */ + + /* Jump straight to node 0 */ + if (keyrot_rx_lookup(b, s0, &p, n, &rx) != 0) + goto fail_b; + + memcpy(k_first, p, SYMMKEYSZ); + + /* Jump forward to node 5 (simulates a burst skip) */ + if (keyrot_rx_lookup(b, s5, &p, n, &rx) != 0) + goto fail_b; + + memcpy(k_node5, p, SYMMKEYSZ); + + /* Different nodes must yield different keys */ + if (memcmp(k_first, k_node5, SYMMKEYSZ) == 0) + goto fail_b; + + /* Jump back to node 0: still works, identical (no wedge) */ + if (keyrot_rx_lookup(b, s0, &p, n, &rx) != 0) + goto fail_b; + + if (memcmp(k_first, p, SYMMKEYSZ) != 0) + goto fail_b; + + /* Out-of-range node must be rejected */ + mk_sel(0, KEY_NODE_COUNT, 0, s0); + if (keyrot_rx_lookup(b, s0, &p, n, &rx) == 0) + goto fail_b; + + keyrot_destroy(b); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_b: + keyrot_destroy(b); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* + * Look up and commit one within-node counter on epoch 0. Returns 0 on + * accept, 1 on a rejected commit (replay or too old), and -1 if the + * lookup itself failed - kept distinct so a reject assertion can never + * pass on an unrelated lookup miss. + */ +static int commit_ctr(struct keyrot * kr, + uint32_t ctr) +{ + uint8_t sel[KR_SELECTOR_LEN]; + uint8_t n[KR_NONCE_LEN]; + const uint8_t * k; + struct kr_rx rx; + + mk_sel(0, 0, ctr, sel); + + if (keyrot_rx_lookup(kr, sel, &k, n, &rx) != 0) + return -1; + + return keyrot_rx_commit(kr, &rx) == 0 ? 0 : 1; +} + +static int test_replay_window(void) +{ + struct keyrot * b; + struct keyrot * c; + uint32_t base; + uint32_t jump; + + TEST_START(); + + b = keyrot_create(SEED_A, 0, 1); + if (b == NULL) + goto fail; + + /* Fresh counters accepted; an immediate replay is rejected. */ + if (commit_ctr(b, 100) != 0) + goto fail_b; + + if (commit_ctr(b, 100) != 1) + goto fail_b; + + /* In-window reorder: accepted once, rejected on replay. */ + if (commit_ctr(b, 105) != 0) + goto fail_b; + + if (commit_ctr(b, 102) != 0) + goto fail_b; + + if (commit_ctr(b, 102) != 1) + goto fail_b; + + /* Too-old boundary: the window edge is rejected, just inside is not. */ + base = 4 * KEY_REPLAY_WINDOW; + if (commit_ctr(b, base) != 0) + goto fail_b; + + if (commit_ctr(b, base - (KEY_REPLAY_WINDOW - 64)) != 1) + goto fail_b; + + if (commit_ctr(b, base - (KEY_REPLAY_WINDOW - 64) + 1) != 0) + goto fail_b; + + /* + * RFC 6479 slack-word regression: two low counters, then a + * forward jump of a full bitmap that aliases their slot, then a + * replay of a low counter. Without the reserved slack word this + * replay is wrongly accepted. + */ + c = keyrot_create(SEED_A, 0, 1); + if (c == NULL) + goto fail_b; + + if (commit_ctr(c, 70) != 0) + goto fail_c; + + if (commit_ctr(c, 74) != 0) + goto fail_c; + + jump = KEY_REPLAY_WINDOW + 63; + if (commit_ctr(c, jump) != 0) + goto fail_c; + + if (commit_ctr(c, 74) != 1) + goto fail_c; + + keyrot_destroy(c); + keyrot_destroy(b); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_c: + keyrot_destroy(c); + fail_b: + keyrot_destroy(b); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_lookup_no_commit(void) +{ + struct keyrot * b; + uint8_t sel[KR_SELECTOR_LEN]; + uint8_t n[KR_NONCE_LEN]; + const uint8_t * k; + struct kr_rx rx; + int i; + + TEST_START(); + + b = keyrot_create(SEED_A, 0, 1); + if (b == NULL) + goto fail; + + mk_sel(0, 0, 100, sel); + + /* Repeated lookups are pre-AEAD and must not consume the slot. */ + for (i = 0; i < 4; i++) { + if (keyrot_rx_lookup(b, sel, &k, n, &rx) != 0) + goto fail_b; + } + + /* The slot is still fresh, so the first commit accepts ... */ + if (keyrot_rx_commit(b, &rx) != 0) + goto fail_b; + + /* ... and only the commit advanced it, so the next is a replay. */ + if (keyrot_rx_commit(b, &rx) == 0) + goto fail_b; + + keyrot_destroy(b); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_b: + keyrot_destroy(b); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_commit_prev_batch(void) +{ + struct keyrot * b; + uint8_t sel[KR_SELECTOR_LEN]; + uint8_t n[KR_NONCE_LEN]; + const uint8_t * k; + struct kr_rx rx; + + TEST_START(); + + b = keyrot_create(SEED_A, 0, 1); + if (b == NULL) + goto fail; + + /* Capture a packet under cur (epoch 0). */ + mk_sel(0, 0, 7, sel); + if (keyrot_rx_lookup(b, sel, &k, n, &rx) != 0) + goto fail_b; + + /* Re-key: the captured batch becomes prev and the flag clears. */ + if (keyrot_rekey(b, SEED_B, 1) != 0) + goto fail_b; + + /* The straggler commits under prev without claiming a switch. */ + if (keyrot_rx_commit(b, &rx) != 0) + goto fail_b; + + if (keyrot_peer_switched(b)) + goto fail_b; + + /* prev still holds a replay window: its replay is rejected. */ + if (keyrot_rx_commit(b, &rx) == 0) + goto fail_b; + + keyrot_destroy(b); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_b: + keyrot_destroy(b); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_replay_forward_clear(void) +{ + struct keyrot * d; + uint32_t low; + uint32_t alias; + uint32_t jump; + + TEST_START(); + + d = keyrot_create(SEED_A, 0, 1); + if (d == NULL) + goto fail; + + /* alias shares low's slot a window away; the jump must clear it. */ + low = 10; + alias = low + KEY_REPLAY_WINDOW; + jump = alias + KEY_REPLAY_WINDOW / 2; + + if (commit_ctr(d, low) != 0) + goto fail_d; + + if (commit_ctr(d, jump) != 0) + goto fail_d; + + if (commit_ctr(d, alias) != 0) + goto fail_d; + + if (commit_ctr(d, alias) != 1) + goto fail_d; + + keyrot_destroy(d); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_d: + keyrot_destroy(d); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_rekey_overlap(void) +{ + struct keyrot * a; /* role 0 */ + struct keyrot * b; /* role 1 */ + uint8_t old_sel[KR_SELECTOR_LEN]; + uint8_t sel[KR_SELECTOR_LEN]; + uint8_t ntx[KR_NONCE_LEN]; + uint8_t nrx[KR_NONCE_LEN]; + uint8_t ktx[SYMMKEYSZ]; + const uint8_t * ptx; + const uint8_t * prx; + struct kr_rx rx; + + TEST_START(); + + a = keyrot_create(SEED_A, 0, 0); + if (a == NULL) + goto fail; + + b = keyrot_create(SEED_A, 0, 1); + if (b == NULL) + goto fail_a; + + /* Send one gen-0 packet; keep its selector for the overlap. */ + if (keyrot_tx_next(a, old_sel, &ptx, ntx) != 0) + goto fail_b; + + memcpy(ktx, ptx, SYMMKEYSZ); + if (keyrot_rx_lookup(b, old_sel, &prx, nrx, &rx) != 0) + goto fail_b; + + if (memcmp(ktx, prx, SYMMKEYSZ) != 0) + goto fail_b; + + /* Both ends re-key to epoch 1 with a fresh seed. */ + if (keyrot_rekey(a, SEED_B, 1) != 0) + goto fail_b; + + if (keyrot_rekey(b, SEED_B, 1) != 0) + goto fail_b; + + /* TX is gated until promotion; promote a to emit the new epoch. */ + keyrot_tx_promote(a); + + /* New gen-1 traffic works. */ + if (keyrot_tx_next(a, sel, &ptx, ntx) != 0) + goto fail_b; + + memcpy(ktx, ptx, SYMMKEYSZ); + if (keyrot_rx_lookup(b, sel, &prx, nrx, &rx) != 0) + goto fail_b; + + if (memcmp(ktx, prx, SYMMKEYSZ) != 0) + goto fail_b; + + /* A straggling gen-0 packet still decrypts (overlap window). */ + if (keyrot_rx_lookup(b, old_sel, &prx, nrx, &rx) != 0) + goto fail_b; + + /* An unknown epoch is rejected. */ + mk_sel(7, 0, 0, sel); + if (keyrot_rx_lookup(b, sel, &prx, nrx, &rx) == 0) + goto fail_b; + + keyrot_destroy(b); + keyrot_destroy(a); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_b: + keyrot_destroy(b); + fail_a: + keyrot_destroy(a); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_tx_gate(void) +{ + struct keyrot * a; /* role 0 */ + struct keyrot * b; /* role 1 */ + uint8_t sel[KR_SELECTOR_LEN]; + uint8_t n[KR_NONCE_LEN]; + const uint8_t * p; + struct kr_rx rx; + + TEST_START(); + + a = keyrot_create(SEED_A, 0, 0); + if (a == NULL) + goto fail; + + b = keyrot_create(SEED_A, 0, 1); + if (b == NULL) + goto fail_a; + + /* Both re-key to epoch 1; TX must stay on epoch 0 until promoted. */ + if (keyrot_rekey(a, SEED_B, 1) != 0) + goto fail_b; + + if (keyrot_rekey(b, SEED_B, 1) != 0) + goto fail_b; + + /* a's TX still stamps the old epoch (0). */ + if (keyrot_tx_next(a, sel, &p, n) != 0) + goto fail_b; + + if ((sel[0] >> 4) != 0) + goto fail_b; + + /* b decrypts the old-epoch packet via its prev batch. */ + if (keyrot_rx_lookup(b, sel, &p, n, &rx) != 0) + goto fail_b; + + if (keyrot_rx_commit(b, &rx) != 0) + goto fail_b; + + /* b has not yet seen the new epoch from a. */ + if (keyrot_peer_switched(b)) + goto fail_b; + + /* a promotes; its TX now stamps the new epoch (1). */ + keyrot_tx_promote(a); + if (keyrot_tx_next(a, sel, &p, n) != 0) + goto fail_b; + + if ((sel[0] >> 4) != 1) + goto fail_b; + + /* b sees the new epoch and reports the peer switched. */ + if (keyrot_rx_lookup(b, sel, &p, n, &rx) != 0) + goto fail_b; + + if (keyrot_rx_commit(b, &rx) != 0) + goto fail_b; + + if (!keyrot_peer_switched(b)) + goto fail_b; + + keyrot_destroy(b); + keyrot_destroy(a); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_b: + keyrot_destroy(b); + fail_a: + keyrot_destroy(a); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_peer_switched_commit_only(void) +{ + struct keyrot * b; + uint8_t sel[KR_SELECTOR_LEN]; + uint8_t n[KR_NONCE_LEN]; + const uint8_t * k; + struct kr_rx rx; + + TEST_START(); + + b = keyrot_create(SEED_A, 0, 1); + if (b == NULL) + goto fail; + + /* A re-key clears the flag until a packet is seen on cur. */ + if (keyrot_rekey(b, SEED_B, 1) != 0) + goto fail_b; + + if (keyrot_peer_switched(b)) + goto fail_b; + + mk_sel(1, 0, 0, sel); + + /* Lookup is pre-AEAD: selecting a key must not flip the flag. */ + if (keyrot_rx_lookup(b, sel, &k, n, &rx) != 0) + goto fail_b; + + if (keyrot_peer_switched(b)) + goto fail_b; + + /* Commit runs post-AEAD and is what records the peer switched. */ + if (keyrot_rx_commit(b, &rx) != 0) + goto fail_b; + + if (!keyrot_peer_switched(b)) + goto fail_b; + + keyrot_destroy(b); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_b: + keyrot_destroy(b); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_commit_evicted(void) +{ + struct keyrot * b; + uint8_t sel[KR_SELECTOR_LEN]; + uint8_t n[KR_NONCE_LEN]; + const uint8_t * k; + struct kr_rx rx; + + TEST_START(); + + b = keyrot_create(SEED_A, 0, 1); + if (b == NULL) + goto fail; + + mk_sel(0, 0, 3, sel); + if (keyrot_rx_lookup(b, sel, &k, n, &rx) != 0) + goto fail_b; + + /* Two re-keys drop the captured batch from both cur and prev. */ + if (keyrot_rekey(b, SEED_B, 1) != 0) + goto fail_b; + + if (keyrot_rekey(b, SEED_A, 2) != 0) + goto fail_b; + + /* Commit on an evicted batch is a silent no-op, not a fault. */ + if (keyrot_rx_commit(b, &rx) != 0) + goto fail_b; + + keyrot_destroy(b); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_b: + keyrot_destroy(b); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* TX fails closed when the tx_epoch batch is evicted, until promote. */ +static int test_tx_fail_closed(void) +{ + struct keyrot * b; + uint8_t sel[KR_SELECTOR_LEN]; + uint8_t n[KR_NONCE_LEN]; + const uint8_t * k; + + TEST_START(); + + b = keyrot_create(SEED_A, 0, 0); + if (b == NULL) + goto fail; + + if (keyrot_rekey(b, SEED_B, 1) != 0) + goto fail_b; + + if (keyrot_tx_next(b, sel, &k, n) != 0) { + printf("TX should keep the old epoch after one re-key.\n"); + goto fail_b; + } + + /* Second re-key without promote evicts the TX epoch-0 batch. */ + if (keyrot_rekey(b, SEED_A, 2) != 0) + goto fail_b; + + if (keyrot_tx_next(b, sel, &k, n) == 0) { + printf("TX should fail closed with tx_epoch evicted.\n"); + goto fail_b; + } + + keyrot_tx_promote(b); + + if (keyrot_tx_next(b, sel, &k, n) != 0) { + printf("TX should resync after promote.\n"); + goto fail_b; + } + + keyrot_destroy(b); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_b: + keyrot_destroy(b); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* + * Concurrency: many TX threads + RX + re-key share one keyrot. The + * (epoch, counter) the TX side stamps must be globally unique (no AEAD + * nonce reuse). Capped below 16 re-keys so epoch maps 1:1 to a batch and + * the wire epoch never wraps (a wrapped epoch under a fresh key is not + * reuse but would false-trip the uniqueness check). Run under TSan to + * catch data races the static reviews can't. + */ +#define CT_THREADS 4 +#define CT_PKTS 2000 +#define CT_REKEYS 8 + +struct ct_rec { + uint8_t epoch; + uint64_t ctr; +}; + +struct ct_arg { + struct keyrot * kr; + struct ct_rec * recs; + size_t n; +}; + +static void * ct_tx_thread(void * a) +{ + struct ct_arg * arg = a; + uint8_t sel[KR_SELECTOR_LEN]; + uint8_t nonce[KR_NONCE_LEN]; + const uint8_t * k; + uint64_t ctr; + size_t i; + size_t j; + + for (i = 0; i < CT_PKTS; i++) { + if (keyrot_tx_next(arg->kr, sel, &k, nonce) != 0) + continue; + + ctr = 0; + for (j = 0; j < 8; j++) + ctr = (ctr << 8) | nonce[j]; + + arg->recs[arg->n].epoch = (uint8_t) (sel[0] >> 4); + arg->recs[arg->n].ctr = ctr; + arg->n++; + } + + return NULL; +} + +static void * ct_rx_thread(void * a) +{ + struct keyrot * kr = a; + uint8_t sel[KR_SELECTOR_LEN]; + uint8_t nonce[KR_NONCE_LEN]; + const uint8_t * k; + struct kr_rx rx; + size_t i; + + /* Exercise rx_lookup against re-key reclaim; results ignored. */ + for (i = 0; i < CT_PKTS; i++) { + mk_sel((uint8_t) (i % 16), 0, (uint32_t) i, sel); + if (keyrot_rx_lookup(kr, sel, &k, nonce, &rx) == 0) + (void) keyrot_rx_commit(kr, &rx); + } + + return NULL; +} + +static void * ct_rekey_thread(void * a) +{ + struct keyrot * kr = a; + struct timespec t; + int e; + + t.tv_sec = 0; + t.tv_nsec = 2 * 1000 * 1000; /* 2 ms */ + + for (e = 1; e <= CT_REKEYS; e++) { + nanosleep(&t, NULL); + if (keyrot_rekey(kr, (e & 1) ? SEED_B : SEED_A, + (uint8_t) e) != 0) + break; + keyrot_tx_promote(kr); + } + + return NULL; +} + +static int ct_cmp(const void * x, + const void * y) +{ + const struct ct_rec * a = x; + const struct ct_rec * b = y; + + if (a->epoch != b->epoch) + return a->epoch < b->epoch ? -1 : 1; + + if (a->ctr != b->ctr) + return a->ctr < b->ctr ? -1 : 1; + + return 0; +} + +static int test_concurrent_nonce_unique(void) +{ + struct keyrot * kr; + struct ct_arg arg[CT_THREADS]; + pthread_t tx[CT_THREADS]; + pthread_t rx; + pthread_t rk; + struct ct_rec * all; + size_t total; + size_t i; + bool reuse = false; + + TEST_START(); + + kr = keyrot_create(SEED_A, 0, 0); + if (kr == NULL) + goto fail; + + all = malloc(sizeof(*all) * CT_THREADS * CT_PKTS); + if (all == NULL) + goto fail_kr; + + for (i = 0; i < CT_THREADS; i++) { + arg[i].kr = kr; + arg[i].n = 0; + arg[i].recs = all + i * CT_PKTS; + } + + for (i = 0; i < CT_THREADS; i++) + pthread_create(&tx[i], NULL, ct_tx_thread, &arg[i]); + + pthread_create(&rx, NULL, ct_rx_thread, kr); + pthread_create(&rk, NULL, ct_rekey_thread, kr); + + for (i = 0; i < CT_THREADS; i++) + pthread_join(tx[i], NULL); + + pthread_join(rx, NULL); + pthread_join(rk, NULL); + + total = 0; + for (i = 0; i < CT_THREADS; i++) { + memmove(all + total, all + i * CT_PKTS, + arg[i].n * sizeof(*all)); + total += arg[i].n; + } + + qsort(all, total, sizeof(*all), ct_cmp); + + for (i = 1; i < total; i++) + if (ct_cmp(&all[i - 1], &all[i]) == 0) { + printf("(epoch %u, ctr %llu) reused\n", + all[i].epoch, + (unsigned long long) all[i].ctr); + reuse = true; + break; + } + + free(all); + + if (reuse) + goto fail_kr; + + keyrot_destroy(kr); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_kr: + keyrot_destroy(kr); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} +#endif /* HAVE_OPENSSL */ + +int keyrot_test(int argc, + char ** argv) +{ + int ret = 0; + + (void) argc; + (void) argv; + +#ifdef HAVE_OPENSSL + ret |= test_create_destroy(); + ret |= test_epoch_range(); + ret |= test_rekey_dup_epoch(); + ret |= test_rekey_epoch_wrap(); + ret |= test_tx_deterministic(); + ret |= test_selector_layout(); + ret |= test_nodes_left_initial(); + ret |= test_roundtrip(); + ret |= test_direction_separation(); + ret |= test_random_access(); + ret |= test_peer_switched_commit_only(); + ret |= test_commit_evicted(); + ret |= test_tx_fail_closed(); + ret |= test_replay_window(); + ret |= test_lookup_no_commit(); + ret |= test_commit_prev_batch(); + ret |= test_replay_forward_clear(); + ret |= test_rekey_overlap(); + ret |= test_tx_gate(); + ret |= test_concurrent_nonce_unique(); +#endif + return ret; +} diff --git a/src/lib/tests/tpm_test.c b/src/lib/tests/tpm_test.c index df1d8850..7cc049cd 100644 --- a/src/lib/tests/tpm_test.c +++ b/src/lib/tests/tpm_test.c @@ -21,7 +21,7 @@ */ -#include "tpm.c" +#include <ouroboros/tpm.h> #include <test/test.h> diff --git a/src/lib/tests/tw_test.c b/src/lib/tests/tw_test.c new file mode 100644 index 00000000..32c302c4 --- /dev/null +++ b/src/lib/tests/tw_test.c @@ -0,0 +1,663 @@ +/* + * Ouroboros - Copyright (C) 2016 - 2026 + * + * Generic timing-wheel tests + * + * Dimitri Staessens <dimitri@ouroboros.rocks> + * Sander Vrijders <sander@ouroboros.rocks> + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 as + * published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., http://www.fsf.org/about/contact/. + */ + +#if defined(__linux__) || defined(__CYGWIN__) +#define _DEFAULT_SOURCE +#else +#define _POSIX_C_SOURCE 200809L +#endif + +#include "config.h" + +#include <test/test.h> + +#include <ouroboros/time.h> +#include <ouroboros/tw.h> + +#include <stdint.h> +#include <stdio.h> +#include <time.h> + +struct payload { + struct tw_entry tw; + int fired; +}; + +struct cancel_payload { + struct tw_entry tw; + int fired; + struct tw_entry * sibling; +}; + +struct repost_payload { + struct tw_entry tw; + int fired; + struct payload * sibling; + uint64_t repost_at; +}; + +static void cb_count(void * arg) +{ + struct payload * p = arg; + p->fired++; +} + +static void cb_cancel_sibling(void * arg) +{ + struct cancel_payload * p = arg; + p->fired++; + tw_cancel(p->sibling); +} + +static void cb_repost_sibling(void * arg) +{ + struct repost_payload * p = arg; + p->fired++; + tw_post(&p->sibling->tw, p->repost_at, cb_count, p->sibling); +} + +static uint64_t now_ns(void) +{ + struct timespec ts; + clock_gettime(PTHREAD_COND_CLOCK, &ts); + return TS_TO_UINT64(ts); +} + +static void sleep_ns(uint64_t ns) +{ + struct timespec ts; + UINT64_TO_TS(ns, &ts); + nanosleep(&ts, NULL); +} + +static int test_tw_init_fini(void) +{ + TEST_START(); + + if (tw_init() < 0) { + printf("tw_init failed.\n"); + goto fail; + } + + tw_fini(); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_tw_post_fires_after_deadline(void) +{ + struct payload p; + + TEST_START(); + + if (tw_init() < 0) + goto fail; + + tw_init_entry(&p.tw); + p.fired = 0; + + tw_post(&p.tw, now_ns() + 5 * MILLION, cb_count, &p); + + sleep_ns(20 * MILLION); + tw_move(); + + if (p.fired != 1) { + printf("expected 1 fire, got %d\n", p.fired); + goto fail_post; + } + + tw_cancel(&p.tw); + tw_fini(); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_post: + tw_cancel(&p.tw); + tw_fini(); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_tw_no_fire_before_deadline(void) +{ + struct payload p; + + TEST_START(); + + if (tw_init() < 0) + goto fail; + + tw_init_entry(&p.tw); + p.fired = 0; + + tw_post(&p.tw, now_ns() + 100 * MILLION, cb_count, &p); + + sleep_ns(2 * MILLION); + tw_move(); + + if (p.fired != 0) { + printf("expected 0 fires, got %d\n", p.fired); + goto fail_post; + } + + tw_cancel(&p.tw); + tw_fini(); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_post: + tw_cancel(&p.tw); + tw_fini(); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_tw_cancel_prevents_fire(void) +{ + struct payload p; + + TEST_START(); + + if (tw_init() < 0) + goto fail; + + tw_init_entry(&p.tw); + p.fired = 0; + + tw_post(&p.tw, now_ns() + 5 * MILLION, cb_count, &p); + tw_cancel(&p.tw); + + sleep_ns(20 * MILLION); + tw_move(); + + if (p.fired != 0) { + printf("cancelled entry fired %d times\n", p.fired); + goto fail_init; + } + + tw_fini(); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_init: + tw_fini(); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_tw_cancel_unposted_is_noop(void) +{ + struct tw_entry e; + + TEST_START(); + + if (tw_init() < 0) + goto fail; + + tw_init_entry(&e); + tw_cancel(&e); + tw_cancel(&e); + + tw_fini(); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_tw_fire_only_once(void) +{ + struct payload p; + + TEST_START(); + + if (tw_init() < 0) + goto fail; + + tw_init_entry(&p.tw); + p.fired = 0; + + tw_post(&p.tw, now_ns() + 3 * MILLION, cb_count, &p); + + sleep_ns(20 * MILLION); + tw_move(); + tw_move(); + tw_move(); + + if (p.fired != 1) { + printf("expected 1 fire, got %d after 3 moves\n", p.fired); + goto fail_post; + } + + tw_cancel(&p.tw); + tw_fini(); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_post: + tw_cancel(&p.tw); + tw_fini(); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* Multi-level: post a level-1 (>= 256ms) deadline; should still fire. */ +static int test_tw_post_level1_fires(void) +{ + struct payload p; + + TEST_START(); + + if (tw_init() < 0) + goto fail; + + tw_init_entry(&p.tw); + p.fired = 0; + + tw_post(&p.tw, now_ns() + 300 * MILLION, cb_count, &p); + + if (p.tw.lvl != 1) { + printf("expected level 1 placement, got %zu\n", p.tw.lvl); + goto fail_post; + } + + sleep_ns(320 * MILLION); + tw_move(); + + if (p.fired != 1) { + printf("level-1 entry didn't fire (got %d)\n", p.fired); + goto fail_post; + } + + tw_cancel(&p.tw); + tw_fini(); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_post: + tw_cancel(&p.tw); + tw_fini(); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +static int test_tw_many_entries_all_fire(void) +{ + struct payload pl[16]; + size_t i; + size_t total = 0; + + TEST_START(); + + if (tw_init() < 0) + goto fail; + + for (i = 0; i < 16; ++i) { + tw_init_entry(&pl[i].tw); + pl[i].fired = 0; + tw_post(&pl[i].tw, now_ns() + (1 + i) * MILLION, + cb_count, &pl[i]); + } + + sleep_ns(40 * MILLION); + tw_move(); + + for (i = 0; i < 16; ++i) + total += pl[i].fired; + + if (total != 16) { + printf("expected 16 fires, got %zu\n", total); + goto fail_post; + } + + for (i = 0; i < 16; ++i) + tw_cancel(&pl[i].tw); + tw_fini(); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_post: + for (i = 0; i < 16; ++i) + tw_cancel(&pl[i].tw); + tw_fini(); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* tw_next_expiry signals empty wheel via tv_nsec == -1. */ +static int test_tw_next_expiry_empty(void) +{ + struct timespec out; + + TEST_START(); + + if (tw_init() < 0) + goto fail; + + tw_next_expiry(&out); + if (out.tv_nsec != -1) { + printf("expected tv_nsec=-1, got %ld\n", (long) out.tv_nsec); + goto fail_init; + } + + tw_fini(); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_init: + tw_fini(); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* tw_next_expiry returns a deadline within the right ballpark. */ +static int test_tw_next_expiry_returns_deadline(void) +{ + struct payload p; + struct timespec out; + uint64_t target; + uint64_t out_ns; + int64_t skew; + + TEST_START(); + + if (tw_init() < 0) + goto fail; + + tw_init_entry(&p.tw); + p.fired = 0; + + target = now_ns() + 50 * MILLION; + tw_post(&p.tw, target, cb_count, &p); + + tw_next_expiry(&out); + out_ns = TS_TO_UINT64(out); + + /* Level-0 quantization gives ±1 slot of skew. */ + skew = (int64_t)(out_ns) - (int64_t)(target); + if (skew < -2 * MILLION || skew > 4 * MILLION) { + printf("deadline not in -2..+4 ms, skew=%ld ns\n", (long) skew); + goto fail_post; + } + + tw_cancel(&p.tw); + tw_fini(); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_post: + tw_cancel(&p.tw); + tw_fini(); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* Repost: fire, then post again. */ +static int test_tw_repost_after_fire(void) +{ + struct payload p; + + TEST_START(); + + if (tw_init() < 0) + goto fail; + + tw_init_entry(&p.tw); + p.fired = 0; + + tw_post(&p.tw, now_ns() + 3 * MILLION, cb_count, &p); + sleep_ns(20 * MILLION); + tw_move(); + if (p.fired != 1) { + printf("first fire missed\n"); + goto fail_post; + } + + tw_post(&p.tw, now_ns() + 3 * MILLION, cb_count, &p); + sleep_ns(20 * MILLION); + tw_move(); + if (p.fired != 2) { + printf("second fire missed (fired=%d)\n", p.fired); + goto fail_post; + } + + tw_cancel(&p.tw); + tw_fini(); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_post: + tw_cancel(&p.tw); + tw_fini(); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* Double-post replaces the schedule; only the second fires. */ +static int test_tw_double_post_replaces(void) +{ + struct payload p; + + TEST_START(); + + if (tw_init() < 0) + goto fail; + + tw_init_entry(&p.tw); + p.fired = 0; + + tw_post(&p.tw, now_ns() + 30 * MILLION, cb_count, &p); + tw_post(&p.tw, now_ns() + 3 * MILLION, cb_count, &p); + + sleep_ns(20 * MILLION); + tw_move(); + + if (p.fired != 1) { + printf("expected 1 fire after replace, got %d\n", p.fired); + goto fail_post; + } + + sleep_ns(40 * MILLION); + tw_move(); + + if (p.fired != 1) { + printf("first schedule fired after replace (got %d)\n", + p.fired); + goto fail_post; + } + + tw_cancel(&p.tw); + tw_fini(); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_post: + tw_cancel(&p.tw); + tw_fini(); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* Fire callback may safely cancel a sibling in the same slot. */ +static int test_tw_fire_cancels_sibling(void) +{ + struct cancel_payload a; + struct payload b; + uint64_t deadline; + + TEST_START(); + + if (tw_init() < 0) + goto fail; + + tw_init_entry(&a.tw); + tw_init_entry(&b.tw); + a.fired = 0; + a.sibling = &b.tw; + b.fired = 0; + + deadline = now_ns() + 3 * MILLION; + tw_post(&a.tw, deadline, cb_cancel_sibling, &a); + tw_post(&b.tw, deadline, cb_count, &b); + + sleep_ns(20 * MILLION); + tw_move(); + + if (a.fired != 1) { + printf("a expected 1 fire, got %d\n", a.fired); + goto fail_post; + } + if (b.fired != 0) { + printf("b should not have fired (got %d)\n", b.fired); + goto fail_post; + } + + tw_cancel(&a.tw); + tw_cancel(&b.tw); + tw_fini(); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_post: + tw_cancel(&a.tw); + tw_cancel(&b.tw); + tw_fini(); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +/* Fire callback may safely repost a sibling to a future slot. */ +static int test_tw_fire_posts_sibling(void) +{ + struct repost_payload a; + struct payload b; + uint64_t deadline; + + TEST_START(); + + if (tw_init() < 0) + goto fail; + + tw_init_entry(&a.tw); + tw_init_entry(&b.tw); + a.fired = 0; + a.sibling = &b; + a.repost_at = now_ns() + 30 * MILLION; + b.fired = 0; + + deadline = now_ns() + 3 * MILLION; + tw_post(&a.tw, deadline, cb_repost_sibling, &a); + tw_post(&b.tw, deadline, cb_count, &b); + + sleep_ns(20 * MILLION); + tw_move(); + + if (a.fired != 1) { + printf("a expected 1 fire, got %d\n", a.fired); + goto fail_post; + } + if (b.fired != 0) { + printf("b fired before reposted deadline (got %d)\n", + b.fired); + goto fail_post; + } + + sleep_ns(25 * MILLION); + tw_move(); + + if (b.fired != 1) { + printf("b expected 1 fire after repost, got %d\n", + b.fired); + goto fail_post; + } + + tw_cancel(&a.tw); + tw_cancel(&b.tw); + tw_fini(); + + TEST_SUCCESS(); + + return TEST_RC_SUCCESS; + fail_post: + tw_cancel(&a.tw); + tw_cancel(&b.tw); + tw_fini(); + fail: + TEST_FAIL(); + return TEST_RC_FAIL; +} + +int tw_test(int argc, + char ** argv) +{ + int ret = 0; + + (void) argc; + (void) argv; + + ret |= test_tw_init_fini(); + ret |= test_tw_post_fires_after_deadline(); + ret |= test_tw_no_fire_before_deadline(); + ret |= test_tw_cancel_prevents_fire(); + ret |= test_tw_cancel_unposted_is_noop(); + ret |= test_tw_fire_only_once(); + ret |= test_tw_post_level1_fires(); + ret |= test_tw_many_entries_all_fire(); + ret |= test_tw_next_expiry_empty(); + ret |= test_tw_next_expiry_returns_deadline(); + ret |= test_tw_repost_after_fire(); + ret |= test_tw_double_post_replaces(); + ret |= test_tw_fire_cancels_sibling(); + ret |= test_tw_fire_posts_sibling(); + + return ret; +} diff --git a/src/lib/timerwheel.c b/src/lib/timerwheel.c deleted file mode 100644 index 2c796c96..00000000 --- a/src/lib/timerwheel.c +++ /dev/null @@ -1,414 +0,0 @@ -/* - * Ouroboros - Copyright (C) 2016 - 2026 - * - * Timerwheel - * - * Dimitri Staessens <dimitri@ouroboros.rocks> - * Sander Vrijders <sander@ouroboros.rocks> - * - * This library is free software; you can redistribute it and/or - * modify it under the terms of the GNU Lesser General Public License - * version 2.1 as published by the Free Software Foundation. - * - * This library is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU - * Lesser General Public License for more details. - * - * You should have received a copy of the GNU Lesser General Public - * License along with this library; if not, write to the Free Software - * Foundation, Inc., http://www.fsf.org/about/contact/. - */ - -#include <ouroboros/list.h> - -/* Overflow limits range to about 6 hours. */ -#define ts_to_ns(ts) (ts.tv_sec * BILLION + ts.tv_nsec) -#define ts_to_rxm_slot(ts) (ts_to_ns(ts) >> RXMQ_RES) -#define ts_to_ack_slot(ts) (ts_to_ns(ts) >> ACKQ_RES) - -struct rxm { - struct list_head next; - uint32_t seqno; -#ifndef RXM_BUFFER_ON_HEAP - struct ssm_pk_buff * spb; -#endif - struct frct_pci * pkt; - size_t len; - time_t t0; /* Time when original was sent (us). */ - struct frcti * frcti; - int fd; - int flow_id; /* Prevent rtx when fd reused. */ -}; - -struct ack { - struct list_head next; - struct frcti * frcti; - int fd; - int flow_id; -}; - -struct { - /* - * At a 1 ms min resolution, every level bumps the - * resolution by a factor of 16. - */ - struct list_head rxms[RXMQ_LVLS][RXMQ_SLOTS]; - - struct list_head acks[ACKQ_SLOTS]; - bool map[ACKQ_SLOTS][PROG_MAX_FLOWS]; - - size_t prv_rxm[RXMQ_LVLS]; /* Last processed rxm slots. */ - size_t prv_ack; /* Last processed ack slot. */ - pthread_mutex_t lock; -} rw; - -static void timerwheel_fini(void) -{ - size_t i; - size_t j; - struct list_head * p; - struct list_head * h; - - pthread_mutex_lock(&rw.lock); - - for (i = 0; i < RXMQ_LVLS; ++i) { - for (j = 0; j < RXMQ_SLOTS; j++) { - list_for_each_safe(p, h, &rw.rxms[i][j]) { - struct rxm * rxm; - rxm = list_entry(p, struct rxm, next); - list_del(&rxm->next); -#ifdef RXM_BUFFER_ON_HEAP - free(rxm->pkt); -#else - ssm_pk_buff_ack(rxm->spb); - ipcp_spb_release(rxm->spb); -#endif - free(rxm); - } - } - } - - for (i = 0; i < ACKQ_SLOTS; ++i) { - list_for_each_safe(p, h, &rw.acks[i]) { - struct ack * a = list_entry(p, struct ack, next); - list_del(&a->next); - free(a); - } - } - - pthread_mutex_unlock(&rw.lock); - - pthread_mutex_destroy(&rw.lock); -} - -static int timerwheel_init(void) -{ - struct timespec now; - size_t i; - size_t j; - - if (pthread_mutex_init(&rw.lock, NULL)) - return -1; - - clock_gettime(PTHREAD_COND_CLOCK, &now); - - for (i = 0; i < RXMQ_LVLS; ++i) { - rw.prv_rxm[i] = (ts_to_rxm_slot(now) - 1); - rw.prv_rxm[i] >>= (RXMQ_BUMP * i); - rw.prv_rxm[i] &= (RXMQ_SLOTS - 1); - for (j = 0; j < RXMQ_SLOTS; ++j) - list_head_init(&rw.rxms[i][j]); - } - - rw.prv_ack = (ts_to_ack_slot(now) - 1) & (ACKQ_SLOTS - 1); - for (i = 0; i < ACKQ_SLOTS; ++i) - list_head_init(&rw.acks[i]); - - return 0; -} - -static void timerwheel_move(void) -{ - struct timespec now; - struct list_head * p; - struct list_head * h; - size_t rxm_slot; - size_t ack_slot; - size_t i; - size_t j; - - pthread_mutex_lock(&rw.lock); - - pthread_cleanup_push(__cleanup_mutex_unlock, &rw.lock); - - clock_gettime(PTHREAD_COND_CLOCK, &now); - - rxm_slot = ts_to_rxm_slot(now); - - for (i = 0; i < RXMQ_LVLS; ++i) { - size_t j_max_slot = rxm_slot & (RXMQ_SLOTS - 1); - j = rw.prv_rxm[i]; - if (j_max_slot < j) - j_max_slot += RXMQ_SLOTS; - while (j++ < j_max_slot) { - list_for_each_safe(p, h, - &rw.rxms[i][j & (RXMQ_SLOTS - 1)]) { - struct rxm * r; - struct frct_cr * snd_cr; - struct frct_cr * rcv_cr; - size_t slot; - size_t rslot; - ssize_t idx; - struct ssm_pk_buff * spb; - struct frct_pci * pci; - struct flow * f; - uint32_t snd_lwe; - uint32_t rcv_lwe; - size_t lvl = 0; - - r = list_entry(p, struct rxm, next); - - list_del(&r->next); - - snd_cr = &r->frcti->snd_cr; - rcv_cr = &r->frcti->rcv_cr; - f = &proc.flows[r->fd]; -#ifndef RXM_BUFFER_ON_HEAP - ssm_pk_buff_ack(r->spb); -#endif - if (f->frcti == NULL - || f->info.id != r->flow_id) - goto cleanup; - - pthread_rwlock_rdlock(&r->frcti->lock); - - snd_lwe = snd_cr->lwe; - rcv_lwe = rcv_cr->lwe; - - pthread_rwlock_unlock(&r->frcti->lock); - - /* Has been ack'd, remove. */ - if (before(r->seqno, snd_lwe)) - goto cleanup; - - /* Check for r-timer expiry. */ - if (ts_to_ns(now) - r->t0 > r->frcti->r) - goto flow_down; - - pthread_rwlock_wrlock(&r->frcti->lock); - - if (r->seqno == r->frcti->rttseq) { - r->frcti->rto += - r->frcti->rto >> RTO_DIV; - r->frcti->probe = false; - } -#ifdef PROC_FLOW_STATS - r->frcti->n_rtx++; -#endif - rslot = r->frcti->rto >> RXMQ_RES; - - pthread_rwlock_unlock(&r->frcti->lock); - - /* Schedule at least in the next time slot. */ - slot = ts_to_ns(now) >> RXMQ_RES; - - while (rslot >= RXMQ_SLOTS) { - ++lvl; - rslot >>= RXMQ_BUMP; - slot >>= RXMQ_BUMP; - } - - if (lvl >= RXMQ_LVLS) /* Can't reschedule */ - goto flow_down; - - rslot = (rslot + slot + 1) & (RXMQ_SLOTS - 1); -#ifdef RXM_BLOCKING - if (ipcp_spb_reserve(&spb, r->len) < 0) -#else - if (ssm_pool_alloc(proc.pool, r->len, NULL, - &spb) < 0) -#endif - goto reschedule; /* rdrbuff full */ - - pci = (struct frct_pci *) ssm_pk_buff_head(spb); - memcpy(pci, r->pkt, r->len); -#ifndef RXM_BUFFER_ON_HEAP - ipcp_spb_release(r->spb); - r->spb = spb; - r->pkt = pci; - ssm_pk_buff_wait_ack(spb); -#endif - idx = ssm_pk_buff_get_idx(spb); - - /* Retransmit the copy. */ - pci->ackno = hton32(rcv_lwe); -#ifdef RXM_BLOCKING - if (ssm_rbuff_write_b(f->tx_rb, idx, NULL) < 0) -#else - if (ssm_rbuff_write(f->tx_rb, idx) < 0) -#endif - goto flow_down; - ssm_flow_set_notify(f->set, f->info.id, - FLOW_PKT); - reschedule: - list_add(&r->next, &rw.rxms[lvl][rslot]); - continue; - - flow_down: - ssm_rbuff_set_acl(f->tx_rb, ACL_FLOWDOWN); - ssm_rbuff_set_acl(f->rx_rb, ACL_FLOWDOWN); - cleanup: -#ifdef RXM_BUFFER_ON_HEAP - free(r->pkt); -#else - ipcp_spb_release(r->spb); -#endif - free(r); - } - } - rw.prv_rxm[i] = rxm_slot & (RXMQ_SLOTS - 1); - /* Move up a level in the wheel. */ - rxm_slot >>= RXMQ_BUMP; - } - - ack_slot = ts_to_ack_slot(now) & (ACKQ_SLOTS - 1) ; - - j = rw.prv_ack; - - if (ack_slot < j) - ack_slot += ACKQ_SLOTS; - - while (j++ < ack_slot) { - list_for_each_safe(p, h, &rw.acks[j & (ACKQ_SLOTS - 1)]) { - struct ack * a; - struct flow * f; - - a = list_entry(p, struct ack, next); - - list_del(&a->next); - - f = &proc.flows[a->fd]; - - rw.map[j & (ACKQ_SLOTS - 1)][a->fd] = false; - - if (f->info.id == a->flow_id && f->frcti != NULL) - send_frct_pkt(a->frcti); - - free(a); - } - } - - rw.prv_ack = ack_slot & (ACKQ_SLOTS - 1); - - pthread_cleanup_pop(true); -} - -static int timerwheel_rxm(struct frcti * frcti, - uint32_t seqno, - struct ssm_pk_buff * spb) -{ - struct timespec now; - struct rxm * r; - size_t slot; - size_t lvl = 0; - time_t rto_slot; - - r = malloc(sizeof(*r)); - if (r == NULL) - return -ENOMEM; - - clock_gettime(PTHREAD_COND_CLOCK, &now); - - r->t0 = ts_to_ns(now); - r->seqno = seqno; - r->frcti = frcti; - r->len = ssm_pk_buff_len(spb); -#ifdef RXM_BUFFER_ON_HEAP - r->pkt = malloc(r->len); - if (r->pkt == NULL) { - free(r); - return -ENOMEM; - } - memcpy(r->pkt, ssm_pk_buff_head(spb), r->len); -#else - r->spb = spb; - r->pkt = (struct frct_pci *) ssm_pk_buff_head(spb); -#endif - pthread_rwlock_rdlock(&r->frcti->lock); - - rto_slot = frcti->rto >> RXMQ_RES; - slot = r->t0 >> RXMQ_RES; - - r->fd = frcti->fd; - r->flow_id = proc.flows[r->fd].info.id; - - pthread_rwlock_unlock(&r->frcti->lock); - - while (rto_slot >= RXMQ_SLOTS) { - ++lvl; - rto_slot >>= RXMQ_BUMP; - slot >>= RXMQ_BUMP; - } - - if (lvl >= RXMQ_LVLS) { /* Out of timerwheel range. */ -#ifdef RXM_BUFFER_ON_HEAP - free(r->pkt); -#endif - free(r); - return -EPERM; - } - - slot = (slot + rto_slot + 1) & (RXMQ_SLOTS - 1); - - pthread_mutex_lock(&rw.lock); - - list_add_tail(&r->next, &rw.rxms[lvl][slot]); -#ifndef RXM_BUFFER_ON_HEAP - ssm_pk_buff_wait_ack(spb); -#endif - pthread_mutex_unlock(&rw.lock); - - return 0; -} - -static int timerwheel_delayed_ack(int fd, - struct frcti * frcti) -{ - struct timespec now; - struct ack * a; - size_t slot; - - a = malloc(sizeof(*a)); - if (a == NULL) - return -ENOMEM; - - clock_gettime(PTHREAD_COND_CLOCK, &now); - - pthread_rwlock_rdlock(&frcti->lock); - - slot = (((ts_to_ns(now) + (TICTIME << 1)) >> ACKQ_RES) + 1) - & (ACKQ_SLOTS - 1); - - pthread_rwlock_unlock(&frcti->lock); - - a->fd = fd; - a->frcti = frcti; - a->flow_id = proc.flows[fd].info.id; - - pthread_mutex_lock(&rw.lock); - - if (rw.map[slot][fd]) { - pthread_mutex_unlock(&rw.lock); - free(a); - return 0; - } - - rw.map[slot][fd] = true; - - list_add_tail(&a->next, &rw.acks[slot]); - - pthread_mutex_unlock(&rw.lock); - - return 0; -} diff --git a/src/lib/tw.c b/src/lib/tw.c new file mode 100644 index 00000000..ccde7dd1 --- /dev/null +++ b/src/lib/tw.c @@ -0,0 +1,307 @@ +/* + * Ouroboros - Copyright (C) 2016 - 2026 + * + * Generic deadline-ordered callback queue (timing wheel) + * + * Dimitri Staessens <dimitri@ouroboros.rocks> + * Sander Vrijders <sander@ouroboros.rocks> + * + * This library is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public License + * version 2.1 as published by the Free Software Foundation. + * + * This library is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with this library; if not, write to the Free Software + * Foundation, Inc., http://www.fsf.org/about/contact/. + */ + +#if defined(__linux__) || defined(__CYGWIN__) +#define _DEFAULT_SOURCE +#else +#define _POSIX_C_SOURCE 200809L +#endif + +#include "config.h" + +#include <ouroboros/list.h> +#include <ouroboros/pthread.h> +#include <ouroboros/time.h> +#include <ouroboros/tw.h> + +#include <assert.h> +#include <stdbool.h> +#include <stdint.h> + +/* 3 levels × 256 slots, 1 ms / 16 ms / 256 ms per-slot resolution. */ +#define TW_LVLS 3 +#define TW_SLOTS 256 +#define TW_BUMP 4 +#define TW_RES 20 /* 2^20 ns ≈ 1 ms per slot at level 0. */ + +#define TW_SLOT(x) ((x) & (TW_SLOTS - 1)) + +static struct { + struct list_head levels[TW_LVLS][TW_SLOTS]; + size_t prv[TW_LVLS]; + pthread_mutex_t mtx; + pthread_mutex_t move_mtx; + bool initialised; +} tw; + +static size_t tw_lvl_res(size_t lvl) +{ + return TW_RES + TW_BUMP * lvl; +} + +/* Smallest level whose slot range covers the deadline. */ +static size_t tw_pick_lvl(uint64_t now_ns, + uint64_t deadline_ns) +{ + uint64_t delta; + size_t lvl; + + delta = deadline_ns > now_ns ? deadline_ns - now_ns : 0; + lvl = 0; + + while (lvl < TW_LVLS - 1 && (delta >> tw_lvl_res(lvl)) >= TW_SLOTS) + ++lvl; + + return lvl; +} + +static size_t tw_slot(uint64_t ns, + size_t lvl) +{ + return TW_SLOT(ns >> tw_lvl_res(lvl)); +} + +int tw_init(void) +{ + struct timespec now; + size_t i; + size_t j; + + assert(!tw.initialised); + + if (pthread_mutex_init(&tw.mtx, NULL)) + goto fail_mtx; + + if (pthread_mutex_init(&tw.move_mtx, NULL)) + goto fail_move_mtx; + + clock_gettime(PTHREAD_COND_CLOCK, &now); + + for (i = 0; i < TW_LVLS; ++i) { + tw.prv[i] = TW_SLOT(tw_slot(TS_TO_UINT64(now), i) - 1); + for (j = 0; j < TW_SLOTS; ++j) + list_head_init(&tw.levels[i][j]); + } + + tw.initialised = true; + + return 0; + + fail_move_mtx: + pthread_mutex_destroy(&tw.mtx); + fail_mtx: + return -1; +} + +void tw_fini(void) +{ + size_t i; + size_t j; + + assert(tw.initialised); + + for (i = 0; i < TW_LVLS; ++i) { + for (j = 0; j < TW_SLOTS; ++j) + assert(list_is_empty(&tw.levels[i][j])); + } + + pthread_mutex_destroy(&tw.move_mtx); + pthread_mutex_destroy(&tw.mtx); + + tw.initialised = false; +} + +void tw_init_entry(struct tw_entry * e) +{ + list_head_init(&e->next); + + e->deadline_ns = 0; + e->fire = NULL; + e->arg = NULL; + e->lvl = 0; +} + +void tw_post(struct tw_entry * e, + uint64_t deadline_ns, + tw_fire_fn_t fire, + void * arg) +{ + struct timespec now; + size_t lvl; + size_t slot; + + assert(tw.initialised); + + clock_gettime(PTHREAD_COND_CLOCK, &now); + + lvl = tw_pick_lvl(TS_TO_UINT64(now), deadline_ns); + /* +1 so deadline <= slot_start; lands later in slot. */ + slot = TW_SLOT(tw_slot(deadline_ns, lvl) + 1); + + e->deadline_ns = deadline_ns; + e->fire = fire; + e->arg = arg; + e->lvl = lvl; + + pthread_mutex_lock(&tw.mtx); + + if (!list_is_empty(&e->next)) + list_del(&e->next); + + list_add_tail(&e->next, &tw.levels[lvl][slot]); + + pthread_mutex_unlock(&tw.mtx); +} + +void tw_cancel(struct tw_entry * e) +{ + if (e == NULL) + return; + + assert(tw.initialised); + + pthread_mutex_lock(&tw.mtx); + + if (!list_is_empty(&e->next)) { + list_del(&e->next); + list_head_init(&e->next); + } + + pthread_mutex_unlock(&tw.mtx); +} + +void tw_move(void) +{ + struct timespec now; + struct list_head deferred; + struct list_head * p; + uint64_t now_ns; + size_t i; + size_t j; + size_t cur; + + assert(tw.initialised); + + if (pthread_mutex_trylock(&tw.move_mtx) != 0) + return; + + pthread_cleanup_push(__cleanup_mutex_unlock, &tw.move_mtx); + + pthread_mutex_lock(&tw.mtx); + + pthread_cleanup_push(__cleanup_mutex_unlock, &tw.mtx); + + clock_gettime(PTHREAD_COND_CLOCK, &now); + now_ns = TS_TO_UINT64(now); + + for (i = 0; i < TW_LVLS; ++i) { + cur = tw_slot(now_ns, i); + + j = tw.prv[i]; + if (cur < j) + cur += TW_SLOTS; + + while (j++ < cur) { + size_t s = TW_SLOT(j); + + /* Pop-front so fire may mutate any entry. */ + list_head_init(&deferred); + + while (!list_is_empty(&tw.levels[i][s])) { + struct tw_entry * e; + p = tw.levels[i][s].nxt; + e = list_entry(p, struct tw_entry, next); + list_del(&e->next); + + if (e->deadline_ns > now_ns) { + list_add_tail(&e->next, &deferred); + continue; + } + + pthread_mutex_unlock(&tw.mtx); + e->fire(e->arg); + pthread_mutex_lock(&tw.mtx); + } + + while (!list_is_empty(&deferred)) { + p = deferred.nxt; + list_del(p); + list_add_tail(p, &tw.levels[i][s]); + } + } + + tw.prv[i] = TW_SLOT(cur); + } + + pthread_cleanup_pop(true); /* tw.mtx */ + pthread_cleanup_pop(true); /* tw.move_mtx */ +} + +/* Earliest pending deadline at level lvl, INT64_MAX if level is empty. */ +static int64_t tw_lvl_earliest(size_t lvl, + uint64_t now_ns) +{ + size_t cur = tw_slot(now_ns, lvl); + size_t j; + + for (j = 1; j <= TW_SLOTS; ++j) { + size_t s = TW_SLOT(cur + j); + + if (list_is_empty(&tw.levels[lvl][s])) + continue; + + return (int64_t)(now_ns + ((uint64_t) j << tw_lvl_res(lvl))); + } + + return INT64_MAX; +} + +void tw_next_expiry(struct timespec * out) +{ + struct timespec now; + uint64_t now_ns; + int64_t earliest = INT64_MAX; + size_t i; + + assert(tw.initialised); + + clock_gettime(PTHREAD_COND_CLOCK, &now); + now_ns = TS_TO_UINT64(now); + + pthread_mutex_lock(&tw.mtx); + + for (i = 0; i < TW_LVLS; ++i) { + int64_t dl = tw_lvl_earliest(i, now_ns); + if (dl < earliest) + earliest = dl; + } + + pthread_mutex_unlock(&tw.mtx); + + if (earliest == INT64_MAX) { + /* Empty wheel: tv_nsec=-1 is an invalid normalised value. */ + out->tv_sec = 0; + out->tv_nsec = -1; + } else { + UINT64_TO_TS((uint64_t) earliest, out); + } +} |