diff options
Diffstat (limited to 'link')
| -rw-r--r-- | link/crypto-openssl.c | 284 | ||||
| -rw-r--r-- | link/crypto.c | 167 | ||||
| -rw-r--r-- | link/crypto.h | 49 | ||||
| -rw-r--r-- | link/link.c | 1862 | ||||
| -rw-r--r-- | link/link.h | 118 | ||||
| -rw-r--r-- | link/seal.h | 80 | ||||
| -rw-r--r-- | link/tls.h | 18 |
7 files changed, 2578 insertions, 0 deletions
diff --git a/link/crypto-openssl.c b/link/crypto-openssl.c new file mode 100644 index 0000000..6bdf6fb --- /dev/null +++ b/link/crypto-openssl.c @@ -0,0 +1,284 @@ +#define _GNU_SOURCE +/* pure-crypto.so: what pure-serve asks of OpenSSL (link/crypto.h), mapped + * only when a server seals, opens a seal or reaches an https receiver. It + * exports one table, pure_crypto. */ +#include "crypto.h" +#include <arpa/inet.h> +#include <errno.h> +#include <fcntl.h> +#include <openssl/crypto.h> +#include <openssl/evp.h> +#include <openssl/core_names.h> +#include <openssl/hpke.h> +#include <openssl/kdf.h> +#include <openssl/ssl.h> +#include <openssl/x509v3.h> +#include <limits.h> +#include <pthread.h> +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <sys/stat.h> +#include <unistd.h> + +/* ---- TLS: one client context, verifying the receiver's certificate against + * the system's CAs or the file given, and its name. */ +static SSL_CTX *tls_context; + +static bool tls_start(const char *ca_file) { + SSL_CTX *context = SSL_CTX_new(TLS_client_method()); + if (!context) return false; + if (!SSL_CTX_set_min_proto_version(context, TLS1_2_VERSION) || + !(ca_file ? SSL_CTX_load_verify_locations(context, ca_file, NULL) : SSL_CTX_set_default_verify_paths(context))) { + SSL_CTX_free(context); + return false; + } + SSL_CTX_set_verify(context, SSL_VERIFY_PEER, NULL); + SSL_CTX_set_mode(context, SSL_MODE_AUTO_RETRY); + tls_context = context; + return true; +} + +/* TLS over a connected socket: the receiver's certificate must name the host + * (a DNS name, or an IP address written as one). */ +static PureTls *tls_open(const char *host, int fd) { + SSL *ssl = SSL_new(tls_context); + if (!ssl) return NULL; + unsigned char address[16]; + bool literal = inet_pton(AF_INET, host, address) == 1 || inet_pton(AF_INET6, host, address) == 1; + bool named = literal ? X509_VERIFY_PARAM_set1_ip_asc(SSL_get0_param(ssl), host) == 1 : + SSL_set_tlsext_host_name(ssl, host) == 1 && SSL_set1_host(ssl, host) == 1; + if (!named || SSL_set_fd(ssl, fd) != 1 || SSL_connect(ssl) != 1 || SSL_get_verify_result(ssl) != X509_V_OK) { + SSL_free(ssl); + return NULL; + } + return (PureTls *)ssl; +} + +static int tls_read(PureTls *tls, void *bytes, int size) { return SSL_read((SSL *)tls, bytes, size); } +static int tls_write(PureTls *tls, const void *bytes, int size) { return SSL_write((SSL *)tls, bytes, size); } +static void tls_close(PureTls *tls) { SSL_free((SSL *)tls); } + +/* ---- Sealing: HPKE (RFC 9180), X25519 and HKDF-SHA256, with the AEAD the + * suite names; auth mode when the sender seals with a key of its own. A + * session is one encapsulation, its seals numbered by the context. */ +struct PureSealKey { EVP_PKEY *key; }; +struct PureSealer { OSSL_HPKE_CTX *context; }; +struct PureOpener { OSSL_HPKE_CTX *context; }; + +static const unsigned char info[] = "pure sealed send 2"; + +static bool suite_named(const char *name, OSSL_HPKE_SUITE *out) { + OSSL_HPKE_SUITE suite = {OSSL_HPKE_KEM_ID_X25519, OSSL_HPKE_KDF_ID_HKDF_SHA256, 0}; + if (!strcmp(name, "x25519-sha256-aes128gcm")) suite.aead_id = OSSL_HPKE_AEAD_ID_AES_GCM_128; + else if (!strcmp(name, "x25519-sha256-chacha20poly1305")) suite.aead_id = OSSL_HPKE_AEAD_ID_CHACHA_POLY1305; + else return false; + *out = suite; + return true; +} + +static pthread_once_t secure_once = PTHREAD_ONCE_INIT; +static void secure_heap(void) { CRYPTO_secure_malloc_init(16384, 32); } + +static PureSealKey *seal_key(const char *path, char public_hex[2 * PURE_SEAL_KEY + 1], char *error, size_t size) { + pthread_once(&secure_once, secure_heap); + unsigned char *raw = OPENSSL_secure_malloc(PURE_SEAL_KEY), public_key[PURE_SEAL_KEY]; + PureSealKey *made = calloc(1, sizeof(*made)); + EVP_PKEY *key = NULL; + size_t length = PURE_SEAL_KEY; + if (!raw || !made) { snprintf(error, size, "out of memory"); goto fail; } + int fd = open(path, O_RDONLY | O_NOFOLLOW | O_CLOEXEC); + struct stat info_; + if (fd >= 0) { + bool ok = !fstat(fd, &info_) && S_ISREG(info_.st_mode) && !(info_.st_mode & 077) && + read(fd, raw, PURE_SEAL_KEY) == PURE_SEAL_KEY && info_.st_size == PURE_SEAL_KEY; + close(fd); + if (!ok) { snprintf(error, size, "%s: a key is 32 bytes that only its owner may read (0600)", path); goto fail; } + key = EVP_PKEY_new_raw_private_key(EVP_PKEY_X25519, NULL, raw, PURE_SEAL_KEY); + } else if (errno == ENOENT) { + /* A new key: made, written whole and synced before it is used. */ + key = EVP_PKEY_Q_keygen(NULL, NULL, "X25519"); + fd = open(path, O_WRONLY | O_CREAT | O_EXCL | O_NOFOLLOW | O_CLOEXEC, 0600); + bool ok = key && fd >= 0 && EVP_PKEY_get_raw_private_key(key, raw, &length) == 1 && length == PURE_SEAL_KEY && + write(fd, raw, PURE_SEAL_KEY) == PURE_SEAL_KEY && !fsync(fd); + if (fd >= 0 && close(fd)) ok = false; + if (!ok) { snprintf(error, size, "%s: cannot make a key there", path); unlink(path); goto fail; } + } else { snprintf(error, size, "%s: %s", path, strerror(errno)); goto fail; } + length = PURE_SEAL_KEY; + if (!key || EVP_PKEY_get_raw_public_key(key, public_key, &length) != 1 || length != PURE_SEAL_KEY) { + snprintf(error, size, "%s: not an X25519 key", path); + goto fail; + } + for (size_t i = 0; i < PURE_SEAL_KEY; ++i) snprintf(public_hex + 2 * i, 3, "%02x", public_key[i]); + OPENSSL_secure_clear_free(raw, PURE_SEAL_KEY); + made->key = key; + return made; +fail: + EVP_PKEY_free(key); + if (raw) OPENSSL_secure_clear_free(raw, PURE_SEAL_KEY); + free(made); + return NULL; +} + +static void seal_key_free(PureSealKey *key) { + if (!key) return; + EVP_PKEY_free(key->key); + free(key); +} + +static PureSealer *sealer(const char *suite_name, const unsigned char peer[PURE_SEAL_KEY], const PureSealKey *own, + unsigned char enc[PURE_SEAL_KEY]) { + OSSL_HPKE_SUITE suite; + PureSealer *made = suite_named(suite_name, &suite) ? calloc(1, sizeof(*made)) : NULL; + if (!made) return NULL; + made->context = OSSL_HPKE_CTX_new(own ? OSSL_HPKE_MODE_AUTH : OSSL_HPKE_MODE_BASE, suite, OSSL_HPKE_ROLE_SENDER, + NULL, NULL); + size_t length = PURE_SEAL_KEY; + if (!made->context || (own && OSSL_HPKE_CTX_set1_authpriv(made->context, own->key) != 1) || + OSSL_HPKE_encap(made->context, enc, &length, peer, PURE_SEAL_KEY, info, sizeof(info) - 1) != 1 || + length != PURE_SEAL_KEY) { + OSSL_HPKE_CTX_free(made->context); + free(made); + return NULL; + } + return made; +} + +static bool seal(PureSealer *s, const char *aad, const unsigned char *plain, size_t size, unsigned char *out, + uint64_t *n) { + size_t sealed = size + PURE_SEAL_OVERHEAD; + return OSSL_HPKE_CTX_get_seq(s->context, n) == 1 && + OSSL_HPKE_seal(s->context, out, &sealed, (const unsigned char *)aad, strlen(aad), plain, size) == 1 && + sealed == size + PURE_SEAL_OVERHEAD; +} + +static void sealer_free(PureSealer *s) { + if (!s) return; + OSSL_HPKE_CTX_free(s->context); + free(s); +} + +static PureOpener *opener(const PureSealKey *key, const char *suite_name, const unsigned char enc[PURE_SEAL_KEY], + const unsigned char *from) { + OSSL_HPKE_SUITE suite; + PureOpener *made = key && suite_named(suite_name, &suite) ? calloc(1, sizeof(*made)) : NULL; + if (!made) return NULL; + made->context = OSSL_HPKE_CTX_new(from ? OSSL_HPKE_MODE_AUTH : OSSL_HPKE_MODE_BASE, suite, OSSL_HPKE_ROLE_RECEIVER, + NULL, NULL); + if (!made->context || (from && OSSL_HPKE_CTX_set1_authpub(made->context, from, PURE_SEAL_KEY) != 1) || + OSSL_HPKE_decap(made->context, enc, PURE_SEAL_KEY, key->key, info, sizeof(info) - 1) != 1) { + OSSL_HPKE_CTX_free(made->context); + free(made); + return NULL; + } + return made; +} + +/* A receiver may set the number (a sender may not: it would reuse a nonce). */ +static bool unseal(PureOpener *o, uint64_t n, const char *aad, const unsigned char *sealed, size_t size, + unsigned char *plain) { + size_t opened = size - PURE_SEAL_OVERHEAD; + return size >= PURE_SEAL_OVERHEAD && OSSL_HPKE_CTX_set_seq(o->context, n) == 1 && + OSSL_HPKE_open(o->context, plain, &opened, (const unsigned char *)aad, strlen(aad), sealed, size) == 1 && + opened == size - PURE_SEAL_OVERHEAD; +} + +static void opener_free(PureOpener *o) { + if (!o) return; + OSSL_HPKE_CTX_free(o->context); + free(o); +} + +/* ---- Links: X25519 between two nodes' keys, HKDF-SHA256, AES-128-GCM with + * the caller's nonce (one context sealing and one opening, each keeping its + * key's schedule). */ +struct PureAead { EVP_CIPHER_CTX *seal, *open; }; + +static bool link_secret(const PureSealKey *own, const unsigned char peer[PURE_SEAL_KEY], unsigned char out[32]) { + EVP_PKEY *theirs = EVP_PKEY_new_raw_public_key(EVP_PKEY_X25519, NULL, peer, PURE_SEAL_KEY); + EVP_PKEY_CTX *context = theirs && own ? EVP_PKEY_CTX_new(own->key, NULL) : NULL; + size_t length = 32; + bool ok = context && EVP_PKEY_derive_init(context) == 1 && EVP_PKEY_derive_set_peer(context, theirs) == 1 && + EVP_PKEY_derive(context, out, &length) == 1 && length == 32; + EVP_PKEY_CTX_free(context); + EVP_PKEY_free(theirs); + unsigned char any = 0; /* a small-order key makes zeros: no secret */ + for (size_t i = 0; ok && i < 32; ++i) any |= out[i]; + return ok && any; +} + +/* HKDF, fetched once. */ +static EVP_KDF *hkdf(void) { + static EVP_KDF *kept; + EVP_KDF *kdf = __atomic_load_n(&kept, __ATOMIC_ACQUIRE), *none = NULL; + if (kdf) return kdf; + if ((kdf = EVP_KDF_fetch(NULL, "HKDF", NULL)) && + !__atomic_compare_exchange_n(&kept, &none, kdf, false, __ATOMIC_ACQ_REL, __ATOMIC_ACQUIRE)) { + EVP_KDF_free(kdf); /* another thread's is kept */ + kdf = none; + } + return kdf; +} + +static bool link_derive(const unsigned char secret[32], const unsigned char *info, size_t info_size, unsigned char *out, + size_t out_size) { + static const char salt[] = "pure link 0"; + EVP_KDF *kdf = hkdf(); + EVP_KDF_CTX *context = kdf ? EVP_KDF_CTX_new(kdf) : NULL; + OSSL_PARAM params[] = {OSSL_PARAM_construct_utf8_string(OSSL_KDF_PARAM_DIGEST, (char *)"SHA256", 0), + OSSL_PARAM_construct_octet_string(OSSL_KDF_PARAM_KEY, (void *)secret, 32), + OSSL_PARAM_construct_octet_string(OSSL_KDF_PARAM_SALT, (void *)salt, sizeof(salt) - 1), + OSSL_PARAM_construct_octet_string(OSSL_KDF_PARAM_INFO, (void *)info, info_size), + OSSL_PARAM_construct_end()}; + bool ok = context && EVP_KDF_derive(context, out, out_size, params) == 1; + EVP_KDF_CTX_free(context); + return ok; +} + +static void aead_free(PureAead *a) { + if (!a) return; + EVP_CIPHER_CTX_free(a->seal); + EVP_CIPHER_CTX_free(a->open); + free(a); +} + +static PureAead *aead(const unsigned char key[PURE_AEAD_KEY]) { + PureAead *a = calloc(1, sizeof(*a)); + if (!a) return NULL; + a->seal = EVP_CIPHER_CTX_new(); + a->open = EVP_CIPHER_CTX_new(); + if (!a->seal || !a->open || EVP_EncryptInit_ex(a->seal, EVP_aes_128_gcm(), NULL, key, NULL) != 1 || + EVP_DecryptInit_ex(a->open, EVP_aes_128_gcm(), NULL, key, NULL) != 1) { + aead_free(a); + return NULL; + } + return a; +} + +static bool aead_seal(PureAead *a, const unsigned char *key, const unsigned char nonce[PURE_AEAD_NONCE], + const unsigned char *aad, size_t aad_size, const unsigned char *plain, size_t size, + unsigned char *out) { + int n = 0, last = 0; + return size <= INT_MAX && EVP_EncryptInit_ex(a->seal, NULL, NULL, key, nonce) == 1 && + (!aad_size || EVP_EncryptUpdate(a->seal, NULL, &n, aad, (int)aad_size) == 1) && + (!size || EVP_EncryptUpdate(a->seal, out, &n, plain, (int)size) == 1) && + EVP_EncryptFinal_ex(a->seal, out + size, &last) == 1 && + EVP_CIPHER_CTX_ctrl(a->seal, EVP_CTRL_GCM_GET_TAG, PURE_AEAD_TAG, out + size) == 1; +} + +static bool aead_open(PureAead *a, const unsigned char *key, const unsigned char nonce[PURE_AEAD_NONCE], + const unsigned char *aad, size_t aad_size, const unsigned char *sealed, size_t size, + unsigned char *plain) { + int n = 0, last = 0; + if (size < PURE_AEAD_TAG || size - PURE_AEAD_TAG > INT_MAX) return false; + size_t body = size - PURE_AEAD_TAG; + return EVP_DecryptInit_ex(a->open, NULL, NULL, key, nonce) == 1 && + (!aad_size || EVP_DecryptUpdate(a->open, NULL, &n, aad, (int)aad_size) == 1) && + (!body || EVP_DecryptUpdate(a->open, plain, &n, sealed, (int)body) == 1) && + EVP_CIPHER_CTX_ctrl(a->open, EVP_CTRL_GCM_SET_TAG, PURE_AEAD_TAG, (void *)(sealed + body)) == 1 && + EVP_DecryptFinal_ex(a->open, plain + body, &last) == 1; +} + +const PureCrypto pure_crypto = {PURE_CRYPTO_VERSION, tls_start, tls_open, tls_read, tls_write, tls_close, + seal_key, seal_key_free, sealer, seal, sealer_free, opener, unseal, opener_free, + link_secret, link_derive, aead, aead_seal, aead_open, aead_free}; diff --git a/link/crypto.c b/link/crypto.c new file mode 100644 index 0000000..2fa7d44 --- /dev/null +++ b/link/crypto.c @@ -0,0 +1,167 @@ +#define _GNU_SOURCE +/* Sealing and TLS for pure-serve (link/seal.h, link/tls.h), by way of + * pure-crypto.so (link/crypto.h): mapped, and OpenSSL with it, the first + * time a server needs it. What needs no OpenSSL (a key's hex, the suite + * this machine seals with) is here. */ +#include "crypto.h" +#include <dlfcn.h> +#include <pthread.h> +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <unistd.h> +#if defined(__aarch64__) && defined(__linux__) +#include <asm/hwcap.h> +#include <sys/auxv.h> +#endif + +static pthread_once_t once = PTHREAD_ONCE_INIT; +static const PureCrypto *crypto; +static char why[512]; + +/* pure-crypto.so beside the executable, or where PURE_CRYPTO names it. */ +static void load(void) { + char path[4096]; + const char *named = getenv("PURE_CRYPTO"); + if (named) snprintf(path, sizeof(path), "%s", named); + else { + ssize_t length = readlink("/proc/self/exe", path, sizeof(path) - 32); + char *slash = length > 0 ? memrchr(path, '/', (size_t)length) : NULL; + if (!slash) { snprintf(why, sizeof(why), "pure-crypto.so: cannot find pure-serve's directory"); return; } + strcpy(slash + 1, "pure-crypto.so"); + } + void *library = dlopen(path, RTLD_NOW | RTLD_LOCAL); + const PureCrypto *table = library ? dlsym(library, "pure_crypto") : NULL; + if (!table || table->version != PURE_CRYPTO_VERSION) { + snprintf(why, sizeof(why), "%s", !library ? dlerror() : "pure-crypto.so is not this pure-serve's"); + return; + } + crypto = table; +} + +static const PureCrypto *loaded(char *error, size_t size) { + pthread_once(&once, load); + if (!crypto && error) snprintf(error, size, "%s", why); + return crypto; +} + +/* ---- TLS */ + +bool pure_tls_start(const char *ca_file, char *error, size_t size) { + static int started; /* 1 made, -1 failed; by the thread that opens wires */ + const PureCrypto *c = loaded(error, size); + if (!c) return false; + if (!started) started = c->tls_start(ca_file) ? 1 : -1; + if (started < 0) snprintf(error, size, "no TLS context: %s cannot be read", ca_file ? ca_file : "the system's CAs"); + return started > 0; +} + +/* A session exists only once the table is loaded. */ +PureTls *pure_tls_open(const char *host, int fd) { return crypto ? crypto->tls_open(host, fd) : NULL; } +int pure_tls_read(PureTls *tls, void *bytes, int size) { return crypto->tls_read(tls, bytes, size); } +int pure_tls_write(PureTls *tls, const void *bytes, int size) { return crypto->tls_write(tls, bytes, size); } +void pure_tls_close(PureTls *tls) { if (tls) crypto->tls_close(tls); } + +/* ---- Sealing */ + +static const char aes_name[] = "x25519-sha256-aes128gcm", chacha_name[] = "x25519-sha256-chacha20poly1305"; + +/* AES-GCM where AES is in the CPU (two to three times ChaCha20-Poly1305 from + * a kilobyte up on a Cortex-A76, research 086); ChaCha20-Poly1305, fast in + * plain code, elsewhere. */ +static bool has_aes(void) { +#if defined(__aarch64__) && defined(__linux__) + return (getauxval(AT_HWCAP) & HWCAP_AES) && (getauxval(AT_HWCAP) & HWCAP_PMULL); +#elif defined(__x86_64__) && (defined(__GNUC__) || defined(__clang__)) + return __builtin_cpu_supports("aes") && __builtin_cpu_supports("pclmul"); +#else + return false; +#endif +} + +const char *pure_seal_suite(void) { return has_aes() ? aes_name : chacha_name; } + +static bool hex_value(char c, unsigned *v) { + if (c >= '0' && c <= '9') *v = (unsigned)(c - '0'); + else if (c >= 'a' && c <= 'f') *v = (unsigned)(c - 'a' + 10); + else if (c >= 'A' && c <= 'F') *v = (unsigned)(c - 'A' + 10); + else return false; + return true; +} + +bool pure_seal_peer(const char *hex, unsigned char out[PURE_SEAL_KEY]) { + if (strlen(hex) != 2 * PURE_SEAL_KEY) return false; + for (size_t i = 0; i < PURE_SEAL_KEY; ++i) { + unsigned high, low; + if (!hex_value(hex[2 * i], &high) || !hex_value(hex[2 * i + 1], &low)) return false; + out[i] = (unsigned char)(high << 4 | low); + } + return true; +} + +bool pure_seal_ready(char *error, size_t size) { return loaded(error, size) != NULL; } + +PureSealKey *pure_seal_key(const char *path, char public_hex[2 * PURE_SEAL_KEY + 1], char *error, size_t size) { + const PureCrypto *c = loaded(error, size); + return c ? c->seal_key(path, public_hex, error, size) : NULL; +} + +/* A key or a session exists only once the table is loaded. */ +void pure_seal_key_free(PureSealKey *key) { if (key) crypto->seal_key_free(key); } + +PureSealer *pure_sealer(const unsigned char peer[PURE_SEAL_KEY], const PureSealKey *own, + unsigned char enc[PURE_SEAL_KEY]) { + const PureCrypto *c = loaded(NULL, 0); + return c ? c->sealer(pure_seal_suite(), peer, own, enc) : NULL; +} + +bool pure_seal(PureSealer *s, const char *aad, const unsigned char *plain, size_t size, unsigned char *out, + uint64_t *n) { + return crypto->seal(s, aad, plain, size, out, n); +} + +void pure_sealer_free(PureSealer *s) { if (s) crypto->sealer_free(s); } + +PureOpener *pure_opener(const PureSealKey *key, const char *suite, const unsigned char enc[PURE_SEAL_KEY], + const unsigned char *from) { + return key ? crypto->opener(key, suite, enc, from) : NULL; +} + +bool pure_unseal(PureOpener *o, uint64_t n, const char *aad, const unsigned char *sealed, size_t size, + unsigned char *plain) { + return crypto->unseal(o, n, aad, sealed, size, plain); +} + +void pure_opener_free(PureOpener *o) { if (o) crypto->opener_free(o); } + +/* ---- Links */ + +bool pure_link_secret(const PureSealKey *own, const unsigned char peer[PURE_SEAL_KEY], unsigned char out[32]) { + const PureCrypto *c = loaded(NULL, 0); + return c && own && c->link_secret(own, peer, out); +} + +bool pure_link_derive(const unsigned char secret[32], const unsigned char *info, size_t info_size, unsigned char *out, + size_t out_size) { + const PureCrypto *c = loaded(NULL, 0); + return c && c->link_derive(secret, info, info_size, out, out_size); +} + +PureAead *pure_aead(const unsigned char key[PURE_AEAD_KEY]) { + const PureCrypto *c = loaded(NULL, 0); + return c ? c->aead(key) : NULL; +} + +bool pure_aead_seal(PureAead *a, const unsigned char *key, const unsigned char nonce[PURE_AEAD_NONCE], + const unsigned char *aad, size_t aad_size, const unsigned char *plain, size_t size, + unsigned char *out) { + return crypto->aead_seal(a, key, nonce, aad, aad_size, plain, size, out); +} + +bool pure_aead_open(PureAead *a, const unsigned char *key, const unsigned char nonce[PURE_AEAD_NONCE], + const unsigned char *aad, size_t aad_size, const unsigned char *sealed, size_t size, + unsigned char *plain) { + return crypto->aead_open(a, key, nonce, aad, aad_size, sealed, size, plain); +} + +void pure_aead_free(PureAead *a) { if (a) crypto->aead_free(a); } diff --git a/link/crypto.h b/link/crypto.h new file mode 100644 index 0000000..53ac785 --- /dev/null +++ b/link/crypto.h @@ -0,0 +1,49 @@ +#ifndef PURE_CRYPTO_H +#define PURE_CRYPTO_H +#include "seal.h" +#include "tls.h" + +/* What pure-serve asks of OpenSSL, as the table pure-crypto.so exports + * (link/crypto-openssl.c). pure-serve maps it, and OpenSSL with it, the + * first time a server seals, opens a seal or connects to an https receiver + * (link/crypto.c): one that does none of these maps no OpenSSL, 2.3 MiB + * less resident and 0.6 MiB less of its own. */ +enum { PURE_CRYPTO_VERSION = 4 }; + +typedef struct { + unsigned version; + /* TLS client: the context (verifying against ca_file, or the system's + * CAs when NULL), sessions over connected sockets, their reads and + * writes (as SSL_read and SSL_write: bytes, or <= 0), and their end. */ + bool (*tls_start)(const char *ca_file); + PureTls *(*tls_open)(const char *host, int fd); + int (*tls_read)(PureTls *, void *bytes, int size); + int (*tls_write)(PureTls *, const void *bytes, int size); + void (*tls_close)(PureTls *); + /* Sealing, as link/seal.h says; a sender's sessions in the suite named. */ + PureSealKey *(*seal_key)(const char *path, char public_hex[2 * PURE_SEAL_KEY + 1], char *error, size_t size); + void (*seal_key_free)(PureSealKey *); + PureSealer *(*sealer)(const char *suite, const unsigned char peer[PURE_SEAL_KEY], const PureSealKey *own, + unsigned char enc[PURE_SEAL_KEY]); + bool (*seal)(PureSealer *, const char *aad, const unsigned char *plain, size_t size, unsigned char *out, + uint64_t *n); + void (*sealer_free)(PureSealer *); + PureOpener *(*opener)(const PureSealKey *, const char *suite, const unsigned char enc[PURE_SEAL_KEY], + const unsigned char *from); + bool (*unseal)(PureOpener *, uint64_t n, const char *aad, const unsigned char *sealed, size_t size, + unsigned char *plain); + void (*opener_free)(PureOpener *); + /* Links, as link/seal.h says. */ + bool (*link_secret)(const PureSealKey *, const unsigned char peer[PURE_SEAL_KEY], unsigned char out[32]); + bool (*link_derive)(const unsigned char secret[32], const unsigned char *info, size_t info_size, unsigned char *out, + size_t out_size); + PureAead *(*aead)(const unsigned char key[PURE_AEAD_KEY]); + bool (*aead_seal)(PureAead *, const unsigned char *key, const unsigned char nonce[PURE_AEAD_NONCE], + const unsigned char *aad, size_t aad_size, const unsigned char *plain, size_t size, + unsigned char *out); + bool (*aead_open)(PureAead *, const unsigned char *key, const unsigned char nonce[PURE_AEAD_NONCE], + const unsigned char *aad, size_t aad_size, const unsigned char *sealed, size_t size, + unsigned char *plain); + void (*aead_free)(PureAead *); +} PureCrypto; +#endif diff --git a/link/link.c b/link/link.c new file mode 100644 index 0000000..9c6794b --- /dev/null +++ b/link/link.c @@ -0,0 +1,1862 @@ +#define _GNU_SOURCE +/* Links without connections (link/link.h, docs/design/011-pure-transport.md). */ +#include "link.h" +#include <arpa/inet.h> +#include <errno.h> +#include <fcntl.h> +#include <malloc.h> +#include <netdb.h> +#include <netinet/in.h> +#include <netinet/udp.h> +#include <poll.h> +#include <pthread.h> +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <sys/eventfd.h> +#include <sys/mman.h> +#include <sys/random.h> +#include <sys/socket.h> +#include <sys/stat.h> +#include <time.h> +#include <unistd.h> + +enum { VERSION = 0x70, HEADER = 21, DATAGRAM = 1200, BODY = DATAGRAM - HEADER - PURE_AEAD_TAG, PIECE = 1000 }; +enum { F_PADDING = 0, F_STREAM = 1, F_RECORD = 2, F_ACK = 3, F_STORED = 4, F_PING = 5, F_RECORDS = 6, STORED_GAP = 1 }; +enum { RECORDS_HEAD = 10 }; /* type, the first number, the count */ +/* Packets sent at once with UDP GSO: each but the last is a whole datagram + * (a full one padded), and 54 of them fit a UDP datagram's 65,507 bytes. */ +enum { BURST = 54 }; +#ifndef UDP_SEGMENT +#define UDP_SEGMENT 103 +#endif +#ifndef UDP_GRO +#define UDP_GRO 104 +#endif +enum { RANGES = 32, RUNS = 2, UNITS = 64, CREDIT = 1024, BATCH = 256, RECORD_HEAD = 19 }; +enum { WINDOW_FIRST = 10 * DATAGRAM, WINDOW_LEAST = 2 * DATAGRAM }; +#define MS UINT64_C(1000000) +/* RFC 9002's values, but a first RTT of 100 ms (a LAN's, not the 333 ms of + * the Internet at large); a probe at least every second while a receiver is + * away; a record received but not stored after a second is sent again. */ +static const uint64_t FIRST_RTT = 100 * MS, ACK_DELAY_MOST = 25 * MS, GRANULARITY = 1 * MS, PTO_MOST = 1000 * MS, + STORED_WAIT = 1000 * MS, RECORD_MOST = UINT64_C(64) << 20, HELD_MOST = UINT64_C(16) << 20; + +static void put16(unsigned char *p, uint32_t v) { p[0] = (unsigned char)v, p[1] = (unsigned char)(v >> 8); } +static void put32(unsigned char *p, uint32_t v) { for (int i = 0; i < 4; ++i) p[i] = (unsigned char)(v >> 8 * i); } +static void put64(unsigned char *p, uint64_t v) { for (int i = 0; i < 8; ++i) p[i] = (unsigned char)(v >> 8 * i); } +static uint32_t get16(const unsigned char *p) { return (uint32_t)p[0] | (uint32_t)p[1] << 8; } +static uint32_t get32(const unsigned char *p) { + return (uint32_t)p[0] | (uint32_t)p[1] << 8 | (uint32_t)p[2] << 16 | (uint32_t)p[3] << 24; +} +static uint64_t get64(const unsigned char *p) { return (uint64_t)get32(p) | (uint64_t)get32(p + 4) << 32; } +/* LEB128: seven bits a byte, low first. */ +static size_t put_varint(unsigned char *p, uint64_t v) { + size_t n = 0; + for (; v >= 0x80; v >>= 7) p[n++] = (unsigned char)(v | 0x80); + p[n++] = (unsigned char)v; + return n; +} +static size_t varint_size(uint64_t v) { + size_t n = 1; + for (; v >= 0x80; v >>= 7) ++n; + return n; +} +/* A varint from p (size bytes): how many bytes it took, or 0. */ +static size_t get_varint(const unsigned char *p, size_t size, uint64_t *v) { + *v = 0; + for (size_t n = 0; n < size && n < 10; ++n) { + *v |= (uint64_t)(p[n] & 0x7f) << (7 * n); + if (!(p[n] & 0x80)) return n + 1; + } + return 0; +} + +static uint64_t least(uint64_t a, uint64_t b) { return a < b ? a : b; } +static uint64_t most(uint64_t a, uint64_t b) { return a > b ? a : b; } + +static uint64_t now_ns(void) { + struct timespec t; + clock_gettime(CLOCK_MONOTONIC, &t); + return (uint64_t)t.tv_sec * 1000000000u + (uint64_t)t.tv_nsec; +} + +static uint32_t fnv(uint32_t h, const void *bytes, size_t size) { + const unsigned char *p = bytes; + for (size_t i = 0; i < size; ++i) h = (h ^ p[i]) * 16777619u; + return h; +} + +static uint32_t key_id(const unsigned char key[PURE_SEAL_KEY]) { return fnv(2166136261u, key, PURE_SEAL_KEY); } + +/* What a message of input looks like: its fields' names, types and places. */ +static uint32_t shape_of(const PureModule *m, uint32_t input) { + const PureInputShape *s = &m->input[input]; + uint32_t h = fnv(2166136261u, &s->prefix, sizeof(s->prefix)); + for (uint32_t f = 0; f < s->field_count; ++f) { + const PureField *field = &m->field[s->field_start + f]; + unsigned char fixed[7] = {field->type, (unsigned char)field->rows, (unsigned char)(field->rows >> 8)}; + put32(fixed + 3, field->offset); + h = fnv(fnv(h, field->name, strlen(field->name) + 1), fixed, sizeof(fixed)); + } + return h; +} + +static uint64_t random64(void) { + uint64_t value = 0; + while (getrandom(&value, sizeof(value), 0) != sizeof(value) || !value) {} + return value; +} + +/* The key and nonce base of packets from `from` (in its run from_run) to + * `to`, answering to_run (0 for none). */ +static bool derive(const unsigned char secret[32], const unsigned char from[PURE_SEAL_KEY], + const unsigned char to[PURE_SEAL_KEY], uint64_t from_run, uint64_t to_run, + unsigned char key[PURE_AEAD_KEY], unsigned char base[PURE_AEAD_NONCE]) { + unsigned char info[16 + 2 * PURE_SEAL_KEY + 16], out[PURE_AEAD_KEY + PURE_AEAD_NONCE]; + memcpy(info, "pure link 0 keys", 16); + memcpy(info + 16, from, PURE_SEAL_KEY); + memcpy(info + 16 + PURE_SEAL_KEY, to, PURE_SEAL_KEY); + put64(info + 16 + 2 * PURE_SEAL_KEY, from_run); + put64(info + 24 + 2 * PURE_SEAL_KEY, to_run); + if (!pure_link_derive(secret, info, sizeof(info), out, sizeof(out))) return false; + memcpy(key, out, PURE_AEAD_KEY); + memcpy(base, out + PURE_AEAD_KEY, PURE_AEAD_NONCE); + explicit_bzero(out, sizeof(out)); + return true; +} + +/* The same as a context of its own, for a sender's one link. */ +static PureAead *aead_of(const unsigned char secret[32], const unsigned char from[PURE_SEAL_KEY], + const unsigned char to[PURE_SEAL_KEY], uint64_t from_run, uint64_t to_run, + unsigned char base[PURE_AEAD_NONCE]) { + unsigned char key[PURE_AEAD_KEY]; + PureAead *aead = derive(secret, from, to, from_run, to_run, key, base) ? pure_aead(key) : NULL; + explicit_bzero(key, sizeof(key)); + return aead; +} + +static void nonce_of(unsigned char out[PURE_AEAD_NONCE], const unsigned char base[PURE_AEAD_NONCE], uint64_t pn) { + memcpy(out, base, PURE_AEAD_NONCE); + for (int i = 0; i < 8; ++i) out[4 + i] ^= (unsigned char)(pn >> (56 - 8 * i)); +} + +/* A packet: the header, then body sealed (with key, or the context's own + * when NULL); its size, or 0. */ +static size_t seal_packet(PureAead *aead, const unsigned char *key, const unsigned char base[PURE_AEAD_NONCE], uint32_t id, + uint64_t run, uint64_t pn, const unsigned char *body, size_t size, unsigned char *out) { + unsigned char nonce[PURE_AEAD_NONCE]; + out[0] = VERSION; + put32(out + 1, id); + put64(out + 5, run); + put64(out + 13, pn); + nonce_of(nonce, base, pn); + return pure_aead_seal(aead, key, nonce, out, HEADER, body, size, out + HEADER) ? HEADER + size + PURE_AEAD_TAG : 0; +} + +static size_t put_names(unsigned char *p, const char *source, const char *input) { + size_t a = strlen(source), b = strlen(input); + p[0] = (unsigned char)a; + memcpy(p + 1, source, a); + p[1 + a] = (unsigned char)b; + memcpy(p + 2 + a, input, b); + return 2 + a + b; +} + +/* A source and an input's names from p (size bytes); how many it takes, or 0. */ +static size_t get_names(const unsigned char *p, size_t size, char source[PURE_SOURCE + 1], char input[PURE_MODULE_NAME]) { + if (size < 1 || !p[0] || p[0] > PURE_SOURCE || size < 2u + p[0]) return 0; + size_t a = p[0], b = p[1 + a]; + if (!b || b >= PURE_MODULE_NAME || size < 2 + a + b) return 0; + memcpy(source, p + 1, a); + source[a] = 0; + memcpy(input, p + 2 + a, b); + input[b] = 0; + return 2 + a + b; +} + +/* Datagrams taken from a socket at once (recvmmsg). */ +enum { TAKE = 64, SLOT = DATAGRAM + 64 }; +typedef struct { + unsigned char bytes[TAKE][SLOT]; + struct sockaddr_storage from[TAKE]; + struct iovec parts[TAKE]; + struct mmsghdr messages[TAKE]; +} Taking; + +/* With UDP GRO the kernel hands a listener the datagrams of one sender that + * came one after another as one buffer, each the segment's size but the + * last (a burst sent with GSO comes whole, not cut up): up to GATHER such + * buffers at once. */ +enum { GATHER = 8, GATHERED = 65536 }; +typedef struct { + unsigned char bytes[GATHER][GATHERED]; + struct sockaddr_storage from[GATHER]; + struct iovec parts[GATHER]; + struct mmsghdr messages[GATHER]; + _Alignas(8) unsigned char control[GATHER][64]; /* a cmsghdr's alignment */ +} Gathering; + +static int gather(int fd, Gathering *g) { + for (unsigned k = 0; k < GATHER; ++k) { + g->parts[k] = (struct iovec){g->bytes[k], GATHERED}; + g->messages[k] = (struct mmsghdr){.msg_hdr = {.msg_name = &g->from[k], .msg_namelen = sizeof(g->from[k]), + .msg_iov = &g->parts[k], .msg_iovlen = 1, + .msg_control = g->control[k], + .msg_controllen = sizeof(g->control[k])}}; + } + int n = recvmmsg(fd, g->messages, GATHER, MSG_DONTWAIT, NULL); + return n > 0 ? n : 0; +} + +/* The size of each datagram in a buffer gathered (all of it when one). */ +static size_t segment_of(struct msghdr *m, size_t length) { + for (struct cmsghdr *c = CMSG_FIRSTHDR(m); c; c = CMSG_NXTHDR(m, c)) + if (c->cmsg_level == SOL_UDP && c->cmsg_type == UDP_GRO) { + int size; + memcpy(&size, CMSG_DATA(c), sizeof(size)); + if (size > 0) return (size_t)size; + } + return length ? length : 1; +} + +/* Up to TAKE datagrams from fd: how many (0 when none waits). */ +static int take_datagrams(int fd, Taking *t) { + for (unsigned k = 0; k < TAKE; ++k) { + t->parts[k] = (struct iovec){t->bytes[k], SLOT}; + t->messages[k] = (struct mmsghdr){.msg_hdr = {.msg_name = &t->from[k], .msg_namelen = sizeof(t->from[k]), + .msg_iov = &t->parts[k], .msg_iovlen = 1}}; + } + int n = recvmmsg(fd, t->messages, TAKE, MSG_DONTWAIT, NULL); + return n > 0 ? n : 0; +} + +/* ---- The sending side */ + +enum { UNSENT, FLYING, RECEIVED }; +typedef struct { uint64_t pn; uint8_t state; } Piece; +/* A record queued: its bytes at `at` in the link's byte ring, its pieces' + * states (one kept here, more beside). */ +typedef struct { uint64_t seq; size_t at; uint32_t size, pieces; Piece *more, one; } Record; +static Piece *piece_of(Record *r, uint32_t i) { return r->pieces == 1 ? &r->one : &r->more[i]; } +typedef struct { uint64_t seq; uint32_t piece; } Unit; +typedef struct { + uint64_t pn, sent; /* its number; when it went (ns) */ + uint32_t bytes; + bool eliciting, done; /* acknowledged or lost */ + uint16_t count; + Unit units[UNITS]; +} Flight; + +struct PureLink { + const PureModule *module; + uint32_t output, shape, own_id, peer_id; + char host[256], port[8], input[PURE_MODULE_NAME], source[PURE_SOURCE + 1]; + const PureSealKey *own; + unsigned char own_public[PURE_SEAL_KEY], peer[PURE_SEAL_KEY], secret[32]; + bool peered; + pthread_mutex_t lock; + Record *records; /* the sends not yet stored, oldest first (a ring of consecutive numbers) */ + size_t head, count, capacity; + /* Their bytes, one after another as queued (a record's never cut by the + * ring's end): taken back in order as they are stored. */ + unsigned char *ring; + size_t ring_capacity; + uint64_t queued, stored, told; /* the last queued; saved by the receiver; told the owner */ + bool stopping, started; + /* The thread sleeps (in poll), and has been woken since; records were + * queued since it last sent. A queue wakes it only when it sleeps, once. */ + bool sleeping, signaled, queued_new; + int fd, wake; + pthread_t thread; + void (*notify)(void *); + void *context; + /* The thread's own, under the lock. */ + uint64_t run, next_pn, in_run; + PureAead *out, *in; + unsigned char out_base[PURE_AEAD_NONCE], in_base[PURE_AEAD_NONCE]; + Flight *flights; /* one for each number sent, oldest first (a ring) */ + size_t flight_head, flight_count, flight_capacity; + uint64_t in_flight, window, threshold, recovery_start, eliciting; + uint64_t largest_acked, latest_rtt, smoothed_rtt, rtt_var, min_rtt; + bool acked_any, sampled; + uint64_t loss_time, last_eliciting, credit, stored_changed; + uint64_t received; /* pieces received and not yet stored */ + uint64_t unsent_from; /* no record before this number has a piece unsent */ + Taking *taking; /* answers, received at once */ + unsigned char *burst; /* packets sealed and not yet sent (BURST datagrams) */ + size_t burst_size; + unsigned burst_count; + uint64_t burst_pn[BURST]; + bool gso; + unsigned pto_count, probes; + PureLinkCounts counts; +}; + +PureLink *pure_link_open(const PureModule *m, const char *output, const char *url, const char *source, + const PureSealKey *own, const unsigned char own_public[PURE_SEAL_KEY], char *error, + size_t size) { + int index = pure_module_input(m, output); + if (index < 0 || m->input[index].kind != PURE_KIND_OUTPUT) { + snprintf(error, size, "no output '%s'", output); + return NULL; + } + if (!own) { + snprintf(error, size, "a pure:// link needs the node's --key"); + return NULL; + } + size_t length = strlen(source); + const char *host = url + 7, *slash = strchr(host, '/'), *colon = slash ? memchr(host, ':', (size_t)(slash - host)) : NULL; + if (!length || length > PURE_SOURCE || strncmp(url, "pure://", 7) || !slash || !colon || colon == host || + (size_t)(colon - host) >= 256 || slash - colon - 1 < 1 || slash - colon - 1 > 5 || !slash[1] || + strlen(slash + 1) >= PURE_MODULE_NAME) { + snprintf(error, size, "a link is pure://host:port/INPUT"); + return NULL; + } + PureLink *l = calloc(1, sizeof(*l)); + if (!l || pthread_mutex_init(&l->lock, NULL)) { + free(l); + snprintf(error, size, "out of memory"); + return NULL; + } + l->module = m; + l->output = (uint32_t)index; + l->shape = shape_of(m, (uint32_t)index); + memcpy(l->host, host, (size_t)(colon - host)); + memcpy(l->port, colon + 1, (size_t)(slash - colon - 1)); + strcpy(l->input, slash + 1); + strcpy(l->source, source); + l->own = own; + memcpy(l->own_public, own_public, PURE_SEAL_KEY); + l->own_id = key_id(own_public); + l->fd = l->wake = -1; + return l; +} + +void pure_link_peer(PureLink *l, const unsigned char peer[PURE_SEAL_KEY]) { + memcpy(l->peer, peer, PURE_SEAL_KEY); + l->peer_id = key_id(peer); + l->peered = true; +} + +uint32_t pure_link_output(const PureLink *l) { return l->output; } + +static Record *record_at(PureLink *l, size_t k) { return &l->records[(l->head + k) % l->capacity]; } + +/* The queued record numbered seq, or NULL (stored already, or not queued). */ +static Record *record_of(PureLink *l, uint64_t seq) { + if (!l->count || seq < record_at(l, 0)->seq || seq - record_at(l, 0)->seq >= l->count) return NULL; + return record_at(l, (size_t)(seq - record_at(l, 0)->seq)); +} + +/* Every change of a piece's state goes through here: the link knows at once + * how many pieces wait to be stored, and where the unsent ones begin. */ +static void set_piece(PureLink *l, Record *r, uint32_t i, uint8_t state) { + uint8_t was = (*piece_of(r, i)).state; + if (was == state) return; + if (was == RECEIVED) --l->received; + if (state == RECEIVED) ++l->received; + if (state == UNSENT && r->seq < l->unsent_from) l->unsent_from = r->seq; + (*piece_of(r, i)).state = state; +} + +static void wake(int fd) { + uint64_t one = 1; + if (fd >= 0 && write(fd, &one, sizeof(one)) != sizeof(one)) { /* full: it is readable already */ } +} + +/* Where size bytes go at the end of the byte ring (after the newest record, + * or from its start when that is free and the end is not), or SIZE_MAX. */ +static size_t ring_place(PureLink *l, size_t size) { + if (!l->count) return size <= l->ring_capacity ? 0 : SIZE_MAX; + Record *oldest = record_at(l, 0), *newest = record_at(l, l->count - 1); + size_t head = oldest->at, tail = newest->at + newest->size; + if (newest->at >= head) /* the bytes are [head, tail): after them, or before them from the start */ + return l->ring_capacity - tail >= size ? tail : size <= head ? 0 : SIZE_MAX; + return head - tail >= size ? tail : SIZE_MAX; /* around the end: between the newest and the oldest */ +} + +/* A ring room for size bytes more: twice as large (or more), the records + * moved into it one after another. */ +static bool ring_grow(PureLink *l, size_t size) { + size_t used = 0; + for (size_t k = 0; k < l->count; ++k) used += record_at(l, k)->size; + size_t capacity = l->ring_capacity ? 2 * l->ring_capacity : 64 << 10; + while (capacity < used + size + 1) capacity *= 2; + unsigned char *ring = malloc(capacity); + if (!ring) return false; + for (size_t k = 0, at = 0; k < l->count; ++k) { + Record *r = record_at(l, k); + if (r->size) memcpy(ring + at, l->ring + r->at, r->size); + r->at = at; + at += r->size; + } + free(l->ring); + l->ring = ring; + l->ring_capacity = capacity; + return true; +} + +bool pure_link_queue(PureLink *l, uint64_t seq, PureSpan payload) { + Record r = {.seq = seq, .size = (uint32_t)payload.size}; + r.pieces = payload.size ? (uint32_t)((payload.size + PIECE - 1) / PIECE) : 1; + if (payload.size > RECORD_MOST || (r.pieces > 1 && !(r.more = calloc(r.pieces, sizeof(Piece))))) return false; + pthread_mutex_lock(&l->lock); + bool ok = seq > l->queued && (!l->count || seq == l->queued + 1); + size_t at = ok ? ring_place(l, payload.size) : SIZE_MAX; + if (ok && at == SIZE_MAX) ok = ring_grow(l, payload.size) && (at = ring_place(l, payload.size)) != SIZE_MAX; + if (ok) { + r.at = at; + if (payload.size) memcpy(l->ring + at, payload.data, payload.size); + } + if (ok && l->count == l->capacity) { + size_t capacity = l->capacity ? 2 * l->capacity : 64; + Record *grown = malloc(capacity * sizeof(*grown)); + if (grown) { + for (size_t k = 0; k < l->count; ++k) grown[k] = *record_at(l, k); + free(l->records); + l->records = grown; + l->capacity = capacity; + l->head = 0; + } else + ok = false; + } + bool signal = false; + if (ok) { + *record_at(l, l->count++) = r; + l->queued = seq; + l->queued_new = true; + signal = l->sleeping && !l->signaled; + l->signaled = l->signaled || signal; + } + pthread_mutex_unlock(&l->lock); + if (!ok) { + free(r.more); + return false; + } + if (signal) wake(l->wake); + return true; +} + +uint64_t pure_link_queued(PureLink *l) { + pthread_mutex_lock(&l->lock); + uint64_t queued = l->queued; + pthread_mutex_unlock(&l->lock); + return queued; +} + +uint64_t pure_link_delivered(PureLink *l) { + pthread_mutex_lock(&l->lock); + uint64_t stored = l->stored; + pthread_mutex_unlock(&l->lock); + return stored; +} + +static Flight *flight_of(PureLink *l, uint64_t pn) { + if (!l->flight_count) return NULL; + uint64_t first = l->flights[l->flight_head].pn; + if (pn < first || pn - first >= l->flight_count) return NULL; + return &l->flights[(l->flight_head + (size_t)(pn - first)) % l->flight_capacity]; +} + +static bool push_flight(PureLink *l, const Flight *f) { + if (l->flight_count == l->flight_capacity) { + size_t capacity = l->flight_capacity ? 2 * l->flight_capacity : 256; + Flight *grown = malloc(capacity * sizeof(*grown)); + if (!grown) return false; + for (size_t k = 0; k < l->flight_count; ++k) grown[k] = l->flights[(l->flight_head + k) % l->flight_capacity]; + free(l->flights); + l->flights = grown; + l->flight_capacity = capacity; + l->flight_head = 0; + } + l->flights[(l->flight_head + l->flight_count++) % l->flight_capacity] = *f; + return true; +} + +static void settle(PureLink *l, Flight *f) { + f->done = true; + if (f->eliciting) { + --l->eliciting; + l->in_flight -= f->bytes; + } +} + +/* A packet lost: what it carried and was not received otherwise goes again. */ +static void lose(PureLink *l, Flight *f, uint64_t now) { + settle(l, f); + ++l->counts.lost; + for (uint16_t u = 0; u < f->count; ++u) { + Record *r = record_of(l, f->units[u].seq); + Piece *p = r ? &(*piece_of(r, f->units[u].piece)) : NULL; + if (p && p->state == FLYING && p->pn == f->pn) { + set_piece(l, r, f->units[u].piece, UNSENT); + ++l->counts.resent; + } + } + if (f->sent > l->recovery_start) { /* one congestion event per round trip */ + l->recovery_start = now; + l->window = most(l->window / 2, WINDOW_LEAST); + l->threshold = l->window; + } +} + +static void detect_lost(PureLink *l, uint64_t now) { + l->loss_time = 0; + if (!l->acked_any) return; + uint64_t rtt = most(l->latest_rtt, l->sampled ? l->smoothed_rtt : FIRST_RTT); + uint64_t delay = most(rtt * 9 / 8, GRANULARITY); + for (size_t k = 0; k < l->flight_count; ++k) { + Flight *f = &l->flights[(l->flight_head + k) % l->flight_capacity]; + if (f->pn >= l->largest_acked) break; + if (f->done) continue; + if (f->sent + delay <= now || l->largest_acked >= f->pn + 3) lose(l, f, now); + else if (!l->loss_time || f->sent + delay < l->loss_time) l->loss_time = f->sent + delay; + } +} + +static void sample_rtt(PureLink *l, uint64_t latest, uint64_t ack_delay) { + l->latest_rtt = latest; + if (!l->sampled) { + l->min_rtt = l->smoothed_rtt = latest; + l->rtt_var = latest / 2; + l->sampled = true; + return; + } + l->min_rtt = least(l->min_rtt, latest); + ack_delay = least(ack_delay, ACK_DELAY_MOST); + uint64_t adjusted = latest >= l->min_rtt + ack_delay ? latest - ack_delay : latest; + uint64_t difference = l->smoothed_rtt > adjusted ? l->smoothed_rtt - adjusted : adjusted - l->smoothed_rtt; + l->rtt_var = (3 * l->rtt_var + difference) / 4; + l->smoothed_rtt = (7 * l->smoothed_rtt + adjusted) / 8; +} + +static void on_ack(PureLink *l, const unsigned char *p, size_t size, uint64_t now) { + if (size < 13 || size < 13 + 16u * p[12]) return; + uint64_t largest = get64(p), delay = (uint64_t)get32(p + 8) * 1000; + Flight *top = NULL; + for (unsigned k = 0; k < p[12]; ++k) { + uint64_t first = get64(p + 13 + 16 * k), last = get64(p + 21 + 16 * k); + if (first > last || !l->flight_count) continue; + uint64_t oldest = l->flights[l->flight_head].pn, newest = oldest + l->flight_count - 1; + for (uint64_t pn = most(first, oldest); pn <= least(last, newest); ++pn) { + Flight *f = flight_of(l, pn); + if (!f || f->done) continue; + settle(l, f); + ++l->counts.acknowledged; + for (uint16_t u = 0; u < f->count; ++u) { + Record *r = record_of(l, f->units[u].seq); + if (r && (*piece_of(r, f->units[u].piece)).state != RECEIVED) { + set_piece(l, r, f->units[u].piece, RECEIVED); + l->stored_changed = now; /* waits for its save from now */ + } + } + if (f->eliciting && f->sent > l->recovery_start) + l->window += l->window < l->threshold ? f->bytes : (uint64_t)DATAGRAM * f->bytes / l->window; + if (pn == largest) top = f; + } + } + if (!l->acked_any || largest > l->largest_acked) l->largest_acked = largest; + l->acked_any = true; + if (top && top->eliciting) sample_rtt(l, now - top->sent, delay); + l->pto_count = 0; + detect_lost(l, now); + while (l->flight_count && l->flights[l->flight_head].done) { + l->flight_head = (l->flight_head + 1) % l->flight_capacity; + --l->flight_count; + } +} + +static void on_stored(PureLink *l, const unsigned char *p, size_t size, uint64_t now) { + char source[PURE_SOURCE + 1], input[PURE_MODULE_NAME]; + size_t names = get_names(p, size, source, input); + if (!names || size < names + 13 || strcmp(source, l->source) || strcmp(input, l->input)) return; + uint64_t heard = get64(p + names); + unsigned flags = p[names + 8]; + uint32_t credit = get32(p + names + 9); + if (heard > l->stored) { + while (l->count && record_at(l, 0)->seq <= heard) { + Record *r = record_at(l, 0); + for (uint32_t i = 0; i < r->pieces; ++i) l->received -= (*piece_of(r, i)).state == RECEIVED; + free(r->more); /* its bytes go back with the ring's head */ + l->head = (l->head + 1) % l->capacity; + --l->count; + } + l->stored = heard; + l->stored_changed = now; + } + if (flags & STORED_GAP) /* it did not keep what came after a gap: send it again */ + for (size_t k = 0; k < l->count && l->received; ++k) { + Record *r = record_at(l, k); + for (uint32_t i = 0; i < r->pieces; ++i) + if ((*piece_of(r, i)).state == RECEIVED) set_piece(l, r, i, UNSENT); + } + l->credit = credit ? least(credit, 65536) : 1; +} + +static void take_answer(PureLink *l, unsigned char *packet, ssize_t n, uint64_t now); + +static void take_answers(PureLink *l, uint64_t now) { + for (unsigned rounds = 0, n = TAKE; rounds < 4 && n == TAKE; ++rounds) { + n = (unsigned)take_datagrams(l->fd, l->taking); + for (unsigned k = 0; k < n; ++k) take_answer(l, l->taking->bytes[k], l->taking->messages[k].msg_len, now); + } +} + +static void take_answer(PureLink *l, unsigned char *packet, ssize_t n, uint64_t now) { + unsigned char plain[DATAGRAM], nonce[PURE_AEAD_NONCE]; + if (n < HEADER + PURE_AEAD_TAG || (size_t)n > DATAGRAM || packet[0] != VERSION || get32(packet + 1) != l->peer_id) + return; /* not ours, or an error from an earlier send (nobody listening) */ + uint64_t run = get64(packet + 5), pn = get64(packet + 13); + if (!l->in || l->in_run != run) { + pure_aead_free(l->in); + l->in = aead_of(l->secret, l->peer, l->own_public, run, l->run, l->in_base); + l->in_run = run; + if (!l->in) return; + } + nonce_of(nonce, l->in_base, pn); + size_t size = (size_t)n - HEADER - PURE_AEAD_TAG; + if (!pure_aead_open(l->in, NULL, nonce, packet, HEADER, packet + HEADER, (size_t)n - HEADER, plain)) return; + ++l->counts.received; + for (size_t at = 0; at < size;) { + unsigned type = plain[at++]; + if (type == F_ACK && at + 13 <= size && at + 13 + 16u * plain[at + 12] <= size) { + on_ack(l, plain + at, size - at, now); + at += 13 + 16u * plain[at + 12]; + } else if (type == F_STORED) { + char source[PURE_SOURCE + 1], input[PURE_MODULE_NAME]; + size_t names = get_names(plain + at, size - at, source, input); + if (!names || at + names + 13 > size) break; + on_stored(l, plain + at, size - at, now); + at += names + 13; + } else if (type != F_PING) + break; /* padding, or what this side does not know: the rest */ + } +} + +/* A packet of pieces not sent yet, in order, within the receiver's credit: + * its body's size, or 0 when there is none. */ +static size_t fill(PureLink *l, unsigned char *body, Flight *f) { + size_t at = 0; + body[at++] = F_STREAM; + at += put_names(body + at, l->source, l->input); + put32(body + at, l->shape); + at += 4; + /* The credit counts from the oldest record not stored: after a restart + * the sender does not know yet how far the receiver stored. */ + uint64_t base = l->count ? record_at(l, 0)->seq - 1 : l->stored; + /* From the first record that may have an unsent piece: the ones before + * it are sent, and every one found sent moves that mark on. */ + size_t k = l->count && l->unsent_from > base + 1 ? (size_t)(l->unsent_from - base - 1) : 0; + bool before = true; + /* A record none of whose pieces went yet goes whole when it fits what is + * left of the packet: whole records whose numbers follow on share a + * RECORDS frame (its count at run). Else its unsent pieces go, each in a + * RECORD frame. */ + size_t run = 0; + uint64_t next = 0; + for (; k < l->count && f->count < UNITS; ++k) { + Record *r = record_at(l, k); + if (r->seq > base + l->credit) break; + bool unsent = false, fresh = true, joins = run && r->seq == next; + for (uint32_t i = 0; i < r->pieces && fresh; ++i) fresh = piece_of(r, i)->state == UNSENT; + size_t whole = (joins ? 0 : RECORDS_HEAD) + varint_size(r->size) + r->size; + if (fresh && f->count + r->pieces <= UNITS && at + whole <= BODY) { + if (!joins) { + body[at] = F_RECORDS; + put64(body + at + 1, r->seq); + body[at + 9] = 0; + run = at + 9; + at += RECORDS_HEAD; + } + at += put_varint(body + at, r->size); + if (r->size) memcpy(body + at, l->ring + r->at, r->size); + at += r->size; + ++body[run]; + next = r->seq + 1; + for (uint32_t i = 0; i < r->pieces; ++i) { + set_piece(l, r, i, FLYING); + piece_of(r, i)->pn = f->pn; + f->units[f->count++] = (Unit){r->seq, i}; + } + ++l->counts.records; + } else if (fresh && r->pieces == 1) + unsent = true; /* too large for what is left: the next packet */ + else + for (uint32_t i = 0; i < r->pieces; ++i) { + if (piece_of(r, i)->state != UNSENT) continue; + uint32_t offset = i * PIECE, length = (uint32_t)least(PIECE, r->size - offset); + if (f->count == UNITS || at + RECORD_HEAD + length > BODY) { + unsent = true; + break; + } + body[at] = F_RECORD; + put64(body + at + 1, r->seq); + put32(body + at + 9, r->size); + put32(body + at + 13, offset); + put16(body + at + 17, length); + if (length) memcpy(body + at + RECORD_HEAD, l->ring + r->at + offset, length); + at += RECORD_HEAD + length; + run = 0; + set_piece(l, r, i, FLYING); + piece_of(r, i)->pn = f->pn; + f->units[f->count++] = (Unit){r->seq, i}; + ++l->counts.records; + } + if (before && !unsent) l->unsent_from = r->seq + 1; + before = before && !unsent; + if (unsent) { /* the packet is full: whole, so it can go in a burst */ + if (l->gso && f->count) { + memset(body + at, F_PADDING, BODY - at); + at = BODY; + } + break; + } + } + return f->count ? at : 0; +} + +/* Seal and send f with body; its number is spent even when the socket + * refuses it (what it carried is then marked unsent again). */ +/* A packet the socket refused: what it carried is unsent again, and it + * counts as never sent. Its number is spent all the same. */ +static void refused(PureLink *l, uint64_t pn) { + Flight *f = flight_of(l, pn); + if (!f || f->done) return; + for (uint16_t u = 0; u < f->count; ++u) { + Record *r = record_of(l, f->units[u].seq); + if (r && (*piece_of(r, f->units[u].piece)).pn == pn) set_piece(l, r, f->units[u].piece, UNSENT); + } + settle(l, f); + f->eliciting = false; + f->count = 0; + --l->counts.packets; +} + +static bool transient(void) { return errno == EAGAIN || errno == EWOULDBLOCK || errno == ENOBUFS || errno == EINTR; } + +/* The packets of the burst to the socket (each but the last a whole + * datagram): with GSO in one call, else one by one. */ +static void flush(PureLink *l) { + unsigned n = l->burst_count, sent = 0; + bool whole = n > 1 && l->gso; + if (whole && send(l->fd, l->burst, l->burst_size, 0) >= 0) sent = n; + else if (whole && !transient()) l->gso = false, whole = false; /* no GSO here: one by one from now on */ + for (size_t at = 0; !whole && sent < n; ++sent) { + size_t size = sent + 1 < n ? DATAGRAM : l->burst_size - at; + if (send(l->fd, l->burst + at, size, 0) < 0 && transient()) break; + at += size; + } + for (unsigned k = sent; k < n; ++k) refused(l, l->burst_pn[k]); + l->burst_count = 0; + l->burst_size = 0; +} + +/* Seal f with body into the burst (a whole datagram each but the last), and + * count it sent; its number is spent even when the socket then refuses it. */ +static bool transmit(PureLink *l, Flight *f, const unsigned char *body, size_t size, uint64_t now) { + if (l->burst_count && l->burst_size != (size_t)l->burst_count * DATAGRAM) flush(l); /* the last was short */ + size_t total = seal_packet(l->out, NULL, l->out_base, l->own_id, l->run, f->pn, body, size, l->burst + l->burst_size); + ++l->next_pn; + f->sent = now; + f->bytes = (uint32_t)total; + if (!total) { /* not sealed: what it carried is unsent again */ + for (uint16_t u = 0; u < f->count; ++u) { + Record *r = record_of(l, f->units[u].seq); + if (r && (*piece_of(r, f->units[u].piece)).pn == f->pn) set_piece(l, r, f->units[u].piece, UNSENT); + } + f->done = true; + f->eliciting = false; + f->count = 0; + push_flight(l, f); + return false; + } + if (!push_flight(l, f)) return false; + l->burst_pn[l->burst_count++] = f->pn; + l->burst_size += total; + ++l->counts.packets; + if (f->eliciting) { + ++l->eliciting; + l->in_flight += f->bytes; + l->last_eliciting = now; + } + if (l->burst_count == BURST) flush(l); + return true; +} + +static void send_packets(PureLink *l, uint64_t now) { + unsigned char body[BODY]; + l->queued_new = false; /* all queued is looked at now */ + for (unsigned n = 0; n < 128; ++n) { + bool probe = l->probes > 0; + if (!probe && l->in_flight + DATAGRAM > l->window) break; + Flight f = {.pn = l->next_pn, .eliciting = true}; + size_t size = fill(l, body, &f); + if (!size) { + if (!probe) break; + body[0] = F_PING; /* nothing to send again: a probe all the same */ + size = 1; + } + if (!transmit(l, &f, body, size, now)) break; + if (probe) --l->probes; + } + flush(l); +} + +static uint64_t pto_period(const PureLink *l) { + uint64_t smoothed = l->sampled ? l->smoothed_rtt : FIRST_RTT, var = l->sampled ? l->rtt_var : FIRST_RTT / 2; + uint64_t period = smoothed + most(4 * var, GRANULARITY) + ACK_DELAY_MOST; + for (unsigned i = 0; i < l->pto_count && period < PTO_MOST; ++i) period *= 2; + return least(period, PTO_MOST); +} + +static bool waiting_for_save(PureLink *l) { return l->received > 0; } + +/* The next time a timer is due (ns), or 0. */ +static uint64_t next_timer(PureLink *l) { + uint64_t due = l->loss_time; + if (l->eliciting) { + uint64_t pto = l->last_eliciting + pto_period(l); + if (!due || pto < due) due = pto; + } + if (waiting_for_save(l)) { + uint64_t stored = l->stored_changed + STORED_WAIT; + if (!due || stored < due) due = stored; + } + return due; +} + +static void run_timers(PureLink *l, uint64_t now) { + if (l->loss_time && l->loss_time <= now) detect_lost(l, now); + else if (l->eliciting && l->last_eliciting + pto_period(l) <= now) { + /* Nothing heard for a while: a probe, what flies longest sent again. */ + ++l->pto_count; + l->probes = 1; + for (size_t k = 0; k < l->count; ++k) { + Record *r = record_at(l, k); + uint32_t i = 0; + while (i < r->pieces && (*piece_of(r, i)).state != FLYING) ++i; + if (i < r->pieces) { + set_piece(l, r, i, UNSENT); + ++l->counts.resent; + break; + } + } + } + if (waiting_for_save(l) && l->stored_changed + STORED_WAIT <= now) { + /* Received long ago and still not stored: the receiver lost them (it + * restarted) or its word got lost. Sent again: it takes a number once. */ + for (size_t k = 0; k < l->count && l->received; ++k) { + Record *r = record_at(l, k); + for (uint32_t i = 0; i < r->pieces; ++i) + if ((*piece_of(r, i)).state == RECEIVED) set_piece(l, r, i, UNSENT), ++l->counts.resent; + } + l->stored_changed = now; + } +} + +static void *run_link(void *context) { + PureLink *l = context; + for (;;) { + pthread_mutex_lock(&l->lock); + uint64_t now = now_ns(), due = next_timer(l); + bool stopping = l->stopping, more = l->queued_new; /* queued while it sent: no sleep */ + l->sleeping = !more; + pthread_mutex_unlock(&l->lock); + if (stopping) break; + int timeout = more ? 0 : due ? (int)((due > now ? due - now : 0) / MS) + 1 : -1; + struct pollfd waits[2] = {{l->fd, POLLIN, 0}, {l->wake, POLLIN, 0}}; + if (poll(waits, 2, timeout) < 0 && errno != EINTR) break; + if (waits[1].revents & POLLIN) { + uint64_t signals; + if (read(l->wake, &signals, sizeof(signals)) < 0) { /* cleared already */ } + } + pthread_mutex_lock(&l->lock); + l->sleeping = l->signaled = false; + now = now_ns(); + if (waits[0].revents) take_answers(l, now); + run_timers(l, now); + send_packets(l, now); + bool told = l->stored != l->told; + l->told = l->stored; + pthread_mutex_unlock(&l->lock); + if (told && l->notify) l->notify(l->context); + } + return NULL; +} + +bool pure_link_start(PureLink *l, void (*notify)(void *), void *context, char *error, size_t size) { + if (!l->peered) { + snprintf(error, size, "pure://%s:%s/%s: no receiver's key (--seal)", l->host, l->port, l->input); + return false; + } + struct addrinfo hints = {.ai_family = AF_UNSPEC, .ai_socktype = SOCK_DGRAM}, *found = NULL; + int resolved = getaddrinfo(l->host, l->port, &hints, &found); + if (resolved) { + snprintf(error, size, "%s: %s", l->host, gai_strerror(resolved)); + return false; + } + l->fd = socket(found->ai_family, SOCK_DGRAM | SOCK_NONBLOCK | SOCK_CLOEXEC, 0); + bool ok = l->fd >= 0 && !connect(l->fd, found->ai_addr, found->ai_addrlen); + freeaddrinfo(found); + int buffer = 4 << 20; + if (ok) { + setsockopt(l->fd, SOL_SOCKET, SO_SNDBUF, &buffer, sizeof(buffer)); + setsockopt(l->fd, SOL_SOCKET, SO_RCVBUF, &buffer, sizeof(buffer)); + } + int segment = DATAGRAM; + l->gso = ok && !setsockopt(l->fd, SOL_UDP, UDP_SEGMENT, &segment, sizeof(segment)); + ok = ok && (l->burst = malloc((size_t)BURST * DATAGRAM)) && (l->taking = malloc(sizeof(*l->taking))); + l->run = random64(); + l->window = WINDOW_FIRST; + l->threshold = UINT64_MAX; + l->credit = CREDIT; + ok = ok && pure_link_secret(l->own, l->peer, l->secret) && + (l->out = aead_of(l->secret, l->own_public, l->peer, l->run, 0, l->out_base)) && + (l->wake = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC)) >= 0; + l->notify = notify; + l->context = context; + if (!ok || pthread_create(&l->thread, NULL, run_link, l)) { + snprintf(error, size, "pure://%s:%s/%s: cannot start (a socket, the keys)", l->host, l->port, l->input); + return false; + } + pthread_setname_np(l->thread, "pure-link"); /* a link's CPU, seen apart */ + l->started = true; + return true; +} + +void pure_link_counts(PureLink *l, PureLinkCounts *out) { + pthread_mutex_lock(&l->lock); + *out = l->counts; + out->stored = l->stored; + pthread_mutex_unlock(&l->lock); +} + +void pure_link_close(PureLink *l) { + if (!l) return; + if (l->started) { + pthread_mutex_lock(&l->lock); + l->stopping = true; + pthread_mutex_unlock(&l->lock); + wake(l->wake); + pthread_join(l->thread, NULL); + } + for (size_t k = 0; k < l->count; ++k) { + free(record_at(l, k)->more); + } + free(l->records); + free(l->ring); + free(l->flights); + free(l->burst); + free(l->taking); + pure_aead_free(l->out); + pure_aead_free(l->in); + explicit_bzero(l->secret, sizeof(l->secret)); + if (l->fd >= 0) close(l->fd); + if (l->wake >= 0) close(l->wake); + pthread_mutex_destroy(&l->lock); + free(l); +} + +/* ---- The receiving side + * + * A sender allowed is a peer: its key and the sources it may send as, a few + * bytes each, found by its key's id. What a sending peer needs besides (the + * secret its key shares with the node's, its runs' keys and addresses, its + * streams) is its active state: made at its first packet (one X25519), let + * go once the peer has been quiet for the idle period with nothing of it + * waiting here. So a link at rest costs the node its allowing, and the next + * packet goes on without a handshake. */ + +typedef struct { uint64_t first, last; } Range; +typedef struct Active Active; +typedef struct { + Active *active; + uint64_t run, seen; /* the sender's run; when it last sent */ + bool used, ack_due, due; + unsigned char in_key[PURE_AEAD_KEY], out_key[PURE_AEAD_KEY], in_base[PURE_AEAD_NONCE], out_base[PURE_AEAD_NONCE]; + uint64_t next_pn, largest_at; + struct sockaddr_in6 address; /* or an IPv4 one */ + socklen_t address_size; + Range got[RANGES]; /* the numbers received, newest first */ + unsigned got_count; +} Run; +typedef struct { + uint64_t seq; + unsigned char *bytes; /* its size, then a byte a piece: had */ + uint32_t size, pieces, got; +} Held; +typedef struct Stream { + Active *active; + char source[PURE_SOURCE + 1]; + uint32_t input; + uint64_t heard; /* the last number heard (stored), once known */ + bool known, out, stored_due, gap, ready; + /* The line: once the last heard is known, whole records from the one + * after it on, in order, as a delivery carries them (u32 size, bytes), + * from line_head: handed over as they are, let go by moving the head. + * Records that come early or in pieces, or before the last heard is + * known, wait in held until the line reaches them. */ + unsigned char *line; + size_t line_head, line_size, line_capacity; + uint64_t line_first; + unsigned line_count; + Held *held; /* records not yet stored, by number: held_count from held_first */ + size_t held_first, held_count, held_capacity, held_bytes; + Run *run; /* the run that sent to it last: where STORED goes */ + uint64_t run_id; +} Stream; +struct Active { + uint32_t peer; /* its place among the peers */ + unsigned char secret[32]; + uint64_t seen; + size_t place; /* among the active */ + Run runs[RUNS]; + Stream **streams; + unsigned stream_count, stream_capacity; +}; +typedef struct { + unsigned char key[PURE_SEAL_KEY]; + uint32_t id, allowed; /* allowed: the place + 1 of its first source (0: none) */ + Active *active; +} Peer; +typedef struct { uint32_t name, next; } Allowed; /* name: where in names (u8 length, bytes); next: a place + 1 */ +typedef struct { Stream *stream; uint64_t heard; bool stored, gap; } Done; + +/* A peer quiet this long lets its active state go; the active are looked at + * that often; what they held goes back to the system once TRIM_AFTER have + * gone, or TRIM_WAIT after any went. */ +static const uint64_t IDLE_FIRST = 30000 * MS, SWEEP_EVERY = 250 * MS, TRIM_WAIT = 10000 * MS; +enum { TRIM_AFTER = 1024 }; + +struct PureLinkListener { + const PureModule *module; + const PureSealKey *own; + unsigned char own_public[PURE_SEAL_KEY]; + uint32_t own_id; + unsigned port; + uint64_t run, idle, swept, trimmed, untrimmed; + PureAead *aead; /* keyed by each packet's run */ + unsigned char open_key[PURE_AEAD_KEY], seal_key[PURE_AEAD_KEY]; /* the keys it has now */ + bool open_keyed, seal_keyed; + unsigned char *memo; /* the secrets remembered (pure_link_remember), mapped */ + size_t memo_size; + int fd, wake; + pthread_t thread; + bool started; + pthread_mutex_t lock; /* the done list and stopping */ + Done *done; + size_t done_count, done_capacity; + bool stopping; + Peer *peers; + size_t peer_count, peer_capacity; + uint32_t *index; /* the peers by key id: open addressing, a place + 1 (0: empty) */ + size_t index_size; + Allowed *allowed; + size_t allowed_count, allowed_capacity; + unsigned char *names; + size_t names_size, names_capacity; + Active **actives; + size_t active_count, active_capacity; + Stream **ready; /* streams that may have whole records to hand over */ + size_t ready_count, ready_capacity; + Run **due; /* runs to answer */ + size_t due_count, due_capacity; + bool (*deliver)(void *, const PureLinkDelivery *); + void *context; + unsigned char *batch; + size_t batch_capacity; + PureLinkCounts counts; +}; + +PureLinkListener *pure_link_listen(const PureModule *m, unsigned port, const PureSealKey *own, + const unsigned char own_public[PURE_SEAL_KEY], + bool (*deliver)(void *, const PureLinkDelivery *), void *context, char *error, + size_t size) { + PureLinkListener *s = own ? calloc(1, sizeof(*s)) : NULL; + if (!s || pthread_mutex_init(&s->lock, NULL)) { + free(s); + snprintf(error, size, own ? "out of memory" : "--link-port needs the node's --key"); + return NULL; + } + s->module = m; + s->own = own; + memcpy(s->own_public, own_public, PURE_SEAL_KEY); + s->own_id = key_id(own_public); + s->port = port; + s->idle = IDLE_FIRST; + s->deliver = deliver; + s->context = context; + s->fd = s->wake = -1; + return s; +} + +void pure_link_idle(PureLinkListener *s, uint64_t ms) { s->idle = ms * MS; } + +/* items (count used of capacity) with room for one more: the same or grown, + * or NULL when memory runs out (items stay as they were). */ +static void *room(void *items, size_t *capacity, size_t count, size_t item_size, size_t first) { + if (count < *capacity) return items; + size_t more = *capacity ? 2 * *capacity : first; + void *grown = realloc(items, more * item_size); + if (grown) *capacity = more; + return grown; +} + +/* The peers by key id again, in an index twice as large. */ +static bool grow_index(PureLinkListener *s) { + size_t size = s->index_size ? 2 * s->index_size : 1024; + uint32_t *index = calloc(size, sizeof(*index)); + if (!index) return false; + for (size_t p = 0; p < s->peer_count; ++p) { + size_t slot = s->peers[p].id & (size - 1); + while (index[slot]) slot = (slot + 1) & (size - 1); + index[slot] = (uint32_t)p + 1; + } + free(s->index); + s->index = index; + s->index_size = size; + return true; +} + +static bool same_name(const PureLinkListener *s, uint32_t name, const char *source, size_t length) { + return s->names[name] == length && !memcmp(s->names + name + 1, source, length); +} + +/* Keys are found by their ids, which are 32 bits: two keys may share one. */ +bool pure_link_allow(PureLinkListener *s, const char *source, const unsigned char key[PURE_SEAL_KEY], char *error, + size_t size) { + size_t length = strlen(source), p = SIZE_MAX; + uint32_t id = key_id(key); + if (!length || length > PURE_SOURCE) { + snprintf(error, size, "a source's name is 1 to %d bytes", PURE_SOURCE); + return false; + } + for (size_t slot = s->index_size ? id & (s->index_size - 1) : 0; s->index_size && s->index[slot]; + slot = (slot + 1) & (s->index_size - 1)) { + const Peer *peer = &s->peers[s->index[slot] - 1]; + if (peer->id == id && !memcmp(peer->key, key, PURE_SEAL_KEY)) { + p = s->index[slot] - 1; + break; + } + } + if (p == SIZE_MAX) { + Peer *peers = s->peer_count < UINT32_MAX - 1 && ((s->peer_count + 1) * 2 <= s->index_size || grow_index(s)) ? + room(s->peers, &s->peer_capacity, s->peer_count, sizeof(Peer), 64) : NULL; + if (!peers) goto full; + s->peers = peers; + p = s->peer_count++; + s->peers[p] = (Peer){.id = id}; + memcpy(s->peers[p].key, key, PURE_SEAL_KEY); + size_t slot = id & (s->index_size - 1); + while (s->index[slot]) slot = (slot + 1) & (s->index_size - 1); + s->index[slot] = (uint32_t)p + 1; + } + for (uint32_t a = s->peers[p].allowed; a; a = s->allowed[a - 1].next) + if (same_name(s, s->allowed[a - 1].name, source, length)) return true; + Allowed *kept = s->allowed_count < UINT32_MAX - 1 && s->names_size + 1 + length <= UINT32_MAX ? + room(s->allowed, &s->allowed_capacity, s->allowed_count, sizeof(Allowed), 64) : NULL; + if (!kept) goto full; + s->allowed = kept; + if (s->names_size + 1 + length > s->names_capacity) { + size_t more = s->names_capacity ? 2 * s->names_capacity : 4096; + unsigned char *names = realloc(s->names, more); + if (!names) goto full; + s->names = names; + s->names_capacity = more; + } + s->names[s->names_size] = (unsigned char)length; + memcpy(s->names + s->names_size + 1, source, length); + s->allowed[s->allowed_count] = (Allowed){(uint32_t)s->names_size, s->peers[p].allowed}; + s->peers[p].allowed = (uint32_t)++s->allowed_count; + s->names_size += 1 + length; + return true; +full: + snprintf(error, size, "out of memory"); + return false; +} + +static Held *held_at(Stream *st, size_t k) { return &st->held[st->held_first + k]; } + +static uint64_t line_end(const Stream *st) { return st->line_first + st->line_count; } + +/* Room for more bytes at the line's end: its front let go given back + * first when that is at least half of it (so no byte moves twice in a + * doubling), else a longer line. */ +static bool line_room(Stream *st, size_t more) { + if (st->line_size + more <= st->line_capacity) return true; + if (st->line_head && st->line_head >= st->line_capacity / 2) { + memmove(st->line, st->line + st->line_head, st->line_size - st->line_head); + st->line_size -= st->line_head; + st->line_head = 0; + if (st->line_size + more <= st->line_capacity) return true; + } + size_t capacity = st->line_capacity ? st->line_capacity : 4096; + while (capacity < st->line_size + more) capacity *= 2; + unsigned char *grown = realloc(st->line, capacity); + if (!grown) return false; + st->line = grown; + st->line_capacity = capacity; + return true; +} + +static bool line_append(Stream *st, const unsigned char *bytes, uint32_t size) { + if (!line_room(st, 4 + (size_t)size)) return false; + put32(st->line + st->line_size, size); + if (size) memcpy(st->line + st->line_size + 4, bytes, size); + st->line_size += 4 + (size_t)size; + ++st->line_count; + return true; +} + +/* The line's records up to heard let go (all of them, when heard is past it). */ +static void line_drop(Stream *st, uint64_t heard) { + if (heard < st->line_first) return; + if (heard >= line_end(st)) { + st->line_head = st->line_size = 0; + st->line_count = 0; + st->line_first = heard + 1; + return; + } + for (uint64_t n = heard - st->line_first + 1; n; --n, --st->line_count, ++st->line_first) + st->line_head += 4 + (size_t)get32(st->line + st->line_head); +} + +/* Records held that the line reached, moved onto it; ones behind it let go. */ +static void drop_first(Stream *st); +static void line_pull(Stream *st) { + while (st->held_count) { + Held *h = held_at(st, 0); + if (h->seq < line_end(st)) drop_first(st); + else if (h->seq == line_end(st) && h->got == h->pieces && line_append(st, h->bytes, h->size)) drop_first(st); + else break; + } +} + +/* The first record held let go: the ones after it stay where they are. */ +static void drop_first(Stream *st) { + st->held_bytes -= held_at(st, 0)->size; + free(held_at(st, 0)->bytes); + ++st->held_first; + if (!--st->held_count) st->held_first = 0; +} + +/* The memo of secrets (pure_link_remember): a header, then a place a peer + * in the order allowed, each the first 28 bytes of its key, a check (FNV-1a + * of its whole key and the secret) and the secret. A place whose check + * fails (never made, another key, cut short by a crash) is made again. */ +enum { MEMO_HEAD = 64, MEMO_PLACE = 64 }; +static const unsigned char memo_magic[8] = {'P', 'L', 'M', '1', 0, 0, 0, 0}; + +static uint32_t memo_check(const unsigned char key[PURE_SEAL_KEY], const unsigned char secret[32]) { + return fnv(fnv(2166136261u, key, PURE_SEAL_KEY), secret, 32); +} + +bool pure_link_remember(PureLinkListener *s, const char *path, char *error, size_t size) { + int fd = open(path, O_RDWR | O_CREAT | O_NOFOLLOW | O_CLOEXEC, 0600); + struct stat info; + unsigned char head[MEMO_HEAD]; + size_t needed = MEMO_HEAD + (s->peer_count ? s->peer_count : 1) * MEMO_PLACE; + bool ok = fd >= 0 && !fstat(fd, &info) && S_ISREG(info.st_mode) && !(info.st_mode & 077); + /* Another node's key, or none yet: begun again. */ + bool ours = ok && pread(fd, head, MEMO_HEAD, 0) == MEMO_HEAD && !memcmp(head, memo_magic, 8) && + !memcmp(head + 8, s->own_public, PURE_SEAL_KEY); + if (ok && !ours) { + memset(head, 0, MEMO_HEAD); + memcpy(head, memo_magic, 8); + memcpy(head + 8, s->own_public, PURE_SEAL_KEY); + ok = !ftruncate(fd, 0) && pwrite(fd, head, MEMO_HEAD, 0) == MEMO_HEAD; + } + ok = ok && (ours && (size_t)info.st_size >= needed ? true : !ftruncate(fd, (off_t)needed)); + void *map = ok ? mmap(NULL, needed, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0) : MAP_FAILED; + if (fd >= 0) close(fd); + if (map == MAP_FAILED) { + snprintf(error, size, "%s: cannot keep secrets there (a file only its owner may read)", path); + return false; + } + s->memo = map; + s->memo_size = needed; + return true; +} + +/* A peer's active state, made at its first packet: the secret its key + * shares with ours, remembered or made (one X25519, then remembered: the + * secret first, the key that makes the place its own last). */ +static Active *activate(PureLinkListener *s, Peer *peer, uint64_t now) { + Active *a = calloc(1, sizeof(*a)), **actives = room(s->actives, &s->active_capacity, s->active_count, sizeof(*actives), 64); + if (actives) s->actives = actives; + if (!a || !actives) { + free(a); + return NULL; + } + size_t index = (size_t)(peer - s->peers); + unsigned char *place = s->memo && MEMO_HEAD + (index + 1) * MEMO_PLACE <= s->memo_size ? + s->memo + MEMO_HEAD + index * MEMO_PLACE : NULL; + if (place && !memcmp(place, peer->key, 28) && get32(place + 28) == memo_check(peer->key, place + 32)) { + memcpy(a->secret, place + 32, 32); + ++s->counts.remembered; + } else if (pure_link_secret(s->own, peer->key, a->secret)) { + if (place) { + memset(place, 0, 28); + __atomic_thread_fence(__ATOMIC_RELEASE); + memcpy(place + 32, a->secret, 32); + put32(place + 28, memo_check(peer->key, a->secret)); + __atomic_thread_fence(__ATOMIC_RELEASE); + memcpy(place, peer->key, 28); + } + } else { + explicit_bzero(a, sizeof(*a)); + free(a); + return NULL; + } + a->peer = (uint32_t)(peer - s->peers); + a->seen = now; + a->place = s->active_count; + s->actives[s->active_count++] = a; + peer->active = a; + return a; +} + +static void let_go(PureLinkListener *s, Active *a) { + for (unsigned k = 0; k < a->stream_count; ++k) { + while (a->streams[k]->held_count) drop_first(a->streams[k]); + free(a->streams[k]->held); + free(a->streams[k]->line); + free(a->streams[k]); + } + free(a->streams); + s->peers[a->peer].active = NULL; + s->actives[a->place] = s->actives[--s->active_count]; + s->actives[a->place]->place = a->place; + explicit_bzero(a, sizeof(*a)); + free(a); +} + +/* Whether any of a peer's records or answers waits here. */ +static bool busy(const Active *a) { + for (unsigned k = 0; k < a->stream_count; ++k) + if (a->streams[k]->out || a->streams[k]->held_count || a->streams[k]->line_count || a->streams[k]->ready) return true; + for (unsigned i = 0; i < RUNS; ++i) + if (a->runs[i].due) return true; + return false; +} + +/* Let go the active state of peers quiet for the idle period; what idle + * links held goes back to the system. */ +static void sweep(PureLinkListener *s, uint64_t now) { + if (now - s->swept < SWEEP_EVERY) return; + s->swept = now; + size_t freed = 0; + for (size_t k = 0; k < s->active_count;) { + Active *a = s->actives[k]; + if (now - a->seen < s->idle || busy(a)) ++k; + else let_go(s, a), ++freed; + } + s->counts.swept += freed; + s->untrimmed += freed; + if (s->untrimmed >= TRIM_AFTER || (s->untrimmed && now - s->trimmed >= TRIM_WAIT)) { + malloc_trim(0); + s->untrimmed = 0; + s->trimmed = now; + } +} + +static void mark_ready(PureLinkListener *s, Stream *st) { + Stream **ready = st->ready ? NULL : room(s->ready, &s->ready_capacity, s->ready_count, sizeof(*ready), 64); + if (!ready) return; + s->ready = ready; + s->ready[s->ready_count++] = st; + st->ready = true; +} + +static void mark_due(PureLinkListener *s, Run *r) { + Run **due = r->due ? NULL : room(s->due, &s->due_capacity, s->due_count, sizeof(*due), 64); + if (!due) return; + s->due = due; + s->due[s->due_count++] = r; + r->due = true; +} + +static Run *run_found(Active *a, uint64_t run) { + for (unsigned i = 0; i < RUNS; ++i) + if (a->runs[i].used && a->runs[i].run == run) return &a->runs[i]; + return NULL; +} + +/* A new run's keys, in the slot unused longest: only once its first packet + * opened with them, so a forged one displaces nothing. */ +static Run *install_run(Active *a, const Run *made) { + Run *slot = &a->runs[0]; + for (unsigned i = 0; i < RUNS; ++i) { + Run *r = &a->runs[i]; + if (!r->used || (slot->used && r->seen < slot->seen)) slot = r; + } + bool due = slot->due; /* the answer list may hold the slot */ + *slot = *made; + slot->due = due; + return slot; +} + +static bool had(const Run *r, uint64_t pn) { + for (unsigned i = 0; i < r->got_count; ++i) + if (pn >= r->got[i].first && pn <= r->got[i].last) return true; + return false; +} + +static void note(Run *r, uint64_t pn, uint64_t now) { + unsigned i = 0, n = r->got_count; + while (i < n && r->got[i].first > pn + 1) ++i; + if (i < n && r->got[i].first == pn + 1) { /* joins the range above it */ + r->got[i].first = pn; + if (i + 1 < n && r->got[i + 1].last + 1 == pn) { + r->got[i].first = r->got[i + 1].first; + memmove(&r->got[i + 1], &r->got[i + 2], (n - i - 2) * sizeof(Range)); + --r->got_count; + } + } else if (i < n && r->got[i].last + 1 == pn) { + r->got[i].last = pn; + } else if (i < RANGES) { + if (n == RANGES) --n; /* the oldest range is forgotten */ + memmove(&r->got[i + 1], &r->got[i], (n - i) * sizeof(Range)); + r->got[i] = (Range){pn, pn}; + r->got_count = n + 1; + } + if (r->got_count && r->got[0].last == pn) r->largest_at = now; +} + +static bool allowed(const PureLinkListener *s, const Peer *peer, const char *source) { + for (uint32_t a = peer->allowed; a; a = s->allowed[a - 1].next) + if (same_name(s, s->allowed[a - 1].name, source, strlen(source))) return true; + return false; +} + +/* The stream of a peer's source to input, made the first time: NULL when the + * source is not allowed from the peer, or input takes no such messages. */ +static Stream *stream_of(PureLinkListener *s, Active *a, const char *source, const char *input, uint32_t shape) { + int index = pure_module_input(s->module, input); + if (index < 0) return NULL; + for (unsigned k = 0; k < a->stream_count; ++k) { + Stream *st = a->streams[k]; + if (st->input == (uint32_t)index && !strcmp(st->source, source)) return st; + } + const PureInputShape *in = &s->module->input[index]; + if (!allowed(s, &s->peers[a->peer], source) || in->kind != PURE_KIND_MESSAGE || in->keyed || + shape_of(s->module, (uint32_t)index) != shape) + return NULL; + size_t capacity = a->stream_capacity; + Stream **streams = room(a->streams, &capacity, a->stream_count, sizeof(*streams), 2), *st; + if (!streams) return NULL; + a->streams = streams; + a->stream_capacity = (unsigned)capacity; + if (!(st = calloc(1, sizeof(*st)))) return NULL; + st->active = a; + st->input = (uint32_t)index; + snprintf(st->source, sizeof(st->source), "%s", source); + a->streams[a->stream_count++] = st; + return st; +} + +/* A record's piece, kept until the record is whole and handed over. */ +static void hold(Stream *st, uint64_t seq, uint32_t size, uint32_t offset, const unsigned char *bytes, uint32_t length) { + if (st->known && seq <= st->heard) { + st->stored_due = true; /* sent again: say again how far it is stored */ + return; + } + if ((st->known && seq > st->heard + CREDIT) || size > RECORD_MOST) return; + if (st->known && seq < line_end(st)) return; /* on the line already */ + if (st->known && seq == line_end(st) && !offset && length == size) { /* the next, whole */ + if (st->held_bytes + (st->line_size - st->line_head) + size <= HELD_MOST && line_append(st, bytes, size)) + line_pull(st); + return; + } + /* In order it goes last; out of order, its place is found by halving. */ + size_t k = st->held_count; + if (k && held_at(st, k - 1)->seq >= seq) { + size_t low = 0; + while (low < k) { + size_t middle = low + (k - low) / 2; + if (held_at(st, middle)->seq < seq) low = middle + 1; + else k = middle; + } + } + if (k == st->held_count || held_at(st, k)->seq != seq) { + if (st->held_bytes + size > HELD_MOST) return; + if (st->held_first && st->held_first + st->held_count == st->held_capacity) { /* the front is free: used again */ + memmove(st->held, held_at(st, 0), st->held_count * sizeof(Held)); + st->held_first = 0; + } + Held *kept = room(st->held, &st->held_capacity, st->held_first + st->held_count, sizeof(Held), 16); + if (!kept) return; + st->held = kept; + Held h = {.seq = seq, .size = size, .pieces = size ? (size + PIECE - 1) / PIECE : 1}; + if (!(h.bytes = malloc(size + h.pieces))) return; + memset(h.bytes + size, 0, h.pieces); + memmove(held_at(st, k + 1), held_at(st, k), (st->held_count - k) * sizeof(Held)); + *held_at(st, k) = h; + ++st->held_count; + st->held_bytes += size; + } + Held *h = held_at(st, k); + if (h->size != size) return; + if (!offset && length == size) { /* whole: every piece had */ + if (size) memcpy(h->bytes, bytes, size); + memset(h->bytes + size, 1, h->pieces); + h->got = h->pieces; + return; + } + uint32_t piece = offset / PIECE; + if (offset % PIECE || piece >= h->pieces || length != least(PIECE, size - offset) || h->bytes[size + piece]) return; + if (length) memcpy(h->bytes + offset, bytes, length); + h->bytes[size + piece] = 1; + ++h->got; +} + +/* Whole records from the one after the last heard, in order, to the host: + * one delivery a stream at a time. Before the last heard is known (after a + * restart) they go from the first held: the host says how far it heard. + * False when the host could not take them now. */ +static bool hand_over(PureLinkListener *s, Stream *st) { + if (st->out) return true; + if (st->known) { /* the line as it is, up to BATCH records */ + line_pull(st); + if (!st->line_count) return true; + unsigned n = 0; + size_t bytes = 0; + for (; n < st->line_count && n < BATCH; ++n) bytes += 4 + (size_t)get32(st->line + st->line_head + bytes); + PureLinkDelivery d = {st->input, st->source, st->line_first, n, st->line + st->line_head, bytes, st}; + if (!s->deliver(s->context, &d)) return false; + st->out = true; + ++s->counts.deliveries; + return true; + } + if (!st->held_count || (st->known && held_at(st, 0)->seq != st->heard + 1)) return true; + size_t n = 0, bytes = 0; + while (n < st->held_count && n < BATCH && held_at(st, n)->got == held_at(st, n)->pieces && + held_at(st, n)->seq == held_at(st, 0)->seq + n) + bytes += 4 + held_at(st, n++)->size; + if (!n) return true; + if (bytes > s->batch_capacity) { + unsigned char *grown = realloc(s->batch, bytes); + if (!grown) return false; + s->batch = grown; + s->batch_capacity = bytes; + } + size_t at = 0; + for (size_t k = 0; k < n; ++k) { + put32(s->batch + at, held_at(st, k)->size); + memcpy(s->batch + at + 4, held_at(st, k)->bytes, held_at(st, k)->size); + at += 4 + held_at(st, k)->size; + } + PureLinkDelivery d = {st->input, st->source, held_at(st, 0)->seq, (unsigned)n, s->batch, bytes, st}; + if (!s->deliver(s->context, &d)) return false; + st->out = true; + ++s->counts.deliveries; + return true; +} + +/* A packet from peer, if it opens with the keys of peer's run (made for a + * run not seen yet, and for a peer at rest its active state): its frames. */ +static bool take_from(PureLinkListener *s, Peer *peer, unsigned char *packet, size_t n, + const struct sockaddr_storage *from, socklen_t from_size, uint64_t now) { + uint64_t pn = get64(packet + 13), run = get64(packet + 5); + Active *a = peer->active; + Run *r = a ? run_found(a, run) : NULL; + if (r && had(r, pn)) { + r->ack_due = true; /* its acknowledgment was lost, perhaps */ + mark_due(s, r); + return true; + } + bool woke = !a; + if (woke && !(a = activate(s, peer, now))) return false; + Run made = {.active = a, .run = run, .seen = now, .used = true}; + bool keyed = r || (derive(a->secret, peer->key, s->own_public, run, 0, made.in_key, made.in_base) && + derive(a->secret, s->own_public, peer->key, s->run, run, made.out_key, made.out_base)); + unsigned char plain[DATAGRAM], nonce[PURE_AEAD_NONCE]; + nonce_of(nonce, r ? r->in_base : made.in_base, pn); + size_t size = n - HEADER - PURE_AEAD_TAG; + const unsigned char *key = r ? r->in_key : made.in_key, *give = key; + if (s->open_keyed && !memcmp(s->open_key, key, PURE_AEAD_KEY)) give = NULL; /* the key it has */ + else if (keyed) memcpy(s->open_key, key, PURE_AEAD_KEY), s->open_keyed = true; + if (!keyed || !pure_aead_open(s->aead, give, nonce, packet, HEADER, packet + HEADER, n - HEADER, plain)) { + explicit_bzero(&made, sizeof(made)); + if (woke) let_go(s, a); + return false; + } + if (woke) ++s->counts.woken; + if (!r) r = install_run(a, &made); + a->seen = r->seen = now; + note(r, pn, now); + r->address_size = from_size <= sizeof(r->address) ? from_size : sizeof(r->address); + memcpy(&r->address, from, r->address_size); /* where it is now: answers go there */ + ++s->counts.received; + Stream *st = NULL; + bool eliciting = false; + for (size_t at = 0; at < size;) { + unsigned type = plain[at++]; + if (type == F_STREAM) { + char source[PURE_SOURCE + 1], input[PURE_MODULE_NAME]; + size_t names = get_names(plain + at, size - at, source, input); + if (!names || at + names + 4 > size) break; + st = stream_of(s, a, source, input, get32(plain + at + names)); + if (st) { + st->run = r; + st->run_id = r->run; + } + at += names + 4; + eliciting = true; + } else if (type == F_RECORD) { + if (at + RECORD_HEAD - 1 > size) break; + uint32_t length = get16(plain + at + 16); + if (at + RECORD_HEAD - 1 + length > size) break; + if (st) { + hold(st, get64(plain + at), get32(plain + at + 8), get32(plain + at + 12), plain + at + 18, length); + mark_ready(s, st); + } + at += RECORD_HEAD - 1 + length; + eliciting = true; + } else if (type == F_RECORDS) { + if (at + RECORDS_HEAD - 1 > size) break; + uint64_t first = get64(plain + at), length = 0; + unsigned count = plain[at + 8]; + at += RECORDS_HEAD - 1; + unsigned i = 0; + for (size_t used; i < count && (used = get_varint(plain + at, size - at, &length)) && + length <= size - at - used; ++i) { + if (st) hold(st, first + i, (uint32_t)length, 0, plain + at + used, (uint32_t)length); + at += used + length; + } + if (st) mark_ready(s, st); + eliciting = true; + if (i < count) break; /* cut short: the rest is not read */ + } else if (type == F_PING) { + eliciting = true; + } else + break; /* padding, or what this side does not know: the rest */ + } + if (eliciting) { + r->ack_due = true; + mark_due(s, r); + } + return true; +} + +static void take_packet(PureLinkListener *s, unsigned char *packet, size_t n, const struct sockaddr_storage *from, + socklen_t from_size, uint64_t now) { + if (n < HEADER + PURE_AEAD_TAG || n > DATAGRAM || packet[0] != VERSION || !s->index_size) return; + uint32_t id = get32(packet + 1); + size_t mask = s->index_size - 1; + uint64_t run = get64(packet + 5); + /* The peers with the packet's key id: first one whose run is known (no + * key to make), then any. */ + for (size_t slot = id & mask; s->index[slot]; slot = (slot + 1) & mask) { + Peer *peer = &s->peers[s->index[slot] - 1]; + if (peer->id == id && peer->active && run_found(peer->active, run) && + take_from(s, peer, packet, n, from, from_size, now)) + return; + } + for (size_t slot = id & mask; s->index[slot]; slot = (slot + 1) & mask) { + Peer *peer = &s->peers[s->index[slot] - 1]; + if (peer->id == id && !(peer->active && run_found(peer->active, run)) && + take_from(s, peer, packet, n, from, from_size, now)) + return; + } +} + +/* Acknowledgments and how far each stream is stored, a packet to each run + * that has any. How far a stream is stored goes with every acknowledgment + * to its run, so a lost word is said again at the next packet. */ +static void answer(PureLinkListener *s, uint64_t now) { + unsigned char body[BODY], packet[DATAGRAM]; + for (size_t k = 0; k < s->due_count; ++k) { + Run *r = s->due[k]; + Active *a = r->active; + r->due = false; + if (!r->used) continue; + size_t at = 0; + if (r->ack_due && r->got_count) { + body[at++] = F_ACK; + put64(body + at, r->got[0].last); + put32(body + at + 8, (uint32_t)least((now - r->largest_at) / 1000, UINT32_MAX)); + body[at + 12] = (unsigned char)r->got_count; + at += 13; + for (unsigned i = 0; i < r->got_count; ++i, at += 16) { + put64(body + at, r->got[i].first); + put64(body + at + 8, r->got[i].last); + } + } + for (unsigned j = 0; j < a->stream_count; ++j) { + Stream *st = a->streams[j]; + if (!(st->stored_due || r->ack_due) || st->run != r || st->run_id != r->run || !st->known) continue; + if (at + 2 + PURE_SOURCE + PURE_MODULE_NAME + 14 > BODY) break; + body[at++] = F_STORED; + at += put_names(body + at, st->source, s->module->input_names[st->input]); + put64(body + at, st->heard); + body[at + 8] = st->gap ? STORED_GAP : 0; + size_t room = st->held_bytes + (st->line_size - st->line_head) < HELD_MOST ? CREDIT : 1; + put32(body + at + 9, (uint32_t)room); + at += 13; + st->stored_due = st->gap = false; + } + r->ack_due = false; + if (!at) continue; + const unsigned char *give = r->out_key; + if (s->seal_keyed && !memcmp(s->seal_key, r->out_key, PURE_AEAD_KEY)) give = NULL; /* the key it has */ + else memcpy(s->seal_key, r->out_key, PURE_AEAD_KEY), s->seal_keyed = true; + size_t total = seal_packet(s->aead, give, r->out_base, s->own_id, s->run, r->next_pn++, body, at, packet); + if (total && sendto(s->fd, packet, total, 0, (const struct sockaddr *)&r->address, r->address_size) >= 0) + ++s->counts.packets; + } + s->due_count = 0; +} + +static void take_done(PureLinkListener *s) { + pthread_mutex_lock(&s->lock); + Done *done = s->done; + size_t count = s->done_count; + s->done = NULL; + s->done_count = s->done_capacity = 0; + pthread_mutex_unlock(&s->lock); + for (size_t k = 0; k < count; ++k) { + Stream *st = done[k].stream; + st->out = false; + mark_ready(s, st); + if (!done[k].stored) continue; /* not saved: handed over again */ + if (!st->known || done[k].heard > st->heard) st->heard = done[k].heard; + if (!st->known) st->line_first = st->heard + 1; /* the line begins */ + line_drop(st, st->heard); + while (st->held_count && held_at(st, 0)->seq <= st->heard) drop_first(st); + st->known = true; + st->gap = done[k].gap; + st->stored_due = true; + if (st->run && st->run->used && st->run_id == st->run->run) mark_due(s, st->run); + s->counts.stored = st->heard; + } + free(done); +} + +/* The streams ready, each handed over; one the host could not take now is + * tried again at the next turn. */ +static void hand_over_ready(PureLinkListener *s) { + size_t count = s->ready_count; + for (size_t k = 0; k < count; ++k) { + Stream *st = s->ready[k]; + st->ready = false; + if (!hand_over(s, st)) mark_ready(s, st); + } + memmove(s->ready, s->ready + count, (s->ready_count - count) * sizeof(*s->ready)); + s->ready_count -= count; +} + +static void *run_listener(void *context) { + PureLinkListener *s = context; + Gathering *gathering = malloc(sizeof(*gathering)); + if (!gathering) return NULL; + for (;;) { + struct pollfd waits[2] = {{s->fd, POLLIN, 0}, {s->wake, POLLIN, 0}}; + int timeout = s->active_count || s->ready_count ? (int)(SWEEP_EVERY / MS) : -1; + if (poll(waits, 2, timeout) < 0 && errno != EINTR) break; + pthread_mutex_lock(&s->lock); + bool stopping = s->stopping; + pthread_mutex_unlock(&s->lock); + if (stopping) break; + if (waits[1].revents & POLLIN) { + uint64_t signals; + if (read(s->wake, &signals, sizeof(signals)) < 0) { /* cleared already */ } + } + uint64_t now = now_ns(); + for (unsigned rounds = 0, n = GATHER; rounds < 8 && n == GATHER; ++rounds) { + n = (unsigned)gather(s->fd, gathering); + for (unsigned k = 0; k < n; ++k) { + struct mmsghdr *m = &gathering->messages[k]; + size_t length = m->msg_len, step = segment_of(&m->msg_hdr, length); + for (size_t at = 0; at < length; at += step) + take_packet(s, gathering->bytes[k] + at, least(step, length - at), &gathering->from[k], + m->msg_hdr.msg_namelen, now); + } + } + take_done(s); + hand_over_ready(s); + answer(s, now); + sweep(s, now); + } + free(gathering); + return NULL; +} + +bool pure_link_serve(PureLinkListener *s, char *error, size_t size) { + struct sockaddr_in6 six = {.sin6_family = AF_INET6, .sin6_port = htons((uint16_t)s->port), .sin6_addr = in6addr_any}; + int off = 0, buffer = 4 << 20; + s->fd = socket(AF_INET6, SOCK_DGRAM | SOCK_NONBLOCK | SOCK_CLOEXEC, 0); + bool bound = s->fd >= 0 && !setsockopt(s->fd, IPPROTO_IPV6, IPV6_V6ONLY, &off, sizeof(off)) && + !bind(s->fd, (struct sockaddr *)&six, sizeof(six)); + if (!bound) { /* no IPv6 here */ + if (s->fd >= 0) close(s->fd); + struct sockaddr_in four = {.sin_family = AF_INET, .sin_port = htons((uint16_t)s->port), .sin_addr = {INADDR_ANY}}; + s->fd = socket(AF_INET, SOCK_DGRAM | SOCK_NONBLOCK | SOCK_CLOEXEC, 0); + bound = s->fd >= 0 && !bind(s->fd, (struct sockaddr *)&four, sizeof(four)); + } + if (!bound) { + snprintf(error, size, "--link-port %u: %s", s->port, strerror(errno)); + return false; + } + setsockopt(s->fd, SOL_SOCKET, SO_RCVBUF, &buffer, sizeof(buffer)); + setsockopt(s->fd, SOL_SOCKET, SO_SNDBUF, &buffer, sizeof(buffer)); + int one = 1; /* without GRO, a buffer is a datagram */ + setsockopt(s->fd, SOL_UDP, UDP_GRO, &one, sizeof(one)); + s->run = random64(); + s->swept = s->trimmed = now_ns(); + s->wake = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC); + unsigned char key[PURE_AEAD_KEY] = {0}; /* each packet brings its run's */ + if (s->wake < 0 || !(s->aead = pure_aead(key)) || pthread_create(&s->thread, NULL, run_listener, s)) { + snprintf(error, size, "--link-port %u: cannot start", s->port); + return false; + } + pthread_setname_np(s->thread, "pure-listen"); + s->started = true; + return true; +} + +void pure_link_done(PureLinkListener *s, void *token, uint64_t heard, bool stored, bool gap) { + pthread_mutex_lock(&s->lock); + if (s->done_count == s->done_capacity) { + size_t capacity = s->done_capacity ? 2 * s->done_capacity : 16; + Done *grown = realloc(s->done, capacity * sizeof(*grown)); + if (!grown) abort(); /* a delivery's word cannot be lost */ + s->done = grown; + s->done_capacity = capacity; + } + s->done[s->done_count++] = (Done){token, heard, stored, gap}; + pthread_mutex_unlock(&s->lock); + wake(s->wake); +} + +void pure_link_listener_close(PureLinkListener *s, PureLinkCounts *counts) { + if (!s) return; + if (s->started) { + pthread_mutex_lock(&s->lock); + s->stopping = true; + pthread_mutex_unlock(&s->lock); + wake(s->wake); + pthread_join(s->thread, NULL); + } + if (counts) { + *counts = s->counts; + counts->active = s->active_count; + } + while (s->active_count) let_go(s, s->actives[0]); + pure_aead_free(s->aead); + if (s->memo) munmap(s->memo, s->memo_size); + free(s->actives); + free(s->peers); + free(s->index); + free(s->allowed); + free(s->names); + free(s->ready); + free(s->due); + free(s->batch); + free(s->done); + if (s->fd >= 0) close(s->fd); + if (s->wake >= 0) close(s->wake); + pthread_mutex_destroy(&s->lock); + free(s); +} diff --git a/link/link.h b/link/link.h new file mode 100644 index 0000000..83cc645 --- /dev/null +++ b/link/link.h @@ -0,0 +1,118 @@ +#ifndef PURE_LINK_H +#define PURE_LINK_H +#include "module.h" +#include "seal.h" + +/* Links without connections (docs/design/011-pure-transport.md): one node's + * saved sends to another node's input over UDP, sealed with a key the two + * nodes' keys share, with no handshake: the first packet carries records. + * A link's lasting state is the apps' own: the sender's outbox (what to send + * and send again) and the receiver's heard numbers (what it applied, saved + * with the events). What either side keeps in memory is a cache, so either + * may restart and the other goes on with the next packet. + * + * A packet is a header in the clear, then frames sealed with AES-128-GCM + * (the header as associated data): + * u8 0x70 (version 0), u32le the key id of the node that sent it (FNV-1a + * of its public key; two keys may share an id, and a receiver tries the + * keys it allows with it), u64le that node's run (random, new each time it + * starts, so no nonce repeats under a key across restarts), u64le the + * packet's number in its run. + * Keys come from the secret the two nodes' keys share (X25519) by + * HKDF-SHA256, with both public keys and the sending node's run: a sender's + * packets use its own run; a receiver's answers add the run they answer, + * so an answer opens only in that run. The nonce is the key's base with the + * packet's number in its last eight bytes (XOR, big-endian, as QUIC does). + * + * Frames (numbers little-endian): + * 0x01 STREAM source (u8 length, bytes), input (u8 length, bytes), u32 the + * shape of the messages (the sender's output and the + * receiver's input must have the same): the stream the + * records after it are of + * 0x02 RECORD u64 number, u32 size, u32 offset, u16 length, bytes: a piece + * of a saved send (pieces are 1000 bytes, the last shorter) + * 0x03 ACK u64 largest number, u32 delay (µs), u8 ranges, each as u64 + * first and u64 last, newest first + * 0x04 STORED source, input (as STREAM), u64 the last number heard + * (applied and saved), u8 flags (1: a record after a gap was + * not taken: send what follows it again), u32 credit (records + * it takes past that) + * 0x05 PING to be acknowledged + * 0x06 RECORDS u64 the first one's number, u8 count, then whole records one + * after another (numbers following on), each its size (a + * varint: LEB128) and its bytes: a send that fits the packet + * whole, with 1 to 3 bytes of framing, not the 19 of RECORD + * 0x00 PADDING to the end of the packet: a full packet sent in a burst is + * a whole datagram + * Losses are detected as RFC 9002 says; congestion control is its NewReno + * (RFC 9002 §7). */ + +/* The sending side: one output's saved sends to url, pure://host:port/INPUT, + * as source (a Pure receiver's name for it), from own (this node's key, + * whose public half is own_public) to the receiver's public key, given by + * pure_link_peer before pure_link_start. The owner queues and trims as it + * does a wire's (serve/deliver.h): a send counts as delivered once the + * receiver says it saved it. */ +typedef struct PureLink PureLink; +PureLink *pure_link_open(const PureModule *m, const char *output, const char *url, const char *source, + const PureSealKey *own, const unsigned char own_public[PURE_SEAL_KEY], char *error, + size_t size); +void pure_link_peer(PureLink *, const unsigned char peer[PURE_SEAL_KEY]); +uint32_t pure_link_output(const PureLink *); +bool pure_link_queue(PureLink *, uint64_t seq, PureSpan payload); +uint64_t pure_link_queued(PureLink *); +/* The last number the receiver saved; any thread. */ +uint64_t pure_link_delivered(PureLink *); +/* Send on a thread of the link's own, calling notify when more was saved. */ +bool pure_link_start(PureLink *, void (*notify)(void *), void *context, char *error, size_t size); +void pure_link_close(PureLink *); + +/* The receiving side: packets on a UDP port, from the senders allowed (a + * source name and the key it seals with, a few bytes each however many + * there are: see pure_link_idle). Whole records of a stream, in + * order, go to deliver (on the listener's thread) as one delivery; the host + * applies them from the one after the source's last heard number, stops at + * a gap, saves, and then calls pure_link_done (any thread) with the last + * number heard. Records wait here until then, and only the saved ones are + * acknowledged as stored. */ +typedef struct { + uint32_t input; + const char *source; + uint64_t first; + unsigned count; + const unsigned char *records; /* count of: u32le size, bytes */ + size_t size; + void *token; /* for pure_link_done */ +} PureLinkDelivery; +typedef struct PureLinkListener PureLinkListener; +PureLinkListener *pure_link_listen(const PureModule *m, unsigned port, const PureSealKey *own, + const unsigned char own_public[PURE_SEAL_KEY], + bool (*deliver)(void *, const PureLinkDelivery *), void *context, char *error, + size_t size); +bool pure_link_allow(PureLinkListener *, const char *source, const unsigned char key[PURE_SEAL_KEY], char *error, + size_t size); +/* A sender quiet for ms (30 s at first) with nothing of it waiting lets its + * keys, addresses and streams go: an idle link costs the listener only its + * allowing, and the sender's next packet makes them again (one X25519). */ +void pure_link_idle(PureLinkListener *, uint64_t ms); +/* Remember the secret each sender's key shares with the node's in a file + * (path, 0600; begun again for another node's key): a place a sender in + * the order allowed, mapped, so a sender at rest takes no memory for it. + * A node that starts again then makes no X25519 for a sender it heard from + * before. After the senders are allowed, before serving. */ +bool pure_link_remember(PureLinkListener *, const char *path, char *error, size_t size); +bool pure_link_serve(PureLinkListener *, char *error, size_t size); +/* stored: the delivery's save succeeded; gap: it stopped at a gap. */ +void pure_link_done(PureLinkListener *, void *token, uint64_t heard, bool stored, bool gap); + +/* Counts for a host's report: packets and records sent, records sent again, + * packets lost and acknowledged, packets received, deliveries handed over, + * the last number stored; a listener's senders woken (their active state + * made), of them with a secret remembered (no X25519), let go when idle, + * and active when it closed. A listener's are given when it closes. */ +typedef struct { + uint64_t packets, records, resent, lost, acknowledged, received, deliveries, stored, woken, remembered, swept, active; +} PureLinkCounts; +void pure_link_counts(PureLink *, PureLinkCounts *); +void pure_link_listener_close(PureLinkListener *, PureLinkCounts *counts); +#endif diff --git a/link/seal.h b/link/seal.h new file mode 100644 index 0000000..9245f25 --- /dev/null +++ b/link/seal.h @@ -0,0 +1,80 @@ +#ifndef PURE_SEAL_H +#define PURE_SEAL_H +#include <stdbool.h> +#include <stddef.h> +#include <stdint.h> + +/* Sends sealed end to end (docs/research/086-sealed-sends.md, 088): a + * receiving node's key opens what a sender sealed to its public key, and + * nothing between them (a relay, a proxy, a mailbox) can read or change it. + * HPKE (RFC 9180): X25519, HKDF-SHA256, and AES-128-GCM where the CPU has AES + * instructions, else ChaCha20-Poly1305; auth mode when the sender has a key + * of its own, so the receiver knows who sealed. A sender keeps one session + * (one encapsulation) and numbers its seals in it; a receiver opens a + * session once and then each seal by its number. A private key lives in + * OpenSSL's secure heap: locked in memory, out of core dumps, cleared when + * freed. OpenSSL comes with pure-crypto.so (link/crypto.c), mapped the + * first time a key is read or a session made. */ +enum { PURE_SEAL_KEY = 32, PURE_SEAL_OVERHEAD = 16 }; /* a public key; a seal's tag */ + +typedef struct PureSealKey PureSealKey; +typedef struct PureSealer PureSealer; /* a sender's session */ +typedef struct PureOpener PureOpener; /* a receiver's session */ + +/* Whether this machine can seal (pure-crypto.so maps), with why not in error. */ +bool pure_seal_ready(char *error, size_t size); +/* A node's key: read from path, or made and written there (0600) when there + * is none; its public half in hex. NULL with why in error. */ +PureSealKey *pure_seal_key(const char *path, char public_hex[2 * PURE_SEAL_KEY + 1], char *error, size_t size); +void pure_seal_key_free(PureSealKey *); +/* A public key from its hex. */ +bool pure_seal_peer(const char *hex, unsigned char out[PURE_SEAL_KEY]); +/* The suite this machine seals with, as the X-Pure-Seal header names it. */ +const char *pure_seal_suite(void); + +/* A session sealing to peer, authenticated by own when given: its + * encapsulated key into enc (what a receiver opens the session with). */ +PureSealer *pure_sealer(const unsigned char peer[PURE_SEAL_KEY], const PureSealKey *own, + unsigned char enc[PURE_SEAL_KEY]); +/* plain sealed as the session's next seal, bound to aad: size + + * PURE_SEAL_OVERHEAD bytes into out, and its number into n. */ +bool pure_seal(PureSealer *, const char *aad, const unsigned char *plain, size_t size, unsigned char *out, + uint64_t *n); +void pure_sealer_free(PureSealer *); + +/* A session a sender made to key's public half, in suite (this machine's or + * the other known one), from its enc; from is the sender's public key when + * it sealed with its own (NULL otherwise). NULL when it does not open. */ +PureOpener *pure_opener(const PureSealKey *key, const char *suite, const unsigned char enc[PURE_SEAL_KEY], + const unsigned char *from); +/* The session's seal number n: size - PURE_SEAL_OVERHEAD bytes into plain; + * false when it does not open (another session, another aad, or changed on + * the way). */ +bool pure_unseal(PureOpener *, uint64_t n, const char *aad, const unsigned char *sealed, size_t size, + unsigned char *plain); +void pure_opener_free(PureOpener *); + +/* Links (link/link.h): the secret two nodes share (X25519 of this node's + * private key and the other's public key: the same on both sides, and + * never all zeros), keys from it (HKDF-SHA256 with the info given), and + * AES-128-GCM under such a key with the nonce the caller gives (sealing may + * be in place: out may be plain). A context is used by one thread at a time; + * sealing or opening with a key given uses that key from then on, so one + * context serves many links (a key costs about as much as a small packet), + * and NULL keeps the key it has. */ +typedef struct PureAead PureAead; +enum { PURE_AEAD_KEY = 16, PURE_AEAD_NONCE = 12, PURE_AEAD_TAG = 16 }; +bool pure_link_secret(const PureSealKey *own, const unsigned char peer[PURE_SEAL_KEY], unsigned char out[32]); +bool pure_link_derive(const unsigned char secret[32], const unsigned char *info, size_t info_size, unsigned char *out, + size_t out_size); +PureAead *pure_aead(const unsigned char key[PURE_AEAD_KEY]); +/* size + PURE_AEAD_TAG bytes into out. */ +bool pure_aead_seal(PureAead *, const unsigned char *key, const unsigned char nonce[PURE_AEAD_NONCE], + const unsigned char *aad, size_t aad_size, const unsigned char *plain, size_t size, + unsigned char *out); +/* size - PURE_AEAD_TAG bytes into plain; false when it does not open. */ +bool pure_aead_open(PureAead *, const unsigned char *key, const unsigned char nonce[PURE_AEAD_NONCE], + const unsigned char *aad, size_t aad_size, const unsigned char *sealed, size_t size, + unsigned char *plain); +void pure_aead_free(PureAead *); +#endif diff --git a/link/tls.h b/link/tls.h new file mode 100644 index 0000000..3939f21 --- /dev/null +++ b/link/tls.h @@ -0,0 +1,18 @@ +#ifndef PURE_TLS_H +#define PURE_TLS_H +#include <stdbool.h> +#include <stddef.h> + +/* TLS to https receivers (link/crypto.c): one client context, verifying a + * receiver's certificate against the system's CAs or ca_file, and its name. */ +typedef struct PureTls PureTls; +/* The context, made once; false with why in error when it cannot be. */ +bool pure_tls_start(const char *ca_file, char *error, size_t size); +/* A session over a connected socket to host (a DNS name, or an IP address + * written as one), its certificate verified; NULL when it fails. */ +PureTls *pure_tls_open(const char *host, int fd); +/* As SSL_read and SSL_write: the bytes moved, or <= 0. */ +int pure_tls_read(PureTls *, void *bytes, int size); +int pure_tls_write(PureTls *, const void *bytes, int size); +void pure_tls_close(PureTls *); +#endif |