Index: head/sys/kern/uipc_ktls.c =================================================================== --- head/sys/kern/uipc_ktls.c (revision 352813) +++ head/sys/kern/uipc_ktls.c (revision 352814) @@ -1,1450 +1,1480 @@ /*- * SPDX-License-Identifier: BSD-2-Clause * * Copyright (c) 2014-2019 Netflix Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_rss.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) #include #endif #include #ifdef RSS #include #include #endif #if defined(INET) || defined(INET6) #include #include #endif #include #include #include #include #include #include struct ktls_wq { struct mtx mtx; STAILQ_HEAD(, mbuf_ext_pgs) head; bool running; } __aligned(CACHE_LINE_SIZE); static struct ktls_wq *ktls_wq; static struct proc *ktls_proc; LIST_HEAD(, ktls_crypto_backend) ktls_backends; static struct rmlock ktls_backends_lock; static uma_zone_t ktls_session_zone; static uint16_t ktls_cpuid_lookup[MAXCPU]; SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW, 0, "Kernel TLS offload"); SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW, 0, "Kernel TLS offload stats"); static int ktls_allow_unload; SYSCTL_INT(_kern_ipc_tls, OID_AUTO, allow_unload, CTLFLAG_RDTUN, &ktls_allow_unload, 0, "Allow software crypto modules to unload"); #ifdef RSS static int ktls_bind_threads = 1; #else static int ktls_bind_threads; #endif SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN, &ktls_bind_threads, 0, "Bind crypto threads to cores or domains at boot"); static u_int ktls_maxlen = 16384; SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RWTUN, &ktls_maxlen, 0, "Maximum TLS record size"); static int ktls_number_threads; SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD, &ktls_number_threads, 0, "Number of TLS threads in thread-pool"); static bool ktls_offload_enable; SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RW, &ktls_offload_enable, 0, "Enable support for kernel TLS offload"); static bool ktls_cbc_enable = true; SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RW, &ktls_cbc_enable, 1, "Enable Support of AES-CBC crypto for kernel TLS"); static counter_u64_t ktls_tasks_active; SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD, &ktls_tasks_active, "Number of active tasks"); static counter_u64_t ktls_cnt_on; SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, so_inqueue, CTLFLAG_RD, &ktls_cnt_on, "Number of TLS records in queue to tasks for SW crypto"); static counter_u64_t ktls_offload_total; SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total, CTLFLAG_RD, &ktls_offload_total, "Total successful TLS setups (parameters set)"); static counter_u64_t ktls_offload_enable_calls; SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls, CTLFLAG_RD, &ktls_offload_enable_calls, "Total number of TLS enable calls made"); static counter_u64_t ktls_offload_active; SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD, &ktls_offload_active, "Total Active TLS sessions"); static counter_u64_t ktls_offload_failed_crypto; SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD, &ktls_offload_failed_crypto, "Total TLS crypto failures"); static counter_u64_t ktls_switch_to_ifnet; SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD, &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet"); static counter_u64_t ktls_switch_to_sw; SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD, &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW"); static counter_u64_t ktls_switch_failed; SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD, &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet"); SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD, 0, "Software TLS session stats"); SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD, 0, "Hardware (ifnet) TLS session stats"); static counter_u64_t ktls_sw_cbc; SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc, "Active number of software TLS sessions using AES-CBC"); static counter_u64_t ktls_sw_gcm; SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm, "Active number of software TLS sessions using AES-GCM"); static counter_u64_t ktls_ifnet_cbc; SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD, &ktls_ifnet_cbc, "Active number of ifnet TLS sessions using AES-CBC"); static counter_u64_t ktls_ifnet_gcm; SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD, &ktls_ifnet_gcm, "Active number of ifnet TLS sessions using AES-GCM"); static counter_u64_t ktls_ifnet_reset; SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD, &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag"); static counter_u64_t ktls_ifnet_reset_dropped; SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD, &ktls_ifnet_reset_dropped, "TLS sessions dropped after failing to update ifnet send tag"); static counter_u64_t ktls_ifnet_reset_failed; SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD, &ktls_ifnet_reset_failed, "TLS sessions that failed to allocate a new ifnet send tag"); static int ktls_ifnet_permitted; SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN, &ktls_ifnet_permitted, 1, "Whether to permit hardware (ifnet) TLS sessions"); static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS"); static void ktls_cleanup(struct ktls_session *tls); #if defined(INET) || defined(INET6) static void ktls_reset_send_tag(void *context, int pending); #endif static void ktls_work_thread(void *ctx); int ktls_crypto_backend_register(struct ktls_crypto_backend *be) { struct ktls_crypto_backend *curr_be, *tmp; if (be->api_version != KTLS_API_VERSION) { printf("KTLS: API version mismatch (%d vs %d) for %s\n", be->api_version, KTLS_API_VERSION, be->name); return (EINVAL); } rm_wlock(&ktls_backends_lock); printf("KTLS: Registering crypto method %s with prio %d\n", be->name, be->prio); if (LIST_EMPTY(&ktls_backends)) { LIST_INSERT_HEAD(&ktls_backends, be, next); } else { LIST_FOREACH_SAFE(curr_be, &ktls_backends, next, tmp) { if (curr_be->prio < be->prio) { LIST_INSERT_BEFORE(curr_be, be, next); break; } if (LIST_NEXT(curr_be, next) == NULL) { LIST_INSERT_AFTER(curr_be, be, next); break; } } } rm_wunlock(&ktls_backends_lock); return (0); } int ktls_crypto_backend_deregister(struct ktls_crypto_backend *be) { struct ktls_crypto_backend *tmp; /* * Don't error if the backend isn't registered. This permits * MOD_UNLOAD handlers to use this function unconditionally. */ rm_wlock(&ktls_backends_lock); LIST_FOREACH(tmp, &ktls_backends, next) { if (tmp == be) break; } if (tmp == NULL) { rm_wunlock(&ktls_backends_lock); return (0); } if (!ktls_allow_unload) { rm_wunlock(&ktls_backends_lock); printf( "KTLS: Deregistering crypto method %s is not supported\n", be->name); return (EBUSY); } if (be->use_count) { rm_wunlock(&ktls_backends_lock); return (EBUSY); } LIST_REMOVE(be, next); rm_wunlock(&ktls_backends_lock); return (0); } #if defined(INET) || defined(INET6) static uint16_t ktls_get_cpu(struct socket *so) { struct inpcb *inp; uint16_t cpuid; inp = sotoinpcb(so); #ifdef RSS cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); if (cpuid != NETISR_CPUID_NONE) return (cpuid); #endif /* * Just use the flowid to shard connections in a repeatable * fashion. Note that some crypto backends rely on the * serialization provided by having the same connection use * the same queue. */ cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads]; return (cpuid); } #endif static void ktls_init(void *dummy __unused) { struct thread *td; struct pcpu *pc; cpuset_t mask; int error, i; ktls_tasks_active = counter_u64_alloc(M_WAITOK); ktls_cnt_on = counter_u64_alloc(M_WAITOK); ktls_offload_total = counter_u64_alloc(M_WAITOK); ktls_offload_enable_calls = counter_u64_alloc(M_WAITOK); ktls_offload_active = counter_u64_alloc(M_WAITOK); ktls_offload_failed_crypto = counter_u64_alloc(M_WAITOK); ktls_switch_to_ifnet = counter_u64_alloc(M_WAITOK); ktls_switch_to_sw = counter_u64_alloc(M_WAITOK); ktls_switch_failed = counter_u64_alloc(M_WAITOK); ktls_sw_cbc = counter_u64_alloc(M_WAITOK); ktls_sw_gcm = counter_u64_alloc(M_WAITOK); ktls_ifnet_cbc = counter_u64_alloc(M_WAITOK); ktls_ifnet_gcm = counter_u64_alloc(M_WAITOK); ktls_ifnet_reset = counter_u64_alloc(M_WAITOK); ktls_ifnet_reset_dropped = counter_u64_alloc(M_WAITOK); ktls_ifnet_reset_failed = counter_u64_alloc(M_WAITOK); rm_init(&ktls_backends_lock, "ktls backends"); LIST_INIT(&ktls_backends); ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS, M_WAITOK | M_ZERO); ktls_session_zone = uma_zcreate("ktls_session", sizeof(struct ktls_session), #ifdef INVARIANTS trash_ctor, trash_dtor, trash_init, trash_fini, #else NULL, NULL, NULL, NULL, #endif UMA_ALIGN_CACHE, 0); /* * Initialize the workqueues to run the TLS work. We create a * work queue for each CPU. */ CPU_FOREACH(i) { STAILQ_INIT(&ktls_wq[i].head); mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF); error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i], &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i); if (error) panic("Can't add KTLS thread %d error %d", i, error); /* * Bind threads to cores. If ktls_bind_threads is > * 1, then we bind to the NUMA domain. */ if (ktls_bind_threads) { if (ktls_bind_threads > 1) { pc = pcpu_find(i); CPU_COPY(&cpuset_domain[pc->pc_domain], &mask); } else { CPU_SETOF(i, &mask); } error = cpuset_setthread(td->td_tid, &mask); if (error) panic( "Unable to bind KTLS thread for CPU %d error %d", i, error); } ktls_cpuid_lookup[ktls_number_threads] = i; ktls_number_threads++; } printf("KTLS: Initialized %d threads\n", ktls_number_threads); } SYSINIT(ktls, SI_SUB_SMP + 1, SI_ORDER_ANY, ktls_init, NULL); #if defined(INET) || defined(INET6) static int ktls_create_session(struct socket *so, struct tls_enable *en, struct ktls_session **tlsp) { struct ktls_session *tls; int error; /* Only TLS 1.0 - 1.2 are supported. */ if (en->tls_vmajor != TLS_MAJOR_VER_ONE) return (EINVAL); if (en->tls_vminor < TLS_MINOR_VER_ZERO || - en->tls_vminor > TLS_MINOR_VER_TWO) + en->tls_vminor > TLS_MINOR_VER_THREE) return (EINVAL); if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE) return (EINVAL); if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE) return (EINVAL); - if (en->iv_len < 0 || en->iv_len > TLS_MAX_PARAM_SIZE) + if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv)) return (EINVAL); /* All supported algorithms require a cipher key. */ if (en->cipher_key_len == 0) return (EINVAL); /* No flags are currently supported. */ if (en->flags != 0) return (EINVAL); /* Common checks for supported algorithms. */ switch (en->cipher_algorithm) { case CRYPTO_AES_NIST_GCM_16: /* * auth_algorithm isn't used, but permit GMAC values * for compatibility. */ switch (en->auth_algorithm) { case 0: case CRYPTO_AES_128_NIST_GMAC: case CRYPTO_AES_192_NIST_GMAC: case CRYPTO_AES_256_NIST_GMAC: break; default: return (EINVAL); } if (en->auth_key_len != 0) return (EINVAL); - if (en->iv_len != TLS_AEAD_GCM_LEN) + if ((en->tls_vminor == TLS_MINOR_VER_TWO && + en->iv_len != TLS_AEAD_GCM_LEN) || + (en->tls_vminor == TLS_MINOR_VER_THREE && + en->iv_len != TLS_1_3_GCM_IV_LEN)) return (EINVAL); break; case CRYPTO_AES_CBC: switch (en->auth_algorithm) { case CRYPTO_SHA1_HMAC: /* * TLS 1.0 requires an implicit IV. TLS 1.1+ * all use explicit IVs. */ if (en->tls_vminor == TLS_MINOR_VER_ZERO) { if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN) return (EINVAL); break; } /* FALLTHROUGH */ case CRYPTO_SHA2_256_HMAC: case CRYPTO_SHA2_384_HMAC: /* Ignore any supplied IV. */ en->iv_len = 0; break; default: return (EINVAL); } if (en->auth_key_len == 0) return (EINVAL); break; default: return (EINVAL); } tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); counter_u64_add(ktls_offload_active, 1); refcount_init(&tls->refcount, 1); TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls); tls->wq_index = ktls_get_cpu(so); tls->params.cipher_algorithm = en->cipher_algorithm; tls->params.auth_algorithm = en->auth_algorithm; tls->params.tls_vmajor = en->tls_vmajor; tls->params.tls_vminor = en->tls_vminor; tls->params.flags = en->flags; tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen); /* Set the header and trailer lengths. */ tls->params.tls_hlen = sizeof(struct tls_record_layer); switch (en->cipher_algorithm) { case CRYPTO_AES_NIST_GCM_16: - tls->params.tls_hlen += 8; + /* + * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte + * nonce. TLS 1.3 uses a 12 byte implicit IV. + */ + if (en->tls_vminor < TLS_MINOR_VER_THREE) + tls->params.tls_hlen += sizeof(uint64_t); tls->params.tls_tlen = AES_GMAC_HASH_LEN; + + /* + * TLS 1.3 includes optional padding which we + * do not support, and also puts the "real" record + * type at the end of the encrypted data. + */ + if (en->tls_vminor == TLS_MINOR_VER_THREE) + tls->params.tls_tlen += sizeof(uint8_t); + tls->params.tls_bs = 1; break; case CRYPTO_AES_CBC: switch (en->auth_algorithm) { case CRYPTO_SHA1_HMAC: if (en->tls_vminor == TLS_MINOR_VER_ZERO) { /* Implicit IV, no nonce. */ } else { tls->params.tls_hlen += AES_BLOCK_LEN; } tls->params.tls_tlen = AES_BLOCK_LEN + SHA1_HASH_LEN; break; case CRYPTO_SHA2_256_HMAC: tls->params.tls_hlen += AES_BLOCK_LEN; tls->params.tls_tlen = AES_BLOCK_LEN + SHA2_256_HASH_LEN; break; case CRYPTO_SHA2_384_HMAC: tls->params.tls_hlen += AES_BLOCK_LEN; tls->params.tls_tlen = AES_BLOCK_LEN + SHA2_384_HASH_LEN; break; default: panic("invalid hmac"); } tls->params.tls_bs = AES_BLOCK_LEN; break; default: panic("invalid cipher"); } KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN, ("TLS header length too long: %d", tls->params.tls_hlen)); KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN, ("TLS trailer length too long: %d", tls->params.tls_tlen)); if (en->auth_key_len != 0) { tls->params.auth_key_len = en->auth_key_len; tls->params.auth_key = malloc(en->auth_key_len, M_KTLS, M_WAITOK); error = copyin(en->auth_key, tls->params.auth_key, en->auth_key_len); if (error) goto out; } tls->params.cipher_key_len = en->cipher_key_len; tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK); error = copyin(en->cipher_key, tls->params.cipher_key, en->cipher_key_len); if (error) goto out; /* * This holds the implicit portion of the nonce for GCM and * the initial implicit IV for TLS 1.0. The explicit portions * of the IV are generated in ktls_frame() and ktls_seq(). */ if (en->iv_len != 0) { - MPASS(en->iv_len <= sizeof(tls->params.iv)); tls->params.iv_len = en->iv_len; error = copyin(en->iv, tls->params.iv, en->iv_len); if (error) goto out; } *tlsp = tls; return (0); out: ktls_cleanup(tls); return (error); } static struct ktls_session * ktls_clone_session(struct ktls_session *tls) { struct ktls_session *tls_new; tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); counter_u64_add(ktls_offload_active, 1); refcount_init(&tls_new->refcount, 1); /* Copy fields from existing session. */ tls_new->params = tls->params; tls_new->wq_index = tls->wq_index; /* Deep copy keys. */ if (tls_new->params.auth_key != NULL) { tls_new->params.auth_key = malloc(tls->params.auth_key_len, M_KTLS, M_WAITOK); memcpy(tls_new->params.auth_key, tls->params.auth_key, tls->params.auth_key_len); } tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS, M_WAITOK); memcpy(tls_new->params.cipher_key, tls->params.cipher_key, tls->params.cipher_key_len); return (tls_new); } #endif static void ktls_cleanup(struct ktls_session *tls) { counter_u64_add(ktls_offload_active, -1); if (tls->free != NULL) { MPASS(tls->be != NULL); switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: counter_u64_add(ktls_sw_cbc, -1); break; case CRYPTO_AES_NIST_GCM_16: counter_u64_add(ktls_sw_gcm, -1); break; } tls->free(tls); } else if (tls->snd_tag != NULL) { switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: counter_u64_add(ktls_ifnet_cbc, -1); break; case CRYPTO_AES_NIST_GCM_16: counter_u64_add(ktls_ifnet_gcm, -1); break; } m_snd_tag_rele(tls->snd_tag); } if (tls->params.auth_key != NULL) { explicit_bzero(tls->params.auth_key, tls->params.auth_key_len); free(tls->params.auth_key, M_KTLS); tls->params.auth_key = NULL; tls->params.auth_key_len = 0; } if (tls->params.cipher_key != NULL) { explicit_bzero(tls->params.cipher_key, tls->params.cipher_key_len); free(tls->params.cipher_key, M_KTLS); tls->params.cipher_key = NULL; tls->params.cipher_key_len = 0; } explicit_bzero(tls->params.iv, sizeof(tls->params.iv)); } #if defined(INET) || defined(INET6) /* * Common code used when first enabling ifnet TLS on a connection or * when allocating a new ifnet TLS session due to a routing change. * This function allocates a new TLS send tag on whatever interface * the connection is currently routed over. */ static int ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force, struct m_snd_tag **mstp) { union if_snd_tag_alloc_params params; struct ifnet *ifp; struct rtentry *rt; struct tcpcb *tp; int error; INP_RLOCK(inp); if (inp->inp_flags2 & INP_FREED) { INP_RUNLOCK(inp); return (ECONNRESET); } if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { INP_RUNLOCK(inp); return (ECONNRESET); } if (inp->inp_socket == NULL) { INP_RUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); /* * Check administrative controls on ifnet TLS to determine if * ifnet TLS should be denied. * * - Always permit 'force' requests. * - ktls_ifnet_permitted == 0: always deny. */ if (!force && ktls_ifnet_permitted == 0) { INP_RUNLOCK(inp); return (ENXIO); } /* * XXX: Use the cached route in the inpcb to find the * interface. This should perhaps instead use * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only * enabled after a connection has completed key negotiation in * userland, the cached route will be present in practice. */ rt = inp->inp_route.ro_rt; if (rt == NULL || rt->rt_ifp == NULL) { INP_RUNLOCK(inp); return (ENXIO); } ifp = rt->rt_ifp; if_ref(ifp); params.hdr.type = IF_SND_TAG_TYPE_TLS; params.hdr.flowid = inp->inp_flowid; params.hdr.flowtype = inp->inp_flowtype; params.tls.inp = inp; params.tls.tls = tls; INP_RUNLOCK(inp); if (ifp->if_snd_tag_alloc == NULL) { error = EOPNOTSUPP; goto out; } if ((ifp->if_capenable & IFCAP_NOMAP) == 0) { error = EOPNOTSUPP; goto out; } if (inp->inp_vflag & INP_IPV6) { if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) { error = EOPNOTSUPP; goto out; } } else { if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) { error = EOPNOTSUPP; goto out; } } error = ifp->if_snd_tag_alloc(ifp, ¶ms, mstp); out: if_rele(ifp); return (error); } static int ktls_try_ifnet(struct socket *so, struct ktls_session *tls, bool force) { struct m_snd_tag *mst; int error; error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst); if (error == 0) { tls->snd_tag = mst; switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: counter_u64_add(ktls_ifnet_cbc, 1); break; case CRYPTO_AES_NIST_GCM_16: counter_u64_add(ktls_ifnet_gcm, 1); break; } } return (error); } static int ktls_try_sw(struct socket *so, struct ktls_session *tls) { struct rm_priotracker prio; struct ktls_crypto_backend *be; /* * Choose the best software crypto backend. Backends are * stored in sorted priority order (larget value == most * important at the head of the list), so this just stops on * the first backend that claims the session by returning * success. */ if (ktls_allow_unload) rm_rlock(&ktls_backends_lock, &prio); LIST_FOREACH(be, &ktls_backends, next) { if (be->try(so, tls) == 0) break; KASSERT(tls->cipher == NULL, ("ktls backend leaked a cipher pointer")); } if (be != NULL) { if (ktls_allow_unload) be->use_count++; tls->be = be; } if (ktls_allow_unload) rm_runlock(&ktls_backends_lock, &prio); if (be == NULL) return (EOPNOTSUPP); switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: counter_u64_add(ktls_sw_cbc, 1); break; case CRYPTO_AES_NIST_GCM_16: counter_u64_add(ktls_sw_gcm, 1); break; } return (0); } int ktls_enable_tx(struct socket *so, struct tls_enable *en) { struct ktls_session *tls; int error; if (!ktls_offload_enable) return (ENOTSUP); counter_u64_add(ktls_offload_enable_calls, 1); /* * This should always be true since only the TCP socket option * invokes this function. */ if (so->so_proto->pr_protocol != IPPROTO_TCP) return (EINVAL); /* * XXX: Don't overwrite existing sessions. We should permit * this to support rekeying in the future. */ if (so->so_snd.sb_tls_info != NULL) return (EALREADY); if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) return (ENOTSUP); /* TLS requires ext pgs */ if (mb_use_ext_pgs == 0) return (ENXIO); error = ktls_create_session(so, en, &tls); if (error) return (error); /* Prefer ifnet TLS over software TLS. */ error = ktls_try_ifnet(so, tls, false); if (error) error = ktls_try_sw(so, tls); if (error) { ktls_cleanup(tls); return (error); } error = sblock(&so->so_snd, SBL_WAIT); if (error) { ktls_cleanup(tls); return (error); } SOCKBUF_LOCK(&so->so_snd); so->so_snd.sb_tls_info = tls; if (tls->sw_encrypt == NULL) so->so_snd.sb_flags |= SB_TLS_IFNET; SOCKBUF_UNLOCK(&so->so_snd); sbunlock(&so->so_snd); counter_u64_add(ktls_offload_total, 1); return (0); } int ktls_get_tx_mode(struct socket *so) { struct ktls_session *tls; struct inpcb *inp; int mode; inp = so->so_pcb; INP_WLOCK_ASSERT(inp); SOCKBUF_LOCK(&so->so_snd); tls = so->so_snd.sb_tls_info; if (tls == NULL) mode = TCP_TLS_MODE_NONE; else if (tls->sw_encrypt != NULL) mode = TCP_TLS_MODE_SW; else mode = TCP_TLS_MODE_IFNET; SOCKBUF_UNLOCK(&so->so_snd); return (mode); } /* * Switch between SW and ifnet TLS sessions as requested. */ int ktls_set_tx_mode(struct socket *so, int mode) { struct ktls_session *tls, *tls_new; struct inpcb *inp; int error; MPASS(mode == TCP_TLS_MODE_SW || mode == TCP_TLS_MODE_IFNET); inp = so->so_pcb; INP_WLOCK_ASSERT(inp); SOCKBUF_LOCK(&so->so_snd); tls = so->so_snd.sb_tls_info; if (tls == NULL) { SOCKBUF_UNLOCK(&so->so_snd); return (0); } if ((tls->sw_encrypt != NULL && mode == TCP_TLS_MODE_SW) || (tls->sw_encrypt == NULL && mode == TCP_TLS_MODE_IFNET)) { SOCKBUF_UNLOCK(&so->so_snd); return (0); } tls = ktls_hold(tls); SOCKBUF_UNLOCK(&so->so_snd); INP_WUNLOCK(inp); tls_new = ktls_clone_session(tls); if (mode == TCP_TLS_MODE_IFNET) error = ktls_try_ifnet(so, tls_new, true); else error = ktls_try_sw(so, tls_new); if (error) { counter_u64_add(ktls_switch_failed, 1); ktls_free(tls_new); ktls_free(tls); INP_WLOCK(inp); return (error); } error = sblock(&so->so_snd, SBL_WAIT); if (error) { counter_u64_add(ktls_switch_failed, 1); ktls_free(tls_new); ktls_free(tls); INP_WLOCK(inp); return (error); } /* * If we raced with another session change, keep the existing * session. */ if (tls != so->so_snd.sb_tls_info) { counter_u64_add(ktls_switch_failed, 1); sbunlock(&so->so_snd); ktls_free(tls_new); ktls_free(tls); INP_WLOCK(inp); return (EBUSY); } SOCKBUF_LOCK(&so->so_snd); so->so_snd.sb_tls_info = tls_new; if (tls_new->sw_encrypt == NULL) so->so_snd.sb_flags |= SB_TLS_IFNET; SOCKBUF_UNLOCK(&so->so_snd); sbunlock(&so->so_snd); /* * Drop two references on 'tls'. The first is for the * ktls_hold() above. The second drops the reference from the * socket buffer. */ KASSERT(tls->refcount >= 2, ("too few references on old session")); ktls_free(tls); ktls_free(tls); if (mode == TCP_TLS_MODE_IFNET) counter_u64_add(ktls_switch_to_ifnet, 1); else counter_u64_add(ktls_switch_to_sw, 1); INP_WLOCK(inp); return (0); } /* * Try to allocate a new TLS send tag. This task is scheduled when * ip_output detects a route change while trying to transmit a packet * holding a TLS record. If a new tag is allocated, replace the tag * in the TLS session. Subsequent packets on the connection will use * the new tag. If a new tag cannot be allocated, drop the * connection. */ static void ktls_reset_send_tag(void *context, int pending) { struct epoch_tracker et; struct ktls_session *tls; struct m_snd_tag *old, *new; struct inpcb *inp; struct tcpcb *tp; int error; MPASS(pending == 1); tls = context; inp = tls->inp; /* * Free the old tag first before allocating a new one. * ip[6]_output_send() will treat a NULL send tag the same as * an ifp mismatch and drop packets until a new tag is * allocated. * * Write-lock the INP when changing tls->snd_tag since * ip[6]_output_send() holds a read-lock when reading the * pointer. */ INP_WLOCK(inp); old = tls->snd_tag; tls->snd_tag = NULL; INP_WUNLOCK(inp); if (old != NULL) m_snd_tag_rele(old); error = ktls_alloc_snd_tag(inp, tls, true, &new); if (error == 0) { INP_WLOCK(inp); tls->snd_tag = new; mtx_pool_lock(mtxpool_sleep, tls); tls->reset_pending = false; mtx_pool_unlock(mtxpool_sleep, tls); if (!in_pcbrele_wlocked(inp)) INP_WUNLOCK(inp); counter_u64_add(ktls_ifnet_reset, 1); /* * XXX: Should we kick tcp_output explicitly now that * the send tag is fixed or just rely on timers? */ } else { INP_INFO_RLOCK_ET(&V_tcbinfo, et); INP_WLOCK(inp); if (!in_pcbrele_wlocked(inp)) { if (!(inp->inp_flags & INP_TIMEWAIT) && !(inp->inp_flags & INP_DROPPED)) { tp = intotcpcb(inp); tp = tcp_drop(tp, ECONNABORTED); if (tp != NULL) INP_WUNLOCK(inp); counter_u64_add(ktls_ifnet_reset_dropped, 1); } else INP_WUNLOCK(inp); } INP_INFO_RUNLOCK_ET(&V_tcbinfo, et); counter_u64_add(ktls_ifnet_reset_failed, 1); /* * Leave reset_pending true to avoid future tasks while * the socket goes away. */ } ktls_free(tls); } int ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls) { if (inp == NULL) return (ENOBUFS); INP_LOCK_ASSERT(inp); /* * See if we should schedule a task to update the send tag for * this session. */ mtx_pool_lock(mtxpool_sleep, tls); if (!tls->reset_pending) { (void) ktls_hold(tls); in_pcbref(inp); tls->inp = inp; tls->reset_pending = true; taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task); } mtx_pool_unlock(mtxpool_sleep, tls); return (ENOBUFS); } #endif void ktls_destroy(struct ktls_session *tls) { struct rm_priotracker prio; ktls_cleanup(tls); if (tls->be != NULL && ktls_allow_unload) { rm_rlock(&ktls_backends_lock, &prio); tls->be->use_count--; rm_runlock(&ktls_backends_lock, &prio); } uma_zfree(ktls_session_zone, tls); } void ktls_seq(struct sockbuf *sb, struct mbuf *m) { struct mbuf_ext_pgs *pgs; struct tls_record_layer *tlshdr; uint64_t seqno; for (; m != NULL; m = m->m_next) { KASSERT((m->m_flags & M_NOMAP) != 0, ("ktls_seq: mapped mbuf %p", m)); pgs = m->m_ext.ext_pgs; pgs->seqno = sb->sb_tls_seqno; /* * Store the sequence number in the TLS header as the * explicit part of the IV for GCM. */ if (pgs->tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16) { tlshdr = (void *)pgs->hdr; seqno = htobe64(pgs->seqno); memcpy(tlshdr + 1, &seqno, sizeof(seqno)); } sb->sb_tls_seqno++; } } /* * Add TLS framing (headers and trailers) to a chain of mbufs. Each * mbuf in the chain must be an unmapped mbuf. The payload of the * mbuf must be populated with the payload of each TLS record. * * The record_type argument specifies the TLS record type used when * populating the TLS header. * * The enq_count argument on return is set to the number of pages of * payload data for this entire chain that need to be encrypted via SW * encryption. The returned value should be passed to ktls_enqueue * when scheduling encryption of this chain of mbufs. */ int ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt, uint8_t record_type) { struct tls_record_layer *tlshdr; struct mbuf *m; struct mbuf_ext_pgs *pgs; uint16_t tls_len; int maxlen; maxlen = tls->params.max_frame_len; *enq_cnt = 0; for (m = top; m != NULL; m = m->m_next) { /* * All mbufs in the chain should be non-empty TLS * records whose payload does not exceed the maximum * frame length. */ if (m->m_len > maxlen || m->m_len == 0) return (EINVAL); tls_len = m->m_len; /* * TLS frames require unmapped mbufs to store session * info. */ KASSERT((m->m_flags & M_NOMAP) != 0, ("ktls_frame: mapped mbuf %p (top = %p)\n", m, top)); pgs = m->m_ext.ext_pgs; /* Save a reference to the session. */ pgs->tls = ktls_hold(tls); pgs->hdr_len = tls->params.tls_hlen; pgs->trail_len = tls->params.tls_tlen; if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) { int bs, delta; /* * AES-CBC pads messages to a multiple of the * block size. Note that the padding is * applied after the digest and the encryption * is done on the "plaintext || mac || padding". * At least one byte of padding is always * present. * * Compute the final trailer length assuming * at most one block of padding. * tls->params.sb_tls_tlen is the maximum * possible trailer length (padding + digest). * delta holds the number of excess padding * bytes if the maximum were used. Those * extra bytes are removed. */ bs = tls->params.tls_bs; delta = (tls_len + tls->params.tls_tlen) & (bs - 1); pgs->trail_len -= delta; } m->m_len += pgs->hdr_len + pgs->trail_len; /* Populate the TLS header. */ tlshdr = (void *)pgs->hdr; tlshdr->tls_vmajor = tls->params.tls_vmajor; - tlshdr->tls_vminor = tls->params.tls_vminor; - tlshdr->tls_type = record_type; + + /* + * TLS 1.3 masquarades as TLS 1.2 with a record type + * of TLS_RLTYPE_APP. + */ + if (tls->params.tls_vminor == TLS_MINOR_VER_THREE && + tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) { + tlshdr->tls_vminor = TLS_MINOR_VER_TWO; + tlshdr->tls_type = TLS_RLTYPE_APP; + /* save the real record type for later */ + pgs->record_type = record_type; + } else { + tlshdr->tls_vminor = tls->params.tls_vminor; + tlshdr->tls_type = record_type; + } tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr)); /* * For GCM, the sequence number is stored in the * header by ktls_seq(). For CBC, a random nonce is * inserted for TLS 1.1+. */ if (tls->params.cipher_algorithm == CRYPTO_AES_CBC && tls->params.tls_vminor >= TLS_MINOR_VER_ONE) arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0); /* * When using SW encryption, mark the mbuf not ready. * It will be marked ready via sbready() after the * record has been encrypted. * * When using ifnet TLS, unencrypted TLS records are * sent down the stack to the NIC. */ if (tls->sw_encrypt != NULL) { m->m_flags |= M_NOTREADY; pgs->nrdy = pgs->npgs; *enq_cnt += pgs->npgs; } } return (0); } void ktls_enqueue_to_free(struct mbuf_ext_pgs *pgs) { struct ktls_wq *wq; bool running; /* Mark it for freeing. */ pgs->mbuf = NULL; wq = &ktls_wq[pgs->tls->wq_index]; mtx_lock(&wq->mtx); STAILQ_INSERT_TAIL(&wq->head, pgs, stailq); running = wq->running; mtx_unlock(&wq->mtx); if (!running) wakeup(wq); } void ktls_enqueue(struct mbuf *m, struct socket *so, int page_count) { struct mbuf_ext_pgs *pgs; struct ktls_wq *wq; bool running; KASSERT(((m->m_flags & (M_NOMAP | M_NOTREADY)) == (M_NOMAP | M_NOTREADY)), ("ktls_enqueue: %p not unready & nomap mbuf\n", m)); KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count")); pgs = m->m_ext.ext_pgs; KASSERT(pgs->tls->sw_encrypt != NULL, ("ifnet TLS mbuf")); pgs->enc_cnt = page_count; pgs->mbuf = m; /* * Save a pointer to the socket. The caller is responsible * for taking an additional reference via soref(). */ pgs->so = so; wq = &ktls_wq[pgs->tls->wq_index]; mtx_lock(&wq->mtx); STAILQ_INSERT_TAIL(&wq->head, pgs, stailq); running = wq->running; mtx_unlock(&wq->mtx); if (!running) wakeup(wq); counter_u64_add(ktls_cnt_on, 1); } static __noinline void ktls_encrypt(struct mbuf_ext_pgs *pgs) { struct ktls_session *tls; struct socket *so; struct mbuf *m, *top; vm_paddr_t parray[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)]; struct iovec src_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)]; struct iovec dst_iov[1 + btoc(TLS_MAX_MSG_SIZE_V10_2)]; vm_page_t pg; int error, i, len, npages, off, total_pages; bool is_anon; so = pgs->so; tls = pgs->tls; top = pgs->mbuf; KASSERT(tls != NULL, ("tls = NULL, top = %p, pgs = %p\n", top, pgs)); KASSERT(so != NULL, ("so = NULL, top = %p, pgs = %p\n", top, pgs)); #ifdef INVARIANTS pgs->so = NULL; pgs->mbuf = NULL; #endif total_pages = pgs->enc_cnt; npages = 0; /* * Encrypt the TLS records in the chain of mbufs starting with * 'top'. 'total_pages' gives us a total count of pages and is * used to know when we have finished encrypting the TLS * records originally queued with 'top'. * * NB: These mbufs are queued in the socket buffer and * 'm_next' is traversing the mbufs in the socket buffer. The * socket buffer lock is not held while traversing this chain. * Since the mbufs are all marked M_NOTREADY their 'm_next' * pointers should be stable. However, the 'm_next' of the * last mbuf encrypted is not necessarily NULL. It can point * to other mbufs appended while 'top' was on the TLS work * queue. * * Each mbuf holds an entire TLS record. */ error = 0; for (m = top; npages != total_pages; m = m->m_next) { pgs = m->m_ext.ext_pgs; KASSERT(pgs->tls == tls, ("different TLS sessions in a single mbuf chain: %p vs %p", tls, pgs->tls)); KASSERT((m->m_flags & (M_NOMAP | M_NOTREADY)) == (M_NOMAP | M_NOTREADY), ("%p not unready & nomap mbuf (top = %p)\n", m, top)); KASSERT(npages + pgs->npgs <= total_pages, ("page count mismatch: top %p, total_pages %d, m %p", top, total_pages, m)); /* * Generate source and destination ivoecs to pass to * the SW encryption backend. For writable mbufs, the * destination iovec is a copy of the source and * encryption is done in place. For file-backed mbufs * (from sendfile), anonymous wired pages are * allocated and assigned to the destination iovec. */ is_anon = M_WRITABLE(m); off = pgs->first_pg_off; for (i = 0; i < pgs->npgs; i++, off = 0) { len = mbuf_ext_pg_len(pgs, i, off); src_iov[i].iov_len = len; src_iov[i].iov_base = (char *)(void *)PHYS_TO_DMAP(pgs->pa[i]) + off; if (is_anon) { dst_iov[i].iov_base = src_iov[i].iov_base; dst_iov[i].iov_len = src_iov[i].iov_len; continue; } retry_page: pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_NODUMP | VM_ALLOC_WIRED); if (pg == NULL) { vm_wait(NULL); goto retry_page; } parray[i] = VM_PAGE_TO_PHYS(pg); dst_iov[i].iov_base = (char *)(void *)PHYS_TO_DMAP(parray[i]) + off; dst_iov[i].iov_len = len; } npages += i; error = (*tls->sw_encrypt)(tls, (const struct tls_record_layer *)pgs->hdr, - pgs->trail, src_iov, dst_iov, i, pgs->seqno); + pgs->trail, src_iov, dst_iov, i, pgs->seqno, + pgs->record_type); if (error) { counter_u64_add(ktls_offload_failed_crypto, 1); break; } /* * For file-backed mbufs, release the file-backed * pages and replace them in the ext_pgs array with * the anonymous wired pages allocated above. */ if (!is_anon) { /* Free the old pages. */ m->m_ext.ext_free(m); /* Replace them with the new pages. */ for (i = 0; i < pgs->npgs; i++) pgs->pa[i] = parray[i]; /* Use the basic free routine. */ m->m_ext.ext_free = mb_free_mext_pgs; } /* * Drop a reference to the session now that it is no * longer needed. Existing code depends on encrypted * records having no associated session vs * yet-to-be-encrypted records having an associated * session. */ pgs->tls = NULL; ktls_free(tls); } CURVNET_SET(so->so_vnet); if (error == 0) { (void)(*so->so_proto->pr_usrreqs->pru_ready)(so, top, npages); } else { so->so_proto->pr_usrreqs->pru_abort(so); so->so_error = EIO; mb_free_notready(top, total_pages); } SOCK_LOCK(so); sorele(so); CURVNET_RESTORE(); } static void ktls_work_thread(void *ctx) { struct ktls_wq *wq = ctx; struct mbuf_ext_pgs *p, *n; struct ktls_session *tls; STAILQ_HEAD(, mbuf_ext_pgs) local_head; #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) fpu_kern_thread(0); #endif for (;;) { mtx_lock(&wq->mtx); while (STAILQ_EMPTY(&wq->head)) { wq->running = false; mtx_sleep(wq, &wq->mtx, 0, "-", 0); wq->running = true; } STAILQ_INIT(&local_head); STAILQ_CONCAT(&local_head, &wq->head); mtx_unlock(&wq->mtx); STAILQ_FOREACH_SAFE(p, &local_head, stailq, n) { if (p->mbuf != NULL) { ktls_encrypt(p); counter_u64_add(ktls_cnt_on, -1); } else { tls = p->tls; ktls_free(tls); uma_zfree(zone_extpgs, p); } } } } Index: head/sys/net/iflib.c =================================================================== --- head/sys/net/iflib.c (revision 352813) +++ head/sys/net/iflib.c (revision 352814) @@ -1,6800 +1,6802 @@ /*- * Copyright (c) 2014-2018, Matthew Macy * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, * this list of conditions and the following disclaimer. * * 2. Neither the name of Matthew Macy nor the names of its * contributors may be used to endorse or promote products derived from * this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE * POSSIBILITY OF SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_acpi.h" #include "opt_sched.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "ifdi_if.h" #ifdef PCI_IOV #include #endif #include /* * enable accounting of every mbuf as it comes in to and goes out of * iflib's software descriptor references */ #define MEMORY_LOGGING 0 /* * Enable mbuf vectors for compressing long mbuf chains */ /* * NB: * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead * we prefetch needs to be determined by the time spent in m_free vis a vis * the cost of a prefetch. This will of course vary based on the workload: * - NFLX's m_free path is dominated by vm-based M_EXT manipulation which * is quite expensive, thus suggesting very little prefetch. * - small packet forwarding which is just returning a single mbuf to * UMA will typically be very fast vis a vis the cost of a memory * access. */ /* * File organization: * - private structures * - iflib private utility functions * - ifnet functions * - vlan registry and other exported functions * - iflib public core functions * * */ MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library"); struct iflib_txq; typedef struct iflib_txq *iflib_txq_t; struct iflib_rxq; typedef struct iflib_rxq *iflib_rxq_t; struct iflib_fl; typedef struct iflib_fl *iflib_fl_t; struct iflib_ctx; static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid); static void iflib_timer(void *arg); typedef struct iflib_filter_info { driver_filter_t *ifi_filter; void *ifi_filter_arg; struct grouptask *ifi_task; void *ifi_ctx; } *iflib_filter_info_t; struct iflib_ctx { KOBJ_FIELDS; /* * Pointer to hardware driver's softc */ void *ifc_softc; device_t ifc_dev; if_t ifc_ifp; cpuset_t ifc_cpus; if_shared_ctx_t ifc_sctx; struct if_softc_ctx ifc_softc_ctx; struct sx ifc_ctx_sx; struct mtx ifc_state_mtx; iflib_txq_t ifc_txqs; iflib_rxq_t ifc_rxqs; uint32_t ifc_if_flags; uint32_t ifc_flags; uint32_t ifc_max_fl_buf_size; uint32_t ifc_rx_mbuf_sz; int ifc_link_state; int ifc_watchdog_events; struct cdev *ifc_led_dev; struct resource *ifc_msix_mem; struct if_irq ifc_legacy_irq; struct grouptask ifc_admin_task; struct grouptask ifc_vflr_task; struct iflib_filter_info ifc_filter_info; struct ifmedia ifc_media; struct ifmedia *ifc_mediap; struct sysctl_oid *ifc_sysctl_node; uint16_t ifc_sysctl_ntxqs; uint16_t ifc_sysctl_nrxqs; uint16_t ifc_sysctl_qs_eq_override; uint16_t ifc_sysctl_rx_budget; uint16_t ifc_sysctl_tx_abdicate; uint16_t ifc_sysctl_core_offset; #define CORE_OFFSET_UNSPECIFIED 0xffff uint8_t ifc_sysctl_separate_txrx; qidx_t ifc_sysctl_ntxds[8]; qidx_t ifc_sysctl_nrxds[8]; struct if_txrx ifc_txrx; #define isc_txd_encap ifc_txrx.ift_txd_encap #define isc_txd_flush ifc_txrx.ift_txd_flush #define isc_txd_credits_update ifc_txrx.ift_txd_credits_update #define isc_rxd_available ifc_txrx.ift_rxd_available #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get #define isc_rxd_refill ifc_txrx.ift_rxd_refill #define isc_rxd_flush ifc_txrx.ift_rxd_flush #define isc_rxd_refill ifc_txrx.ift_rxd_refill #define isc_rxd_refill ifc_txrx.ift_rxd_refill #define isc_legacy_intr ifc_txrx.ift_legacy_intr eventhandler_tag ifc_vlan_attach_event; eventhandler_tag ifc_vlan_detach_event; struct ether_addr ifc_mac; }; void * iflib_get_softc(if_ctx_t ctx) { return (ctx->ifc_softc); } device_t iflib_get_dev(if_ctx_t ctx) { return (ctx->ifc_dev); } if_t iflib_get_ifp(if_ctx_t ctx) { return (ctx->ifc_ifp); } struct ifmedia * iflib_get_media(if_ctx_t ctx) { return (ctx->ifc_mediap); } uint32_t iflib_get_flags(if_ctx_t ctx) { return (ctx->ifc_flags); } void iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN]) { bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN); } if_softc_ctx_t iflib_get_softc_ctx(if_ctx_t ctx) { return (&ctx->ifc_softc_ctx); } if_shared_ctx_t iflib_get_sctx(if_ctx_t ctx) { return (ctx->ifc_sctx); } #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2) #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*)) #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1))) #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP) #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF) typedef struct iflib_sw_rx_desc_array { bus_dmamap_t *ifsd_map; /* bus_dma maps for packet */ struct mbuf **ifsd_m; /* pkthdr mbufs */ caddr_t *ifsd_cl; /* direct cluster pointer for rx */ bus_addr_t *ifsd_ba; /* bus addr of cluster for rx */ } iflib_rxsd_array_t; typedef struct iflib_sw_tx_desc_array { bus_dmamap_t *ifsd_map; /* bus_dma maps for packet */ bus_dmamap_t *ifsd_tso_map; /* bus_dma maps for TSO packet */ struct mbuf **ifsd_m; /* pkthdr mbufs */ } if_txsd_vec_t; /* magic number that should be high enough for any hardware */ #define IFLIB_MAX_TX_SEGS 128 #define IFLIB_RX_COPY_THRESH 128 #define IFLIB_MAX_RX_REFRESH 32 /* The minimum descriptors per second before we start coalescing */ #define IFLIB_MIN_DESC_SEC 16384 #define IFLIB_DEFAULT_TX_UPDATE_FREQ 16 #define IFLIB_QUEUE_IDLE 0 #define IFLIB_QUEUE_HUNG 1 #define IFLIB_QUEUE_WORKING 2 /* maximum number of txqs that can share an rx interrupt */ #define IFLIB_MAX_TX_SHARED_INTR 4 /* this should really scale with ring size - this is a fairly arbitrary value */ #define TX_BATCH_SIZE 32 #define IFLIB_RESTART_BUDGET 8 #define CSUM_OFFLOAD (CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \ CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \ CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP) struct iflib_txq { qidx_t ift_in_use; qidx_t ift_cidx; qidx_t ift_cidx_processed; qidx_t ift_pidx; uint8_t ift_gen; uint8_t ift_br_offset; uint16_t ift_npending; uint16_t ift_db_pending; uint16_t ift_rs_pending; /* implicit pad */ uint8_t ift_txd_size[8]; uint64_t ift_processed; uint64_t ift_cleaned; uint64_t ift_cleaned_prev; #if MEMORY_LOGGING uint64_t ift_enqueued; uint64_t ift_dequeued; #endif uint64_t ift_no_tx_dma_setup; uint64_t ift_no_desc_avail; uint64_t ift_mbuf_defrag_failed; uint64_t ift_mbuf_defrag; uint64_t ift_map_failed; uint64_t ift_txd_encap_efbig; uint64_t ift_pullups; uint64_t ift_last_timer_tick; struct mtx ift_mtx; struct mtx ift_db_mtx; /* constant values */ if_ctx_t ift_ctx; struct ifmp_ring *ift_br; struct grouptask ift_task; qidx_t ift_size; uint16_t ift_id; struct callout ift_timer; if_txsd_vec_t ift_sds; uint8_t ift_qstatus; uint8_t ift_closed; uint8_t ift_update_freq; struct iflib_filter_info ift_filter_info; bus_dma_tag_t ift_buf_tag; bus_dma_tag_t ift_tso_buf_tag; iflib_dma_info_t ift_ifdi; #define MTX_NAME_LEN 16 char ift_mtx_name[MTX_NAME_LEN]; bus_dma_segment_t ift_segs[IFLIB_MAX_TX_SEGS] __aligned(CACHE_LINE_SIZE); #ifdef IFLIB_DIAGNOSTICS uint64_t ift_cpu_exec_count[256]; #endif } __aligned(CACHE_LINE_SIZE); struct iflib_fl { qidx_t ifl_cidx; qidx_t ifl_pidx; qidx_t ifl_credits; uint8_t ifl_gen; uint8_t ifl_rxd_size; #if MEMORY_LOGGING uint64_t ifl_m_enqueued; uint64_t ifl_m_dequeued; uint64_t ifl_cl_enqueued; uint64_t ifl_cl_dequeued; #endif /* implicit pad */ bitstr_t *ifl_rx_bitmap; qidx_t ifl_fragidx; /* constant */ qidx_t ifl_size; uint16_t ifl_buf_size; uint16_t ifl_cltype; uma_zone_t ifl_zone; iflib_rxsd_array_t ifl_sds; iflib_rxq_t ifl_rxq; uint8_t ifl_id; bus_dma_tag_t ifl_buf_tag; iflib_dma_info_t ifl_ifdi; uint64_t ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE); caddr_t ifl_vm_addrs[IFLIB_MAX_RX_REFRESH]; qidx_t ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH]; } __aligned(CACHE_LINE_SIZE); static inline qidx_t get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen) { qidx_t used; if (pidx > cidx) used = pidx - cidx; else if (pidx < cidx) used = size - cidx + pidx; else if (gen == 0 && pidx == cidx) used = 0; else if (gen == 1 && pidx == cidx) used = size; else panic("bad state"); return (used); } #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen)) #define IDXDIFF(head, tail, wrap) \ ((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head)) struct iflib_rxq { if_ctx_t ifr_ctx; iflib_fl_t ifr_fl; uint64_t ifr_rx_irq; struct pfil_head *pfil; /* * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is * the command queue consumer index. Otherwise it's unused. */ qidx_t ifr_cq_cidx; uint16_t ifr_id; uint8_t ifr_nfl; uint8_t ifr_ntxqirq; uint8_t ifr_txqid[IFLIB_MAX_TX_SHARED_INTR]; uint8_t ifr_fl_offset; struct lro_ctrl ifr_lc; struct grouptask ifr_task; struct iflib_filter_info ifr_filter_info; iflib_dma_info_t ifr_ifdi; /* dynamically allocate if any drivers need a value substantially larger than this */ struct if_rxd_frag ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE); #ifdef IFLIB_DIAGNOSTICS uint64_t ifr_cpu_exec_count[256]; #endif } __aligned(CACHE_LINE_SIZE); typedef struct if_rxsd { caddr_t *ifsd_cl; iflib_fl_t ifsd_fl; qidx_t ifsd_cidx; } *if_rxsd_t; /* multiple of word size */ #ifdef __LP64__ #define PKT_INFO_SIZE 6 #define RXD_INFO_SIZE 5 #define PKT_TYPE uint64_t #else #define PKT_INFO_SIZE 11 #define RXD_INFO_SIZE 8 #define PKT_TYPE uint32_t #endif #define PKT_LOOP_BOUND ((PKT_INFO_SIZE/3)*3) #define RXD_LOOP_BOUND ((RXD_INFO_SIZE/4)*4) typedef struct if_pkt_info_pad { PKT_TYPE pkt_val[PKT_INFO_SIZE]; } *if_pkt_info_pad_t; typedef struct if_rxd_info_pad { PKT_TYPE rxd_val[RXD_INFO_SIZE]; } *if_rxd_info_pad_t; CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info)); CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info)); static inline void pkt_info_zero(if_pkt_info_t pi) { if_pkt_info_pad_t pi_pad; pi_pad = (if_pkt_info_pad_t)pi; pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0; pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0; #ifndef __LP64__ pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0; pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0; #endif } static device_method_t iflib_pseudo_methods[] = { DEVMETHOD(device_attach, noop_attach), DEVMETHOD(device_detach, iflib_pseudo_detach), DEVMETHOD_END }; driver_t iflib_pseudodriver = { "iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx), }; static inline void rxd_info_zero(if_rxd_info_t ri) { if_rxd_info_pad_t ri_pad; int i; ri_pad = (if_rxd_info_pad_t)ri; for (i = 0; i < RXD_LOOP_BOUND; i += 4) { ri_pad->rxd_val[i] = 0; ri_pad->rxd_val[i+1] = 0; ri_pad->rxd_val[i+2] = 0; ri_pad->rxd_val[i+3] = 0; } #ifdef __LP64__ ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0; #endif } /* * Only allow a single packet to take up most 1/nth of the tx ring */ #define MAX_SINGLE_PACKET_FRACTION 12 #define IF_BAD_DMA (bus_addr_t)-1 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING)) #define CTX_LOCK_INIT(_sc) sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock") #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx) #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx) #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx) #define STATE_LOCK_INIT(_sc, _name) mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF) #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx) #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx) #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx) #define CALLOUT_LOCK(txq) mtx_lock(&txq->ift_mtx) #define CALLOUT_UNLOCK(txq) mtx_unlock(&txq->ift_mtx) void iflib_set_detach(if_ctx_t ctx) { STATE_LOCK(ctx); ctx->ifc_flags |= IFC_IN_DETACH; STATE_UNLOCK(ctx); } /* Our boot-time initialization hook */ static int iflib_module_event_handler(module_t, int, void *); static moduledata_t iflib_moduledata = { "iflib", iflib_module_event_handler, NULL }; DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY); MODULE_VERSION(iflib, 1); MODULE_DEPEND(iflib, pci, 1, 1, 1); MODULE_DEPEND(iflib, ether, 1, 1, 1); TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1); TASKQGROUP_DEFINE(if_config_tqg, 1, 1); #ifndef IFLIB_DEBUG_COUNTERS #ifdef INVARIANTS #define IFLIB_DEBUG_COUNTERS 1 #else #define IFLIB_DEBUG_COUNTERS 0 #endif /* !INVARIANTS */ #endif static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0, "iflib driver parameters"); /* * XXX need to ensure that this can't accidentally cause the head to be moved backwards */ static int iflib_min_tx_latency = 0; SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW, &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput"); static int iflib_no_tx_batch = 0; SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW, &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput"); #if IFLIB_DEBUG_COUNTERS static int iflib_tx_seen; static int iflib_tx_sent; static int iflib_tx_encap; static int iflib_rx_allocs; static int iflib_fl_refills; static int iflib_fl_refills_large; static int iflib_tx_frees; SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD, &iflib_tx_seen, 0, "# TX mbufs seen"); SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD, &iflib_tx_sent, 0, "# TX mbufs sent"); SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD, &iflib_tx_encap, 0, "# TX mbufs encapped"); SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD, &iflib_tx_frees, 0, "# TX frees"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD, &iflib_rx_allocs, 0, "# RX allocations"); SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD, &iflib_fl_refills, 0, "# refills"); SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD, &iflib_fl_refills_large, 0, "# large refills"); static int iflib_txq_drain_flushing; static int iflib_txq_drain_oactive; static int iflib_txq_drain_notready; SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD, &iflib_txq_drain_flushing, 0, "# drain flushes"); SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD, &iflib_txq_drain_oactive, 0, "# drain oactives"); SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD, &iflib_txq_drain_notready, 0, "# drain notready"); static int iflib_encap_load_mbuf_fail; static int iflib_encap_pad_mbuf_fail; static int iflib_encap_txq_avail_fail; static int iflib_encap_txd_encap_fail; SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD, &iflib_encap_load_mbuf_fail, 0, "# busdma load failures"); SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD, &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures"); SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD, &iflib_encap_txq_avail_fail, 0, "# txq avail failures"); SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD, &iflib_encap_txd_encap_fail, 0, "# driver encap failures"); static int iflib_task_fn_rxs; static int iflib_rx_intr_enables; static int iflib_fast_intrs; static int iflib_rx_unavail; static int iflib_rx_ctx_inactive; static int iflib_rx_if_input; static int iflib_rxd_flush; static int iflib_verbose_debug; SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD, &iflib_task_fn_rxs, 0, "# task_fn_rx calls"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD, &iflib_rx_intr_enables, 0, "# RX intr enables"); SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD, &iflib_fast_intrs, 0, "# fast_intr calls"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD, &iflib_rx_unavail, 0, "# times rxeof called with no available data"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD, &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD, &iflib_rx_if_input, 0, "# times rxeof called if_input"); SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD, &iflib_rxd_flush, 0, "# times rxd_flush called"); SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW, &iflib_verbose_debug, 0, "enable verbose debugging"); #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1) static void iflib_debug_reset(void) { iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs = iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees = iflib_txq_drain_flushing = iflib_txq_drain_oactive = iflib_txq_drain_notready = iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail = iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail = iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs = iflib_rx_unavail = iflib_rx_ctx_inactive = iflib_rx_if_input = iflib_rxd_flush = 0; } #else #define DBG_COUNTER_INC(name) static void iflib_debug_reset(void) {} #endif #define IFLIB_DEBUG 0 static void iflib_tx_structures_free(if_ctx_t ctx); static void iflib_rx_structures_free(if_ctx_t ctx); static int iflib_queues_alloc(if_ctx_t ctx); static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq); static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget); static int iflib_qset_structures_setup(if_ctx_t ctx); static int iflib_msix_init(if_ctx_t ctx); static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str); static void iflib_txq_check_drain(iflib_txq_t txq, int budget); static uint32_t iflib_txq_can_drain(struct ifmp_ring *); #ifdef ALTQ static void iflib_altq_if_start(if_t ifp); static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m); #endif static int iflib_register(if_ctx_t); static void iflib_deregister(if_ctx_t); static void iflib_init_locked(if_ctx_t ctx); static void iflib_add_device_sysctl_pre(if_ctx_t ctx); static void iflib_add_device_sysctl_post(if_ctx_t ctx); static void iflib_ifmp_purge(iflib_txq_t txq); static void _iflib_pre_assert(if_softc_ctx_t scctx); static void iflib_if_init_locked(if_ctx_t ctx); static void iflib_free_intr_mem(if_ctx_t ctx); #ifndef __NO_STRICT_ALIGNMENT static struct mbuf * iflib_fixup_rx(struct mbuf *m); #endif static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets = SLIST_HEAD_INITIALIZER(cpu_offsets); struct cpu_offset { SLIST_ENTRY(cpu_offset) entries; cpuset_t set; unsigned int refcount; uint16_t offset; }; static struct mtx cpu_offset_mtx; MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock", MTX_DEF); NETDUMP_DEFINE(iflib); #ifdef DEV_NETMAP #include #include #include MODULE_DEPEND(iflib, netmap, 1, 1, 1); static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init); /* * device-specific sysctl variables: * * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it. * During regular operations the CRC is stripped, but on some * hardware reception of frames not multiple of 64 is slower, * so using crcstrip=0 helps in benchmarks. * * iflib_rx_miss, iflib_rx_miss_bufs: * count packets that might be missed due to lost interrupts. */ SYSCTL_DECL(_dev_netmap); /* * The xl driver by default strips CRCs and we do not override it. */ int iflib_crcstrip = 1; SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip, CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames"); int iflib_rx_miss, iflib_rx_miss_bufs; SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss, CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr"); SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs, CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs"); /* * Register/unregister. We are already under netmap lock. * Only called on the first register or the last unregister. */ static int iflib_netmap_register(struct netmap_adapter *na, int onoff) { if_t ifp = na->ifp; if_ctx_t ctx = ifp->if_softc; int status; CTX_LOCK(ctx); IFDI_INTR_DISABLE(ctx); /* Tell the stack that the interface is no longer active */ ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE); if (!CTX_IS_VF(ctx)) IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); /* enable or disable flags and callbacks in na and ifp */ if (onoff) { nm_set_native_flags(na); } else { nm_clear_native_flags(na); } iflib_stop(ctx); iflib_init_locked(ctx); IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ? status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1; if (status) nm_clear_native_flags(na); CTX_UNLOCK(ctx); return (status); } static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init) { struct netmap_adapter *na = kring->na; u_int const lim = kring->nkr_num_slots - 1; u_int head = kring->rhead; struct netmap_ring *ring = kring->ring; bus_dmamap_t *map; struct if_rxd_update iru; if_ctx_t ctx = rxq->ifr_ctx; iflib_fl_t fl = &rxq->ifr_fl[0]; uint32_t refill_pidx, nic_i; #if IFLIB_DEBUG_COUNTERS int rf_count = 0; #endif if (nm_i == head && __predict_true(!init)) return 0; iru_init(&iru, rxq, 0 /* flid */); map = fl->ifl_sds.ifsd_map; refill_pidx = netmap_idx_k2n(kring, nm_i); /* * IMPORTANT: we must leave one free slot in the ring, * so move head back by one unit */ head = nm_prev(head, lim); nic_i = UINT_MAX; DBG_COUNTER_INC(fl_refills); while (nm_i != head) { #if IFLIB_DEBUG_COUNTERS if (++rf_count == 9) DBG_COUNTER_INC(fl_refills_large); #endif for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) { struct netmap_slot *slot = &ring->slot[nm_i]; void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]); uint32_t nic_i_dma = refill_pidx; nic_i = netmap_idx_k2n(kring, nm_i); MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH); if (addr == NETMAP_BUF_BASE(na)) /* bad buf */ return netmap_ring_reinit(kring); fl->ifl_vm_addrs[tmp_pidx] = addr; if (__predict_false(init)) { netmap_load_map(na, fl->ifl_buf_tag, map[nic_i], addr); } else if (slot->flags & NS_BUF_CHANGED) { /* buffer has changed, reload map */ netmap_reload_map(na, fl->ifl_buf_tag, map[nic_i], addr); } slot->flags &= ~NS_BUF_CHANGED; nm_i = nm_next(nm_i, lim); fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim); if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1) continue; iru.iru_pidx = refill_pidx; iru.iru_count = tmp_pidx+1; ctx->isc_rxd_refill(ctx->ifc_softc, &iru); refill_pidx = nic_i; for (int n = 0; n < iru.iru_count; n++) { bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i_dma], BUS_DMASYNC_PREREAD); /* XXX - change this to not use the netmap func*/ nic_i_dma = nm_next(nic_i_dma, lim); } } } kring->nr_hwcur = head; bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); if (__predict_true(nic_i != UINT_MAX)) { ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i); DBG_COUNTER_INC(rxd_flush); } return (0); } /* * Reconcile kernel and user view of the transmit ring. * * All information is in the kring. * Userspace wants to send packets up to the one before kring->rhead, * kernel knows kring->nr_hwcur is the first unsent packet. * * Here we push packets out (as many as possible), and possibly * reclaim buffers from previously completed transmission. * * The caller (netmap) guarantees that there is only one instance * running at any time. Any interference with other driver * methods should be handled by the individual drivers. */ static int iflib_netmap_txsync(struct netmap_kring *kring, int flags) { struct netmap_adapter *na = kring->na; if_t ifp = na->ifp; struct netmap_ring *ring = kring->ring; u_int nm_i; /* index into the netmap kring */ u_int nic_i; /* index into the NIC ring */ u_int n; u_int const lim = kring->nkr_num_slots - 1; u_int const head = kring->rhead; struct if_pkt_info pi; /* * interrupts on every tx packet are expensive so request * them every half ring, or where NS_REPORT is set */ u_int report_frequency = kring->nkr_num_slots >> 1; /* device-specific */ if_ctx_t ctx = ifp->if_softc; iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id]; bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); /* * First part: process new packets to send. * nm_i is the current index in the netmap kring, * nic_i is the corresponding index in the NIC ring. * * If we have packets to send (nm_i != head) * iterate over the netmap ring, fetch length and update * the corresponding slot in the NIC ring. Some drivers also * need to update the buffer's physical address in the NIC slot * even NS_BUF_CHANGED is not set (PNMB computes the addresses). * * The netmap_reload_map() calls is especially expensive, * even when (as in this case) the tag is 0, so do only * when the buffer has actually changed. * * If possible do not set the report/intr bit on all slots, * but only a few times per ring or when NS_REPORT is set. * * Finally, on 10G and faster drivers, it might be useful * to prefetch the next slot and txr entry. */ nm_i = kring->nr_hwcur; if (nm_i != head) { /* we have new packets to send */ pkt_info_zero(&pi); pi.ipi_segs = txq->ift_segs; pi.ipi_qsidx = kring->ring_id; nic_i = netmap_idx_k2n(kring, nm_i); __builtin_prefetch(&ring->slot[nm_i]); __builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]); __builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]); for (n = 0; nm_i != head; n++) { struct netmap_slot *slot = &ring->slot[nm_i]; u_int len = slot->len; uint64_t paddr; void *addr = PNMB(na, slot, &paddr); int flags = (slot->flags & NS_REPORT || nic_i == 0 || nic_i == report_frequency) ? IPI_TX_INTR : 0; /* device-specific */ pi.ipi_len = len; pi.ipi_segs[0].ds_addr = paddr; pi.ipi_segs[0].ds_len = len; pi.ipi_nsegs = 1; pi.ipi_ndescs = 0; pi.ipi_pidx = nic_i; pi.ipi_flags = flags; /* Fill the slot in the NIC ring. */ ctx->isc_txd_encap(ctx->ifc_softc, &pi); DBG_COUNTER_INC(tx_encap); /* prefetch for next round */ __builtin_prefetch(&ring->slot[nm_i + 1]); __builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]); __builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]); NM_CHECK_ADDR_LEN(na, addr, len); if (slot->flags & NS_BUF_CHANGED) { /* buffer has changed, reload map */ netmap_reload_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[nic_i], addr); } /* make sure changes to the buffer are synced */ bus_dmamap_sync(txq->ift_buf_tag, txq->ift_sds.ifsd_map[nic_i], BUS_DMASYNC_PREWRITE); slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED); nm_i = nm_next(nm_i, lim); nic_i = nm_next(nic_i, lim); } kring->nr_hwcur = nm_i; /* synchronize the NIC ring */ bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); /* (re)start the tx unit up to slot nic_i (excluded) */ ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i); } /* * Second part: reclaim buffers for completed transmissions. * * If there are unclaimed buffers, attempt to reclaim them. * If none are reclaimed, and TX IRQs are not in use, do an initial * minimal delay, then trigger the tx handler which will spin in the * group task queue. */ if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) { if (iflib_tx_credits_update(ctx, txq)) { /* some tx completed, increment avail */ nic_i = txq->ift_cidx_processed; kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim); } } if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) { callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000, iflib_timer, txq, txq->ift_timer.c_cpu); } return (0); } /* * Reconcile kernel and user view of the receive ring. * Same as for the txsync, this routine must be efficient. * The caller guarantees a single invocations, but races against * the rest of the driver should be handled here. * * On call, kring->rhead is the first packet that userspace wants * to keep, and kring->rcur is the wakeup point. * The kernel has previously reported packets up to kring->rtail. * * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective * of whether or not we received an interrupt. */ static int iflib_netmap_rxsync(struct netmap_kring *kring, int flags) { struct netmap_adapter *na = kring->na; struct netmap_ring *ring = kring->ring; if_t ifp = na->ifp; iflib_fl_t fl; uint32_t nm_i; /* index into the netmap ring */ uint32_t nic_i; /* index into the NIC ring */ u_int i, n; u_int const lim = kring->nkr_num_slots - 1; u_int const head = kring->rhead; int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR; struct if_rxd_info ri; if_ctx_t ctx = ifp->if_softc; iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id]; if (head > lim) return netmap_ring_reinit(kring); /* * XXX netmap_fl_refill() only ever (re)fills free list 0 so far. */ for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) { bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); } /* * First part: import newly received packets. * * nm_i is the index of the next free slot in the netmap ring, * nic_i is the index of the next received packet in the NIC ring, * and they may differ in case if_init() has been called while * in netmap mode. For the receive ring we have * * nic_i = rxr->next_check; * nm_i = kring->nr_hwtail (previous) * and * nm_i == (nic_i + kring->nkr_hwofs) % ring_size * * rxr->next_check is set to 0 on a ring reinit */ if (netmap_no_pendintr || force_update) { int crclen = iflib_crcstrip ? 0 : 4; int error, avail; for (i = 0; i < rxq->ifr_nfl; i++) { fl = &rxq->ifr_fl[i]; nic_i = fl->ifl_cidx; nm_i = netmap_idx_n2k(kring, nic_i); avail = ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, nic_i, USHRT_MAX); for (n = 0; avail > 0; n++, avail--) { rxd_info_zero(&ri); ri.iri_frags = rxq->ifr_frags; ri.iri_qsidx = kring->ring_id; ri.iri_ifp = ctx->ifc_ifp; ri.iri_cidx = nic_i; error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri); ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen; ring->slot[nm_i].flags = 0; bus_dmamap_sync(fl->ifl_buf_tag, fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD); nm_i = nm_next(nm_i, lim); nic_i = nm_next(nic_i, lim); } if (n) { /* update the state variables */ if (netmap_no_pendintr && !force_update) { /* diagnostics */ iflib_rx_miss ++; iflib_rx_miss_bufs += n; } fl->ifl_cidx = nic_i; kring->nr_hwtail = nm_i; } kring->nr_kflags &= ~NKR_PENDINTR; } } /* * Second part: skip past packets that userspace has released. * (kring->nr_hwcur to head excluded), * and make the buffers available for reception. * As usual nm_i is the index in the netmap ring, * nic_i is the index in the NIC ring, and * nm_i == (nic_i + kring->nkr_hwofs) % ring_size */ /* XXX not sure how this will work with multiple free lists */ nm_i = kring->nr_hwcur; return (netmap_fl_refill(rxq, kring, nm_i, false)); } static void iflib_netmap_intr(struct netmap_adapter *na, int onoff) { if_ctx_t ctx = na->ifp->if_softc; CTX_LOCK(ctx); if (onoff) { IFDI_INTR_ENABLE(ctx); } else { IFDI_INTR_DISABLE(ctx); } CTX_UNLOCK(ctx); } static int iflib_netmap_attach(if_ctx_t ctx) { struct netmap_adapter na; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; bzero(&na, sizeof(na)); na.ifp = ctx->ifc_ifp; na.na_flags = NAF_BDG_MAYSLEEP; MPASS(ctx->ifc_softc_ctx.isc_ntxqsets); MPASS(ctx->ifc_softc_ctx.isc_nrxqsets); na.num_tx_desc = scctx->isc_ntxd[0]; na.num_rx_desc = scctx->isc_nrxd[0]; na.nm_txsync = iflib_netmap_txsync; na.nm_rxsync = iflib_netmap_rxsync; na.nm_register = iflib_netmap_register; na.nm_intr = iflib_netmap_intr; na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets; na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets; return (netmap_attach(&na)); } static void iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq) { struct netmap_adapter *na = NA(ctx->ifc_ifp); struct netmap_slot *slot; slot = netmap_reset(na, NR_TX, txq->ift_id, 0); if (slot == NULL) return; for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) { /* * In netmap mode, set the map for the packet buffer. * NOTE: Some drivers (not this one) also need to set * the physical buffer address in the NIC ring. * netmap_idx_n2k() maps a nic index, i, into the corresponding * netmap slot index, si */ int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i); netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si)); } } static void iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq) { struct netmap_adapter *na = NA(ctx->ifc_ifp); struct netmap_kring *kring = na->rx_rings[rxq->ifr_id]; struct netmap_slot *slot; uint32_t nm_i; slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0); if (slot == NULL) return; nm_i = netmap_idx_n2k(kring, 0); netmap_fl_refill(rxq, kring, nm_i, true); } static void iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on) { struct netmap_kring *kring; uint16_t txqid; txqid = txq->ift_id; kring = NA(ctx->ifc_ifp)->tx_rings[txqid]; if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) { bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_POSTREAD); if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) netmap_tx_irq(ctx->ifc_ifp, txqid); if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) { if (hz < 2000) *reset_on = 1; else *reset_on = hz / 1000; } } } #define iflib_netmap_detach(ifp) netmap_detach(ifp) #else #define iflib_netmap_txq_init(ctx, txq) #define iflib_netmap_rxq_init(ctx, rxq) #define iflib_netmap_detach(ifp) #define iflib_netmap_attach(ctx) (0) #define netmap_rx_irq(ifp, qid, budget) (0) #define netmap_tx_irq(ifp, qid) do {} while (0) #define iflib_netmap_timer_adjust(ctx, txq, reset_on) #endif #if defined(__i386__) || defined(__amd64__) static __inline void prefetch(void *x) { __asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x)); } static __inline void prefetch2cachelines(void *x) { __asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x)); #if (CACHE_LINE_SIZE < 128) __asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long))))); #endif } #else #define prefetch(x) #define prefetch2cachelines(x) #endif static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid) { iflib_fl_t fl; fl = &rxq->ifr_fl[flid]; iru->iru_paddrs = fl->ifl_bus_addrs; iru->iru_vaddrs = &fl->ifl_vm_addrs[0]; iru->iru_idxs = fl->ifl_rxd_idxs; iru->iru_qsidx = rxq->ifr_id; iru->iru_buf_size = fl->ifl_buf_size; iru->iru_flidx = fl->ifl_id; } static void _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err) { if (err) return; *(bus_addr_t *) arg = segs[0].ds_addr; } int iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags) { int err; device_t dev = ctx->ifc_dev; err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ align, 0, /* alignment, bounds */ BUS_SPACE_MAXADDR, /* lowaddr */ BUS_SPACE_MAXADDR, /* highaddr */ NULL, NULL, /* filter, filterarg */ size, /* maxsize */ 1, /* nsegments */ size, /* maxsegsize */ BUS_DMA_ALLOCNOW, /* flags */ NULL, /* lockfunc */ NULL, /* lockarg */ &dma->idi_tag); if (err) { device_printf(dev, "%s: bus_dma_tag_create failed: %d\n", __func__, err); goto fail_0; } err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr, BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map); if (err) { device_printf(dev, "%s: bus_dmamem_alloc(%ju) failed: %d\n", __func__, (uintmax_t)size, err); goto fail_1; } dma->idi_paddr = IF_BAD_DMA; err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr, size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT); if (err || dma->idi_paddr == IF_BAD_DMA) { device_printf(dev, "%s: bus_dmamap_load failed: %d\n", __func__, err); goto fail_2; } dma->idi_size = size; return (0); fail_2: bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map); fail_1: bus_dma_tag_destroy(dma->idi_tag); fail_0: dma->idi_tag = NULL; return (err); } int iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags) { if_shared_ctx_t sctx = ctx->ifc_sctx; KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized")); return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags)); } int iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count) { int i, err; iflib_dma_info_t *dmaiter; dmaiter = dmalist; for (i = 0; i < count; i++, dmaiter++) { if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0) break; } if (err) iflib_dma_free_multi(dmalist, i); return (err); } void iflib_dma_free(iflib_dma_info_t dma) { if (dma->idi_tag == NULL) return; if (dma->idi_paddr != IF_BAD_DMA) { bus_dmamap_sync(dma->idi_tag, dma->idi_map, BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(dma->idi_tag, dma->idi_map); dma->idi_paddr = IF_BAD_DMA; } if (dma->idi_vaddr != NULL) { bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map); dma->idi_vaddr = NULL; } bus_dma_tag_destroy(dma->idi_tag); dma->idi_tag = NULL; } void iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count) { int i; iflib_dma_info_t *dmaiter = dmalist; for (i = 0; i < count; i++, dmaiter++) iflib_dma_free(*dmaiter); } #ifdef EARLY_AP_STARTUP static const int iflib_started = 1; #else /* * We used to abuse the smp_started flag to decide if the queues have been * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()). * That gave bad races, since the SYSINIT() runs strictly after smp_started * is set. Run a SYSINIT() strictly after that to just set a usable * completion flag. */ static int iflib_started; static void iflib_record_started(void *arg) { iflib_started = 1; } SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST, iflib_record_started, NULL); #endif static int iflib_fast_intr(void *arg) { iflib_filter_info_t info = arg; struct grouptask *gtask = info->ifi_task; int result; if (!iflib_started) return (FILTER_STRAY); DBG_COUNTER_INC(fast_intrs); if (info->ifi_filter != NULL) { result = info->ifi_filter(info->ifi_filter_arg); if ((result & FILTER_SCHEDULE_THREAD) == 0) return (result); } GROUPTASK_ENQUEUE(gtask); return (FILTER_HANDLED); } static int iflib_fast_intr_rxtx(void *arg) { iflib_filter_info_t info = arg; struct grouptask *gtask = info->ifi_task; if_ctx_t ctx; iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx; iflib_txq_t txq; void *sc; int i, cidx, result; qidx_t txqid; bool intr_enable, intr_legacy; if (!iflib_started) return (FILTER_STRAY); DBG_COUNTER_INC(fast_intrs); if (info->ifi_filter != NULL) { result = info->ifi_filter(info->ifi_filter_arg); if ((result & FILTER_SCHEDULE_THREAD) == 0) return (result); } ctx = rxq->ifr_ctx; sc = ctx->ifc_softc; intr_enable = false; intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY); MPASS(rxq->ifr_ntxqirq); for (i = 0; i < rxq->ifr_ntxqirq; i++) { txqid = rxq->ifr_txqid[i]; txq = &ctx->ifc_txqs[txqid]; bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_POSTREAD); if (!ctx->isc_txd_credits_update(sc, txqid, false)) { if (intr_legacy) intr_enable = true; else IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid); continue; } GROUPTASK_ENQUEUE(&txq->ift_task); } if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ) cidx = rxq->ifr_cq_cidx; else cidx = rxq->ifr_fl[0].ifl_cidx; if (iflib_rxd_avail(ctx, rxq, cidx, 1)) GROUPTASK_ENQUEUE(gtask); else { if (intr_legacy) intr_enable = true; else IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id); DBG_COUNTER_INC(rx_intr_enables); } if (intr_enable) IFDI_INTR_ENABLE(ctx); return (FILTER_HANDLED); } static int iflib_fast_intr_ctx(void *arg) { iflib_filter_info_t info = arg; struct grouptask *gtask = info->ifi_task; int result; if (!iflib_started) return (FILTER_STRAY); DBG_COUNTER_INC(fast_intrs); if (info->ifi_filter != NULL) { result = info->ifi_filter(info->ifi_filter_arg); if ((result & FILTER_SCHEDULE_THREAD) == 0) return (result); } GROUPTASK_ENQUEUE(gtask); return (FILTER_HANDLED); } static int _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid, driver_filter_t filter, driver_intr_t handler, void *arg, const char *name) { struct resource *res; void *tag = NULL; device_t dev = ctx->ifc_dev; int flags, i, rc; flags = RF_ACTIVE; if (ctx->ifc_flags & IFC_LEGACY) flags |= RF_SHAREABLE; MPASS(rid < 512); i = rid; res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags); if (res == NULL) { device_printf(dev, "failed to allocate IRQ for rid %d, name %s.\n", rid, name); return (ENOMEM); } irq->ii_res = res; KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL")); rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET, filter, handler, arg, &tag); if (rc != 0) { device_printf(dev, "failed to setup interrupt for rid %d, name %s: %d\n", rid, name ? name : "unknown", rc); return (rc); } else if (name) bus_describe_intr(dev, res, tag, "%s", name); irq->ii_tag = tag; return (0); } /********************************************************************* * * Allocate DMA resources for TX buffers as well as memory for the TX * mbuf map. TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a * iflib_sw_tx_desc_array structure, storing all the information that * is needed to transmit a packet on the wire. This is called only * once at attach, setup is done every reset. * **********************************************************************/ static int iflib_txsd_alloc(iflib_txq_t txq) { if_ctx_t ctx = txq->ift_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; device_t dev = ctx->ifc_dev; bus_size_t tsomaxsize; int err, nsegments, ntsosegments; bool tso; nsegments = scctx->isc_tx_nsegments; ntsosegments = scctx->isc_tx_tso_segments_max; tsomaxsize = scctx->isc_tx_tso_size_max; if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU) tsomaxsize += sizeof(struct ether_vlan_header); MPASS(scctx->isc_ntxd[0] > 0); MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0); MPASS(nsegments > 0); if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) { MPASS(ntsosegments > 0); MPASS(sctx->isc_tso_maxsize >= tsomaxsize); } /* * Set up DMA tags for TX buffers. */ if ((err = bus_dma_tag_create(bus_get_dma_tag(dev), 1, 0, /* alignment, bounds */ BUS_SPACE_MAXADDR, /* lowaddr */ BUS_SPACE_MAXADDR, /* highaddr */ NULL, NULL, /* filter, filterarg */ sctx->isc_tx_maxsize, /* maxsize */ nsegments, /* nsegments */ sctx->isc_tx_maxsegsize, /* maxsegsize */ 0, /* flags */ NULL, /* lockfunc */ NULL, /* lockfuncarg */ &txq->ift_buf_tag))) { device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err); device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n", (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize); goto fail; } tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0; if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev), 1, 0, /* alignment, bounds */ BUS_SPACE_MAXADDR, /* lowaddr */ BUS_SPACE_MAXADDR, /* highaddr */ NULL, NULL, /* filter, filterarg */ tsomaxsize, /* maxsize */ ntsosegments, /* nsegments */ sctx->isc_tso_maxsegsize,/* maxsegsize */ 0, /* flags */ NULL, /* lockfunc */ NULL, /* lockfuncarg */ &txq->ift_tso_buf_tag))) { device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n", err); goto fail; } /* Allocate memory for the TX mbuf map. */ if (!(txq->ift_sds.ifsd_m = (struct mbuf **) malloc(sizeof(struct mbuf *) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate TX mbuf map memory\n"); err = ENOMEM; goto fail; } /* * Create the DMA maps for TX buffers. */ if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc( sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) { device_printf(dev, "Unable to allocate TX buffer DMA map memory\n"); err = ENOMEM; goto fail; } if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc( sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) { device_printf(dev, "Unable to allocate TSO TX buffer map memory\n"); err = ENOMEM; goto fail; } for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) { err = bus_dmamap_create(txq->ift_buf_tag, 0, &txq->ift_sds.ifsd_map[i]); if (err != 0) { device_printf(dev, "Unable to create TX DMA map\n"); goto fail; } if (!tso) continue; err = bus_dmamap_create(txq->ift_tso_buf_tag, 0, &txq->ift_sds.ifsd_tso_map[i]); if (err != 0) { device_printf(dev, "Unable to create TSO TX DMA map\n"); goto fail; } } return (0); fail: /* We free all, it handles case where we are in the middle */ iflib_tx_structures_free(ctx); return (err); } static void iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i) { bus_dmamap_t map; map = NULL; if (txq->ift_sds.ifsd_map != NULL) map = txq->ift_sds.ifsd_map[i]; if (map != NULL) { bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(txq->ift_buf_tag, map); bus_dmamap_destroy(txq->ift_buf_tag, map); txq->ift_sds.ifsd_map[i] = NULL; } map = NULL; if (txq->ift_sds.ifsd_tso_map != NULL) map = txq->ift_sds.ifsd_tso_map[i]; if (map != NULL) { bus_dmamap_sync(txq->ift_tso_buf_tag, map, BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(txq->ift_tso_buf_tag, map); bus_dmamap_destroy(txq->ift_tso_buf_tag, map); txq->ift_sds.ifsd_tso_map[i] = NULL; } } static void iflib_txq_destroy(iflib_txq_t txq) { if_ctx_t ctx = txq->ift_ctx; for (int i = 0; i < txq->ift_size; i++) iflib_txsd_destroy(ctx, txq, i); if (txq->ift_sds.ifsd_map != NULL) { free(txq->ift_sds.ifsd_map, M_IFLIB); txq->ift_sds.ifsd_map = NULL; } if (txq->ift_sds.ifsd_tso_map != NULL) { free(txq->ift_sds.ifsd_tso_map, M_IFLIB); txq->ift_sds.ifsd_tso_map = NULL; } if (txq->ift_sds.ifsd_m != NULL) { free(txq->ift_sds.ifsd_m, M_IFLIB); txq->ift_sds.ifsd_m = NULL; } if (txq->ift_buf_tag != NULL) { bus_dma_tag_destroy(txq->ift_buf_tag); txq->ift_buf_tag = NULL; } if (txq->ift_tso_buf_tag != NULL) { bus_dma_tag_destroy(txq->ift_tso_buf_tag); txq->ift_tso_buf_tag = NULL; } } static void iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i) { struct mbuf **mp; mp = &txq->ift_sds.ifsd_m[i]; if (*mp == NULL) return; if (txq->ift_sds.ifsd_map != NULL) { bus_dmamap_sync(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]); } if (txq->ift_sds.ifsd_tso_map != NULL) { bus_dmamap_sync(txq->ift_tso_buf_tag, txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(txq->ift_tso_buf_tag, txq->ift_sds.ifsd_tso_map[i]); } m_free(*mp); DBG_COUNTER_INC(tx_frees); *mp = NULL; } static int iflib_txq_setup(iflib_txq_t txq) { if_ctx_t ctx = txq->ift_ctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; iflib_dma_info_t di; int i; /* Set number of descriptors available */ txq->ift_qstatus = IFLIB_QUEUE_IDLE; /* XXX make configurable */ txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ; /* Reset indices */ txq->ift_cidx_processed = 0; txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0; txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset]; for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++) bzero((void *)di->idi_vaddr, di->idi_size); IFDI_TXQ_SETUP(ctx, txq->ift_id); for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++) bus_dmamap_sync(di->idi_tag, di->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); return (0); } /********************************************************************* * * Allocate DMA resources for RX buffers as well as memory for the RX * mbuf map, direct RX cluster pointer map and RX cluster bus address * map. RX DMA map, RX mbuf map, direct RX cluster pointer map and * RX cluster map are kept in a iflib_sw_rx_desc_array structure. * Since we use use one entry in iflib_sw_rx_desc_array per received * packet, the maximum number of entries we'll need is equal to the * number of hardware receive descriptors that we've allocated. * **********************************************************************/ static int iflib_rxsd_alloc(iflib_rxq_t rxq) { if_ctx_t ctx = rxq->ifr_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; device_t dev = ctx->ifc_dev; iflib_fl_t fl; int err; MPASS(scctx->isc_nrxd[0] > 0); MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0); fl = rxq->ifr_fl; for (int i = 0; i < rxq->ifr_nfl; i++, fl++) { fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */ /* Set up DMA tag for RX buffers. */ err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ 1, 0, /* alignment, bounds */ BUS_SPACE_MAXADDR, /* lowaddr */ BUS_SPACE_MAXADDR, /* highaddr */ NULL, NULL, /* filter, filterarg */ sctx->isc_rx_maxsize, /* maxsize */ sctx->isc_rx_nsegments, /* nsegments */ sctx->isc_rx_maxsegsize, /* maxsegsize */ 0, /* flags */ NULL, /* lockfunc */ NULL, /* lockarg */ &fl->ifl_buf_tag); if (err) { device_printf(dev, "Unable to allocate RX DMA tag: %d\n", err); goto fail; } /* Allocate memory for the RX mbuf map. */ if (!(fl->ifl_sds.ifsd_m = (struct mbuf **) malloc(sizeof(struct mbuf *) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate RX mbuf map memory\n"); err = ENOMEM; goto fail; } /* Allocate memory for the direct RX cluster pointer map. */ if (!(fl->ifl_sds.ifsd_cl = (caddr_t *) malloc(sizeof(caddr_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate RX cluster map memory\n"); err = ENOMEM; goto fail; } /* Allocate memory for the RX cluster bus address map. */ if (!(fl->ifl_sds.ifsd_ba = (bus_addr_t *) malloc(sizeof(bus_addr_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate RX bus address map memory\n"); err = ENOMEM; goto fail; } /* * Create the DMA maps for RX buffers. */ if (!(fl->ifl_sds.ifsd_map = (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate RX buffer DMA map memory\n"); err = ENOMEM; goto fail; } for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) { err = bus_dmamap_create(fl->ifl_buf_tag, 0, &fl->ifl_sds.ifsd_map[i]); if (err != 0) { device_printf(dev, "Unable to create RX buffer DMA map\n"); goto fail; } } } return (0); fail: iflib_rx_structures_free(ctx); return (err); } /* * Internal service routines */ struct rxq_refill_cb_arg { int error; bus_dma_segment_t seg; int nseg; }; static void _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error) { struct rxq_refill_cb_arg *cb_arg = arg; cb_arg->error = error; cb_arg->seg = segs[0]; cb_arg->nseg = nseg; } /** * _iflib_fl_refill - refill an rxq free-buffer list * @ctx: the iflib context * @fl: the free list to refill * @count: the number of new buffers to allocate * * (Re)populate an rxq free-buffer list with up to @count new packet buffers. * The caller must assure that @count does not exceed the queue's capacity. */ static void _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count) { struct if_rxd_update iru; struct rxq_refill_cb_arg cb_arg; struct mbuf *m; caddr_t cl, *sd_cl; struct mbuf **sd_m; bus_dmamap_t *sd_map; bus_addr_t bus_addr, *sd_ba; int err, frag_idx, i, idx, n, pidx; qidx_t credits; sd_m = fl->ifl_sds.ifsd_m; sd_map = fl->ifl_sds.ifsd_map; sd_cl = fl->ifl_sds.ifsd_cl; sd_ba = fl->ifl_sds.ifsd_ba; pidx = fl->ifl_pidx; idx = pidx; frag_idx = fl->ifl_fragidx; credits = fl->ifl_credits; i = 0; n = count; MPASS(n > 0); MPASS(credits + n <= fl->ifl_size); if (pidx < fl->ifl_cidx) MPASS(pidx + n <= fl->ifl_cidx); if (pidx == fl->ifl_cidx && (credits < fl->ifl_size)) MPASS(fl->ifl_gen == 0); if (pidx > fl->ifl_cidx) MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx); DBG_COUNTER_INC(fl_refills); if (n > 8) DBG_COUNTER_INC(fl_refills_large); iru_init(&iru, fl->ifl_rxq, fl->ifl_id); while (n--) { /* * We allocate an uninitialized mbuf + cluster, mbuf is * initialized after rx. * * If the cluster is still set then we know a minimum sized packet was received */ bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size, &frag_idx); if (frag_idx < 0) bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx); MPASS(frag_idx >= 0); if ((cl = sd_cl[frag_idx]) == NULL) { if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL) break; cb_arg.error = 0; MPASS(sd_map != NULL); err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx], cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, BUS_DMA_NOWAIT); if (err != 0 || cb_arg.error) { /* * !zone_pack ? */ if (fl->ifl_zone == zone_pack) uma_zfree(fl->ifl_zone, cl); break; } sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr; sd_cl[frag_idx] = cl; #if MEMORY_LOGGING fl->ifl_cl_enqueued++; #endif } else { bus_addr = sd_ba[frag_idx]; } bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx], BUS_DMASYNC_PREREAD); if (sd_m[frag_idx] == NULL) { if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) { break; } sd_m[frag_idx] = m; } bit_set(fl->ifl_rx_bitmap, frag_idx); #if MEMORY_LOGGING fl->ifl_m_enqueued++; #endif DBG_COUNTER_INC(rx_allocs); fl->ifl_rxd_idxs[i] = frag_idx; fl->ifl_bus_addrs[i] = bus_addr; fl->ifl_vm_addrs[i] = cl; credits++; i++; MPASS(credits <= fl->ifl_size); if (++idx == fl->ifl_size) { fl->ifl_gen = 1; idx = 0; } if (n == 0 || i == IFLIB_MAX_RX_REFRESH) { iru.iru_pidx = pidx; iru.iru_count = i; ctx->isc_rxd_refill(ctx->ifc_softc, &iru); i = 0; pidx = idx; fl->ifl_pidx = idx; fl->ifl_credits = credits; } } if (i) { iru.iru_pidx = pidx; iru.iru_count = i; ctx->isc_rxd_refill(ctx->ifc_softc, &iru); fl->ifl_pidx = idx; fl->ifl_credits = credits; } DBG_COUNTER_INC(rxd_flush); if (fl->ifl_pidx == 0) pidx = fl->ifl_size - 1; else pidx = fl->ifl_pidx - 1; bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx); fl->ifl_fragidx = frag_idx; } static __inline void __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max) { /* we avoid allowing pidx to catch up with cidx as it confuses ixl */ int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1; #ifdef INVARIANTS int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1; #endif MPASS(fl->ifl_credits <= fl->ifl_size); MPASS(reclaimable == delta); if (reclaimable > 0) _iflib_fl_refill(ctx, fl, min(max, reclaimable)); } uint8_t iflib_in_detach(if_ctx_t ctx) { bool in_detach; STATE_LOCK(ctx); in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH); STATE_UNLOCK(ctx); return (in_detach); } static void iflib_fl_bufs_free(iflib_fl_t fl) { iflib_dma_info_t idi = fl->ifl_ifdi; bus_dmamap_t sd_map; uint32_t i; for (i = 0; i < fl->ifl_size; i++) { struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i]; caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i]; if (*sd_cl != NULL) { sd_map = fl->ifl_sds.ifsd_map[i]; bus_dmamap_sync(fl->ifl_buf_tag, sd_map, BUS_DMASYNC_POSTREAD); bus_dmamap_unload(fl->ifl_buf_tag, sd_map); if (*sd_cl != NULL) uma_zfree(fl->ifl_zone, *sd_cl); // XXX: Should this get moved out? if (iflib_in_detach(fl->ifl_rxq->ifr_ctx)) bus_dmamap_destroy(fl->ifl_buf_tag, sd_map); if (*sd_m != NULL) { m_init(*sd_m, M_NOWAIT, MT_DATA, 0); uma_zfree(zone_mbuf, *sd_m); } } else { MPASS(*sd_cl == NULL); MPASS(*sd_m == NULL); } #if MEMORY_LOGGING fl->ifl_m_dequeued++; fl->ifl_cl_dequeued++; #endif *sd_cl = NULL; *sd_m = NULL; } #ifdef INVARIANTS for (i = 0; i < fl->ifl_size; i++) { MPASS(fl->ifl_sds.ifsd_cl[i] == NULL); MPASS(fl->ifl_sds.ifsd_m[i] == NULL); } #endif /* * Reset free list values */ fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0; bzero(idi->idi_vaddr, idi->idi_size); } /********************************************************************* * * Initialize a free list and its buffers. * **********************************************************************/ static int iflib_fl_setup(iflib_fl_t fl) { iflib_rxq_t rxq = fl->ifl_rxq; if_ctx_t ctx = rxq->ifr_ctx; bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1); /* ** Free current RX buffer structs and their mbufs */ iflib_fl_bufs_free(fl); /* Now replenish the mbufs */ MPASS(fl->ifl_credits == 0); fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz; if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size) ctx->ifc_max_fl_buf_size = fl->ifl_buf_size; fl->ifl_cltype = m_gettype(fl->ifl_buf_size); fl->ifl_zone = m_getzone(fl->ifl_buf_size); /* avoid pre-allocating zillions of clusters to an idle card * potentially speeding up attach */ _iflib_fl_refill(ctx, fl, min(128, fl->ifl_size)); MPASS(min(128, fl->ifl_size) == fl->ifl_credits); if (min(128, fl->ifl_size) != fl->ifl_credits) return (ENOBUFS); /* * handle failure */ MPASS(rxq != NULL); MPASS(fl->ifl_ifdi != NULL); bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); return (0); } /********************************************************************* * * Free receive ring data structures * **********************************************************************/ static void iflib_rx_sds_free(iflib_rxq_t rxq) { iflib_fl_t fl; int i, j; if (rxq->ifr_fl != NULL) { for (i = 0; i < rxq->ifr_nfl; i++) { fl = &rxq->ifr_fl[i]; if (fl->ifl_buf_tag != NULL) { if (fl->ifl_sds.ifsd_map != NULL) { for (j = 0; j < fl->ifl_size; j++) { if (fl->ifl_sds.ifsd_map[j] == NULL) continue; bus_dmamap_sync( fl->ifl_buf_tag, fl->ifl_sds.ifsd_map[j], BUS_DMASYNC_POSTREAD); bus_dmamap_unload( fl->ifl_buf_tag, fl->ifl_sds.ifsd_map[j]); } } bus_dma_tag_destroy(fl->ifl_buf_tag); fl->ifl_buf_tag = NULL; } free(fl->ifl_sds.ifsd_m, M_IFLIB); free(fl->ifl_sds.ifsd_cl, M_IFLIB); free(fl->ifl_sds.ifsd_ba, M_IFLIB); free(fl->ifl_sds.ifsd_map, M_IFLIB); fl->ifl_sds.ifsd_m = NULL; fl->ifl_sds.ifsd_cl = NULL; fl->ifl_sds.ifsd_ba = NULL; fl->ifl_sds.ifsd_map = NULL; } free(rxq->ifr_fl, M_IFLIB); rxq->ifr_fl = NULL; rxq->ifr_cq_cidx = 0; } } /* * Timer routine */ static void iflib_timer(void *arg) { iflib_txq_t txq = arg; if_ctx_t ctx = txq->ift_ctx; if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; uint64_t this_tick = ticks; uint32_t reset_on = hz / 2; if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) return; /* ** Check on the state of the TX queue(s), this ** can be done without the lock because its RO ** and the HUNG state will be static if set. */ if (this_tick - txq->ift_last_timer_tick >= hz / 2) { txq->ift_last_timer_tick = this_tick; IFDI_TIMER(ctx, txq->ift_id); if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) && ((txq->ift_cleaned_prev == txq->ift_cleaned) || (sctx->isc_pause_frames == 0))) goto hung; if (ifmp_ring_is_stalled(txq->ift_br)) txq->ift_qstatus = IFLIB_QUEUE_HUNG; txq->ift_cleaned_prev = txq->ift_cleaned; } #ifdef DEV_NETMAP if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) iflib_netmap_timer_adjust(ctx, txq, &reset_on); #endif /* handle any laggards */ if (txq->ift_db_pending) GROUPTASK_ENQUEUE(&txq->ift_task); sctx->isc_pause_frames = 0; if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu); return; hung: device_printf(ctx->ifc_dev, "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n", txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx); STATE_LOCK(ctx); if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET); iflib_admin_intr_deferred(ctx); STATE_UNLOCK(ctx); } static void iflib_calc_rx_mbuf_sz(if_ctx_t ctx) { if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; /* * XXX don't set the max_frame_size to larger * than the hardware can handle */ if (sctx->isc_max_frame_size <= MCLBYTES) ctx->ifc_rx_mbuf_sz = MCLBYTES; else ctx->ifc_rx_mbuf_sz = MJUMPAGESIZE; } uint32_t iflib_get_rx_mbuf_sz(if_ctx_t ctx) { return (ctx->ifc_rx_mbuf_sz); } static void iflib_init_locked(if_ctx_t ctx) { if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; if_t ifp = ctx->ifc_ifp; iflib_fl_t fl; iflib_txq_t txq; iflib_rxq_t rxq; int i, j, tx_ip_csum_flags, tx_ip6_csum_flags; if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); IFDI_INTR_DISABLE(ctx); tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP); tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP); /* Set hardware offload abilities */ if_clearhwassist(ifp); if (if_getcapenable(ifp) & IFCAP_TXCSUM) if_sethwassistbits(ifp, tx_ip_csum_flags, 0); if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6) if_sethwassistbits(ifp, tx_ip6_csum_flags, 0); if (if_getcapenable(ifp) & IFCAP_TSO4) if_sethwassistbits(ifp, CSUM_IP_TSO, 0); if (if_getcapenable(ifp) & IFCAP_TSO6) if_sethwassistbits(ifp, CSUM_IP6_TSO, 0); for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) { CALLOUT_LOCK(txq); callout_stop(&txq->ift_timer); CALLOUT_UNLOCK(txq); iflib_netmap_txq_init(ctx, txq); } /* * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so * that drivers can use the value when setting up the hardware receive * buffers. */ iflib_calc_rx_mbuf_sz(ctx); #ifdef INVARIANTS i = if_getdrvflags(ifp); #endif IFDI_INIT(ctx); MPASS(if_getdrvflags(ifp) == i); for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) { /* XXX this should really be done on a per-queue basis */ if (if_getcapenable(ifp) & IFCAP_NETMAP) { MPASS(rxq->ifr_id == i); iflib_netmap_rxq_init(ctx, rxq); continue; } for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) { if (iflib_fl_setup(fl)) { device_printf(ctx->ifc_dev, "setting up free list %d failed - " "check cluster settings\n", j); goto done; } } } done: if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE); IFDI_INTR_ENABLE(ctx); txq = ctx->ifc_txqs; for (i = 0; i < sctx->isc_ntxqsets; i++, txq++) callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu); } static int iflib_media_change(if_t ifp) { if_ctx_t ctx = if_getsoftc(ifp); int err; CTX_LOCK(ctx); if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0) iflib_init_locked(ctx); CTX_UNLOCK(ctx); return (err); } static void iflib_media_status(if_t ifp, struct ifmediareq *ifmr) { if_ctx_t ctx = if_getsoftc(ifp); CTX_LOCK(ctx); IFDI_UPDATE_ADMIN_STATUS(ctx); IFDI_MEDIA_STATUS(ctx, ifmr); CTX_UNLOCK(ctx); } void iflib_stop(if_ctx_t ctx) { iflib_txq_t txq = ctx->ifc_txqs; iflib_rxq_t rxq = ctx->ifc_rxqs; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; iflib_dma_info_t di; iflib_fl_t fl; int i, j; /* Tell the stack that the interface is no longer active */ if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); IFDI_INTR_DISABLE(ctx); DELAY(1000); IFDI_STOP(ctx); DELAY(1000); iflib_debug_reset(); /* Wait for current tx queue users to exit to disarm watchdog timer. */ for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) { /* make sure all transmitters have completed before proceeding XXX */ CALLOUT_LOCK(txq); callout_stop(&txq->ift_timer); CALLOUT_UNLOCK(txq); /* clean any enqueued buffers */ iflib_ifmp_purge(txq); /* Free any existing tx buffers. */ for (j = 0; j < txq->ift_size; j++) { iflib_txsd_free(ctx, txq, j); } txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0; txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0; txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0; txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0; txq->ift_pullups = 0; ifmp_ring_reset_stats(txq->ift_br); for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++) bzero((void *)di->idi_vaddr, di->idi_size); } for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) { /* make sure all transmitters have completed before proceeding XXX */ rxq->ifr_cq_cidx = 0; for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++) bzero((void *)di->idi_vaddr, di->idi_size); /* also resets the free lists pidx/cidx */ for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) iflib_fl_bufs_free(fl); } } static inline caddr_t calc_next_rxd(iflib_fl_t fl, int cidx) { qidx_t size; int nrxd; caddr_t start, end, cur, next; nrxd = fl->ifl_size; size = fl->ifl_rxd_size; start = fl->ifl_ifdi->idi_vaddr; if (__predict_false(size == 0)) return (start); cur = start + size*cidx; end = start + size*nrxd; next = CACHE_PTR_NEXT(cur); return (next < end ? next : start); } static inline void prefetch_pkts(iflib_fl_t fl, int cidx) { int nextptr; int nrxd = fl->ifl_size; caddr_t next_rxd; nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1); prefetch(&fl->ifl_sds.ifsd_m[nextptr]); prefetch(&fl->ifl_sds.ifsd_cl[nextptr]); next_rxd = calc_next_rxd(fl, cidx); prefetch(next_rxd); prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]); prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]); prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]); prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]); prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]); prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]); prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]); prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]); } static struct mbuf * rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd, int *pf_rv, if_rxd_info_t ri) { bus_dmamap_t map; iflib_fl_t fl; caddr_t payload; struct mbuf *m; int flid, cidx, len, next; map = NULL; flid = irf->irf_flid; cidx = irf->irf_idx; fl = &rxq->ifr_fl[flid]; sd->ifsd_fl = fl; sd->ifsd_cidx = cidx; m = fl->ifl_sds.ifsd_m[cidx]; sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx]; fl->ifl_credits--; #if MEMORY_LOGGING fl->ifl_m_dequeued++; #endif if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH) prefetch_pkts(fl, cidx); next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1); prefetch(&fl->ifl_sds.ifsd_map[next]); map = fl->ifl_sds.ifsd_map[cidx]; next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1); /* not valid assert if bxe really does SGE from non-contiguous elements */ MPASS(fl->ifl_cidx == cidx); bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD); if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL) { payload = *sd->ifsd_cl; payload += ri->iri_pad; len = ri->iri_len - ri->iri_pad; *pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp, len | PFIL_MEMPTR | PFIL_IN, NULL); switch (*pf_rv) { case PFIL_DROPPED: case PFIL_CONSUMED: /* * The filter ate it. Everything is recycled. */ m = NULL; unload = 0; break; case PFIL_REALLOCED: /* * The filter copied it. Everything is recycled. */ m = pfil_mem2mbuf(payload); unload = 0; break; case PFIL_PASS: /* * Filter said it was OK, so receive like * normal */ fl->ifl_sds.ifsd_m[cidx] = NULL; break; default: MPASS(0); } } else { fl->ifl_sds.ifsd_m[cidx] = NULL; *pf_rv = PFIL_PASS; } if (unload) bus_dmamap_unload(fl->ifl_buf_tag, map); fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1); if (__predict_false(fl->ifl_cidx == 0)) fl->ifl_gen = 0; bit_clear(fl->ifl_rx_bitmap, cidx); return (m); } static struct mbuf * assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv) { struct mbuf *m, *mh, *mt; caddr_t cl; int *pf_rv_ptr, flags, i, padlen; bool consumed; i = 0; mh = NULL; consumed = false; *pf_rv = PFIL_PASS; pf_rv_ptr = pf_rv; do { m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd, pf_rv_ptr, ri); MPASS(*sd->ifsd_cl != NULL); /* * Exclude zero-length frags & frags from * packets the filter has consumed or dropped */ if (ri->iri_frags[i].irf_len == 0 || consumed || *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) { if (mh == NULL) { /* everything saved here */ consumed = true; pf_rv_ptr = NULL; continue; } /* XXX we can save the cluster here, but not the mbuf */ m_init(m, M_NOWAIT, MT_DATA, 0); m_free(m); continue; } if (mh == NULL) { flags = M_PKTHDR|M_EXT; mh = mt = m; padlen = ri->iri_pad; } else { flags = M_EXT; mt->m_next = m; mt = m; /* assuming padding is only on the first fragment */ padlen = 0; } cl = *sd->ifsd_cl; *sd->ifsd_cl = NULL; /* Can these two be made one ? */ m_init(m, M_NOWAIT, MT_DATA, flags); m_cljset(m, cl, sd->ifsd_fl->ifl_cltype); /* * These must follow m_init and m_cljset */ m->m_data += padlen; ri->iri_len -= padlen; m->m_len = ri->iri_frags[i].irf_len; } while (++i < ri->iri_nfrags); return (mh); } /* * Process one software descriptor */ static struct mbuf * iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri) { struct if_rxsd sd; struct mbuf *m; int pf_rv; /* should I merge this back in now that the two paths are basically duplicated? */ if (ri->iri_nfrags == 1 && ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) { m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd, &pf_rv, ri); if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED) return (m); if (pf_rv == PFIL_PASS) { m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR); #ifndef __NO_STRICT_ALIGNMENT if (!IP_ALIGNED(m)) m->m_data += 2; #endif memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len); m->m_len = ri->iri_frags[0].irf_len; } } else { m = assemble_segments(rxq, ri, &sd, &pf_rv); if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED) return (m); } m->m_pkthdr.len = ri->iri_len; m->m_pkthdr.rcvif = ri->iri_ifp; m->m_flags |= ri->iri_flags; m->m_pkthdr.ether_vtag = ri->iri_vtag; m->m_pkthdr.flowid = ri->iri_flowid; M_HASHTYPE_SET(m, ri->iri_rsstype); m->m_pkthdr.csum_flags = ri->iri_csum_flags; m->m_pkthdr.csum_data = ri->iri_csum_data; return (m); } #if defined(INET6) || defined(INET) static void iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6) { CURVNET_SET(lc->ifp->if_vnet); #if defined(INET6) *v6 = V_ip6_forwarding; #endif #if defined(INET) *v4 = V_ipforwarding; #endif CURVNET_RESTORE(); } /* * Returns true if it's possible this packet could be LROed. * if it returns false, it is guaranteed that tcp_lro_rx() * would not return zero. */ static bool iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding) { struct ether_header *eh; eh = mtod(m, struct ether_header *); switch (eh->ether_type) { #if defined(INET6) case htons(ETHERTYPE_IPV6): return (!v6_forwarding); #endif #if defined (INET) case htons(ETHERTYPE_IP): return (!v4_forwarding); #endif } return false; } #else static void iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused) { } #endif static bool iflib_rxeof(iflib_rxq_t rxq, qidx_t budget) { if_t ifp; if_ctx_t ctx = rxq->ifr_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; int avail, i; qidx_t *cidxp; struct if_rxd_info ri; int err, budget_left, rx_bytes, rx_pkts; iflib_fl_t fl; int lro_enabled; bool v4_forwarding, v6_forwarding, lro_possible; /* * XXX early demux data packets so that if_input processing only handles * acks in interrupt context */ struct mbuf *m, *mh, *mt, *mf; lro_possible = v4_forwarding = v6_forwarding = false; ifp = ctx->ifc_ifp; mh = mt = NULL; MPASS(budget > 0); rx_pkts = rx_bytes = 0; if (sctx->isc_flags & IFLIB_HAS_RXCQ) cidxp = &rxq->ifr_cq_cidx; else cidxp = &rxq->ifr_fl[0].ifl_cidx; if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) { for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++) __iflib_fl_refill_lt(ctx, fl, budget + 8); DBG_COUNTER_INC(rx_unavail); return (false); } /* pfil needs the vnet to be set */ CURVNET_SET_QUIET(ifp->if_vnet); for (budget_left = budget; budget_left > 0 && avail > 0;) { if (__predict_false(!CTX_ACTIVE(ctx))) { DBG_COUNTER_INC(rx_ctx_inactive); break; } /* * Reset client set fields to their default values */ rxd_info_zero(&ri); ri.iri_qsidx = rxq->ifr_id; ri.iri_cidx = *cidxp; ri.iri_ifp = ifp; ri.iri_frags = rxq->ifr_frags; err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri); if (err) goto err; rx_pkts += 1; rx_bytes += ri.iri_len; if (sctx->isc_flags & IFLIB_HAS_RXCQ) { *cidxp = ri.iri_cidx; /* Update our consumer index */ /* XXX NB: shurd - check if this is still safe */ while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) rxq->ifr_cq_cidx -= scctx->isc_nrxd[0]; /* was this only a completion queue message? */ if (__predict_false(ri.iri_nfrags == 0)) continue; } MPASS(ri.iri_nfrags != 0); MPASS(ri.iri_len != 0); /* will advance the cidx on the corresponding free lists */ m = iflib_rxd_pkt_get(rxq, &ri); avail--; budget_left--; if (avail == 0 && budget_left) avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left); if (__predict_false(m == NULL)) continue; /* imm_pkt: -- cxgb */ if (mh == NULL) mh = mt = m; else { mt->m_nextpkt = m; mt = m; } } CURVNET_RESTORE(); /* make sure that we can refill faster than drain */ for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++) __iflib_fl_refill_lt(ctx, fl, budget + 8); lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO); if (lro_enabled) iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding); mt = mf = NULL; while (mh != NULL) { m = mh; mh = mh->m_nextpkt; m->m_nextpkt = NULL; #ifndef __NO_STRICT_ALIGNMENT if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL) continue; #endif rx_bytes += m->m_pkthdr.len; rx_pkts++; #if defined(INET6) || defined(INET) if (lro_enabled) { if (!lro_possible) { lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding); if (lro_possible && mf != NULL) { ifp->if_input(ifp, mf); DBG_COUNTER_INC(rx_if_input); mt = mf = NULL; } } if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) == (CSUM_L4_CALC|CSUM_L4_VALID)) { if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0) continue; } } #endif if (lro_possible) { ifp->if_input(ifp, m); DBG_COUNTER_INC(rx_if_input); continue; } if (mf == NULL) mf = m; if (mt != NULL) mt->m_nextpkt = m; mt = m; } if (mf != NULL) { ifp->if_input(ifp, mf); DBG_COUNTER_INC(rx_if_input); } if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes); if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts); /* * Flush any outstanding LRO work */ #if defined(INET6) || defined(INET) tcp_lro_flush_all(&rxq->ifr_lc); #endif if (avail) return true; return (iflib_rxd_avail(ctx, rxq, *cidxp, 1)); err: STATE_LOCK(ctx); ctx->ifc_flags |= IFC_DO_RESET; iflib_admin_intr_deferred(ctx); STATE_UNLOCK(ctx); return (false); } #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1) static inline qidx_t txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use) { qidx_t notify_count = TXD_NOTIFY_COUNT(txq); qidx_t minthresh = txq->ift_size / 8; if (in_use > 4*minthresh) return (notify_count); if (in_use > 2*minthresh) return (notify_count >> 1); if (in_use > minthresh) return (notify_count >> 3); return (0); } static inline qidx_t txq_max_rs_deferred(iflib_txq_t txq) { qidx_t notify_count = TXD_NOTIFY_COUNT(txq); qidx_t minthresh = txq->ift_size / 8; if (txq->ift_in_use > 4*minthresh) return (notify_count); if (txq->ift_in_use > 2*minthresh) return (notify_count >> 1); if (txq->ift_in_use > minthresh) return (notify_count >> 2); return (2); } #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags) #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG) #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use)) #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq) #define TXQ_MAX_DB_CONSUMED(size) (size >> 4) /* forward compatibility for cxgb */ #define FIRST_QSET(ctx) 0 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets) #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets) #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx)) #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments)) /* XXX we should be setting this to something other than zero */ #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh) #define MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \ (ctx)->ifc_softc_ctx.isc_tx_nsegments) static inline bool iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use) { qidx_t dbval, max; bool rang; rang = false; max = TXQ_MAX_DB_DEFERRED(txq, in_use); if (ring || txq->ift_db_pending >= max) { dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx; bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval); txq->ift_db_pending = txq->ift_npending = 0; rang = true; } return (rang); } #ifdef PKT_DEBUG static void print_pkt(if_pkt_info_t pi) { printf("pi len: %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n", pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx); printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n", pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag); printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n", pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto); } #endif #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO) #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO)) #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO) #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO)) static int iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp) { if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx; struct ether_vlan_header *eh; struct mbuf *m; m = *mp; if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) && M_WRITABLE(m) == 0) { if ((m = m_dup(m, M_NOWAIT)) == NULL) { return (ENOMEM); } else { m_freem(*mp); DBG_COUNTER_INC(tx_frees); *mp = m; } } /* * Determine where frame payload starts. * Jump over vlan headers if already present, * helpful for QinQ too. */ if (__predict_false(m->m_len < sizeof(*eh))) { txq->ift_pullups++; if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL)) return (ENOMEM); } eh = mtod(m, struct ether_vlan_header *); if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) { pi->ipi_etype = ntohs(eh->evl_proto); pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN; } else { pi->ipi_etype = ntohs(eh->evl_encap_proto); pi->ipi_ehdrlen = ETHER_HDR_LEN; } switch (pi->ipi_etype) { #ifdef INET case ETHERTYPE_IP: { struct mbuf *n; struct ip *ip = NULL; struct tcphdr *th = NULL; int minthlen; minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th)); if (__predict_false(m->m_len < minthlen)) { /* * if this code bloat is causing too much of a hit * move it to a separate function and mark it noinline */ if (m->m_len == pi->ipi_ehdrlen) { n = m->m_next; MPASS(n); if (n->m_len >= sizeof(*ip)) { ip = (struct ip *)n->m_data; if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th)) th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2)); } else { txq->ift_pullups++; if (__predict_false((m = m_pullup(m, minthlen)) == NULL)) return (ENOMEM); ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen); } } else { txq->ift_pullups++; if (__predict_false((m = m_pullup(m, minthlen)) == NULL)) return (ENOMEM); ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen); if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th)) th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2)); } } else { ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen); if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th)) th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2)); } pi->ipi_ip_hlen = ip->ip_hl << 2; pi->ipi_ipproto = ip->ip_p; pi->ipi_flags |= IPI_TX_IPV4; /* TCP checksum offload may require TCP header length */ if (IS_TX_OFFLOAD4(pi)) { if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) { if (__predict_false(th == NULL)) { txq->ift_pullups++; if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL)) return (ENOMEM); th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen); } pi->ipi_tcp_hflags = th->th_flags; pi->ipi_tcp_hlen = th->th_off << 2; pi->ipi_tcp_seq = th->th_seq; } if (IS_TSO4(pi)) { if (__predict_false(ip->ip_p != IPPROTO_TCP)) return (ENXIO); /* * TSO always requires hardware checksum offload. */ pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP); th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(IPPROTO_TCP)); pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz; if (sctx->isc_flags & IFLIB_TSO_INIT_IP) { ip->ip_sum = 0; ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz); } } } if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP)) ip->ip_sum = 0; break; } #endif #ifdef INET6 case ETHERTYPE_IPV6: { struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen); struct tcphdr *th; pi->ipi_ip_hlen = sizeof(struct ip6_hdr); if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) { txq->ift_pullups++; if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL)) return (ENOMEM); } th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen); /* XXX-BZ this will go badly in case of ext hdrs. */ pi->ipi_ipproto = ip6->ip6_nxt; pi->ipi_flags |= IPI_TX_IPV6; /* TCP checksum offload may require TCP header length */ if (IS_TX_OFFLOAD6(pi)) { if (pi->ipi_ipproto == IPPROTO_TCP) { if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) { txq->ift_pullups++; if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL)) return (ENOMEM); } pi->ipi_tcp_hflags = th->th_flags; pi->ipi_tcp_hlen = th->th_off << 2; pi->ipi_tcp_seq = th->th_seq; } if (IS_TSO6(pi)) { if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP)) return (ENXIO); /* * TSO always requires hardware checksum offload. */ pi->ipi_csum_flags |= CSUM_IP6_TCP; th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0); pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz; } } break; } #endif default: pi->ipi_csum_flags &= ~CSUM_OFFLOAD; pi->ipi_ip_hlen = 0; break; } *mp = m; return (0); } /* * If dodgy hardware rejects the scatter gather chain we've handed it * we'll need to remove the mbuf chain from ifsg_m[] before we can add the * m_defrag'd mbufs */ static __noinline struct mbuf * iflib_remove_mbuf(iflib_txq_t txq) { int ntxd, pidx; struct mbuf *m, **ifsd_m; ifsd_m = txq->ift_sds.ifsd_m; ntxd = txq->ift_size; pidx = txq->ift_pidx & (ntxd - 1); ifsd_m = txq->ift_sds.ifsd_m; m = ifsd_m[pidx]; ifsd_m[pidx] = NULL; bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]); if (txq->ift_sds.ifsd_tso_map != NULL) bus_dmamap_unload(txq->ift_tso_buf_tag, txq->ift_sds.ifsd_tso_map[pidx]); #if MEMORY_LOGGING txq->ift_dequeued++; #endif return (m); } static inline caddr_t calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid) { qidx_t size; int ntxd; caddr_t start, end, cur, next; ntxd = txq->ift_size; size = txq->ift_txd_size[qid]; start = txq->ift_ifdi[qid].idi_vaddr; if (__predict_false(size == 0)) return (start); cur = start + size*cidx; end = start + size*ntxd; next = CACHE_PTR_NEXT(cur); return (next < end ? next : start); } /* * Pad an mbuf to ensure a minimum ethernet frame size. * min_frame_size is the frame size (less CRC) to pad the mbuf to */ static __noinline int iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size) { /* * 18 is enough bytes to pad an ARP packet to 46 bytes, and * and ARP message is the smallest common payload I can think of */ static char pad[18]; /* just zeros */ int n; struct mbuf *new_head; if (!M_WRITABLE(*m_head)) { new_head = m_dup(*m_head, M_NOWAIT); if (new_head == NULL) { m_freem(*m_head); device_printf(dev, "cannot pad short frame, m_dup() failed"); DBG_COUNTER_INC(encap_pad_mbuf_fail); DBG_COUNTER_INC(tx_frees); return ENOMEM; } m_freem(*m_head); *m_head = new_head; } for (n = min_frame_size - (*m_head)->m_pkthdr.len; n > 0; n -= sizeof(pad)) if (!m_append(*m_head, min(n, sizeof(pad)), pad)) break; if (n > 0) { m_freem(*m_head); device_printf(dev, "cannot pad short frame\n"); DBG_COUNTER_INC(encap_pad_mbuf_fail); DBG_COUNTER_INC(tx_frees); return (ENOBUFS); } return 0; } static int iflib_encap(iflib_txq_t txq, struct mbuf **m_headp) { if_ctx_t ctx; if_shared_ctx_t sctx; if_softc_ctx_t scctx; bus_dma_tag_t buf_tag; bus_dma_segment_t *segs; struct mbuf *m_head, **ifsd_m; void *next_txd; bus_dmamap_t map; struct if_pkt_info pi; int remap = 0; int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd; ctx = txq->ift_ctx; sctx = ctx->ifc_sctx; scctx = &ctx->ifc_softc_ctx; segs = txq->ift_segs; ntxd = txq->ift_size; m_head = *m_headp; map = NULL; /* * If we're doing TSO the next descriptor to clean may be quite far ahead */ cidx = txq->ift_cidx; pidx = txq->ift_pidx; if (ctx->ifc_flags & IFC_PREFETCH) { next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1); if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) { next_txd = calc_next_txd(txq, cidx, 0); prefetch(next_txd); } /* prefetch the next cache line of mbuf pointers and flags */ prefetch(&txq->ift_sds.ifsd_m[next]); prefetch(&txq->ift_sds.ifsd_map[next]); next = (cidx + CACHE_LINE_SIZE) & (ntxd-1); } map = txq->ift_sds.ifsd_map[pidx]; ifsd_m = txq->ift_sds.ifsd_m; if (m_head->m_pkthdr.csum_flags & CSUM_TSO) { buf_tag = txq->ift_tso_buf_tag; max_segs = scctx->isc_tx_tso_segments_max; map = txq->ift_sds.ifsd_tso_map[pidx]; MPASS(buf_tag != NULL); MPASS(max_segs > 0); } else { buf_tag = txq->ift_buf_tag; max_segs = scctx->isc_tx_nsegments; map = txq->ift_sds.ifsd_map[pidx]; } if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) && __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) { err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size); if (err) { DBG_COUNTER_INC(encap_txd_encap_fail); return err; } } m_head = *m_headp; pkt_info_zero(&pi); pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST)); pi.ipi_pidx = pidx; pi.ipi_qsidx = txq->ift_id; pi.ipi_len = m_head->m_pkthdr.len; pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags; pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0; /* deliberate bitwise OR to make one condition */ if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) { if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) { DBG_COUNTER_INC(encap_txd_encap_fail); return (err); } m_head = *m_headp; } retry: err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs, BUS_DMA_NOWAIT); defrag: if (__predict_false(err)) { switch (err) { case EFBIG: /* try collapse once and defrag once */ if (remap == 0) { m_head = m_collapse(*m_headp, M_NOWAIT, max_segs); /* try defrag if collapsing fails */ if (m_head == NULL) remap++; } if (remap == 1) { txq->ift_mbuf_defrag++; m_head = m_defrag(*m_headp, M_NOWAIT); } /* * remap should never be >1 unless bus_dmamap_load_mbuf_sg * failed to map an mbuf that was run through m_defrag */ MPASS(remap <= 1); if (__predict_false(m_head == NULL || remap > 1)) goto defrag_failed; remap++; *m_headp = m_head; goto retry; break; case ENOMEM: txq->ift_no_tx_dma_setup++; break; default: txq->ift_no_tx_dma_setup++; m_freem(*m_headp); DBG_COUNTER_INC(tx_frees); *m_headp = NULL; break; } txq->ift_map_failed++; DBG_COUNTER_INC(encap_load_mbuf_fail); DBG_COUNTER_INC(encap_txd_encap_fail); return (err); } ifsd_m[pidx] = m_head; /* * XXX assumes a 1 to 1 relationship between segments and * descriptors - this does not hold true on all drivers, e.g. * cxgb */ if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) { txq->ift_no_desc_avail++; bus_dmamap_unload(buf_tag, map); DBG_COUNTER_INC(encap_txq_avail_fail); DBG_COUNTER_INC(encap_txd_encap_fail); if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0) GROUPTASK_ENQUEUE(&txq->ift_task); return (ENOBUFS); } /* * On Intel cards we can greatly reduce the number of TX interrupts * we see by only setting report status on every Nth descriptor. * However, this also means that the driver will need to keep track * of the descriptors that RS was set on to check them for the DD bit. */ txq->ift_rs_pending += nsegs + 1; if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) || iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) { pi.ipi_flags |= IPI_TX_INTR; txq->ift_rs_pending = 0; } pi.ipi_segs = segs; pi.ipi_nsegs = nsegs; MPASS(pidx >= 0 && pidx < txq->ift_size); #ifdef PKT_DEBUG print_pkt(&pi); #endif if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) { bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE); DBG_COUNTER_INC(tx_encap); MPASS(pi.ipi_new_pidx < txq->ift_size); ndesc = pi.ipi_new_pidx - pi.ipi_pidx; if (pi.ipi_new_pidx < pi.ipi_pidx) { ndesc += txq->ift_size; txq->ift_gen = 1; } /* * drivers can need as many as * two sentinels */ MPASS(ndesc <= pi.ipi_nsegs + 2); MPASS(pi.ipi_new_pidx != pidx); MPASS(ndesc > 0); txq->ift_in_use += ndesc; /* * We update the last software descriptor again here because there may * be a sentinel and/or there may be more mbufs than segments */ txq->ift_pidx = pi.ipi_new_pidx; txq->ift_npending += pi.ipi_ndescs; } else { *m_headp = m_head = iflib_remove_mbuf(txq); if (err == EFBIG) { txq->ift_txd_encap_efbig++; if (remap < 2) { remap = 1; goto defrag; } } goto defrag_failed; } /* * err can't possibly be non-zero here, so we don't neet to test it * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail). */ return (err); defrag_failed: txq->ift_mbuf_defrag_failed++; txq->ift_map_failed++; m_freem(*m_headp); DBG_COUNTER_INC(tx_frees); *m_headp = NULL; DBG_COUNTER_INC(encap_txd_encap_fail); return (ENOMEM); } static void iflib_tx_desc_free(iflib_txq_t txq, int n) { uint32_t qsize, cidx, mask, gen; struct mbuf *m, **ifsd_m; bool do_prefetch; cidx = txq->ift_cidx; gen = txq->ift_gen; qsize = txq->ift_size; mask = qsize-1; ifsd_m = txq->ift_sds.ifsd_m; do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH); while (n-- > 0) { if (do_prefetch) { prefetch(ifsd_m[(cidx + 3) & mask]); prefetch(ifsd_m[(cidx + 4) & mask]); } if ((m = ifsd_m[cidx]) != NULL) { prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]); if (m->m_pkthdr.csum_flags & CSUM_TSO) { bus_dmamap_sync(txq->ift_tso_buf_tag, txq->ift_sds.ifsd_tso_map[cidx], BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(txq->ift_tso_buf_tag, txq->ift_sds.ifsd_tso_map[cidx]); } else { bus_dmamap_sync(txq->ift_buf_tag, txq->ift_sds.ifsd_map[cidx], BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[cidx]); } /* XXX we don't support any drivers that batch packets yet */ MPASS(m->m_nextpkt == NULL); m_freem(m); ifsd_m[cidx] = NULL; #if MEMORY_LOGGING txq->ift_dequeued++; #endif DBG_COUNTER_INC(tx_frees); } if (__predict_false(++cidx == qsize)) { cidx = 0; gen = 0; } } txq->ift_cidx = cidx; txq->ift_gen = gen; } static __inline int iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh) { int reclaim; if_ctx_t ctx = txq->ift_ctx; KASSERT(thresh >= 0, ("invalid threshold to reclaim")); MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size); /* * Need a rate-limiting check so that this isn't called every time */ iflib_tx_credits_update(ctx, txq); reclaim = DESC_RECLAIMABLE(txq); if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) { #ifdef INVARIANTS if (iflib_verbose_debug) { printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__, txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments, reclaim, thresh); } #endif return (0); } iflib_tx_desc_free(txq, reclaim); txq->ift_cleaned += reclaim; txq->ift_in_use -= reclaim; return (reclaim); } static struct mbuf ** _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining) { int next, size; struct mbuf **items; size = r->size; next = (cidx + CACHE_PTR_INCREMENT) & (size-1); items = __DEVOLATILE(struct mbuf **, &r->items[0]); prefetch(items[(cidx + offset) & (size-1)]); if (remaining > 1) { prefetch2cachelines(&items[next]); prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]); prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]); prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]); } return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)])); } static void iflib_txq_check_drain(iflib_txq_t txq, int budget) { ifmp_ring_check_drainage(txq->ift_br, budget); } static uint32_t iflib_txq_can_drain(struct ifmp_ring *r) { iflib_txq_t txq = r->cookie; if_ctx_t ctx = txq->ift_ctx; if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) return (1); bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_POSTREAD); return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false)); } static uint32_t iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx) { iflib_txq_t txq = r->cookie; if_ctx_t ctx = txq->ift_ctx; if_t ifp = ctx->ifc_ifp; struct mbuf *m, **mp; int avail, bytes_sent, consumed, count, err, i, in_use_prev; int mcast_sent, pkt_sent, reclaimed, txq_avail; bool do_prefetch, rang, ring; if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) || !LINK_ACTIVE(ctx))) { DBG_COUNTER_INC(txq_drain_notready); return (0); } reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx)); rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use); avail = IDXDIFF(pidx, cidx, r->size); if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) { DBG_COUNTER_INC(txq_drain_flushing); for (i = 0; i < avail; i++) { if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq)) m_free(r->items[(cidx + i) & (r->size-1)]); r->items[(cidx + i) & (r->size-1)] = NULL; } return (avail); } if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) { txq->ift_qstatus = IFLIB_QUEUE_IDLE; CALLOUT_LOCK(txq); callout_stop(&txq->ift_timer); CALLOUT_UNLOCK(txq); DBG_COUNTER_INC(txq_drain_oactive); return (0); } if (reclaimed) txq->ift_qstatus = IFLIB_QUEUE_IDLE; consumed = mcast_sent = bytes_sent = pkt_sent = 0; count = MIN(avail, TX_BATCH_SIZE); #ifdef INVARIANTS if (iflib_verbose_debug) printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__, avail, ctx->ifc_flags, TXQ_AVAIL(txq)); #endif do_prefetch = (ctx->ifc_flags & IFC_PREFETCH); txq_avail = TXQ_AVAIL(txq); err = 0; for (i = 0; i < count && txq_avail > MAX_TX_DESC(ctx) + 2; i++) { int rem = do_prefetch ? count - i : 0; mp = _ring_peek_one(r, cidx, i, rem); MPASS(mp != NULL && *mp != NULL); if (__predict_false(*mp == (struct mbuf *)txq)) { consumed++; continue; } in_use_prev = txq->ift_in_use; err = iflib_encap(txq, mp); if (__predict_false(err)) { /* no room - bail out */ if (err == ENOBUFS) break; consumed++; /* we can't send this packet - skip it */ continue; } consumed++; pkt_sent++; m = *mp; DBG_COUNTER_INC(tx_sent); bytes_sent += m->m_pkthdr.len; mcast_sent += !!(m->m_flags & M_MCAST); txq_avail = TXQ_AVAIL(txq); txq->ift_db_pending += (txq->ift_in_use - in_use_prev); ETHER_BPF_MTAP(ifp, m); if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING))) break; rang = iflib_txd_db_check(ctx, txq, false, in_use_prev); } /* deliberate use of bitwise or to avoid gratuitous short-circuit */ ring = rang ? false : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx)); iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use); if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent); if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent); if (mcast_sent) if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent); #ifdef INVARIANTS if (iflib_verbose_debug) printf("consumed=%d\n", consumed); #endif return (consumed); } static uint32_t iflib_txq_drain_always(struct ifmp_ring *r) { return (1); } static uint32_t iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx) { int i, avail; struct mbuf **mp; iflib_txq_t txq; txq = r->cookie; txq->ift_qstatus = IFLIB_QUEUE_IDLE; CALLOUT_LOCK(txq); callout_stop(&txq->ift_timer); CALLOUT_UNLOCK(txq); avail = IDXDIFF(pidx, cidx, r->size); for (i = 0; i < avail; i++) { mp = _ring_peek_one(r, cidx, i, avail - i); if (__predict_false(*mp == (struct mbuf *)txq)) continue; m_freem(*mp); DBG_COUNTER_INC(tx_frees); } MPASS(ifmp_ring_is_stalled(r) == 0); return (avail); } static void iflib_ifmp_purge(iflib_txq_t txq) { struct ifmp_ring *r; r = txq->ift_br; r->drain = iflib_txq_drain_free; r->can_drain = iflib_txq_drain_always; ifmp_ring_check_drainage(r, r->size); r->drain = iflib_txq_drain; r->can_drain = iflib_txq_can_drain; } static void _task_fn_tx(void *context) { iflib_txq_t txq = context; if_ctx_t ctx = txq->ift_ctx; #if defined(ALTQ) || defined(DEV_NETMAP) if_t ifp = ctx->ifc_ifp; #endif int abdicate = ctx->ifc_sysctl_tx_abdicate; #ifdef IFLIB_DIAGNOSTICS txq->ift_cpu_exec_count[curcpu]++; #endif if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) return; #ifdef DEV_NETMAP if (if_getcapenable(ifp) & IFCAP_NETMAP) { bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_POSTREAD); if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false)) netmap_tx_irq(ifp, txq->ift_id); if (ctx->ifc_flags & IFC_LEGACY) IFDI_INTR_ENABLE(ctx); else IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id); return; } #endif #ifdef ALTQ if (ALTQ_IS_ENABLED(&ifp->if_snd)) iflib_altq_if_start(ifp); #endif if (txq->ift_db_pending) ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate); else if (!abdicate) ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE); /* * When abdicating, we always need to check drainage, not just when we don't enqueue */ if (abdicate) ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE); if (ctx->ifc_flags & IFC_LEGACY) IFDI_INTR_ENABLE(ctx); else IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id); } static void _task_fn_rx(void *context) { iflib_rxq_t rxq = context; if_ctx_t ctx = rxq->ifr_ctx; bool more; uint16_t budget; #ifdef IFLIB_DIAGNOSTICS rxq->ifr_cpu_exec_count[curcpu]++; #endif DBG_COUNTER_INC(task_fn_rxs); if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))) return; more = true; #ifdef DEV_NETMAP if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) { u_int work = 0; if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) { more = false; } } #endif budget = ctx->ifc_sysctl_rx_budget; if (budget == 0) budget = 16; /* XXX */ if (more == false || (more = iflib_rxeof(rxq, budget)) == false) { if (ctx->ifc_flags & IFC_LEGACY) IFDI_INTR_ENABLE(ctx); else IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id); DBG_COUNTER_INC(rx_intr_enables); } if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))) return; if (more) GROUPTASK_ENQUEUE(&rxq->ifr_task); } static void _task_fn_admin(void *context) { if_ctx_t ctx = context; if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; iflib_txq_t txq; int i; bool oactive, running, do_reset, do_watchdog, in_detach; uint32_t reset_on = hz / 2; STATE_LOCK(ctx); running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING); oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE); do_reset = (ctx->ifc_flags & IFC_DO_RESET); do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG); in_detach = (ctx->ifc_flags & IFC_IN_DETACH); ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG); STATE_UNLOCK(ctx); if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN)) return; if (in_detach) return; CTX_LOCK(ctx); for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) { CALLOUT_LOCK(txq); callout_stop(&txq->ift_timer); CALLOUT_UNLOCK(txq); } if (do_watchdog) { ctx->ifc_watchdog_events++; IFDI_WATCHDOG_RESET(ctx); } IFDI_UPDATE_ADMIN_STATUS(ctx); for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) { #ifdef DEV_NETMAP reset_on = hz / 2; if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) iflib_netmap_timer_adjust(ctx, txq, &reset_on); #endif callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu); } IFDI_LINK_INTR_ENABLE(ctx); if (do_reset) iflib_if_init_locked(ctx); CTX_UNLOCK(ctx); if (LINK_ACTIVE(ctx) == 0) return; for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET); } static void _task_fn_iov(void *context) { if_ctx_t ctx = context; if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN)) return; CTX_LOCK(ctx); IFDI_VFLR_HANDLE(ctx); CTX_UNLOCK(ctx); } static int iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS) { int err; if_int_delay_info_t info; if_ctx_t ctx; info = (if_int_delay_info_t)arg1; ctx = info->iidi_ctx; info->iidi_req = req; info->iidi_oidp = oidp; CTX_LOCK(ctx); err = IFDI_SYSCTL_INT_DELAY(ctx, info); CTX_UNLOCK(ctx); return (err); } /********************************************************************* * * IFNET FUNCTIONS * **********************************************************************/ static void iflib_if_init_locked(if_ctx_t ctx) { iflib_stop(ctx); iflib_init_locked(ctx); } static void iflib_if_init(void *arg) { if_ctx_t ctx = arg; CTX_LOCK(ctx); iflib_if_init_locked(ctx); CTX_UNLOCK(ctx); } static int iflib_if_transmit(if_t ifp, struct mbuf *m) { if_ctx_t ctx = if_getsoftc(ifp); iflib_txq_t txq; int err, qidx; int abdicate = ctx->ifc_sysctl_tx_abdicate; if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) { DBG_COUNTER_INC(tx_frees); m_freem(m); return (ENETDOWN); } MPASS(m->m_nextpkt == NULL); /* ALTQ-enabled interfaces always use queue 0. */ qidx = 0; if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd)) qidx = QIDX(ctx, m); /* * XXX calculate buf_ring based on flowid (divvy up bits?) */ txq = &ctx->ifc_txqs[qidx]; #ifdef DRIVER_BACKPRESSURE if (txq->ift_closed) { while (m != NULL) { next = m->m_nextpkt; m->m_nextpkt = NULL; m_freem(m); DBG_COUNTER_INC(tx_frees); m = next; } return (ENOBUFS); } #endif #ifdef notyet qidx = count = 0; mp = marr; next = m; do { count++; next = next->m_nextpkt; } while (next != NULL); if (count > nitems(marr)) if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) { /* XXX check nextpkt */ m_freem(m); /* XXX simplify for now */ DBG_COUNTER_INC(tx_frees); return (ENOBUFS); } for (next = m, i = 0; next != NULL; i++) { mp[i] = next; next = next->m_nextpkt; mp[i]->m_nextpkt = NULL; } #endif DBG_COUNTER_INC(tx_seen); err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate); if (abdicate) GROUPTASK_ENQUEUE(&txq->ift_task); if (err) { if (!abdicate) GROUPTASK_ENQUEUE(&txq->ift_task); /* support forthcoming later */ #ifdef DRIVER_BACKPRESSURE txq->ift_closed = TRUE; #endif ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE); m_freem(m); DBG_COUNTER_INC(tx_frees); } return (err); } #ifdef ALTQ /* * The overall approach to integrating iflib with ALTQ is to continue to use * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware * ring. Technically, when using ALTQ, queueing to an intermediate mp_ring * is redundant/unnecessary, but doing so minimizes the amount of * ALTQ-specific code required in iflib. It is assumed that the overhead of * redundantly queueing to an intermediate mp_ring is swamped by the * performance limitations inherent in using ALTQ. * * When ALTQ support is compiled in, all iflib drivers will use a transmit * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the * given interface. If ALTQ is enabled for an interface, then all * transmitted packets for that interface will be submitted to the ALTQ * subsystem via IFQ_ENQUEUE(). We don't use the legacy if_transmit() * implementation because it uses IFQ_HANDOFF(), which will duplicatively * update stats that the iflib machinery handles, and which is sensitve to * the disused IFF_DRV_OACTIVE flag. Additionally, iflib_altq_if_start() * will be installed as the start routine for use by ALTQ facilities that * need to trigger queue drains on a scheduled basis. * */ static void iflib_altq_if_start(if_t ifp) { struct ifaltq *ifq = &ifp->if_snd; struct mbuf *m; IFQ_LOCK(ifq); IFQ_DEQUEUE_NOLOCK(ifq, m); while (m != NULL) { iflib_if_transmit(ifp, m); IFQ_DEQUEUE_NOLOCK(ifq, m); } IFQ_UNLOCK(ifq); } static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m) { int err; if (ALTQ_IS_ENABLED(&ifp->if_snd)) { IFQ_ENQUEUE(&ifp->if_snd, m, err); if (err == 0) iflib_altq_if_start(ifp); } else err = iflib_if_transmit(ifp, m); return (err); } #endif /* ALTQ */ static void iflib_if_qflush(if_t ifp) { if_ctx_t ctx = if_getsoftc(ifp); iflib_txq_t txq = ctx->ifc_txqs; int i; STATE_LOCK(ctx); ctx->ifc_flags |= IFC_QFLUSH; STATE_UNLOCK(ctx); for (i = 0; i < NTXQSETS(ctx); i++, txq++) while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br))) iflib_txq_check_drain(txq, 0); STATE_LOCK(ctx); ctx->ifc_flags &= ~IFC_QFLUSH; STATE_UNLOCK(ctx); /* * When ALTQ is enabled, this will also take care of purging the * ALTQ queue(s). */ if_qflush(ifp); } #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \ IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \ IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \ - IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM) + IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_NOMAP) static int iflib_if_ioctl(if_t ifp, u_long command, caddr_t data) { if_ctx_t ctx = if_getsoftc(ifp); struct ifreq *ifr = (struct ifreq *)data; #if defined(INET) || defined(INET6) struct ifaddr *ifa = (struct ifaddr *)data; #endif bool avoid_reset = false; int err = 0, reinit = 0, bits; switch (command) { case SIOCSIFADDR: #ifdef INET if (ifa->ifa_addr->sa_family == AF_INET) avoid_reset = true; #endif #ifdef INET6 if (ifa->ifa_addr->sa_family == AF_INET6) avoid_reset = true; #endif /* ** Calling init results in link renegotiation, ** so we avoid doing it when possible. */ if (avoid_reset) { if_setflagbits(ifp, IFF_UP,0); if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING)) reinit = 1; #ifdef INET if (!(if_getflags(ifp) & IFF_NOARP)) arp_ifinit(ifp, ifa); #endif } else err = ether_ioctl(ifp, command, data); break; case SIOCSIFMTU: CTX_LOCK(ctx); if (ifr->ifr_mtu == if_getmtu(ifp)) { CTX_UNLOCK(ctx); break; } bits = if_getdrvflags(ifp); /* stop the driver and free any clusters before proceeding */ iflib_stop(ctx); if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) { STATE_LOCK(ctx); if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size) ctx->ifc_flags |= IFC_MULTISEG; else ctx->ifc_flags &= ~IFC_MULTISEG; STATE_UNLOCK(ctx); err = if_setmtu(ifp, ifr->ifr_mtu); } iflib_init_locked(ctx); STATE_LOCK(ctx); if_setdrvflags(ifp, bits); STATE_UNLOCK(ctx); CTX_UNLOCK(ctx); break; case SIOCSIFFLAGS: CTX_LOCK(ctx); if (if_getflags(ifp) & IFF_UP) { if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { if ((if_getflags(ifp) ^ ctx->ifc_if_flags) & (IFF_PROMISC | IFF_ALLMULTI)) { err = IFDI_PROMISC_SET(ctx, if_getflags(ifp)); } } else reinit = 1; } else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { iflib_stop(ctx); } ctx->ifc_if_flags = if_getflags(ifp); CTX_UNLOCK(ctx); break; case SIOCADDMULTI: case SIOCDELMULTI: if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { CTX_LOCK(ctx); IFDI_INTR_DISABLE(ctx); IFDI_MULTI_SET(ctx); IFDI_INTR_ENABLE(ctx); CTX_UNLOCK(ctx); } break; case SIOCSIFMEDIA: CTX_LOCK(ctx); IFDI_MEDIA_SET(ctx); CTX_UNLOCK(ctx); /* FALLTHROUGH */ case SIOCGIFMEDIA: case SIOCGIFXMEDIA: err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command); break; case SIOCGI2C: { struct ifi2creq i2c; err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c)); if (err != 0) break; if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) { err = EINVAL; break; } if (i2c.len > sizeof(i2c.data)) { err = EINVAL; break; } if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0) err = copyout(&i2c, ifr_data_get_ptr(ifr), sizeof(i2c)); break; } case SIOCSIFCAP: { int mask, setmask, oldmask; oldmask = if_getcapenable(ifp); mask = ifr->ifr_reqcap ^ oldmask; - mask &= ctx->ifc_softc_ctx.isc_capabilities; + mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_NOMAP; setmask = 0; #ifdef TCP_OFFLOAD setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6); #endif setmask |= (mask & IFCAP_FLAGS); setmask |= (mask & IFCAP_WOL); /* * If any RX csum has changed, change all the ones that * are supported by the driver. */ if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) { setmask |= ctx->ifc_softc_ctx.isc_capabilities & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6); } /* * want to ensure that traffic has stopped before we change any of the flags */ if (setmask) { CTX_LOCK(ctx); bits = if_getdrvflags(ifp); if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL) iflib_stop(ctx); STATE_LOCK(ctx); if_togglecapenable(ifp, setmask); STATE_UNLOCK(ctx); if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL) iflib_init_locked(ctx); STATE_LOCK(ctx); if_setdrvflags(ifp, bits); STATE_UNLOCK(ctx); CTX_UNLOCK(ctx); } if_vlancap(ifp); break; } case SIOCGPRIVATE_0: case SIOCSDRVSPEC: case SIOCGDRVSPEC: CTX_LOCK(ctx); err = IFDI_PRIV_IOCTL(ctx, command, data); CTX_UNLOCK(ctx); break; default: err = ether_ioctl(ifp, command, data); break; } if (reinit) iflib_if_init(ctx); return (err); } static uint64_t iflib_if_get_counter(if_t ifp, ift_counter cnt) { if_ctx_t ctx = if_getsoftc(ifp); return (IFDI_GET_COUNTER(ctx, cnt)); } /********************************************************************* * * OTHER FUNCTIONS EXPORTED TO THE STACK * **********************************************************************/ static void iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag) { if_ctx_t ctx = if_getsoftc(ifp); if ((void *)ctx != arg) return; if ((vtag == 0) || (vtag > 4095)) return; if (iflib_in_detach(ctx)) return; CTX_LOCK(ctx); IFDI_VLAN_REGISTER(ctx, vtag); /* Re-init to load the changes */ if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER) iflib_if_init_locked(ctx); CTX_UNLOCK(ctx); } static void iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag) { if_ctx_t ctx = if_getsoftc(ifp); if ((void *)ctx != arg) return; if ((vtag == 0) || (vtag > 4095)) return; CTX_LOCK(ctx); IFDI_VLAN_UNREGISTER(ctx, vtag); /* Re-init to load the changes */ if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER) iflib_if_init_locked(ctx); CTX_UNLOCK(ctx); } static void iflib_led_func(void *arg, int onoff) { if_ctx_t ctx = arg; CTX_LOCK(ctx); IFDI_LED_FUNC(ctx, onoff); CTX_UNLOCK(ctx); } /********************************************************************* * * BUS FUNCTION DEFINITIONS * **********************************************************************/ int iflib_device_probe(device_t dev) { const pci_vendor_info_t *ent; if_shared_ctx_t sctx; uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id; uint16_t pci_vendor_id; if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC) return (ENOTSUP); pci_vendor_id = pci_get_vendor(dev); pci_device_id = pci_get_device(dev); pci_subvendor_id = pci_get_subvendor(dev); pci_subdevice_id = pci_get_subdevice(dev); pci_rev_id = pci_get_revid(dev); if (sctx->isc_parse_devinfo != NULL) sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id); ent = sctx->isc_vendor_info; while (ent->pvi_vendor_id != 0) { if (pci_vendor_id != ent->pvi_vendor_id) { ent++; continue; } if ((pci_device_id == ent->pvi_device_id) && ((pci_subvendor_id == ent->pvi_subvendor_id) || (ent->pvi_subvendor_id == 0)) && ((pci_subdevice_id == ent->pvi_subdevice_id) || (ent->pvi_subdevice_id == 0)) && ((pci_rev_id == ent->pvi_rev_id) || (ent->pvi_rev_id == 0))) { device_set_desc_copy(dev, ent->pvi_name); /* this needs to be changed to zero if the bus probing code * ever stops re-probing on best match because the sctx * may have its values over written by register calls * in subsequent probes */ return (BUS_PROBE_DEFAULT); } ent++; } return (ENXIO); } int iflib_device_probe_vendor(device_t dev) { int probe; probe = iflib_device_probe(dev); if (probe == BUS_PROBE_DEFAULT) return (BUS_PROBE_VENDOR); else return (probe); } static void iflib_reset_qvalues(if_ctx_t ctx) { if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; device_t dev = ctx->ifc_dev; int i; if (ctx->ifc_sysctl_ntxqs != 0) scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs; if (ctx->ifc_sysctl_nrxqs != 0) scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs; for (i = 0; i < sctx->isc_ntxqs; i++) { if (ctx->ifc_sysctl_ntxds[i] != 0) scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i]; else scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i]; } for (i = 0; i < sctx->isc_nrxqs; i++) { if (ctx->ifc_sysctl_nrxds[i] != 0) scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i]; else scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i]; } for (i = 0; i < sctx->isc_nrxqs; i++) { if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) { device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n", i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]); scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i]; } if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) { device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n", i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]); scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i]; } if (!powerof2(scctx->isc_nrxd[i])) { device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n", i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]); scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i]; } } for (i = 0; i < sctx->isc_ntxqs; i++) { if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) { device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n", i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]); scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i]; } if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) { device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n", i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]); scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i]; } if (!powerof2(scctx->isc_ntxd[i])) { device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n", i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]); scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i]; } } } static void iflib_add_pfil(if_ctx_t ctx) { struct pfil_head *pfil; struct pfil_head_args pa; iflib_rxq_t rxq; int i; pa.pa_version = PFIL_VERSION; pa.pa_flags = PFIL_IN; pa.pa_type = PFIL_TYPE_ETHERNET; pa.pa_headname = ctx->ifc_ifp->if_xname; pfil = pfil_head_register(&pa); for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) { rxq->pfil = pfil; } } static void iflib_rem_pfil(if_ctx_t ctx) { struct pfil_head *pfil; iflib_rxq_t rxq; int i; rxq = ctx->ifc_rxqs; pfil = rxq->pfil; for (i = 0; i < NRXQSETS(ctx); i++, rxq++) { rxq->pfil = NULL; } pfil_head_unregister(pfil); } static uint16_t get_ctx_core_offset(if_ctx_t ctx) { if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; struct cpu_offset *op; uint16_t qc; uint16_t ret = ctx->ifc_sysctl_core_offset; if (ret != CORE_OFFSET_UNSPECIFIED) return (ret); if (ctx->ifc_sysctl_separate_txrx) qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets; else qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets); mtx_lock(&cpu_offset_mtx); SLIST_FOREACH(op, &cpu_offsets, entries) { if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) { ret = op->offset; op->offset += qc; MPASS(op->refcount < UINT_MAX); op->refcount++; break; } } if (ret == CORE_OFFSET_UNSPECIFIED) { ret = 0; op = malloc(sizeof(struct cpu_offset), M_IFLIB, M_NOWAIT | M_ZERO); if (op == NULL) { device_printf(ctx->ifc_dev, "allocation for cpu offset failed.\n"); } else { op->offset = qc; op->refcount = 1; CPU_COPY(&ctx->ifc_cpus, &op->set); SLIST_INSERT_HEAD(&cpu_offsets, op, entries); } } mtx_unlock(&cpu_offset_mtx); return (ret); } static void unref_ctx_core_offset(if_ctx_t ctx) { struct cpu_offset *op, *top; mtx_lock(&cpu_offset_mtx); SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) { if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) { MPASS(op->refcount > 0); op->refcount--; if (op->refcount == 0) { SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries); free(op, M_IFLIB); } break; } } mtx_unlock(&cpu_offset_mtx); } int iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp) { if_ctx_t ctx; if_t ifp; if_softc_ctx_t scctx; kobjop_desc_t kobj_desc; kobj_method_t *kobj_method; int err, msix, rid; uint16_t main_rxq, main_txq; ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO); if (sc == NULL) { sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO); device_set_softc(dev, ctx); ctx->ifc_flags |= IFC_SC_ALLOCATED; } ctx->ifc_sctx = sctx; ctx->ifc_dev = dev; ctx->ifc_softc = sc; if ((err = iflib_register(ctx)) != 0) { device_printf(dev, "iflib_register failed %d\n", err); goto fail_ctx_free; } iflib_add_device_sysctl_pre(ctx); scctx = &ctx->ifc_softc_ctx; ifp = ctx->ifc_ifp; iflib_reset_qvalues(ctx); CTX_LOCK(ctx); if ((err = IFDI_ATTACH_PRE(ctx)) != 0) { device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err); goto fail_unlock; } _iflib_pre_assert(scctx); ctx->ifc_txrx = *scctx->isc_txrx; if (sctx->isc_flags & IFLIB_DRIVER_MEDIA) ctx->ifc_mediap = scctx->isc_media; #ifdef INVARIANTS if (scctx->isc_capabilities & IFCAP_TXCSUM) MPASS(scctx->isc_tx_csum_flags); #endif - if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS); - if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS); + if_setcapabilities(ifp, + scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_NOMAP); + if_setcapenable(ifp, + scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_NOMAP); if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets)) scctx->isc_ntxqsets = scctx->isc_ntxqsets_max; if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets)) scctx->isc_nrxqsets = scctx->isc_nrxqsets_max; main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0; main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0; /* XXX change for per-queue sizes */ device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n", scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]); if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION) scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION); if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION) scctx->isc_tx_tso_segments_max = max(1, scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION); /* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */ if (if_getcapabilities(ifp) & IFCAP_TSO) { /* * The stack can't handle a TSO size larger than IP_MAXPACKET, * but some MACs do. */ if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max, IP_MAXPACKET)); /* * Take maximum number of m_pullup(9)'s in iflib_parse_header() * into account. In the worst case, each of these calls will * add another mbuf and, thus, the requirement for another DMA * segment. So for best performance, it doesn't make sense to * advertize a maximum of TSO segments that typically will * require defragmentation in iflib_encap(). */ if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3); if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max); } if (scctx->isc_rss_table_size == 0) scctx->isc_rss_table_size = 64; scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1; GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx); /* XXX format name */ taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, NULL, NULL, "admin"); /* Set up cpu set. If it fails, use the set of all CPUs. */ if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) { device_printf(dev, "Unable to fetch CPU list\n"); CPU_COPY(&all_cpus, &ctx->ifc_cpus); } MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0); /* ** Now set up MSI or MSI-X, should return us the number of supported ** vectors (will be 1 for a legacy interrupt and MSI). */ if (sctx->isc_flags & IFLIB_SKIP_MSIX) { msix = scctx->isc_vectors; } else if (scctx->isc_msix_bar != 0) /* * The simple fact that isc_msix_bar is not 0 does not mean we * we have a good value there that is known to work. */ msix = iflib_msix_init(ctx); else { scctx->isc_vectors = 1; scctx->isc_ntxqsets = 1; scctx->isc_nrxqsets = 1; scctx->isc_intr = IFLIB_INTR_LEGACY; msix = 0; } /* Get memory for the station queues */ if ((err = iflib_queues_alloc(ctx))) { device_printf(dev, "Unable to allocate queue memory\n"); goto fail_intr_free; } if ((err = iflib_qset_structures_setup(ctx))) goto fail_queues; /* * Now that we know how many queues there are, get the core offset. */ ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx); /* * Group taskqueues aren't properly set up until SMP is started, * so we disable interrupts until we can handle them post * SI_SUB_SMP. * * XXX: disabling interrupts doesn't actually work, at least for * the non-MSI case. When they occur before SI_SUB_SMP completes, * we do null handling and depend on this not causing too large an * interrupt storm. */ IFDI_INTR_DISABLE(ctx); if (msix > 1) { /* * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable * aren't the default NULL implementation. */ kobj_desc = &ifdi_rx_queue_intr_enable_desc; kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL, kobj_desc); if (kobj_method == &kobj_desc->deflt) { device_printf(dev, "MSI-X requires ifdi_rx_queue_intr_enable method"); err = EOPNOTSUPP; goto fail_queues; } kobj_desc = &ifdi_tx_queue_intr_enable_desc; kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL, kobj_desc); if (kobj_method == &kobj_desc->deflt) { device_printf(dev, "MSI-X requires ifdi_tx_queue_intr_enable method"); err = EOPNOTSUPP; goto fail_queues; } /* * Assign the MSI-X vectors. * Note that the default NULL ifdi_msix_intr_assign method will * fail here, too. */ err = IFDI_MSIX_INTR_ASSIGN(ctx, msix); if (err != 0) { device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err); goto fail_queues; } } else if (scctx->isc_intr != IFLIB_INTR_MSIX) { rid = 0; if (scctx->isc_intr == IFLIB_INTR_MSI) { MPASS(msix == 1); rid = 1; } if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) { device_printf(dev, "iflib_legacy_setup failed %d\n", err); goto fail_queues; } } else { device_printf(dev, "Cannot use iflib with only 1 MSI-X interrupt!\n"); err = ENODEV; goto fail_intr_free; } ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet); if ((err = IFDI_ATTACH_POST(ctx)) != 0) { device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err); goto fail_detach; } /* * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported. * This must appear after the call to ether_ifattach() because * ether_ifattach() sets if_hdrlen to the default value. */ if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU) if_setifheaderlen(ifp, sizeof(struct ether_vlan_header)); if ((err = iflib_netmap_attach(ctx))) { device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err); goto fail_detach; } *ctxp = ctx; NETDUMP_SET(ctx->ifc_ifp, iflib); if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter); iflib_add_device_sysctl_post(ctx); iflib_add_pfil(ctx); ctx->ifc_flags |= IFC_INIT_DONE; CTX_UNLOCK(ctx); return (0); fail_detach: ether_ifdetach(ctx->ifc_ifp); fail_intr_free: iflib_free_intr_mem(ctx); fail_queues: iflib_tx_structures_free(ctx); iflib_rx_structures_free(ctx); taskqgroup_detach(qgroup_if_config_tqg, &ctx->ifc_admin_task); IFDI_DETACH(ctx); fail_unlock: CTX_UNLOCK(ctx); iflib_deregister(ctx); fail_ctx_free: device_set_softc(ctx->ifc_dev, NULL); if (ctx->ifc_flags & IFC_SC_ALLOCATED) free(ctx->ifc_softc, M_IFLIB); free(ctx, M_IFLIB); return (err); } int iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp, struct iflib_cloneattach_ctx *clctx) { int err; if_ctx_t ctx; if_t ifp; if_softc_ctx_t scctx; int i; void *sc; uint16_t main_txq; uint16_t main_rxq; ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO); sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO); ctx->ifc_flags |= IFC_SC_ALLOCATED; if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL)) ctx->ifc_flags |= IFC_PSEUDO; ctx->ifc_sctx = sctx; ctx->ifc_softc = sc; ctx->ifc_dev = dev; if ((err = iflib_register(ctx)) != 0) { device_printf(dev, "%s: iflib_register failed %d\n", __func__, err); goto fail_ctx_free; } iflib_add_device_sysctl_pre(ctx); scctx = &ctx->ifc_softc_ctx; ifp = ctx->ifc_ifp; iflib_reset_qvalues(ctx); CTX_LOCK(ctx); if ((err = IFDI_ATTACH_PRE(ctx)) != 0) { device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err); goto fail_unlock; } if (sctx->isc_flags & IFLIB_GEN_MAC) ether_gen_addr(ifp, &ctx->ifc_mac); if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name, clctx->cc_params)) != 0) { device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err); goto fail_ctx_free; } ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL); ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL); ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO); #ifdef INVARIANTS if (scctx->isc_capabilities & IFCAP_TXCSUM) MPASS(scctx->isc_tx_csum_flags); #endif if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE); if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE); ifp->if_flags |= IFF_NOGROUP; if (sctx->isc_flags & IFLIB_PSEUDO) { ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet); if ((err = IFDI_ATTACH_POST(ctx)) != 0) { device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err); goto fail_detach; } *ctxp = ctx; /* * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported. * This must appear after the call to ether_ifattach() because * ether_ifattach() sets if_hdrlen to the default value. */ if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU) if_setifheaderlen(ifp, sizeof(struct ether_vlan_header)); if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter); iflib_add_device_sysctl_post(ctx); ctx->ifc_flags |= IFC_INIT_DONE; return (0); } _iflib_pre_assert(scctx); ctx->ifc_txrx = *scctx->isc_txrx; if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets)) scctx->isc_ntxqsets = scctx->isc_ntxqsets_max; if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets)) scctx->isc_nrxqsets = scctx->isc_nrxqsets_max; main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0; main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0; /* XXX change for per-queue sizes */ device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n", scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]); if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION) scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION); if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION) scctx->isc_tx_tso_segments_max = max(1, scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION); /* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */ if (if_getcapabilities(ifp) & IFCAP_TSO) { /* * The stack can't handle a TSO size larger than IP_MAXPACKET, * but some MACs do. */ if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max, IP_MAXPACKET)); /* * Take maximum number of m_pullup(9)'s in iflib_parse_header() * into account. In the worst case, each of these calls will * add another mbuf and, thus, the requirement for another DMA * segment. So for best performance, it doesn't make sense to * advertize a maximum of TSO segments that typically will * require defragmentation in iflib_encap(). */ if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3); if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max); } if (scctx->isc_rss_table_size == 0) scctx->isc_rss_table_size = 64; scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1; GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx); /* XXX format name */ taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, NULL, NULL, "admin"); /* XXX --- can support > 1 -- but keep it simple for now */ scctx->isc_intr = IFLIB_INTR_LEGACY; /* Get memory for the station queues */ if ((err = iflib_queues_alloc(ctx))) { device_printf(dev, "Unable to allocate queue memory\n"); goto fail_iflib_detach; } if ((err = iflib_qset_structures_setup(ctx))) { device_printf(dev, "qset structure setup failed %d\n", err); goto fail_queues; } /* * XXX What if anything do we want to do about interrupts? */ ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet); if ((err = IFDI_ATTACH_POST(ctx)) != 0) { device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err); goto fail_detach; } /* * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported. * This must appear after the call to ether_ifattach() because * ether_ifattach() sets if_hdrlen to the default value. */ if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU) if_setifheaderlen(ifp, sizeof(struct ether_vlan_header)); /* XXX handle more than one queue */ for (i = 0; i < scctx->isc_nrxqsets; i++) IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl); *ctxp = ctx; if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter); iflib_add_device_sysctl_post(ctx); ctx->ifc_flags |= IFC_INIT_DONE; CTX_UNLOCK(ctx); return (0); fail_detach: ether_ifdetach(ctx->ifc_ifp); fail_queues: iflib_tx_structures_free(ctx); iflib_rx_structures_free(ctx); fail_iflib_detach: IFDI_DETACH(ctx); fail_unlock: CTX_UNLOCK(ctx); iflib_deregister(ctx); fail_ctx_free: free(ctx->ifc_softc, M_IFLIB); free(ctx, M_IFLIB); return (err); } int iflib_pseudo_deregister(if_ctx_t ctx) { if_t ifp = ctx->ifc_ifp; iflib_txq_t txq; iflib_rxq_t rxq; int i, j; struct taskqgroup *tqg; iflib_fl_t fl; ether_ifdetach(ifp); /* XXX drain any dependent tasks */ tqg = qgroup_if_io_tqg; for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) { callout_drain(&txq->ift_timer); if (txq->ift_task.gt_uniq != NULL) taskqgroup_detach(tqg, &txq->ift_task); } for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) { if (rxq->ifr_task.gt_uniq != NULL) taskqgroup_detach(tqg, &rxq->ifr_task); for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) free(fl->ifl_rx_bitmap, M_IFLIB); } tqg = qgroup_if_config_tqg; if (ctx->ifc_admin_task.gt_uniq != NULL) taskqgroup_detach(tqg, &ctx->ifc_admin_task); if (ctx->ifc_vflr_task.gt_uniq != NULL) taskqgroup_detach(tqg, &ctx->ifc_vflr_task); iflib_tx_structures_free(ctx); iflib_rx_structures_free(ctx); iflib_deregister(ctx); if (ctx->ifc_flags & IFC_SC_ALLOCATED) free(ctx->ifc_softc, M_IFLIB); free(ctx, M_IFLIB); return (0); } int iflib_device_attach(device_t dev) { if_ctx_t ctx; if_shared_ctx_t sctx; if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC) return (ENOTSUP); pci_enable_busmaster(dev); return (iflib_device_register(dev, NULL, sctx, &ctx)); } int iflib_device_deregister(if_ctx_t ctx) { if_t ifp = ctx->ifc_ifp; iflib_txq_t txq; iflib_rxq_t rxq; device_t dev = ctx->ifc_dev; int i, j; struct taskqgroup *tqg; iflib_fl_t fl; /* Make sure VLANS are not using driver */ if (if_vlantrunkinuse(ifp)) { device_printf(dev, "Vlan in use, detach first\n"); return (EBUSY); } #ifdef PCI_IOV if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) { device_printf(dev, "SR-IOV in use; detach first.\n"); return (EBUSY); } #endif STATE_LOCK(ctx); ctx->ifc_flags |= IFC_IN_DETACH; STATE_UNLOCK(ctx); CTX_LOCK(ctx); iflib_stop(ctx); CTX_UNLOCK(ctx); iflib_netmap_detach(ifp); ether_ifdetach(ifp); iflib_rem_pfil(ctx); if (ctx->ifc_led_dev != NULL) led_destroy(ctx->ifc_led_dev); /* XXX drain any dependent tasks */ tqg = qgroup_if_io_tqg; for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) { callout_drain(&txq->ift_timer); if (txq->ift_task.gt_uniq != NULL) taskqgroup_detach(tqg, &txq->ift_task); } for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) { if (rxq->ifr_task.gt_uniq != NULL) taskqgroup_detach(tqg, &rxq->ifr_task); for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) free(fl->ifl_rx_bitmap, M_IFLIB); } tqg = qgroup_if_config_tqg; if (ctx->ifc_admin_task.gt_uniq != NULL) taskqgroup_detach(tqg, &ctx->ifc_admin_task); if (ctx->ifc_vflr_task.gt_uniq != NULL) taskqgroup_detach(tqg, &ctx->ifc_vflr_task); CTX_LOCK(ctx); IFDI_DETACH(ctx); CTX_UNLOCK(ctx); /* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/ iflib_free_intr_mem(ctx); bus_generic_detach(dev); iflib_tx_structures_free(ctx); iflib_rx_structures_free(ctx); iflib_deregister(ctx); device_set_softc(ctx->ifc_dev, NULL); if (ctx->ifc_flags & IFC_SC_ALLOCATED) free(ctx->ifc_softc, M_IFLIB); unref_ctx_core_offset(ctx); free(ctx, M_IFLIB); return (0); } static void iflib_free_intr_mem(if_ctx_t ctx) { if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) { iflib_irq_free(ctx, &ctx->ifc_legacy_irq); } if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) { pci_release_msi(ctx->ifc_dev); } if (ctx->ifc_msix_mem != NULL) { bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY, rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem); ctx->ifc_msix_mem = NULL; } } int iflib_device_detach(device_t dev) { if_ctx_t ctx = device_get_softc(dev); return (iflib_device_deregister(ctx)); } int iflib_device_suspend(device_t dev) { if_ctx_t ctx = device_get_softc(dev); CTX_LOCK(ctx); IFDI_SUSPEND(ctx); CTX_UNLOCK(ctx); return bus_generic_suspend(dev); } int iflib_device_shutdown(device_t dev) { if_ctx_t ctx = device_get_softc(dev); CTX_LOCK(ctx); IFDI_SHUTDOWN(ctx); CTX_UNLOCK(ctx); return bus_generic_suspend(dev); } int iflib_device_resume(device_t dev) { if_ctx_t ctx = device_get_softc(dev); iflib_txq_t txq = ctx->ifc_txqs; CTX_LOCK(ctx); IFDI_RESUME(ctx); iflib_if_init_locked(ctx); CTX_UNLOCK(ctx); for (int i = 0; i < NTXQSETS(ctx); i++, txq++) iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET); return (bus_generic_resume(dev)); } int iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params) { int error; if_ctx_t ctx = device_get_softc(dev); CTX_LOCK(ctx); error = IFDI_IOV_INIT(ctx, num_vfs, params); CTX_UNLOCK(ctx); return (error); } void iflib_device_iov_uninit(device_t dev) { if_ctx_t ctx = device_get_softc(dev); CTX_LOCK(ctx); IFDI_IOV_UNINIT(ctx); CTX_UNLOCK(ctx); } int iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params) { int error; if_ctx_t ctx = device_get_softc(dev); CTX_LOCK(ctx); error = IFDI_IOV_VF_ADD(ctx, vfnum, params); CTX_UNLOCK(ctx); return (error); } /********************************************************************* * * MODULE FUNCTION DEFINITIONS * **********************************************************************/ /* * - Start a fast taskqueue thread for each core * - Start a taskqueue for control operations */ static int iflib_module_init(void) { return (0); } static int iflib_module_event_handler(module_t mod, int what, void *arg) { int err; switch (what) { case MOD_LOAD: if ((err = iflib_module_init()) != 0) return (err); break; case MOD_UNLOAD: return (EBUSY); default: return (EOPNOTSUPP); } return (0); } /********************************************************************* * * PUBLIC FUNCTION DEFINITIONS * ordered as in iflib.h * **********************************************************************/ static void _iflib_assert(if_shared_ctx_t sctx) { int i; MPASS(sctx->isc_tx_maxsize); MPASS(sctx->isc_tx_maxsegsize); MPASS(sctx->isc_rx_maxsize); MPASS(sctx->isc_rx_nsegments); MPASS(sctx->isc_rx_maxsegsize); MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8); for (i = 0; i < sctx->isc_nrxqs; i++) { MPASS(sctx->isc_nrxd_min[i]); MPASS(powerof2(sctx->isc_nrxd_min[i])); MPASS(sctx->isc_nrxd_max[i]); MPASS(powerof2(sctx->isc_nrxd_max[i])); MPASS(sctx->isc_nrxd_default[i]); MPASS(powerof2(sctx->isc_nrxd_default[i])); } MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8); for (i = 0; i < sctx->isc_ntxqs; i++) { MPASS(sctx->isc_ntxd_min[i]); MPASS(powerof2(sctx->isc_ntxd_min[i])); MPASS(sctx->isc_ntxd_max[i]); MPASS(powerof2(sctx->isc_ntxd_max[i])); MPASS(sctx->isc_ntxd_default[i]); MPASS(powerof2(sctx->isc_ntxd_default[i])); } } static void _iflib_pre_assert(if_softc_ctx_t scctx) { MPASS(scctx->isc_txrx->ift_txd_encap); MPASS(scctx->isc_txrx->ift_txd_flush); MPASS(scctx->isc_txrx->ift_txd_credits_update); MPASS(scctx->isc_txrx->ift_rxd_available); MPASS(scctx->isc_txrx->ift_rxd_pkt_get); MPASS(scctx->isc_txrx->ift_rxd_refill); MPASS(scctx->isc_txrx->ift_rxd_flush); } static int iflib_register(if_ctx_t ctx) { if_shared_ctx_t sctx = ctx->ifc_sctx; driver_t *driver = sctx->isc_driver; device_t dev = ctx->ifc_dev; if_t ifp; _iflib_assert(sctx); CTX_LOCK_INIT(ctx); STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev)); ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER); if (ifp == NULL) { device_printf(dev, "can not allocate ifnet structure\n"); return (ENOMEM); } /* * Initialize our context's device specific methods */ kobj_init((kobj_t) ctx, (kobj_class_t) driver); kobj_class_compile((kobj_class_t) driver); if_initname(ifp, device_get_name(dev), device_get_unit(dev)); if_setsoftc(ifp, ctx); if_setdev(ifp, dev); if_setinitfn(ifp, iflib_if_init); if_setioctlfn(ifp, iflib_if_ioctl); #ifdef ALTQ if_setstartfn(ifp, iflib_altq_if_start); if_settransmitfn(ifp, iflib_altq_if_transmit); if_setsendqready(ifp); #else if_settransmitfn(ifp, iflib_if_transmit); #endif if_setqflushfn(ifp, iflib_if_qflush); if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST); ctx->ifc_vlan_attach_event = EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx, EVENTHANDLER_PRI_FIRST); ctx->ifc_vlan_detach_event = EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx, EVENTHANDLER_PRI_FIRST); if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) { ctx->ifc_mediap = &ctx->ifc_media; ifmedia_init(ctx->ifc_mediap, IFM_IMASK, iflib_media_change, iflib_media_status); } return (0); } static void iflib_deregister(if_ctx_t ctx) { if_t ifp = ctx->ifc_ifp; /* Remove all media */ ifmedia_removeall(&ctx->ifc_media); /* Unregister VLAN events */ if (ctx->ifc_vlan_attach_event != NULL) { EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event); ctx->ifc_vlan_attach_event = NULL; } if (ctx->ifc_vlan_detach_event != NULL) { EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event); ctx->ifc_vlan_detach_event = NULL; } /* Release kobject reference */ kobj_delete((kobj_t) ctx, NULL); /* Free the ifnet structure */ if_free(ifp); STATE_LOCK_DESTROY(ctx); /* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/ CTX_LOCK_DESTROY(ctx); } static int iflib_queues_alloc(if_ctx_t ctx) { if_shared_ctx_t sctx = ctx->ifc_sctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; device_t dev = ctx->ifc_dev; int nrxqsets = scctx->isc_nrxqsets; int ntxqsets = scctx->isc_ntxqsets; iflib_txq_t txq; iflib_rxq_t rxq; iflib_fl_t fl = NULL; int i, j, cpu, err, txconf, rxconf; iflib_dma_info_t ifdip; uint32_t *rxqsizes = scctx->isc_rxqsizes; uint32_t *txqsizes = scctx->isc_txqsizes; uint8_t nrxqs = sctx->isc_nrxqs; uint8_t ntxqs = sctx->isc_ntxqs; int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1; caddr_t *vaddrs; uint64_t *paddrs; KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1")); KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1")); /* Allocate the TX ring struct memory */ if (!(ctx->ifc_txqs = (iflib_txq_t) malloc(sizeof(struct iflib_txq) * ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate TX ring memory\n"); err = ENOMEM; goto fail; } /* Now allocate the RX */ if (!(ctx->ifc_rxqs = (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) * nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate RX ring memory\n"); err = ENOMEM; goto rx_fail; } txq = ctx->ifc_txqs; rxq = ctx->ifc_rxqs; /* * XXX handle allocation failure */ for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) { /* Set up some basics */ if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) { device_printf(dev, "Unable to allocate TX DMA info memory\n"); err = ENOMEM; goto err_tx_desc; } txq->ift_ifdi = ifdip; for (j = 0; j < ntxqs; j++, ifdip++) { if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) { device_printf(dev, "Unable to allocate TX descriptors\n"); err = ENOMEM; goto err_tx_desc; } txq->ift_txd_size[j] = scctx->isc_txd_size[j]; bzero((void *)ifdip->idi_vaddr, txqsizes[j]); } txq->ift_ctx = ctx; txq->ift_id = i; if (sctx->isc_flags & IFLIB_HAS_TXCQ) { txq->ift_br_offset = 1; } else { txq->ift_br_offset = 0; } /* XXX fix this */ txq->ift_timer.c_cpu = cpu; if (iflib_txsd_alloc(txq)) { device_printf(dev, "Critical Failure setting up TX buffers\n"); err = ENOMEM; goto err_tx_desc; } /* Initialize the TX lock */ snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout", device_get_nameunit(dev), txq->ift_id); mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF); callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0); err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain, iflib_txq_can_drain, M_IFLIB, M_WAITOK); if (err) { /* XXX free any allocated rings */ device_printf(dev, "Unable to allocate buf_ring\n"); goto err_tx_desc; } } for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) { /* Set up some basics */ if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) { device_printf(dev, "Unable to allocate RX DMA info memory\n"); err = ENOMEM; goto err_tx_desc; } rxq->ifr_ifdi = ifdip; /* XXX this needs to be changed if #rx queues != #tx queues */ rxq->ifr_ntxqirq = 1; rxq->ifr_txqid[0] = i; for (j = 0; j < nrxqs; j++, ifdip++) { if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) { device_printf(dev, "Unable to allocate RX descriptors\n"); err = ENOMEM; goto err_tx_desc; } bzero((void *)ifdip->idi_vaddr, rxqsizes[j]); } rxq->ifr_ctx = ctx; rxq->ifr_id = i; if (sctx->isc_flags & IFLIB_HAS_RXCQ) { rxq->ifr_fl_offset = 1; } else { rxq->ifr_fl_offset = 0; } rxq->ifr_nfl = nfree_lists; if (!(fl = (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate free list memory\n"); err = ENOMEM; goto err_tx_desc; } rxq->ifr_fl = fl; for (j = 0; j < nfree_lists; j++) { fl[j].ifl_rxq = rxq; fl[j].ifl_id = j; fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset]; fl[j].ifl_rxd_size = scctx->isc_rxd_size[j]; } /* Allocate receive buffers for the ring */ if (iflib_rxsd_alloc(rxq)) { device_printf(dev, "Critical Failure setting up receive buffers\n"); err = ENOMEM; goto err_rx_desc; } for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK); } /* TXQs */ vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK); paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK); for (i = 0; i < ntxqsets; i++) { iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi; for (j = 0; j < ntxqs; j++, di++) { vaddrs[i*ntxqs + j] = di->idi_vaddr; paddrs[i*ntxqs + j] = di->idi_paddr; } } if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) { device_printf(ctx->ifc_dev, "Unable to allocate device TX queue\n"); iflib_tx_structures_free(ctx); free(vaddrs, M_IFLIB); free(paddrs, M_IFLIB); goto err_rx_desc; } free(vaddrs, M_IFLIB); free(paddrs, M_IFLIB); /* RXQs */ vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK); paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK); for (i = 0; i < nrxqsets; i++) { iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi; for (j = 0; j < nrxqs; j++, di++) { vaddrs[i*nrxqs + j] = di->idi_vaddr; paddrs[i*nrxqs + j] = di->idi_paddr; } } if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) { device_printf(ctx->ifc_dev, "Unable to allocate device RX queue\n"); iflib_tx_structures_free(ctx); free(vaddrs, M_IFLIB); free(paddrs, M_IFLIB); goto err_rx_desc; } free(vaddrs, M_IFLIB); free(paddrs, M_IFLIB); return (0); /* XXX handle allocation failure changes */ err_rx_desc: err_tx_desc: rx_fail: if (ctx->ifc_rxqs != NULL) free(ctx->ifc_rxqs, M_IFLIB); ctx->ifc_rxqs = NULL; if (ctx->ifc_txqs != NULL) free(ctx->ifc_txqs, M_IFLIB); ctx->ifc_txqs = NULL; fail: return (err); } static int iflib_tx_structures_setup(if_ctx_t ctx) { iflib_txq_t txq = ctx->ifc_txqs; int i; for (i = 0; i < NTXQSETS(ctx); i++, txq++) iflib_txq_setup(txq); return (0); } static void iflib_tx_structures_free(if_ctx_t ctx) { iflib_txq_t txq = ctx->ifc_txqs; if_shared_ctx_t sctx = ctx->ifc_sctx; int i, j; for (i = 0; i < NTXQSETS(ctx); i++, txq++) { iflib_txq_destroy(txq); for (j = 0; j < sctx->isc_ntxqs; j++) iflib_dma_free(&txq->ift_ifdi[j]); } free(ctx->ifc_txqs, M_IFLIB); ctx->ifc_txqs = NULL; IFDI_QUEUES_FREE(ctx); } /********************************************************************* * * Initialize all receive rings. * **********************************************************************/ static int iflib_rx_structures_setup(if_ctx_t ctx) { iflib_rxq_t rxq = ctx->ifc_rxqs; int q; #if defined(INET6) || defined(INET) int err, i; #endif for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) { #if defined(INET6) || defined(INET) if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) { err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp, TCP_LRO_ENTRIES, min(1024, ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset])); if (err != 0) { device_printf(ctx->ifc_dev, "LRO Initialization failed!\n"); goto fail; } } #endif IFDI_RXQ_SETUP(ctx, rxq->ifr_id); } return (0); #if defined(INET6) || defined(INET) fail: /* * Free LRO resources allocated so far, we will only handle * the rings that completed, the failing case will have * cleaned up for itself. 'q' failed, so its the terminus. */ rxq = ctx->ifc_rxqs; for (i = 0; i < q; ++i, rxq++) { if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) tcp_lro_free(&rxq->ifr_lc); } return (err); #endif } /********************************************************************* * * Free all receive rings. * **********************************************************************/ static void iflib_rx_structures_free(if_ctx_t ctx) { iflib_rxq_t rxq = ctx->ifc_rxqs; int i; for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) { iflib_rx_sds_free(rxq); #if defined(INET6) || defined(INET) if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) tcp_lro_free(&rxq->ifr_lc); #endif } free(ctx->ifc_rxqs, M_IFLIB); ctx->ifc_rxqs = NULL; } static int iflib_qset_structures_setup(if_ctx_t ctx) { int err; /* * It is expected that the caller takes care of freeing queues if this * fails. */ if ((err = iflib_tx_structures_setup(ctx)) != 0) { device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err); return (err); } if ((err = iflib_rx_structures_setup(ctx)) != 0) device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err); return (err); } int iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid, driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name) { return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name)); } #ifdef SMP static int find_nth(if_ctx_t ctx, int qid) { cpuset_t cpus; int i, cpuid, eqid, count; CPU_COPY(&ctx->ifc_cpus, &cpus); count = CPU_COUNT(&cpus); eqid = qid % count; /* clear up to the qid'th bit */ for (i = 0; i < eqid; i++) { cpuid = CPU_FFS(&cpus); MPASS(cpuid != 0); CPU_CLR(cpuid-1, &cpus); } cpuid = CPU_FFS(&cpus); MPASS(cpuid != 0); return (cpuid-1); } #ifdef SCHED_ULE extern struct cpu_group *cpu_top; /* CPU topology */ static int find_child_with_core(int cpu, struct cpu_group *grp) { int i; if (grp->cg_children == 0) return -1; MPASS(grp->cg_child); for (i = 0; i < grp->cg_children; i++) { if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask)) return i; } return -1; } /* * Find the nth "close" core to the specified core * "close" is defined as the deepest level that shares * at least an L2 cache. With threads, this will be * threads on the same core. If the shared cache is L3 * or higher, simply returns the same core. */ static int find_close_core(int cpu, int core_offset) { struct cpu_group *grp; int i; int fcpu; cpuset_t cs; grp = cpu_top; if (grp == NULL) return cpu; i = 0; while ((i = find_child_with_core(cpu, grp)) != -1) { /* If the child only has one cpu, don't descend */ if (grp->cg_child[i].cg_count <= 1) break; grp = &grp->cg_child[i]; } /* If they don't share at least an L2 cache, use the same CPU */ if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE) return cpu; /* Now pick one */ CPU_COPY(&grp->cg_mask, &cs); /* Add the selected CPU offset to core offset. */ for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) { if (fcpu - 1 == cpu) break; CPU_CLR(fcpu - 1, &cs); } MPASS(fcpu); core_offset += i; CPU_COPY(&grp->cg_mask, &cs); for (i = core_offset % grp->cg_count; i > 0; i--) { MPASS(CPU_FFS(&cs)); CPU_CLR(CPU_FFS(&cs) - 1, &cs); } MPASS(CPU_FFS(&cs)); return CPU_FFS(&cs) - 1; } #else static int find_close_core(int cpu, int core_offset __unused) { return cpu; } #endif static int get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid) { switch (type) { case IFLIB_INTR_TX: /* TX queues get cores which share at least an L2 cache with the corresponding RX queue */ /* XXX handle multiple RX threads per core and more than two core per L2 group */ return qid / CPU_COUNT(&ctx->ifc_cpus) + 1; case IFLIB_INTR_RX: case IFLIB_INTR_RXTX: /* RX queues get the specified core */ return qid / CPU_COUNT(&ctx->ifc_cpus); default: return -1; } } #else #define get_core_offset(ctx, type, qid) CPU_FIRST() #define find_close_core(cpuid, tid) CPU_FIRST() #define find_nth(ctx, gid) CPU_FIRST() #endif /* Just to avoid copy/paste */ static inline int iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, const char *name) { device_t dev; int co, cpuid, err, tid; dev = ctx->ifc_dev; co = ctx->ifc_sysctl_core_offset; if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX) co += ctx->ifc_softc_ctx.isc_nrxqsets; cpuid = find_nth(ctx, qid + co); tid = get_core_offset(ctx, type, qid); if (tid < 0) { device_printf(dev, "get_core_offset failed\n"); return (EOPNOTSUPP); } cpuid = find_close_core(cpuid, tid); err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res, name); if (err) { device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err); return (err); } #ifdef notyet if (cpuid > ctx->ifc_cpuid_highest) ctx->ifc_cpuid_highest = cpuid; #endif return (0); } int iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid, iflib_intr_type_t type, driver_filter_t *filter, void *filter_arg, int qid, const char *name) { device_t dev; struct grouptask *gtask; struct taskqgroup *tqg; iflib_filter_info_t info; gtask_fn_t *fn; int tqrid, err; driver_filter_t *intr_fast; void *q; info = &ctx->ifc_filter_info; tqrid = rid; switch (type) { /* XXX merge tx/rx for netmap? */ case IFLIB_INTR_TX: q = &ctx->ifc_txqs[qid]; info = &ctx->ifc_txqs[qid].ift_filter_info; gtask = &ctx->ifc_txqs[qid].ift_task; tqg = qgroup_if_io_tqg; fn = _task_fn_tx; intr_fast = iflib_fast_intr; GROUPTASK_INIT(gtask, 0, fn, q); ctx->ifc_flags |= IFC_NETMAP_TX_IRQ; break; case IFLIB_INTR_RX: q = &ctx->ifc_rxqs[qid]; info = &ctx->ifc_rxqs[qid].ifr_filter_info; gtask = &ctx->ifc_rxqs[qid].ifr_task; tqg = qgroup_if_io_tqg; fn = _task_fn_rx; intr_fast = iflib_fast_intr; GROUPTASK_INIT(gtask, 0, fn, q); break; case IFLIB_INTR_RXTX: q = &ctx->ifc_rxqs[qid]; info = &ctx->ifc_rxqs[qid].ifr_filter_info; gtask = &ctx->ifc_rxqs[qid].ifr_task; tqg = qgroup_if_io_tqg; fn = _task_fn_rx; intr_fast = iflib_fast_intr_rxtx; GROUPTASK_INIT(gtask, 0, fn, q); break; case IFLIB_INTR_ADMIN: q = ctx; tqrid = -1; info = &ctx->ifc_filter_info; gtask = &ctx->ifc_admin_task; tqg = qgroup_if_config_tqg; fn = _task_fn_admin; intr_fast = iflib_fast_intr_ctx; break; default: device_printf(ctx->ifc_dev, "%s: unknown net intr type\n", __func__); return (EINVAL); } info->ifi_filter = filter; info->ifi_filter_arg = filter_arg; info->ifi_task = gtask; info->ifi_ctx = q; dev = ctx->ifc_dev; err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info, name); if (err != 0) { device_printf(dev, "_iflib_irq_alloc failed %d\n", err); return (err); } if (type == IFLIB_INTR_ADMIN) return (0); if (tqrid != -1) { err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name); if (err) return (err); } else { taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name); } return (0); } void iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name) { struct grouptask *gtask; struct taskqgroup *tqg; gtask_fn_t *fn; void *q; int err; switch (type) { case IFLIB_INTR_TX: q = &ctx->ifc_txqs[qid]; gtask = &ctx->ifc_txqs[qid].ift_task; tqg = qgroup_if_io_tqg; fn = _task_fn_tx; break; case IFLIB_INTR_RX: q = &ctx->ifc_rxqs[qid]; gtask = &ctx->ifc_rxqs[qid].ifr_task; tqg = qgroup_if_io_tqg; fn = _task_fn_rx; break; case IFLIB_INTR_IOV: q = ctx; gtask = &ctx->ifc_vflr_task; tqg = qgroup_if_config_tqg; fn = _task_fn_iov; break; default: panic("unknown net intr type"); } GROUPTASK_INIT(gtask, 0, fn, q); if (irq != NULL) { err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name); if (err) taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev, irq->ii_res, name); } else { taskqgroup_attach(tqg, gtask, q, NULL, NULL, name); } } void iflib_irq_free(if_ctx_t ctx, if_irq_t irq) { if (irq->ii_tag) bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag); if (irq->ii_res) bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, rman_get_rid(irq->ii_res), irq->ii_res); } static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name) { iflib_txq_t txq = ctx->ifc_txqs; iflib_rxq_t rxq = ctx->ifc_rxqs; if_irq_t irq = &ctx->ifc_legacy_irq; iflib_filter_info_t info; device_t dev; struct grouptask *gtask; struct resource *res; struct taskqgroup *tqg; gtask_fn_t *fn; void *q; int err, tqrid; q = &ctx->ifc_rxqs[0]; info = &rxq[0].ifr_filter_info; gtask = &rxq[0].ifr_task; tqg = qgroup_if_io_tqg; tqrid = *rid; fn = _task_fn_rx; ctx->ifc_flags |= IFC_LEGACY; info->ifi_filter = filter; info->ifi_filter_arg = filter_arg; info->ifi_task = gtask; info->ifi_ctx = q; dev = ctx->ifc_dev; /* We allocate a single interrupt resource */ if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_rxtx, NULL, info, name)) != 0) return (err); GROUPTASK_INIT(gtask, 0, fn, q); res = irq->ii_res; taskqgroup_attach(tqg, gtask, q, dev, res, name); GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq); taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res, "tx"); return (0); } void iflib_led_create(if_ctx_t ctx) { ctx->ifc_led_dev = led_create(iflib_led_func, ctx, device_get_nameunit(ctx->ifc_dev)); } void iflib_tx_intr_deferred(if_ctx_t ctx, int txqid) { GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task); } void iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid) { GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task); } void iflib_admin_intr_deferred(if_ctx_t ctx) { #ifdef INVARIANTS struct grouptask *gtask; gtask = &ctx->ifc_admin_task; MPASS(gtask != NULL && gtask->gt_taskqueue != NULL); #endif GROUPTASK_ENQUEUE(&ctx->ifc_admin_task); } void iflib_iov_intr_deferred(if_ctx_t ctx) { GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task); } void iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name) { taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL, name); } void iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn, const char *name) { GROUPTASK_INIT(gtask, 0, fn, ctx); taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL, name); } void iflib_config_gtask_deinit(struct grouptask *gtask) { taskqgroup_detach(qgroup_if_config_tqg, gtask); } void iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate) { if_t ifp = ctx->ifc_ifp; iflib_txq_t txq = ctx->ifc_txqs; if_setbaudrate(ifp, baudrate); if (baudrate >= IF_Gbps(10)) { STATE_LOCK(ctx); ctx->ifc_flags |= IFC_PREFETCH; STATE_UNLOCK(ctx); } /* If link down, disable watchdog */ if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) { for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++) txq->ift_qstatus = IFLIB_QUEUE_IDLE; } ctx->ifc_link_state = link_state; if_link_state_change(ifp, link_state); } static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq) { int credits; #ifdef INVARIANTS int credits_pre = txq->ift_cidx_processed; #endif bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_POSTREAD); if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0) return (0); txq->ift_processed += credits; txq->ift_cidx_processed += credits; MPASS(credits_pre + credits == txq->ift_cidx_processed); if (txq->ift_cidx_processed >= txq->ift_size) txq->ift_cidx_processed -= txq->ift_size; return (credits); } static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget) { iflib_fl_t fl; u_int i; for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++) bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx, budget)); } void iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name, const char *description, if_int_delay_info_t info, int offset, int value) { info->iidi_ctx = ctx; info->iidi_offset = offset; info->iidi_value = value; SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev), SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)), OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW, info, 0, iflib_sysctl_int_delay, "I", description); } struct sx * iflib_ctx_lock_get(if_ctx_t ctx) { return (&ctx->ifc_ctx_sx); } static int iflib_msix_init(if_ctx_t ctx) { device_t dev = ctx->ifc_dev; if_shared_ctx_t sctx = ctx->ifc_sctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues; int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors; iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs; iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs; if (bootverbose) device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets)); /* Override by tuneable */ if (scctx->isc_disable_msix) goto msi; /* First try MSI-X */ if ((msgs = pci_msix_count(dev)) == 0) { if (bootverbose) device_printf(dev, "MSI-X not supported or disabled\n"); goto msi; } bar = ctx->ifc_softc_ctx.isc_msix_bar; /* * bar == -1 => "trust me I know what I'm doing" * Some drivers are for hardware that is so shoddily * documented that no one knows which bars are which * so the developer has to map all bars. This hack * allows shoddy garbage to use MSI-X in this framework. */ if (bar != -1) { ctx->ifc_msix_mem = bus_alloc_resource_any(dev, SYS_RES_MEMORY, &bar, RF_ACTIVE); if (ctx->ifc_msix_mem == NULL) { device_printf(dev, "Unable to map MSI-X table\n"); goto msi; } } admincnt = sctx->isc_admin_intrcnt; #if IFLIB_DEBUG /* use only 1 qset in debug mode */ queuemsgs = min(msgs - admincnt, 1); #else queuemsgs = msgs - admincnt; #endif #ifdef RSS queues = imin(queuemsgs, rss_getnumbuckets()); #else queues = queuemsgs; #endif queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues); if (bootverbose) device_printf(dev, "intr CPUs: %d queue msgs: %d admincnt: %d\n", CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt); #ifdef RSS /* If we're doing RSS, clamp at the number of RSS buckets */ if (queues > rss_getnumbuckets()) queues = rss_getnumbuckets(); #endif if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt) rx_queues = iflib_num_rx_queues; else rx_queues = queues; if (rx_queues > scctx->isc_nrxqsets) rx_queues = scctx->isc_nrxqsets; /* * We want this to be all logical CPUs by default */ if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues) tx_queues = iflib_num_tx_queues; else tx_queues = mp_ncpus; if (tx_queues > scctx->isc_ntxqsets) tx_queues = scctx->isc_ntxqsets; if (ctx->ifc_sysctl_qs_eq_override == 0) { #ifdef INVARIANTS if (tx_queues != rx_queues) device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n", min(rx_queues, tx_queues), min(rx_queues, tx_queues)); #endif tx_queues = min(rx_queues, tx_queues); rx_queues = min(rx_queues, tx_queues); } vectors = rx_queues + admincnt; if (msgs < vectors) { device_printf(dev, "insufficient number of MSI-X vectors " "(supported %d, need %d)\n", msgs, vectors); goto msi; } device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues, tx_queues); msgs = vectors; if ((err = pci_alloc_msix(dev, &vectors)) == 0) { if (vectors != msgs) { device_printf(dev, "Unable to allocate sufficient MSI-X vectors " "(got %d, need %d)\n", vectors, msgs); pci_release_msi(dev); if (bar != -1) { bus_release_resource(dev, SYS_RES_MEMORY, bar, ctx->ifc_msix_mem); ctx->ifc_msix_mem = NULL; } goto msi; } device_printf(dev, "Using MSI-X interrupts with %d vectors\n", vectors); scctx->isc_vectors = vectors; scctx->isc_nrxqsets = rx_queues; scctx->isc_ntxqsets = tx_queues; scctx->isc_intr = IFLIB_INTR_MSIX; return (vectors); } else { device_printf(dev, "failed to allocate %d MSI-X vectors, err: %d\n", vectors, err); if (bar != -1) { bus_release_resource(dev, SYS_RES_MEMORY, bar, ctx->ifc_msix_mem); ctx->ifc_msix_mem = NULL; } } msi: vectors = pci_msi_count(dev); scctx->isc_nrxqsets = 1; scctx->isc_ntxqsets = 1; scctx->isc_vectors = vectors; if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) { device_printf(dev,"Using an MSI interrupt\n"); scctx->isc_intr = IFLIB_INTR_MSI; } else { scctx->isc_vectors = 1; device_printf(dev,"Using a Legacy interrupt\n"); scctx->isc_intr = IFLIB_INTR_LEGACY; } return (vectors); } static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" }; static int mp_ring_state_handler(SYSCTL_HANDLER_ARGS) { int rc; uint16_t *state = ((uint16_t *)oidp->oid_arg1); struct sbuf *sb; const char *ring_state = "UNKNOWN"; /* XXX needed ? */ rc = sysctl_wire_old_buffer(req, 0); MPASS(rc == 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 80, req); MPASS(sb != NULL); if (sb == NULL) return (ENOMEM); if (state[3] <= 3) ring_state = ring_states[state[3]]; sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s", state[0], state[1], state[2], ring_state); rc = sbuf_finish(sb); sbuf_delete(sb); return(rc); } enum iflib_ndesc_handler { IFLIB_NTXD_HANDLER, IFLIB_NRXD_HANDLER, }; static int mp_ndesc_handler(SYSCTL_HANDLER_ARGS) { if_ctx_t ctx = (void *)arg1; enum iflib_ndesc_handler type = arg2; char buf[256] = {0}; qidx_t *ndesc; char *p, *next; int nqs, rc, i; nqs = 8; switch(type) { case IFLIB_NTXD_HANDLER: ndesc = ctx->ifc_sysctl_ntxds; if (ctx->ifc_sctx) nqs = ctx->ifc_sctx->isc_ntxqs; break; case IFLIB_NRXD_HANDLER: ndesc = ctx->ifc_sysctl_nrxds; if (ctx->ifc_sctx) nqs = ctx->ifc_sctx->isc_nrxqs; break; default: printf("%s: unhandled type\n", __func__); return (EINVAL); } if (nqs == 0) nqs = 8; for (i=0; i<8; i++) { if (i >= nqs) break; if (i) strcat(buf, ","); sprintf(strchr(buf, 0), "%d", ndesc[i]); } rc = sysctl_handle_string(oidp, buf, sizeof(buf), req); if (rc || req->newptr == NULL) return rc; for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p; i++, p = strsep(&next, " ,")) { ndesc[i] = strtoul(p, NULL, 10); } return(rc); } #define NAME_BUFLEN 32 static void iflib_add_device_sysctl_pre(if_ctx_t ctx) { device_t dev = iflib_get_dev(ctx); struct sysctl_oid_list *child, *oid_list; struct sysctl_ctx_list *ctx_list; struct sysctl_oid *node; ctx_list = device_get_sysctl_ctx(dev); child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev)); ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib", CTLFLAG_RD, NULL, "IFLIB fields"); oid_list = SYSCTL_CHILDREN(node); SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version", CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, "driver version"); SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs", CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0, "# of txqs to use, 0 => use default #"); SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs", CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0, "# of rxqs to use, 0 => use default #"); SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable", CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0, "permit #txq != #rxq"); SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix", CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0, "disable MSI-X (default 0)"); SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget", CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0, "set the RX budget"); SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate", CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0, "cause TX to abdicate instead of running to completion"); ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED; SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset", CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0, "offset to start using cores at"); SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx", CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0, "use separate cores for TX and RX"); /* XXX change for per-queue sizes */ SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds", CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A", "list of # of TX descriptors to use, 0 = use default #"); SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds", CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A", "list of # of RX descriptors to use, 0 = use default #"); } static void iflib_add_device_sysctl_post(if_ctx_t ctx) { if_shared_ctx_t sctx = ctx->ifc_sctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; device_t dev = iflib_get_dev(ctx); struct sysctl_oid_list *child; struct sysctl_ctx_list *ctx_list; iflib_fl_t fl; iflib_txq_t txq; iflib_rxq_t rxq; int i, j; char namebuf[NAME_BUFLEN]; char *qfmt; struct sysctl_oid *queue_node, *fl_node, *node; struct sysctl_oid_list *queue_list, *fl_list; ctx_list = device_get_sysctl_ctx(dev); node = ctx->ifc_sysctl_node; child = SYSCTL_CHILDREN(node); if (scctx->isc_ntxqsets > 100) qfmt = "txq%03d"; else if (scctx->isc_ntxqsets > 10) qfmt = "txq%02d"; else qfmt = "txq%d"; for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) { snprintf(namebuf, NAME_BUFLEN, qfmt, i); queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf, CTLFLAG_RD, NULL, "Queue Name"); queue_list = SYSCTL_CHILDREN(queue_node); #if MEMORY_LOGGING SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued", CTLFLAG_RD, &txq->ift_dequeued, "total mbufs freed"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued", CTLFLAG_RD, &txq->ift_enqueued, "total mbufs enqueued"); #endif SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag", CTLFLAG_RD, &txq->ift_mbuf_defrag, "# of times m_defrag was called"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups", CTLFLAG_RD, &txq->ift_pullups, "# of times m_pullup was called"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed", CTLFLAG_RD, &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail", CTLFLAG_RD, &txq->ift_no_desc_avail, "# of times no descriptors were available"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed", CTLFLAG_RD, &txq->ift_map_failed, "# of times DMA map failed"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig", CTLFLAG_RD, &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup", CTLFLAG_RD, &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG"); SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx", CTLFLAG_RD, &txq->ift_pidx, 1, "Producer Index"); SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx", CTLFLAG_RD, &txq->ift_cidx, 1, "Consumer Index"); SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed", CTLFLAG_RD, &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update"); SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use", CTLFLAG_RD, &txq->ift_in_use, 1, "descriptors in use"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed", CTLFLAG_RD, &txq->ift_processed, "descriptors procesed for clean"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned", CTLFLAG_RD, &txq->ift_cleaned, "total cleaned"); SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state", CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0, mp_ring_state_handler, "A", "soft ring state"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues", CTLFLAG_RD, &txq->ift_br->enqueues, "# of enqueues to the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops", CTLFLAG_RD, &txq->ift_br->drops, "# of drops in the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts", CTLFLAG_RD, &txq->ift_br->starts, "# of normal consumer starts in the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls", CTLFLAG_RD, &txq->ift_br->stalls, "# of consumer stalls in the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts", CTLFLAG_RD, &txq->ift_br->restarts, "# of consumer restarts in the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications", CTLFLAG_RD, &txq->ift_br->abdications, "# of consumer abdications in the mp_ring for this queue"); } if (scctx->isc_nrxqsets > 100) qfmt = "rxq%03d"; else if (scctx->isc_nrxqsets > 10) qfmt = "rxq%02d"; else qfmt = "rxq%d"; for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) { snprintf(namebuf, NAME_BUFLEN, qfmt, i); queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf, CTLFLAG_RD, NULL, "Queue Name"); queue_list = SYSCTL_CHILDREN(queue_node); if (sctx->isc_flags & IFLIB_HAS_RXCQ) { SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx", CTLFLAG_RD, &rxq->ifr_cq_cidx, 1, "Consumer Index"); } for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) { snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j); fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf, CTLFLAG_RD, NULL, "freelist Name"); fl_list = SYSCTL_CHILDREN(fl_node); SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx", CTLFLAG_RD, &fl->ifl_pidx, 1, "Producer Index"); SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx", CTLFLAG_RD, &fl->ifl_cidx, 1, "Consumer Index"); SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits", CTLFLAG_RD, &fl->ifl_credits, 1, "credits available"); #if MEMORY_LOGGING SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued", CTLFLAG_RD, &fl->ifl_m_enqueued, "mbufs allocated"); SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued", CTLFLAG_RD, &fl->ifl_m_dequeued, "mbufs freed"); SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued", CTLFLAG_RD, &fl->ifl_cl_enqueued, "clusters allocated"); SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued", CTLFLAG_RD, &fl->ifl_cl_dequeued, "clusters freed"); #endif } } } void iflib_request_reset(if_ctx_t ctx) { STATE_LOCK(ctx); ctx->ifc_flags |= IFC_DO_RESET; STATE_UNLOCK(ctx); } #ifndef __NO_STRICT_ALIGNMENT static struct mbuf * iflib_fixup_rx(struct mbuf *m) { struct mbuf *n; if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) { bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len); m->m_data += ETHER_HDR_LEN; n = m; } else { MGETHDR(n, M_NOWAIT, MT_DATA); if (n == NULL) { m_freem(m); return (NULL); } bcopy(m->m_data, n->m_data, ETHER_HDR_LEN); m->m_data += ETHER_HDR_LEN; m->m_len -= ETHER_HDR_LEN; n->m_len = ETHER_HDR_LEN; M_MOVE_PKTHDR(n, m); n->m_next = m; } return (n); } #endif #ifdef NETDUMP static void iflib_netdump_init(if_t ifp, int *nrxr, int *ncl, int *clsize) { if_ctx_t ctx; ctx = if_getsoftc(ifp); CTX_LOCK(ctx); *nrxr = NRXQSETS(ctx); *ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size; *clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size; CTX_UNLOCK(ctx); } static void iflib_netdump_event(if_t ifp, enum netdump_ev event) { if_ctx_t ctx; if_softc_ctx_t scctx; iflib_fl_t fl; iflib_rxq_t rxq; int i, j; ctx = if_getsoftc(ifp); scctx = &ctx->ifc_softc_ctx; switch (event) { case NETDUMP_START: for (i = 0; i < scctx->isc_nrxqsets; i++) { rxq = &ctx->ifc_rxqs[i]; for (j = 0; j < rxq->ifr_nfl; j++) { fl = rxq->ifr_fl; fl->ifl_zone = m_getzone(fl->ifl_buf_size); } } iflib_no_tx_batch = 1; break; default: break; } } static int iflib_netdump_transmit(if_t ifp, struct mbuf *m) { if_ctx_t ctx; iflib_txq_t txq; int error; ctx = if_getsoftc(ifp); if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) != IFF_DRV_RUNNING) return (EBUSY); txq = &ctx->ifc_txqs[0]; error = iflib_encap(txq, &m); if (error == 0) (void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use); return (error); } static int iflib_netdump_poll(if_t ifp, int count) { if_ctx_t ctx; if_softc_ctx_t scctx; iflib_txq_t txq; int i; ctx = if_getsoftc(ifp); scctx = &ctx->ifc_softc_ctx; if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) != IFF_DRV_RUNNING) return (EBUSY); txq = &ctx->ifc_txqs[0]; (void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx)); for (i = 0; i < scctx->isc_nrxqsets; i++) (void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */); return (0); } #endif /* NETDUMP */ Index: head/sys/opencrypto/ktls_ocf.c =================================================================== --- head/sys/opencrypto/ktls_ocf.c (revision 352813) +++ head/sys/opencrypto/ktls_ocf.c (revision 352814) @@ -1,308 +1,308 @@ /*- * SPDX-License-Identifier: BSD-2-Clause * * Copyright (c) 2019 Netflix Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include struct ocf_session { crypto_session_t sid; int crda_alg; struct mtx lock; }; struct ocf_operation { struct ocf_session *os; bool done; struct iovec iov[0]; }; static MALLOC_DEFINE(M_KTLS_OCF, "ktls_ocf", "OCF KTLS"); SYSCTL_DECL(_kern_ipc_tls); SYSCTL_DECL(_kern_ipc_tls_stats); static counter_u64_t ocf_gcm_crypts; SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ocf_gcm_crypts, CTLFLAG_RD, &ocf_gcm_crypts, "Total number of OCF GCM encryption operations"); static counter_u64_t ocf_retries; SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ocf_retries, CTLFLAG_RD, &ocf_retries, "Number of OCF encryption operation retries"); static int ktls_ocf_callback(struct cryptop *crp) { struct ocf_operation *oo; oo = crp->crp_opaque; mtx_lock(&oo->os->lock); oo->done = true; mtx_unlock(&oo->os->lock); wakeup(oo); return (0); } static int ktls_ocf_encrypt(struct ktls_session *tls, const struct tls_record_layer *hdr, uint8_t *trailer, struct iovec *iniov, struct iovec *outiov, int iovcnt, - uint64_t seqno) + uint64_t seqno, uint8_t record_type __unused) { struct uio uio; struct tls_aead_data ad; struct tls_nonce_data nd; struct cryptodesc *crde, *crda; struct cryptop *crp; struct ocf_session *os; struct ocf_operation *oo; struct iovec *iov; int i, error; uint16_t tls_comp_len; os = tls->cipher; oo = malloc(sizeof(*oo) + (iovcnt + 2) * sizeof(*iov), M_KTLS_OCF, M_WAITOK | M_ZERO); oo->os = os; iov = oo->iov; crp = crypto_getreq(2); if (crp == NULL) { free(oo, M_KTLS_OCF); return (ENOMEM); } /* Setup the IV. */ memcpy(nd.fixed, tls->params.iv, TLS_AEAD_GCM_LEN); memcpy(&nd.seq, hdr + 1, sizeof(nd.seq)); /* Setup the AAD. */ tls_comp_len = ntohs(hdr->tls_length) - (AES_GMAC_HASH_LEN + sizeof(nd.seq)); ad.seq = htobe64(seqno); ad.type = hdr->tls_type; ad.tls_vmajor = hdr->tls_vmajor; ad.tls_vminor = hdr->tls_vminor; ad.tls_length = htons(tls_comp_len); iov[0].iov_base = &ad; iov[0].iov_len = sizeof(ad); uio.uio_resid = sizeof(ad); /* * OCF always does encryption in place, so copy the data if * needed. Ugh. */ for (i = 0; i < iovcnt; i++) { iov[i + 1] = outiov[i]; if (iniov[i].iov_base != outiov[i].iov_base) memcpy(outiov[i].iov_base, iniov[i].iov_base, outiov[i].iov_len); uio.uio_resid += outiov[i].iov_len; } iov[iovcnt + 1].iov_base = trailer; iov[iovcnt + 1].iov_len = AES_GMAC_HASH_LEN; uio.uio_resid += AES_GMAC_HASH_LEN; uio.uio_iov = iov; uio.uio_iovcnt = iovcnt + 2; uio.uio_offset = 0; uio.uio_segflg = UIO_SYSSPACE; uio.uio_td = curthread; crp->crp_session = os->sid; crp->crp_flags = CRYPTO_F_IOV | CRYPTO_F_CBIMM; crp->crp_uio = &uio; crp->crp_ilen = uio.uio_resid; crp->crp_opaque = oo; crp->crp_callback = ktls_ocf_callback; crde = crp->crp_desc; crda = crde->crd_next; crda->crd_alg = os->crda_alg; crda->crd_skip = 0; crda->crd_len = sizeof(ad); crda->crd_inject = crp->crp_ilen - AES_GMAC_HASH_LEN; crde->crd_alg = CRYPTO_AES_NIST_GCM_16; crde->crd_skip = sizeof(ad); crde->crd_len = crp->crp_ilen - (sizeof(ad) + AES_GMAC_HASH_LEN); crde->crd_flags = CRD_F_ENCRYPT | CRD_F_IV_EXPLICIT | CRD_F_IV_PRESENT; memcpy(crde->crd_iv, &nd, sizeof(nd)); counter_u64_add(ocf_gcm_crypts, 1); for (;;) { error = crypto_dispatch(crp); if (error) break; mtx_lock(&os->lock); while (!oo->done) mtx_sleep(oo, &os->lock, 0, "ocfktls", 0); mtx_unlock(&os->lock); if (crp->crp_etype != EAGAIN) { error = crp->crp_etype; break; } crp->crp_etype = 0; crp->crp_flags &= ~CRYPTO_F_DONE; oo->done = false; counter_u64_add(ocf_retries, 1); } crypto_freereq(crp); free(oo, M_KTLS_OCF); return (error); } static void ktls_ocf_free(struct ktls_session *tls) { struct ocf_session *os; os = tls->cipher; mtx_destroy(&os->lock); explicit_bzero(os, sizeof(*os)); free(os, M_KTLS_OCF); } static int ktls_ocf_try(struct socket *so, struct ktls_session *tls) { struct cryptoini cria, crie; struct ocf_session *os; int error; memset(&cria, 0, sizeof(cria)); memset(&crie, 0, sizeof(crie)); switch (tls->params.cipher_algorithm) { case CRYPTO_AES_NIST_GCM_16: if (tls->params.iv_len != TLS_AEAD_GCM_LEN) return (EINVAL); switch (tls->params.cipher_key_len) { case 128 / 8: cria.cri_alg = CRYPTO_AES_128_NIST_GMAC; break; case 256 / 8: cria.cri_alg = CRYPTO_AES_256_NIST_GMAC; break; default: return (EINVAL); } cria.cri_key = tls->params.cipher_key; cria.cri_klen = tls->params.cipher_key_len * 8; break; default: return (EPROTONOSUPPORT); } /* Only TLS 1.1 and TLS 1.2 are currently supported. */ if (tls->params.tls_vmajor != TLS_MAJOR_VER_ONE || tls->params.tls_vminor < TLS_MINOR_VER_ONE || tls->params.tls_vminor > TLS_MINOR_VER_TWO) return (EPROTONOSUPPORT); os = malloc(sizeof(*os), M_KTLS_OCF, M_NOWAIT | M_ZERO); if (os == NULL) return (ENOMEM); crie.cri_alg = tls->params.cipher_algorithm; crie.cri_key = tls->params.cipher_key; crie.cri_klen = tls->params.cipher_key_len * 8; crie.cri_next = &cria; error = crypto_newsession(&os->sid, &crie, CRYPTO_FLAG_HARDWARE | CRYPTO_FLAG_SOFTWARE); if (error) { free(os, M_KTLS_OCF); return (error); } os->crda_alg = cria.cri_alg; mtx_init(&os->lock, "ktls_ocf", NULL, MTX_DEF); tls->cipher = os; tls->sw_encrypt = ktls_ocf_encrypt; tls->free = ktls_ocf_free; return (0); } struct ktls_crypto_backend ocf_backend = { .name = "OCF", .prio = 5, .api_version = KTLS_API_VERSION, .try = ktls_ocf_try, }; static int ktls_ocf_modevent(module_t mod, int what, void *arg) { int error; switch (what) { case MOD_LOAD: ocf_gcm_crypts = counter_u64_alloc(M_WAITOK); ocf_retries = counter_u64_alloc(M_WAITOK); return (ktls_crypto_backend_register(&ocf_backend)); case MOD_UNLOAD: error = ktls_crypto_backend_deregister(&ocf_backend); if (error) return (error); counter_u64_free(ocf_gcm_crypts); counter_u64_free(ocf_retries); return (0); default: return (EOPNOTSUPP); } } static moduledata_t ktls_ocf_moduledata = { "ktls_ocf", ktls_ocf_modevent, NULL }; DECLARE_MODULE(ktls_ocf, ktls_ocf_moduledata, SI_SUB_PROTO_END, SI_ORDER_ANY); Index: head/sys/sys/ktls.h =================================================================== --- head/sys/sys/ktls.h (revision 352813) +++ head/sys/sys/ktls.h (revision 352814) @@ -1,194 +1,196 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2014-2019 Netflix Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #ifndef _SYS_KTLS_H_ #define _SYS_KTLS_H_ #include #include struct tls_record_layer { uint8_t tls_type; uint8_t tls_vmajor; uint8_t tls_vminor; uint16_t tls_length; uint8_t tls_data[0]; } __attribute__ ((packed)); #define TLS_MAX_MSG_SIZE_V10_2 16384 #define TLS_MAX_PARAM_SIZE 1024 /* Max key/mac/iv in sockopt */ #define TLS_AEAD_GCM_LEN 4 +#define TLS_1_3_GCM_IV_LEN 12 #define TLS_CBC_IMPLICIT_IV_LEN 16 /* Type values for the record layer */ #define TLS_RLTYPE_APP 23 /* * Nonce for GCM. */ struct tls_nonce_data { uint8_t fixed[TLS_AEAD_GCM_LEN]; uint64_t seq; } __packed; /* * AEAD additional data format per RFC. */ struct tls_aead_data { uint64_t seq; /* In network order */ uint8_t type; uint8_t tls_vmajor; uint8_t tls_vminor; uint16_t tls_length; } __packed; /* * Stream Cipher MAC additional data input. This does not match the * exact data on the wire (the sequence number is not placed on the * wire, and any explicit IV after the record header is not covered by * the MAC). */ struct tls_mac_data { uint64_t seq; uint8_t type; uint8_t tls_vmajor; uint8_t tls_vminor; uint16_t tls_length; } __packed; #define TLS_MAJOR_VER_ONE 3 #define TLS_MINOR_VER_ZERO 1 /* 3, 1 */ #define TLS_MINOR_VER_ONE 2 /* 3, 2 */ #define TLS_MINOR_VER_TWO 3 /* 3, 3 */ +#define TLS_MINOR_VER_THREE 4 /* 3, 4 */ /* For TCP_TXTLS_ENABLE */ struct tls_enable { const uint8_t *cipher_key; const uint8_t *iv; /* Implicit IV. */ const uint8_t *auth_key; int cipher_algorithm; /* e.g. CRYPTO_AES_CBC */ int cipher_key_len; int iv_len; int auth_algorithm; /* e.g. CRYPTO_SHA2_256_HMAC */ int auth_key_len; int flags; uint8_t tls_vmajor; uint8_t tls_vminor; }; struct tls_session_params { uint8_t *cipher_key; uint8_t *auth_key; uint8_t iv[TLS_CBC_IMPLICIT_IV_LEN]; int cipher_algorithm; int auth_algorithm; uint16_t cipher_key_len; uint16_t iv_len; uint16_t auth_key_len; uint16_t max_frame_len; uint8_t tls_vmajor; uint8_t tls_vminor; uint8_t tls_hlen; uint8_t tls_tlen; uint8_t tls_bs; uint8_t flags; }; #ifdef _KERNEL -#define KTLS_API_VERSION 5 +#define KTLS_API_VERSION 6 struct iovec; struct ktls_session; struct m_snd_tag; struct mbuf; struct mbuf_ext_pgs; struct sockbuf; struct socket; struct ktls_crypto_backend { LIST_ENTRY(ktls_crypto_backend) next; int (*try)(struct socket *so, struct ktls_session *tls); int prio; int api_version; int use_count; const char *name; }; struct ktls_session { int (*sw_encrypt)(struct ktls_session *tls, const struct tls_record_layer *hdr, uint8_t *trailer, struct iovec *src, struct iovec *dst, int iovcnt, - uint64_t seqno); + uint64_t seqno, uint8_t record_type); union { void *cipher; struct m_snd_tag *snd_tag; }; struct ktls_crypto_backend *be; void (*free)(struct ktls_session *tls); struct tls_session_params params; u_int wq_index; volatile u_int refcount; struct task reset_tag_task; struct inpcb *inp; bool reset_pending; } __aligned(CACHE_LINE_SIZE); int ktls_crypto_backend_register(struct ktls_crypto_backend *be); int ktls_crypto_backend_deregister(struct ktls_crypto_backend *be); int ktls_enable_tx(struct socket *so, struct tls_enable *en); void ktls_destroy(struct ktls_session *tls); int ktls_frame(struct mbuf *m, struct ktls_session *tls, int *enqueue_cnt, uint8_t record_type); void ktls_seq(struct sockbuf *sb, struct mbuf *m); void ktls_enqueue(struct mbuf *m, struct socket *so, int page_count); void ktls_enqueue_to_free(struct mbuf_ext_pgs *pgs); int ktls_set_tx_mode(struct socket *so, int mode); int ktls_get_tx_mode(struct socket *so); int ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls); static inline struct ktls_session * ktls_hold(struct ktls_session *tls) { if (tls != NULL) refcount_acquire(&tls->refcount); return (tls); } static inline void ktls_free(struct ktls_session *tls) { if (refcount_release(&tls->refcount)) ktls_destroy(tls); } #endif /* !_KERNEL */ #endif /* !_SYS_KTLS_H_ */ Index: head/sys/sys/mbuf.h =================================================================== --- head/sys/sys/mbuf.h (revision 352813) +++ head/sys/sys/mbuf.h (revision 352814) @@ -1,1526 +1,1527 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1988, 1993 * The Regents of the University of California. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)mbuf.h 8.5 (Berkeley) 2/19/95 * $FreeBSD$ */ #ifndef _SYS_MBUF_H_ #define _SYS_MBUF_H_ /* XXX: These includes suck. Sorry! */ #include #ifdef _KERNEL #include #include #include #ifdef WITNESS #include #endif #endif #ifdef _KERNEL #include #define MBUF_PROBE1(probe, arg0) \ SDT_PROBE1(sdt, , , probe, arg0) #define MBUF_PROBE2(probe, arg0, arg1) \ SDT_PROBE2(sdt, , , probe, arg0, arg1) #define MBUF_PROBE3(probe, arg0, arg1, arg2) \ SDT_PROBE3(sdt, , , probe, arg0, arg1, arg2) #define MBUF_PROBE4(probe, arg0, arg1, arg2, arg3) \ SDT_PROBE4(sdt, , , probe, arg0, arg1, arg2, arg3) #define MBUF_PROBE5(probe, arg0, arg1, arg2, arg3, arg4) \ SDT_PROBE5(sdt, , , probe, arg0, arg1, arg2, arg3, arg4) SDT_PROBE_DECLARE(sdt, , , m__init); SDT_PROBE_DECLARE(sdt, , , m__gethdr); SDT_PROBE_DECLARE(sdt, , , m__get); SDT_PROBE_DECLARE(sdt, , , m__getcl); SDT_PROBE_DECLARE(sdt, , , m__clget); SDT_PROBE_DECLARE(sdt, , , m__cljget); SDT_PROBE_DECLARE(sdt, , , m__cljset); SDT_PROBE_DECLARE(sdt, , , m__free); SDT_PROBE_DECLARE(sdt, , , m__freem); #endif /* _KERNEL */ /* * Mbufs are of a single size, MSIZE (sys/param.h), which includes overhead. * An mbuf may add a single "mbuf cluster" of size MCLBYTES (also in * sys/param.h), which has no additional overhead and is used instead of the * internal data area; this is done when at least MINCLSIZE of data must be * stored. Additionally, it is possible to allocate a separate buffer * externally and attach it to the mbuf in a way similar to that of mbuf * clusters. * * NB: These calculation do not take actual compiler-induced alignment and * padding inside the complete struct mbuf into account. Appropriate * attention is required when changing members of struct mbuf. * * MLEN is data length in a normal mbuf. * MHLEN is data length in an mbuf with pktheader. * MINCLSIZE is a smallest amount of data that should be put into cluster. * * Compile-time assertions in uipc_mbuf.c test these values to ensure that * they are sensible. */ struct mbuf; #define MHSIZE offsetof(struct mbuf, m_dat) #define MPKTHSIZE offsetof(struct mbuf, m_pktdat) #define MLEN ((int)(MSIZE - MHSIZE)) #define MHLEN ((int)(MSIZE - MPKTHSIZE)) #define MINCLSIZE (MHLEN + 1) #define M_NODOM 255 #ifdef _KERNEL /*- * Macro for type conversion: convert mbuf pointer to data pointer of correct * type: * * mtod(m, t) -- Convert mbuf pointer to data pointer of correct type. * mtodo(m, o) -- Same as above but with offset 'o' into data. */ #define mtod(m, t) ((t)((m)->m_data)) #define mtodo(m, o) ((void *)(((m)->m_data) + (o))) /* * Argument structure passed to UMA routines during mbuf and packet * allocations. */ struct mb_args { int flags; /* Flags for mbuf being allocated */ short type; /* Type of mbuf being allocated */ }; #endif /* _KERNEL */ /* * Packet tag structure (see below for details). */ struct m_tag { SLIST_ENTRY(m_tag) m_tag_link; /* List of packet tags */ u_int16_t m_tag_id; /* Tag ID */ u_int16_t m_tag_len; /* Length of data */ u_int32_t m_tag_cookie; /* ABI/Module ID */ void (*m_tag_free)(struct m_tag *); }; /* * Static network interface owned tag. * Allocated through ifp->if_snd_tag_alloc(). */ struct m_snd_tag { struct ifnet *ifp; /* network interface tag belongs to */ volatile u_int refcount; }; /* * Record/packet header in first mbuf of chain; valid only if M_PKTHDR is set. * Size ILP32: 48 * LP64: 56 * Compile-time assertions in uipc_mbuf.c test these values to ensure that * they are correct. */ struct pkthdr { union { struct m_snd_tag *snd_tag; /* send tag, if any */ struct ifnet *rcvif; /* rcv interface */ }; SLIST_HEAD(packet_tags, m_tag) tags; /* list of packet tags */ int32_t len; /* total packet length */ /* Layer crossing persistent information. */ uint32_t flowid; /* packet's 4-tuple system */ uint32_t csum_flags; /* checksum and offload features */ uint16_t fibnum; /* this packet should use this fib */ uint8_t numa_domain; /* NUMA domain of recvd pkt */ uint8_t rsstype; /* hash type */ union { uint64_t rcv_tstmp; /* timestamp in ns */ struct { uint8_t l2hlen; /* layer 2 hdr len */ uint8_t l3hlen; /* layer 3 hdr len */ uint8_t l4hlen; /* layer 4 hdr len */ uint8_t l5hlen; /* layer 5 hdr len */ uint32_t spare; }; }; union { uint8_t eight[8]; uint16_t sixteen[4]; uint32_t thirtytwo[2]; uint64_t sixtyfour[1]; uintptr_t unintptr[1]; void *ptr; } PH_per; /* Layer specific non-persistent local storage for reassembly, etc. */ union { uint8_t eight[8]; uint16_t sixteen[4]; uint32_t thirtytwo[2]; uint64_t sixtyfour[1]; uintptr_t unintptr[1]; void *ptr; } PH_loc; }; #define ether_vtag PH_per.sixteen[0] #define PH_vt PH_per #define vt_nrecs sixteen[0] #define tso_segsz PH_per.sixteen[1] #define lro_nsegs tso_segsz #define csum_phsum PH_per.sixteen[2] #define csum_data PH_per.thirtytwo[1] #define lro_len PH_per.sixteen[0] /* inbound during LRO */ #define lro_csum PH_per.sixteen[1] /* inbound during LRO */ #define pace_thoff PH_loc.sixteen[0] #define pace_tlen PH_loc.sixteen[1] #define pace_drphdrlen PH_loc.sixteen[2] #define pace_tos PH_loc.eight[6] #define pace_lock PH_loc.eight[7] /* * Description of external storage mapped into mbuf; valid only if M_EXT is * set. * Size ILP32: 28 * LP64: 48 * Compile-time assertions in uipc_mbuf.c test these values to ensure that * they are correct. */ typedef void m_ext_free_t(struct mbuf *); struct m_ext { union { /* * If EXT_FLAG_EMBREF is set, then we use refcount in the * mbuf, the 'ext_count' member. Otherwise, we have a * shadow copy and we use pointer 'ext_cnt'. The original * mbuf is responsible to carry the pointer to free routine * and its arguments. They aren't copied into shadows in * mb_dupcl() to avoid dereferencing next cachelines. */ volatile u_int ext_count; volatile u_int *ext_cnt; }; union { /* * If ext_type == EXT_PGS, 'ext_pgs' points to a * structure describing the buffer. Otherwise, * 'ext_buf' points to the start of the buffer. */ struct mbuf_ext_pgs *ext_pgs; char *ext_buf; }; uint32_t ext_size; /* size of buffer, for ext_free */ uint32_t ext_type:8, /* type of external storage */ ext_flags:24; /* external storage mbuf flags */ /* * Fields below store the free context for the external storage. * They are valid only in the refcount carrying mbuf, the one with * EXT_FLAG_EMBREF flag, with exclusion for EXT_EXTREF type, where * the free context is copied into all mbufs that use same external * storage. */ #define m_ext_copylen offsetof(struct m_ext, ext_free) m_ext_free_t *ext_free; /* free routine if not the usual */ void *ext_arg1; /* optional argument pointer */ void *ext_arg2; /* optional argument pointer */ }; /* * The core of the mbuf object along with some shortcut defines for practical * purposes. */ struct mbuf { /* * Header present at the beginning of every mbuf. * Size ILP32: 24 * LP64: 32 * Compile-time assertions in uipc_mbuf.c test these values to ensure * that they are correct. */ union { /* next buffer in chain */ struct mbuf *m_next; SLIST_ENTRY(mbuf) m_slist; STAILQ_ENTRY(mbuf) m_stailq; }; union { /* next chain in queue/record */ struct mbuf *m_nextpkt; SLIST_ENTRY(mbuf) m_slistpkt; STAILQ_ENTRY(mbuf) m_stailqpkt; }; caddr_t m_data; /* location of data */ int32_t m_len; /* amount of data in this mbuf */ uint32_t m_type:8, /* type of data in this mbuf */ m_flags:24; /* flags; see below */ #if !defined(__LP64__) uint32_t m_pad; /* pad for 64bit alignment */ #endif /* * A set of optional headers (packet header, external storage header) * and internal data storage. Historically, these arrays were sized * to MHLEN (space left after a packet header) and MLEN (space left * after only a regular mbuf header); they are now variable size in * order to support future work on variable-size mbufs. */ union { struct { struct pkthdr m_pkthdr; /* M_PKTHDR set */ union { struct m_ext m_ext; /* M_EXT set */ char m_pktdat[0]; }; }; char m_dat[0]; /* !M_PKTHDR, !M_EXT */ }; }; struct ktls_session; struct socket; /* * TLS records for TLS 1.0-1.2 can have the following header lengths: * - 5 (AES-CBC with implicit IV) * - 21 (AES-CBC with explicit IV) * - 13 (AES-GCM with 8 byte explicit IV) */ #define MBUF_PEXT_HDR_LEN 24 /* * TLS records for TLS 1.0-1.2 can have the following maximum trailer * lengths: * - 16 (AES-GCM) * - 36 (AES-CBC with SHA1 and up to 16 bytes of padding) * - 48 (AES-CBC with SHA2-256 and up to 16 bytes of padding) * - 64 (AES-CBC with SHA2-384 and up to 16 bytes of padding) */ #define MBUF_PEXT_TRAIL_LEN 64 #ifdef __LP64__ #define MBUF_PEXT_MAX_PGS (152 / sizeof(vm_paddr_t)) #else #define MBUF_PEXT_MAX_PGS (156 / sizeof(vm_paddr_t)) #endif #define MBUF_PEXT_MAX_BYTES \ (MBUF_PEXT_MAX_PGS * PAGE_SIZE + MBUF_PEXT_HDR_LEN + MBUF_PEXT_TRAIL_LEN) /* * This struct is 256 bytes in size and is arranged so that the most * common case (accessing the first 4 pages of a 16KB TLS record) will * fit in a single 64 byte cacheline. */ struct mbuf_ext_pgs { uint8_t npgs; /* Number of attached pages */ uint8_t nrdy; /* Pages with I/O pending */ uint8_t hdr_len; /* TLS header length */ uint8_t trail_len; /* TLS trailer length */ uint16_t first_pg_off; /* Offset into 1st page */ uint16_t last_pg_len; /* Length of last page */ vm_paddr_t pa[MBUF_PEXT_MAX_PGS]; /* phys addrs of pages */ char hdr[MBUF_PEXT_HDR_LEN]; /* TLS header */ struct ktls_session *tls; /* TLS session */ #if defined(__i386__) || \ (defined(__powerpc__) && !defined(__powerpc64__) && defined(BOOKE)) /* * i386 and Book-E PowerPC have 64-bit vm_paddr_t, so there is * a 4 byte remainder from the space allocated for pa[]. */ uint32_t pad; #endif union { char trail[MBUF_PEXT_TRAIL_LEN]; /* TLS trailer */ struct { + uint8_t record_type; /* Must be first */ struct socket *so; struct mbuf *mbuf; uint64_t seqno; STAILQ_ENTRY(mbuf_ext_pgs) stailq; int enc_cnt; }; }; }; #ifdef _KERNEL static inline int mbuf_ext_pg_len(struct mbuf_ext_pgs *ext_pgs, int pidx, int pgoff) { KASSERT(pgoff == 0 || pidx == 0, ("page %d with non-zero offset %d in %p", pidx, pgoff, ext_pgs)); if (pidx == ext_pgs->npgs - 1) { return (ext_pgs->last_pg_len); } else { return (PAGE_SIZE - pgoff); } } #ifdef INVARIANT_SUPPORT void mb_ext_pgs_check(struct mbuf_ext_pgs *ext_pgs); #endif #ifdef INVARIANTS #define MBUF_EXT_PGS_ASSERT_SANITY(ext_pgs) mb_ext_pgs_check((ext_pgs)) #else #define MBUF_EXT_PGS_ASSERT_SANITY(ext_pgs) #endif #endif /* * mbuf flags of global significance and layer crossing. * Those of only protocol/layer specific significance are to be mapped * to M_PROTO[1-11] and cleared at layer handoff boundaries. * NB: Limited to the lower 24 bits. */ #define M_EXT 0x00000001 /* has associated external storage */ #define M_PKTHDR 0x00000002 /* start of record */ #define M_EOR 0x00000004 /* end of record */ #define M_RDONLY 0x00000008 /* associated data is marked read-only */ #define M_BCAST 0x00000010 /* send/received as link-level broadcast */ #define M_MCAST 0x00000020 /* send/received as link-level multicast */ #define M_PROMISC 0x00000040 /* packet was not for us */ #define M_VLANTAG 0x00000080 /* ether_vtag is valid */ #define M_NOMAP 0x00000100 /* mbuf data is unmapped */ #define M_NOFREE 0x00000200 /* do not free mbuf, embedded in cluster */ #define M_TSTMP 0x00000400 /* rcv_tstmp field is valid */ #define M_TSTMP_HPREC 0x00000800 /* rcv_tstmp is high-prec, typically hw-stamped on port (useful for IEEE 1588 and 802.1AS) */ #define M_TSTMP_LRO 0x00001000 /* Time LRO pushed in pkt is valid in (PH_loc) */ #define M_PROTO1 0x00002000 /* protocol-specific */ #define M_PROTO2 0x00004000 /* protocol-specific */ #define M_PROTO3 0x00008000 /* protocol-specific */ #define M_PROTO4 0x00010000 /* protocol-specific */ #define M_PROTO5 0x00020000 /* protocol-specific */ #define M_PROTO6 0x00040000 /* protocol-specific */ #define M_PROTO7 0x00080000 /* protocol-specific */ #define M_PROTO8 0x00100000 /* protocol-specific */ #define M_PROTO9 0x00200000 /* protocol-specific */ #define M_PROTO10 0x00400000 /* protocol-specific */ #define M_PROTO11 0x00800000 /* protocol-specific */ #define MB_DTOR_SKIP 0x1 /* don't pollute the cache by touching a freed mbuf */ /* * Flags to purge when crossing layers. */ #define M_PROTOFLAGS \ (M_PROTO1|M_PROTO2|M_PROTO3|M_PROTO4|M_PROTO5|M_PROTO6|M_PROTO7|M_PROTO8|\ M_PROTO9|M_PROTO10|M_PROTO11) /* * Flags preserved when copying m_pkthdr. */ #define M_COPYFLAGS \ (M_PKTHDR|M_EOR|M_RDONLY|M_BCAST|M_MCAST|M_PROMISC|M_VLANTAG|M_TSTMP| \ M_TSTMP_HPREC|M_TSTMP_LRO|M_PROTOFLAGS) /* * Mbuf flag description for use with printf(9) %b identifier. */ #define M_FLAG_BITS \ "\20\1M_EXT\2M_PKTHDR\3M_EOR\4M_RDONLY\5M_BCAST\6M_MCAST" \ "\7M_PROMISC\10M_VLANTAG\11M_NOMAP\12M_NOFREE\13M_TSTMP\14M_TSTMP_HPREC\15M_TSTMP_LRO" #define M_FLAG_PROTOBITS \ "\16M_PROTO1\17M_PROTO2\20M_PROTO3\21M_PROTO4" \ "\22M_PROTO5\23M_PROTO6\24M_PROTO7\25M_PROTO8\26M_PROTO9" \ "\27M_PROTO10\28M_PROTO11" #define M_FLAG_PRINTF (M_FLAG_BITS M_FLAG_PROTOBITS) /* * Network interface cards are able to hash protocol fields (such as IPv4 * addresses and TCP port numbers) classify packets into flows. These flows * can then be used to maintain ordering while delivering packets to the OS * via parallel input queues, as well as to provide a stateless affinity * model. NIC drivers can pass up the hash via m->m_pkthdr.flowid, and set * m_flag fields to indicate how the hash should be interpreted by the * network stack. * * Most NICs support RSS, which provides ordering and explicit affinity, and * use the hash m_flag bits to indicate what header fields were covered by * the hash. M_HASHTYPE_OPAQUE and M_HASHTYPE_OPAQUE_HASH can be set by non- * RSS cards or configurations that provide an opaque flow identifier, allowing * for ordering and distribution without explicit affinity. Additionally, * M_HASHTYPE_OPAQUE_HASH indicates that the flow identifier has hash * properties. * * The meaning of the IPV6_EX suffix: * "o Home address from the home address option in the IPv6 destination * options header. If the extension header is not present, use the Source * IPv6 Address. * o IPv6 address that is contained in the Routing-Header-Type-2 from the * associated extension header. If the extension header is not present, * use the Destination IPv6 Address." * Quoted from: * https://docs.microsoft.com/en-us/windows-hardware/drivers/network/rss-hashing-types#ndishashipv6ex */ #define M_HASHTYPE_HASHPROP 0x80 /* has hash properties */ #define M_HASHTYPE_HASH(t) (M_HASHTYPE_HASHPROP | (t)) /* Microsoft RSS standard hash types */ #define M_HASHTYPE_NONE 0 #define M_HASHTYPE_RSS_IPV4 M_HASHTYPE_HASH(1) /* IPv4 2-tuple */ #define M_HASHTYPE_RSS_TCP_IPV4 M_HASHTYPE_HASH(2) /* TCPv4 4-tuple */ #define M_HASHTYPE_RSS_IPV6 M_HASHTYPE_HASH(3) /* IPv6 2-tuple */ #define M_HASHTYPE_RSS_TCP_IPV6 M_HASHTYPE_HASH(4) /* TCPv6 4-tuple */ #define M_HASHTYPE_RSS_IPV6_EX M_HASHTYPE_HASH(5) /* IPv6 2-tuple + * ext hdrs */ #define M_HASHTYPE_RSS_TCP_IPV6_EX M_HASHTYPE_HASH(6) /* TCPv6 4-tuple + * ext hdrs */ #define M_HASHTYPE_RSS_UDP_IPV4 M_HASHTYPE_HASH(7) /* IPv4 UDP 4-tuple*/ #define M_HASHTYPE_RSS_UDP_IPV6 M_HASHTYPE_HASH(9) /* IPv6 UDP 4-tuple*/ #define M_HASHTYPE_RSS_UDP_IPV6_EX M_HASHTYPE_HASH(10)/* IPv6 UDP 4-tuple + * ext hdrs */ #define M_HASHTYPE_OPAQUE 63 /* ordering, not affinity */ #define M_HASHTYPE_OPAQUE_HASH M_HASHTYPE_HASH(M_HASHTYPE_OPAQUE) /* ordering+hash, not affinity*/ #define M_HASHTYPE_CLEAR(m) ((m)->m_pkthdr.rsstype = 0) #define M_HASHTYPE_GET(m) ((m)->m_pkthdr.rsstype) #define M_HASHTYPE_SET(m, v) ((m)->m_pkthdr.rsstype = (v)) #define M_HASHTYPE_TEST(m, v) (M_HASHTYPE_GET(m) == (v)) #define M_HASHTYPE_ISHASH(m) (M_HASHTYPE_GET(m) & M_HASHTYPE_HASHPROP) /* * External mbuf storage buffer types. */ #define EXT_CLUSTER 1 /* mbuf cluster */ #define EXT_SFBUF 2 /* sendfile(2)'s sf_buf */ #define EXT_JUMBOP 3 /* jumbo cluster page sized */ #define EXT_JUMBO9 4 /* jumbo cluster 9216 bytes */ #define EXT_JUMBO16 5 /* jumbo cluster 16184 bytes */ #define EXT_PACKET 6 /* mbuf+cluster from packet zone */ #define EXT_MBUF 7 /* external mbuf reference */ #define EXT_RXRING 8 /* data in NIC receive ring */ #define EXT_PGS 9 /* array of unmapped pages */ #define EXT_VENDOR1 224 /* for vendor-internal use */ #define EXT_VENDOR2 225 /* for vendor-internal use */ #define EXT_VENDOR3 226 /* for vendor-internal use */ #define EXT_VENDOR4 227 /* for vendor-internal use */ #define EXT_EXP1 244 /* for experimental use */ #define EXT_EXP2 245 /* for experimental use */ #define EXT_EXP3 246 /* for experimental use */ #define EXT_EXP4 247 /* for experimental use */ #define EXT_NET_DRV 252 /* custom ext_buf provided by net driver(s) */ #define EXT_MOD_TYPE 253 /* custom module's ext_buf type */ #define EXT_DISPOSABLE 254 /* can throw this buffer away w/page flipping */ #define EXT_EXTREF 255 /* has externally maintained ext_cnt ptr */ /* * Flags for external mbuf buffer types. * NB: limited to the lower 24 bits. */ #define EXT_FLAG_EMBREF 0x000001 /* embedded ext_count */ #define EXT_FLAG_EXTREF 0x000002 /* external ext_cnt, notyet */ #define EXT_FLAG_NOFREE 0x000010 /* don't free mbuf to pool, notyet */ #define EXT_FLAG_VENDOR1 0x010000 /* These flags are vendor */ #define EXT_FLAG_VENDOR2 0x020000 /* or submodule specific, */ #define EXT_FLAG_VENDOR3 0x040000 /* not used by mbuf code. */ #define EXT_FLAG_VENDOR4 0x080000 /* Set/read by submodule. */ #define EXT_FLAG_EXP1 0x100000 /* for experimental use */ #define EXT_FLAG_EXP2 0x200000 /* for experimental use */ #define EXT_FLAG_EXP3 0x400000 /* for experimental use */ #define EXT_FLAG_EXP4 0x800000 /* for experimental use */ /* * EXT flag description for use with printf(9) %b identifier. */ #define EXT_FLAG_BITS \ "\20\1EXT_FLAG_EMBREF\2EXT_FLAG_EXTREF\5EXT_FLAG_NOFREE" \ "\21EXT_FLAG_VENDOR1\22EXT_FLAG_VENDOR2\23EXT_FLAG_VENDOR3" \ "\24EXT_FLAG_VENDOR4\25EXT_FLAG_EXP1\26EXT_FLAG_EXP2\27EXT_FLAG_EXP3" \ "\30EXT_FLAG_EXP4" #define MBUF_EXT_PGS_ASSERT(m) \ KASSERT((((m)->m_flags & M_EXT) != 0) && \ ((m)->m_ext.ext_type == EXT_PGS), \ ("%s: m %p !M_EXT or !EXT_PGS", __func__, m)) /* * Flags indicating checksum, segmentation and other offload work to be * done, or already done, by hardware or lower layers. It is split into * separate inbound and outbound flags. * * Outbound flags that are set by upper protocol layers requesting lower * layers, or ideally the hardware, to perform these offloading tasks. * For outbound packets this field and its flags can be directly tested * against ifnet if_hwassist. */ #define CSUM_IP 0x00000001 /* IP header checksum offload */ #define CSUM_IP_UDP 0x00000002 /* UDP checksum offload */ #define CSUM_IP_TCP 0x00000004 /* TCP checksum offload */ #define CSUM_IP_SCTP 0x00000008 /* SCTP checksum offload */ #define CSUM_IP_TSO 0x00000010 /* TCP segmentation offload */ #define CSUM_IP_ISCSI 0x00000020 /* iSCSI checksum offload */ #define CSUM_IP6_UDP 0x00000200 /* UDP checksum offload */ #define CSUM_IP6_TCP 0x00000400 /* TCP checksum offload */ #define CSUM_IP6_SCTP 0x00000800 /* SCTP checksum offload */ #define CSUM_IP6_TSO 0x00001000 /* TCP segmentation offload */ #define CSUM_IP6_ISCSI 0x00002000 /* iSCSI checksum offload */ /* Inbound checksum support where the checksum was verified by hardware. */ #define CSUM_L3_CALC 0x01000000 /* calculated layer 3 csum */ #define CSUM_L3_VALID 0x02000000 /* checksum is correct */ #define CSUM_L4_CALC 0x04000000 /* calculated layer 4 csum */ #define CSUM_L4_VALID 0x08000000 /* checksum is correct */ #define CSUM_L5_CALC 0x10000000 /* calculated layer 5 csum */ #define CSUM_L5_VALID 0x20000000 /* checksum is correct */ #define CSUM_COALESCED 0x40000000 /* contains merged segments */ #define CSUM_SND_TAG 0x80000000 /* Packet header has send tag */ /* * CSUM flag description for use with printf(9) %b identifier. */ #define CSUM_BITS \ "\20\1CSUM_IP\2CSUM_IP_UDP\3CSUM_IP_TCP\4CSUM_IP_SCTP\5CSUM_IP_TSO" \ "\6CSUM_IP_ISCSI" \ "\12CSUM_IP6_UDP\13CSUM_IP6_TCP\14CSUM_IP6_SCTP\15CSUM_IP6_TSO" \ "\16CSUM_IP6_ISCSI" \ "\31CSUM_L3_CALC\32CSUM_L3_VALID\33CSUM_L4_CALC\34CSUM_L4_VALID" \ "\35CSUM_L5_CALC\36CSUM_L5_VALID\37CSUM_COALESCED\40CSUM_SND_TAG" /* CSUM flags compatibility mappings. */ #define CSUM_IP_CHECKED CSUM_L3_CALC #define CSUM_IP_VALID CSUM_L3_VALID #define CSUM_DATA_VALID CSUM_L4_VALID #define CSUM_PSEUDO_HDR CSUM_L4_CALC #define CSUM_SCTP_VALID CSUM_L4_VALID #define CSUM_DELAY_DATA (CSUM_TCP|CSUM_UDP) #define CSUM_DELAY_IP CSUM_IP /* Only v4, no v6 IP hdr csum */ #define CSUM_DELAY_DATA_IPV6 (CSUM_TCP_IPV6|CSUM_UDP_IPV6) #define CSUM_DATA_VALID_IPV6 CSUM_DATA_VALID #define CSUM_TCP CSUM_IP_TCP #define CSUM_UDP CSUM_IP_UDP #define CSUM_SCTP CSUM_IP_SCTP #define CSUM_TSO (CSUM_IP_TSO|CSUM_IP6_TSO) #define CSUM_UDP_IPV6 CSUM_IP6_UDP #define CSUM_TCP_IPV6 CSUM_IP6_TCP #define CSUM_SCTP_IPV6 CSUM_IP6_SCTP /* * mbuf types describing the content of the mbuf (including external storage). */ #define MT_NOTMBUF 0 /* USED INTERNALLY ONLY! Object is not mbuf */ #define MT_DATA 1 /* dynamic (data) allocation */ #define MT_HEADER MT_DATA /* packet header, use M_PKTHDR instead */ #define MT_VENDOR1 4 /* for vendor-internal use */ #define MT_VENDOR2 5 /* for vendor-internal use */ #define MT_VENDOR3 6 /* for vendor-internal use */ #define MT_VENDOR4 7 /* for vendor-internal use */ #define MT_SONAME 8 /* socket name */ #define MT_EXP1 9 /* for experimental use */ #define MT_EXP2 10 /* for experimental use */ #define MT_EXP3 11 /* for experimental use */ #define MT_EXP4 12 /* for experimental use */ #define MT_CONTROL 14 /* extra-data protocol message */ #define MT_EXTCONTROL 15 /* control message with externalized contents */ #define MT_OOBDATA 16 /* expedited data */ #define MT_NOINIT 255 /* Not a type but a flag to allocate a non-initialized mbuf */ /* * String names of mbuf-related UMA(9) and malloc(9) types. Exposed to * !_KERNEL so that monitoring tools can look up the zones with * libmemstat(3). */ #define MBUF_MEM_NAME "mbuf" #define MBUF_CLUSTER_MEM_NAME "mbuf_cluster" #define MBUF_PACKET_MEM_NAME "mbuf_packet" #define MBUF_JUMBOP_MEM_NAME "mbuf_jumbo_page" #define MBUF_JUMBO9_MEM_NAME "mbuf_jumbo_9k" #define MBUF_JUMBO16_MEM_NAME "mbuf_jumbo_16k" #define MBUF_TAG_MEM_NAME "mbuf_tag" #define MBUF_EXTREFCNT_MEM_NAME "mbuf_ext_refcnt" #define MBUF_EXTPGS_MEM_NAME "mbuf_extpgs" #ifdef _KERNEL #ifdef WITNESS #define MBUF_CHECKSLEEP(how) do { \ if (how == M_WAITOK) \ WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, \ "Sleeping in \"%s\"", __func__); \ } while (0) #else #define MBUF_CHECKSLEEP(how) #endif /* * Network buffer allocation API * * The rest of it is defined in kern/kern_mbuf.c */ extern uma_zone_t zone_mbuf; extern uma_zone_t zone_clust; extern uma_zone_t zone_pack; extern uma_zone_t zone_jumbop; extern uma_zone_t zone_jumbo9; extern uma_zone_t zone_jumbo16; extern uma_zone_t zone_extpgs; void mb_dupcl(struct mbuf *, struct mbuf *); void mb_free_ext(struct mbuf *); void mb_free_mext_pgs(struct mbuf *); struct mbuf *mb_alloc_ext_pgs(int, bool, m_ext_free_t); int mb_unmapped_compress(struct mbuf *m); struct mbuf *mb_unmapped_to_ext(struct mbuf *m); void mb_free_notready(struct mbuf *m, int count); void m_adj(struct mbuf *, int); int m_apply(struct mbuf *, int, int, int (*)(void *, void *, u_int), void *); int m_append(struct mbuf *, int, c_caddr_t); void m_cat(struct mbuf *, struct mbuf *); void m_catpkt(struct mbuf *, struct mbuf *); int m_clget(struct mbuf *m, int how); void *m_cljget(struct mbuf *m, int how, int size); struct mbuf *m_collapse(struct mbuf *, int, int); void m_copyback(struct mbuf *, int, int, c_caddr_t); void m_copydata(const struct mbuf *, int, int, caddr_t); struct mbuf *m_copym(struct mbuf *, int, int, int); struct mbuf *m_copypacket(struct mbuf *, int); void m_copy_pkthdr(struct mbuf *, struct mbuf *); struct mbuf *m_copyup(struct mbuf *, int, int); struct mbuf *m_defrag(struct mbuf *, int); void m_demote_pkthdr(struct mbuf *); void m_demote(struct mbuf *, int, int); struct mbuf *m_devget(char *, int, int, struct ifnet *, void (*)(char *, caddr_t, u_int)); void m_dispose_extcontrolm(struct mbuf *m); struct mbuf *m_dup(const struct mbuf *, int); int m_dup_pkthdr(struct mbuf *, const struct mbuf *, int); void m_extadd(struct mbuf *, char *, u_int, m_ext_free_t, void *, void *, int, int); u_int m_fixhdr(struct mbuf *); struct mbuf *m_fragment(struct mbuf *, int, int); void m_freem(struct mbuf *); struct mbuf *m_get2(int, int, short, int); struct mbuf *m_getjcl(int, short, int, int); struct mbuf *m_getm2(struct mbuf *, int, int, short, int); struct mbuf *m_getptr(struct mbuf *, int, int *); u_int m_length(struct mbuf *, struct mbuf **); int m_mbuftouio(struct uio *, const struct mbuf *, int); int m_unmappedtouio(const struct mbuf *, int, struct uio *, int); void m_move_pkthdr(struct mbuf *, struct mbuf *); int m_pkthdr_init(struct mbuf *, int); struct mbuf *m_prepend(struct mbuf *, int, int); void m_print(const struct mbuf *, int); struct mbuf *m_pulldown(struct mbuf *, int, int, int *); struct mbuf *m_pullup(struct mbuf *, int); int m_sanity(struct mbuf *, int); struct mbuf *m_split(struct mbuf *, int, int); struct mbuf *m_uiotombuf(struct uio *, int, int, int, int); struct mbuf *m_unshare(struct mbuf *, int); void m_snd_tag_init(struct m_snd_tag *, struct ifnet *); void m_snd_tag_destroy(struct m_snd_tag *); static __inline int m_gettype(int size) { int type; switch (size) { case MSIZE: type = EXT_MBUF; break; case MCLBYTES: type = EXT_CLUSTER; break; #if MJUMPAGESIZE != MCLBYTES case MJUMPAGESIZE: type = EXT_JUMBOP; break; #endif case MJUM9BYTES: type = EXT_JUMBO9; break; case MJUM16BYTES: type = EXT_JUMBO16; break; default: panic("%s: invalid cluster size %d", __func__, size); } return (type); } /* * Associated an external reference counted buffer with an mbuf. */ static __inline void m_extaddref(struct mbuf *m, char *buf, u_int size, u_int *ref_cnt, m_ext_free_t freef, void *arg1, void *arg2) { KASSERT(ref_cnt != NULL, ("%s: ref_cnt not provided", __func__)); atomic_add_int(ref_cnt, 1); m->m_flags |= M_EXT; m->m_ext.ext_buf = buf; m->m_ext.ext_cnt = ref_cnt; m->m_data = m->m_ext.ext_buf; m->m_ext.ext_size = size; m->m_ext.ext_free = freef; m->m_ext.ext_arg1 = arg1; m->m_ext.ext_arg2 = arg2; m->m_ext.ext_type = EXT_EXTREF; m->m_ext.ext_flags = 0; } static __inline uma_zone_t m_getzone(int size) { uma_zone_t zone; switch (size) { case MCLBYTES: zone = zone_clust; break; #if MJUMPAGESIZE != MCLBYTES case MJUMPAGESIZE: zone = zone_jumbop; break; #endif case MJUM9BYTES: zone = zone_jumbo9; break; case MJUM16BYTES: zone = zone_jumbo16; break; default: panic("%s: invalid cluster size %d", __func__, size); } return (zone); } /* * Initialize an mbuf with linear storage. * * Inline because the consumer text overhead will be roughly the same to * initialize or call a function with this many parameters and M_PKTHDR * should go away with constant propagation for !MGETHDR. */ static __inline int m_init(struct mbuf *m, int how, short type, int flags) { int error; m->m_next = NULL; m->m_nextpkt = NULL; m->m_data = m->m_dat; m->m_len = 0; m->m_flags = flags; m->m_type = type; if (flags & M_PKTHDR) error = m_pkthdr_init(m, how); else error = 0; MBUF_PROBE5(m__init, m, how, type, flags, error); return (error); } static __inline struct mbuf * m_get(int how, short type) { struct mbuf *m; struct mb_args args; args.flags = 0; args.type = type; m = uma_zalloc_arg(zone_mbuf, &args, how); MBUF_PROBE3(m__get, how, type, m); return (m); } static __inline struct mbuf * m_gethdr(int how, short type) { struct mbuf *m; struct mb_args args; args.flags = M_PKTHDR; args.type = type; m = uma_zalloc_arg(zone_mbuf, &args, how); MBUF_PROBE3(m__gethdr, how, type, m); return (m); } static __inline struct mbuf * m_getcl(int how, short type, int flags) { struct mbuf *m; struct mb_args args; args.flags = flags; args.type = type; m = uma_zalloc_arg(zone_pack, &args, how); MBUF_PROBE4(m__getcl, how, type, flags, m); return (m); } /* * XXX: m_cljset() is a dangerous API. One must attach only a new, * unreferenced cluster to an mbuf(9). It is not possible to assert * that, so care can be taken only by users of the API. */ static __inline void m_cljset(struct mbuf *m, void *cl, int type) { int size; switch (type) { case EXT_CLUSTER: size = MCLBYTES; break; #if MJUMPAGESIZE != MCLBYTES case EXT_JUMBOP: size = MJUMPAGESIZE; break; #endif case EXT_JUMBO9: size = MJUM9BYTES; break; case EXT_JUMBO16: size = MJUM16BYTES; break; default: panic("%s: unknown cluster type %d", __func__, type); break; } m->m_data = m->m_ext.ext_buf = cl; m->m_ext.ext_free = m->m_ext.ext_arg1 = m->m_ext.ext_arg2 = NULL; m->m_ext.ext_size = size; m->m_ext.ext_type = type; m->m_ext.ext_flags = EXT_FLAG_EMBREF; m->m_ext.ext_count = 1; m->m_flags |= M_EXT; MBUF_PROBE3(m__cljset, m, cl, type); } static __inline void m_chtype(struct mbuf *m, short new_type) { m->m_type = new_type; } static __inline void m_clrprotoflags(struct mbuf *m) { while (m) { m->m_flags &= ~M_PROTOFLAGS; m = m->m_next; } } static __inline struct mbuf * m_last(struct mbuf *m) { while (m->m_next) m = m->m_next; return (m); } static inline u_int m_extrefcnt(struct mbuf *m) { KASSERT(m->m_flags & M_EXT, ("%s: M_EXT missing", __func__)); return ((m->m_ext.ext_flags & EXT_FLAG_EMBREF) ? m->m_ext.ext_count : *m->m_ext.ext_cnt); } /* * mbuf, cluster, and external object allocation macros (for compatibility * purposes). */ #define M_MOVE_PKTHDR(to, from) m_move_pkthdr((to), (from)) #define MGET(m, how, type) ((m) = m_get((how), (type))) #define MGETHDR(m, how, type) ((m) = m_gethdr((how), (type))) #define MCLGET(m, how) m_clget((m), (how)) #define MEXTADD(m, buf, size, free, arg1, arg2, flags, type) \ m_extadd((m), (char *)(buf), (size), (free), (arg1), (arg2), \ (flags), (type)) #define m_getm(m, len, how, type) \ m_getm2((m), (len), (how), (type), M_PKTHDR) /* * Evaluate TRUE if it's safe to write to the mbuf m's data region (this can * be both the local data payload, or an external buffer area, depending on * whether M_EXT is set). */ #define M_WRITABLE(m) (((m)->m_flags & (M_RDONLY | M_NOMAP)) == 0 && \ (!(((m)->m_flags & M_EXT)) || \ (m_extrefcnt(m) == 1))) /* Check if the supplied mbuf has a packet header, or else panic. */ #define M_ASSERTPKTHDR(m) \ KASSERT((m) != NULL && (m)->m_flags & M_PKTHDR, \ ("%s: no mbuf packet header!", __func__)) /* * Ensure that the supplied mbuf is a valid, non-free mbuf. * * XXX: Broken at the moment. Need some UMA magic to make it work again. */ #define M_ASSERTVALID(m) \ KASSERT((((struct mbuf *)m)->m_flags & 0) == 0, \ ("%s: attempted use of a free mbuf!", __func__)) /* * Return the address of the start of the buffer associated with an mbuf, * handling external storage, packet-header mbufs, and regular data mbufs. */ #define M_START(m) \ (((m)->m_flags & M_NOMAP) ? NULL : \ ((m)->m_flags & M_EXT) ? (m)->m_ext.ext_buf : \ ((m)->m_flags & M_PKTHDR) ? &(m)->m_pktdat[0] : \ &(m)->m_dat[0]) /* * Return the size of the buffer associated with an mbuf, handling external * storage, packet-header mbufs, and regular data mbufs. */ #define M_SIZE(m) \ (((m)->m_flags & M_EXT) ? (m)->m_ext.ext_size : \ ((m)->m_flags & M_PKTHDR) ? MHLEN : \ MLEN) /* * Set the m_data pointer of a newly allocated mbuf to place an object of the * specified size at the end of the mbuf, longword aligned. * * NB: Historically, we had M_ALIGN(), MH_ALIGN(), and MEXT_ALIGN() as * separate macros, each asserting that it was called at the proper moment. * This required callers to themselves test the storage type and call the * right one. Rather than require callers to be aware of those layout * decisions, we centralize here. */ static __inline void m_align(struct mbuf *m, int len) { #ifdef INVARIANTS const char *msg = "%s: not a virgin mbuf"; #endif int adjust; KASSERT(m->m_data == M_START(m), (msg, __func__)); adjust = M_SIZE(m) - len; m->m_data += adjust &~ (sizeof(long)-1); } #define M_ALIGN(m, len) m_align(m, len) #define MH_ALIGN(m, len) m_align(m, len) #define MEXT_ALIGN(m, len) m_align(m, len) /* * Compute the amount of space available before the current start of data in * an mbuf. * * The M_WRITABLE() is a temporary, conservative safety measure: the burden * of checking writability of the mbuf data area rests solely with the caller. * * NB: In previous versions, M_LEADINGSPACE() would only check M_WRITABLE() * for mbufs with external storage. We now allow mbuf-embedded data to be * read-only as well. */ #define M_LEADINGSPACE(m) \ (M_WRITABLE(m) ? ((m)->m_data - M_START(m)) : 0) /* * Compute the amount of space available after the end of data in an mbuf. * * The M_WRITABLE() is a temporary, conservative safety measure: the burden * of checking writability of the mbuf data area rests solely with the caller. * * NB: In previous versions, M_TRAILINGSPACE() would only check M_WRITABLE() * for mbufs with external storage. We now allow mbuf-embedded data to be * read-only as well. */ #define M_TRAILINGSPACE(m) \ (M_WRITABLE(m) ? \ ((M_START(m) + M_SIZE(m)) - ((m)->m_data + (m)->m_len)) : 0) /* * Arrange to prepend space of size plen to mbuf m. If a new mbuf must be * allocated, how specifies whether to wait. If the allocation fails, the * original mbuf chain is freed and m is set to NULL. */ #define M_PREPEND(m, plen, how) do { \ struct mbuf **_mmp = &(m); \ struct mbuf *_mm = *_mmp; \ int _mplen = (plen); \ int __mhow = (how); \ \ MBUF_CHECKSLEEP(how); \ if (M_LEADINGSPACE(_mm) >= _mplen) { \ _mm->m_data -= _mplen; \ _mm->m_len += _mplen; \ } else \ _mm = m_prepend(_mm, _mplen, __mhow); \ if (_mm != NULL && _mm->m_flags & M_PKTHDR) \ _mm->m_pkthdr.len += _mplen; \ *_mmp = _mm; \ } while (0) /* * Change mbuf to new type. This is a relatively expensive operation and * should be avoided. */ #define MCHTYPE(m, t) m_chtype((m), (t)) /* Return the rcvif of a packet header. */ static __inline struct ifnet * m_rcvif(struct mbuf *m) { M_ASSERTPKTHDR(m); if (m->m_pkthdr.csum_flags & CSUM_SND_TAG) return (NULL); return (m->m_pkthdr.rcvif); } /* Length to m_copy to copy all. */ #define M_COPYALL 1000000000 extern int max_datalen; /* MHLEN - max_hdr */ extern int max_hdr; /* Largest link + protocol header */ extern int max_linkhdr; /* Largest link-level header */ extern int max_protohdr; /* Largest protocol header */ extern int nmbclusters; /* Maximum number of clusters */ extern bool mb_use_ext_pgs; /* Use ext_pgs for sendfile */ /*- * Network packets may have annotations attached by affixing a list of * "packet tags" to the pkthdr structure. Packet tags are dynamically * allocated semi-opaque data structures that have a fixed header * (struct m_tag) that specifies the size of the memory block and a * pair that identifies it. The cookie is a 32-bit unique * unsigned value used to identify a module or ABI. By convention this value * is chosen as the date+time that the module is created, expressed as the * number of seconds since the epoch (e.g., using date -u +'%s'). The type * value is an ABI/module-specific value that identifies a particular * annotation and is private to the module. For compatibility with systems * like OpenBSD that define packet tags w/o an ABI/module cookie, the value * PACKET_ABI_COMPAT is used to implement m_tag_get and m_tag_find * compatibility shim functions and several tag types are defined below. * Users that do not require compatibility should use a private cookie value * so that packet tag-related definitions can be maintained privately. * * Note that the packet tag returned by m_tag_alloc has the default memory * alignment implemented by malloc. To reference private data one can use a * construct like: * * struct m_tag *mtag = m_tag_alloc(...); * struct foo *p = (struct foo *)(mtag+1); * * if the alignment of struct m_tag is sufficient for referencing members of * struct foo. Otherwise it is necessary to embed struct m_tag within the * private data structure to insure proper alignment; e.g., * * struct foo { * struct m_tag tag; * ... * }; * struct foo *p = (struct foo *) m_tag_alloc(...); * struct m_tag *mtag = &p->tag; */ /* * Persistent tags stay with an mbuf until the mbuf is reclaimed. Otherwise * tags are expected to ``vanish'' when they pass through a network * interface. For most interfaces this happens normally as the tags are * reclaimed when the mbuf is free'd. However in some special cases * reclaiming must be done manually. An example is packets that pass through * the loopback interface. Also, one must be careful to do this when * ``turning around'' packets (e.g., icmp_reflect). * * To mark a tag persistent bit-or this flag in when defining the tag id. * The tag will then be treated as described above. */ #define MTAG_PERSISTENT 0x800 #define PACKET_TAG_NONE 0 /* Nadda */ /* Packet tags for use with PACKET_ABI_COMPAT. */ #define PACKET_TAG_IPSEC_IN_DONE 1 /* IPsec applied, in */ #define PACKET_TAG_IPSEC_OUT_DONE 2 /* IPsec applied, out */ #define PACKET_TAG_IPSEC_IN_CRYPTO_DONE 3 /* NIC IPsec crypto done */ #define PACKET_TAG_IPSEC_OUT_CRYPTO_NEEDED 4 /* NIC IPsec crypto req'ed */ #define PACKET_TAG_IPSEC_IN_COULD_DO_CRYPTO 5 /* NIC notifies IPsec */ #define PACKET_TAG_IPSEC_PENDING_TDB 6 /* Reminder to do IPsec */ #define PACKET_TAG_BRIDGE 7 /* Bridge processing done */ #define PACKET_TAG_GIF 8 /* GIF processing done */ #define PACKET_TAG_GRE 9 /* GRE processing done */ #define PACKET_TAG_IN_PACKET_CHECKSUM 10 /* NIC checksumming done */ #define PACKET_TAG_ENCAP 11 /* Encap. processing */ #define PACKET_TAG_IPSEC_SOCKET 12 /* IPSEC socket ref */ #define PACKET_TAG_IPSEC_HISTORY 13 /* IPSEC history */ #define PACKET_TAG_IPV6_INPUT 14 /* IPV6 input processing */ #define PACKET_TAG_DUMMYNET 15 /* dummynet info */ #define PACKET_TAG_DIVERT 17 /* divert info */ #define PACKET_TAG_IPFORWARD 18 /* ipforward info */ #define PACKET_TAG_MACLABEL (19 | MTAG_PERSISTENT) /* MAC label */ #define PACKET_TAG_PF (21 | MTAG_PERSISTENT) /* PF/ALTQ information */ #define PACKET_TAG_RTSOCKFAM 25 /* rtsock sa family */ #define PACKET_TAG_IPOPTIONS 27 /* Saved IP options */ #define PACKET_TAG_CARP 28 /* CARP info */ #define PACKET_TAG_IPSEC_NAT_T_PORTS 29 /* two uint16_t */ #define PACKET_TAG_ND_OUTGOING 30 /* ND outgoing */ /* Specific cookies and tags. */ /* Packet tag routines. */ struct m_tag *m_tag_alloc(u_int32_t, int, int, int); void m_tag_delete(struct mbuf *, struct m_tag *); void m_tag_delete_chain(struct mbuf *, struct m_tag *); void m_tag_free_default(struct m_tag *); struct m_tag *m_tag_locate(struct mbuf *, u_int32_t, int, struct m_tag *); struct m_tag *m_tag_copy(struct m_tag *, int); int m_tag_copy_chain(struct mbuf *, const struct mbuf *, int); void m_tag_delete_nonpersistent(struct mbuf *); /* * Initialize the list of tags associated with an mbuf. */ static __inline void m_tag_init(struct mbuf *m) { SLIST_INIT(&m->m_pkthdr.tags); } /* * Set up the contents of a tag. Note that this does not fill in the free * method; the caller is expected to do that. * * XXX probably should be called m_tag_init, but that was already taken. */ static __inline void m_tag_setup(struct m_tag *t, u_int32_t cookie, int type, int len) { t->m_tag_id = type; t->m_tag_len = len; t->m_tag_cookie = cookie; } /* * Reclaim resources associated with a tag. */ static __inline void m_tag_free(struct m_tag *t) { (*t->m_tag_free)(t); } /* * Return the first tag associated with an mbuf. */ static __inline struct m_tag * m_tag_first(struct mbuf *m) { return (SLIST_FIRST(&m->m_pkthdr.tags)); } /* * Return the next tag in the list of tags associated with an mbuf. */ static __inline struct m_tag * m_tag_next(struct mbuf *m __unused, struct m_tag *t) { return (SLIST_NEXT(t, m_tag_link)); } /* * Prepend a tag to the list of tags associated with an mbuf. */ static __inline void m_tag_prepend(struct mbuf *m, struct m_tag *t) { SLIST_INSERT_HEAD(&m->m_pkthdr.tags, t, m_tag_link); } /* * Unlink a tag from the list of tags associated with an mbuf. */ static __inline void m_tag_unlink(struct mbuf *m, struct m_tag *t) { SLIST_REMOVE(&m->m_pkthdr.tags, t, m_tag, m_tag_link); } /* These are for OpenBSD compatibility. */ #define MTAG_ABI_COMPAT 0 /* compatibility ABI */ static __inline struct m_tag * m_tag_get(int type, int length, int wait) { return (m_tag_alloc(MTAG_ABI_COMPAT, type, length, wait)); } static __inline struct m_tag * m_tag_find(struct mbuf *m, int type, struct m_tag *start) { return (SLIST_EMPTY(&m->m_pkthdr.tags) ? (struct m_tag *)NULL : m_tag_locate(m, MTAG_ABI_COMPAT, type, start)); } static inline struct m_snd_tag * m_snd_tag_ref(struct m_snd_tag *mst) { refcount_acquire(&mst->refcount); return (mst); } static inline void m_snd_tag_rele(struct m_snd_tag *mst) { if (refcount_release(&mst->refcount)) m_snd_tag_destroy(mst); } static __inline struct mbuf * m_free(struct mbuf *m) { struct mbuf *n = m->m_next; MBUF_PROBE1(m__free, m); if ((m->m_flags & (M_PKTHDR|M_NOFREE)) == (M_PKTHDR|M_NOFREE)) m_tag_delete_chain(m, NULL); if (m->m_flags & M_PKTHDR && m->m_pkthdr.csum_flags & CSUM_SND_TAG) m_snd_tag_rele(m->m_pkthdr.snd_tag); if (m->m_flags & M_EXT) mb_free_ext(m); else if ((m->m_flags & M_NOFREE) == 0) uma_zfree(zone_mbuf, m); return (n); } static __inline int rt_m_getfib(struct mbuf *m) { KASSERT(m->m_flags & M_PKTHDR , ("Attempt to get FIB from non header mbuf.")); return (m->m_pkthdr.fibnum); } #define M_GETFIB(_m) rt_m_getfib(_m) #define M_SETFIB(_m, _fib) do { \ KASSERT((_m)->m_flags & M_PKTHDR, ("Attempt to set FIB on non header mbuf.")); \ ((_m)->m_pkthdr.fibnum) = (_fib); \ } while (0) /* flags passed as first argument for "m_ether_tcpip_hash()" */ #define MBUF_HASHFLAG_L2 (1 << 2) #define MBUF_HASHFLAG_L3 (1 << 3) #define MBUF_HASHFLAG_L4 (1 << 4) /* mbuf hashing helper routines */ uint32_t m_ether_tcpip_hash_init(void); uint32_t m_ether_tcpip_hash(const uint32_t, const struct mbuf *, const uint32_t); #ifdef MBUF_PROFILING void m_profile(struct mbuf *m); #define M_PROFILE(m) m_profile(m) #else #define M_PROFILE(m) #endif struct mbufq { STAILQ_HEAD(, mbuf) mq_head; int mq_len; int mq_maxlen; }; static inline void mbufq_init(struct mbufq *mq, int maxlen) { STAILQ_INIT(&mq->mq_head); mq->mq_maxlen = maxlen; mq->mq_len = 0; } static inline struct mbuf * mbufq_flush(struct mbufq *mq) { struct mbuf *m; m = STAILQ_FIRST(&mq->mq_head); STAILQ_INIT(&mq->mq_head); mq->mq_len = 0; return (m); } static inline void mbufq_drain(struct mbufq *mq) { struct mbuf *m, *n; n = mbufq_flush(mq); while ((m = n) != NULL) { n = STAILQ_NEXT(m, m_stailqpkt); m_freem(m); } } static inline struct mbuf * mbufq_first(const struct mbufq *mq) { return (STAILQ_FIRST(&mq->mq_head)); } static inline struct mbuf * mbufq_last(const struct mbufq *mq) { return (STAILQ_LAST(&mq->mq_head, mbuf, m_stailqpkt)); } static inline int mbufq_full(const struct mbufq *mq) { return (mq->mq_maxlen > 0 && mq->mq_len >= mq->mq_maxlen); } static inline int mbufq_len(const struct mbufq *mq) { return (mq->mq_len); } static inline int mbufq_enqueue(struct mbufq *mq, struct mbuf *m) { if (mbufq_full(mq)) return (ENOBUFS); STAILQ_INSERT_TAIL(&mq->mq_head, m, m_stailqpkt); mq->mq_len++; return (0); } static inline struct mbuf * mbufq_dequeue(struct mbufq *mq) { struct mbuf *m; m = STAILQ_FIRST(&mq->mq_head); if (m) { STAILQ_REMOVE_HEAD(&mq->mq_head, m_stailqpkt); m->m_nextpkt = NULL; mq->mq_len--; } return (m); } static inline void mbufq_prepend(struct mbufq *mq, struct mbuf *m) { STAILQ_INSERT_HEAD(&mq->mq_head, m, m_stailqpkt); mq->mq_len++; } /* * Note: this doesn't enforce the maximum list size for dst. */ static inline void mbufq_concat(struct mbufq *mq_dst, struct mbufq *mq_src) { mq_dst->mq_len += mq_src->mq_len; STAILQ_CONCAT(&mq_dst->mq_head, &mq_src->mq_head); mq_src->mq_len = 0; } #ifdef _SYS_TIMESPEC_H_ static inline void mbuf_tstmp2timespec(struct mbuf *m, struct timespec *ts) { KASSERT((m->m_flags & M_PKTHDR) != 0, ("mbuf %p no M_PKTHDR", m)); KASSERT((m->m_flags & (M_TSTMP|M_TSTMP_LRO)) != 0, ("mbuf %p no M_TSTMP or M_TSTMP_LRO", m)); ts->tv_sec = m->m_pkthdr.rcv_tstmp / 1000000000; ts->tv_nsec = m->m_pkthdr.rcv_tstmp % 1000000000; } #endif #ifdef NETDUMP /* Invoked from the netdump client code. */ void netdump_mbuf_drain(void); void netdump_mbuf_dump(void); void netdump_mbuf_reinit(int nmbuf, int nclust, int clsize); #endif static inline bool mbuf_has_tls_session(struct mbuf *m) { if (m->m_flags & M_NOMAP) { MBUF_EXT_PGS_ASSERT(m); if (m->m_ext.ext_pgs->tls != NULL) { return (true); } } return (false); } #endif /* _KERNEL */ #endif /* !_SYS_MBUF_H_ */