Date: Thu, 23 Jul 2020 23:48:19 +0000 (UTC) From: John Baldwin <jhb@FreeBSD.org> To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-head@freebsd.org Subject: svn commit: r363464 - in head: share/man/man4 sys/kern sys/opencrypto sys/sys Message-ID: <202007232348.06NNmJ6o063845@repo.freebsd.org>
next in thread | raw e-mail | index | archive | help
Author: jhb Date: Thu Jul 23 23:48:18 2020 New Revision: 363464 URL: https://svnweb.freebsd.org/changeset/base/363464 Log: Add support for KTLS RX via software decryption. Allow TLS records to be decrypted in the kernel after being received by a NIC. At a high level this is somewhat similar to software KTLS for the transmit path except in reverse. Protocols enqueue mbufs containing encrypted TLS records (or portions of records) into the tail of a socket buffer and the KTLS layer decrypts those records before returning them to userland applications. However, there is an important difference: - In the transmit case, the socket buffer is always a single "record" holding a chain of mbufs. Not-yet-encrypted mbufs are marked not ready (M_NOTREADY) and released to protocols for transmit by marking mbufs ready once their data is encrypted. - In the receive case, incoming (encrypted) data appended to the socket buffer is still a single stream of data from the protocol, but decrypted TLS records are stored as separate records in the socket buffer and read individually via recvmsg(). Initially I tried to make this work by marking incoming mbufs as M_NOTREADY, but there didn't seemed to be a non-gross way to deal with picking a portion of the mbuf chain and turning it into a new record in the socket buffer after decrypting the TLS record it contained (along with prepending a control message). Also, such mbufs would also need to be "pinned" in some way while they are being decrypted such that a concurrent sbcut() wouldn't free them out from under the thread performing decryption. As such, I settled on the following solution: - Socket buffers now contain an additional chain of mbufs (sb_mtls, sb_mtlstail, and sb_tlscc) containing encrypted mbufs appended by the protocol layer. These mbufs are still marked M_NOTREADY, but soreceive*() generally don't know about them (except that they will block waiting for data to be decrypted for a blocking read). - Each time a new mbuf is appended to this TLS mbuf chain, the socket buffer peeks at the TLS record header at the head of the chain to determine the encrypted record's length. If enough data is queued for the TLS record, the socket is placed on a per-CPU TLS workqueue (reusing the existing KTLS workqueues and worker threads). - The worker thread loops over the TLS mbuf chain decrypting records until it runs out of data. Each record is detached from the TLS mbuf chain while it is being decrypted to keep the mbufs "pinned". However, a new sb_dtlscc field tracks the character count of the detached record and sbcut()/sbdrop() is updated to account for the detached record. After the record is decrypted, the worker thread first checks to see if sbcut() dropped the record. If so, it is freed (can happen when a socket is closed with pending data). Otherwise, the header and trailer are stripped from the original mbufs, a control message is created holding the decrypted TLS header, and the decrypted TLS record is appended to the "normal" socket buffer chain. (Side note: the SBCHECK() infrastucture was very useful as I was able to add assertions there about the TLS chain that caught several bugs during development.) Tested by: rmacklem (various versions) Relnotes: yes Sponsored by: Chelsio Communications Differential Revision: https://reviews.freebsd.org/D24628 Modified: head/share/man/man4/tcp.4 head/sys/kern/uipc_ktls.c head/sys/kern/uipc_sockbuf.c head/sys/kern/uipc_socket.c head/sys/opencrypto/ktls_ocf.c head/sys/sys/ktls.h head/sys/sys/sockbuf.h head/sys/sys/socketvar.h Modified: head/share/man/man4/tcp.4 ============================================================================== --- head/share/man/man4/tcp.4 Thu Jul 23 23:29:50 2020 (r363463) +++ head/share/man/man4/tcp.4 Thu Jul 23 23:48:18 2020 (r363464) @@ -34,7 +34,7 @@ .\" From: @(#)tcp.4 8.1 (Berkeley) 6/5/93 .\" $FreeBSD$ .\" -.Dd April 27, 2020 +.Dd July 23, 2020 .Dt TCP 4 .Os .Sh NAME @@ -356,10 +356,22 @@ control message along with the decrypted payload. The control message contains a .Vt struct tls_get_record which includes fields from the TLS record header. -If a corrupted TLS record is received, +If an invalid or corrupted TLS record is received, recvmsg 2 -will fail with -.Dv EBADMSG . +will fail with one of the following errors: +.Bl -tag -width Er +.It Bq Er EINVAL +The version fields in a TLS record's header did not match the version required +by the +.Vt struct tls_so_enable +structure used to enable in-kernel TLS. +.It Bq Er EMSGSIZE +A TLS record's length was either too small or too large. +.It Bq Er EMSGSIZE +The connection was closed after sending a truncated TLS record. +.It Bq Er EBADMSG +The TLS record failed to match the included authentication tag. +.El .Pp At present, only a single receive key may be set on a socket. As such, users of this option must disable rekeying. Modified: head/sys/kern/uipc_ktls.c ============================================================================== --- head/sys/kern/uipc_ktls.c Thu Jul 23 23:29:50 2020 (r363463) +++ head/sys/kern/uipc_ktls.c Thu Jul 23 23:48:18 2020 (r363464) @@ -78,7 +78,8 @@ __FBSDID("$FreeBSD$"); struct ktls_wq { struct mtx mtx; - STAILQ_HEAD(, mbuf) head; + STAILQ_HEAD(, mbuf) m_head; + STAILQ_HEAD(, socket) so_head; bool running; } __aligned(CACHE_LINE_SIZE); @@ -130,10 +131,16 @@ 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_cnt_tx_queued; +SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD, + &ktls_cnt_tx_queued, + "Number of TLS records in queue to tasks for SW encryption"); +static counter_u64_t ktls_cnt_rx_queued; +SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD, + &ktls_cnt_rx_queued, + "Number of TLS sockets in queue to tasks for SW decryption"); + static counter_u64_t ktls_offload_total; SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total, CTLFLAG_RD, &ktls_offload_total, @@ -148,6 +155,10 @@ 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_corrupted_records; +SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD, + &ktls_offload_corrupted_records, "Total corrupted TLS records received"); + 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"); @@ -333,10 +344,12 @@ ktls_init(void *dummy __unused) int error, i; ktls_tasks_active = counter_u64_alloc(M_WAITOK); - ktls_cnt_on = counter_u64_alloc(M_WAITOK); + ktls_cnt_tx_queued = counter_u64_alloc(M_WAITOK); + ktls_cnt_rx_queued = 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_corrupted_records = 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); @@ -369,7 +382,8 @@ ktls_init(void *dummy __unused) * work queue for each CPU. */ CPU_FOREACH(i) { - STAILQ_INIT(&ktls_wq[i].head); + STAILQ_INIT(&ktls_wq[i].m_head); + STAILQ_INIT(&ktls_wq[i].so_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); @@ -855,7 +869,7 @@ ktls_try_ifnet(struct socket *so, struct ktls_session } static int -ktls_try_sw(struct socket *so, struct ktls_session *tls) +ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction) { struct rm_priotracker prio; struct ktls_crypto_backend *be; @@ -870,7 +884,7 @@ ktls_try_sw(struct socket *so, struct ktls_session *tl if (ktls_allow_unload) rm_rlock(&ktls_backends_lock, &prio); LIST_FOREACH(be, &ktls_backends, next) { - if (be->try(so, tls) == 0) + if (be->try(so, tls, direction) == 0) break; KASSERT(tls->cipher == NULL, ("ktls backend leaked a cipher pointer")); @@ -896,6 +910,61 @@ ktls_try_sw(struct socket *so, struct ktls_session *tl return (0); } +/* + * KTLS RX stores data in the socket buffer as a list of TLS records, + * where each record is stored as a control message containg the TLS + * header followed by data mbufs containing the decrypted data. This + * is different from KTLS TX which always uses an mb_ext_pgs mbuf for + * both encrypted and decrypted data. TLS records decrypted by a NIC + * should be queued to the socket buffer as records, but encrypted + * data which needs to be decrypted by software arrives as a stream of + * regular mbufs which need to be converted. In addition, there may + * already be pending encrypted data in the socket buffer when KTLS RX + * is enabled. + * + * To manage not-yet-decrypted data for KTLS RX, the following scheme + * is used: + * + * - A single chain of NOTREADY mbufs is hung off of sb_mtls. + * + * - ktls_check_rx checks this chain of mbufs reading the TLS header + * from the first mbuf. Once all of the data for that TLS record is + * queued, the socket is queued to a worker thread. + * + * - The worker thread calls ktls_decrypt to decrypt TLS records in + * the TLS chain. Each TLS record is detached from the TLS chain, + * decrypted, and inserted into the regular socket buffer chain as + * record starting with a control message holding the TLS header and + * a chain of mbufs holding the encrypted data. + */ + +static void +sb_mark_notready(struct sockbuf *sb) +{ + struct mbuf *m; + + m = sb->sb_mb; + sb->sb_mtls = m; + sb->sb_mb = NULL; + sb->sb_mbtail = NULL; + sb->sb_lastrecord = NULL; + for (; m != NULL; m = m->m_next) { + KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL", + __func__)); + KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail", + __func__)); + KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len", + __func__)); + m->m_flags |= M_NOTREADY; + sb->sb_acc -= m->m_len; + sb->sb_tlscc += m->m_len; + sb->sb_mtlstail = m; + } + KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc, + ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc, + sb->sb_ccc)); +} + int ktls_enable_rx(struct socket *so, struct tls_enable *en) { @@ -924,16 +993,20 @@ ktls_enable_rx(struct socket *so, struct tls_enable *e if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) return (ENOTSUP); + /* TLS 1.3 is not yet supported. */ + if (en->tls_vmajor == TLS_MAJOR_VER_ONE && + en->tls_vminor == TLS_MINOR_VER_THREE) + return (ENOTSUP); + error = ktls_create_session(so, en, &tls); if (error) return (error); - /* TLS RX offload is only supported on TOE currently. */ #ifdef TCP_OFFLOAD error = ktls_try_toe(so, tls, KTLS_RX); -#else - error = EOPNOTSUPP; + if (error) #endif + error = ktls_try_sw(so, tls, KTLS_RX); if (error) { ktls_cleanup(tls); @@ -942,7 +1015,13 @@ ktls_enable_rx(struct socket *so, struct tls_enable *e /* Mark the socket as using TLS offload. */ SOCKBUF_LOCK(&so->so_rcv); + so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq); so->so_rcv.sb_tls_info = tls; + so->so_rcv.sb_flags |= SB_TLS_RX; + + /* Mark existing data as not ready until it can be decrypted. */ + sb_mark_notready(&so->so_rcv); + ktls_check_rx(&so->so_rcv); SOCKBUF_UNLOCK(&so->so_rcv); counter_u64_add(ktls_offload_total, 1); @@ -993,7 +1072,7 @@ ktls_enable_tx(struct socket *so, struct tls_enable *e #endif error = ktls_try_ifnet(so, tls, false); if (error) - error = ktls_try_sw(so, tls); + error = ktls_try_sw(so, tls, KTLS_TX); if (error) { ktls_cleanup(tls); @@ -1098,7 +1177,7 @@ ktls_set_tx_mode(struct socket *so, int mode) if (mode == TCP_TLS_MODE_IFNET) error = ktls_try_ifnet(so, tls_new, true); else - error = ktls_try_sw(so, tls_new); + error = ktls_try_sw(so, tls_new, KTLS_TX); if (error) { counter_u64_add(ktls_switch_failed, 1); ktls_free(tls_new); @@ -1422,6 +1501,345 @@ ktls_frame(struct mbuf *top, struct ktls_session *tls, } void +ktls_check_rx(struct sockbuf *sb) +{ + struct tls_record_layer hdr; + struct ktls_wq *wq; + struct socket *so; + bool running; + + SOCKBUF_LOCK_ASSERT(sb); + KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX", + __func__, sb)); + so = __containerof(sb, struct socket, so_rcv); + + if (sb->sb_flags & SB_TLS_RX_RUNNING) + return; + + /* Is there enough queued for a TLS header? */ + if (sb->sb_tlscc < sizeof(hdr)) { + if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0) + so->so_error = EMSGSIZE; + return; + } + + m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr); + + /* Is the entire record queued? */ + if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) { + if ((sb->sb_state & SBS_CANTRCVMORE) != 0) + so->so_error = EMSGSIZE; + return; + } + + sb->sb_flags |= SB_TLS_RX_RUNNING; + + soref(so); + wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index]; + mtx_lock(&wq->mtx); + STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list); + running = wq->running; + mtx_unlock(&wq->mtx); + if (!running) + wakeup(wq); + counter_u64_add(ktls_cnt_rx_queued, 1); +} + +static struct mbuf * +ktls_detach_record(struct sockbuf *sb, int len) +{ + struct mbuf *m, *n, *top; + int remain; + + SOCKBUF_LOCK_ASSERT(sb); + MPASS(len <= sb->sb_tlscc); + + /* + * If TLS chain is the exact size of the record, + * just grab the whole record. + */ + top = sb->sb_mtls; + if (sb->sb_tlscc == len) { + sb->sb_mtls = NULL; + sb->sb_mtlstail = NULL; + goto out; + } + + /* + * While it would be nice to use m_split() here, we need + * to know exactly what m_split() allocates to update the + * accounting, so do it inline instead. + */ + remain = len; + for (m = top; remain > m->m_len; m = m->m_next) + remain -= m->m_len; + + /* Easy case: don't have to split 'm'. */ + if (remain == m->m_len) { + sb->sb_mtls = m->m_next; + if (sb->sb_mtls == NULL) + sb->sb_mtlstail = NULL; + m->m_next = NULL; + goto out; + } + + /* + * Need to allocate an mbuf to hold the remainder of 'm'. Try + * with M_NOWAIT first. + */ + n = m_get(M_NOWAIT, MT_DATA); + if (n == NULL) { + /* + * Use M_WAITOK with socket buffer unlocked. If + * 'sb_mtls' changes while the lock is dropped, return + * NULL to force the caller to retry. + */ + SOCKBUF_UNLOCK(sb); + + n = m_get(M_WAITOK, MT_DATA); + + SOCKBUF_LOCK(sb); + if (sb->sb_mtls != top) { + m_free(n); + return (NULL); + } + } + n->m_flags |= M_NOTREADY; + + /* Store remainder in 'n'. */ + n->m_len = m->m_len - remain; + if (m->m_flags & M_EXT) { + n->m_data = m->m_data + remain; + mb_dupcl(n, m); + } else { + bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len); + } + + /* Trim 'm' and update accounting. */ + m->m_len -= n->m_len; + sb->sb_tlscc -= n->m_len; + sb->sb_ccc -= n->m_len; + + /* Account for 'n'. */ + sballoc_ktls_rx(sb, n); + + /* Insert 'n' into the TLS chain. */ + sb->sb_mtls = n; + n->m_next = m->m_next; + if (sb->sb_mtlstail == m) + sb->sb_mtlstail = n; + + /* Detach the record from the TLS chain. */ + m->m_next = NULL; + +out: + MPASS(m_length(top, NULL) == len); + for (m = top; m != NULL; m = m->m_next) + sbfree_ktls_rx(sb, m); + sb->sb_tlsdcc = len; + sb->sb_ccc += len; + SBCHECK(sb); + return (top); +} + +static int +m_segments(struct mbuf *m, int skip) +{ + int count; + + while (skip >= m->m_len) { + skip -= m->m_len; + m = m->m_next; + } + + for (count = 0; m != NULL; count++) + m = m->m_next; + return (count); +} + +static void +ktls_decrypt(struct socket *so) +{ + char tls_header[MBUF_PEXT_HDR_LEN]; + struct ktls_session *tls; + struct sockbuf *sb; + struct tls_record_layer *hdr; + struct tls_get_record tgr; + struct mbuf *control, *data, *m; + uint64_t seqno; + int error, remain, tls_len, trail_len; + + hdr = (struct tls_record_layer *)tls_header; + sb = &so->so_rcv; + SOCKBUF_LOCK(sb); + KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING, + ("%s: socket %p not running", __func__, so)); + + tls = sb->sb_tls_info; + MPASS(tls != NULL); + + for (;;) { + /* Is there enough queued for a TLS header? */ + if (sb->sb_tlscc < tls->params.tls_hlen) + break; + + m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header); + tls_len = sizeof(*hdr) + ntohs(hdr->tls_length); + + if (hdr->tls_vmajor != tls->params.tls_vmajor || + hdr->tls_vminor != tls->params.tls_vminor) + error = EINVAL; + else if (tls_len < tls->params.tls_hlen || tls_len > + tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 + + tls->params.tls_tlen) + error = EMSGSIZE; + else + error = 0; + if (__predict_false(error != 0)) { + /* + * We have a corrupted record and are likely + * out of sync. The connection isn't + * recoverable at this point, so abort it. + */ + SOCKBUF_UNLOCK(sb); + counter_u64_add(ktls_offload_corrupted_records, 1); + + CURVNET_SET(so->so_vnet); + so->so_proto->pr_usrreqs->pru_abort(so); + so->so_error = error; + CURVNET_RESTORE(); + goto deref; + } + + /* Is the entire record queued? */ + if (sb->sb_tlscc < tls_len) + break; + + /* + * Split out the portion of the mbuf chain containing + * this TLS record. + */ + data = ktls_detach_record(sb, tls_len); + if (data == NULL) + continue; + MPASS(sb->sb_tlsdcc == tls_len); + + seqno = sb->sb_tls_seqno; + sb->sb_tls_seqno++; + SBCHECK(sb); + SOCKBUF_UNLOCK(sb); + + error = tls->sw_decrypt(tls, hdr, data, seqno, &trail_len); + if (error) { + counter_u64_add(ktls_offload_failed_crypto, 1); + + SOCKBUF_LOCK(sb); + if (sb->sb_tlsdcc == 0) { + /* + * sbcut/drop/flush discarded these + * mbufs. + */ + m_freem(data); + break; + } + + /* + * Drop this TLS record's data, but keep + * decrypting subsequent records. + */ + sb->sb_ccc -= tls_len; + sb->sb_tlsdcc = 0; + + CURVNET_SET(so->so_vnet); + so->so_error = EBADMSG; + sorwakeup_locked(so); + CURVNET_RESTORE(); + + m_freem(data); + + SOCKBUF_LOCK(sb); + continue; + } + + /* Allocate the control mbuf. */ + tgr.tls_type = hdr->tls_type; + tgr.tls_vmajor = hdr->tls_vmajor; + tgr.tls_vminor = hdr->tls_vminor; + tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen - + trail_len); + control = sbcreatecontrol_how(&tgr, sizeof(tgr), + TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK); + + SOCKBUF_LOCK(sb); + if (sb->sb_tlsdcc == 0) { + /* sbcut/drop/flush discarded these mbufs. */ + MPASS(sb->sb_tlscc == 0); + m_freem(data); + m_freem(control); + break; + } + + /* + * Clear the 'dcc' accounting in preparation for + * adding the decrypted record. + */ + sb->sb_ccc -= tls_len; + sb->sb_tlsdcc = 0; + SBCHECK(sb); + + /* If there is no payload, drop all of the data. */ + if (tgr.tls_length == htobe16(0)) { + m_freem(data); + data = NULL; + } else { + /* Trim header. */ + remain = tls->params.tls_hlen; + while (remain > 0) { + if (data->m_len > remain) { + data->m_data += remain; + data->m_len -= remain; + break; + } + remain -= data->m_len; + data = m_free(data); + } + + /* Trim trailer and clear M_NOTREADY. */ + remain = be16toh(tgr.tls_length); + m = data; + for (m = data; remain > m->m_len; m = m->m_next) { + m->m_flags &= ~M_NOTREADY; + remain -= m->m_len; + } + m->m_len = remain; + m_freem(m->m_next); + m->m_next = NULL; + m->m_flags &= ~M_NOTREADY; + + /* Set EOR on the final mbuf. */ + m->m_flags |= M_EOR; + } + + sbappendcontrol_locked(sb, data, control, 0); + } + + sb->sb_flags &= ~SB_TLS_RX_RUNNING; + + if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0) + so->so_error = EMSGSIZE; + + sorwakeup_locked(so); + +deref: + SOCKBUF_UNLOCK_ASSERT(sb); + + CURVNET_SET(so->so_vnet); + SOCK_LOCK(so); + sorele(so); + CURVNET_RESTORE(); +} + +void ktls_enqueue_to_free(struct mbuf *m) { struct ktls_wq *wq; @@ -1431,7 +1849,7 @@ ktls_enqueue_to_free(struct mbuf *m) m->m_epg_flags |= EPG_FLAG_2FREE; wq = &ktls_wq[m->m_epg_tls->wq_index]; mtx_lock(&wq->mtx); - STAILQ_INSERT_TAIL(&wq->head, m, m_epg_stailq); + STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); running = wq->running; mtx_unlock(&wq->mtx); if (!running) @@ -1461,12 +1879,12 @@ ktls_enqueue(struct mbuf *m, struct socket *so, int pa wq = &ktls_wq[m->m_epg_tls->wq_index]; mtx_lock(&wq->mtx); - STAILQ_INSERT_TAIL(&wq->head, m, m_epg_stailq); + STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); running = wq->running; mtx_unlock(&wq->mtx); if (!running) wakeup(wq); - counter_u64_add(ktls_cnt_on, 1); + counter_u64_add(ktls_cnt_tx_queued, 1); } static __noinline void @@ -1618,31 +2036,41 @@ ktls_work_thread(void *ctx) { struct ktls_wq *wq = ctx; struct mbuf *m, *n; - STAILQ_HEAD(, mbuf) local_head; + struct socket *so, *son; + STAILQ_HEAD(, mbuf) local_m_head; + STAILQ_HEAD(, socket) local_so_head; #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) fpu_kern_thread(0); #endif for (;;) { mtx_lock(&wq->mtx); - while (STAILQ_EMPTY(&wq->head)) { + while (STAILQ_EMPTY(&wq->m_head) && + STAILQ_EMPTY(&wq->so_head)) { wq->running = false; mtx_sleep(wq, &wq->mtx, 0, "-", 0); wq->running = true; } - STAILQ_INIT(&local_head); - STAILQ_CONCAT(&local_head, &wq->head); + STAILQ_INIT(&local_m_head); + STAILQ_CONCAT(&local_m_head, &wq->m_head); + STAILQ_INIT(&local_so_head); + STAILQ_CONCAT(&local_so_head, &wq->so_head); mtx_unlock(&wq->mtx); - STAILQ_FOREACH_SAFE(m, &local_head, m_epg_stailq, n) { + STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) { if (m->m_epg_flags & EPG_FLAG_2FREE) { ktls_free(m->m_epg_tls); uma_zfree(zone_mbuf, m); } else { ktls_encrypt(m); - counter_u64_add(ktls_cnt_on, -1); + counter_u64_add(ktls_cnt_tx_queued, -1); } + } + + STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) { + ktls_decrypt(so); + counter_u64_add(ktls_cnt_rx_queued, -1); } } } Modified: head/sys/kern/uipc_sockbuf.c ============================================================================== --- head/sys/kern/uipc_sockbuf.c Thu Jul 23 23:29:50 2020 (r363463) +++ head/sys/kern/uipc_sockbuf.c Thu Jul 23 23:48:18 2020 (r363464) @@ -70,6 +70,8 @@ u_long sb_max_adj = static u_long sb_efficiency = 8; /* parameter for sbreserve() */ +static void sbcompress_ktls_rx(struct sockbuf *sb, struct mbuf *m, + struct mbuf *n); static struct mbuf *sbcut_internal(struct sockbuf *sb, int len); static void sbflush_internal(struct sockbuf *sb); @@ -334,7 +336,52 @@ sbfree(struct sockbuf *sb, struct mbuf *m) sb->sb_sndptroff -= m->m_len; } +#ifdef KERN_TLS /* + * Similar to sballoc/sbfree but does not adjust state associated with + * the sb_mb chain such as sb_fnrdy or sb_sndptr*. Also assumes mbufs + * are not ready. + */ +void +sballoc_ktls_rx(struct sockbuf *sb, struct mbuf *m) +{ + + SOCKBUF_LOCK_ASSERT(sb); + + sb->sb_ccc += m->m_len; + sb->sb_tlscc += m->m_len; + + sb->sb_mbcnt += MSIZE; + sb->sb_mcnt += 1; + + if (m->m_flags & M_EXT) { + sb->sb_mbcnt += m->m_ext.ext_size; + sb->sb_ccnt += 1; + } +} + +void +sbfree_ktls_rx(struct sockbuf *sb, struct mbuf *m) +{ + +#if 0 /* XXX: not yet: soclose() call path comes here w/o lock. */ + SOCKBUF_LOCK_ASSERT(sb); +#endif + + sb->sb_ccc -= m->m_len; + sb->sb_tlscc -= m->m_len; + + sb->sb_mbcnt -= MSIZE; + sb->sb_mcnt -= 1; + + if (m->m_flags & M_EXT) { + sb->sb_mbcnt -= m->m_ext.ext_size; + sb->sb_ccnt -= 1; + } +} +#endif + +/* * Socantsendmore indicates that no more data will be sent on the socket; it * would normally be applied to a socket when the user informs the system * that no more data is to be sent, by the protocol code (in case @@ -370,6 +417,10 @@ socantrcvmore_locked(struct socket *so) SOCKBUF_LOCK_ASSERT(&so->so_rcv); so->so_rcv.sb_state |= SBS_CANTRCVMORE; +#ifdef KERN_TLS + if (so->so_rcv.sb_flags & SB_TLS_RX) + ktls_check_rx(&so->so_rcv); +#endif sorwakeup_locked(so); mtx_assert(SOCKBUF_MTX(&so->so_rcv), MA_NOTOWNED); } @@ -770,6 +821,24 @@ sblastmbufchk(struct sockbuf *sb, const char *file, in } panic("%s from %s:%u", __func__, file, line); } + +#ifdef KERN_TLS + m = sb->sb_mtls; + while (m && m->m_next) + m = m->m_next; + + if (m != sb->sb_mtlstail) { + printf("%s: sb_mtls %p sb_mtlstail %p last %p\n", + __func__, sb->sb_mtls, sb->sb_mtlstail, m); + printf("TLS packet tree:\n"); + printf("\t"); + for (m = sb->sb_mtls; m != NULL; m = m->m_next) { + printf("%p ", m); + } + printf("\n"); + panic("%s from %s:%u", __func__, file, line); + } +#endif } #endif /* SOCKBUF_DEBUG */ @@ -847,7 +916,30 @@ sbappend(struct sockbuf *sb, struct mbuf *m, int flags SOCKBUF_UNLOCK(sb); } +#ifdef KERN_TLS /* + * Append an mbuf containing encrypted TLS data. The data + * is marked M_NOTREADY until it has been decrypted and + * stored as a TLS record. + */ +static void +sbappend_ktls_rx(struct sockbuf *sb, struct mbuf *m) +{ + struct mbuf *n; + + SBLASTMBUFCHK(sb); + + /* Remove all packet headers and mbuf tags to get a pure data chain. */ + m_demote(m, 1, 0); + + for (n = m; n != NULL; n = n->m_next) + n->m_flags |= M_NOTREADY; + sbcompress_ktls_rx(sb, m, sb->sb_mtlstail); + ktls_check_rx(sb); +} +#endif + +/* * This version of sbappend() should only be used when the caller absolutely * knows that there will never be more than one record in the socket buffer, * that is, a stream protocol (such as TCP). @@ -858,6 +950,19 @@ sbappendstream_locked(struct sockbuf *sb, struct mbuf SOCKBUF_LOCK_ASSERT(sb); KASSERT(m->m_nextpkt == NULL,("sbappendstream 0")); + +#ifdef KERN_TLS + /* + * Decrypted TLS records are appended as records via + * sbappendrecord(). TCP passes encrypted TLS records to this + * function which must be scheduled for decryption. + */ + if (sb->sb_flags & SB_TLS_RX) { + sbappend_ktls_rx(sb, m); + return; + } +#endif + KASSERT(sb->sb_mb == sb->sb_lastrecord,("sbappendstream 1")); SBLASTMBUFCHK(sb); @@ -896,6 +1001,9 @@ sbcheck(struct sockbuf *sb, const char *file, int line { struct mbuf *m, *n, *fnrdy; u_long acc, ccc, mbcnt; +#ifdef KERN_TLS + u_long tlscc; +#endif SOCKBUF_LOCK_ASSERT(sb); @@ -931,9 +1039,46 @@ sbcheck(struct sockbuf *sb, const char *file, int line mbcnt += m->m_ext.ext_size; } } +#ifdef KERN_TLS + /* + * Account for mbufs "detached" by ktls_detach_record() while + * they are decrypted by ktls_decrypt(). tlsdcc gives a count + * of the detached bytes that are included in ccc. The mbufs + * and clusters are not included in the socket buffer + * accounting. + */ + ccc += sb->sb_tlsdcc; + + tlscc = 0; + for (m = sb->sb_mtls; m; m = m->m_next) { + if (m->m_nextpkt != NULL) { + printf("sb %p TLS mbuf %p with nextpkt\n", sb, m); + goto fail; + } + if ((m->m_flags & M_NOTREADY) == 0) { + printf("sb %p TLS mbuf %p ready\n", sb, m); + goto fail; + } + tlscc += m->m_len; + ccc += m->m_len; + mbcnt += MSIZE; + if (m->m_flags & M_EXT) /*XXX*/ /* pretty sure this is bogus */ + mbcnt += m->m_ext.ext_size; + } + + if (sb->sb_tlscc != tlscc) { + printf("tlscc %ld/%u dcc %u\n", tlscc, sb->sb_tlscc, + sb->sb_tlsdcc); + goto fail; + } +#endif if (acc != sb->sb_acc || ccc != sb->sb_ccc || mbcnt != sb->sb_mbcnt) { printf("acc %ld/%u ccc %ld/%u mbcnt %ld/%u\n", acc, sb->sb_acc, ccc, sb->sb_ccc, mbcnt, sb->sb_mbcnt); +#ifdef KERN_TLS + printf("tlscc %ld/%u dcc %u\n", tlscc, sb->sb_tlscc, + sb->sb_tlsdcc); +#endif goto fail; } return; @@ -1209,14 +1354,72 @@ sbcompress(struct sockbuf *sb, struct mbuf *m, struct SBLASTMBUFCHK(sb); } +#ifdef KERN_TLS /* + * A version of sbcompress() for encrypted TLS RX mbufs. These mbufs + * are appended to the 'sb_mtls' chain instead of 'sb_mb' and are also + * a bit simpler (no EOR markers, always MT_DATA, etc.). + */ +static void +sbcompress_ktls_rx(struct sockbuf *sb, struct mbuf *m, struct mbuf *n) +{ + + SOCKBUF_LOCK_ASSERT(sb); + + while (m) { + KASSERT((m->m_flags & M_EOR) == 0, + ("TLS RX mbuf %p with EOR", m)); + KASSERT(m->m_type == MT_DATA, + ("TLS RX mbuf %p is not MT_DATA", m)); + KASSERT((m->m_flags & M_NOTREADY) != 0, + ("TLS RX mbuf %p ready", m)); + KASSERT((m->m_flags & M_EXTPG) == 0, + ("TLS RX mbuf %p unmapped", m)); + + if (m->m_len == 0) { + m = m_free(m); + continue; + } + + /* + * Even though both 'n' and 'm' are NOTREADY, it's ok + * to coalesce the data. + */ + if (n && + M_WRITABLE(n) && + ((sb->sb_flags & SB_NOCOALESCE) == 0) && + !(n->m_flags & (M_EXTPG)) && + m->m_len <= MCLBYTES / 4 && /* XXX: Don't copy too much */ + m->m_len <= M_TRAILINGSPACE(n)) { + m_copydata(m, 0, m->m_len, mtodo(n, n->m_len)); + n->m_len += m->m_len; + sb->sb_ccc += m->m_len; + sb->sb_tlscc += m->m_len; + m = m_free(m); + continue; + } + if (n) + n->m_next = m; + else + sb->sb_mtls = m; + sb->sb_mtlstail = m; + sballoc_ktls_rx(sb, m); + n = m; + m = m->m_next; + n->m_next = NULL; + } + SBLASTMBUFCHK(sb); +} +#endif + +/* * Free all mbufs in a sockbuf. Check that all resources are reclaimed. */ static void sbflush_internal(struct sockbuf *sb) { - while (sb->sb_mbcnt) { + while (sb->sb_mbcnt || sb->sb_tlsdcc) { /* * Don't call sbcut(sb, 0) if the leading mbuf is non-empty: * we would loop forever. Panic instead. @@ -1254,6 +1457,7 @@ static struct mbuf * sbcut_internal(struct sockbuf *sb, int len) { struct mbuf *m, *next, *mfree; + bool is_tls; KASSERT(len >= 0, ("%s: len is %d but it is supposed to be >= 0", __func__, len)); @@ -1261,10 +1465,25 @@ sbcut_internal(struct sockbuf *sb, int len) __func__, len, sb->sb_ccc)); next = (m = sb->sb_mb) ? m->m_nextpkt : 0; + is_tls = false; mfree = NULL; while (len > 0) { if (m == NULL) { +#ifdef KERN_TLS + if (next == NULL && !is_tls) { + if (sb->sb_tlsdcc != 0) { + MPASS(len >= sb->sb_tlsdcc); + len -= sb->sb_tlsdcc; + sb->sb_ccc -= sb->sb_tlsdcc; + sb->sb_tlsdcc = 0; + if (len == 0) + break; + } + next = sb->sb_mtls; + is_tls = true; + } +#endif KASSERT(next, ("%s: no next, len %d", __func__, len)); m = next; next = m->m_nextpkt; @@ -1283,12 +1502,17 @@ sbcut_internal(struct sockbuf *sb, int len) break; } len -= m->m_len; - sbfree(sb, m); +#ifdef KERN_TLS + if (is_tls) + sbfree_ktls_rx(sb, m); + else +#endif + sbfree(sb, m); /* * Do not put M_NOTREADY buffers to the free list, they * are referenced from outside. */ - if (m->m_flags & M_NOTREADY) + if (m->m_flags & M_NOTREADY && !is_tls) m = m->m_next; else { struct mbuf *n; @@ -1314,6 +1538,14 @@ sbcut_internal(struct sockbuf *sb, int len) mfree = m; m = n; } +#ifdef KERN_TLS + if (is_tls) { *** DIFF OUTPUT TRUNCATED AT 1000 LINES ***
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?202007232348.06NNmJ6o063845>