From owner-svn-src-all@freebsd.org Mon Feb 25 10:38:38 2019 Return-Path: Delivered-To: svn-src-all@mailman.ysv.freebsd.org Received: from mx1.freebsd.org (mx1.freebsd.org [IPv6:2610:1c1:1:606c::19:1]) by mailman.ysv.freebsd.org (Postfix) with ESMTP id 9BDE41505024; Mon, 25 Feb 2019 10:38:38 +0000 (UTC) (envelope-from tuexen@FreeBSD.org) Received: from mxrelay.nyi.freebsd.org (mxrelay.nyi.freebsd.org [IPv6:2610:1c1:1:606c::19:3]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) server-signature RSA-PSS (4096 bits) client-signature RSA-PSS (4096 bits) client-digest SHA256) (Client CN "mxrelay.nyi.freebsd.org", Issuer "Let's Encrypt Authority X3" (verified OK)) by mx1.freebsd.org (Postfix) with ESMTPS id 32D6A861AA; Mon, 25 Feb 2019 10:38:38 +0000 (UTC) (envelope-from tuexen@FreeBSD.org) Received: from repo.freebsd.org (repo.freebsd.org [IPv6:2610:1c1:1:6068::e6a:0]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (Client did not present a certificate) by mxrelay.nyi.freebsd.org (Postfix) with ESMTPS id 1CFCD1BB80; Mon, 25 Feb 2019 10:38:38 +0000 (UTC) (envelope-from tuexen@FreeBSD.org) Received: from repo.freebsd.org ([127.0.1.37]) by repo.freebsd.org (8.15.2/8.15.2) with ESMTP id x1PAccbE064477; Mon, 25 Feb 2019 10:38:38 GMT (envelope-from tuexen@FreeBSD.org) Received: (from tuexen@localhost) by repo.freebsd.org (8.15.2/8.15.2/Submit) id x1PAcbpZ064474; Mon, 25 Feb 2019 10:38:37 GMT (envelope-from tuexen@FreeBSD.org) Message-Id: <201902251038.x1PAcbpZ064474@repo.freebsd.org> X-Authentication-Warning: repo.freebsd.org: tuexen set sender to tuexen@FreeBSD.org using -f From: Michael Tuexen Date: Mon, 25 Feb 2019 10:38:37 +0000 (UTC) To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-stable@freebsd.org, svn-src-stable-11@freebsd.org Subject: svn commit: r344511 - in stable/11/sys: netinet netinet/tcp_stacks sys X-SVN-Group: stable-11 X-SVN-Commit-Author: tuexen X-SVN-Commit-Paths: in stable/11/sys: netinet netinet/tcp_stacks sys X-SVN-Commit-Revision: 344511 X-SVN-Commit-Repository: base MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit X-Rspamd-Queue-Id: 32D6A861AA X-Spamd-Bar: -- Authentication-Results: mx1.freebsd.org X-Spamd-Result: default: False [-2.95 / 15.00]; local_wl_from(0.00)[FreeBSD.org]; NEURAL_HAM_MEDIUM(-1.00)[-0.998,0]; NEURAL_HAM_LONG(-1.00)[-0.999,0]; NEURAL_HAM_SHORT(-0.95)[-0.955,0]; ASN(0.00)[asn:11403, ipnet:2610:1c1:1::/48, country:US] X-BeenThere: svn-src-all@freebsd.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: "SVN commit messages for the entire src tree \(except for " user" and " projects" \)" List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Mon, 25 Feb 2019 10:38:39 -0000 Author: tuexen Date: Mon Feb 25 10:38:37 2019 New Revision: 344511 URL: https://svnweb.freebsd.org/changeset/base/344511 Log: Backport the new TCP reassembly code from head to stable/11. In particular: * Cherry pick the changes in sys/queue.h from r334804. * MFC r338102 with manually removing changes to file not existent in stable/11 and resolve conflicts in tcp_var.h. This change represents a substantial restructure of the way we reassembly inbound tcp segments. The old algorithm just blindly dropped in segments without coalescing. This meant that every segment could take up greater and greater room on the linked list of segments. This of course is now subject to a tighter limit (100) of segments which in a high BDP situation will cause us to be a lot more in-efficent as we drop segments beyond 100 entries that we receive. What this restructure does is cause the reassembly buffer to coalesce segments putting an emphasis on the two common cases (which avoid walking the list of segments) i.e. where we add to the back of the queue of segments and where we add to the front. We also have the reassembly buffer supporting a couple of debug options (black box logging as well as counters for code coverage). These are compiled out by default but can be added by uncommenting the defines. * Manually fix tcp_stacks/fastopen.c, since it does not exist anymore in head. * MFC r342280: Fix a regression in the TCP handling of received segments. When receiving TCP segments the stack protects itself by limiting the resources allocated for a TCP connections. This patch adds an exception to these limitations for the TCP segement which is the next expected in-sequence segment. Without this patch, TCP connections may stall and finally fail in some cases of packet loss. * MFC r343439: Don't include two header files when not needed. This allows the part of the rewrite of TCP reassembly in this files to be MFCed to stable/11 with manual change. * MFC r344428: This patch addresses an issue brought up by bz@ in D18968: When TCP_REASS_LOGGING is defined, a NULL pointer dereference would happen, if user data was received during the TCP handshake and BB logging is used. A KASSERT is also added to detect tcp_reass() calls with illegal parameter combinations. Reviewed by: bz@, jtl, rrs@ Sponsored by: Netflix, Inc. Differential Revision: https://reviews.freebsd.org/D18960 Modified: stable/11/sys/netinet/tcp_input.c stable/11/sys/netinet/tcp_reass.c stable/11/sys/netinet/tcp_stacks/fastpath.c stable/11/sys/netinet/tcp_subr.c stable/11/sys/netinet/tcp_usrreq.c stable/11/sys/netinet/tcp_var.h stable/11/sys/sys/queue.h Directory Properties: stable/11/ (props changed) Modified: stable/11/sys/netinet/tcp_input.c ============================================================================== --- stable/11/sys/netinet/tcp_input.c Mon Feb 25 09:57:06 2019 (r344510) +++ stable/11/sys/netinet/tcp_input.c Mon Feb 25 10:38:37 2019 (r344511) @@ -1760,7 +1760,7 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, stru tp->snd_nxt == tp->snd_max && tiwin && tiwin == tp->snd_wnd && ((tp->t_flags & (TF_NEEDSYN|TF_NEEDFIN)) == 0) && - LIST_EMPTY(&tp->t_segq) && + SEGQ_EMPTY(tp) && ((to.to_flags & TOF_TS) == 0 || TSTMP_GEQ(to.to_tsval, tp->ts_recent)) ) { @@ -2473,7 +2473,7 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, stru * later; if not, do so now to pass queued data to user. */ if (tlen == 0 && (thflags & TH_FIN) == 0) - (void) tcp_reass(tp, (struct tcphdr *)0, 0, + (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0, (struct mbuf *)0); tp->snd_wl1 = th->th_seq - 1; /* FALLTHROUGH */ @@ -3044,7 +3044,7 @@ dodata: /* XXX */ * fast retransmit can work). */ if (th->th_seq == tp->rcv_nxt && - LIST_EMPTY(&tp->t_segq) && + SEGQ_EMPTY(tp) && (TCPS_HAVEESTABLISHED(tp->t_state) || tfo_syn)) { if (DELAY_ACK(tp, tlen) || tfo_syn) @@ -3069,7 +3069,7 @@ dodata: /* XXX */ * m_adj() doesn't actually frees any mbufs * when trimming from the head. */ - thflags = tcp_reass(tp, th, &tlen, m); + thflags = tcp_reass(tp, th, &save_start, &tlen, m); tp->t_flags |= TF_ACKNOW; } if (tlen > 0 && (tp->t_flags & TF_SACK_PERMIT)) Modified: stable/11/sys/netinet/tcp_reass.c ============================================================================== --- stable/11/sys/netinet/tcp_reass.c Mon Feb 25 09:57:06 2019 (r344510) +++ stable/11/sys/netinet/tcp_reass.c Mon Feb 25 10:38:37 2019 (r344511) @@ -36,6 +36,10 @@ __FBSDID("$FreeBSD$"); #include "opt_inet6.h" #include "opt_tcpdebug.h" +/* For debugging we want counters and BB logging */ +/* #define TCP_REASS_COUNTERS 1 */ +/* #define TCP_REASS_LOGGING 1 */ + #include #include #include @@ -70,15 +74,35 @@ __FBSDID("$FreeBSD$"); #include #include #include +#ifdef TCP_REASS_LOGGING +#include +#include +#endif #include #include #ifdef TCPDEBUG #include #endif /* TCPDEBUG */ +#define TCP_R_LOG_ADD 1 +#define TCP_R_LOG_LIMIT_REACHED 2 +#define TCP_R_LOG_APPEND 3 +#define TCP_R_LOG_PREPEND 4 +#define TCP_R_LOG_REPLACE 5 +#define TCP_R_LOG_MERGE_INTO 6 +#define TCP_R_LOG_NEW_ENTRY 7 +#define TCP_R_LOG_READ 8 +#define TCP_R_LOG_ZERO 9 +#define TCP_R_LOG_DUMP 10 +#define TCP_R_LOG_TRIM 11 + static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, reass, CTLFLAG_RW, 0, "TCP Segment Reassembly Queue"); +static SYSCTL_NODE(_net_inet_tcp_reass, OID_AUTO, stats, CTLFLAG_RW, 0, + "TCP Segment Reassembly stats"); + + static int tcp_reass_maxseg = 0; SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, maxsegments, CTLFLAG_RDTUN, &tcp_reass_maxseg, 0, @@ -94,6 +118,72 @@ SYSCTL_UINT(_net_inet_tcp_reass, OID_AUTO, maxqueuelen &tcp_reass_maxqueuelen, 0, "Maximum number of TCP Segments per Reassembly Queue"); +static int tcp_new_limits = 0; +SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, new_limit, CTLFLAG_RWTUN, + &tcp_new_limits, 0, + "Do we use the new limit method we are discussing?"); + +static u_int tcp_reass_queue_guard = 16; +SYSCTL_UINT(_net_inet_tcp_reass, OID_AUTO, queueguard, CTLFLAG_RWTUN, + &tcp_reass_queue_guard, 16, + "Number of TCP Segments in Reassembly Queue where we flip over to guard mode"); + +#ifdef TCP_REASS_COUNTERS + +counter_u64_t reass_entry; +SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, entry, CTLFLAG_RD, + &reass_entry, "A segment entered reassembly "); + +counter_u64_t reass_path1; +SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path1, CTLFLAG_RD, + &reass_path1, "Took path 1"); + +counter_u64_t reass_path2; +SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path2, CTLFLAG_RD, + &reass_path2, "Took path 2"); + +counter_u64_t reass_path3; +SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path3, CTLFLAG_RD, + &reass_path3, "Took path 3"); + +counter_u64_t reass_path4; +SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path4, CTLFLAG_RD, + &reass_path4, "Took path 4"); + +counter_u64_t reass_path5; +SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path5, CTLFLAG_RD, + &reass_path5, "Took path 5"); + +counter_u64_t reass_path6; +SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path6, CTLFLAG_RD, + &reass_path6, "Took path 6"); + +counter_u64_t reass_path7; +SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path7, CTLFLAG_RD, + &reass_path7, "Took path 7"); + +counter_u64_t reass_fullwalk; +SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, fullwalk, CTLFLAG_RD, + &reass_fullwalk, "Took a full walk "); + +counter_u64_t reass_nospace; +SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, nospace, CTLFLAG_RD, + &reass_nospace, "Had no mbuf capacity "); + +counter_u64_t merge_fwd; +SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, merge_fwd, CTLFLAG_RD, + &merge_fwd, "Ran merge fwd"); + +counter_u64_t merge_into; +SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, merge_into, CTLFLAG_RD, + &merge_into, "Ran merge into"); + +counter_u64_t tcp_zero_input; +SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, zero_input, CTLFLAG_RD, + &tcp_zero_input, "The reassembly buffer saw a zero len segment etc"); + +#endif + /* Initialize TCP reassembly queue */ static void tcp_reass_zone_change(void *tag) @@ -105,6 +195,77 @@ tcp_reass_zone_change(void *tag) tcp_reass_maxseg); } +#ifdef TCP_REASS_LOGGING + +static void +tcp_log_reassm(struct tcpcb *tp, struct tseg_qent *q, struct tseg_qent *p, + tcp_seq seq, int len, uint8_t action, int instance) +{ + uint32_t cts; + struct timeval tv; + + if (tp->t_logstate != TCP_LOG_STATE_OFF) { + union tcp_log_stackspecific log; + + memset(&log, 0, sizeof(log)); + cts = tcp_get_usecs(&tv); + log.u_bbr.flex1 = seq; + log.u_bbr.cur_del_rate = (uint64_t)q; + log.u_bbr.delRate = (uint64_t)p; + if (q != NULL) { + log.u_bbr.flex2 = q->tqe_start; + log.u_bbr.flex3 = q->tqe_len; + log.u_bbr.flex4 = q->tqe_mbuf_cnt; + log.u_bbr.hptsi_gain = q->tqe_flags; + } + if (p != NULL) { + log.u_bbr.flex5 = p->tqe_start; + log.u_bbr.pkts_out = p->tqe_len; + log.u_bbr.epoch = p->tqe_mbuf_cnt; + log.u_bbr.cwnd_gain = p->tqe_flags; + } + log.u_bbr.flex6 = tp->t_segqmbuflen; + log.u_bbr.flex7 = instance; + log.u_bbr.flex8 = action; + log.u_bbr.timeStamp = cts; + TCP_LOG_EVENTP(tp, NULL, + &tp->t_inpcb->inp_socket->so_rcv, + &tp->t_inpcb->inp_socket->so_snd, + TCP_LOG_REASS, 0, + len, &log, false, &tv); + } +} + +static void +tcp_reass_log_dump(struct tcpcb *tp) +{ + struct tseg_qent *q; + + if (tp->t_logstate != TCP_LOG_STATE_OFF) { + TAILQ_FOREACH(q, &tp->t_segq, tqe_q) { + tcp_log_reassm(tp, q, NULL, q->tqe_start, q->tqe_len, TCP_R_LOG_DUMP, 0); + } + }; +} + +static void +tcp_reass_log_new_in(struct tcpcb *tp, tcp_seq seq, int len, struct mbuf *m, + int logval, struct tseg_qent *q) +{ + int cnt; + struct mbuf *t; + + cnt = 0; + t = m; + while (t) { + cnt += t->m_len; + t = t->m_next; + } + tcp_log_reassm(tp, q, NULL, seq, len, logval, cnt); +} + +#endif + void tcp_reass_global_init(void) { @@ -117,8 +278,24 @@ tcp_reass_global_init(void) /* Set the zone limit and read back the effective value. */ tcp_reass_maxseg = uma_zone_set_max(tcp_reass_zone, tcp_reass_maxseg); +#ifdef TCP_REASS_COUNTERS + reass_path1 = counter_u64_alloc(M_WAITOK); + reass_path2 = counter_u64_alloc(M_WAITOK); + reass_path3 = counter_u64_alloc(M_WAITOK); + reass_path4 = counter_u64_alloc(M_WAITOK); + reass_path5 = counter_u64_alloc(M_WAITOK); + reass_path6 = counter_u64_alloc(M_WAITOK); + reass_path7 = counter_u64_alloc(M_WAITOK); + reass_fullwalk = counter_u64_alloc(M_WAITOK); + reass_nospace = counter_u64_alloc(M_WAITOK); + reass_entry = counter_u64_alloc(M_WAITOK); + merge_fwd = counter_u64_alloc(M_WAITOK); + merge_into = counter_u64_alloc(M_WAITOK); + tcp_zero_input = counter_u64_alloc(M_WAITOK); +#endif EVENTHANDLER_REGISTER(nmbclusters_change, tcp_reass_zone_change, NULL, EVENTHANDLER_PRI_ANY); + } void @@ -128,191 +305,739 @@ tcp_reass_flush(struct tcpcb *tp) INP_WLOCK_ASSERT(tp->t_inpcb); - while ((qe = LIST_FIRST(&tp->t_segq)) != NULL) { - LIST_REMOVE(qe, tqe_q); + while ((qe = TAILQ_FIRST(&tp->t_segq)) != NULL) { + TAILQ_REMOVE(&tp->t_segq, qe, tqe_q); m_freem(qe->tqe_m); uma_zfree(tcp_reass_zone, qe); tp->t_segqlen--; } - + tp->t_segqmbuflen = 0; KASSERT((tp->t_segqlen == 0), ("TCP reass queue %p segment count is %d instead of 0 after flush.", tp, tp->t_segqlen)); } +static void +tcp_reass_append(struct tcpcb *tp, struct tseg_qent *last, + struct mbuf *m, struct tcphdr *th, int tlen, + struct mbuf *mlast, int lenofoh) +{ + +#ifdef TCP_REASS_LOGGING + tcp_log_reassm(tp, last, NULL, th->th_seq, tlen, TCP_R_LOG_APPEND, 0); +#endif + last->tqe_len += tlen; + last->tqe_m->m_pkthdr.len += tlen; + /* Preserve the FIN bit if its there */ + last->tqe_flags |= (th->th_flags & TH_FIN); + last->tqe_last->m_next = m; + last->tqe_last = mlast; + last->tqe_mbuf_cnt += lenofoh; + tp->t_rcvoopack++; + TCPSTAT_INC(tcps_rcvoopack); + TCPSTAT_ADD(tcps_rcvoobyte, tlen); +#ifdef TCP_REASS_LOGGING + tcp_reass_log_new_in(tp, last->tqe_start, lenofoh, last->tqe_m, + TCP_R_LOG_APPEND, + last); +#endif +} + +static void +tcp_reass_prepend(struct tcpcb *tp, struct tseg_qent *first, struct mbuf *m, struct tcphdr *th, + int tlen, struct mbuf *mlast, int lenofoh) +{ + int i; + +#ifdef TCP_REASS_LOGGING + tcp_log_reassm(tp, first, NULL, th->th_seq, tlen, TCP_R_LOG_PREPEND, 0); +#endif + if (SEQ_GT((th->th_seq + tlen), first->tqe_start)) { + /* The new data overlaps into the old */ + i = (th->th_seq + tlen) - first->tqe_start; +#ifdef TCP_REASS_LOGGING + tcp_log_reassm(tp, first, NULL, 0, i, TCP_R_LOG_TRIM, 1); +#endif + m_adj(first->tqe_m, i); + first->tqe_len -= i; + first->tqe_start += i; + } + /* Ok now setup our chain to point to the old first */ + mlast->m_next = first->tqe_m; + first->tqe_m = m; + first->tqe_len += tlen; + first->tqe_start = th->th_seq; + first->tqe_m->m_pkthdr.len = first->tqe_len; + first->tqe_mbuf_cnt += lenofoh; + tp->t_rcvoopack++; + TCPSTAT_INC(tcps_rcvoopack); + TCPSTAT_ADD(tcps_rcvoobyte, tlen); +#ifdef TCP_REASS_LOGGING + tcp_reass_log_new_in(tp, first->tqe_start, lenofoh, first->tqe_m, + TCP_R_LOG_PREPEND, + first); +#endif +} + +static void +tcp_reass_replace(struct tcpcb *tp, struct tseg_qent *q, struct mbuf *m, + tcp_seq seq, int len, struct mbuf *mlast, int mbufoh, uint8_t flags) +{ + /* + * Free the data in q, and replace + * it with the new segment. + */ + int len_dif; + +#ifdef TCP_REASS_LOGGING + tcp_log_reassm(tp, q, NULL, seq, len, TCP_R_LOG_REPLACE, 0); +#endif + m_freem(q->tqe_m); + KASSERT(tp->t_segqmbuflen >= q->tqe_mbuf_cnt, + ("Tp:%p seg queue goes negative", tp)); + tp->t_segqmbuflen -= q->tqe_mbuf_cnt; + q->tqe_mbuf_cnt = mbufoh; + q->tqe_m = m; + q->tqe_last = mlast; + q->tqe_start = seq; + if (len > q->tqe_len) + len_dif = len - q->tqe_len; + else + len_dif = 0; + tp->t_rcvoopack++; + TCPSTAT_INC(tcps_rcvoopack); + TCPSTAT_ADD(tcps_rcvoobyte, len_dif); + q->tqe_len = len; + q->tqe_flags = (flags & TH_FIN); + q->tqe_m->m_pkthdr.len = q->tqe_len; + tp->t_segqmbuflen += mbufoh; + +} + +static void +tcp_reass_merge_into(struct tcpcb *tp, struct tseg_qent *ent, + struct tseg_qent *q) +{ + /* + * Merge q into ent and free q from the list. + */ +#ifdef TCP_REASS_LOGGING + tcp_log_reassm(tp, q, ent, 0, 0, TCP_R_LOG_MERGE_INTO, 0); +#endif +#ifdef TCP_REASS_COUNTERS + counter_u64_add(merge_into, 1); +#endif + ent->tqe_last->m_next = q->tqe_m; + ent->tqe_last = q->tqe_last; + ent->tqe_len += q->tqe_len; + ent->tqe_mbuf_cnt += q->tqe_mbuf_cnt; + ent->tqe_m->m_pkthdr.len += q->tqe_len; + ent->tqe_flags |= (q->tqe_flags & TH_FIN); + TAILQ_REMOVE(&tp->t_segq, q, tqe_q); + uma_zfree(tcp_reass_zone, q); + tp->t_segqlen--; + +} + +static void +tcp_reass_merge_forward(struct tcpcb *tp, struct tseg_qent *ent) +{ + struct tseg_qent *q, *qtmp; + int i; + tcp_seq max; + /* + * Given an entry merge forward anyplace + * that ent overlaps forward. + */ + + max = ent->tqe_start + ent->tqe_len; + q = TAILQ_NEXT(ent, tqe_q); + if (q == NULL) { + /* Nothing left */ + return; + } + TAILQ_FOREACH_FROM_SAFE(q, &tp->t_segq, tqe_q, qtmp) { + if (SEQ_GT(q->tqe_start, max)) { + /* Beyond q */ + break; + } + /* We have some or all that are overlapping */ + if (SEQ_GEQ(max, (q->tqe_start + q->tqe_len))) { + /* It consumes it all */ + tp->t_segqmbuflen -= q->tqe_mbuf_cnt; + m_freem(q->tqe_m); + TAILQ_REMOVE(&tp->t_segq, q, tqe_q); + uma_zfree(tcp_reass_zone, q); + tp->t_segqlen--; + continue; + } + /* + * Trim the q entry to dovetail to this one + * and then merge q into ent updating max + * in the process. + */ + i = max - q->tqe_start; +#ifdef TCP_REASS_LOGGING + tcp_log_reassm(tp, q, NULL, 0, i, TCP_R_LOG_TRIM, 2); +#endif + m_adj(q->tqe_m, i); + q->tqe_len -= i; + q->tqe_start += i; + tcp_reass_merge_into(tp, ent, q); + max = ent->tqe_start + ent->tqe_len; + } +#ifdef TCP_REASS_COUNTERS + counter_u64_add(merge_fwd, 1); +#endif +} + +static int +tcp_reass_overhead_of_chain(struct mbuf *m, struct mbuf **mlast) +{ + int len = MSIZE; + + if (m->m_flags & M_EXT) + len += m->m_ext.ext_size; + while (m->m_next != NULL) { + m = m->m_next; + len += MSIZE; + if (m->m_flags & M_EXT) + len += m->m_ext.ext_size; + } + *mlast = m; + return (len); +} + + +/* + * NOTE!!! the new tcp-reassembly code *must not* use + * m_adj() with a negative index. That alters the chain + * of mbufs (by possibly chopping trailing mbufs). At + * the front of tcp_reass we count the mbuf overhead + * and setup the tail pointer. If we use m_adj(m, -5) + * we could corrupt the tail pointer. Currently the + * code only uses m_adj(m, postive-num). If this + * changes appropriate changes to update mlast would + * be needed. + */ int -tcp_reass(struct tcpcb *tp, struct tcphdr *th, int *tlenp, struct mbuf *m) +tcp_reass(struct tcpcb *tp, struct tcphdr *th, tcp_seq *seq_start, + int *tlenp, struct mbuf *m) { - struct tseg_qent *q; + struct tseg_qent *q, *last, *first; struct tseg_qent *p = NULL; - struct tseg_qent *nq; + struct tseg_qent *nq = NULL; struct tseg_qent *te = NULL; + struct tseg_qent tqs; + struct mbuf *mlast = NULL; + struct sockbuf *sb; struct socket *so = tp->t_inpcb->inp_socket; char *s = NULL; - int flags; - struct tseg_qent tqs; + int flags, i, lenofoh; INP_WLOCK_ASSERT(tp->t_inpcb); - /* * XXX: tcp_reass() is rather inefficient with its data structures * and should be rewritten (see NetBSD for optimizations). */ + KASSERT(th == NULL || (seq_start != NULL && tlenp != NULL), + ("tcp_reass called with illegal parameter combination " + "(tp=%p, th=%p, seq_start=%p, tlenp=%p, m=%p)", + tp, th, seq_start, tlenp, m)); /* * Call with th==NULL after become established to * force pre-ESTABLISHED data up to user socket. */ if (th == NULL) goto present; - + KASSERT(SEQ_GEQ(th->th_seq, tp->rcv_nxt), + ("Attempt to add old entry to reassembly queue (th=%p, tp=%p)", + th, tp)); +#ifdef TCP_REASS_LOGGING + tcp_reass_log_new_in(tp, th->th_seq, *tlenp, m, TCP_R_LOG_ADD, NULL); +#endif +#ifdef TCP_REASS_COUNTERS + counter_u64_add(reass_entry, 1); +#endif /* - * Limit the number of segments that can be queued to reduce the - * potential for mbuf exhaustion. For best performance, we want to be - * able to queue a full window's worth of segments. The size of the - * socket receive buffer determines our advertised window and grows - * automatically when socket buffer autotuning is enabled. Use it as the - * basis for our queue limit. - * - * However, allow the user to specify a ceiling for the number of - * segments in each queue. - * - * Always let the missing segment through which caused this queue. - * NB: Access to the socket buffer is left intentionally unlocked as we - * can tolerate stale information here. - * - * XXXLAS: Using sbspace(so->so_rcv) instead of so->so_rcv.sb_hiwat - * should work but causes packets to be dropped when they shouldn't. - * Investigate why and re-evaluate the below limit after the behaviour - * is understood. + * Check for zero length data. */ + if ((*tlenp == 0) && ((th->th_flags & TH_FIN) == 0)) { + /* + * A zero length segment does no + * one any good. We could check + * the rcv_nxt <-> rcv_wnd but thats + * already done for us by the caller. + */ +#ifdef TCP_REASS_COUNTERS + counter_u64_add(tcp_zero_input, 1); +#endif + m_freem(m); +#ifdef TCP_REASS_LOGGING + tcp_reass_log_dump(tp); +#endif + return (0); + } + /* + * Will it fit? + */ + lenofoh = tcp_reass_overhead_of_chain(m, &mlast); + sb = &tp->t_inpcb->inp_socket->so_rcv; if ((th->th_seq != tp->rcv_nxt || !TCPS_HAVEESTABLISHED(tp->t_state)) && - tp->t_segqlen >= min((so->so_rcv.sb_hiwat / tp->t_maxseg) + 1, - tcp_reass_maxqueuelen)) { + (sb->sb_mbcnt + tp->t_segqmbuflen + lenofoh) > sb->sb_mbmax) { + /* No room */ TCPSTAT_INC(tcps_rcvreassfull); - *tlenp = 0; +#ifdef TCP_REASS_COUNTERS + counter_u64_add(reass_nospace, 1); +#endif +#ifdef TCP_REASS_LOGGING + tcp_log_reassm(tp, NULL, NULL, th->th_seq, lenofoh, TCP_R_LOG_LIMIT_REACHED, 0); +#endif if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL, NULL))) { - log(LOG_DEBUG, "%s; %s: queue limit reached, " + log(LOG_DEBUG, "%s; %s: mbuf count limit reached, " "segment dropped\n", s, __func__); free(s, M_TCPLOG); } m_freem(m); + *tlenp = 0; +#ifdef TCP_REASS_LOGGING + tcp_reass_log_dump(tp); +#endif return (0); } - /* - * Allocate a new queue entry. If we can't, or hit the zone limit - * just drop the pkt. - * - * Use a temporary structure on the stack for the missing segment - * when the zone is exhausted. Otherwise we may get stuck. + * First lets deal with two common cases, the + * segment appends to the back of our collected + * segments. Or the segment is the next in line. */ - te = uma_zalloc(tcp_reass_zone, M_NOWAIT); - if (te == NULL) { - if (th->th_seq != tp->rcv_nxt || !TCPS_HAVEESTABLISHED(tp->t_state)) { - TCPSTAT_INC(tcps_rcvmemdrop); - m_freem(m); + last = TAILQ_LAST_FAST(&tp->t_segq, tseg_qent, tqe_q); + if (last != NULL) { + if ((th->th_flags & TH_FIN) && + SEQ_LT((th->th_seq + *tlenp), (last->tqe_start + last->tqe_len))) { + /* + * Someone is trying to game us, dump + * the segment. + */ *tlenp = 0; - if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL, - NULL))) { - log(LOG_DEBUG, "%s; %s: global zone limit " - "reached, segment dropped\n", s, __func__); - free(s, M_TCPLOG); + m_freem(m); + return (0); + } + if ((SEQ_GEQ(th->th_seq, last->tqe_start)) && + (SEQ_GEQ((last->tqe_start + last->tqe_len), th->th_seq))) { + /* Common case, trailing segment is added */ + /** + * +--last + * v + * reassembly buffer |---| |---| |---| + * new segment |---| + */ +#ifdef TCP_REASS_COUNTERS + counter_u64_add(reass_path1, 1); +#endif + if (SEQ_GT((last->tqe_start + last->tqe_len), th->th_seq)) { + i = (last->tqe_start + last->tqe_len) - th->th_seq; + if (i < *tlenp) { +#ifdef TCP_REASS_LOGGING + tcp_log_reassm(tp, last, NULL, 0, i, TCP_R_LOG_TRIM, 3); + th->th_seq += i; +#endif + m_adj(m, i); + *tlenp -= i; + } else { + /* Complete overlap */ + TCPSTAT_INC(tcps_rcvduppack); + TCPSTAT_ADD(tcps_rcvdupbyte, *tlenp); + m_freem(m); + *tlenp = last->tqe_len; + *seq_start = last->tqe_start; + return (0); + } } + if (last->tqe_flags & TH_FIN) { + /* + * We have data after the FIN on the last? + */ + *tlenp = 0; + m_freem(m); + return(0); + } + tcp_reass_append(tp, last, m, th, *tlenp, mlast, lenofoh); + tp->t_segqmbuflen += lenofoh; + *seq_start = last->tqe_start; + *tlenp = last->tqe_len; return (0); - } else { - bzero(&tqs, sizeof(struct tseg_qent)); - te = &tqs; - if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL, - NULL))) { - log(LOG_DEBUG, - "%s; %s: global zone limit reached, using " - "stack for missing segment\n", s, __func__); - free(s, M_TCPLOG); + } else if (SEQ_GT(th->th_seq, (last->tqe_start + last->tqe_len))) { + /* + * Second common case, we missed + * another one and have something more + * for the end. + */ + /** + * +--last + * v + * reassembly buffer |---| |---| |---| + * new segment |---| + */ + if (last->tqe_flags & TH_FIN) { + /* + * We have data after the FIN on the last? + */ + *tlenp = 0; + m_freem(m); + return(0); } +#ifdef TCP_REASS_COUNTERS + counter_u64_add(reass_path2, 1); +#endif + p = last; + goto new_entry; } + } else { + /* First segment (it's NULL). */ + goto new_entry; } - tp->t_segqlen++; + first = TAILQ_FIRST(&tp->t_segq); + if (SEQ_LT(th->th_seq, first->tqe_start) && + SEQ_GEQ((th->th_seq + *tlenp),first->tqe_start) && + SEQ_LT((th->th_seq + *tlenp), (first->tqe_start + first->tqe_len))) { + /* + * The head of the queue is prepended by this and + * it may be the one I want most. + */ + /** + * first-------+ + * v + * rea: |---| |---| |---| + * new: |---| + * Note the case we do not deal with here is: + * rea= |---| |---| |---| + * new= |----| + * Due to the fact that it could be + * new |--------------------| + * And we might need to merge forward. + */ +#ifdef INVARIANTS + struct mbuf *firstmbuf; +#endif +#ifdef TCP_REASS_COUNTERS + counter_u64_add(reass_path3, 1); +#endif + if (SEQ_LT(th->th_seq, tp->rcv_nxt)) { + /* + * The resend was even before + * what we have. We need to trim it. + * Note TSNH (it should be trimmed + * before the call to tcp_reass()). + */ +#ifdef INVARIANTS + panic("th->th_seq:%u rcv_nxt:%u tp:%p not pre-trimmed", + th->th_seq, tp->rcv_nxt, tp); +#else + i = tp->rcv_nxt - th->th_seq; +#ifdef TCP_REASS_LOGGING + tcp_log_reassm(tp, first, NULL, 0, i, TCP_R_LOG_TRIM, 4); +#endif + m_adj(m, i); + th->th_seq += i; + *tlenp -= i; +#endif + } +#ifdef INVARIANTS + firstmbuf = first->tqe_m; +#endif + tcp_reass_prepend(tp, first, m, th, *tlenp, mlast, lenofoh); +#ifdef INVARIANTS + if (firstmbuf == first->tqe_m) { + panic("First stayed same m:%p foobar:%p first->tqe_m:%p tp:%p first:%p", + m, firstmbuf, first->tqe_m, tp, first); + } else if (first->tqe_m != m) { + panic("First did not change to m:%p foobar:%p first->tqe_m:%p tp:%p first:%p", + m, firstmbuf, first->tqe_m, tp, first); + } +#endif + tp->t_segqmbuflen += lenofoh; + *seq_start = first->tqe_start; + *tlenp = first->tqe_len; + goto present; + } else if (SEQ_LT((th->th_seq + *tlenp), first->tqe_start)) { + /* New segment is before our earliest segment. */ + /** + * first---->+ + * v + * rea= |---| .... + * new" |---| + * + */ + goto new_entry; + } /* * Find a segment which begins after this one does. */ - LIST_FOREACH(q, &tp->t_segq, tqe_q) { - if (SEQ_GT(q->tqe_th->th_seq, th->th_seq)) +#ifdef TCP_REASS_COUNTERS + counter_u64_add(reass_fullwalk, 1); +#endif + TAILQ_FOREACH(q, &tp->t_segq, tqe_q) { + if (SEQ_GT(q->tqe_start, th->th_seq)) break; - p = q; } - - /* - * If there is a preceding segment, it may provide some of - * our data already. If so, drop the data from the incoming - * segment. If it provides all of our data, drop us. + p = TAILQ_PREV(q, tsegqe_head, tqe_q); + /** + * Now is this fit just in-between only? + * i.e.: + * p---+ +----q + * v v + * res= |--| |--| |--| + * nee |-| */ - if (p != NULL) { - int i; + if (SEQ_LT((th->th_seq + *tlenp), q->tqe_start) && + ((p == NULL) || (SEQ_GT(th->th_seq, (p->tqe_start + p->tqe_len))))) { + /* Yep no overlap */ + goto new_entry; + } + /** + * If we reach here we have some (possibly all) overlap + * such as: + * res= |--| |--| |--| + * new= |----| + * or new= |-----------------| + * or new= |--------| + * or new= |---| + * or new= |-----------| + */ + if ((p != NULL) && + (SEQ_LEQ(th->th_seq, (p->tqe_start + p->tqe_len)))) { /* conversion to int (in i) handles seq wraparound */ - i = p->tqe_th->th_seq + p->tqe_len - th->th_seq; - if (i > 0) { + +#ifdef TCP_REASS_COUNTERS + counter_u64_add(reass_path4, 1); +#endif + i = p->tqe_start + p->tqe_len - th->th_seq; + if (i >= 0) { if (i >= *tlenp) { + /** + * prev seg---->+ + * v + * reassembly buffer |---| + * new segment |-| + */ TCPSTAT_INC(tcps_rcvduppack); TCPSTAT_ADD(tcps_rcvdupbyte, *tlenp); + *tlenp = p->tqe_len; + *seq_start = p->tqe_start; m_freem(m); - if (te != &tqs) - uma_zfree(tcp_reass_zone, te); - tp->t_segqlen--; /* * Try to present any queued data * at the left window edge to the user. * This is needed after the 3-WHS - * completes. + * completes. Note this probably + * will not work and we will return. */ - goto present; /* ??? */ + return (0); } - m_adj(m, i); - *tlenp -= i; - th->th_seq += i; + if (i > 0) { + /** + * prev seg---->+ + * v + * reassembly buffer |---| + * new segment |-----| + */ +#ifdef TCP_REASS_COUNTERS + counter_u64_add(reass_path5, 1); +#endif +#ifdef TCP_REASS_LOGGING + tcp_log_reassm(tp, p, NULL, 0, i, TCP_R_LOG_TRIM, 5); +#endif + m_adj(m, i); + *tlenp -= i; + th->th_seq += i; + } } + if (th->th_seq == (p->tqe_start + p->tqe_len)) { + /* + * If dovetails in with this one + * append it. + */ + /** + * prev seg---->+ + * v + * reassembly buffer |--| |---| + * new segment |--| + * (note: it was trimmed above if it overlapped) + */ + tcp_reass_append(tp, p, m, th, *tlenp, mlast, lenofoh); + tp->t_segqmbuflen += lenofoh; + } else { +#ifdef INVARIANTS + panic("Impossible cut th_seq:%u p->seq:%u(%d) p:%p tp:%p", + th->th_seq, p->tqe_start, p->tqe_len, + p, tp); +#endif + *tlenp = 0; + m_freem(m); + return (0); + } + q = p; + } else { + /* + * The new data runs over the + * top of previously sack'd data (in q). + * It may be partially overlapping, or + * it may overlap the entire segment. + */ +#ifdef TCP_REASS_COUNTERS + counter_u64_add(reass_path6, 1); +#endif + if (SEQ_GEQ((th->th_seq + *tlenp), (q->tqe_start + q->tqe_len))) { + /* It consumes it all */ + /** + * next seg---->+ + * v + * reassembly buffer |--| |---| + * new segment |----------| + */ +#ifdef TCP_REASS_COUNTERS + counter_u64_add(reass_path7, 1); +#endif + tcp_reass_replace(tp, q, m, th->th_seq, *tlenp, mlast, lenofoh, th->th_flags); + } else { + /* + * We just need to prepend the data + * to this. It does not overrun + * the end. + */ + /** + * next seg---->+ + * v + * reassembly buffer |--| |---| + * new segment |----------| + */ + tcp_reass_prepend(tp, q, m, th, *tlenp, mlast, lenofoh); + tp->t_segqmbuflen += lenofoh; + } } - tp->t_rcvoopack++; - TCPSTAT_INC(tcps_rcvoopack); - TCPSTAT_ADD(tcps_rcvoobyte, *tlenp); + /* Now does it go further than that? */ + tcp_reass_merge_forward(tp, q); + *seq_start = q->tqe_start; + *tlenp = q->tqe_len; + goto present; + /* + * When we reach here we can't combine it + * with any existing segment. + * + * Limit the number of segments that can be queued to reduce the + * potential for mbuf exhaustion. For best performance, we want to be + * able to queue a full window's worth of segments. The size of the + * socket receive buffer determines our advertised window and grows + * automatically when socket buffer autotuning is enabled. Use it as the + * basis for our queue limit. + * + * However, allow the user to specify a ceiling for the number of + * segments in each queue. + * + * Always let the missing segment through which caused this queue. + * NB: Access to the socket buffer is left intentionally unlocked as we + * can tolerate stale information here. + * + * XXXLAS: Using sbspace(so->so_rcv) instead of so->so_rcv.sb_hiwat + * should work but causes packets to be dropped when they shouldn't. + * Investigate why and re-evaluate the below limit after the behaviour + * is understood. + */ +new_entry: + if (th->th_seq == tp->rcv_nxt && TCPS_HAVEESTABLISHED(tp->t_state)) { + tp->rcv_nxt += *tlenp; + flags = th->th_flags & TH_FIN; + TCPSTAT_INC(tcps_rcvoopack); + TCPSTAT_ADD(tcps_rcvoobyte, *tlenp); + SOCKBUF_LOCK(&so->so_rcv); + if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { + m_freem(m); + } else { + sbappendstream_locked(&so->so_rcv, m, 0); *** DIFF OUTPUT TRUNCATED AT 1000 LINES ***