2 * SPDX-License-Identifier: BSD-2-Clause
4 * Copyright (c) 2007, Myricom Inc.
5 * Copyright (c) 2008, Intel Corporation.
6 * Copyright (c) 2012 The FreeBSD Foundation
7 * Copyright (c) 2016-2021 Mellanox Technologies.
10 * Portions of this software were developed by Bjoern Zeeb
11 * under sponsorship from the FreeBSD Foundation.
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
22 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 #include <sys/cdefs.h>
37 #include "opt_inet6.h"
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/malloc.h>
44 #include <sys/socket.h>
45 #include <sys/socketvar.h>
46 #include <sys/sockbuf.h>
47 #include <sys/sysctl.h>
50 #include <net/if_var.h>
51 #include <net/ethernet.h>
54 #include <net/if_dl.h>
55 #include <net/if_media.h>
56 #include <net/if_private.h>
57 #include <net/if_types.h>
58 #include <net/infiniband.h>
59 #include <net/if_lagg.h>
61 #include <netinet/in_systm.h>
62 #include <netinet/in.h>
63 #include <netinet/ip6.h>
64 #include <netinet/ip.h>
65 #include <netinet/ip_var.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet6/in6_pcb.h>
68 #include <netinet/tcp.h>
69 #include <netinet/tcp_seq.h>
70 #include <netinet/tcp_lro.h>
71 #include <netinet/tcp_var.h>
72 #include <netinet/tcpip.h>
73 #include <netinet/tcp_hpts.h>
74 #include <netinet/tcp_log_buf.h>
75 #include <netinet/tcp_fsm.h>
76 #include <netinet/udp.h>
77 #include <netinet6/ip6_var.h>
79 #include <machine/in_cksum.h>
81 static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures");
83 #define TCP_LRO_TS_OPTION \
84 ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
85 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
87 static void tcp_lro_rx_done(struct lro_ctrl *lc);
88 static int tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m,
89 uint32_t csum, bool use_hash);
92 static bool do_bpf_strip_and_compress(struct tcpcb *, struct lro_ctrl *,
93 struct lro_entry *, struct mbuf **, struct mbuf **, struct mbuf **,
94 bool *, bool, bool, struct ifnet *, bool);
98 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
101 static long tcplro_stacks_wanting_mbufq;
102 counter_u64_t tcp_inp_lro_direct_queue;
103 counter_u64_t tcp_inp_lro_wokeup_queue;
104 counter_u64_t tcp_inp_lro_compressed;
105 counter_u64_t tcp_inp_lro_locks_taken;
106 counter_u64_t tcp_extra_mbuf;
107 counter_u64_t tcp_would_have_but;
108 counter_u64_t tcp_comp_total;
109 counter_u64_t tcp_uncomp_total;
110 counter_u64_t tcp_bad_csums;
112 static unsigned tcp_lro_entries = TCP_LRO_ENTRIES;
113 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries,
114 CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_entries, 0,
115 "default number of LRO entries");
117 static uint32_t tcp_lro_cpu_set_thresh = TCP_LRO_CPU_DECLARATION_THRESH;
118 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_cpu_threshold,
119 CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_cpu_set_thresh, 0,
120 "Number of interrupts in a row on the same CPU that will make us declare an 'affinity' cpu?");
122 static uint32_t tcp_less_accurate_lro_ts = 0;
123 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_less_accurate,
124 CTLFLAG_MPSAFE, &tcp_less_accurate_lro_ts, 0,
125 "Do we trade off efficency by doing less timestamp operations for time accuracy?");
127 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, fullqueue, CTLFLAG_RD,
128 &tcp_inp_lro_direct_queue, "Number of lro's fully queued to transport");
129 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, wokeup, CTLFLAG_RD,
130 &tcp_inp_lro_wokeup_queue, "Number of lro's where we woke up transport via hpts");
131 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, compressed, CTLFLAG_RD,
132 &tcp_inp_lro_compressed, "Number of lro's compressed and sent to transport");
133 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lockcnt, CTLFLAG_RD,
134 &tcp_inp_lro_locks_taken, "Number of lro's inp_wlocks taken");
135 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, extra_mbuf, CTLFLAG_RD,
136 &tcp_extra_mbuf, "Number of times we had an extra compressed ack dropped into the tp");
137 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, would_have_but, CTLFLAG_RD,
138 &tcp_would_have_but, "Number of times we would have had an extra compressed, but mget failed");
139 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, with_m_ackcmp, CTLFLAG_RD,
140 &tcp_comp_total, "Number of mbufs queued with M_ACKCMP flags set");
141 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, without_m_ackcmp, CTLFLAG_RD,
142 &tcp_uncomp_total, "Number of mbufs queued without M_ACKCMP");
143 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lro_badcsum, CTLFLAG_RD,
144 &tcp_bad_csums, "Number of packets that the common code saw with bad csums");
147 tcp_lro_reg_mbufq(void)
149 atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, 1);
153 tcp_lro_dereg_mbufq(void)
155 atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, -1);
159 tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket,
160 struct lro_entry *le)
163 LIST_INSERT_HEAD(&lc->lro_active, le, next);
164 LIST_INSERT_HEAD(bucket, le, hash_next);
168 tcp_lro_active_remove(struct lro_entry *le)
171 LIST_REMOVE(le, next); /* active list */
172 LIST_REMOVE(le, hash_next); /* hash bucket */
176 tcp_lro_init(struct lro_ctrl *lc)
178 return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0));
182 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
183 unsigned lro_entries, unsigned lro_mbufs)
185 struct lro_entry *le;
187 unsigned i, elements;
189 lc->lro_bad_csum = 0;
192 lc->lro_mbuf_count = 0;
193 lc->lro_mbuf_max = lro_mbufs;
194 lc->lro_cnt = lro_entries;
195 lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
196 lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
198 LIST_INIT(&lc->lro_free);
199 LIST_INIT(&lc->lro_active);
201 /* create hash table to accelerate entry lookup */
202 if (lro_entries > lro_mbufs)
203 elements = lro_entries;
205 elements = lro_mbufs;
206 lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz,
208 if (lc->lro_hash == NULL) {
209 memset(lc, 0, sizeof(*lc));
213 /* compute size to allocate */
214 size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
215 (lro_entries * sizeof(*le));
216 lc->lro_mbuf_data = (struct lro_mbuf_sort *)
217 malloc(size, M_LRO, M_NOWAIT | M_ZERO);
219 /* check for out of memory */
220 if (lc->lro_mbuf_data == NULL) {
221 free(lc->lro_hash, M_LRO);
222 memset(lc, 0, sizeof(*lc));
225 /* compute offset for LRO entries */
226 le = (struct lro_entry *)
227 (lc->lro_mbuf_data + lro_mbufs);
229 /* setup linked list */
230 for (i = 0; i != lro_entries; i++)
231 LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
236 struct vxlan_header {
242 tcp_lro_low_level_parser(void *ptr, struct lro_parser *parser, bool update_data, bool is_vxlan, int mlen)
244 const struct ether_vlan_header *eh;
249 memset(parser, 0, sizeof(*parser));
254 const struct vxlan_header *vxh;
256 ptr = (uint8_t *)ptr + sizeof(*vxh);
258 parser->data.vxlan_vni =
259 vxh->vxlh_vni & htonl(0xffffff00);
264 if (__predict_false(eh->evl_encap_proto == htons(ETHERTYPE_VLAN))) {
265 eth_type = eh->evl_proto;
267 /* strip priority and keep VLAN ID only */
268 parser->data.vlan_id = eh->evl_tag & htons(EVL_VLID_MASK);
270 /* advance to next header */
271 ptr = (uint8_t *)ptr + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
272 mlen -= (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
274 eth_type = eh->evl_encap_proto;
275 /* advance to next header */
276 mlen -= ETHER_HDR_LEN;
277 ptr = (uint8_t *)ptr + ETHER_HDR_LEN;
279 if (__predict_false(mlen <= 0))
283 case htons(ETHERTYPE_IP):
285 if (__predict_false(mlen < sizeof(struct ip)))
287 /* Ensure there are no IPv4 options. */
288 if ((parser->ip4->ip_hl << 2) != sizeof (*parser->ip4))
290 /* .. and the packet is not fragmented. */
291 if (parser->ip4->ip_off & htons(IP_MF|IP_OFFMASK))
293 /* .. and the packet has valid src/dst addrs */
294 if (__predict_false(parser->ip4->ip_src.s_addr == INADDR_ANY ||
295 parser->ip4->ip_dst.s_addr == INADDR_ANY))
297 ptr = (uint8_t *)ptr + (parser->ip4->ip_hl << 2);
298 mlen -= sizeof(struct ip);
300 parser->data.s_addr.v4 = parser->ip4->ip_src;
301 parser->data.d_addr.v4 = parser->ip4->ip_dst;
303 switch (parser->ip4->ip_p) {
305 if (__predict_false(mlen < sizeof(struct udphdr)))
309 parser->data.lro_type = LRO_TYPE_IPV4_UDP;
310 parser->data.s_port = parser->udp->uh_sport;
311 parser->data.d_port = parser->udp->uh_dport;
313 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_UDP);
315 ptr = ((uint8_t *)ptr + sizeof(*parser->udp));
316 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
320 if (__predict_false(mlen < sizeof(struct tcphdr)))
323 parser->data.lro_type = LRO_TYPE_IPV4_TCP;
324 parser->data.s_port = parser->tcp->th_sport;
325 parser->data.d_port = parser->tcp->th_dport;
327 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_TCP);
329 if (__predict_false(mlen < (parser->tcp->th_off << 2)))
331 ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
332 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
340 case htons(ETHERTYPE_IPV6):
342 if (__predict_false(mlen < sizeof(struct ip6_hdr)))
344 /* Ensure the packet has valid src/dst addrs */
345 if (__predict_false(IN6_IS_ADDR_UNSPECIFIED(&parser->ip6->ip6_src) ||
346 IN6_IS_ADDR_UNSPECIFIED(&parser->ip6->ip6_dst)))
348 ptr = (uint8_t *)ptr + sizeof(*parser->ip6);
350 parser->data.s_addr.v6 = parser->ip6->ip6_src;
351 parser->data.d_addr.v6 = parser->ip6->ip6_dst;
353 mlen -= sizeof(struct ip6_hdr);
354 switch (parser->ip6->ip6_nxt) {
356 if (__predict_false(mlen < sizeof(struct udphdr)))
360 parser->data.lro_type = LRO_TYPE_IPV6_UDP;
361 parser->data.s_port = parser->udp->uh_sport;
362 parser->data.d_port = parser->udp->uh_dport;
364 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_UDP);
366 ptr = (uint8_t *)ptr + sizeof(*parser->udp);
367 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
370 if (__predict_false(mlen < sizeof(struct tcphdr)))
374 parser->data.lro_type = LRO_TYPE_IPV6_TCP;
375 parser->data.s_port = parser->tcp->th_sport;
376 parser->data.d_port = parser->tcp->th_dport;
378 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_TCP);
380 if (__predict_false(mlen < (parser->tcp->th_off << 2)))
382 ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
383 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
393 /* Invalid packet - cannot parse */
397 static const int vxlan_csum = CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
398 CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID;
400 static inline struct lro_parser *
401 tcp_lro_parser(struct mbuf *m, struct lro_parser *po, struct lro_parser *pi, bool update_data)
405 /* Try to parse outer headers first. */
406 data_ptr = tcp_lro_low_level_parser(m->m_data, po, update_data, false, m->m_len);
407 if (data_ptr == NULL || po->total_hdr_len > m->m_len)
411 /* Store VLAN ID, if any. */
412 if (__predict_false(m->m_flags & M_VLANTAG)) {
414 htons(m->m_pkthdr.ether_vtag) & htons(EVL_VLID_MASK);
416 /* Store decrypted flag, if any. */
417 if (__predict_false((m->m_pkthdr.csum_flags &
418 CSUM_TLS_MASK) == CSUM_TLS_DECRYPTED))
419 po->data.lro_flags |= LRO_FLAG_DECRYPTED;
422 switch (po->data.lro_type) {
423 case LRO_TYPE_IPV4_UDP:
424 case LRO_TYPE_IPV6_UDP:
425 /* Check for VXLAN headers. */
426 if ((m->m_pkthdr.csum_flags & vxlan_csum) != vxlan_csum)
429 /* Try to parse inner headers. */
430 data_ptr = tcp_lro_low_level_parser(data_ptr, pi, update_data, true,
431 (m->m_len - ((caddr_t)data_ptr - m->m_data)));
432 if (data_ptr == NULL || (pi->total_hdr_len + po->total_hdr_len) > m->m_len)
435 /* Verify supported header types. */
436 switch (pi->data.lro_type) {
437 case LRO_TYPE_IPV4_TCP:
438 case LRO_TYPE_IPV6_TCP:
444 case LRO_TYPE_IPV4_TCP:
445 case LRO_TYPE_IPV6_TCP:
447 memset(pi, 0, sizeof(*pi));
456 tcp_lro_trim_mbuf_chain(struct mbuf *m, const struct lro_parser *po)
460 switch (po->data.lro_type) {
462 case LRO_TYPE_IPV4_TCP:
463 len = ((uint8_t *)po->ip4 - (uint8_t *)m->m_data) +
464 ntohs(po->ip4->ip_len);
468 case LRO_TYPE_IPV6_TCP:
469 len = ((uint8_t *)po->ip6 - (uint8_t *)m->m_data) +
470 ntohs(po->ip6->ip6_plen) + sizeof(*po->ip6);
474 return (TCP_LRO_CANNOT);
478 * If the frame is padded beyond the end of the IP packet,
479 * then trim the extra bytes off:
481 if (__predict_true(m->m_pkthdr.len == len)) {
483 } else if (m->m_pkthdr.len > len) {
484 m_adj(m, len - m->m_pkthdr.len);
487 return (TCP_LRO_CANNOT);
490 static struct tcphdr *
491 tcp_lro_get_th(struct mbuf *m)
493 return ((struct tcphdr *)((uint8_t *)m->m_data + m->m_pkthdr.lro_tcp_h_off));
497 lro_free_mbuf_chain(struct mbuf *m)
510 tcp_lro_free(struct lro_ctrl *lc)
512 struct lro_entry *le;
515 /* reset LRO free list */
516 LIST_INIT(&lc->lro_free);
518 /* free active mbufs, if any */
519 while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
520 tcp_lro_active_remove(le);
521 lro_free_mbuf_chain(le->m_head);
524 /* free hash table */
525 free(lc->lro_hash, M_LRO);
529 /* free mbuf array, if any */
530 for (x = 0; x != lc->lro_mbuf_count; x++)
531 m_freem(lc->lro_mbuf_data[x].mb);
532 lc->lro_mbuf_count = 0;
534 /* free allocated memory, if any */
535 free(lc->lro_mbuf_data, M_LRO);
536 lc->lro_mbuf_data = NULL;
540 tcp_lro_rx_csum_tcphdr(const struct tcphdr *th)
546 csum = -th->th_sum; /* exclude checksum field */
548 ptr = (const uint16_t *)th;
555 while (csum > 0xffff)
556 csum = (csum >> 16) + (csum & 0xffff);
562 tcp_lro_rx_csum_data(const struct lro_parser *pa, uint16_t tcp_csum)
569 switch (pa->data.lro_type) {
571 case LRO_TYPE_IPV6_TCP:
572 /* Compute full pseudo IPv6 header checksum. */
573 cs = in6_cksum_pseudo(pa->ip6, ntohs(pa->ip6->ip6_plen), pa->ip6->ip6_nxt, 0);
577 case LRO_TYPE_IPV4_TCP:
578 /* Compute full pseudo IPv4 header checsum. */
579 cs = in_addword(ntohs(pa->ip4->ip_len) - sizeof(*pa->ip4), IPPROTO_TCP);
580 cs = in_pseudo(pa->ip4->ip_src.s_addr, pa->ip4->ip_dst.s_addr, htons(cs));
584 cs = 0; /* Keep compiler happy. */
588 /* Complement checksum. */
592 /* Remove TCP header checksum. */
593 cs = ~tcp_lro_rx_csum_tcphdr(pa->tcp);
596 /* Compute checksum remainder. */
598 c = (c >> 16) + (c & 0xffff);
604 tcp_lro_rx_done(struct lro_ctrl *lc)
606 struct lro_entry *le;
608 while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
609 tcp_lro_active_remove(le);
610 tcp_lro_flush(lc, le);
615 tcp_lro_flush_active(struct lro_ctrl *lc)
617 struct lro_entry *le;
620 * Walk through the list of le entries, and
621 * any one that does have packets flush. This
622 * is called because we have an inbound packet
623 * (e.g. SYN) that has to have all others flushed
624 * in front of it. Note we have to do the remove
625 * because tcp_lro_flush() assumes that the entry
626 * is being freed. This is ok it will just get
627 * reallocated again like it was new.
629 LIST_FOREACH(le, &lc->lro_active, next) {
630 if (le->m_head != NULL) {
631 tcp_lro_active_remove(le);
632 tcp_lro_flush(lc, le);
638 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
640 struct lro_entry *le, *le_tmp;
645 if (LIST_EMPTY(&lc->lro_active))
648 /* get timeout time and current time in ns */
650 now = bintime2ns(&bt);
651 tov = ((timeout->tv_sec * 1000000000) + (timeout->tv_usec * 1000));
652 LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
653 if (now >= (bintime2ns(&le->alloc_time) + tov)) {
654 tcp_lro_active_remove(le);
655 tcp_lro_flush(lc, le);
662 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4)
666 /* Legacy IP has a header checksum that needs to be correct. */
667 if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) {
668 if (__predict_false((m->m_pkthdr.csum_flags & CSUM_IP_VALID) == 0)) {
670 return (TCP_LRO_CANNOT);
673 csum = in_cksum_hdr(ip4);
674 if (__predict_false(csum != 0)) {
676 return (TCP_LRO_CANNOT);
685 tcp_lro_log(struct tcpcb *tp, const struct lro_ctrl *lc,
686 const struct lro_entry *le, const struct mbuf *m,
687 int frm, int32_t tcp_data_len, uint32_t th_seq,
688 uint32_t th_ack, uint16_t th_win)
690 if (tcp_bblogging_on(tp)) {
691 union tcp_log_stackspecific log;
692 struct timeval tv, btv;
695 cts = tcp_get_usecs(&tv);
696 memset(&log, 0, sizeof(union tcp_log_stackspecific));
697 log.u_bbr.flex8 = frm;
698 log.u_bbr.flex1 = tcp_data_len;
700 log.u_bbr.flex2 = m->m_pkthdr.len;
704 log.u_bbr.flex3 = le->m_head->m_pkthdr.lro_nsegs;
705 log.u_bbr.flex4 = le->m_head->m_pkthdr.lro_tcp_d_len;
706 log.u_bbr.flex5 = le->m_head->m_pkthdr.len;
707 log.u_bbr.delRate = le->m_head->m_flags;
708 log.u_bbr.rttProp = le->m_head->m_pkthdr.rcv_tstmp;
710 log.u_bbr.inflight = th_seq;
711 log.u_bbr.delivered = th_ack;
712 log.u_bbr.timeStamp = cts;
713 log.u_bbr.epoch = le->next_seq;
714 log.u_bbr.lt_epoch = le->ack_seq;
715 log.u_bbr.pacing_gain = th_win;
716 log.u_bbr.cwnd_gain = le->window;
717 log.u_bbr.lost = curcpu;
718 log.u_bbr.cur_del_rate = (uintptr_t)m;
719 log.u_bbr.bw_inuse = (uintptr_t)le->m_head;
720 bintime2timeval(&lc->lro_last_queue_time, &btv);
721 log.u_bbr.flex6 = tcp_tv_to_usectick(&btv);
722 log.u_bbr.flex7 = le->compressed;
723 log.u_bbr.pacing_gain = le->uncompressed;
724 if (in_epoch(net_epoch_preempt))
725 log.u_bbr.inhpts = 1;
727 log.u_bbr.inhpts = 0;
728 TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
729 &tptosocket(tp)->so_snd,
730 TCP_LOG_LRO, 0, 0, &log, false, &tv);
736 tcp_lro_assign_and_checksum_16(uint16_t *ptr, uint16_t value, uint16_t *psum)
740 csum = 0xffff - *ptr + value;
741 while (csum > 0xffff)
742 csum = (csum >> 16) + (csum & 0xffff);
748 tcp_lro_update_checksum(const struct lro_parser *pa, const struct lro_entry *le,
749 uint16_t payload_len, uint16_t delta_sum)
753 uint16_t temp[5] = {};
755 switch (pa->data.lro_type) {
756 case LRO_TYPE_IPV4_TCP:
757 /* Compute new IPv4 length. */
758 tlen = (pa->ip4->ip_hl << 2) + (pa->tcp->th_off << 2) + payload_len;
759 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
761 /* Subtract delta from current IPv4 checksum. */
762 csum = pa->ip4->ip_sum + 0xffff - temp[0];
763 while (csum > 0xffff)
764 csum = (csum >> 16) + (csum & 0xffff);
765 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
766 goto update_tcp_header;
768 case LRO_TYPE_IPV6_TCP:
769 /* Compute new IPv6 length. */
770 tlen = (pa->tcp->th_off << 2) + payload_len;
771 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
772 goto update_tcp_header;
774 case LRO_TYPE_IPV4_UDP:
775 /* Compute new IPv4 length. */
776 tlen = (pa->ip4->ip_hl << 2) + sizeof(*pa->udp) + payload_len;
777 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
779 /* Subtract delta from current IPv4 checksum. */
780 csum = pa->ip4->ip_sum + 0xffff - temp[0];
781 while (csum > 0xffff)
782 csum = (csum >> 16) + (csum & 0xffff);
783 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
784 goto update_udp_header;
786 case LRO_TYPE_IPV6_UDP:
787 /* Compute new IPv6 length. */
788 tlen = sizeof(*pa->udp) + payload_len;
789 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
790 goto update_udp_header;
797 /* Compute current TCP header checksum. */
798 temp[2] = tcp_lro_rx_csum_tcphdr(pa->tcp);
800 /* Incorporate the latest ACK into the TCP header. */
801 pa->tcp->th_ack = le->ack_seq;
802 pa->tcp->th_win = le->window;
804 /* Incorporate latest timestamp into the TCP header. */
805 if (le->timestamp != 0) {
808 ts_ptr = (uint32_t *)(pa->tcp + 1);
809 ts_ptr[1] = htonl(le->tsval);
810 ts_ptr[2] = le->tsecr;
813 /* Compute new TCP header checksum. */
814 temp[3] = tcp_lro_rx_csum_tcphdr(pa->tcp);
816 /* Compute new TCP checksum. */
817 csum = pa->tcp->th_sum + 0xffff - delta_sum +
818 0xffff - temp[0] + 0xffff - temp[3] + temp[2];
819 while (csum > 0xffff)
820 csum = (csum >> 16) + (csum & 0xffff);
822 /* Assign new TCP checksum. */
823 tcp_lro_assign_and_checksum_16(&pa->tcp->th_sum, csum, &temp[4]);
825 /* Compute all modififications affecting next checksum. */
826 csum = temp[0] + temp[1] + 0xffff - temp[2] +
827 temp[3] + temp[4] + delta_sum;
828 while (csum > 0xffff)
829 csum = (csum >> 16) + (csum & 0xffff);
831 /* Return delta checksum to next stage, if any. */
835 tlen = sizeof(*pa->udp) + payload_len;
836 /* Assign new UDP length and compute checksum delta. */
837 tcp_lro_assign_and_checksum_16(&pa->udp->uh_ulen, htons(tlen), &temp[2]);
839 /* Check if there is a UDP checksum. */
840 if (__predict_false(pa->udp->uh_sum != 0)) {
841 /* Compute new UDP checksum. */
842 csum = pa->udp->uh_sum + 0xffff - delta_sum +
843 0xffff - temp[0] + 0xffff - temp[2];
844 while (csum > 0xffff)
845 csum = (csum >> 16) + (csum & 0xffff);
846 /* Assign new UDP checksum. */
847 tcp_lro_assign_and_checksum_16(&pa->udp->uh_sum, csum, &temp[3]);
850 /* Compute all modififications affecting next checksum. */
851 csum = temp[0] + temp[1] + temp[2] + temp[3] + delta_sum;
852 while (csum > 0xffff)
853 csum = (csum >> 16) + (csum & 0xffff);
855 /* Return delta checksum to next stage, if any. */
860 tcp_flush_out_entry(struct lro_ctrl *lc, struct lro_entry *le)
862 /* Check if we need to recompute any checksums. */
863 if (le->needs_merge) {
866 switch (le->inner.data.lro_type) {
867 case LRO_TYPE_IPV4_TCP:
868 csum = tcp_lro_update_checksum(&le->inner, le,
869 le->m_head->m_pkthdr.lro_tcp_d_len,
870 le->m_head->m_pkthdr.lro_tcp_d_csum);
871 csum = tcp_lro_update_checksum(&le->outer, NULL,
872 le->m_head->m_pkthdr.lro_tcp_d_len +
873 le->inner.total_hdr_len, csum);
874 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
875 CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
876 le->m_head->m_pkthdr.csum_data = 0xffff;
877 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
878 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
880 case LRO_TYPE_IPV6_TCP:
881 csum = tcp_lro_update_checksum(&le->inner, le,
882 le->m_head->m_pkthdr.lro_tcp_d_len,
883 le->m_head->m_pkthdr.lro_tcp_d_csum);
884 csum = tcp_lro_update_checksum(&le->outer, NULL,
885 le->m_head->m_pkthdr.lro_tcp_d_len +
886 le->inner.total_hdr_len, csum);
887 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
889 le->m_head->m_pkthdr.csum_data = 0xffff;
890 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
891 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
894 switch (le->outer.data.lro_type) {
895 case LRO_TYPE_IPV4_TCP:
896 csum = tcp_lro_update_checksum(&le->outer, le,
897 le->m_head->m_pkthdr.lro_tcp_d_len,
898 le->m_head->m_pkthdr.lro_tcp_d_csum);
899 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
900 CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
901 le->m_head->m_pkthdr.csum_data = 0xffff;
902 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
903 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
905 case LRO_TYPE_IPV6_TCP:
906 csum = tcp_lro_update_checksum(&le->outer, le,
907 le->m_head->m_pkthdr.lro_tcp_d_len,
908 le->m_head->m_pkthdr.lro_tcp_d_csum);
909 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
911 le->m_head->m_pkthdr.csum_data = 0xffff;
912 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
913 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
925 * Break any chain, this is not set to NULL on the singleton
926 * case m_nextpkt points to m_head. Other case set them
927 * m_nextpkt to NULL in push_and_replace.
929 le->m_head->m_nextpkt = NULL;
930 lc->lro_queued += le->m_head->m_pkthdr.lro_nsegs;
931 (*lc->ifp->if_input)(lc->ifp, le->m_head);
935 tcp_set_entry_to_mbuf(struct lro_ctrl *lc, struct lro_entry *le,
936 struct mbuf *m, struct tcphdr *th)
939 uint16_t tcp_data_len;
940 uint16_t tcp_opt_len;
942 ts_ptr = (uint32_t *)(th + 1);
943 tcp_opt_len = (th->th_off << 2);
944 tcp_opt_len -= sizeof(*th);
946 /* Check if there is a timestamp option. */
947 if (tcp_opt_len == 0 ||
948 __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
949 *ts_ptr != TCP_LRO_TS_OPTION)) {
950 /* We failed to find the timestamp option. */
954 le->tsval = ntohl(*(ts_ptr + 1));
955 le->tsecr = *(ts_ptr + 2);
958 tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
960 /* Pull out TCP sequence numbers and window size. */
961 le->next_seq = ntohl(th->th_seq) + tcp_data_len;
962 le->ack_seq = th->th_ack;
963 le->window = th->th_win;
964 le->flags = tcp_get_flags(th);
967 /* Setup new data pointers. */
969 le->m_tail = m_last(m);
973 tcp_push_and_replace(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m)
975 struct lro_parser *pa;
978 * Push up the stack of the current entry
979 * and replace it with "m".
983 /* Grab off the next and save it */
984 msave = le->m_head->m_nextpkt;
985 le->m_head->m_nextpkt = NULL;
987 /* Now push out the old entry */
988 tcp_flush_out_entry(lc, le);
990 /* Re-parse new header, should not fail. */
991 pa = tcp_lro_parser(m, &le->outer, &le->inner, false);
993 ("tcp_push_and_replace: LRO parser failed on m=%p\n", m));
996 * Now to replace the data properly in the entry
997 * we have to reset the TCP header and
1000 tcp_set_entry_to_mbuf(lc, le, m, pa->tcp);
1002 /* Restore the next list */
1003 m->m_nextpkt = msave;
1007 tcp_lro_mbuf_append_pkthdr(struct lro_entry *le, const struct mbuf *p)
1013 if (m->m_pkthdr.lro_nsegs == 1) {
1014 /* Compute relative checksum. */
1015 csum = p->m_pkthdr.lro_tcp_d_csum;
1017 /* Merge TCP data checksums. */
1018 csum = (uint32_t)m->m_pkthdr.lro_tcp_d_csum +
1019 (uint32_t)p->m_pkthdr.lro_tcp_d_csum;
1020 while (csum > 0xffff)
1021 csum = (csum >> 16) + (csum & 0xffff);
1024 /* Update various counters. */
1025 m->m_pkthdr.len += p->m_pkthdr.lro_tcp_d_len;
1026 m->m_pkthdr.lro_tcp_d_csum = csum;
1027 m->m_pkthdr.lro_tcp_d_len += p->m_pkthdr.lro_tcp_d_len;
1028 m->m_pkthdr.lro_nsegs += p->m_pkthdr.lro_nsegs;
1029 le->needs_merge = 1;
1033 tcp_lro_condense(struct lro_ctrl *lc, struct lro_entry *le)
1036 * Walk through the mbuf chain we
1037 * have on tap and compress/condense
1043 uint32_t tcp_data_len_total;
1044 uint32_t tcp_data_seg_total;
1045 uint16_t tcp_data_len;
1046 uint16_t tcp_opt_len;
1049 * First we must check the lead (m_head)
1050 * we must make sure that it is *not*
1051 * something that should be sent up
1052 * right away (sack etc).
1055 m = le->m_head->m_nextpkt;
1057 /* Just one left. */
1061 th = tcp_lro_get_th(m);
1062 tcp_opt_len = (th->th_off << 2);
1063 tcp_opt_len -= sizeof(*th);
1064 ts_ptr = (uint32_t *)(th + 1);
1066 if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1067 *ts_ptr != TCP_LRO_TS_OPTION)) {
1069 * Its not the timestamp. We can't
1070 * use this guy as the head.
1072 le->m_head->m_nextpkt = m->m_nextpkt;
1073 tcp_push_and_replace(lc, le, m);
1076 if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1078 * Make sure that previously seen segments/ACKs are delivered
1079 * before this segment, e.g. FIN.
1081 le->m_head->m_nextpkt = m->m_nextpkt;
1082 tcp_push_and_replace(lc, le, m);
1085 while((m = le->m_head->m_nextpkt) != NULL) {
1087 * condense m into le, first
1088 * pull m out of the list.
1090 le->m_head->m_nextpkt = m->m_nextpkt;
1091 m->m_nextpkt = NULL;
1093 tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
1094 th = tcp_lro_get_th(m);
1095 ts_ptr = (uint32_t *)(th + 1);
1096 tcp_opt_len = (th->th_off << 2);
1097 tcp_opt_len -= sizeof(*th);
1098 tcp_data_len_total = le->m_head->m_pkthdr.lro_tcp_d_len + tcp_data_len;
1099 tcp_data_seg_total = le->m_head->m_pkthdr.lro_nsegs + m->m_pkthdr.lro_nsegs;
1101 if (tcp_data_seg_total >= lc->lro_ackcnt_lim ||
1102 tcp_data_len_total >= lc->lro_length_lim) {
1103 /* Flush now if appending will result in overflow. */
1104 tcp_push_and_replace(lc, le, m);
1107 if (tcp_opt_len != 0 &&
1108 __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1109 *ts_ptr != TCP_LRO_TS_OPTION)) {
1111 * Maybe a sack in the new one? We need to
1112 * start all over after flushing the
1113 * current le. We will go up to the beginning
1114 * and flush it (calling the replace again possibly
1115 * or just returning).
1117 tcp_push_and_replace(lc, le, m);
1120 if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1121 tcp_push_and_replace(lc, le, m);
1124 if (tcp_opt_len != 0) {
1125 uint32_t tsval = ntohl(*(ts_ptr + 1));
1126 /* Make sure timestamp values are increasing. */
1127 if (TSTMP_GT(le->tsval, tsval)) {
1128 tcp_push_and_replace(lc, le, m);
1132 le->tsecr = *(ts_ptr + 2);
1134 /* Try to append the new segment. */
1135 if (__predict_false(ntohl(th->th_seq) != le->next_seq ||
1136 ((tcp_get_flags(th) & TH_ACK) !=
1137 (le->flags & TH_ACK)) ||
1138 (tcp_data_len == 0 &&
1139 le->ack_seq == th->th_ack &&
1140 le->window == th->th_win))) {
1141 /* Out of order packet, non-ACK + ACK or dup ACK. */
1142 tcp_push_and_replace(lc, le, m);
1145 if (tcp_data_len != 0 ||
1146 SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1147 le->next_seq += tcp_data_len;
1148 le->ack_seq = th->th_ack;
1149 le->window = th->th_win;
1150 le->needs_merge = 1;
1151 } else if (th->th_ack == le->ack_seq) {
1152 if (WIN_GT(th->th_win, le->window)) {
1153 le->window = th->th_win;
1154 le->needs_merge = 1;
1158 if (tcp_data_len == 0) {
1163 /* Merge TCP data checksum and length to head mbuf. */
1164 tcp_lro_mbuf_append_pkthdr(le, m);
1167 * Adjust the mbuf so that m_data points to the first byte of
1168 * the ULP payload. Adjust the mbuf to avoid complications and
1169 * append new segment to existing mbuf chain.
1171 m_adj(m, m->m_pkthdr.len - tcp_data_len);
1173 le->m_tail->m_next = m;
1174 le->m_tail = m_last(m);
1180 tcp_queue_pkts(struct tcpcb *tp, struct lro_entry *le)
1183 INP_WLOCK_ASSERT(tptoinpcb(tp));
1185 STAILQ_HEAD(, mbuf) q = { le->m_head,
1186 &STAILQ_NEXT(le->m_last_mbuf, m_stailqpkt) };
1187 STAILQ_CONCAT(&tp->t_inqueue, &q);
1189 le->m_last_mbuf = NULL;
1193 tcp_lro_check_wake_status(struct tcpcb *tp)
1196 if (tp->t_fb->tfb_early_wake_check != NULL)
1197 return ((tp->t_fb->tfb_early_wake_check)(tp));
1201 static struct mbuf *
1202 tcp_lro_get_last_if_ackcmp(struct lro_ctrl *lc, struct lro_entry *le,
1203 struct tcpcb *tp, int32_t *new_m, bool can_append_old_cmp)
1207 /* Look at the last mbuf if any in queue */
1208 if (can_append_old_cmp) {
1209 m = STAILQ_LAST(&tp->t_inqueue, mbuf, m_stailqpkt);
1210 if (m != NULL && (m->m_flags & M_ACKCMP) != 0) {
1211 if (M_TRAILINGSPACE(m) >= sizeof(struct tcp_ackent)) {
1212 tcp_lro_log(tp, lc, le, NULL, 23, 0, 0, 0, 0);
1214 counter_u64_add(tcp_extra_mbuf, 1);
1217 /* Mark we ran out of space */
1218 tp->t_flags2 |= TF2_MBUF_L_ACKS;
1222 /* Decide mbuf size. */
1223 tcp_lro_log(tp, lc, le, NULL, 21, 0, 0, 0, 0);
1224 if (tp->t_flags2 & TF2_MBUF_L_ACKS)
1225 m = m_getcl(M_NOWAIT, MT_DATA, M_ACKCMP | M_PKTHDR);
1227 m = m_gethdr(M_NOWAIT, MT_DATA);
1229 if (__predict_false(m == NULL)) {
1230 counter_u64_add(tcp_would_have_but, 1);
1233 counter_u64_add(tcp_comp_total, 1);
1234 m->m_pkthdr.rcvif = lc->ifp;
1235 m->m_flags |= M_ACKCMP;
1240 static struct tcpcb *
1241 tcp_lro_lookup(struct ifnet *ifp, struct lro_parser *pa)
1245 switch (pa->data.lro_type) {
1247 case LRO_TYPE_IPV6_TCP:
1248 inp = in6_pcblookup(&V_tcbinfo,
1249 &pa->data.s_addr.v6,
1251 &pa->data.d_addr.v6,
1258 case LRO_TYPE_IPV4_TCP:
1259 inp = in_pcblookup(&V_tcbinfo,
1272 return (intotcpcb(inp));
1276 tcp_lro_ack_valid(struct mbuf *m, struct tcphdr *th, uint32_t **ppts, bool *other_opts)
1279 * This function returns two bits of valuable information.
1280 * a) Is what is present capable of being ack-compressed,
1281 * we can ack-compress if there is no options or just
1282 * a timestamp option, and of course the th_flags must
1283 * be correct as well.
1284 * b) Our other options present such as SACK. This is
1285 * used to determine if we want to wakeup or not.
1289 switch (th->th_off << 2) {
1290 case (sizeof(*th) + TCPOLEN_TSTAMP_APPA):
1291 *ppts = (uint32_t *)(th + 1);
1292 /* Check if we have only one timestamp option. */
1293 if (**ppts == TCP_LRO_TS_OPTION)
1294 *other_opts = false;
1303 *other_opts = false;
1311 /* For ACKCMP we only accept ACK, PUSH, ECE and CWR. */
1312 if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH | TH_ECE | TH_CWR)) != 0)
1314 /* If it has data on it we cannot compress it */
1315 if (m->m_pkthdr.lro_tcp_d_len)
1318 /* ACK flag must be set. */
1319 if (!(tcp_get_flags(th) & TH_ACK))
1325 tcp_lro_flush_tcphpts(struct lro_ctrl *lc, struct lro_entry *le)
1328 struct mbuf **pp, *cmp, *mv_to;
1329 struct ifnet *lagg_ifp;
1330 bool bpf_req, lagg_bpf_req, should_wake, can_append_old_cmp;
1332 /* Check if packet doesn't belongs to our network interface. */
1333 if ((tcplro_stacks_wanting_mbufq == 0) ||
1334 (le->outer.data.vlan_id != 0) ||
1335 (le->inner.data.lro_type != LRO_TYPE_NONE))
1336 return (TCP_LRO_CANNOT);
1340 * Be proactive about unspecified IPv6 address in source. As
1341 * we use all-zero to indicate unbounded/unconnected pcb,
1342 * unspecified IPv6 address can be used to confuse us.
1344 * Note that packets with unspecified IPv6 destination is
1345 * already dropped in ip6_input.
1347 if (__predict_false(le->outer.data.lro_type == LRO_TYPE_IPV6_TCP &&
1348 IN6_IS_ADDR_UNSPECIFIED(&le->outer.data.s_addr.v6)))
1349 return (TCP_LRO_CANNOT);
1351 if (__predict_false(le->inner.data.lro_type == LRO_TYPE_IPV6_TCP &&
1352 IN6_IS_ADDR_UNSPECIFIED(&le->inner.data.s_addr.v6)))
1353 return (TCP_LRO_CANNOT);
1355 /* Lookup inp, if any. Returns locked TCP inpcb. */
1356 tp = tcp_lro_lookup(lc->ifp,
1357 (le->inner.data.lro_type == LRO_TYPE_NONE) ? &le->outer : &le->inner);
1359 return (TCP_LRO_CANNOT);
1361 counter_u64_add(tcp_inp_lro_locks_taken, 1);
1363 /* Check if the inp is dead, Jim. */
1364 if (tp->t_state == TCPS_TIME_WAIT) {
1365 INP_WUNLOCK(tptoinpcb(tp));
1366 return (TCP_LRO_CANNOT);
1368 if (tp->t_lro_cpu == HPTS_CPU_NONE && lc->lro_cpu_is_set == 1)
1369 tp->t_lro_cpu = lc->lro_last_cpu;
1370 /* Check if the transport doesn't support the needed optimizations. */
1371 if ((tp->t_flags2 & (TF2_SUPPORTS_MBUFQ | TF2_MBUF_ACKCMP)) == 0) {
1372 INP_WUNLOCK(tptoinpcb(tp));
1373 return (TCP_LRO_CANNOT);
1376 if (tp->t_flags2 & TF2_MBUF_QUEUE_READY)
1377 should_wake = false;
1380 /* Check if packets should be tapped to BPF. */
1381 bpf_req = bpf_peers_present(lc->ifp->if_bpf);
1382 lagg_bpf_req = false;
1384 if (lc->ifp->if_type == IFT_IEEE8023ADLAG ||
1385 lc->ifp->if_type == IFT_INFINIBANDLAG) {
1386 struct lagg_port *lp = lc->ifp->if_lagg;
1387 struct lagg_softc *sc = lp->lp_softc;
1389 lagg_ifp = sc->sc_ifp;
1390 if (lagg_ifp != NULL)
1391 lagg_bpf_req = bpf_peers_present(lagg_ifp->if_bpf);
1394 /* Strip and compress all the incoming packets. */
1395 can_append_old_cmp = true;
1397 for (pp = &le->m_head; *pp != NULL; ) {
1399 if (do_bpf_strip_and_compress(tp, lc, le, pp,
1400 &cmp, &mv_to, &should_wake, bpf_req,
1401 lagg_bpf_req, lagg_ifp, can_append_old_cmp) == false) {
1402 /* Advance to next mbuf. */
1403 pp = &(*pp)->m_nextpkt;
1405 * Once we have appended we can't look in the pending
1406 * inbound packets for a compressed ack to append to.
1408 can_append_old_cmp = false;
1410 * Once we append we also need to stop adding to any
1411 * compressed ack we were remembering. A new cmp
1412 * ack will be required.
1415 tcp_lro_log(tp, lc, le, NULL, 25, 0, 0, 0, 0);
1416 } else if (mv_to != NULL) {
1417 /* We are asked to move pp up */
1418 pp = &mv_to->m_nextpkt;
1419 tcp_lro_log(tp, lc, le, NULL, 24, 0, 0, 0, 0);
1421 tcp_lro_log(tp, lc, le, NULL, 26, 0, 0, 0, 0);
1423 /* Update "m_last_mbuf", if any. */
1424 if (pp == &le->m_head)
1425 le->m_last_mbuf = *pp;
1427 le->m_last_mbuf = __containerof(pp, struct mbuf, m_nextpkt);
1429 /* Check if any data mbufs left. */
1430 if (le->m_head != NULL) {
1431 counter_u64_add(tcp_inp_lro_direct_queue, 1);
1432 tcp_lro_log(tp, lc, le, NULL, 22, 1, tp->t_flags2, 0, 1);
1433 tcp_queue_pkts(tp, le);
1437 counter_u64_add(tcp_inp_lro_wokeup_queue, 1);
1438 if ((*tp->t_fb->tfb_do_queued_segments)(tp, 0))
1439 /* TCP cb gone and unlocked. */
1442 INP_WUNLOCK(tptoinpcb(tp));
1444 return (0); /* Success. */
1449 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
1451 /* Only optimise if there are multiple packets waiting. */
1458 CURVNET_SET(lc->ifp->if_vnet);
1459 error = tcp_lro_flush_tcphpts(lc, le);
1463 tcp_lro_condense(lc, le);
1464 tcp_flush_out_entry(lc, le);
1469 bzero(le, sizeof(*le));
1470 LIST_INSERT_HEAD(&lc->lro_free, le, next);
1473 #define tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
1476 * The tcp_lro_sort() routine is comparable to qsort(), except it has
1477 * a worst case complexity limit of O(MIN(N,64)*N), where N is the
1478 * number of elements to sort and 64 is the number of sequence bits
1479 * available. The algorithm is bit-slicing the 64-bit sequence number,
1480 * sorting one bit at a time from the most significant bit until the
1481 * least significant one, skipping the constant bits. This is
1482 * typically called a radix sort.
1485 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
1487 struct lro_mbuf_sort temp;
1494 /* for small arrays insertion sort is faster */
1496 for (x = 1; x < size; x++) {
1498 for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
1499 parray[y] = parray[y - 1];
1505 /* compute sequence bits which are constant */
1508 for (x = 0; x != size; x++) {
1509 ones |= parray[x].seq;
1510 zeros |= ~parray[x].seq;
1513 /* compute bits which are not constant into "ones" */
1518 /* pick the most significant bit which is not constant */
1519 ones = tcp_lro_msb_64(ones);
1522 * Move entries having cleared sequence bits to the beginning
1525 for (x = y = 0; y != size; y++) {
1527 if (parray[y].seq & ones)
1531 parray[x] = parray[y];
1536 KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
1539 tcp_lro_sort(parray, x);
1548 tcp_lro_flush_all(struct lro_ctrl *lc)
1555 /* check if no mbufs to flush */
1556 if (lc->lro_mbuf_count == 0)
1558 if (lc->lro_cpu_is_set == 0) {
1559 if (lc->lro_last_cpu == curcpu) {
1560 lc->lro_cnt_of_same_cpu++;
1561 /* Have we reached the threshold to declare a cpu? */
1562 if (lc->lro_cnt_of_same_cpu > tcp_lro_cpu_set_thresh)
1563 lc->lro_cpu_is_set = 1;
1565 lc->lro_last_cpu = curcpu;
1566 lc->lro_cnt_of_same_cpu = 0;
1569 CURVNET_SET(lc->ifp->if_vnet);
1571 /* get current time */
1572 binuptime(&lc->lro_last_queue_time);
1574 /* sort all mbufs according to stream */
1575 tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
1577 /* input data into LRO engine, stream by stream */
1579 for (x = 0; x != lc->lro_mbuf_count; x++) {
1583 mb = lc->lro_mbuf_data[x].mb;
1585 /* get sequence number, masking away the packet index */
1586 nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
1588 /* check for new stream */
1592 /* flush active streams */
1593 tcp_lro_rx_done(lc);
1596 /* add packet to LRO engine */
1597 if (tcp_lro_rx_common(lc, mb, 0, false) != 0) {
1598 /* Flush anything we have acummulated */
1599 tcp_lro_flush_active(lc);
1600 /* input packet to network layer */
1601 (*lc->ifp->if_input)(lc->ifp, mb);
1608 /* flush active streams */
1609 tcp_lro_rx_done(lc);
1614 lc->lro_mbuf_count = 0;
1619 build_ack_entry(struct tcp_ackent *ae, struct tcphdr *th, struct mbuf *m,
1620 uint32_t *ts_ptr, uint16_t iptos)
1623 * Given a TCP ACK, summarize it down into the small TCP ACK
1626 ae->timestamp = m->m_pkthdr.rcv_tstmp;
1628 if (m->m_flags & M_TSTMP_LRO)
1629 ae->flags |= TSTMP_LRO;
1630 else if (m->m_flags & M_TSTMP)
1631 ae->flags |= TSTMP_HDWR;
1632 ae->seq = ntohl(th->th_seq);
1633 ae->ack = ntohl(th->th_ack);
1634 ae->flags |= tcp_get_flags(th);
1635 if (ts_ptr != NULL) {
1636 ae->ts_value = ntohl(ts_ptr[1]);
1637 ae->ts_echo = ntohl(ts_ptr[2]);
1638 ae->flags |= HAS_TSTMP;
1640 ae->win = ntohs(th->th_win);
1641 ae->codepoint = iptos;
1645 * Do BPF tap for either ACK_CMP packets or MBUF QUEUE type packets
1646 * and strip all, but the IPv4/IPv6 header.
1649 do_bpf_strip_and_compress(struct tcpcb *tp, struct lro_ctrl *lc,
1650 struct lro_entry *le, struct mbuf **pp, struct mbuf **cmp, struct mbuf **mv_to,
1651 bool *should_wake, bool bpf_req, bool lagg_bpf_req, struct ifnet *lagg_ifp, bool can_append_old_cmp)
1656 struct ip6_hdr *ip6;
1661 struct tcp_ackent *ack_ent;
1664 bool other_opts, can_compress;
1670 /* Get current mbuf. */
1673 /* Let the BPF see the packet */
1674 if (__predict_false(bpf_req))
1675 ETHER_BPF_MTAP(lc->ifp, m);
1677 if (__predict_false(lagg_bpf_req))
1678 ETHER_BPF_MTAP(lagg_ifp, m);
1680 tcp_hdr_offset = m->m_pkthdr.lro_tcp_h_off;
1681 lro_type = le->inner.data.lro_type;
1684 lro_type = le->outer.data.lro_type;
1686 case LRO_TYPE_IPV4_TCP:
1687 tcp_hdr_offset -= sizeof(*le->outer.ip4);
1688 m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1690 case LRO_TYPE_IPV6_TCP:
1691 tcp_hdr_offset -= sizeof(*le->outer.ip6);
1692 m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1698 case LRO_TYPE_IPV4_TCP:
1699 tcp_hdr_offset -= sizeof(*le->outer.ip4);
1700 m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1702 case LRO_TYPE_IPV6_TCP:
1703 tcp_hdr_offset -= sizeof(*le->outer.ip6);
1704 m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1710 MPASS(tcp_hdr_offset >= 0);
1712 m_adj(m, tcp_hdr_offset);
1713 m->m_flags |= M_LRO_EHDRSTRP;
1714 m->m_flags &= ~M_ACKCMP;
1715 m->m_pkthdr.lro_tcp_h_off -= tcp_hdr_offset;
1717 th = tcp_lro_get_th(m);
1719 th->th_sum = 0; /* TCP checksum is valid. */
1721 /* Check if ACK can be compressed */
1722 can_compress = tcp_lro_ack_valid(m, th, &ts_ptr, &other_opts);
1724 /* Now lets look at the should wake states */
1725 if ((other_opts == true) &&
1726 ((tp->t_flags2 & TF2_DONT_SACK_QUEUE) == 0)) {
1728 * If there are other options (SACK?) and the
1729 * tcp endpoint has not expressly told us it does
1730 * not care about SACKS, then we should wake up.
1732 *should_wake = true;
1733 } else if (*should_wake == false) {
1734 /* Wakeup override check if we are false here */
1735 *should_wake = tcp_lro_check_wake_status(tp);
1737 /* Is the ack compressable? */
1738 if (can_compress == false)
1740 /* Does the TCP endpoint support ACK compression? */
1741 if ((tp->t_flags2 & TF2_MBUF_ACKCMP) == 0)
1744 /* Lets get the TOS/traffic class field */
1745 l3.ptr = mtod(m, void *);
1747 case LRO_TYPE_IPV4_TCP:
1748 iptos = l3.ip4->ip_tos;
1750 case LRO_TYPE_IPV6_TCP:
1751 iptos = IPV6_TRAFFIC_CLASS(l3.ip6);
1754 iptos = 0; /* Keep compiler happy. */
1757 /* Now lets get space if we don't have some already */
1760 nm = tcp_lro_get_last_if_ackcmp(lc, le, tp, &n_mbuf,
1761 can_append_old_cmp);
1762 if (__predict_false(nm == NULL))
1767 * Link in the new cmp ack to our in-order place,
1768 * first set our cmp ack's next to where we are.
1773 * Set it up so mv_to is advanced to our
1774 * compressed ack. This way the caller can
1775 * advance pp to the right place.
1779 * Advance it here locally as well.
1781 pp = &nm->m_nextpkt;
1784 /* We have one already we are working on */
1786 if (M_TRAILINGSPACE(nm) < sizeof(struct tcp_ackent)) {
1787 /* We ran out of space */
1788 tp->t_flags2 |= TF2_MBUF_L_ACKS;
1792 MPASS(M_TRAILINGSPACE(nm) >= sizeof(struct tcp_ackent));
1793 counter_u64_add(tcp_inp_lro_compressed, 1);
1795 /* We can add in to the one on the tail */
1796 ack_ent = mtod(nm, struct tcp_ackent *);
1797 idx = (nm->m_len / sizeof(struct tcp_ackent));
1798 build_ack_entry(&ack_ent[idx], th, m, ts_ptr, iptos);
1800 /* Bump the size of both pkt-hdr and len */
1801 nm->m_len += sizeof(struct tcp_ackent);
1802 nm->m_pkthdr.len += sizeof(struct tcp_ackent);
1804 /* Advance to next mbuf before freeing. */
1806 m->m_nextpkt = NULL;
1810 counter_u64_add(tcp_uncomp_total, 1);
1816 static struct lro_head *
1817 tcp_lro_rx_get_bucket(struct lro_ctrl *lc, struct mbuf *m, struct lro_parser *parser)
1821 if (M_HASHTYPE_ISHASH(m)) {
1822 hash = m->m_pkthdr.flowid;
1824 for (unsigned i = hash = 0; i != LRO_RAW_ADDRESS_MAX; i++)
1825 hash += parser->data.raw[i];
1827 return (&lc->lro_hash[hash % lc->lro_hashsz]);
1831 tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash)
1833 struct lro_parser pi; /* inner address data */
1834 struct lro_parser po; /* outer address data */
1835 struct lro_parser *pa; /* current parser for TCP stream */
1836 struct lro_entry *le;
1837 struct lro_head *bucket;
1842 uint16_t tcp_data_sum;
1845 /* Quickly decide if packet cannot be LRO'ed */
1846 if (__predict_false(V_ipforwarding != 0))
1847 return (TCP_LRO_CANNOT);
1850 /* Quickly decide if packet cannot be LRO'ed */
1851 if (__predict_false(V_ip6_forwarding != 0))
1852 return (TCP_LRO_CANNOT);
1854 if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1855 ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1856 (m->m_pkthdr.csum_data != 0xffff)) {
1858 * The checksum either did not have hardware offload
1859 * or it was a bad checksum. We can't LRO such
1862 counter_u64_add(tcp_bad_csums, 1);
1863 return (TCP_LRO_CANNOT);
1865 /* We expect a contiguous header [eh, ip, tcp]. */
1866 pa = tcp_lro_parser(m, &po, &pi, true);
1867 if (__predict_false(pa == NULL))
1868 return (TCP_LRO_NOT_SUPPORTED);
1870 /* We don't expect any padding. */
1871 error = tcp_lro_trim_mbuf_chain(m, pa);
1872 if (__predict_false(error != 0))
1876 switch (pa->data.lro_type) {
1877 case LRO_TYPE_IPV4_TCP:
1878 error = tcp_lro_rx_ipv4(lc, m, pa->ip4);
1879 if (__predict_false(error != 0))
1886 /* If no hardware or arrival stamp on the packet add timestamp */
1887 if ((m->m_flags & (M_TSTMP_LRO | M_TSTMP)) == 0) {
1888 m->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
1889 m->m_flags |= M_TSTMP_LRO;
1892 /* Get pointer to TCP header. */
1895 /* Don't process SYN packets. */
1896 if (__predict_false(tcp_get_flags(th) & TH_SYN))
1897 return (TCP_LRO_CANNOT);
1899 /* Get total TCP header length and compute payload length. */
1900 tcp_opt_len = (th->th_off << 2);
1901 tcp_data_len = m->m_pkthdr.len - ((uint8_t *)th -
1902 (uint8_t *)m->m_data) - tcp_opt_len;
1903 tcp_opt_len -= sizeof(*th);
1905 /* Don't process invalid TCP headers. */
1906 if (__predict_false(tcp_opt_len < 0 || tcp_data_len < 0))
1907 return (TCP_LRO_CANNOT);
1909 /* Compute TCP data only checksum. */
1910 if (tcp_data_len == 0)
1911 tcp_data_sum = 0; /* no data, no checksum */
1912 else if (__predict_false(csum != 0))
1913 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~csum);
1915 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~th->th_sum);
1917 /* Save TCP info in mbuf. */
1918 m->m_nextpkt = NULL;
1919 m->m_pkthdr.rcvif = lc->ifp;
1920 m->m_pkthdr.lro_tcp_d_csum = tcp_data_sum;
1921 m->m_pkthdr.lro_tcp_d_len = tcp_data_len;
1922 m->m_pkthdr.lro_tcp_h_off = ((uint8_t *)th - (uint8_t *)m->m_data);
1923 m->m_pkthdr.lro_nsegs = 1;
1925 /* Get hash bucket. */
1927 bucket = &lc->lro_hash[0];
1929 bucket = tcp_lro_rx_get_bucket(lc, m, pa);
1932 /* Try to find a matching previous segment. */
1933 LIST_FOREACH(le, bucket, hash_next) {
1934 /* Compare addresses and ports. */
1935 if (lro_address_compare(&po.data, &le->outer.data) == false ||
1936 lro_address_compare(&pi.data, &le->inner.data) == false)
1939 /* Check if no data and old ACK. */
1940 if (tcp_data_len == 0 &&
1941 SEQ_LT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1946 /* Mark "m" in the last spot. */
1947 le->m_last_mbuf->m_nextpkt = m;
1948 /* Now set the tail to "m". */
1949 le->m_last_mbuf = m;
1953 /* Try to find an empty slot. */
1954 if (LIST_EMPTY(&lc->lro_free))
1955 return (TCP_LRO_NO_ENTRIES);
1957 /* Start a new segment chain. */
1958 le = LIST_FIRST(&lc->lro_free);
1959 LIST_REMOVE(le, next);
1960 tcp_lro_active_insert(lc, bucket, le);
1962 /* Make sure the headers are set. */
1966 /* Store time this entry was allocated. */
1967 le->alloc_time = lc->lro_last_queue_time;
1969 tcp_set_entry_to_mbuf(lc, le, m, th);
1971 /* Now set the tail to "m". */
1972 le->m_last_mbuf = m;
1978 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
1982 if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1983 ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1984 (m->m_pkthdr.csum_data != 0xffff)) {
1986 * The checksum either did not have hardware offload
1987 * or it was a bad checksum. We can't LRO such
1990 counter_u64_add(tcp_bad_csums, 1);
1991 return (TCP_LRO_CANNOT);
1993 /* get current time */
1994 binuptime(&lc->lro_last_queue_time);
1995 CURVNET_SET(lc->ifp->if_vnet);
1996 error = tcp_lro_rx_common(lc, m, csum, true);
1997 if (__predict_false(error != 0)) {
1999 * Flush anything we have acummulated
2000 * ahead of this packet that can't
2001 * be LRO'd. This preserves order.
2003 tcp_lro_flush_active(lc);
2011 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
2015 if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
2016 lc->lro_mbuf_max == 0)) {
2022 /* check if packet is not LRO capable */
2023 if (__predict_false((lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
2024 /* input packet to network layer */
2025 (*lc->ifp->if_input) (lc->ifp, mb);
2029 /* If no hardware or arrival stamp on the packet add timestamp */
2030 if ((tcplro_stacks_wanting_mbufq > 0) &&
2031 (tcp_less_accurate_lro_ts == 0) &&
2032 ((mb->m_flags & M_TSTMP) == 0)) {
2033 /* Add in an LRO time since no hardware */
2034 binuptime(&lc->lro_last_queue_time);
2035 mb->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
2036 mb->m_flags |= M_TSTMP_LRO;
2039 /* create sequence number */
2040 lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
2041 (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
2042 (((uint64_t)mb->m_pkthdr.flowid) << 24) |
2043 ((uint64_t)lc->lro_mbuf_count);
2046 lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb;
2048 /* flush if array is full */
2049 if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max))
2050 tcp_lro_flush_all(lc);