2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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>
36 __FBSDID("$FreeBSD$");
39 #include "opt_inet6.h"
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/malloc.h>
46 #include <sys/socket.h>
47 #include <sys/socketvar.h>
48 #include <sys/sockbuf.h>
49 #include <sys/sysctl.h>
52 #include <net/if_var.h>
53 #include <net/ethernet.h>
56 #include <net/if_dl.h>
57 #include <net/if_media.h>
58 #include <net/if_types.h>
59 #include <net/infiniband.h>
60 #include <net/if_lagg.h>
62 #include <netinet/in_systm.h>
63 #include <netinet/in.h>
64 #include <netinet/ip6.h>
65 #include <netinet/ip.h>
66 #include <netinet/ip_var.h>
67 #include <netinet/in_pcb.h>
68 #include <netinet6/in6_pcb.h>
69 #include <netinet/tcp.h>
70 #include <netinet/tcp_seq.h>
71 #include <netinet/tcp_lro.h>
72 #include <netinet/tcp_var.h>
73 #include <netinet/tcpip.h>
74 #include <netinet/tcp_hpts.h>
75 #include <netinet/tcp_log_buf.h>
76 #include <netinet/tcp_fsm.h>
77 #include <netinet/udp.h>
78 #include <netinet6/ip6_var.h>
80 #include <machine/in_cksum.h>
82 static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures");
84 #define TCP_LRO_TS_OPTION \
85 ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
86 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
88 static void tcp_lro_rx_done(struct lro_ctrl *lc);
89 static int tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m,
90 uint32_t csum, bool use_hash);
93 static bool do_bpf_strip_and_compress(struct inpcb *, struct lro_ctrl *,
94 struct lro_entry *, struct mbuf **, struct mbuf **, struct mbuf **,
95 bool *, bool, bool, struct ifnet *, bool);
99 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
102 static long tcplro_stacks_wanting_mbufq;
103 counter_u64_t tcp_inp_lro_direct_queue;
104 counter_u64_t tcp_inp_lro_wokeup_queue;
105 counter_u64_t tcp_inp_lro_compressed;
106 counter_u64_t tcp_inp_lro_locks_taken;
107 counter_u64_t tcp_extra_mbuf;
108 counter_u64_t tcp_would_have_but;
109 counter_u64_t tcp_comp_total;
110 counter_u64_t tcp_uncomp_total;
111 counter_u64_t tcp_bad_csums;
113 static unsigned tcp_lro_entries = TCP_LRO_ENTRIES;
114 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries,
115 CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_entries, 0,
116 "default number of LRO entries");
118 static uint32_t tcp_lro_cpu_set_thresh = TCP_LRO_CPU_DECLARATION_THRESH;
119 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_cpu_threshold,
120 CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_cpu_set_thresh, 0,
121 "Number of interrupts in a row on the same CPU that will make us declare an 'affinity' cpu?");
123 static uint32_t tcp_less_accurate_lro_ts = 0;
124 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_less_accurate,
125 CTLFLAG_MPSAFE, &tcp_less_accurate_lro_ts, 0,
126 "Do we trade off efficency by doing less timestamp operations for time accuracy?");
128 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, fullqueue, CTLFLAG_RD,
129 &tcp_inp_lro_direct_queue, "Number of lro's fully queued to transport");
130 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, wokeup, CTLFLAG_RD,
131 &tcp_inp_lro_wokeup_queue, "Number of lro's where we woke up transport via hpts");
132 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, compressed, CTLFLAG_RD,
133 &tcp_inp_lro_compressed, "Number of lro's compressed and sent to transport");
134 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lockcnt, CTLFLAG_RD,
135 &tcp_inp_lro_locks_taken, "Number of lro's inp_wlocks taken");
136 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, extra_mbuf, CTLFLAG_RD,
137 &tcp_extra_mbuf, "Number of times we had an extra compressed ack dropped into the tp");
138 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, would_have_but, CTLFLAG_RD,
139 &tcp_would_have_but, "Number of times we would have had an extra compressed, but mget failed");
140 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, with_m_ackcmp, CTLFLAG_RD,
141 &tcp_comp_total, "Number of mbufs queued with M_ACKCMP flags set");
142 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, without_m_ackcmp, CTLFLAG_RD,
143 &tcp_uncomp_total, "Number of mbufs queued without M_ACKCMP");
144 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lro_badcsum, CTLFLAG_RD,
145 &tcp_bad_csums, "Number of packets that the common code saw with bad csums");
148 tcp_lro_reg_mbufq(void)
150 atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, 1);
154 tcp_lro_dereg_mbufq(void)
156 atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, -1);
160 tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket,
161 struct lro_entry *le)
164 LIST_INSERT_HEAD(&lc->lro_active, le, next);
165 LIST_INSERT_HEAD(bucket, le, hash_next);
169 tcp_lro_active_remove(struct lro_entry *le)
172 LIST_REMOVE(le, next); /* active list */
173 LIST_REMOVE(le, hash_next); /* hash bucket */
177 tcp_lro_init(struct lro_ctrl *lc)
179 return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0));
183 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
184 unsigned lro_entries, unsigned lro_mbufs)
186 struct lro_entry *le;
188 unsigned i, elements;
190 lc->lro_bad_csum = 0;
193 lc->lro_mbuf_count = 0;
194 lc->lro_mbuf_max = lro_mbufs;
195 lc->lro_cnt = lro_entries;
196 lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
197 lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
199 LIST_INIT(&lc->lro_free);
200 LIST_INIT(&lc->lro_active);
202 /* create hash table to accelerate entry lookup */
203 if (lro_entries > lro_mbufs)
204 elements = lro_entries;
206 elements = lro_mbufs;
207 lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz,
209 if (lc->lro_hash == NULL) {
210 memset(lc, 0, sizeof(*lc));
214 /* compute size to allocate */
215 size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
216 (lro_entries * sizeof(*le));
217 lc->lro_mbuf_data = (struct lro_mbuf_sort *)
218 malloc(size, M_LRO, M_NOWAIT | M_ZERO);
220 /* check for out of memory */
221 if (lc->lro_mbuf_data == NULL) {
222 free(lc->lro_hash, M_LRO);
223 memset(lc, 0, sizeof(*lc));
226 /* compute offset for LRO entries */
227 le = (struct lro_entry *)
228 (lc->lro_mbuf_data + lro_mbufs);
230 /* setup linked list */
231 for (i = 0; i != lro_entries; i++)
232 LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
237 struct vxlan_header {
243 tcp_lro_low_level_parser(void *ptr, struct lro_parser *parser, bool update_data, bool is_vxlan, int mlen)
245 const struct ether_vlan_header *eh;
250 memset(parser, 0, sizeof(*parser));
255 const struct vxlan_header *vxh;
257 ptr = (uint8_t *)ptr + sizeof(*vxh);
259 parser->data.vxlan_vni =
260 vxh->vxlh_vni & htonl(0xffffff00);
265 if (__predict_false(eh->evl_encap_proto == htons(ETHERTYPE_VLAN))) {
266 eth_type = eh->evl_proto;
268 /* strip priority and keep VLAN ID only */
269 parser->data.vlan_id = eh->evl_tag & htons(EVL_VLID_MASK);
271 /* advance to next header */
272 ptr = (uint8_t *)ptr + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
273 mlen -= (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
275 eth_type = eh->evl_encap_proto;
276 /* advance to next header */
277 mlen -= ETHER_HDR_LEN;
278 ptr = (uint8_t *)ptr + ETHER_HDR_LEN;
280 if (__predict_false(mlen <= 0))
284 case htons(ETHERTYPE_IP):
286 if (__predict_false(mlen < sizeof(struct ip)))
288 /* Ensure there are no IPv4 options. */
289 if ((parser->ip4->ip_hl << 2) != sizeof (*parser->ip4))
291 /* .. and the packet is not fragmented. */
292 if (parser->ip4->ip_off & htons(IP_MF|IP_OFFMASK))
294 ptr = (uint8_t *)ptr + (parser->ip4->ip_hl << 2);
295 mlen -= sizeof(struct ip);
297 parser->data.s_addr.v4 = parser->ip4->ip_src;
298 parser->data.d_addr.v4 = parser->ip4->ip_dst;
300 switch (parser->ip4->ip_p) {
302 if (__predict_false(mlen < sizeof(struct udphdr)))
306 parser->data.lro_type = LRO_TYPE_IPV4_UDP;
307 parser->data.s_port = parser->udp->uh_sport;
308 parser->data.d_port = parser->udp->uh_dport;
310 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_UDP);
312 ptr = ((uint8_t *)ptr + sizeof(*parser->udp));
313 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
317 if (__predict_false(mlen < sizeof(struct tcphdr)))
320 parser->data.lro_type = LRO_TYPE_IPV4_TCP;
321 parser->data.s_port = parser->tcp->th_sport;
322 parser->data.d_port = parser->tcp->th_dport;
324 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_TCP);
326 if (__predict_false(mlen < (parser->tcp->th_off << 2)))
328 ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
329 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
337 case htons(ETHERTYPE_IPV6):
339 if (__predict_false(mlen < sizeof(struct ip6_hdr)))
341 ptr = (uint8_t *)ptr + sizeof(*parser->ip6);
343 parser->data.s_addr.v6 = parser->ip6->ip6_src;
344 parser->data.d_addr.v6 = parser->ip6->ip6_dst;
346 mlen -= sizeof(struct ip6_hdr);
347 switch (parser->ip6->ip6_nxt) {
349 if (__predict_false(mlen < sizeof(struct udphdr)))
353 parser->data.lro_type = LRO_TYPE_IPV6_UDP;
354 parser->data.s_port = parser->udp->uh_sport;
355 parser->data.d_port = parser->udp->uh_dport;
357 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_UDP);
359 ptr = (uint8_t *)ptr + sizeof(*parser->udp);
360 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
363 if (__predict_false(mlen < sizeof(struct tcphdr)))
367 parser->data.lro_type = LRO_TYPE_IPV6_TCP;
368 parser->data.s_port = parser->tcp->th_sport;
369 parser->data.d_port = parser->tcp->th_dport;
371 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_TCP);
373 if (__predict_false(mlen < (parser->tcp->th_off << 2)))
375 ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
376 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
386 /* Invalid packet - cannot parse */
390 static const int vxlan_csum = CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
391 CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID;
393 static inline struct lro_parser *
394 tcp_lro_parser(struct mbuf *m, struct lro_parser *po, struct lro_parser *pi, bool update_data)
398 /* Try to parse outer headers first. */
399 data_ptr = tcp_lro_low_level_parser(m->m_data, po, update_data, false, m->m_len);
400 if (data_ptr == NULL || po->total_hdr_len > m->m_len)
404 /* Store VLAN ID, if any. */
405 if (__predict_false(m->m_flags & M_VLANTAG)) {
407 htons(m->m_pkthdr.ether_vtag) & htons(EVL_VLID_MASK);
409 /* Store decrypted flag, if any. */
410 if (__predict_false((m->m_pkthdr.csum_flags &
411 CSUM_TLS_MASK) == CSUM_TLS_DECRYPTED))
412 po->data.lro_flags |= LRO_FLAG_DECRYPTED;
415 switch (po->data.lro_type) {
416 case LRO_TYPE_IPV4_UDP:
417 case LRO_TYPE_IPV6_UDP:
418 /* Check for VXLAN headers. */
419 if ((m->m_pkthdr.csum_flags & vxlan_csum) != vxlan_csum)
422 /* Try to parse inner headers. */
423 data_ptr = tcp_lro_low_level_parser(data_ptr, pi, update_data, true,
424 (m->m_len - ((caddr_t)data_ptr - m->m_data)));
425 if (data_ptr == NULL || (pi->total_hdr_len + po->total_hdr_len) > m->m_len)
428 /* Verify supported header types. */
429 switch (pi->data.lro_type) {
430 case LRO_TYPE_IPV4_TCP:
431 case LRO_TYPE_IPV6_TCP:
437 case LRO_TYPE_IPV4_TCP:
438 case LRO_TYPE_IPV6_TCP:
440 memset(pi, 0, sizeof(*pi));
449 tcp_lro_trim_mbuf_chain(struct mbuf *m, const struct lro_parser *po)
453 switch (po->data.lro_type) {
455 case LRO_TYPE_IPV4_TCP:
456 len = ((uint8_t *)po->ip4 - (uint8_t *)m->m_data) +
457 ntohs(po->ip4->ip_len);
461 case LRO_TYPE_IPV6_TCP:
462 len = ((uint8_t *)po->ip6 - (uint8_t *)m->m_data) +
463 ntohs(po->ip6->ip6_plen) + sizeof(*po->ip6);
467 return (TCP_LRO_CANNOT);
471 * If the frame is padded beyond the end of the IP packet,
472 * then trim the extra bytes off:
474 if (__predict_true(m->m_pkthdr.len == len)) {
476 } else if (m->m_pkthdr.len > len) {
477 m_adj(m, len - m->m_pkthdr.len);
480 return (TCP_LRO_CANNOT);
483 static struct tcphdr *
484 tcp_lro_get_th(struct mbuf *m)
486 return ((struct tcphdr *)((uint8_t *)m->m_data + m->m_pkthdr.lro_tcp_h_off));
490 lro_free_mbuf_chain(struct mbuf *m)
503 tcp_lro_free(struct lro_ctrl *lc)
505 struct lro_entry *le;
508 /* reset LRO free list */
509 LIST_INIT(&lc->lro_free);
511 /* free active mbufs, if any */
512 while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
513 tcp_lro_active_remove(le);
514 lro_free_mbuf_chain(le->m_head);
517 /* free hash table */
518 free(lc->lro_hash, M_LRO);
522 /* free mbuf array, if any */
523 for (x = 0; x != lc->lro_mbuf_count; x++)
524 m_freem(lc->lro_mbuf_data[x].mb);
525 lc->lro_mbuf_count = 0;
527 /* free allocated memory, if any */
528 free(lc->lro_mbuf_data, M_LRO);
529 lc->lro_mbuf_data = NULL;
533 tcp_lro_rx_csum_tcphdr(const struct tcphdr *th)
539 csum = -th->th_sum; /* exclude checksum field */
541 ptr = (const uint16_t *)th;
548 while (csum > 0xffff)
549 csum = (csum >> 16) + (csum & 0xffff);
555 tcp_lro_rx_csum_data(const struct lro_parser *pa, uint16_t tcp_csum)
562 switch (pa->data.lro_type) {
564 case LRO_TYPE_IPV6_TCP:
565 /* Compute full pseudo IPv6 header checksum. */
566 cs = in6_cksum_pseudo(pa->ip6, ntohs(pa->ip6->ip6_plen), pa->ip6->ip6_nxt, 0);
570 case LRO_TYPE_IPV4_TCP:
571 /* Compute full pseudo IPv4 header checsum. */
572 cs = in_addword(ntohs(pa->ip4->ip_len) - sizeof(*pa->ip4), IPPROTO_TCP);
573 cs = in_pseudo(pa->ip4->ip_src.s_addr, pa->ip4->ip_dst.s_addr, htons(cs));
577 cs = 0; /* Keep compiler happy. */
581 /* Complement checksum. */
585 /* Remove TCP header checksum. */
586 cs = ~tcp_lro_rx_csum_tcphdr(pa->tcp);
589 /* Compute checksum remainder. */
591 c = (c >> 16) + (c & 0xffff);
597 tcp_lro_rx_done(struct lro_ctrl *lc)
599 struct lro_entry *le;
601 while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
602 tcp_lro_active_remove(le);
603 tcp_lro_flush(lc, le);
608 tcp_lro_flush_active(struct lro_ctrl *lc)
610 struct lro_entry *le;
613 * Walk through the list of le entries, and
614 * any one that does have packets flush. This
615 * is called because we have an inbound packet
616 * (e.g. SYN) that has to have all others flushed
617 * in front of it. Note we have to do the remove
618 * because tcp_lro_flush() assumes that the entry
619 * is being freed. This is ok it will just get
620 * reallocated again like it was new.
622 LIST_FOREACH(le, &lc->lro_active, next) {
623 if (le->m_head != NULL) {
624 tcp_lro_active_remove(le);
625 tcp_lro_flush(lc, le);
631 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
633 struct lro_entry *le, *le_tmp;
638 if (LIST_EMPTY(&lc->lro_active))
641 /* get timeout time and current time in ns */
643 now = bintime2ns(&bt);
644 tov = ((timeout->tv_sec * 1000000000) + (timeout->tv_usec * 1000));
645 LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
646 if (now >= (bintime2ns(&le->alloc_time) + tov)) {
647 tcp_lro_active_remove(le);
648 tcp_lro_flush(lc, le);
655 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4)
659 /* Legacy IP has a header checksum that needs to be correct. */
660 if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) {
661 if (__predict_false((m->m_pkthdr.csum_flags & CSUM_IP_VALID) == 0)) {
663 return (TCP_LRO_CANNOT);
666 csum = in_cksum_hdr(ip4);
667 if (__predict_false(csum != 0)) {
669 return (TCP_LRO_CANNOT);
678 tcp_lro_log(struct tcpcb *tp, const struct lro_ctrl *lc,
679 const struct lro_entry *le, const struct mbuf *m,
680 int frm, int32_t tcp_data_len, uint32_t th_seq,
681 uint32_t th_ack, uint16_t th_win)
683 if (tp->t_logstate != TCP_LOG_STATE_OFF) {
684 union tcp_log_stackspecific log;
685 struct timeval tv, btv;
688 cts = tcp_get_usecs(&tv);
689 memset(&log, 0, sizeof(union tcp_log_stackspecific));
690 log.u_bbr.flex8 = frm;
691 log.u_bbr.flex1 = tcp_data_len;
693 log.u_bbr.flex2 = m->m_pkthdr.len;
697 log.u_bbr.flex3 = le->m_head->m_pkthdr.lro_nsegs;
698 log.u_bbr.flex4 = le->m_head->m_pkthdr.lro_tcp_d_len;
699 log.u_bbr.flex5 = le->m_head->m_pkthdr.len;
700 log.u_bbr.delRate = le->m_head->m_flags;
701 log.u_bbr.rttProp = le->m_head->m_pkthdr.rcv_tstmp;
703 log.u_bbr.inflight = th_seq;
704 log.u_bbr.delivered = th_ack;
705 log.u_bbr.timeStamp = cts;
706 log.u_bbr.epoch = le->next_seq;
707 log.u_bbr.lt_epoch = le->ack_seq;
708 log.u_bbr.pacing_gain = th_win;
709 log.u_bbr.cwnd_gain = le->window;
710 log.u_bbr.lost = curcpu;
711 log.u_bbr.cur_del_rate = (uintptr_t)m;
712 log.u_bbr.bw_inuse = (uintptr_t)le->m_head;
713 bintime2timeval(&lc->lro_last_queue_time, &btv);
714 log.u_bbr.flex6 = tcp_tv_to_usectick(&btv);
715 log.u_bbr.flex7 = le->compressed;
716 log.u_bbr.pacing_gain = le->uncompressed;
717 if (in_epoch(net_epoch_preempt))
718 log.u_bbr.inhpts = 1;
720 log.u_bbr.inhpts = 0;
721 TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
722 &tptosocket(tp)->so_snd,
723 TCP_LOG_LRO, 0, 0, &log, false, &tv);
729 tcp_lro_assign_and_checksum_16(uint16_t *ptr, uint16_t value, uint16_t *psum)
733 csum = 0xffff - *ptr + value;
734 while (csum > 0xffff)
735 csum = (csum >> 16) + (csum & 0xffff);
741 tcp_lro_update_checksum(const struct lro_parser *pa, const struct lro_entry *le,
742 uint16_t payload_len, uint16_t delta_sum)
746 uint16_t temp[5] = {};
748 switch (pa->data.lro_type) {
749 case LRO_TYPE_IPV4_TCP:
750 /* Compute new IPv4 length. */
751 tlen = (pa->ip4->ip_hl << 2) + (pa->tcp->th_off << 2) + payload_len;
752 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
754 /* Subtract delta from current IPv4 checksum. */
755 csum = pa->ip4->ip_sum + 0xffff - temp[0];
756 while (csum > 0xffff)
757 csum = (csum >> 16) + (csum & 0xffff);
758 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
759 goto update_tcp_header;
761 case LRO_TYPE_IPV6_TCP:
762 /* Compute new IPv6 length. */
763 tlen = (pa->tcp->th_off << 2) + payload_len;
764 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
765 goto update_tcp_header;
767 case LRO_TYPE_IPV4_UDP:
768 /* Compute new IPv4 length. */
769 tlen = (pa->ip4->ip_hl << 2) + sizeof(*pa->udp) + payload_len;
770 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
772 /* Subtract delta from current IPv4 checksum. */
773 csum = pa->ip4->ip_sum + 0xffff - temp[0];
774 while (csum > 0xffff)
775 csum = (csum >> 16) + (csum & 0xffff);
776 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
777 goto update_udp_header;
779 case LRO_TYPE_IPV6_UDP:
780 /* Compute new IPv6 length. */
781 tlen = sizeof(*pa->udp) + payload_len;
782 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
783 goto update_udp_header;
790 /* Compute current TCP header checksum. */
791 temp[2] = tcp_lro_rx_csum_tcphdr(pa->tcp);
793 /* Incorporate the latest ACK into the TCP header. */
794 pa->tcp->th_ack = le->ack_seq;
795 pa->tcp->th_win = le->window;
797 /* Incorporate latest timestamp into the TCP header. */
798 if (le->timestamp != 0) {
801 ts_ptr = (uint32_t *)(pa->tcp + 1);
802 ts_ptr[1] = htonl(le->tsval);
803 ts_ptr[2] = le->tsecr;
806 /* Compute new TCP header checksum. */
807 temp[3] = tcp_lro_rx_csum_tcphdr(pa->tcp);
809 /* Compute new TCP checksum. */
810 csum = pa->tcp->th_sum + 0xffff - delta_sum +
811 0xffff - temp[0] + 0xffff - temp[3] + temp[2];
812 while (csum > 0xffff)
813 csum = (csum >> 16) + (csum & 0xffff);
815 /* Assign new TCP checksum. */
816 tcp_lro_assign_and_checksum_16(&pa->tcp->th_sum, csum, &temp[4]);
818 /* Compute all modififications affecting next checksum. */
819 csum = temp[0] + temp[1] + 0xffff - temp[2] +
820 temp[3] + temp[4] + delta_sum;
821 while (csum > 0xffff)
822 csum = (csum >> 16) + (csum & 0xffff);
824 /* Return delta checksum to next stage, if any. */
828 tlen = sizeof(*pa->udp) + payload_len;
829 /* Assign new UDP length and compute checksum delta. */
830 tcp_lro_assign_and_checksum_16(&pa->udp->uh_ulen, htons(tlen), &temp[2]);
832 /* Check if there is a UDP checksum. */
833 if (__predict_false(pa->udp->uh_sum != 0)) {
834 /* Compute new UDP checksum. */
835 csum = pa->udp->uh_sum + 0xffff - delta_sum +
836 0xffff - temp[0] + 0xffff - temp[2];
837 while (csum > 0xffff)
838 csum = (csum >> 16) + (csum & 0xffff);
839 /* Assign new UDP checksum. */
840 tcp_lro_assign_and_checksum_16(&pa->udp->uh_sum, csum, &temp[3]);
843 /* Compute all modififications affecting next checksum. */
844 csum = temp[0] + temp[1] + temp[2] + temp[3] + delta_sum;
845 while (csum > 0xffff)
846 csum = (csum >> 16) + (csum & 0xffff);
848 /* Return delta checksum to next stage, if any. */
853 tcp_flush_out_entry(struct lro_ctrl *lc, struct lro_entry *le)
855 /* Check if we need to recompute any checksums. */
856 if (le->needs_merge) {
859 switch (le->inner.data.lro_type) {
860 case LRO_TYPE_IPV4_TCP:
861 csum = tcp_lro_update_checksum(&le->inner, le,
862 le->m_head->m_pkthdr.lro_tcp_d_len,
863 le->m_head->m_pkthdr.lro_tcp_d_csum);
864 csum = tcp_lro_update_checksum(&le->outer, NULL,
865 le->m_head->m_pkthdr.lro_tcp_d_len +
866 le->inner.total_hdr_len, csum);
867 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
868 CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
869 le->m_head->m_pkthdr.csum_data = 0xffff;
870 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
871 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
873 case LRO_TYPE_IPV6_TCP:
874 csum = tcp_lro_update_checksum(&le->inner, le,
875 le->m_head->m_pkthdr.lro_tcp_d_len,
876 le->m_head->m_pkthdr.lro_tcp_d_csum);
877 csum = tcp_lro_update_checksum(&le->outer, NULL,
878 le->m_head->m_pkthdr.lro_tcp_d_len +
879 le->inner.total_hdr_len, csum);
880 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
882 le->m_head->m_pkthdr.csum_data = 0xffff;
883 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
884 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
887 switch (le->outer.data.lro_type) {
888 case LRO_TYPE_IPV4_TCP:
889 csum = tcp_lro_update_checksum(&le->outer, le,
890 le->m_head->m_pkthdr.lro_tcp_d_len,
891 le->m_head->m_pkthdr.lro_tcp_d_csum);
892 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
893 CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
894 le->m_head->m_pkthdr.csum_data = 0xffff;
895 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
896 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
898 case LRO_TYPE_IPV6_TCP:
899 csum = tcp_lro_update_checksum(&le->outer, le,
900 le->m_head->m_pkthdr.lro_tcp_d_len,
901 le->m_head->m_pkthdr.lro_tcp_d_csum);
902 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
904 le->m_head->m_pkthdr.csum_data = 0xffff;
905 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
906 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
918 * Break any chain, this is not set to NULL on the singleton
919 * case m_nextpkt points to m_head. Other case set them
920 * m_nextpkt to NULL in push_and_replace.
922 le->m_head->m_nextpkt = NULL;
923 lc->lro_queued += le->m_head->m_pkthdr.lro_nsegs;
924 (*lc->ifp->if_input)(lc->ifp, le->m_head);
928 tcp_set_entry_to_mbuf(struct lro_ctrl *lc, struct lro_entry *le,
929 struct mbuf *m, struct tcphdr *th)
932 uint16_t tcp_data_len;
933 uint16_t tcp_opt_len;
935 ts_ptr = (uint32_t *)(th + 1);
936 tcp_opt_len = (th->th_off << 2);
937 tcp_opt_len -= sizeof(*th);
939 /* Check if there is a timestamp option. */
940 if (tcp_opt_len == 0 ||
941 __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
942 *ts_ptr != TCP_LRO_TS_OPTION)) {
943 /* We failed to find the timestamp option. */
947 le->tsval = ntohl(*(ts_ptr + 1));
948 le->tsecr = *(ts_ptr + 2);
951 tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
953 /* Pull out TCP sequence numbers and window size. */
954 le->next_seq = ntohl(th->th_seq) + tcp_data_len;
955 le->ack_seq = th->th_ack;
956 le->window = th->th_win;
957 le->flags = tcp_get_flags(th);
960 /* Setup new data pointers. */
962 le->m_tail = m_last(m);
966 tcp_push_and_replace(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m)
968 struct lro_parser *pa;
971 * Push up the stack of the current entry
972 * and replace it with "m".
976 /* Grab off the next and save it */
977 msave = le->m_head->m_nextpkt;
978 le->m_head->m_nextpkt = NULL;
980 /* Now push out the old entry */
981 tcp_flush_out_entry(lc, le);
983 /* Re-parse new header, should not fail. */
984 pa = tcp_lro_parser(m, &le->outer, &le->inner, false);
986 ("tcp_push_and_replace: LRO parser failed on m=%p\n", m));
989 * Now to replace the data properly in the entry
990 * we have to reset the TCP header and
993 tcp_set_entry_to_mbuf(lc, le, m, pa->tcp);
995 /* Restore the next list */
996 m->m_nextpkt = msave;
1000 tcp_lro_mbuf_append_pkthdr(struct lro_entry *le, const struct mbuf *p)
1006 if (m->m_pkthdr.lro_nsegs == 1) {
1007 /* Compute relative checksum. */
1008 csum = p->m_pkthdr.lro_tcp_d_csum;
1010 /* Merge TCP data checksums. */
1011 csum = (uint32_t)m->m_pkthdr.lro_tcp_d_csum +
1012 (uint32_t)p->m_pkthdr.lro_tcp_d_csum;
1013 while (csum > 0xffff)
1014 csum = (csum >> 16) + (csum & 0xffff);
1017 /* Update various counters. */
1018 m->m_pkthdr.len += p->m_pkthdr.lro_tcp_d_len;
1019 m->m_pkthdr.lro_tcp_d_csum = csum;
1020 m->m_pkthdr.lro_tcp_d_len += p->m_pkthdr.lro_tcp_d_len;
1021 m->m_pkthdr.lro_nsegs += p->m_pkthdr.lro_nsegs;
1022 le->needs_merge = 1;
1026 tcp_lro_condense(struct lro_ctrl *lc, struct lro_entry *le)
1029 * Walk through the mbuf chain we
1030 * have on tap and compress/condense
1036 uint32_t tcp_data_len_total;
1037 uint32_t tcp_data_seg_total;
1038 uint16_t tcp_data_len;
1039 uint16_t tcp_opt_len;
1042 * First we must check the lead (m_head)
1043 * we must make sure that it is *not*
1044 * something that should be sent up
1045 * right away (sack etc).
1048 m = le->m_head->m_nextpkt;
1050 /* Just one left. */
1054 th = tcp_lro_get_th(m);
1055 tcp_opt_len = (th->th_off << 2);
1056 tcp_opt_len -= sizeof(*th);
1057 ts_ptr = (uint32_t *)(th + 1);
1059 if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1060 *ts_ptr != TCP_LRO_TS_OPTION)) {
1062 * Its not the timestamp. We can't
1063 * use this guy as the head.
1065 le->m_head->m_nextpkt = m->m_nextpkt;
1066 tcp_push_and_replace(lc, le, m);
1069 if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1071 * Make sure that previously seen segments/ACKs are delivered
1072 * before this segment, e.g. FIN.
1074 le->m_head->m_nextpkt = m->m_nextpkt;
1075 tcp_push_and_replace(lc, le, m);
1078 while((m = le->m_head->m_nextpkt) != NULL) {
1080 * condense m into le, first
1081 * pull m out of the list.
1083 le->m_head->m_nextpkt = m->m_nextpkt;
1084 m->m_nextpkt = NULL;
1086 tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
1087 th = tcp_lro_get_th(m);
1088 ts_ptr = (uint32_t *)(th + 1);
1089 tcp_opt_len = (th->th_off << 2);
1090 tcp_opt_len -= sizeof(*th);
1091 tcp_data_len_total = le->m_head->m_pkthdr.lro_tcp_d_len + tcp_data_len;
1092 tcp_data_seg_total = le->m_head->m_pkthdr.lro_nsegs + m->m_pkthdr.lro_nsegs;
1094 if (tcp_data_seg_total >= lc->lro_ackcnt_lim ||
1095 tcp_data_len_total >= lc->lro_length_lim) {
1096 /* Flush now if appending will result in overflow. */
1097 tcp_push_and_replace(lc, le, m);
1100 if (tcp_opt_len != 0 &&
1101 __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1102 *ts_ptr != TCP_LRO_TS_OPTION)) {
1104 * Maybe a sack in the new one? We need to
1105 * start all over after flushing the
1106 * current le. We will go up to the beginning
1107 * and flush it (calling the replace again possibly
1108 * or just returning).
1110 tcp_push_and_replace(lc, le, m);
1113 if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1114 tcp_push_and_replace(lc, le, m);
1117 if (tcp_opt_len != 0) {
1118 uint32_t tsval = ntohl(*(ts_ptr + 1));
1119 /* Make sure timestamp values are increasing. */
1120 if (TSTMP_GT(le->tsval, tsval)) {
1121 tcp_push_and_replace(lc, le, m);
1125 le->tsecr = *(ts_ptr + 2);
1127 /* Try to append the new segment. */
1128 if (__predict_false(ntohl(th->th_seq) != le->next_seq ||
1129 ((tcp_get_flags(th) & TH_ACK) !=
1130 (le->flags & TH_ACK)) ||
1131 (tcp_data_len == 0 &&
1132 le->ack_seq == th->th_ack &&
1133 le->window == th->th_win))) {
1134 /* Out of order packet, non-ACK + ACK or dup ACK. */
1135 tcp_push_and_replace(lc, le, m);
1138 if (tcp_data_len != 0 ||
1139 SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1140 le->next_seq += tcp_data_len;
1141 le->ack_seq = th->th_ack;
1142 le->window = th->th_win;
1143 le->needs_merge = 1;
1144 } else if (th->th_ack == le->ack_seq) {
1145 if (WIN_GT(th->th_win, le->window)) {
1146 le->window = th->th_win;
1147 le->needs_merge = 1;
1151 if (tcp_data_len == 0) {
1156 /* Merge TCP data checksum and length to head mbuf. */
1157 tcp_lro_mbuf_append_pkthdr(le, m);
1160 * Adjust the mbuf so that m_data points to the first byte of
1161 * the ULP payload. Adjust the mbuf to avoid complications and
1162 * append new segment to existing mbuf chain.
1164 m_adj(m, m->m_pkthdr.len - tcp_data_len);
1166 le->m_tail->m_next = m;
1167 le->m_tail = m_last(m);
1173 tcp_queue_pkts(struct inpcb *inp, struct tcpcb *tp, struct lro_entry *le)
1175 INP_WLOCK_ASSERT(inp);
1176 if (tp->t_in_pkt == NULL) {
1177 /* Nothing yet there */
1178 tp->t_in_pkt = le->m_head;
1179 tp->t_tail_pkt = le->m_last_mbuf;
1181 /* Already some there */
1182 tp->t_tail_pkt->m_nextpkt = le->m_head;
1183 tp->t_tail_pkt = le->m_last_mbuf;
1186 le->m_last_mbuf = NULL;
1189 static struct mbuf *
1190 tcp_lro_get_last_if_ackcmp(struct lro_ctrl *lc, struct lro_entry *le,
1191 struct inpcb *inp, int32_t *new_m, bool can_append_old_cmp)
1196 tp = intotcpcb(inp);
1197 if (__predict_false(tp == NULL))
1200 /* Look at the last mbuf if any in queue */
1201 if (can_append_old_cmp) {
1203 if (m != NULL && (m->m_flags & M_ACKCMP) != 0) {
1204 if (M_TRAILINGSPACE(m) >= sizeof(struct tcp_ackent)) {
1205 tcp_lro_log(tp, lc, le, NULL, 23, 0, 0, 0, 0);
1207 counter_u64_add(tcp_extra_mbuf, 1);
1210 /* Mark we ran out of space */
1211 inp->inp_flags2 |= INP_MBUF_L_ACKS;
1215 /* Decide mbuf size. */
1216 tcp_lro_log(tp, lc, le, NULL, 21, 0, 0, 0, 0);
1217 if (inp->inp_flags2 & INP_MBUF_L_ACKS)
1218 m = m_getcl(M_NOWAIT, MT_DATA, M_ACKCMP | M_PKTHDR);
1220 m = m_gethdr(M_NOWAIT, MT_DATA);
1222 if (__predict_false(m == NULL)) {
1223 counter_u64_add(tcp_would_have_but, 1);
1226 counter_u64_add(tcp_comp_total, 1);
1227 m->m_pkthdr.rcvif = lc->ifp;
1228 m->m_flags |= M_ACKCMP;
1233 static struct inpcb *
1234 tcp_lro_lookup(struct ifnet *ifp, struct lro_parser *pa)
1238 switch (pa->data.lro_type) {
1240 case LRO_TYPE_IPV6_TCP:
1241 inp = in6_pcblookup(&V_tcbinfo,
1242 &pa->data.s_addr.v6,
1244 &pa->data.d_addr.v6,
1251 case LRO_TYPE_IPV4_TCP:
1252 inp = in_pcblookup(&V_tcbinfo,
1269 tcp_lro_ack_valid(struct mbuf *m, struct tcphdr *th, uint32_t **ppts, bool *other_opts)
1272 * This function returns two bits of valuable information.
1273 * a) Is what is present capable of being ack-compressed,
1274 * we can ack-compress if there is no options or just
1275 * a timestamp option, and of course the th_flags must
1276 * be correct as well.
1277 * b) Our other options present such as SACK. This is
1278 * used to determine if we want to wakeup or not.
1282 switch (th->th_off << 2) {
1283 case (sizeof(*th) + TCPOLEN_TSTAMP_APPA):
1284 *ppts = (uint32_t *)(th + 1);
1285 /* Check if we have only one timestamp option. */
1286 if (**ppts == TCP_LRO_TS_OPTION)
1287 *other_opts = false;
1296 *other_opts = false;
1304 /* For ACKCMP we only accept ACK, PUSH, ECE and CWR. */
1305 if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH | TH_ECE | TH_CWR)) != 0)
1307 /* If it has data on it we cannot compress it */
1308 if (m->m_pkthdr.lro_tcp_d_len)
1311 /* ACK flag must be set. */
1312 if (!(tcp_get_flags(th) & TH_ACK))
1318 tcp_lro_flush_tcphpts(struct lro_ctrl *lc, struct lro_entry *le)
1322 struct mbuf **pp, *cmp, *mv_to;
1323 struct ifnet *lagg_ifp;
1324 bool bpf_req, lagg_bpf_req, should_wake, can_append_old_cmp;
1326 /* Check if packet doesn't belongs to our network interface. */
1327 if ((tcplro_stacks_wanting_mbufq == 0) ||
1328 (le->outer.data.vlan_id != 0) ||
1329 (le->inner.data.lro_type != LRO_TYPE_NONE))
1330 return (TCP_LRO_CANNOT);
1334 * Be proactive about unspecified IPv6 address in source. As
1335 * we use all-zero to indicate unbounded/unconnected pcb,
1336 * unspecified IPv6 address can be used to confuse us.
1338 * Note that packets with unspecified IPv6 destination is
1339 * already dropped in ip6_input.
1341 if (__predict_false(le->outer.data.lro_type == LRO_TYPE_IPV6_TCP &&
1342 IN6_IS_ADDR_UNSPECIFIED(&le->outer.data.s_addr.v6)))
1343 return (TCP_LRO_CANNOT);
1345 if (__predict_false(le->inner.data.lro_type == LRO_TYPE_IPV6_TCP &&
1346 IN6_IS_ADDR_UNSPECIFIED(&le->inner.data.s_addr.v6)))
1347 return (TCP_LRO_CANNOT);
1349 /* Lookup inp, if any. */
1350 inp = tcp_lro_lookup(lc->ifp,
1351 (le->inner.data.lro_type == LRO_TYPE_NONE) ? &le->outer : &le->inner);
1353 return (TCP_LRO_CANNOT);
1355 counter_u64_add(tcp_inp_lro_locks_taken, 1);
1357 /* Get TCP control structure. */
1358 tp = intotcpcb(inp);
1360 /* Check if the inp is dead, Jim. */
1361 if (tp->t_state == TCPS_TIME_WAIT) {
1363 return (TCP_LRO_CANNOT);
1365 if ((inp->inp_irq_cpu_set == 0) && (lc->lro_cpu_is_set == 1)) {
1366 inp->inp_irq_cpu = lc->lro_last_cpu;
1367 inp->inp_irq_cpu_set = 1;
1369 /* Check if the transport doesn't support the needed optimizations. */
1370 if ((inp->inp_flags2 & (INP_SUPPORTS_MBUFQ | INP_MBUF_ACKCMP)) == 0) {
1372 return (TCP_LRO_CANNOT);
1375 if (inp->inp_flags2 & INP_MBUF_QUEUE_READY)
1376 should_wake = false;
1379 /* Check if packets should be tapped to BPF. */
1380 bpf_req = bpf_peers_present(lc->ifp->if_bpf);
1381 lagg_bpf_req = false;
1383 if (lc->ifp->if_type == IFT_IEEE8023ADLAG ||
1384 lc->ifp->if_type == IFT_INFINIBANDLAG) {
1385 struct lagg_port *lp = lc->ifp->if_lagg;
1386 struct lagg_softc *sc = lp->lp_softc;
1388 lagg_ifp = sc->sc_ifp;
1389 if (lagg_ifp != NULL)
1390 lagg_bpf_req = bpf_peers_present(lagg_ifp->if_bpf);
1393 /* Strip and compress all the incoming packets. */
1394 can_append_old_cmp = true;
1396 for (pp = &le->m_head; *pp != NULL; ) {
1398 if (do_bpf_strip_and_compress(inp, lc, le, pp,
1399 &cmp, &mv_to, &should_wake, bpf_req,
1400 lagg_bpf_req, lagg_ifp, can_append_old_cmp) == false) {
1401 /* Advance to next mbuf. */
1402 pp = &(*pp)->m_nextpkt;
1404 * Once we have appended we can't look in the pending
1405 * inbound packets for a compressed ack to append to.
1407 can_append_old_cmp = false;
1409 * Once we append we also need to stop adding to any
1410 * compressed ack we were remembering. A new cmp
1411 * ack will be required.
1414 tcp_lro_log(tp, lc, le, NULL, 25, 0, 0, 0, 0);
1415 } else if (mv_to != NULL) {
1416 /* We are asked to move pp up */
1417 pp = &mv_to->m_nextpkt;
1418 tcp_lro_log(tp, lc, le, NULL, 24, 0, 0, 0, 0);
1420 tcp_lro_log(tp, lc, le, NULL, 26, 0, 0, 0, 0);
1422 /* Update "m_last_mbuf", if any. */
1423 if (pp == &le->m_head)
1424 le->m_last_mbuf = *pp;
1426 le->m_last_mbuf = __containerof(pp, struct mbuf, m_nextpkt);
1428 /* Check if any data mbufs left. */
1429 if (le->m_head != NULL) {
1430 counter_u64_add(tcp_inp_lro_direct_queue, 1);
1431 tcp_lro_log(tp, lc, le, NULL, 22, 1, inp->inp_flags2, 0, 1);
1432 tcp_queue_pkts(inp, tp, le);
1436 counter_u64_add(tcp_inp_lro_wokeup_queue, 1);
1437 if ((*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0))
1442 return (0); /* Success. */
1447 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
1449 /* Only optimise if there are multiple packets waiting. */
1456 CURVNET_SET(lc->ifp->if_vnet);
1457 error = tcp_lro_flush_tcphpts(lc, le);
1461 tcp_lro_condense(lc, le);
1462 tcp_flush_out_entry(lc, le);
1467 bzero(le, sizeof(*le));
1468 LIST_INSERT_HEAD(&lc->lro_free, le, next);
1471 #ifdef HAVE_INLINE_FLSLL
1472 #define tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
1474 static inline uint64_t
1475 tcp_lro_msb_64(uint64_t x)
1483 return (x & ~(x >> 1));
1488 * The tcp_lro_sort() routine is comparable to qsort(), except it has
1489 * a worst case complexity limit of O(MIN(N,64)*N), where N is the
1490 * number of elements to sort and 64 is the number of sequence bits
1491 * available. The algorithm is bit-slicing the 64-bit sequence number,
1492 * sorting one bit at a time from the most significant bit until the
1493 * least significant one, skipping the constant bits. This is
1494 * typically called a radix sort.
1497 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
1499 struct lro_mbuf_sort temp;
1506 /* for small arrays insertion sort is faster */
1508 for (x = 1; x < size; x++) {
1510 for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
1511 parray[y] = parray[y - 1];
1517 /* compute sequence bits which are constant */
1520 for (x = 0; x != size; x++) {
1521 ones |= parray[x].seq;
1522 zeros |= ~parray[x].seq;
1525 /* compute bits which are not constant into "ones" */
1530 /* pick the most significant bit which is not constant */
1531 ones = tcp_lro_msb_64(ones);
1534 * Move entries having cleared sequence bits to the beginning
1537 for (x = y = 0; y != size; y++) {
1539 if (parray[y].seq & ones)
1543 parray[x] = parray[y];
1548 KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
1551 tcp_lro_sort(parray, x);
1560 tcp_lro_flush_all(struct lro_ctrl *lc)
1567 /* check if no mbufs to flush */
1568 if (lc->lro_mbuf_count == 0)
1570 if (lc->lro_cpu_is_set == 0) {
1571 if (lc->lro_last_cpu == curcpu) {
1572 lc->lro_cnt_of_same_cpu++;
1573 /* Have we reached the threshold to declare a cpu? */
1574 if (lc->lro_cnt_of_same_cpu > tcp_lro_cpu_set_thresh)
1575 lc->lro_cpu_is_set = 1;
1577 lc->lro_last_cpu = curcpu;
1578 lc->lro_cnt_of_same_cpu = 0;
1581 CURVNET_SET(lc->ifp->if_vnet);
1583 /* get current time */
1584 binuptime(&lc->lro_last_queue_time);
1586 /* sort all mbufs according to stream */
1587 tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
1589 /* input data into LRO engine, stream by stream */
1591 for (x = 0; x != lc->lro_mbuf_count; x++) {
1595 mb = lc->lro_mbuf_data[x].mb;
1597 /* get sequence number, masking away the packet index */
1598 nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
1600 /* check for new stream */
1604 /* flush active streams */
1605 tcp_lro_rx_done(lc);
1608 /* add packet to LRO engine */
1609 if (tcp_lro_rx_common(lc, mb, 0, false) != 0) {
1610 /* Flush anything we have acummulated */
1611 tcp_lro_flush_active(lc);
1612 /* input packet to network layer */
1613 (*lc->ifp->if_input)(lc->ifp, mb);
1620 /* flush active streams */
1621 tcp_lro_rx_done(lc);
1626 lc->lro_mbuf_count = 0;
1631 build_ack_entry(struct tcp_ackent *ae, struct tcphdr *th, struct mbuf *m,
1632 uint32_t *ts_ptr, uint16_t iptos)
1635 * Given a TCP ACK, summarize it down into the small TCP ACK
1638 ae->timestamp = m->m_pkthdr.rcv_tstmp;
1640 if (m->m_flags & M_TSTMP_LRO)
1641 ae->flags |= TSTMP_LRO;
1642 else if (m->m_flags & M_TSTMP)
1643 ae->flags |= TSTMP_HDWR;
1644 ae->seq = ntohl(th->th_seq);
1645 ae->ack = ntohl(th->th_ack);
1646 ae->flags |= tcp_get_flags(th);
1647 if (ts_ptr != NULL) {
1648 ae->ts_value = ntohl(ts_ptr[1]);
1649 ae->ts_echo = ntohl(ts_ptr[2]);
1650 ae->flags |= HAS_TSTMP;
1652 ae->win = ntohs(th->th_win);
1653 ae->codepoint = iptos;
1657 * Do BPF tap for either ACK_CMP packets or MBUF QUEUE type packets
1658 * and strip all, but the IPv4/IPv6 header.
1661 do_bpf_strip_and_compress(struct inpcb *inp, struct lro_ctrl *lc,
1662 struct lro_entry *le, struct mbuf **pp, struct mbuf **cmp, struct mbuf **mv_to,
1663 bool *should_wake, bool bpf_req, bool lagg_bpf_req, struct ifnet *lagg_ifp, bool can_append_old_cmp)
1668 struct ip6_hdr *ip6;
1673 struct tcp_ackent *ack_ent;
1676 bool other_opts, can_compress;
1682 /* Get current mbuf. */
1685 /* Let the BPF see the packet */
1686 if (__predict_false(bpf_req))
1687 ETHER_BPF_MTAP(lc->ifp, m);
1689 if (__predict_false(lagg_bpf_req))
1690 ETHER_BPF_MTAP(lagg_ifp, m);
1692 tcp_hdr_offset = m->m_pkthdr.lro_tcp_h_off;
1693 lro_type = le->inner.data.lro_type;
1696 lro_type = le->outer.data.lro_type;
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 case LRO_TYPE_IPV4_TCP:
1711 tcp_hdr_offset -= sizeof(*le->outer.ip4);
1712 m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1714 case LRO_TYPE_IPV6_TCP:
1715 tcp_hdr_offset -= sizeof(*le->outer.ip6);
1716 m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1722 MPASS(tcp_hdr_offset >= 0);
1724 m_adj(m, tcp_hdr_offset);
1725 m->m_flags |= M_LRO_EHDRSTRP;
1726 m->m_flags &= ~M_ACKCMP;
1727 m->m_pkthdr.lro_tcp_h_off -= tcp_hdr_offset;
1729 th = tcp_lro_get_th(m);
1731 th->th_sum = 0; /* TCP checksum is valid. */
1733 /* Check if ACK can be compressed */
1734 can_compress = tcp_lro_ack_valid(m, th, &ts_ptr, &other_opts);
1736 /* Now lets look at the should wake states */
1737 if ((other_opts == true) &&
1738 ((inp->inp_flags2 & INP_DONT_SACK_QUEUE) == 0)) {
1740 * If there are other options (SACK?) and the
1741 * tcp endpoint has not expressly told us it does
1742 * not care about SACKS, then we should wake up.
1744 *should_wake = true;
1746 /* Is the ack compressable? */
1747 if (can_compress == false)
1749 /* Does the TCP endpoint support ACK compression? */
1750 if ((inp->inp_flags2 & INP_MBUF_ACKCMP) == 0)
1753 /* Lets get the TOS/traffic class field */
1754 l3.ptr = mtod(m, void *);
1756 case LRO_TYPE_IPV4_TCP:
1757 iptos = l3.ip4->ip_tos;
1759 case LRO_TYPE_IPV6_TCP:
1760 iptos = IPV6_TRAFFIC_CLASS(l3.ip6);
1763 iptos = 0; /* Keep compiler happy. */
1766 /* Now lets get space if we don't have some already */
1769 nm = tcp_lro_get_last_if_ackcmp(lc, le, inp, &n_mbuf, can_append_old_cmp);
1770 if (__predict_false(nm == NULL))
1775 * Link in the new cmp ack to our in-order place,
1776 * first set our cmp ack's next to where we are.
1781 * Set it up so mv_to is advanced to our
1782 * compressed ack. This way the caller can
1783 * advance pp to the right place.
1787 * Advance it here locally as well.
1789 pp = &nm->m_nextpkt;
1792 /* We have one already we are working on */
1794 if (M_TRAILINGSPACE(nm) < sizeof(struct tcp_ackent)) {
1795 /* We ran out of space */
1796 inp->inp_flags2 |= INP_MBUF_L_ACKS;
1800 MPASS(M_TRAILINGSPACE(nm) >= sizeof(struct tcp_ackent));
1801 counter_u64_add(tcp_inp_lro_compressed, 1);
1803 /* We can add in to the one on the tail */
1804 ack_ent = mtod(nm, struct tcp_ackent *);
1805 idx = (nm->m_len / sizeof(struct tcp_ackent));
1806 build_ack_entry(&ack_ent[idx], th, m, ts_ptr, iptos);
1808 /* Bump the size of both pkt-hdr and len */
1809 nm->m_len += sizeof(struct tcp_ackent);
1810 nm->m_pkthdr.len += sizeof(struct tcp_ackent);
1812 /* Advance to next mbuf before freeing. */
1814 m->m_nextpkt = NULL;
1818 counter_u64_add(tcp_uncomp_total, 1);
1824 static struct lro_head *
1825 tcp_lro_rx_get_bucket(struct lro_ctrl *lc, struct mbuf *m, struct lro_parser *parser)
1829 if (M_HASHTYPE_ISHASH(m)) {
1830 hash = m->m_pkthdr.flowid;
1832 for (unsigned i = hash = 0; i != LRO_RAW_ADDRESS_MAX; i++)
1833 hash += parser->data.raw[i];
1835 return (&lc->lro_hash[hash % lc->lro_hashsz]);
1839 tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash)
1841 struct lro_parser pi; /* inner address data */
1842 struct lro_parser po; /* outer address data */
1843 struct lro_parser *pa; /* current parser for TCP stream */
1844 struct lro_entry *le;
1845 struct lro_head *bucket;
1850 uint16_t tcp_data_sum;
1853 /* Quickly decide if packet cannot be LRO'ed */
1854 if (__predict_false(V_ipforwarding != 0))
1855 return (TCP_LRO_CANNOT);
1858 /* Quickly decide if packet cannot be LRO'ed */
1859 if (__predict_false(V_ip6_forwarding != 0))
1860 return (TCP_LRO_CANNOT);
1862 if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1863 ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1864 (m->m_pkthdr.csum_data != 0xffff)) {
1866 * The checksum either did not have hardware offload
1867 * or it was a bad checksum. We can't LRO such
1870 counter_u64_add(tcp_bad_csums, 1);
1871 return (TCP_LRO_CANNOT);
1873 /* We expect a contiguous header [eh, ip, tcp]. */
1874 pa = tcp_lro_parser(m, &po, &pi, true);
1875 if (__predict_false(pa == NULL))
1876 return (TCP_LRO_NOT_SUPPORTED);
1878 /* We don't expect any padding. */
1879 error = tcp_lro_trim_mbuf_chain(m, pa);
1880 if (__predict_false(error != 0))
1884 switch (pa->data.lro_type) {
1885 case LRO_TYPE_IPV4_TCP:
1886 error = tcp_lro_rx_ipv4(lc, m, pa->ip4);
1887 if (__predict_false(error != 0))
1894 /* If no hardware or arrival stamp on the packet add timestamp */
1895 if ((m->m_flags & (M_TSTMP_LRO | M_TSTMP)) == 0) {
1896 m->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
1897 m->m_flags |= M_TSTMP_LRO;
1900 /* Get pointer to TCP header. */
1903 /* Don't process SYN packets. */
1904 if (__predict_false(tcp_get_flags(th) & TH_SYN))
1905 return (TCP_LRO_CANNOT);
1907 /* Get total TCP header length and compute payload length. */
1908 tcp_opt_len = (th->th_off << 2);
1909 tcp_data_len = m->m_pkthdr.len - ((uint8_t *)th -
1910 (uint8_t *)m->m_data) - tcp_opt_len;
1911 tcp_opt_len -= sizeof(*th);
1913 /* Don't process invalid TCP headers. */
1914 if (__predict_false(tcp_opt_len < 0 || tcp_data_len < 0))
1915 return (TCP_LRO_CANNOT);
1917 /* Compute TCP data only checksum. */
1918 if (tcp_data_len == 0)
1919 tcp_data_sum = 0; /* no data, no checksum */
1920 else if (__predict_false(csum != 0))
1921 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~csum);
1923 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~th->th_sum);
1925 /* Save TCP info in mbuf. */
1926 m->m_nextpkt = NULL;
1927 m->m_pkthdr.rcvif = lc->ifp;
1928 m->m_pkthdr.lro_tcp_d_csum = tcp_data_sum;
1929 m->m_pkthdr.lro_tcp_d_len = tcp_data_len;
1930 m->m_pkthdr.lro_tcp_h_off = ((uint8_t *)th - (uint8_t *)m->m_data);
1931 m->m_pkthdr.lro_nsegs = 1;
1933 /* Get hash bucket. */
1935 bucket = &lc->lro_hash[0];
1937 bucket = tcp_lro_rx_get_bucket(lc, m, pa);
1940 /* Try to find a matching previous segment. */
1941 LIST_FOREACH(le, bucket, hash_next) {
1942 /* Compare addresses and ports. */
1943 if (lro_address_compare(&po.data, &le->outer.data) == false ||
1944 lro_address_compare(&pi.data, &le->inner.data) == false)
1947 /* Check if no data and old ACK. */
1948 if (tcp_data_len == 0 &&
1949 SEQ_LT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1954 /* Mark "m" in the last spot. */
1955 le->m_last_mbuf->m_nextpkt = m;
1956 /* Now set the tail to "m". */
1957 le->m_last_mbuf = m;
1961 /* Try to find an empty slot. */
1962 if (LIST_EMPTY(&lc->lro_free))
1963 return (TCP_LRO_NO_ENTRIES);
1965 /* Start a new segment chain. */
1966 le = LIST_FIRST(&lc->lro_free);
1967 LIST_REMOVE(le, next);
1968 tcp_lro_active_insert(lc, bucket, le);
1970 /* Make sure the headers are set. */
1974 /* Store time this entry was allocated. */
1975 le->alloc_time = lc->lro_last_queue_time;
1977 tcp_set_entry_to_mbuf(lc, le, m, th);
1979 /* Now set the tail to "m". */
1980 le->m_last_mbuf = m;
1986 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
1990 if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1991 ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1992 (m->m_pkthdr.csum_data != 0xffff)) {
1994 * The checksum either did not have hardware offload
1995 * or it was a bad checksum. We can't LRO such
1998 counter_u64_add(tcp_bad_csums, 1);
1999 return (TCP_LRO_CANNOT);
2001 /* get current time */
2002 binuptime(&lc->lro_last_queue_time);
2003 CURVNET_SET(lc->ifp->if_vnet);
2004 error = tcp_lro_rx_common(lc, m, csum, true);
2005 if (__predict_false(error != 0)) {
2007 * Flush anything we have acummulated
2008 * ahead of this packet that can't
2009 * be LRO'd. This preserves order.
2011 tcp_lro_flush_active(lc);
2019 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
2023 if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
2024 lc->lro_mbuf_max == 0)) {
2030 /* check if packet is not LRO capable */
2031 if (__predict_false((lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
2032 /* input packet to network layer */
2033 (*lc->ifp->if_input) (lc->ifp, mb);
2037 /* If no hardware or arrival stamp on the packet add timestamp */
2038 if ((tcplro_stacks_wanting_mbufq > 0) &&
2039 (tcp_less_accurate_lro_ts == 0) &&
2040 ((mb->m_flags & M_TSTMP) == 0)) {
2041 /* Add in an LRO time since no hardware */
2042 binuptime(&lc->lro_last_queue_time);
2043 mb->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
2044 mb->m_flags |= M_TSTMP_LRO;
2047 /* create sequence number */
2048 lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
2049 (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
2050 (((uint64_t)mb->m_pkthdr.flowid) << 24) |
2051 ((uint64_t)lc->lro_mbuf_count);
2054 lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb;
2056 /* flush if array is full */
2057 if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max))
2058 tcp_lro_flush_all(lc);