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_private.h>
59 #include <net/if_types.h>
60 #include <net/infiniband.h>
61 #include <net/if_lagg.h>
63 #include <netinet/in_systm.h>
64 #include <netinet/in.h>
65 #include <netinet/ip6.h>
66 #include <netinet/ip.h>
67 #include <netinet/ip_var.h>
68 #include <netinet/in_pcb.h>
69 #include <netinet6/in6_pcb.h>
70 #include <netinet/tcp.h>
71 #include <netinet/tcp_seq.h>
72 #include <netinet/tcp_lro.h>
73 #include <netinet/tcp_var.h>
74 #include <netinet/tcpip.h>
75 #include <netinet/tcp_hpts.h>
76 #include <netinet/tcp_log_buf.h>
77 #include <netinet/tcp_fsm.h>
78 #include <netinet/udp.h>
79 #include <netinet6/ip6_var.h>
81 #include <machine/in_cksum.h>
83 static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures");
85 #define TCP_LRO_TS_OPTION \
86 ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
87 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
89 static void tcp_lro_rx_done(struct lro_ctrl *lc);
90 static int tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m,
91 uint32_t csum, bool use_hash);
94 static bool do_bpf_strip_and_compress(struct inpcb *, struct lro_ctrl *,
95 struct lro_entry *, struct mbuf **, struct mbuf **, struct mbuf **,
96 bool *, bool, bool, struct ifnet *, bool);
100 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
103 static long tcplro_stacks_wanting_mbufq;
104 counter_u64_t tcp_inp_lro_direct_queue;
105 counter_u64_t tcp_inp_lro_wokeup_queue;
106 counter_u64_t tcp_inp_lro_compressed;
107 counter_u64_t tcp_inp_lro_locks_taken;
108 counter_u64_t tcp_extra_mbuf;
109 counter_u64_t tcp_would_have_but;
110 counter_u64_t tcp_comp_total;
111 counter_u64_t tcp_uncomp_total;
112 counter_u64_t tcp_bad_csums;
114 static unsigned tcp_lro_entries = TCP_LRO_ENTRIES;
115 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries,
116 CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_entries, 0,
117 "default number of LRO entries");
119 static uint32_t tcp_lro_cpu_set_thresh = TCP_LRO_CPU_DECLARATION_THRESH;
120 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_cpu_threshold,
121 CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_cpu_set_thresh, 0,
122 "Number of interrupts in a row on the same CPU that will make us declare an 'affinity' cpu?");
124 static uint32_t tcp_less_accurate_lro_ts = 0;
125 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_less_accurate,
126 CTLFLAG_MPSAFE, &tcp_less_accurate_lro_ts, 0,
127 "Do we trade off efficency by doing less timestamp operations for time accuracy?");
129 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, fullqueue, CTLFLAG_RD,
130 &tcp_inp_lro_direct_queue, "Number of lro's fully queued to transport");
131 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, wokeup, CTLFLAG_RD,
132 &tcp_inp_lro_wokeup_queue, "Number of lro's where we woke up transport via hpts");
133 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, compressed, CTLFLAG_RD,
134 &tcp_inp_lro_compressed, "Number of lro's compressed and sent to transport");
135 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lockcnt, CTLFLAG_RD,
136 &tcp_inp_lro_locks_taken, "Number of lro's inp_wlocks taken");
137 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, extra_mbuf, CTLFLAG_RD,
138 &tcp_extra_mbuf, "Number of times we had an extra compressed ack dropped into the tp");
139 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, would_have_but, CTLFLAG_RD,
140 &tcp_would_have_but, "Number of times we would have had an extra compressed, but mget failed");
141 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, with_m_ackcmp, CTLFLAG_RD,
142 &tcp_comp_total, "Number of mbufs queued with M_ACKCMP flags set");
143 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, without_m_ackcmp, CTLFLAG_RD,
144 &tcp_uncomp_total, "Number of mbufs queued without M_ACKCMP");
145 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lro_badcsum, CTLFLAG_RD,
146 &tcp_bad_csums, "Number of packets that the common code saw with bad csums");
149 tcp_lro_reg_mbufq(void)
151 atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, 1);
155 tcp_lro_dereg_mbufq(void)
157 atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, -1);
161 tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket,
162 struct lro_entry *le)
165 LIST_INSERT_HEAD(&lc->lro_active, le, next);
166 LIST_INSERT_HEAD(bucket, le, hash_next);
170 tcp_lro_active_remove(struct lro_entry *le)
173 LIST_REMOVE(le, next); /* active list */
174 LIST_REMOVE(le, hash_next); /* hash bucket */
178 tcp_lro_init(struct lro_ctrl *lc)
180 return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0));
184 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
185 unsigned lro_entries, unsigned lro_mbufs)
187 struct lro_entry *le;
189 unsigned i, elements;
191 lc->lro_bad_csum = 0;
194 lc->lro_mbuf_count = 0;
195 lc->lro_mbuf_max = lro_mbufs;
196 lc->lro_cnt = lro_entries;
197 lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
198 lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
200 LIST_INIT(&lc->lro_free);
201 LIST_INIT(&lc->lro_active);
203 /* create hash table to accelerate entry lookup */
204 if (lro_entries > lro_mbufs)
205 elements = lro_entries;
207 elements = lro_mbufs;
208 lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz,
210 if (lc->lro_hash == NULL) {
211 memset(lc, 0, sizeof(*lc));
215 /* compute size to allocate */
216 size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
217 (lro_entries * sizeof(*le));
218 lc->lro_mbuf_data = (struct lro_mbuf_sort *)
219 malloc(size, M_LRO, M_NOWAIT | M_ZERO);
221 /* check for out of memory */
222 if (lc->lro_mbuf_data == NULL) {
223 free(lc->lro_hash, M_LRO);
224 memset(lc, 0, sizeof(*lc));
227 /* compute offset for LRO entries */
228 le = (struct lro_entry *)
229 (lc->lro_mbuf_data + lro_mbufs);
231 /* setup linked list */
232 for (i = 0; i != lro_entries; i++)
233 LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
238 struct vxlan_header {
244 tcp_lro_low_level_parser(void *ptr, struct lro_parser *parser, bool update_data, bool is_vxlan, int mlen)
246 const struct ether_vlan_header *eh;
251 memset(parser, 0, sizeof(*parser));
256 const struct vxlan_header *vxh;
258 ptr = (uint8_t *)ptr + sizeof(*vxh);
260 parser->data.vxlan_vni =
261 vxh->vxlh_vni & htonl(0xffffff00);
266 if (__predict_false(eh->evl_encap_proto == htons(ETHERTYPE_VLAN))) {
267 eth_type = eh->evl_proto;
269 /* strip priority and keep VLAN ID only */
270 parser->data.vlan_id = eh->evl_tag & htons(EVL_VLID_MASK);
272 /* advance to next header */
273 ptr = (uint8_t *)ptr + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
274 mlen -= (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
276 eth_type = eh->evl_encap_proto;
277 /* advance to next header */
278 mlen -= ETHER_HDR_LEN;
279 ptr = (uint8_t *)ptr + ETHER_HDR_LEN;
281 if (__predict_false(mlen <= 0))
285 case htons(ETHERTYPE_IP):
287 if (__predict_false(mlen < sizeof(struct ip)))
289 /* Ensure there are no IPv4 options. */
290 if ((parser->ip4->ip_hl << 2) != sizeof (*parser->ip4))
292 /* .. and the packet is not fragmented. */
293 if (parser->ip4->ip_off & htons(IP_MF|IP_OFFMASK))
295 ptr = (uint8_t *)ptr + (parser->ip4->ip_hl << 2);
296 mlen -= sizeof(struct ip);
298 parser->data.s_addr.v4 = parser->ip4->ip_src;
299 parser->data.d_addr.v4 = parser->ip4->ip_dst;
301 switch (parser->ip4->ip_p) {
303 if (__predict_false(mlen < sizeof(struct udphdr)))
307 parser->data.lro_type = LRO_TYPE_IPV4_UDP;
308 parser->data.s_port = parser->udp->uh_sport;
309 parser->data.d_port = parser->udp->uh_dport;
311 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_UDP);
313 ptr = ((uint8_t *)ptr + sizeof(*parser->udp));
314 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
318 if (__predict_false(mlen < sizeof(struct tcphdr)))
321 parser->data.lro_type = LRO_TYPE_IPV4_TCP;
322 parser->data.s_port = parser->tcp->th_sport;
323 parser->data.d_port = parser->tcp->th_dport;
325 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_TCP);
327 if (__predict_false(mlen < (parser->tcp->th_off << 2)))
329 ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
330 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
338 case htons(ETHERTYPE_IPV6):
340 if (__predict_false(mlen < sizeof(struct ip6_hdr)))
342 ptr = (uint8_t *)ptr + sizeof(*parser->ip6);
344 parser->data.s_addr.v6 = parser->ip6->ip6_src;
345 parser->data.d_addr.v6 = parser->ip6->ip6_dst;
347 mlen -= sizeof(struct ip6_hdr);
348 switch (parser->ip6->ip6_nxt) {
350 if (__predict_false(mlen < sizeof(struct udphdr)))
354 parser->data.lro_type = LRO_TYPE_IPV6_UDP;
355 parser->data.s_port = parser->udp->uh_sport;
356 parser->data.d_port = parser->udp->uh_dport;
358 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_UDP);
360 ptr = (uint8_t *)ptr + sizeof(*parser->udp);
361 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
364 if (__predict_false(mlen < sizeof(struct tcphdr)))
368 parser->data.lro_type = LRO_TYPE_IPV6_TCP;
369 parser->data.s_port = parser->tcp->th_sport;
370 parser->data.d_port = parser->tcp->th_dport;
372 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_TCP);
374 if (__predict_false(mlen < (parser->tcp->th_off << 2)))
376 ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
377 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
387 /* Invalid packet - cannot parse */
391 static const int vxlan_csum = CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
392 CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID;
394 static inline struct lro_parser *
395 tcp_lro_parser(struct mbuf *m, struct lro_parser *po, struct lro_parser *pi, bool update_data)
399 /* Try to parse outer headers first. */
400 data_ptr = tcp_lro_low_level_parser(m->m_data, po, update_data, false, m->m_len);
401 if (data_ptr == NULL || po->total_hdr_len > m->m_len)
405 /* Store VLAN ID, if any. */
406 if (__predict_false(m->m_flags & M_VLANTAG)) {
408 htons(m->m_pkthdr.ether_vtag) & htons(EVL_VLID_MASK);
410 /* Store decrypted flag, if any. */
411 if (__predict_false((m->m_pkthdr.csum_flags &
412 CSUM_TLS_MASK) == CSUM_TLS_DECRYPTED))
413 po->data.lro_flags |= LRO_FLAG_DECRYPTED;
416 switch (po->data.lro_type) {
417 case LRO_TYPE_IPV4_UDP:
418 case LRO_TYPE_IPV6_UDP:
419 /* Check for VXLAN headers. */
420 if ((m->m_pkthdr.csum_flags & vxlan_csum) != vxlan_csum)
423 /* Try to parse inner headers. */
424 data_ptr = tcp_lro_low_level_parser(data_ptr, pi, update_data, true,
425 (m->m_len - ((caddr_t)data_ptr - m->m_data)));
426 if (data_ptr == NULL || (pi->total_hdr_len + po->total_hdr_len) > m->m_len)
429 /* Verify supported header types. */
430 switch (pi->data.lro_type) {
431 case LRO_TYPE_IPV4_TCP:
432 case LRO_TYPE_IPV6_TCP:
438 case LRO_TYPE_IPV4_TCP:
439 case LRO_TYPE_IPV6_TCP:
441 memset(pi, 0, sizeof(*pi));
450 tcp_lro_trim_mbuf_chain(struct mbuf *m, const struct lro_parser *po)
454 switch (po->data.lro_type) {
456 case LRO_TYPE_IPV4_TCP:
457 len = ((uint8_t *)po->ip4 - (uint8_t *)m->m_data) +
458 ntohs(po->ip4->ip_len);
462 case LRO_TYPE_IPV6_TCP:
463 len = ((uint8_t *)po->ip6 - (uint8_t *)m->m_data) +
464 ntohs(po->ip6->ip6_plen) + sizeof(*po->ip6);
468 return (TCP_LRO_CANNOT);
472 * If the frame is padded beyond the end of the IP packet,
473 * then trim the extra bytes off:
475 if (__predict_true(m->m_pkthdr.len == len)) {
477 } else if (m->m_pkthdr.len > len) {
478 m_adj(m, len - m->m_pkthdr.len);
481 return (TCP_LRO_CANNOT);
484 static struct tcphdr *
485 tcp_lro_get_th(struct mbuf *m)
487 return ((struct tcphdr *)((uint8_t *)m->m_data + m->m_pkthdr.lro_tcp_h_off));
491 lro_free_mbuf_chain(struct mbuf *m)
504 tcp_lro_free(struct lro_ctrl *lc)
506 struct lro_entry *le;
509 /* reset LRO free list */
510 LIST_INIT(&lc->lro_free);
512 /* free active mbufs, if any */
513 while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
514 tcp_lro_active_remove(le);
515 lro_free_mbuf_chain(le->m_head);
518 /* free hash table */
519 free(lc->lro_hash, M_LRO);
523 /* free mbuf array, if any */
524 for (x = 0; x != lc->lro_mbuf_count; x++)
525 m_freem(lc->lro_mbuf_data[x].mb);
526 lc->lro_mbuf_count = 0;
528 /* free allocated memory, if any */
529 free(lc->lro_mbuf_data, M_LRO);
530 lc->lro_mbuf_data = NULL;
534 tcp_lro_rx_csum_tcphdr(const struct tcphdr *th)
540 csum = -th->th_sum; /* exclude checksum field */
542 ptr = (const uint16_t *)th;
549 while (csum > 0xffff)
550 csum = (csum >> 16) + (csum & 0xffff);
556 tcp_lro_rx_csum_data(const struct lro_parser *pa, uint16_t tcp_csum)
563 switch (pa->data.lro_type) {
565 case LRO_TYPE_IPV6_TCP:
566 /* Compute full pseudo IPv6 header checksum. */
567 cs = in6_cksum_pseudo(pa->ip6, ntohs(pa->ip6->ip6_plen), pa->ip6->ip6_nxt, 0);
571 case LRO_TYPE_IPV4_TCP:
572 /* Compute full pseudo IPv4 header checsum. */
573 cs = in_addword(ntohs(pa->ip4->ip_len) - sizeof(*pa->ip4), IPPROTO_TCP);
574 cs = in_pseudo(pa->ip4->ip_src.s_addr, pa->ip4->ip_dst.s_addr, htons(cs));
578 cs = 0; /* Keep compiler happy. */
582 /* Complement checksum. */
586 /* Remove TCP header checksum. */
587 cs = ~tcp_lro_rx_csum_tcphdr(pa->tcp);
590 /* Compute checksum remainder. */
592 c = (c >> 16) + (c & 0xffff);
598 tcp_lro_rx_done(struct lro_ctrl *lc)
600 struct lro_entry *le;
602 while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
603 tcp_lro_active_remove(le);
604 tcp_lro_flush(lc, le);
609 tcp_lro_flush_active(struct lro_ctrl *lc)
611 struct lro_entry *le;
614 * Walk through the list of le entries, and
615 * any one that does have packets flush. This
616 * is called because we have an inbound packet
617 * (e.g. SYN) that has to have all others flushed
618 * in front of it. Note we have to do the remove
619 * because tcp_lro_flush() assumes that the entry
620 * is being freed. This is ok it will just get
621 * reallocated again like it was new.
623 LIST_FOREACH(le, &lc->lro_active, next) {
624 if (le->m_head != NULL) {
625 tcp_lro_active_remove(le);
626 tcp_lro_flush(lc, le);
632 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
634 struct lro_entry *le, *le_tmp;
639 if (LIST_EMPTY(&lc->lro_active))
642 /* get timeout time and current time in ns */
644 now = bintime2ns(&bt);
645 tov = ((timeout->tv_sec * 1000000000) + (timeout->tv_usec * 1000));
646 LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
647 if (now >= (bintime2ns(&le->alloc_time) + tov)) {
648 tcp_lro_active_remove(le);
649 tcp_lro_flush(lc, le);
656 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4)
660 /* Legacy IP has a header checksum that needs to be correct. */
661 if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) {
662 if (__predict_false((m->m_pkthdr.csum_flags & CSUM_IP_VALID) == 0)) {
664 return (TCP_LRO_CANNOT);
667 csum = in_cksum_hdr(ip4);
668 if (__predict_false(csum != 0)) {
670 return (TCP_LRO_CANNOT);
679 tcp_lro_log(struct tcpcb *tp, const struct lro_ctrl *lc,
680 const struct lro_entry *le, const struct mbuf *m,
681 int frm, int32_t tcp_data_len, uint32_t th_seq,
682 uint32_t th_ack, uint16_t th_win)
684 if (tp->t_logstate != TCP_LOG_STATE_OFF) {
685 union tcp_log_stackspecific log;
686 struct timeval tv, btv;
689 cts = tcp_get_usecs(&tv);
690 memset(&log, 0, sizeof(union tcp_log_stackspecific));
691 log.u_bbr.flex8 = frm;
692 log.u_bbr.flex1 = tcp_data_len;
694 log.u_bbr.flex2 = m->m_pkthdr.len;
698 log.u_bbr.flex3 = le->m_head->m_pkthdr.lro_nsegs;
699 log.u_bbr.flex4 = le->m_head->m_pkthdr.lro_tcp_d_len;
700 log.u_bbr.flex5 = le->m_head->m_pkthdr.len;
701 log.u_bbr.delRate = le->m_head->m_flags;
702 log.u_bbr.rttProp = le->m_head->m_pkthdr.rcv_tstmp;
704 log.u_bbr.inflight = th_seq;
705 log.u_bbr.delivered = th_ack;
706 log.u_bbr.timeStamp = cts;
707 log.u_bbr.epoch = le->next_seq;
708 log.u_bbr.lt_epoch = le->ack_seq;
709 log.u_bbr.pacing_gain = th_win;
710 log.u_bbr.cwnd_gain = le->window;
711 log.u_bbr.lost = curcpu;
712 log.u_bbr.cur_del_rate = (uintptr_t)m;
713 log.u_bbr.bw_inuse = (uintptr_t)le->m_head;
714 bintime2timeval(&lc->lro_last_queue_time, &btv);
715 log.u_bbr.flex6 = tcp_tv_to_usectick(&btv);
716 log.u_bbr.flex7 = le->compressed;
717 log.u_bbr.pacing_gain = le->uncompressed;
718 if (in_epoch(net_epoch_preempt))
719 log.u_bbr.inhpts = 1;
721 log.u_bbr.inhpts = 0;
722 TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
723 &tptosocket(tp)->so_snd,
724 TCP_LOG_LRO, 0, 0, &log, false, &tv);
730 tcp_lro_assign_and_checksum_16(uint16_t *ptr, uint16_t value, uint16_t *psum)
734 csum = 0xffff - *ptr + value;
735 while (csum > 0xffff)
736 csum = (csum >> 16) + (csum & 0xffff);
742 tcp_lro_update_checksum(const struct lro_parser *pa, const struct lro_entry *le,
743 uint16_t payload_len, uint16_t delta_sum)
747 uint16_t temp[5] = {};
749 switch (pa->data.lro_type) {
750 case LRO_TYPE_IPV4_TCP:
751 /* Compute new IPv4 length. */
752 tlen = (pa->ip4->ip_hl << 2) + (pa->tcp->th_off << 2) + payload_len;
753 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
755 /* Subtract delta from current IPv4 checksum. */
756 csum = pa->ip4->ip_sum + 0xffff - temp[0];
757 while (csum > 0xffff)
758 csum = (csum >> 16) + (csum & 0xffff);
759 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
760 goto update_tcp_header;
762 case LRO_TYPE_IPV6_TCP:
763 /* Compute new IPv6 length. */
764 tlen = (pa->tcp->th_off << 2) + payload_len;
765 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
766 goto update_tcp_header;
768 case LRO_TYPE_IPV4_UDP:
769 /* Compute new IPv4 length. */
770 tlen = (pa->ip4->ip_hl << 2) + sizeof(*pa->udp) + payload_len;
771 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
773 /* Subtract delta from current IPv4 checksum. */
774 csum = pa->ip4->ip_sum + 0xffff - temp[0];
775 while (csum > 0xffff)
776 csum = (csum >> 16) + (csum & 0xffff);
777 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
778 goto update_udp_header;
780 case LRO_TYPE_IPV6_UDP:
781 /* Compute new IPv6 length. */
782 tlen = sizeof(*pa->udp) + payload_len;
783 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
784 goto update_udp_header;
791 /* Compute current TCP header checksum. */
792 temp[2] = tcp_lro_rx_csum_tcphdr(pa->tcp);
794 /* Incorporate the latest ACK into the TCP header. */
795 pa->tcp->th_ack = le->ack_seq;
796 pa->tcp->th_win = le->window;
798 /* Incorporate latest timestamp into the TCP header. */
799 if (le->timestamp != 0) {
802 ts_ptr = (uint32_t *)(pa->tcp + 1);
803 ts_ptr[1] = htonl(le->tsval);
804 ts_ptr[2] = le->tsecr;
807 /* Compute new TCP header checksum. */
808 temp[3] = tcp_lro_rx_csum_tcphdr(pa->tcp);
810 /* Compute new TCP checksum. */
811 csum = pa->tcp->th_sum + 0xffff - delta_sum +
812 0xffff - temp[0] + 0xffff - temp[3] + temp[2];
813 while (csum > 0xffff)
814 csum = (csum >> 16) + (csum & 0xffff);
816 /* Assign new TCP checksum. */
817 tcp_lro_assign_and_checksum_16(&pa->tcp->th_sum, csum, &temp[4]);
819 /* Compute all modififications affecting next checksum. */
820 csum = temp[0] + temp[1] + 0xffff - temp[2] +
821 temp[3] + temp[4] + delta_sum;
822 while (csum > 0xffff)
823 csum = (csum >> 16) + (csum & 0xffff);
825 /* Return delta checksum to next stage, if any. */
829 tlen = sizeof(*pa->udp) + payload_len;
830 /* Assign new UDP length and compute checksum delta. */
831 tcp_lro_assign_and_checksum_16(&pa->udp->uh_ulen, htons(tlen), &temp[2]);
833 /* Check if there is a UDP checksum. */
834 if (__predict_false(pa->udp->uh_sum != 0)) {
835 /* Compute new UDP checksum. */
836 csum = pa->udp->uh_sum + 0xffff - delta_sum +
837 0xffff - temp[0] + 0xffff - temp[2];
838 while (csum > 0xffff)
839 csum = (csum >> 16) + (csum & 0xffff);
840 /* Assign new UDP checksum. */
841 tcp_lro_assign_and_checksum_16(&pa->udp->uh_sum, csum, &temp[3]);
844 /* Compute all modififications affecting next checksum. */
845 csum = temp[0] + temp[1] + temp[2] + temp[3] + delta_sum;
846 while (csum > 0xffff)
847 csum = (csum >> 16) + (csum & 0xffff);
849 /* Return delta checksum to next stage, if any. */
854 tcp_flush_out_entry(struct lro_ctrl *lc, struct lro_entry *le)
856 /* Check if we need to recompute any checksums. */
857 if (le->needs_merge) {
860 switch (le->inner.data.lro_type) {
861 case LRO_TYPE_IPV4_TCP:
862 csum = tcp_lro_update_checksum(&le->inner, le,
863 le->m_head->m_pkthdr.lro_tcp_d_len,
864 le->m_head->m_pkthdr.lro_tcp_d_csum);
865 csum = tcp_lro_update_checksum(&le->outer, NULL,
866 le->m_head->m_pkthdr.lro_tcp_d_len +
867 le->inner.total_hdr_len, csum);
868 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
869 CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
870 le->m_head->m_pkthdr.csum_data = 0xffff;
871 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
872 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
874 case LRO_TYPE_IPV6_TCP:
875 csum = tcp_lro_update_checksum(&le->inner, le,
876 le->m_head->m_pkthdr.lro_tcp_d_len,
877 le->m_head->m_pkthdr.lro_tcp_d_csum);
878 csum = tcp_lro_update_checksum(&le->outer, NULL,
879 le->m_head->m_pkthdr.lro_tcp_d_len +
880 le->inner.total_hdr_len, csum);
881 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
883 le->m_head->m_pkthdr.csum_data = 0xffff;
884 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
885 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
888 switch (le->outer.data.lro_type) {
889 case LRO_TYPE_IPV4_TCP:
890 csum = tcp_lro_update_checksum(&le->outer, le,
891 le->m_head->m_pkthdr.lro_tcp_d_len,
892 le->m_head->m_pkthdr.lro_tcp_d_csum);
893 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
894 CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
895 le->m_head->m_pkthdr.csum_data = 0xffff;
896 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
897 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
899 case LRO_TYPE_IPV6_TCP:
900 csum = tcp_lro_update_checksum(&le->outer, le,
901 le->m_head->m_pkthdr.lro_tcp_d_len,
902 le->m_head->m_pkthdr.lro_tcp_d_csum);
903 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
905 le->m_head->m_pkthdr.csum_data = 0xffff;
906 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
907 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
919 * Break any chain, this is not set to NULL on the singleton
920 * case m_nextpkt points to m_head. Other case set them
921 * m_nextpkt to NULL in push_and_replace.
923 le->m_head->m_nextpkt = NULL;
924 lc->lro_queued += le->m_head->m_pkthdr.lro_nsegs;
925 (*lc->ifp->if_input)(lc->ifp, le->m_head);
929 tcp_set_entry_to_mbuf(struct lro_ctrl *lc, struct lro_entry *le,
930 struct mbuf *m, struct tcphdr *th)
933 uint16_t tcp_data_len;
934 uint16_t tcp_opt_len;
936 ts_ptr = (uint32_t *)(th + 1);
937 tcp_opt_len = (th->th_off << 2);
938 tcp_opt_len -= sizeof(*th);
940 /* Check if there is a timestamp option. */
941 if (tcp_opt_len == 0 ||
942 __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
943 *ts_ptr != TCP_LRO_TS_OPTION)) {
944 /* We failed to find the timestamp option. */
948 le->tsval = ntohl(*(ts_ptr + 1));
949 le->tsecr = *(ts_ptr + 2);
952 tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
954 /* Pull out TCP sequence numbers and window size. */
955 le->next_seq = ntohl(th->th_seq) + tcp_data_len;
956 le->ack_seq = th->th_ack;
957 le->window = th->th_win;
958 le->flags = tcp_get_flags(th);
961 /* Setup new data pointers. */
963 le->m_tail = m_last(m);
967 tcp_push_and_replace(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m)
969 struct lro_parser *pa;
972 * Push up the stack of the current entry
973 * and replace it with "m".
977 /* Grab off the next and save it */
978 msave = le->m_head->m_nextpkt;
979 le->m_head->m_nextpkt = NULL;
981 /* Now push out the old entry */
982 tcp_flush_out_entry(lc, le);
984 /* Re-parse new header, should not fail. */
985 pa = tcp_lro_parser(m, &le->outer, &le->inner, false);
987 ("tcp_push_and_replace: LRO parser failed on m=%p\n", m));
990 * Now to replace the data properly in the entry
991 * we have to reset the TCP header and
994 tcp_set_entry_to_mbuf(lc, le, m, pa->tcp);
996 /* Restore the next list */
997 m->m_nextpkt = msave;
1001 tcp_lro_mbuf_append_pkthdr(struct lro_entry *le, const struct mbuf *p)
1007 if (m->m_pkthdr.lro_nsegs == 1) {
1008 /* Compute relative checksum. */
1009 csum = p->m_pkthdr.lro_tcp_d_csum;
1011 /* Merge TCP data checksums. */
1012 csum = (uint32_t)m->m_pkthdr.lro_tcp_d_csum +
1013 (uint32_t)p->m_pkthdr.lro_tcp_d_csum;
1014 while (csum > 0xffff)
1015 csum = (csum >> 16) + (csum & 0xffff);
1018 /* Update various counters. */
1019 m->m_pkthdr.len += p->m_pkthdr.lro_tcp_d_len;
1020 m->m_pkthdr.lro_tcp_d_csum = csum;
1021 m->m_pkthdr.lro_tcp_d_len += p->m_pkthdr.lro_tcp_d_len;
1022 m->m_pkthdr.lro_nsegs += p->m_pkthdr.lro_nsegs;
1023 le->needs_merge = 1;
1027 tcp_lro_condense(struct lro_ctrl *lc, struct lro_entry *le)
1030 * Walk through the mbuf chain we
1031 * have on tap and compress/condense
1037 uint32_t tcp_data_len_total;
1038 uint32_t tcp_data_seg_total;
1039 uint16_t tcp_data_len;
1040 uint16_t tcp_opt_len;
1043 * First we must check the lead (m_head)
1044 * we must make sure that it is *not*
1045 * something that should be sent up
1046 * right away (sack etc).
1049 m = le->m_head->m_nextpkt;
1051 /* Just one left. */
1055 th = tcp_lro_get_th(m);
1056 tcp_opt_len = (th->th_off << 2);
1057 tcp_opt_len -= sizeof(*th);
1058 ts_ptr = (uint32_t *)(th + 1);
1060 if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1061 *ts_ptr != TCP_LRO_TS_OPTION)) {
1063 * Its not the timestamp. We can't
1064 * use this guy as the head.
1066 le->m_head->m_nextpkt = m->m_nextpkt;
1067 tcp_push_and_replace(lc, le, m);
1070 if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1072 * Make sure that previously seen segments/ACKs are delivered
1073 * before this segment, e.g. FIN.
1075 le->m_head->m_nextpkt = m->m_nextpkt;
1076 tcp_push_and_replace(lc, le, m);
1079 while((m = le->m_head->m_nextpkt) != NULL) {
1081 * condense m into le, first
1082 * pull m out of the list.
1084 le->m_head->m_nextpkt = m->m_nextpkt;
1085 m->m_nextpkt = NULL;
1087 tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
1088 th = tcp_lro_get_th(m);
1089 ts_ptr = (uint32_t *)(th + 1);
1090 tcp_opt_len = (th->th_off << 2);
1091 tcp_opt_len -= sizeof(*th);
1092 tcp_data_len_total = le->m_head->m_pkthdr.lro_tcp_d_len + tcp_data_len;
1093 tcp_data_seg_total = le->m_head->m_pkthdr.lro_nsegs + m->m_pkthdr.lro_nsegs;
1095 if (tcp_data_seg_total >= lc->lro_ackcnt_lim ||
1096 tcp_data_len_total >= lc->lro_length_lim) {
1097 /* Flush now if appending will result in overflow. */
1098 tcp_push_and_replace(lc, le, m);
1101 if (tcp_opt_len != 0 &&
1102 __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1103 *ts_ptr != TCP_LRO_TS_OPTION)) {
1105 * Maybe a sack in the new one? We need to
1106 * start all over after flushing the
1107 * current le. We will go up to the beginning
1108 * and flush it (calling the replace again possibly
1109 * or just returning).
1111 tcp_push_and_replace(lc, le, m);
1114 if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1115 tcp_push_and_replace(lc, le, m);
1118 if (tcp_opt_len != 0) {
1119 uint32_t tsval = ntohl(*(ts_ptr + 1));
1120 /* Make sure timestamp values are increasing. */
1121 if (TSTMP_GT(le->tsval, tsval)) {
1122 tcp_push_and_replace(lc, le, m);
1126 le->tsecr = *(ts_ptr + 2);
1128 /* Try to append the new segment. */
1129 if (__predict_false(ntohl(th->th_seq) != le->next_seq ||
1130 ((tcp_get_flags(th) & TH_ACK) !=
1131 (le->flags & TH_ACK)) ||
1132 (tcp_data_len == 0 &&
1133 le->ack_seq == th->th_ack &&
1134 le->window == th->th_win))) {
1135 /* Out of order packet, non-ACK + ACK or dup ACK. */
1136 tcp_push_and_replace(lc, le, m);
1139 if (tcp_data_len != 0 ||
1140 SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1141 le->next_seq += tcp_data_len;
1142 le->ack_seq = th->th_ack;
1143 le->window = th->th_win;
1144 le->needs_merge = 1;
1145 } else if (th->th_ack == le->ack_seq) {
1146 if (WIN_GT(th->th_win, le->window)) {
1147 le->window = th->th_win;
1148 le->needs_merge = 1;
1152 if (tcp_data_len == 0) {
1157 /* Merge TCP data checksum and length to head mbuf. */
1158 tcp_lro_mbuf_append_pkthdr(le, m);
1161 * Adjust the mbuf so that m_data points to the first byte of
1162 * the ULP payload. Adjust the mbuf to avoid complications and
1163 * append new segment to existing mbuf chain.
1165 m_adj(m, m->m_pkthdr.len - tcp_data_len);
1167 le->m_tail->m_next = m;
1168 le->m_tail = m_last(m);
1174 tcp_queue_pkts(struct inpcb *inp, struct tcpcb *tp, struct lro_entry *le)
1176 INP_WLOCK_ASSERT(inp);
1177 if (tp->t_in_pkt == NULL) {
1178 /* Nothing yet there */
1179 tp->t_in_pkt = le->m_head;
1180 tp->t_tail_pkt = le->m_last_mbuf;
1182 /* Already some there */
1183 tp->t_tail_pkt->m_nextpkt = le->m_head;
1184 tp->t_tail_pkt = le->m_last_mbuf;
1187 le->m_last_mbuf = NULL;
1190 static struct mbuf *
1191 tcp_lro_get_last_if_ackcmp(struct lro_ctrl *lc, struct lro_entry *le,
1192 struct inpcb *inp, int32_t *new_m, bool can_append_old_cmp)
1197 tp = intotcpcb(inp);
1198 if (__predict_false(tp == NULL))
1201 /* Look at the last mbuf if any in queue */
1202 if (can_append_old_cmp) {
1204 if (m != NULL && (m->m_flags & M_ACKCMP) != 0) {
1205 if (M_TRAILINGSPACE(m) >= sizeof(struct tcp_ackent)) {
1206 tcp_lro_log(tp, lc, le, NULL, 23, 0, 0, 0, 0);
1208 counter_u64_add(tcp_extra_mbuf, 1);
1211 /* Mark we ran out of space */
1212 inp->inp_flags2 |= INP_MBUF_L_ACKS;
1216 /* Decide mbuf size. */
1217 tcp_lro_log(tp, lc, le, NULL, 21, 0, 0, 0, 0);
1218 if (inp->inp_flags2 & INP_MBUF_L_ACKS)
1219 m = m_getcl(M_NOWAIT, MT_DATA, M_ACKCMP | M_PKTHDR);
1221 m = m_gethdr(M_NOWAIT, MT_DATA);
1223 if (__predict_false(m == NULL)) {
1224 counter_u64_add(tcp_would_have_but, 1);
1227 counter_u64_add(tcp_comp_total, 1);
1228 m->m_pkthdr.rcvif = lc->ifp;
1229 m->m_flags |= M_ACKCMP;
1234 static struct inpcb *
1235 tcp_lro_lookup(struct ifnet *ifp, struct lro_parser *pa)
1239 switch (pa->data.lro_type) {
1241 case LRO_TYPE_IPV6_TCP:
1242 inp = in6_pcblookup(&V_tcbinfo,
1243 &pa->data.s_addr.v6,
1245 &pa->data.d_addr.v6,
1252 case LRO_TYPE_IPV4_TCP:
1253 inp = in_pcblookup(&V_tcbinfo,
1270 tcp_lro_ack_valid(struct mbuf *m, struct tcphdr *th, uint32_t **ppts, bool *other_opts)
1273 * This function returns two bits of valuable information.
1274 * a) Is what is present capable of being ack-compressed,
1275 * we can ack-compress if there is no options or just
1276 * a timestamp option, and of course the th_flags must
1277 * be correct as well.
1278 * b) Our other options present such as SACK. This is
1279 * used to determine if we want to wakeup or not.
1283 switch (th->th_off << 2) {
1284 case (sizeof(*th) + TCPOLEN_TSTAMP_APPA):
1285 *ppts = (uint32_t *)(th + 1);
1286 /* Check if we have only one timestamp option. */
1287 if (**ppts == TCP_LRO_TS_OPTION)
1288 *other_opts = false;
1297 *other_opts = false;
1305 /* For ACKCMP we only accept ACK, PUSH, ECE and CWR. */
1306 if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH | TH_ECE | TH_CWR)) != 0)
1308 /* If it has data on it we cannot compress it */
1309 if (m->m_pkthdr.lro_tcp_d_len)
1312 /* ACK flag must be set. */
1313 if (!(tcp_get_flags(th) & TH_ACK))
1319 tcp_lro_flush_tcphpts(struct lro_ctrl *lc, struct lro_entry *le)
1323 struct mbuf **pp, *cmp, *mv_to;
1324 struct ifnet *lagg_ifp;
1325 bool bpf_req, lagg_bpf_req, should_wake, can_append_old_cmp;
1327 /* Check if packet doesn't belongs to our network interface. */
1328 if ((tcplro_stacks_wanting_mbufq == 0) ||
1329 (le->outer.data.vlan_id != 0) ||
1330 (le->inner.data.lro_type != LRO_TYPE_NONE))
1331 return (TCP_LRO_CANNOT);
1335 * Be proactive about unspecified IPv6 address in source. As
1336 * we use all-zero to indicate unbounded/unconnected pcb,
1337 * unspecified IPv6 address can be used to confuse us.
1339 * Note that packets with unspecified IPv6 destination is
1340 * already dropped in ip6_input.
1342 if (__predict_false(le->outer.data.lro_type == LRO_TYPE_IPV6_TCP &&
1343 IN6_IS_ADDR_UNSPECIFIED(&le->outer.data.s_addr.v6)))
1344 return (TCP_LRO_CANNOT);
1346 if (__predict_false(le->inner.data.lro_type == LRO_TYPE_IPV6_TCP &&
1347 IN6_IS_ADDR_UNSPECIFIED(&le->inner.data.s_addr.v6)))
1348 return (TCP_LRO_CANNOT);
1350 /* Lookup inp, if any. */
1351 inp = tcp_lro_lookup(lc->ifp,
1352 (le->inner.data.lro_type == LRO_TYPE_NONE) ? &le->outer : &le->inner);
1354 return (TCP_LRO_CANNOT);
1356 counter_u64_add(tcp_inp_lro_locks_taken, 1);
1358 /* Get TCP control structure. */
1359 tp = intotcpcb(inp);
1361 /* Check if the inp is dead, Jim. */
1362 if (tp->t_state == TCPS_TIME_WAIT) {
1364 return (TCP_LRO_CANNOT);
1366 if ((inp->inp_irq_cpu_set == 0) && (lc->lro_cpu_is_set == 1)) {
1367 inp->inp_irq_cpu = lc->lro_last_cpu;
1368 inp->inp_irq_cpu_set = 1;
1370 /* Check if the transport doesn't support the needed optimizations. */
1371 if ((inp->inp_flags2 & (INP_SUPPORTS_MBUFQ | INP_MBUF_ACKCMP)) == 0) {
1373 return (TCP_LRO_CANNOT);
1376 if (inp->inp_flags2 & INP_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(inp, 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, inp->inp_flags2, 0, 1);
1433 tcp_queue_pkts(inp, tp, le);
1437 counter_u64_add(tcp_inp_lro_wokeup_queue, 1);
1438 if ((*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0))
1443 return (0); /* Success. */
1448 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
1450 /* Only optimise if there are multiple packets waiting. */
1457 CURVNET_SET(lc->ifp->if_vnet);
1458 error = tcp_lro_flush_tcphpts(lc, le);
1462 tcp_lro_condense(lc, le);
1463 tcp_flush_out_entry(lc, le);
1468 bzero(le, sizeof(*le));
1469 LIST_INSERT_HEAD(&lc->lro_free, le, next);
1472 #ifdef HAVE_INLINE_FLSLL
1473 #define tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
1475 static inline uint64_t
1476 tcp_lro_msb_64(uint64_t x)
1484 return (x & ~(x >> 1));
1489 * The tcp_lro_sort() routine is comparable to qsort(), except it has
1490 * a worst case complexity limit of O(MIN(N,64)*N), where N is the
1491 * number of elements to sort and 64 is the number of sequence bits
1492 * available. The algorithm is bit-slicing the 64-bit sequence number,
1493 * sorting one bit at a time from the most significant bit until the
1494 * least significant one, skipping the constant bits. This is
1495 * typically called a radix sort.
1498 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
1500 struct lro_mbuf_sort temp;
1507 /* for small arrays insertion sort is faster */
1509 for (x = 1; x < size; x++) {
1511 for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
1512 parray[y] = parray[y - 1];
1518 /* compute sequence bits which are constant */
1521 for (x = 0; x != size; x++) {
1522 ones |= parray[x].seq;
1523 zeros |= ~parray[x].seq;
1526 /* compute bits which are not constant into "ones" */
1531 /* pick the most significant bit which is not constant */
1532 ones = tcp_lro_msb_64(ones);
1535 * Move entries having cleared sequence bits to the beginning
1538 for (x = y = 0; y != size; y++) {
1540 if (parray[y].seq & ones)
1544 parray[x] = parray[y];
1549 KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
1552 tcp_lro_sort(parray, x);
1561 tcp_lro_flush_all(struct lro_ctrl *lc)
1568 /* check if no mbufs to flush */
1569 if (lc->lro_mbuf_count == 0)
1571 if (lc->lro_cpu_is_set == 0) {
1572 if (lc->lro_last_cpu == curcpu) {
1573 lc->lro_cnt_of_same_cpu++;
1574 /* Have we reached the threshold to declare a cpu? */
1575 if (lc->lro_cnt_of_same_cpu > tcp_lro_cpu_set_thresh)
1576 lc->lro_cpu_is_set = 1;
1578 lc->lro_last_cpu = curcpu;
1579 lc->lro_cnt_of_same_cpu = 0;
1582 CURVNET_SET(lc->ifp->if_vnet);
1584 /* get current time */
1585 binuptime(&lc->lro_last_queue_time);
1587 /* sort all mbufs according to stream */
1588 tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
1590 /* input data into LRO engine, stream by stream */
1592 for (x = 0; x != lc->lro_mbuf_count; x++) {
1596 mb = lc->lro_mbuf_data[x].mb;
1598 /* get sequence number, masking away the packet index */
1599 nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
1601 /* check for new stream */
1605 /* flush active streams */
1606 tcp_lro_rx_done(lc);
1609 /* add packet to LRO engine */
1610 if (tcp_lro_rx_common(lc, mb, 0, false) != 0) {
1611 /* Flush anything we have acummulated */
1612 tcp_lro_flush_active(lc);
1613 /* input packet to network layer */
1614 (*lc->ifp->if_input)(lc->ifp, mb);
1621 /* flush active streams */
1622 tcp_lro_rx_done(lc);
1627 lc->lro_mbuf_count = 0;
1632 build_ack_entry(struct tcp_ackent *ae, struct tcphdr *th, struct mbuf *m,
1633 uint32_t *ts_ptr, uint16_t iptos)
1636 * Given a TCP ACK, summarize it down into the small TCP ACK
1639 ae->timestamp = m->m_pkthdr.rcv_tstmp;
1641 if (m->m_flags & M_TSTMP_LRO)
1642 ae->flags |= TSTMP_LRO;
1643 else if (m->m_flags & M_TSTMP)
1644 ae->flags |= TSTMP_HDWR;
1645 ae->seq = ntohl(th->th_seq);
1646 ae->ack = ntohl(th->th_ack);
1647 ae->flags |= tcp_get_flags(th);
1648 if (ts_ptr != NULL) {
1649 ae->ts_value = ntohl(ts_ptr[1]);
1650 ae->ts_echo = ntohl(ts_ptr[2]);
1651 ae->flags |= HAS_TSTMP;
1653 ae->win = ntohs(th->th_win);
1654 ae->codepoint = iptos;
1658 * Do BPF tap for either ACK_CMP packets or MBUF QUEUE type packets
1659 * and strip all, but the IPv4/IPv6 header.
1662 do_bpf_strip_and_compress(struct inpcb *inp, struct lro_ctrl *lc,
1663 struct lro_entry *le, struct mbuf **pp, struct mbuf **cmp, struct mbuf **mv_to,
1664 bool *should_wake, bool bpf_req, bool lagg_bpf_req, struct ifnet *lagg_ifp, bool can_append_old_cmp)
1669 struct ip6_hdr *ip6;
1674 struct tcp_ackent *ack_ent;
1677 bool other_opts, can_compress;
1683 /* Get current mbuf. */
1686 /* Let the BPF see the packet */
1687 if (__predict_false(bpf_req))
1688 ETHER_BPF_MTAP(lc->ifp, m);
1690 if (__predict_false(lagg_bpf_req))
1691 ETHER_BPF_MTAP(lagg_ifp, m);
1693 tcp_hdr_offset = m->m_pkthdr.lro_tcp_h_off;
1694 lro_type = le->inner.data.lro_type;
1697 lro_type = le->outer.data.lro_type;
1699 case LRO_TYPE_IPV4_TCP:
1700 tcp_hdr_offset -= sizeof(*le->outer.ip4);
1701 m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1703 case LRO_TYPE_IPV6_TCP:
1704 tcp_hdr_offset -= sizeof(*le->outer.ip6);
1705 m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1711 case LRO_TYPE_IPV4_TCP:
1712 tcp_hdr_offset -= sizeof(*le->outer.ip4);
1713 m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1715 case LRO_TYPE_IPV6_TCP:
1716 tcp_hdr_offset -= sizeof(*le->outer.ip6);
1717 m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1723 MPASS(tcp_hdr_offset >= 0);
1725 m_adj(m, tcp_hdr_offset);
1726 m->m_flags |= M_LRO_EHDRSTRP;
1727 m->m_flags &= ~M_ACKCMP;
1728 m->m_pkthdr.lro_tcp_h_off -= tcp_hdr_offset;
1730 th = tcp_lro_get_th(m);
1732 th->th_sum = 0; /* TCP checksum is valid. */
1734 /* Check if ACK can be compressed */
1735 can_compress = tcp_lro_ack_valid(m, th, &ts_ptr, &other_opts);
1737 /* Now lets look at the should wake states */
1738 if ((other_opts == true) &&
1739 ((inp->inp_flags2 & INP_DONT_SACK_QUEUE) == 0)) {
1741 * If there are other options (SACK?) and the
1742 * tcp endpoint has not expressly told us it does
1743 * not care about SACKS, then we should wake up.
1745 *should_wake = true;
1747 /* Is the ack compressable? */
1748 if (can_compress == false)
1750 /* Does the TCP endpoint support ACK compression? */
1751 if ((inp->inp_flags2 & INP_MBUF_ACKCMP) == 0)
1754 /* Lets get the TOS/traffic class field */
1755 l3.ptr = mtod(m, void *);
1757 case LRO_TYPE_IPV4_TCP:
1758 iptos = l3.ip4->ip_tos;
1760 case LRO_TYPE_IPV6_TCP:
1761 iptos = IPV6_TRAFFIC_CLASS(l3.ip6);
1764 iptos = 0; /* Keep compiler happy. */
1767 /* Now lets get space if we don't have some already */
1770 nm = tcp_lro_get_last_if_ackcmp(lc, le, inp, &n_mbuf, can_append_old_cmp);
1771 if (__predict_false(nm == NULL))
1776 * Link in the new cmp ack to our in-order place,
1777 * first set our cmp ack's next to where we are.
1782 * Set it up so mv_to is advanced to our
1783 * compressed ack. This way the caller can
1784 * advance pp to the right place.
1788 * Advance it here locally as well.
1790 pp = &nm->m_nextpkt;
1793 /* We have one already we are working on */
1795 if (M_TRAILINGSPACE(nm) < sizeof(struct tcp_ackent)) {
1796 /* We ran out of space */
1797 inp->inp_flags2 |= INP_MBUF_L_ACKS;
1801 MPASS(M_TRAILINGSPACE(nm) >= sizeof(struct tcp_ackent));
1802 counter_u64_add(tcp_inp_lro_compressed, 1);
1804 /* We can add in to the one on the tail */
1805 ack_ent = mtod(nm, struct tcp_ackent *);
1806 idx = (nm->m_len / sizeof(struct tcp_ackent));
1807 build_ack_entry(&ack_ent[idx], th, m, ts_ptr, iptos);
1809 /* Bump the size of both pkt-hdr and len */
1810 nm->m_len += sizeof(struct tcp_ackent);
1811 nm->m_pkthdr.len += sizeof(struct tcp_ackent);
1813 /* Advance to next mbuf before freeing. */
1815 m->m_nextpkt = NULL;
1819 counter_u64_add(tcp_uncomp_total, 1);
1825 static struct lro_head *
1826 tcp_lro_rx_get_bucket(struct lro_ctrl *lc, struct mbuf *m, struct lro_parser *parser)
1830 if (M_HASHTYPE_ISHASH(m)) {
1831 hash = m->m_pkthdr.flowid;
1833 for (unsigned i = hash = 0; i != LRO_RAW_ADDRESS_MAX; i++)
1834 hash += parser->data.raw[i];
1836 return (&lc->lro_hash[hash % lc->lro_hashsz]);
1840 tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash)
1842 struct lro_parser pi; /* inner address data */
1843 struct lro_parser po; /* outer address data */
1844 struct lro_parser *pa; /* current parser for TCP stream */
1845 struct lro_entry *le;
1846 struct lro_head *bucket;
1851 uint16_t tcp_data_sum;
1854 /* Quickly decide if packet cannot be LRO'ed */
1855 if (__predict_false(V_ipforwarding != 0))
1856 return (TCP_LRO_CANNOT);
1859 /* Quickly decide if packet cannot be LRO'ed */
1860 if (__predict_false(V_ip6_forwarding != 0))
1861 return (TCP_LRO_CANNOT);
1863 if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1864 ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1865 (m->m_pkthdr.csum_data != 0xffff)) {
1867 * The checksum either did not have hardware offload
1868 * or it was a bad checksum. We can't LRO such
1871 counter_u64_add(tcp_bad_csums, 1);
1872 return (TCP_LRO_CANNOT);
1874 /* We expect a contiguous header [eh, ip, tcp]. */
1875 pa = tcp_lro_parser(m, &po, &pi, true);
1876 if (__predict_false(pa == NULL))
1877 return (TCP_LRO_NOT_SUPPORTED);
1879 /* We don't expect any padding. */
1880 error = tcp_lro_trim_mbuf_chain(m, pa);
1881 if (__predict_false(error != 0))
1885 switch (pa->data.lro_type) {
1886 case LRO_TYPE_IPV4_TCP:
1887 error = tcp_lro_rx_ipv4(lc, m, pa->ip4);
1888 if (__predict_false(error != 0))
1895 /* If no hardware or arrival stamp on the packet add timestamp */
1896 if ((m->m_flags & (M_TSTMP_LRO | M_TSTMP)) == 0) {
1897 m->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
1898 m->m_flags |= M_TSTMP_LRO;
1901 /* Get pointer to TCP header. */
1904 /* Don't process SYN packets. */
1905 if (__predict_false(tcp_get_flags(th) & TH_SYN))
1906 return (TCP_LRO_CANNOT);
1908 /* Get total TCP header length and compute payload length. */
1909 tcp_opt_len = (th->th_off << 2);
1910 tcp_data_len = m->m_pkthdr.len - ((uint8_t *)th -
1911 (uint8_t *)m->m_data) - tcp_opt_len;
1912 tcp_opt_len -= sizeof(*th);
1914 /* Don't process invalid TCP headers. */
1915 if (__predict_false(tcp_opt_len < 0 || tcp_data_len < 0))
1916 return (TCP_LRO_CANNOT);
1918 /* Compute TCP data only checksum. */
1919 if (tcp_data_len == 0)
1920 tcp_data_sum = 0; /* no data, no checksum */
1921 else if (__predict_false(csum != 0))
1922 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~csum);
1924 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~th->th_sum);
1926 /* Save TCP info in mbuf. */
1927 m->m_nextpkt = NULL;
1928 m->m_pkthdr.rcvif = lc->ifp;
1929 m->m_pkthdr.lro_tcp_d_csum = tcp_data_sum;
1930 m->m_pkthdr.lro_tcp_d_len = tcp_data_len;
1931 m->m_pkthdr.lro_tcp_h_off = ((uint8_t *)th - (uint8_t *)m->m_data);
1932 m->m_pkthdr.lro_nsegs = 1;
1934 /* Get hash bucket. */
1936 bucket = &lc->lro_hash[0];
1938 bucket = tcp_lro_rx_get_bucket(lc, m, pa);
1941 /* Try to find a matching previous segment. */
1942 LIST_FOREACH(le, bucket, hash_next) {
1943 /* Compare addresses and ports. */
1944 if (lro_address_compare(&po.data, &le->outer.data) == false ||
1945 lro_address_compare(&pi.data, &le->inner.data) == false)
1948 /* Check if no data and old ACK. */
1949 if (tcp_data_len == 0 &&
1950 SEQ_LT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1955 /* Mark "m" in the last spot. */
1956 le->m_last_mbuf->m_nextpkt = m;
1957 /* Now set the tail to "m". */
1958 le->m_last_mbuf = m;
1962 /* Try to find an empty slot. */
1963 if (LIST_EMPTY(&lc->lro_free))
1964 return (TCP_LRO_NO_ENTRIES);
1966 /* Start a new segment chain. */
1967 le = LIST_FIRST(&lc->lro_free);
1968 LIST_REMOVE(le, next);
1969 tcp_lro_active_insert(lc, bucket, le);
1971 /* Make sure the headers are set. */
1975 /* Store time this entry was allocated. */
1976 le->alloc_time = lc->lro_last_queue_time;
1978 tcp_set_entry_to_mbuf(lc, le, m, th);
1980 /* Now set the tail to "m". */
1981 le->m_last_mbuf = m;
1987 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
1991 if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1992 ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1993 (m->m_pkthdr.csum_data != 0xffff)) {
1995 * The checksum either did not have hardware offload
1996 * or it was a bad checksum. We can't LRO such
1999 counter_u64_add(tcp_bad_csums, 1);
2000 return (TCP_LRO_CANNOT);
2002 /* get current time */
2003 binuptime(&lc->lro_last_queue_time);
2004 CURVNET_SET(lc->ifp->if_vnet);
2005 error = tcp_lro_rx_common(lc, m, csum, true);
2006 if (__predict_false(error != 0)) {
2008 * Flush anything we have acummulated
2009 * ahead of this packet that can't
2010 * be LRO'd. This preserves order.
2012 tcp_lro_flush_active(lc);
2020 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
2024 if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
2025 lc->lro_mbuf_max == 0)) {
2031 /* check if packet is not LRO capable */
2032 if (__predict_false((lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
2033 /* input packet to network layer */
2034 (*lc->ifp->if_input) (lc->ifp, mb);
2038 /* If no hardware or arrival stamp on the packet add timestamp */
2039 if ((tcplro_stacks_wanting_mbufq > 0) &&
2040 (tcp_less_accurate_lro_ts == 0) &&
2041 ((mb->m_flags & M_TSTMP) == 0)) {
2042 /* Add in an LRO time since no hardware */
2043 binuptime(&lc->lro_last_queue_time);
2044 mb->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
2045 mb->m_flags |= M_TSTMP_LRO;
2048 /* create sequence number */
2049 lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
2050 (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
2051 (((uint64_t)mb->m_pkthdr.flowid) << 24) |
2052 ((uint64_t)lc->lro_mbuf_count);
2055 lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb;
2057 /* flush if array is full */
2058 if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max))
2059 tcp_lro_flush_all(lc);