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>
57 #include <netinet/in_systm.h>
58 #include <netinet/in.h>
59 #include <netinet/ip6.h>
60 #include <netinet/ip.h>
61 #include <netinet/ip_var.h>
62 #include <netinet/in_pcb.h>
63 #include <netinet6/in6_pcb.h>
64 #include <netinet/tcp.h>
65 #include <netinet/tcp_seq.h>
66 #include <netinet/tcp_lro.h>
67 #include <netinet/tcp_var.h>
68 #include <netinet/tcpip.h>
69 #include <netinet/tcp_hpts.h>
70 #include <netinet/tcp_log_buf.h>
71 #include <netinet/udp.h>
72 #include <netinet6/ip6_var.h>
74 #include <machine/in_cksum.h>
76 static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures");
78 #define TCP_LRO_TS_OPTION \
79 ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
80 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
82 static void tcp_lro_rx_done(struct lro_ctrl *lc);
83 static int tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m,
84 uint32_t csum, bool use_hash);
87 static bool do_bpf_strip_and_compress(struct inpcb *, struct lro_ctrl *,
88 struct lro_entry *, struct mbuf **, struct mbuf **, struct mbuf **, bool *, bool);
92 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
95 static long tcplro_stacks_wanting_mbufq;
96 counter_u64_t tcp_inp_lro_direct_queue;
97 counter_u64_t tcp_inp_lro_wokeup_queue;
98 counter_u64_t tcp_inp_lro_compressed;
99 counter_u64_t tcp_inp_lro_locks_taken;
100 counter_u64_t tcp_extra_mbuf;
101 counter_u64_t tcp_would_have_but;
102 counter_u64_t tcp_comp_total;
103 counter_u64_t tcp_uncomp_total;
105 static unsigned tcp_lro_entries = TCP_LRO_ENTRIES;
106 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries,
107 CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_entries, 0,
108 "default number of LRO entries");
110 static uint32_t tcp_lro_cpu_set_thresh = TCP_LRO_CPU_DECLARATION_THRESH;
111 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_cpu_threshold,
112 CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_cpu_set_thresh, 0,
113 "Number of interrupts in a row on the same CPU that will make us declare an 'affinity' cpu?");
115 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, fullqueue, CTLFLAG_RD,
116 &tcp_inp_lro_direct_queue, "Number of lro's fully queued to transport");
117 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, wokeup, CTLFLAG_RD,
118 &tcp_inp_lro_wokeup_queue, "Number of lro's where we woke up transport via hpts");
119 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, compressed, CTLFLAG_RD,
120 &tcp_inp_lro_compressed, "Number of lro's compressed and sent to transport");
121 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lockcnt, CTLFLAG_RD,
122 &tcp_inp_lro_locks_taken, "Number of lro's inp_wlocks taken");
123 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, extra_mbuf, CTLFLAG_RD,
124 &tcp_extra_mbuf, "Number of times we had an extra compressed ack dropped into the tp");
125 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, would_have_but, CTLFLAG_RD,
126 &tcp_would_have_but, "Number of times we would have had an extra compressed, but mget failed");
127 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, with_m_ackcmp, CTLFLAG_RD,
128 &tcp_comp_total, "Number of mbufs queued with M_ACKCMP flags set");
129 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, without_m_ackcmp, CTLFLAG_RD,
130 &tcp_uncomp_total, "Number of mbufs queued without M_ACKCMP");
133 tcp_lro_reg_mbufq(void)
135 atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, 1);
139 tcp_lro_dereg_mbufq(void)
141 atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, -1);
145 tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket,
146 struct lro_entry *le)
149 LIST_INSERT_HEAD(&lc->lro_active, le, next);
150 LIST_INSERT_HEAD(bucket, le, hash_next);
154 tcp_lro_active_remove(struct lro_entry *le)
157 LIST_REMOVE(le, next); /* active list */
158 LIST_REMOVE(le, hash_next); /* hash bucket */
162 tcp_lro_init(struct lro_ctrl *lc)
164 return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0));
168 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
169 unsigned lro_entries, unsigned lro_mbufs)
171 struct lro_entry *le;
173 unsigned i, elements;
175 lc->lro_bad_csum = 0;
178 lc->lro_mbuf_count = 0;
179 lc->lro_mbuf_max = lro_mbufs;
180 lc->lro_cnt = lro_entries;
181 lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
182 lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
184 LIST_INIT(&lc->lro_free);
185 LIST_INIT(&lc->lro_active);
187 /* create hash table to accelerate entry lookup */
188 if (lro_entries > lro_mbufs)
189 elements = lro_entries;
191 elements = lro_mbufs;
192 lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz,
194 if (lc->lro_hash == NULL) {
195 memset(lc, 0, sizeof(*lc));
199 /* compute size to allocate */
200 size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
201 (lro_entries * sizeof(*le));
202 lc->lro_mbuf_data = (struct lro_mbuf_sort *)
203 malloc(size, M_LRO, M_NOWAIT | M_ZERO);
205 /* check for out of memory */
206 if (lc->lro_mbuf_data == NULL) {
207 free(lc->lro_hash, M_LRO);
208 memset(lc, 0, sizeof(*lc));
211 /* compute offset for LRO entries */
212 le = (struct lro_entry *)
213 (lc->lro_mbuf_data + lro_mbufs);
215 /* setup linked list */
216 for (i = 0; i != lro_entries; i++)
217 LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
222 struct vxlan_header {
228 tcp_lro_low_level_parser(void *ptr, struct lro_parser *parser, bool update_data, bool is_vxlan)
230 const struct ether_vlan_header *eh;
235 memset(parser, 0, sizeof(*parser));
240 const struct vxlan_header *vxh;
242 ptr = (uint8_t *)ptr + sizeof(*vxh);
244 parser->data.vxlan_vni =
245 vxh->vxlh_vni & htonl(0xffffff00);
250 if (__predict_false(eh->evl_encap_proto == htons(ETHERTYPE_VLAN))) {
251 eth_type = eh->evl_proto;
253 /* strip priority and keep VLAN ID only */
254 parser->data.vlan_id = eh->evl_tag & htons(EVL_VLID_MASK);
256 /* advance to next header */
257 ptr = (uint8_t *)ptr + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
259 eth_type = eh->evl_encap_proto;
260 /* advance to next header */
261 ptr = (uint8_t *)ptr + ETHER_HDR_LEN;
266 case htons(ETHERTYPE_IP):
268 /* Ensure there are no IPv4 options. */
269 if ((parser->ip4->ip_hl << 2) != sizeof (*parser->ip4))
271 /* .. and the packet is not fragmented. */
272 if (parser->ip4->ip_off & htons(IP_MF|IP_OFFMASK))
274 ptr = (uint8_t *)ptr + (parser->ip4->ip_hl << 2);
276 parser->data.s_addr.v4 = parser->ip4->ip_src;
277 parser->data.d_addr.v4 = parser->ip4->ip_dst;
279 switch (parser->ip4->ip_p) {
283 parser->data.lro_type = LRO_TYPE_IPV4_UDP;
284 parser->data.s_port = parser->udp->uh_sport;
285 parser->data.d_port = parser->udp->uh_dport;
287 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_UDP);
289 ptr = ((uint8_t *)ptr + sizeof(*parser->udp));
290 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
295 parser->data.lro_type = LRO_TYPE_IPV4_TCP;
296 parser->data.s_port = parser->tcp->th_sport;
297 parser->data.d_port = parser->tcp->th_dport;
299 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_TCP);
301 ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
302 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
310 case htons(ETHERTYPE_IPV6):
312 ptr = (uint8_t *)ptr + sizeof(*parser->ip6);
314 parser->data.s_addr.v6 = parser->ip6->ip6_src;
315 parser->data.d_addr.v6 = parser->ip6->ip6_dst;
317 switch (parser->ip6->ip6_nxt) {
321 parser->data.lro_type = LRO_TYPE_IPV6_UDP;
322 parser->data.s_port = parser->udp->uh_sport;
323 parser->data.d_port = parser->udp->uh_dport;
325 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_UDP);
327 ptr = (uint8_t *)ptr + sizeof(*parser->udp);
328 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
333 parser->data.lro_type = LRO_TYPE_IPV6_TCP;
334 parser->data.s_port = parser->tcp->th_sport;
335 parser->data.d_port = parser->tcp->th_dport;
337 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_TCP);
339 ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
340 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
350 /* Invalid packet - cannot parse */
354 static const int vxlan_csum = CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
355 CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID;
357 static inline struct lro_parser *
358 tcp_lro_parser(struct mbuf *m, struct lro_parser *po, struct lro_parser *pi, bool update_data)
362 /* Try to parse outer headers first. */
363 data_ptr = tcp_lro_low_level_parser(m->m_data, po, update_data, false);
364 if (data_ptr == NULL || po->total_hdr_len > m->m_len)
368 /* Store VLAN ID, if any. */
369 if (__predict_false(m->m_flags & M_VLANTAG)) {
371 htons(m->m_pkthdr.ether_vtag) & htons(EVL_VLID_MASK);
373 /* Store decrypted flag, if any. */
374 if (__predict_false((m->m_pkthdr.csum_flags &
375 CSUM_TLS_MASK) == CSUM_TLS_DECRYPTED))
376 po->data.lro_flags |= LRO_FLAG_DECRYPTED;
379 switch (po->data.lro_type) {
380 case LRO_TYPE_IPV4_UDP:
381 case LRO_TYPE_IPV6_UDP:
382 /* Check for VXLAN headers. */
383 if ((m->m_pkthdr.csum_flags & vxlan_csum) != vxlan_csum)
386 /* Try to parse inner headers. */
387 data_ptr = tcp_lro_low_level_parser(data_ptr, pi, update_data, true);
388 if (data_ptr == NULL || pi->total_hdr_len > m->m_len)
391 /* Verify supported header types. */
392 switch (pi->data.lro_type) {
393 case LRO_TYPE_IPV4_TCP:
394 case LRO_TYPE_IPV6_TCP:
400 case LRO_TYPE_IPV4_TCP:
401 case LRO_TYPE_IPV6_TCP:
403 memset(pi, 0, sizeof(*pi));
412 tcp_lro_trim_mbuf_chain(struct mbuf *m, const struct lro_parser *po)
416 switch (po->data.lro_type) {
418 case LRO_TYPE_IPV4_TCP:
419 len = ((uint8_t *)po->ip4 - (uint8_t *)m->m_data) +
420 ntohs(po->ip4->ip_len);
424 case LRO_TYPE_IPV6_TCP:
425 len = ((uint8_t *)po->ip6 - (uint8_t *)m->m_data) +
426 ntohs(po->ip6->ip6_plen) + sizeof(*po->ip6);
430 return (TCP_LRO_CANNOT);
434 * If the frame is padded beyond the end of the IP packet,
435 * then trim the extra bytes off:
437 if (__predict_true(m->m_pkthdr.len == len)) {
439 } else if (m->m_pkthdr.len > len) {
440 m_adj(m, len - m->m_pkthdr.len);
443 return (TCP_LRO_CANNOT);
446 static struct tcphdr *
447 tcp_lro_get_th(struct mbuf *m)
449 return ((struct tcphdr *)((uint8_t *)m->m_data + m->m_pkthdr.lro_tcp_h_off));
453 lro_free_mbuf_chain(struct mbuf *m)
466 tcp_lro_free(struct lro_ctrl *lc)
468 struct lro_entry *le;
471 /* reset LRO free list */
472 LIST_INIT(&lc->lro_free);
474 /* free active mbufs, if any */
475 while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
476 tcp_lro_active_remove(le);
477 lro_free_mbuf_chain(le->m_head);
480 /* free hash table */
481 free(lc->lro_hash, M_LRO);
485 /* free mbuf array, if any */
486 for (x = 0; x != lc->lro_mbuf_count; x++)
487 m_freem(lc->lro_mbuf_data[x].mb);
488 lc->lro_mbuf_count = 0;
490 /* free allocated memory, if any */
491 free(lc->lro_mbuf_data, M_LRO);
492 lc->lro_mbuf_data = NULL;
496 tcp_lro_rx_csum_tcphdr(const struct tcphdr *th)
502 csum = -th->th_sum; /* exclude checksum field */
504 ptr = (const uint16_t *)th;
511 while (csum > 0xffff)
512 csum = (csum >> 16) + (csum & 0xffff);
518 tcp_lro_rx_csum_data(const struct lro_parser *pa, uint16_t tcp_csum)
525 switch (pa->data.lro_type) {
527 case LRO_TYPE_IPV6_TCP:
528 /* Compute full pseudo IPv6 header checksum. */
529 cs = in6_cksum_pseudo(pa->ip6, ntohs(pa->ip6->ip6_plen), pa->ip6->ip6_nxt, 0);
533 case LRO_TYPE_IPV4_TCP:
534 /* Compute full pseudo IPv4 header checsum. */
535 cs = in_addword(ntohs(pa->ip4->ip_len) - sizeof(*pa->ip4), IPPROTO_TCP);
536 cs = in_pseudo(pa->ip4->ip_src.s_addr, pa->ip4->ip_dst.s_addr, htons(cs));
540 cs = 0; /* Keep compiler happy. */
544 /* Complement checksum. */
548 /* Remove TCP header checksum. */
549 cs = ~tcp_lro_rx_csum_tcphdr(pa->tcp);
552 /* Compute checksum remainder. */
554 c = (c >> 16) + (c & 0xffff);
560 tcp_lro_rx_done(struct lro_ctrl *lc)
562 struct lro_entry *le;
564 while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
565 tcp_lro_active_remove(le);
566 tcp_lro_flush(lc, le);
571 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
573 struct lro_entry *le, *le_tmp;
578 if (LIST_EMPTY(&lc->lro_active))
581 /* get timeout time and current time in ns */
583 now = bintime2ns(&bt);
584 tov = ((timeout->tv_sec * 1000000000) + (timeout->tv_usec * 1000));
585 LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
586 if (now >= (bintime2ns(&le->alloc_time) + tov)) {
587 tcp_lro_active_remove(le);
588 tcp_lro_flush(lc, le);
595 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4)
599 /* Legacy IP has a header checksum that needs to be correct. */
600 if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) {
601 if (__predict_false((m->m_pkthdr.csum_flags & CSUM_IP_VALID) == 0)) {
603 return (TCP_LRO_CANNOT);
606 csum = in_cksum_hdr(ip4);
607 if (__predict_false(csum != 0)) {
609 return (TCP_LRO_CANNOT);
618 tcp_lro_log(struct tcpcb *tp, const struct lro_ctrl *lc,
619 const struct lro_entry *le, const struct mbuf *m,
620 int frm, int32_t tcp_data_len, uint32_t th_seq,
621 uint32_t th_ack, uint16_t th_win)
623 if (tp->t_logstate != TCP_LOG_STATE_OFF) {
624 union tcp_log_stackspecific log;
625 struct timeval tv, btv;
628 cts = tcp_get_usecs(&tv);
629 memset(&log, 0, sizeof(union tcp_log_stackspecific));
630 log.u_bbr.flex8 = frm;
631 log.u_bbr.flex1 = tcp_data_len;
633 log.u_bbr.flex2 = m->m_pkthdr.len;
636 log.u_bbr.flex3 = le->m_head->m_pkthdr.lro_nsegs;
637 log.u_bbr.flex4 = le->m_head->m_pkthdr.lro_tcp_d_len;
639 log.u_bbr.flex5 = le->m_head->m_pkthdr.len;
640 log.u_bbr.delRate = le->m_head->m_flags;
641 log.u_bbr.rttProp = le->m_head->m_pkthdr.rcv_tstmp;
643 log.u_bbr.inflight = th_seq;
644 log.u_bbr.delivered = th_ack;
645 log.u_bbr.timeStamp = cts;
646 log.u_bbr.epoch = le->next_seq;
647 log.u_bbr.lt_epoch = le->ack_seq;
648 log.u_bbr.pacing_gain = th_win;
649 log.u_bbr.cwnd_gain = le->window;
650 log.u_bbr.lost = curcpu;
651 log.u_bbr.cur_del_rate = (uintptr_t)m;
652 log.u_bbr.bw_inuse = (uintptr_t)le->m_head;
653 bintime2timeval(&lc->lro_last_queue_time, &btv);
654 log.u_bbr.flex6 = tcp_tv_to_usectick(&btv);
655 log.u_bbr.flex7 = le->compressed;
656 log.u_bbr.pacing_gain = le->uncompressed;
657 if (in_epoch(net_epoch_preempt))
658 log.u_bbr.inhpts = 1;
660 log.u_bbr.inhpts = 0;
661 TCP_LOG_EVENTP(tp, NULL,
662 &tp->t_inpcb->inp_socket->so_rcv,
663 &tp->t_inpcb->inp_socket->so_snd,
665 0, &log, false, &tv);
671 tcp_lro_assign_and_checksum_16(uint16_t *ptr, uint16_t value, uint16_t *psum)
675 csum = 0xffff - *ptr + value;
676 while (csum > 0xffff)
677 csum = (csum >> 16) + (csum & 0xffff);
683 tcp_lro_update_checksum(const struct lro_parser *pa, const struct lro_entry *le,
684 uint16_t payload_len, uint16_t delta_sum)
688 uint16_t temp[5] = {};
690 switch (pa->data.lro_type) {
691 case LRO_TYPE_IPV4_TCP:
692 /* Compute new IPv4 length. */
693 tlen = (pa->ip4->ip_hl << 2) + (pa->tcp->th_off << 2) + payload_len;
694 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
696 /* Subtract delta from current IPv4 checksum. */
697 csum = pa->ip4->ip_sum + 0xffff - temp[0];
698 while (csum > 0xffff)
699 csum = (csum >> 16) + (csum & 0xffff);
700 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
701 goto update_tcp_header;
703 case LRO_TYPE_IPV6_TCP:
704 /* Compute new IPv6 length. */
705 tlen = (pa->tcp->th_off << 2) + payload_len;
706 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
707 goto update_tcp_header;
709 case LRO_TYPE_IPV4_UDP:
710 /* Compute new IPv4 length. */
711 tlen = (pa->ip4->ip_hl << 2) + sizeof(*pa->udp) + payload_len;
712 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
714 /* Subtract delta from current IPv4 checksum. */
715 csum = pa->ip4->ip_sum + 0xffff - temp[0];
716 while (csum > 0xffff)
717 csum = (csum >> 16) + (csum & 0xffff);
718 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
719 goto update_udp_header;
721 case LRO_TYPE_IPV6_UDP:
722 /* Compute new IPv6 length. */
723 tlen = sizeof(*pa->udp) + payload_len;
724 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
725 goto update_udp_header;
732 /* Compute current TCP header checksum. */
733 temp[2] = tcp_lro_rx_csum_tcphdr(pa->tcp);
735 /* Incorporate the latest ACK into the TCP header. */
736 pa->tcp->th_ack = le->ack_seq;
737 pa->tcp->th_win = le->window;
739 /* Incorporate latest timestamp into the TCP header. */
740 if (le->timestamp != 0) {
743 ts_ptr = (uint32_t *)(pa->tcp + 1);
744 ts_ptr[1] = htonl(le->tsval);
745 ts_ptr[2] = le->tsecr;
748 /* Compute new TCP header checksum. */
749 temp[3] = tcp_lro_rx_csum_tcphdr(pa->tcp);
751 /* Compute new TCP checksum. */
752 csum = pa->tcp->th_sum + 0xffff - delta_sum +
753 0xffff - temp[0] + 0xffff - temp[3] + temp[2];
754 while (csum > 0xffff)
755 csum = (csum >> 16) + (csum & 0xffff);
757 /* Assign new TCP checksum. */
758 tcp_lro_assign_and_checksum_16(&pa->tcp->th_sum, csum, &temp[4]);
760 /* Compute all modififications affecting next checksum. */
761 csum = temp[0] + temp[1] + 0xffff - temp[2] +
762 temp[3] + temp[4] + delta_sum;
763 while (csum > 0xffff)
764 csum = (csum >> 16) + (csum & 0xffff);
766 /* Return delta checksum to next stage, if any. */
770 tlen = sizeof(*pa->udp) + payload_len;
771 /* Assign new UDP length and compute checksum delta. */
772 tcp_lro_assign_and_checksum_16(&pa->udp->uh_ulen, htons(tlen), &temp[2]);
774 /* Check if there is a UDP checksum. */
775 if (__predict_false(pa->udp->uh_sum != 0)) {
776 /* Compute new UDP checksum. */
777 csum = pa->udp->uh_sum + 0xffff - delta_sum +
778 0xffff - temp[0] + 0xffff - temp[2];
779 while (csum > 0xffff)
780 csum = (csum >> 16) + (csum & 0xffff);
781 /* Assign new UDP checksum. */
782 tcp_lro_assign_and_checksum_16(&pa->udp->uh_sum, csum, &temp[3]);
785 /* Compute all modififications affecting next checksum. */
786 csum = temp[0] + temp[1] + temp[2] + temp[3] + delta_sum;
787 while (csum > 0xffff)
788 csum = (csum >> 16) + (csum & 0xffff);
790 /* Return delta checksum to next stage, if any. */
795 tcp_flush_out_entry(struct lro_ctrl *lc, struct lro_entry *le)
797 /* Check if we need to recompute any checksums. */
798 if (le->needs_merge) {
801 switch (le->inner.data.lro_type) {
802 case LRO_TYPE_IPV4_TCP:
803 csum = tcp_lro_update_checksum(&le->inner, le,
804 le->m_head->m_pkthdr.lro_tcp_d_len,
805 le->m_head->m_pkthdr.lro_tcp_d_csum);
806 csum = tcp_lro_update_checksum(&le->outer, NULL,
807 le->m_head->m_pkthdr.lro_tcp_d_len +
808 le->inner.total_hdr_len, csum);
809 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
810 CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
811 le->m_head->m_pkthdr.csum_data = 0xffff;
812 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
813 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
815 case LRO_TYPE_IPV6_TCP:
816 csum = tcp_lro_update_checksum(&le->inner, le,
817 le->m_head->m_pkthdr.lro_tcp_d_len,
818 le->m_head->m_pkthdr.lro_tcp_d_csum);
819 csum = tcp_lro_update_checksum(&le->outer, NULL,
820 le->m_head->m_pkthdr.lro_tcp_d_len +
821 le->inner.total_hdr_len, csum);
822 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
824 le->m_head->m_pkthdr.csum_data = 0xffff;
825 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
826 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
829 switch (le->outer.data.lro_type) {
830 case LRO_TYPE_IPV4_TCP:
831 csum = tcp_lro_update_checksum(&le->outer, le,
832 le->m_head->m_pkthdr.lro_tcp_d_len,
833 le->m_head->m_pkthdr.lro_tcp_d_csum);
834 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
835 CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
836 le->m_head->m_pkthdr.csum_data = 0xffff;
837 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
838 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
840 case LRO_TYPE_IPV6_TCP:
841 csum = tcp_lro_update_checksum(&le->outer, le,
842 le->m_head->m_pkthdr.lro_tcp_d_len,
843 le->m_head->m_pkthdr.lro_tcp_d_csum);
844 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
846 le->m_head->m_pkthdr.csum_data = 0xffff;
847 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
848 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
860 * Break any chain, this is not set to NULL on the singleton
861 * case m_nextpkt points to m_head. Other case set them
862 * m_nextpkt to NULL in push_and_replace.
864 le->m_head->m_nextpkt = NULL;
865 lc->lro_queued += le->m_head->m_pkthdr.lro_nsegs;
866 (*lc->ifp->if_input)(lc->ifp, le->m_head);
870 tcp_set_entry_to_mbuf(struct lro_ctrl *lc, struct lro_entry *le,
871 struct mbuf *m, struct tcphdr *th)
874 uint16_t tcp_data_len;
875 uint16_t tcp_opt_len;
877 ts_ptr = (uint32_t *)(th + 1);
878 tcp_opt_len = (th->th_off << 2);
879 tcp_opt_len -= sizeof(*th);
881 /* Check if there is a timestamp option. */
882 if (tcp_opt_len == 0 ||
883 __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
884 *ts_ptr != TCP_LRO_TS_OPTION)) {
885 /* We failed to find the timestamp option. */
889 le->tsval = ntohl(*(ts_ptr + 1));
890 le->tsecr = *(ts_ptr + 2);
893 tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
895 /* Pull out TCP sequence numbers and window size. */
896 le->next_seq = ntohl(th->th_seq) + tcp_data_len;
897 le->ack_seq = th->th_ack;
898 le->window = th->th_win;
899 le->flags = th->th_flags;
902 /* Setup new data pointers. */
904 le->m_tail = m_last(m);
908 tcp_push_and_replace(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m)
910 struct lro_parser *pa;
913 * Push up the stack of the current entry
914 * and replace it with "m".
918 /* Grab off the next and save it */
919 msave = le->m_head->m_nextpkt;
920 le->m_head->m_nextpkt = NULL;
922 /* Now push out the old entry */
923 tcp_flush_out_entry(lc, le);
925 /* Re-parse new header, should not fail. */
926 pa = tcp_lro_parser(m, &le->outer, &le->inner, false);
928 ("tcp_push_and_replace: LRO parser failed on m=%p\n", m));
931 * Now to replace the data properly in the entry
932 * we have to reset the TCP header and
935 tcp_set_entry_to_mbuf(lc, le, m, pa->tcp);
937 /* Restore the next list */
938 m->m_nextpkt = msave;
942 tcp_lro_mbuf_append_pkthdr(struct lro_entry *le, const struct mbuf *p)
948 if (m->m_pkthdr.lro_nsegs == 1) {
949 /* Compute relative checksum. */
950 csum = p->m_pkthdr.lro_tcp_d_csum;
952 /* Merge TCP data checksums. */
953 csum = (uint32_t)m->m_pkthdr.lro_tcp_d_csum +
954 (uint32_t)p->m_pkthdr.lro_tcp_d_csum;
955 while (csum > 0xffff)
956 csum = (csum >> 16) + (csum & 0xffff);
959 /* Update various counters. */
960 m->m_pkthdr.len += p->m_pkthdr.lro_tcp_d_len;
961 m->m_pkthdr.lro_tcp_d_csum = csum;
962 m->m_pkthdr.lro_tcp_d_len += p->m_pkthdr.lro_tcp_d_len;
963 m->m_pkthdr.lro_nsegs += p->m_pkthdr.lro_nsegs;
968 tcp_lro_condense(struct lro_ctrl *lc, struct lro_entry *le)
971 * Walk through the mbuf chain we
972 * have on tap and compress/condense
978 uint32_t tcp_data_len_total;
979 uint32_t tcp_data_seg_total;
980 uint16_t tcp_data_len;
981 uint16_t tcp_opt_len;
984 * First we must check the lead (m_head)
985 * we must make sure that it is *not*
986 * something that should be sent up
987 * right away (sack etc).
990 m = le->m_head->m_nextpkt;
996 th = tcp_lro_get_th(m);
997 tcp_opt_len = (th->th_off << 2);
998 tcp_opt_len -= sizeof(*th);
999 ts_ptr = (uint32_t *)(th + 1);
1001 if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1002 *ts_ptr != TCP_LRO_TS_OPTION)) {
1004 * Its not the timestamp. We can't
1005 * use this guy as the head.
1007 le->m_head->m_nextpkt = m->m_nextpkt;
1008 tcp_push_and_replace(lc, le, m);
1011 if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0) {
1013 * Make sure that previously seen segments/ACKs are delivered
1014 * before this segment, e.g. FIN.
1016 le->m_head->m_nextpkt = m->m_nextpkt;
1017 tcp_push_and_replace(lc, le, m);
1020 while((m = le->m_head->m_nextpkt) != NULL) {
1022 * condense m into le, first
1023 * pull m out of the list.
1025 le->m_head->m_nextpkt = m->m_nextpkt;
1026 m->m_nextpkt = NULL;
1028 tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
1029 th = tcp_lro_get_th(m);
1030 ts_ptr = (uint32_t *)(th + 1);
1031 tcp_opt_len = (th->th_off << 2);
1032 tcp_opt_len -= sizeof(*th);
1033 tcp_data_len_total = le->m_head->m_pkthdr.lro_tcp_d_len + tcp_data_len;
1034 tcp_data_seg_total = le->m_head->m_pkthdr.lro_nsegs + m->m_pkthdr.lro_nsegs;
1036 if (tcp_data_seg_total >= lc->lro_ackcnt_lim ||
1037 tcp_data_len_total >= lc->lro_length_lim) {
1038 /* Flush now if appending will result in overflow. */
1039 tcp_push_and_replace(lc, le, m);
1042 if (tcp_opt_len != 0 &&
1043 __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1044 *ts_ptr != TCP_LRO_TS_OPTION)) {
1046 * Maybe a sack in the new one? We need to
1047 * start all over after flushing the
1048 * current le. We will go up to the beginning
1049 * and flush it (calling the replace again possibly
1050 * or just returning).
1052 tcp_push_and_replace(lc, le, m);
1055 if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0) {
1056 tcp_push_and_replace(lc, le, m);
1059 if (tcp_opt_len != 0) {
1060 uint32_t tsval = ntohl(*(ts_ptr + 1));
1061 /* Make sure timestamp values are increasing. */
1062 if (TSTMP_GT(le->tsval, tsval)) {
1063 tcp_push_and_replace(lc, le, m);
1067 le->tsecr = *(ts_ptr + 2);
1069 /* Try to append the new segment. */
1070 if (__predict_false(ntohl(th->th_seq) != le->next_seq ||
1071 ((th->th_flags & TH_ACK) !=
1072 (le->flags & TH_ACK)) ||
1073 (tcp_data_len == 0 &&
1074 le->ack_seq == th->th_ack &&
1075 le->window == th->th_win))) {
1076 /* Out of order packet, non-ACK + ACK or dup ACK. */
1077 tcp_push_and_replace(lc, le, m);
1080 if (tcp_data_len != 0 ||
1081 SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1082 le->next_seq += tcp_data_len;
1083 le->ack_seq = th->th_ack;
1084 le->window = th->th_win;
1085 le->needs_merge = 1;
1086 } else if (th->th_ack == le->ack_seq) {
1087 if (WIN_GT(th->th_win, le->window)) {
1088 le->window = th->th_win;
1089 le->needs_merge = 1;
1093 if (tcp_data_len == 0) {
1098 /* Merge TCP data checksum and length to head mbuf. */
1099 tcp_lro_mbuf_append_pkthdr(le, m);
1102 * Adjust the mbuf so that m_data points to the first byte of
1103 * the ULP payload. Adjust the mbuf to avoid complications and
1104 * append new segment to existing mbuf chain.
1106 m_adj(m, m->m_pkthdr.len - tcp_data_len);
1108 le->m_tail->m_next = m;
1109 le->m_tail = m_last(m);
1115 tcp_queue_pkts(struct inpcb *inp, struct tcpcb *tp, struct lro_entry *le)
1117 INP_WLOCK_ASSERT(inp);
1118 if (tp->t_in_pkt == NULL) {
1119 /* Nothing yet there */
1120 tp->t_in_pkt = le->m_head;
1121 tp->t_tail_pkt = le->m_last_mbuf;
1123 /* Already some there */
1124 tp->t_tail_pkt->m_nextpkt = le->m_head;
1125 tp->t_tail_pkt = le->m_last_mbuf;
1128 le->m_last_mbuf = NULL;
1131 static struct mbuf *
1132 tcp_lro_get_last_if_ackcmp(struct lro_ctrl *lc, struct lro_entry *le,
1133 struct inpcb *inp, int32_t *new_m)
1138 tp = intotcpcb(inp);
1139 if (__predict_false(tp == NULL))
1142 /* Look at the last mbuf if any in queue */
1144 if (m != NULL && (m->m_flags & M_ACKCMP) != 0) {
1145 if (M_TRAILINGSPACE(m) >= sizeof(struct tcp_ackent)) {
1146 tcp_lro_log(tp, lc, le, NULL, 23, 0, 0, 0, 0);
1148 counter_u64_add(tcp_extra_mbuf, 1);
1151 /* Mark we ran out of space */
1152 inp->inp_flags2 |= INP_MBUF_L_ACKS;
1155 /* Decide mbuf size. */
1156 if (inp->inp_flags2 & INP_MBUF_L_ACKS)
1157 m = m_getcl(M_NOWAIT, MT_DATA, M_ACKCMP | M_PKTHDR);
1159 m = m_gethdr(M_NOWAIT, MT_DATA);
1161 if (__predict_false(m == NULL)) {
1162 counter_u64_add(tcp_would_have_but, 1);
1165 counter_u64_add(tcp_comp_total, 1);
1166 m->m_flags |= M_ACKCMP;
1171 static struct inpcb *
1172 tcp_lro_lookup(struct ifnet *ifp, struct lro_parser *pa)
1176 switch (pa->data.lro_type) {
1178 case LRO_TYPE_IPV6_TCP:
1179 inp = in6_pcblookup(&V_tcbinfo,
1180 &pa->data.s_addr.v6,
1182 &pa->data.d_addr.v6,
1189 case LRO_TYPE_IPV4_TCP:
1190 inp = in_pcblookup(&V_tcbinfo,
1207 tcp_lro_ack_valid(struct mbuf *m, struct tcphdr *th, uint32_t **ppts, bool *other_opts)
1210 * This function returns two bits of valuable information.
1211 * a) Is what is present capable of being ack-compressed,
1212 * we can ack-compress if there is no options or just
1213 * a timestamp option, and of course the th_flags must
1214 * be correct as well.
1215 * b) Our other options present such as SACK. This is
1216 * used to determine if we want to wakeup or not.
1220 switch (th->th_off << 2) {
1221 case (sizeof(*th) + TCPOLEN_TSTAMP_APPA):
1222 *ppts = (uint32_t *)(th + 1);
1223 /* Check if we have only one timestamp option. */
1224 if (**ppts == TCP_LRO_TS_OPTION)
1225 *other_opts = false;
1234 *other_opts = false;
1242 /* For ACKCMP we only accept ACK, PUSH, ECE and CWR. */
1243 if ((th->th_flags & ~(TH_ACK | TH_PUSH | TH_ECE | TH_CWR)) != 0)
1245 /* If it has data on it we cannot compress it */
1246 if (m->m_pkthdr.lro_tcp_d_len)
1249 /* ACK flag must be set. */
1250 if (!(th->th_flags & TH_ACK))
1256 tcp_lro_flush_tcphpts(struct lro_ctrl *lc, struct lro_entry *le)
1260 struct mbuf **pp, *cmp, *mv_to;
1261 bool bpf_req, should_wake;
1263 /* Check if packet doesn't belongs to our network interface. */
1264 if ((tcplro_stacks_wanting_mbufq == 0) ||
1265 (le->outer.data.vlan_id != 0) ||
1266 (le->inner.data.lro_type != LRO_TYPE_NONE))
1267 return (TCP_LRO_CANNOT);
1271 * Be proactive about unspecified IPv6 address in source. As
1272 * we use all-zero to indicate unbounded/unconnected pcb,
1273 * unspecified IPv6 address can be used to confuse us.
1275 * Note that packets with unspecified IPv6 destination is
1276 * already dropped in ip6_input.
1278 if (__predict_false(le->outer.data.lro_type == LRO_TYPE_IPV6_TCP &&
1279 IN6_IS_ADDR_UNSPECIFIED(&le->outer.data.s_addr.v6)))
1280 return (TCP_LRO_CANNOT);
1282 if (__predict_false(le->inner.data.lro_type == LRO_TYPE_IPV6_TCP &&
1283 IN6_IS_ADDR_UNSPECIFIED(&le->inner.data.s_addr.v6)))
1284 return (TCP_LRO_CANNOT);
1286 /* Lookup inp, if any. */
1287 inp = tcp_lro_lookup(lc->ifp,
1288 (le->inner.data.lro_type == LRO_TYPE_NONE) ? &le->outer : &le->inner);
1290 return (TCP_LRO_CANNOT);
1292 counter_u64_add(tcp_inp_lro_locks_taken, 1);
1294 /* Get TCP control structure. */
1295 tp = intotcpcb(inp);
1297 /* Check if the inp is dead, Jim. */
1299 (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) ||
1300 (inp->inp_flags2 & INP_FREED)) {
1302 return (TCP_LRO_CANNOT);
1304 if ((inp->inp_irq_cpu_set == 0) && (lc->lro_cpu_is_set == 1)) {
1305 inp->inp_irq_cpu = lc->lro_last_cpu;
1306 inp->inp_irq_cpu_set = 1;
1308 /* Check if the transport doesn't support the needed optimizations. */
1309 if ((inp->inp_flags2 & (INP_SUPPORTS_MBUFQ | INP_MBUF_ACKCMP)) == 0) {
1311 return (TCP_LRO_CANNOT);
1314 if (inp->inp_flags2 & INP_MBUF_QUEUE_READY)
1315 should_wake = false;
1318 /* Check if packets should be tapped to BPF. */
1319 bpf_req = bpf_peers_present(lc->ifp->if_bpf);
1321 /* Strip and compress all the incoming packets. */
1323 for (pp = &le->m_head; *pp != NULL; ) {
1325 if (do_bpf_strip_and_compress(inp, lc, le, pp,
1326 &cmp, &mv_to, &should_wake, bpf_req ) == false) {
1327 /* Advance to next mbuf. */
1328 pp = &(*pp)->m_nextpkt;
1329 } else if (mv_to != NULL) {
1330 /* We are asked to move pp up */
1331 pp = &mv_to->m_nextpkt;
1334 /* Update "m_last_mbuf", if any. */
1335 if (pp == &le->m_head)
1336 le->m_last_mbuf = *pp;
1338 le->m_last_mbuf = __containerof(pp, struct mbuf, m_nextpkt);
1340 /* Check if any data mbufs left. */
1341 if (le->m_head != NULL) {
1342 counter_u64_add(tcp_inp_lro_direct_queue, 1);
1343 tcp_lro_log(tp, lc, le, NULL, 22, 1,
1344 inp->inp_flags2, inp->inp_in_input, 1);
1345 tcp_queue_pkts(inp, tp, le);
1349 counter_u64_add(tcp_inp_lro_wokeup_queue, 1);
1350 if ((*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0))
1355 return (0); /* Success. */
1360 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
1362 /* Only optimise if there are multiple packets waiting. */
1369 CURVNET_SET(lc->ifp->if_vnet);
1370 error = tcp_lro_flush_tcphpts(lc, le);
1374 tcp_lro_condense(lc, le);
1375 tcp_flush_out_entry(lc, le);
1380 bzero(le, sizeof(*le));
1381 LIST_INSERT_HEAD(&lc->lro_free, le, next);
1384 #ifdef HAVE_INLINE_FLSLL
1385 #define tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
1387 static inline uint64_t
1388 tcp_lro_msb_64(uint64_t x)
1396 return (x & ~(x >> 1));
1401 * The tcp_lro_sort() routine is comparable to qsort(), except it has
1402 * a worst case complexity limit of O(MIN(N,64)*N), where N is the
1403 * number of elements to sort and 64 is the number of sequence bits
1404 * available. The algorithm is bit-slicing the 64-bit sequence number,
1405 * sorting one bit at a time from the most significant bit until the
1406 * least significant one, skipping the constant bits. This is
1407 * typically called a radix sort.
1410 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
1412 struct lro_mbuf_sort temp;
1419 /* for small arrays insertion sort is faster */
1421 for (x = 1; x < size; x++) {
1423 for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
1424 parray[y] = parray[y - 1];
1430 /* compute sequence bits which are constant */
1433 for (x = 0; x != size; x++) {
1434 ones |= parray[x].seq;
1435 zeros |= ~parray[x].seq;
1438 /* compute bits which are not constant into "ones" */
1443 /* pick the most significant bit which is not constant */
1444 ones = tcp_lro_msb_64(ones);
1447 * Move entries having cleared sequence bits to the beginning
1450 for (x = y = 0; y != size; y++) {
1452 if (parray[y].seq & ones)
1456 parray[x] = parray[y];
1461 KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
1464 tcp_lro_sort(parray, x);
1473 tcp_lro_flush_all(struct lro_ctrl *lc)
1480 /* check if no mbufs to flush */
1481 if (lc->lro_mbuf_count == 0)
1483 if (lc->lro_cpu_is_set == 0) {
1484 if (lc->lro_last_cpu == curcpu) {
1485 lc->lro_cnt_of_same_cpu++;
1486 /* Have we reached the threshold to declare a cpu? */
1487 if (lc->lro_cnt_of_same_cpu > tcp_lro_cpu_set_thresh)
1488 lc->lro_cpu_is_set = 1;
1490 lc->lro_last_cpu = curcpu;
1491 lc->lro_cnt_of_same_cpu = 0;
1494 CURVNET_SET(lc->ifp->if_vnet);
1496 /* get current time */
1497 binuptime(&lc->lro_last_queue_time);
1499 /* sort all mbufs according to stream */
1500 tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
1502 /* input data into LRO engine, stream by stream */
1504 for (x = 0; x != lc->lro_mbuf_count; x++) {
1508 mb = lc->lro_mbuf_data[x].mb;
1510 /* get sequence number, masking away the packet index */
1511 nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
1513 /* check for new stream */
1517 /* flush active streams */
1518 tcp_lro_rx_done(lc);
1521 /* add packet to LRO engine */
1522 if (tcp_lro_rx_common(lc, mb, 0, false) != 0) {
1523 /* input packet to network layer */
1524 (*lc->ifp->if_input)(lc->ifp, mb);
1531 /* flush active streams */
1532 tcp_lro_rx_done(lc);
1537 lc->lro_mbuf_count = 0;
1542 build_ack_entry(struct tcp_ackent *ae, struct tcphdr *th, struct mbuf *m,
1543 uint32_t *ts_ptr, uint16_t iptos)
1546 * Given a TCP ACK, summarize it down into the small TCP ACK
1549 ae->timestamp = m->m_pkthdr.rcv_tstmp;
1550 if (m->m_flags & M_TSTMP_LRO)
1551 ae->flags = TSTMP_LRO;
1552 else if (m->m_flags & M_TSTMP)
1553 ae->flags = TSTMP_HDWR;
1554 ae->seq = ntohl(th->th_seq);
1555 ae->ack = ntohl(th->th_ack);
1556 ae->flags |= th->th_flags;
1557 if (ts_ptr != NULL) {
1558 ae->ts_value = ntohl(ts_ptr[1]);
1559 ae->ts_echo = ntohl(ts_ptr[2]);
1560 ae->flags |= HAS_TSTMP;
1562 ae->win = ntohs(th->th_win);
1563 ae->codepoint = iptos;
1567 * Do BPF tap for either ACK_CMP packets or MBUF QUEUE type packets
1568 * and strip all, but the IPv4/IPv6 header.
1571 do_bpf_strip_and_compress(struct inpcb *inp, struct lro_ctrl *lc,
1572 struct lro_entry *le, struct mbuf **pp, struct mbuf **cmp, struct mbuf **mv_to,
1573 bool *should_wake, bool bpf_req)
1578 struct ip6_hdr *ip6;
1583 struct tcp_ackent *ack_ent;
1586 bool other_opts, can_compress;
1592 /* Get current mbuf. */
1595 /* Let the BPF see the packet */
1596 if (__predict_false(bpf_req))
1597 ETHER_BPF_MTAP(lc->ifp, m);
1599 tcp_hdr_offset = m->m_pkthdr.lro_tcp_h_off;
1600 lro_type = le->inner.data.lro_type;
1603 lro_type = le->outer.data.lro_type;
1605 case LRO_TYPE_IPV4_TCP:
1606 tcp_hdr_offset -= sizeof(*le->outer.ip4);
1607 m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1609 case LRO_TYPE_IPV6_TCP:
1610 tcp_hdr_offset -= sizeof(*le->outer.ip6);
1611 m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1617 case LRO_TYPE_IPV4_TCP:
1618 tcp_hdr_offset -= sizeof(*le->outer.ip4);
1619 m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1621 case LRO_TYPE_IPV6_TCP:
1622 tcp_hdr_offset -= sizeof(*le->outer.ip6);
1623 m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1629 MPASS(tcp_hdr_offset >= 0);
1631 m_adj(m, tcp_hdr_offset);
1632 m->m_flags |= M_LRO_EHDRSTRP;
1633 m->m_flags &= ~M_ACKCMP;
1634 m->m_pkthdr.lro_tcp_h_off -= tcp_hdr_offset;
1636 th = tcp_lro_get_th(m);
1638 th->th_sum = 0; /* TCP checksum is valid. */
1640 /* Check if ACK can be compressed */
1641 can_compress = tcp_lro_ack_valid(m, th, &ts_ptr, &other_opts);
1643 /* Now lets look at the should wake states */
1644 if ((other_opts == true) &&
1645 ((inp->inp_flags2 & INP_DONT_SACK_QUEUE) == 0)) {
1647 * If there are other options (SACK?) and the
1648 * tcp endpoint has not expressly told us it does
1649 * not care about SACKS, then we should wake up.
1651 *should_wake = true;
1653 /* Is the ack compressable? */
1654 if (can_compress == false)
1656 /* Does the TCP endpoint support ACK compression? */
1657 if ((inp->inp_flags2 & INP_MBUF_ACKCMP) == 0)
1660 /* Lets get the TOS/traffic class field */
1661 l3.ptr = mtod(m, void *);
1663 case LRO_TYPE_IPV4_TCP:
1664 iptos = l3.ip4->ip_tos;
1666 case LRO_TYPE_IPV6_TCP:
1667 iptos = IPV6_TRAFFIC_CLASS(l3.ip6);
1670 iptos = 0; /* Keep compiler happy. */
1673 /* Now lets get space if we don't have some already */
1676 nm = tcp_lro_get_last_if_ackcmp(lc, le, inp, &n_mbuf);
1677 if (__predict_false(nm == NULL))
1682 * Link in the new cmp ack to our in-order place,
1683 * first set our cmp ack's next to where we are.
1688 * Set it up so mv_to is advanced to our
1689 * compressed ack. This way the caller can
1690 * advance pp to the right place.
1694 * Advance it here locally as well.
1696 pp = &nm->m_nextpkt;
1699 /* We have one already we are working on */
1701 if (M_TRAILINGSPACE(nm) < sizeof(struct tcp_ackent)) {
1702 /* We ran out of space */
1703 inp->inp_flags2 |= INP_MBUF_L_ACKS;
1707 MPASS(M_TRAILINGSPACE(nm) >= sizeof(struct tcp_ackent));
1708 counter_u64_add(tcp_inp_lro_compressed, 1);
1710 /* We can add in to the one on the tail */
1711 ack_ent = mtod(nm, struct tcp_ackent *);
1712 idx = (nm->m_len / sizeof(struct tcp_ackent));
1713 build_ack_entry(&ack_ent[idx], th, m, ts_ptr, iptos);
1715 /* Bump the size of both pkt-hdr and len */
1716 nm->m_len += sizeof(struct tcp_ackent);
1717 nm->m_pkthdr.len += sizeof(struct tcp_ackent);
1719 /* Advance to next mbuf before freeing. */
1721 m->m_nextpkt = NULL;
1725 counter_u64_add(tcp_uncomp_total, 1);
1731 static struct lro_head *
1732 tcp_lro_rx_get_bucket(struct lro_ctrl *lc, struct mbuf *m, struct lro_parser *parser)
1736 if (M_HASHTYPE_ISHASH(m)) {
1737 hash = m->m_pkthdr.flowid;
1739 for (unsigned i = hash = 0; i != LRO_RAW_ADDRESS_MAX; i++)
1740 hash += parser->data.raw[i];
1742 return (&lc->lro_hash[hash % lc->lro_hashsz]);
1746 tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash)
1748 struct lro_parser pi; /* inner address data */
1749 struct lro_parser po; /* outer address data */
1750 struct lro_parser *pa; /* current parser for TCP stream */
1751 struct lro_entry *le;
1752 struct lro_head *bucket;
1757 uint16_t tcp_data_sum;
1760 /* Quickly decide if packet cannot be LRO'ed */
1761 if (__predict_false(V_ipforwarding != 0))
1762 return (TCP_LRO_CANNOT);
1765 /* Quickly decide if packet cannot be LRO'ed */
1766 if (__predict_false(V_ip6_forwarding != 0))
1767 return (TCP_LRO_CANNOT);
1770 /* We expect a contiguous header [eh, ip, tcp]. */
1771 pa = tcp_lro_parser(m, &po, &pi, true);
1772 if (__predict_false(pa == NULL))
1773 return (TCP_LRO_NOT_SUPPORTED);
1775 /* We don't expect any padding. */
1776 error = tcp_lro_trim_mbuf_chain(m, pa);
1777 if (__predict_false(error != 0))
1781 switch (pa->data.lro_type) {
1782 case LRO_TYPE_IPV4_TCP:
1783 error = tcp_lro_rx_ipv4(lc, m, pa->ip4);
1784 if (__predict_false(error != 0))
1791 /* If no hardware or arrival stamp on the packet add timestamp */
1792 if ((m->m_flags & (M_TSTMP_LRO | M_TSTMP)) == 0) {
1793 m->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
1794 m->m_flags |= M_TSTMP_LRO;
1797 /* Get pointer to TCP header. */
1800 /* Don't process SYN packets. */
1801 if (__predict_false(th->th_flags & TH_SYN))
1802 return (TCP_LRO_CANNOT);
1804 /* Get total TCP header length and compute payload length. */
1805 tcp_opt_len = (th->th_off << 2);
1806 tcp_data_len = m->m_pkthdr.len - ((uint8_t *)th -
1807 (uint8_t *)m->m_data) - tcp_opt_len;
1808 tcp_opt_len -= sizeof(*th);
1810 /* Don't process invalid TCP headers. */
1811 if (__predict_false(tcp_opt_len < 0 || tcp_data_len < 0))
1812 return (TCP_LRO_CANNOT);
1814 /* Compute TCP data only checksum. */
1815 if (tcp_data_len == 0)
1816 tcp_data_sum = 0; /* no data, no checksum */
1817 else if (__predict_false(csum != 0))
1818 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~csum);
1820 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~th->th_sum);
1822 /* Save TCP info in mbuf. */
1823 m->m_nextpkt = NULL;
1824 m->m_pkthdr.rcvif = lc->ifp;
1825 m->m_pkthdr.lro_tcp_d_csum = tcp_data_sum;
1826 m->m_pkthdr.lro_tcp_d_len = tcp_data_len;
1827 m->m_pkthdr.lro_tcp_h_off = ((uint8_t *)th - (uint8_t *)m->m_data);
1828 m->m_pkthdr.lro_nsegs = 1;
1830 /* Get hash bucket. */
1832 bucket = &lc->lro_hash[0];
1834 bucket = tcp_lro_rx_get_bucket(lc, m, pa);
1837 /* Try to find a matching previous segment. */
1838 LIST_FOREACH(le, bucket, hash_next) {
1839 /* Compare addresses and ports. */
1840 if (lro_address_compare(&po.data, &le->outer.data) == false ||
1841 lro_address_compare(&pi.data, &le->inner.data) == false)
1844 /* Check if no data and old ACK. */
1845 if (tcp_data_len == 0 &&
1846 SEQ_LT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1851 /* Mark "m" in the last spot. */
1852 le->m_last_mbuf->m_nextpkt = m;
1853 /* Now set the tail to "m". */
1854 le->m_last_mbuf = m;
1858 /* Try to find an empty slot. */
1859 if (LIST_EMPTY(&lc->lro_free))
1860 return (TCP_LRO_NO_ENTRIES);
1862 /* Start a new segment chain. */
1863 le = LIST_FIRST(&lc->lro_free);
1864 LIST_REMOVE(le, next);
1865 tcp_lro_active_insert(lc, bucket, le);
1867 /* Make sure the headers are set. */
1871 /* Store time this entry was allocated. */
1872 le->alloc_time = lc->lro_last_queue_time;
1874 tcp_set_entry_to_mbuf(lc, le, m, th);
1876 /* Now set the tail to "m". */
1877 le->m_last_mbuf = m;
1883 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
1887 /* get current time */
1888 binuptime(&lc->lro_last_queue_time);
1890 CURVNET_SET(lc->ifp->if_vnet);
1891 error = tcp_lro_rx_common(lc, m, csum, true);
1898 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
1902 if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
1903 lc->lro_mbuf_max == 0)) {
1909 /* check if packet is not LRO capable */
1910 if (__predict_false(mb->m_pkthdr.csum_flags == 0 ||
1911 (lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
1912 /* input packet to network layer */
1913 (*lc->ifp->if_input) (lc->ifp, mb);
1917 /* create sequence number */
1918 lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
1919 (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
1920 (((uint64_t)mb->m_pkthdr.flowid) << 24) |
1921 ((uint64_t)lc->lro_mbuf_count);
1924 lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb;
1926 /* flush if array is full */
1927 if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max))
1928 tcp_lro_flush_all(lc);