2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2007, Myricom Inc.
5 * Copyright (c) 2008, Intel Corporation.
6 * Copyright (c) 2012 The FreeBSD Foundation
7 * Copyright (c) 2016-2021 Mellanox Technologies.
10 * Portions of this software were developed by Bjoern Zeeb
11 * under sponsorship from the FreeBSD Foundation.
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
22 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
39 #include "opt_inet6.h"
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/malloc.h>
46 #include <sys/socket.h>
47 #include <sys/socketvar.h>
48 #include <sys/sockbuf.h>
49 #include <sys/sysctl.h>
52 #include <net/if_var.h>
53 #include <net/ethernet.h>
56 #include <net/if_dl.h>
57 #include <net/if_media.h>
58 #include <net/if_types.h>
59 #include <net/infiniband.h>
60 #include <net/if_lagg.h>
62 #include <netinet/in_systm.h>
63 #include <netinet/in.h>
64 #include <netinet/ip6.h>
65 #include <netinet/ip.h>
66 #include <netinet/ip_var.h>
67 #include <netinet/in_pcb.h>
68 #include <netinet6/in6_pcb.h>
69 #include <netinet/tcp.h>
70 #include <netinet/tcp_seq.h>
71 #include <netinet/tcp_lro.h>
72 #include <netinet/tcp_var.h>
73 #include <netinet/tcpip.h>
74 #include <netinet/tcp_hpts.h>
75 #include <netinet/tcp_log_buf.h>
76 #include <netinet/udp.h>
77 #include <netinet6/ip6_var.h>
79 #include <machine/in_cksum.h>
81 static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures");
83 #define TCP_LRO_TS_OPTION \
84 ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
85 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
87 static void tcp_lro_rx_done(struct lro_ctrl *lc);
88 static int tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m,
89 uint32_t csum, bool use_hash);
92 static bool do_bpf_strip_and_compress(struct inpcb *, struct lro_ctrl *,
93 struct lro_entry *, struct mbuf **, struct mbuf **, struct mbuf **,
94 bool *, bool, bool, struct ifnet *);
98 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
101 static long tcplro_stacks_wanting_mbufq;
102 counter_u64_t tcp_inp_lro_direct_queue;
103 counter_u64_t tcp_inp_lro_wokeup_queue;
104 counter_u64_t tcp_inp_lro_compressed;
105 counter_u64_t tcp_inp_lro_locks_taken;
106 counter_u64_t tcp_extra_mbuf;
107 counter_u64_t tcp_would_have_but;
108 counter_u64_t tcp_comp_total;
109 counter_u64_t tcp_uncomp_total;
110 counter_u64_t tcp_bad_csums;
112 static unsigned tcp_lro_entries = TCP_LRO_ENTRIES;
113 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries,
114 CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_entries, 0,
115 "default number of LRO entries");
117 static uint32_t tcp_lro_cpu_set_thresh = TCP_LRO_CPU_DECLARATION_THRESH;
118 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_cpu_threshold,
119 CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_cpu_set_thresh, 0,
120 "Number of interrupts in a row on the same CPU that will make us declare an 'affinity' cpu?");
122 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, fullqueue, CTLFLAG_RD,
123 &tcp_inp_lro_direct_queue, "Number of lro's fully queued to transport");
124 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, wokeup, CTLFLAG_RD,
125 &tcp_inp_lro_wokeup_queue, "Number of lro's where we woke up transport via hpts");
126 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, compressed, CTLFLAG_RD,
127 &tcp_inp_lro_compressed, "Number of lro's compressed and sent to transport");
128 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lockcnt, CTLFLAG_RD,
129 &tcp_inp_lro_locks_taken, "Number of lro's inp_wlocks taken");
130 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, extra_mbuf, CTLFLAG_RD,
131 &tcp_extra_mbuf, "Number of times we had an extra compressed ack dropped into the tp");
132 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, would_have_but, CTLFLAG_RD,
133 &tcp_would_have_but, "Number of times we would have had an extra compressed, but mget failed");
134 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, with_m_ackcmp, CTLFLAG_RD,
135 &tcp_comp_total, "Number of mbufs queued with M_ACKCMP flags set");
136 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, without_m_ackcmp, CTLFLAG_RD,
137 &tcp_uncomp_total, "Number of mbufs queued without M_ACKCMP");
138 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lro_badcsum, CTLFLAG_RD,
139 &tcp_bad_csums, "Number of packets that the common code saw with bad csums");
142 tcp_lro_reg_mbufq(void)
144 atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, 1);
148 tcp_lro_dereg_mbufq(void)
150 atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, -1);
154 tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket,
155 struct lro_entry *le)
158 LIST_INSERT_HEAD(&lc->lro_active, le, next);
159 LIST_INSERT_HEAD(bucket, le, hash_next);
163 tcp_lro_active_remove(struct lro_entry *le)
166 LIST_REMOVE(le, next); /* active list */
167 LIST_REMOVE(le, hash_next); /* hash bucket */
171 tcp_lro_init(struct lro_ctrl *lc)
173 return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0));
177 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
178 unsigned lro_entries, unsigned lro_mbufs)
180 struct lro_entry *le;
182 unsigned i, elements;
184 lc->lro_bad_csum = 0;
187 lc->lro_mbuf_count = 0;
188 lc->lro_mbuf_max = lro_mbufs;
189 lc->lro_cnt = lro_entries;
190 lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
191 lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
193 LIST_INIT(&lc->lro_free);
194 LIST_INIT(&lc->lro_active);
196 /* create hash table to accelerate entry lookup */
197 if (lro_entries > lro_mbufs)
198 elements = lro_entries;
200 elements = lro_mbufs;
201 lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz,
203 if (lc->lro_hash == NULL) {
204 memset(lc, 0, sizeof(*lc));
208 /* compute size to allocate */
209 size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
210 (lro_entries * sizeof(*le));
211 lc->lro_mbuf_data = (struct lro_mbuf_sort *)
212 malloc(size, M_LRO, M_NOWAIT | M_ZERO);
214 /* check for out of memory */
215 if (lc->lro_mbuf_data == NULL) {
216 free(lc->lro_hash, M_LRO);
217 memset(lc, 0, sizeof(*lc));
220 /* compute offset for LRO entries */
221 le = (struct lro_entry *)
222 (lc->lro_mbuf_data + lro_mbufs);
224 /* setup linked list */
225 for (i = 0; i != lro_entries; i++)
226 LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
231 struct vxlan_header {
237 tcp_lro_low_level_parser(void *ptr, struct lro_parser *parser, bool update_data, bool is_vxlan, int mlen)
239 const struct ether_vlan_header *eh;
244 memset(parser, 0, sizeof(*parser));
249 const struct vxlan_header *vxh;
251 ptr = (uint8_t *)ptr + sizeof(*vxh);
253 parser->data.vxlan_vni =
254 vxh->vxlh_vni & htonl(0xffffff00);
259 if (__predict_false(eh->evl_encap_proto == htons(ETHERTYPE_VLAN))) {
260 eth_type = eh->evl_proto;
262 /* strip priority and keep VLAN ID only */
263 parser->data.vlan_id = eh->evl_tag & htons(EVL_VLID_MASK);
265 /* advance to next header */
266 ptr = (uint8_t *)ptr + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
267 mlen -= (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
269 eth_type = eh->evl_encap_proto;
270 /* advance to next header */
271 mlen -= ETHER_HDR_LEN;
272 ptr = (uint8_t *)ptr + ETHER_HDR_LEN;
274 if (__predict_false(mlen <= 0))
278 case htons(ETHERTYPE_IP):
280 if (__predict_false(mlen < sizeof(struct ip)))
282 /* Ensure there are no IPv4 options. */
283 if ((parser->ip4->ip_hl << 2) != sizeof (*parser->ip4))
285 /* .. and the packet is not fragmented. */
286 if (parser->ip4->ip_off & htons(IP_MF|IP_OFFMASK))
288 ptr = (uint8_t *)ptr + (parser->ip4->ip_hl << 2);
289 mlen -= sizeof(struct ip);
291 parser->data.s_addr.v4 = parser->ip4->ip_src;
292 parser->data.d_addr.v4 = parser->ip4->ip_dst;
294 switch (parser->ip4->ip_p) {
296 if (__predict_false(mlen < sizeof(struct udphdr)))
300 parser->data.lro_type = LRO_TYPE_IPV4_UDP;
301 parser->data.s_port = parser->udp->uh_sport;
302 parser->data.d_port = parser->udp->uh_dport;
304 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_UDP);
306 ptr = ((uint8_t *)ptr + sizeof(*parser->udp));
307 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
311 if (__predict_false(mlen < sizeof(struct tcphdr)))
314 parser->data.lro_type = LRO_TYPE_IPV4_TCP;
315 parser->data.s_port = parser->tcp->th_sport;
316 parser->data.d_port = parser->tcp->th_dport;
318 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_TCP);
320 if (__predict_false(mlen < (parser->tcp->th_off << 2)))
322 ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
323 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
331 case htons(ETHERTYPE_IPV6):
333 if (__predict_false(mlen < sizeof(struct ip6_hdr)))
335 ptr = (uint8_t *)ptr + sizeof(*parser->ip6);
337 parser->data.s_addr.v6 = parser->ip6->ip6_src;
338 parser->data.d_addr.v6 = parser->ip6->ip6_dst;
340 mlen -= sizeof(struct ip6_hdr);
341 switch (parser->ip6->ip6_nxt) {
343 if (__predict_false(mlen < sizeof(struct udphdr)))
347 parser->data.lro_type = LRO_TYPE_IPV6_UDP;
348 parser->data.s_port = parser->udp->uh_sport;
349 parser->data.d_port = parser->udp->uh_dport;
351 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_UDP);
353 ptr = (uint8_t *)ptr + sizeof(*parser->udp);
354 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
357 if (__predict_false(mlen < sizeof(struct tcphdr)))
361 parser->data.lro_type = LRO_TYPE_IPV6_TCP;
362 parser->data.s_port = parser->tcp->th_sport;
363 parser->data.d_port = parser->tcp->th_dport;
365 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_TCP);
367 if (__predict_false(mlen < (parser->tcp->th_off << 2)))
369 ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
370 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
380 /* Invalid packet - cannot parse */
384 static const int vxlan_csum = CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
385 CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID;
387 static inline struct lro_parser *
388 tcp_lro_parser(struct mbuf *m, struct lro_parser *po, struct lro_parser *pi, bool update_data)
392 /* Try to parse outer headers first. */
393 data_ptr = tcp_lro_low_level_parser(m->m_data, po, update_data, false, m->m_len);
394 if (data_ptr == NULL || po->total_hdr_len > m->m_len)
398 /* Store VLAN ID, if any. */
399 if (__predict_false(m->m_flags & M_VLANTAG)) {
401 htons(m->m_pkthdr.ether_vtag) & htons(EVL_VLID_MASK);
403 /* Store decrypted flag, if any. */
404 if (__predict_false((m->m_pkthdr.csum_flags &
405 CSUM_TLS_MASK) == CSUM_TLS_DECRYPTED))
406 po->data.lro_flags |= LRO_FLAG_DECRYPTED;
409 switch (po->data.lro_type) {
410 case LRO_TYPE_IPV4_UDP:
411 case LRO_TYPE_IPV6_UDP:
412 /* Check for VXLAN headers. */
413 if ((m->m_pkthdr.csum_flags & vxlan_csum) != vxlan_csum)
416 /* Try to parse inner headers. */
417 data_ptr = tcp_lro_low_level_parser(data_ptr, pi, update_data, true,
418 (m->m_len - ((caddr_t)data_ptr - m->m_data)));
419 if (data_ptr == NULL || (pi->total_hdr_len + po->total_hdr_len) > m->m_len)
422 /* Verify supported header types. */
423 switch (pi->data.lro_type) {
424 case LRO_TYPE_IPV4_TCP:
425 case LRO_TYPE_IPV6_TCP:
431 case LRO_TYPE_IPV4_TCP:
432 case LRO_TYPE_IPV6_TCP:
434 memset(pi, 0, sizeof(*pi));
443 tcp_lro_trim_mbuf_chain(struct mbuf *m, const struct lro_parser *po)
447 switch (po->data.lro_type) {
449 case LRO_TYPE_IPV4_TCP:
450 len = ((uint8_t *)po->ip4 - (uint8_t *)m->m_data) +
451 ntohs(po->ip4->ip_len);
455 case LRO_TYPE_IPV6_TCP:
456 len = ((uint8_t *)po->ip6 - (uint8_t *)m->m_data) +
457 ntohs(po->ip6->ip6_plen) + sizeof(*po->ip6);
461 return (TCP_LRO_CANNOT);
465 * If the frame is padded beyond the end of the IP packet,
466 * then trim the extra bytes off:
468 if (__predict_true(m->m_pkthdr.len == len)) {
470 } else if (m->m_pkthdr.len > len) {
471 m_adj(m, len - m->m_pkthdr.len);
474 return (TCP_LRO_CANNOT);
477 static struct tcphdr *
478 tcp_lro_get_th(struct mbuf *m)
480 return ((struct tcphdr *)((uint8_t *)m->m_data + m->m_pkthdr.lro_tcp_h_off));
484 lro_free_mbuf_chain(struct mbuf *m)
497 tcp_lro_free(struct lro_ctrl *lc)
499 struct lro_entry *le;
502 /* reset LRO free list */
503 LIST_INIT(&lc->lro_free);
505 /* free active mbufs, if any */
506 while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
507 tcp_lro_active_remove(le);
508 lro_free_mbuf_chain(le->m_head);
511 /* free hash table */
512 free(lc->lro_hash, M_LRO);
516 /* free mbuf array, if any */
517 for (x = 0; x != lc->lro_mbuf_count; x++)
518 m_freem(lc->lro_mbuf_data[x].mb);
519 lc->lro_mbuf_count = 0;
521 /* free allocated memory, if any */
522 free(lc->lro_mbuf_data, M_LRO);
523 lc->lro_mbuf_data = NULL;
527 tcp_lro_rx_csum_tcphdr(const struct tcphdr *th)
533 csum = -th->th_sum; /* exclude checksum field */
535 ptr = (const uint16_t *)th;
542 while (csum > 0xffff)
543 csum = (csum >> 16) + (csum & 0xffff);
549 tcp_lro_rx_csum_data(const struct lro_parser *pa, uint16_t tcp_csum)
556 switch (pa->data.lro_type) {
558 case LRO_TYPE_IPV6_TCP:
559 /* Compute full pseudo IPv6 header checksum. */
560 cs = in6_cksum_pseudo(pa->ip6, ntohs(pa->ip6->ip6_plen), pa->ip6->ip6_nxt, 0);
564 case LRO_TYPE_IPV4_TCP:
565 /* Compute full pseudo IPv4 header checsum. */
566 cs = in_addword(ntohs(pa->ip4->ip_len) - sizeof(*pa->ip4), IPPROTO_TCP);
567 cs = in_pseudo(pa->ip4->ip_src.s_addr, pa->ip4->ip_dst.s_addr, htons(cs));
571 cs = 0; /* Keep compiler happy. */
575 /* Complement checksum. */
579 /* Remove TCP header checksum. */
580 cs = ~tcp_lro_rx_csum_tcphdr(pa->tcp);
583 /* Compute checksum remainder. */
585 c = (c >> 16) + (c & 0xffff);
591 tcp_lro_rx_done(struct lro_ctrl *lc)
593 struct lro_entry *le;
595 while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
596 tcp_lro_active_remove(le);
597 tcp_lro_flush(lc, le);
602 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
604 struct lro_entry *le, *le_tmp;
609 if (LIST_EMPTY(&lc->lro_active))
612 /* get timeout time and current time in ns */
614 now = bintime2ns(&bt);
615 tov = ((timeout->tv_sec * 1000000000) + (timeout->tv_usec * 1000));
616 LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
617 if (now >= (bintime2ns(&le->alloc_time) + tov)) {
618 tcp_lro_active_remove(le);
619 tcp_lro_flush(lc, le);
626 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4)
630 /* Legacy IP has a header checksum that needs to be correct. */
631 if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) {
632 if (__predict_false((m->m_pkthdr.csum_flags & CSUM_IP_VALID) == 0)) {
634 return (TCP_LRO_CANNOT);
637 csum = in_cksum_hdr(ip4);
638 if (__predict_false(csum != 0)) {
640 return (TCP_LRO_CANNOT);
649 tcp_lro_log(struct tcpcb *tp, const struct lro_ctrl *lc,
650 const struct lro_entry *le, const struct mbuf *m,
651 int frm, int32_t tcp_data_len, uint32_t th_seq,
652 uint32_t th_ack, uint16_t th_win)
654 if (tp->t_logstate != TCP_LOG_STATE_OFF) {
655 union tcp_log_stackspecific log;
656 struct timeval tv, btv;
659 cts = tcp_get_usecs(&tv);
660 memset(&log, 0, sizeof(union tcp_log_stackspecific));
661 log.u_bbr.flex8 = frm;
662 log.u_bbr.flex1 = tcp_data_len;
664 log.u_bbr.flex2 = m->m_pkthdr.len;
667 log.u_bbr.flex3 = le->m_head->m_pkthdr.lro_nsegs;
668 log.u_bbr.flex4 = le->m_head->m_pkthdr.lro_tcp_d_len;
670 log.u_bbr.flex5 = le->m_head->m_pkthdr.len;
671 log.u_bbr.delRate = le->m_head->m_flags;
672 log.u_bbr.rttProp = le->m_head->m_pkthdr.rcv_tstmp;
674 log.u_bbr.inflight = th_seq;
675 log.u_bbr.delivered = th_ack;
676 log.u_bbr.timeStamp = cts;
677 log.u_bbr.epoch = le->next_seq;
678 log.u_bbr.lt_epoch = le->ack_seq;
679 log.u_bbr.pacing_gain = th_win;
680 log.u_bbr.cwnd_gain = le->window;
681 log.u_bbr.lost = curcpu;
682 log.u_bbr.cur_del_rate = (uintptr_t)m;
683 log.u_bbr.bw_inuse = (uintptr_t)le->m_head;
684 bintime2timeval(&lc->lro_last_queue_time, &btv);
685 log.u_bbr.flex6 = tcp_tv_to_usectick(&btv);
686 log.u_bbr.flex7 = le->compressed;
687 log.u_bbr.pacing_gain = le->uncompressed;
688 if (in_epoch(net_epoch_preempt))
689 log.u_bbr.inhpts = 1;
691 log.u_bbr.inhpts = 0;
692 TCP_LOG_EVENTP(tp, NULL,
693 &tp->t_inpcb->inp_socket->so_rcv,
694 &tp->t_inpcb->inp_socket->so_snd,
696 0, &log, false, &tv);
702 tcp_lro_assign_and_checksum_16(uint16_t *ptr, uint16_t value, uint16_t *psum)
706 csum = 0xffff - *ptr + value;
707 while (csum > 0xffff)
708 csum = (csum >> 16) + (csum & 0xffff);
714 tcp_lro_update_checksum(const struct lro_parser *pa, const struct lro_entry *le,
715 uint16_t payload_len, uint16_t delta_sum)
719 uint16_t temp[5] = {};
721 switch (pa->data.lro_type) {
722 case LRO_TYPE_IPV4_TCP:
723 /* Compute new IPv4 length. */
724 tlen = (pa->ip4->ip_hl << 2) + (pa->tcp->th_off << 2) + payload_len;
725 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
727 /* Subtract delta from current IPv4 checksum. */
728 csum = pa->ip4->ip_sum + 0xffff - temp[0];
729 while (csum > 0xffff)
730 csum = (csum >> 16) + (csum & 0xffff);
731 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
732 goto update_tcp_header;
734 case LRO_TYPE_IPV6_TCP:
735 /* Compute new IPv6 length. */
736 tlen = (pa->tcp->th_off << 2) + payload_len;
737 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
738 goto update_tcp_header;
740 case LRO_TYPE_IPV4_UDP:
741 /* Compute new IPv4 length. */
742 tlen = (pa->ip4->ip_hl << 2) + sizeof(*pa->udp) + payload_len;
743 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
745 /* Subtract delta from current IPv4 checksum. */
746 csum = pa->ip4->ip_sum + 0xffff - temp[0];
747 while (csum > 0xffff)
748 csum = (csum >> 16) + (csum & 0xffff);
749 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
750 goto update_udp_header;
752 case LRO_TYPE_IPV6_UDP:
753 /* Compute new IPv6 length. */
754 tlen = sizeof(*pa->udp) + payload_len;
755 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
756 goto update_udp_header;
763 /* Compute current TCP header checksum. */
764 temp[2] = tcp_lro_rx_csum_tcphdr(pa->tcp);
766 /* Incorporate the latest ACK into the TCP header. */
767 pa->tcp->th_ack = le->ack_seq;
768 pa->tcp->th_win = le->window;
770 /* Incorporate latest timestamp into the TCP header. */
771 if (le->timestamp != 0) {
774 ts_ptr = (uint32_t *)(pa->tcp + 1);
775 ts_ptr[1] = htonl(le->tsval);
776 ts_ptr[2] = le->tsecr;
779 /* Compute new TCP header checksum. */
780 temp[3] = tcp_lro_rx_csum_tcphdr(pa->tcp);
782 /* Compute new TCP checksum. */
783 csum = pa->tcp->th_sum + 0xffff - delta_sum +
784 0xffff - temp[0] + 0xffff - temp[3] + temp[2];
785 while (csum > 0xffff)
786 csum = (csum >> 16) + (csum & 0xffff);
788 /* Assign new TCP checksum. */
789 tcp_lro_assign_and_checksum_16(&pa->tcp->th_sum, csum, &temp[4]);
791 /* Compute all modififications affecting next checksum. */
792 csum = temp[0] + temp[1] + 0xffff - temp[2] +
793 temp[3] + temp[4] + delta_sum;
794 while (csum > 0xffff)
795 csum = (csum >> 16) + (csum & 0xffff);
797 /* Return delta checksum to next stage, if any. */
801 tlen = sizeof(*pa->udp) + payload_len;
802 /* Assign new UDP length and compute checksum delta. */
803 tcp_lro_assign_and_checksum_16(&pa->udp->uh_ulen, htons(tlen), &temp[2]);
805 /* Check if there is a UDP checksum. */
806 if (__predict_false(pa->udp->uh_sum != 0)) {
807 /* Compute new UDP checksum. */
808 csum = pa->udp->uh_sum + 0xffff - delta_sum +
809 0xffff - temp[0] + 0xffff - temp[2];
810 while (csum > 0xffff)
811 csum = (csum >> 16) + (csum & 0xffff);
812 /* Assign new UDP checksum. */
813 tcp_lro_assign_and_checksum_16(&pa->udp->uh_sum, csum, &temp[3]);
816 /* Compute all modififications affecting next checksum. */
817 csum = temp[0] + temp[1] + temp[2] + temp[3] + delta_sum;
818 while (csum > 0xffff)
819 csum = (csum >> 16) + (csum & 0xffff);
821 /* Return delta checksum to next stage, if any. */
826 tcp_flush_out_entry(struct lro_ctrl *lc, struct lro_entry *le)
828 /* Check if we need to recompute any checksums. */
829 if (le->needs_merge) {
832 switch (le->inner.data.lro_type) {
833 case LRO_TYPE_IPV4_TCP:
834 csum = tcp_lro_update_checksum(&le->inner, le,
835 le->m_head->m_pkthdr.lro_tcp_d_len,
836 le->m_head->m_pkthdr.lro_tcp_d_csum);
837 csum = tcp_lro_update_checksum(&le->outer, NULL,
838 le->m_head->m_pkthdr.lro_tcp_d_len +
839 le->inner.total_hdr_len, csum);
840 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
841 CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
842 le->m_head->m_pkthdr.csum_data = 0xffff;
843 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
844 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
846 case LRO_TYPE_IPV6_TCP:
847 csum = tcp_lro_update_checksum(&le->inner, le,
848 le->m_head->m_pkthdr.lro_tcp_d_len,
849 le->m_head->m_pkthdr.lro_tcp_d_csum);
850 csum = tcp_lro_update_checksum(&le->outer, NULL,
851 le->m_head->m_pkthdr.lro_tcp_d_len +
852 le->inner.total_hdr_len, csum);
853 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
855 le->m_head->m_pkthdr.csum_data = 0xffff;
856 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
857 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
860 switch (le->outer.data.lro_type) {
861 case LRO_TYPE_IPV4_TCP:
862 csum = tcp_lro_update_checksum(&le->outer, le,
863 le->m_head->m_pkthdr.lro_tcp_d_len,
864 le->m_head->m_pkthdr.lro_tcp_d_csum);
865 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
866 CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
867 le->m_head->m_pkthdr.csum_data = 0xffff;
868 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
869 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
871 case LRO_TYPE_IPV6_TCP:
872 csum = tcp_lro_update_checksum(&le->outer, le,
873 le->m_head->m_pkthdr.lro_tcp_d_len,
874 le->m_head->m_pkthdr.lro_tcp_d_csum);
875 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
877 le->m_head->m_pkthdr.csum_data = 0xffff;
878 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
879 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
891 * Break any chain, this is not set to NULL on the singleton
892 * case m_nextpkt points to m_head. Other case set them
893 * m_nextpkt to NULL in push_and_replace.
895 le->m_head->m_nextpkt = NULL;
896 lc->lro_queued += le->m_head->m_pkthdr.lro_nsegs;
897 (*lc->ifp->if_input)(lc->ifp, le->m_head);
901 tcp_set_entry_to_mbuf(struct lro_ctrl *lc, struct lro_entry *le,
902 struct mbuf *m, struct tcphdr *th)
905 uint16_t tcp_data_len;
906 uint16_t tcp_opt_len;
908 ts_ptr = (uint32_t *)(th + 1);
909 tcp_opt_len = (th->th_off << 2);
910 tcp_opt_len -= sizeof(*th);
912 /* Check if there is a timestamp option. */
913 if (tcp_opt_len == 0 ||
914 __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
915 *ts_ptr != TCP_LRO_TS_OPTION)) {
916 /* We failed to find the timestamp option. */
920 le->tsval = ntohl(*(ts_ptr + 1));
921 le->tsecr = *(ts_ptr + 2);
924 tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
926 /* Pull out TCP sequence numbers and window size. */
927 le->next_seq = ntohl(th->th_seq) + tcp_data_len;
928 le->ack_seq = th->th_ack;
929 le->window = th->th_win;
930 le->flags = tcp_get_flags(th);
933 /* Setup new data pointers. */
935 le->m_tail = m_last(m);
939 tcp_push_and_replace(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m)
941 struct lro_parser *pa;
944 * Push up the stack of the current entry
945 * and replace it with "m".
949 /* Grab off the next and save it */
950 msave = le->m_head->m_nextpkt;
951 le->m_head->m_nextpkt = NULL;
953 /* Now push out the old entry */
954 tcp_flush_out_entry(lc, le);
956 /* Re-parse new header, should not fail. */
957 pa = tcp_lro_parser(m, &le->outer, &le->inner, false);
959 ("tcp_push_and_replace: LRO parser failed on m=%p\n", m));
962 * Now to replace the data properly in the entry
963 * we have to reset the TCP header and
966 tcp_set_entry_to_mbuf(lc, le, m, pa->tcp);
968 /* Restore the next list */
969 m->m_nextpkt = msave;
973 tcp_lro_mbuf_append_pkthdr(struct lro_entry *le, const struct mbuf *p)
979 if (m->m_pkthdr.lro_nsegs == 1) {
980 /* Compute relative checksum. */
981 csum = p->m_pkthdr.lro_tcp_d_csum;
983 /* Merge TCP data checksums. */
984 csum = (uint32_t)m->m_pkthdr.lro_tcp_d_csum +
985 (uint32_t)p->m_pkthdr.lro_tcp_d_csum;
986 while (csum > 0xffff)
987 csum = (csum >> 16) + (csum & 0xffff);
990 /* Update various counters. */
991 m->m_pkthdr.len += p->m_pkthdr.lro_tcp_d_len;
992 m->m_pkthdr.lro_tcp_d_csum = csum;
993 m->m_pkthdr.lro_tcp_d_len += p->m_pkthdr.lro_tcp_d_len;
994 m->m_pkthdr.lro_nsegs += p->m_pkthdr.lro_nsegs;
999 tcp_lro_condense(struct lro_ctrl *lc, struct lro_entry *le)
1002 * Walk through the mbuf chain we
1003 * have on tap and compress/condense
1009 uint32_t tcp_data_len_total;
1010 uint32_t tcp_data_seg_total;
1011 uint16_t tcp_data_len;
1012 uint16_t tcp_opt_len;
1015 * First we must check the lead (m_head)
1016 * we must make sure that it is *not*
1017 * something that should be sent up
1018 * right away (sack etc).
1021 m = le->m_head->m_nextpkt;
1023 /* Just one left. */
1027 th = tcp_lro_get_th(m);
1028 tcp_opt_len = (th->th_off << 2);
1029 tcp_opt_len -= sizeof(*th);
1030 ts_ptr = (uint32_t *)(th + 1);
1032 if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1033 *ts_ptr != TCP_LRO_TS_OPTION)) {
1035 * Its not the timestamp. We can't
1036 * use this guy as the head.
1038 le->m_head->m_nextpkt = m->m_nextpkt;
1039 tcp_push_and_replace(lc, le, m);
1042 if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1044 * Make sure that previously seen segments/ACKs are delivered
1045 * before this segment, e.g. FIN.
1047 le->m_head->m_nextpkt = m->m_nextpkt;
1048 tcp_push_and_replace(lc, le, m);
1051 while((m = le->m_head->m_nextpkt) != NULL) {
1053 * condense m into le, first
1054 * pull m out of the list.
1056 le->m_head->m_nextpkt = m->m_nextpkt;
1057 m->m_nextpkt = NULL;
1059 tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
1060 th = tcp_lro_get_th(m);
1061 ts_ptr = (uint32_t *)(th + 1);
1062 tcp_opt_len = (th->th_off << 2);
1063 tcp_opt_len -= sizeof(*th);
1064 tcp_data_len_total = le->m_head->m_pkthdr.lro_tcp_d_len + tcp_data_len;
1065 tcp_data_seg_total = le->m_head->m_pkthdr.lro_nsegs + m->m_pkthdr.lro_nsegs;
1067 if (tcp_data_seg_total >= lc->lro_ackcnt_lim ||
1068 tcp_data_len_total >= lc->lro_length_lim) {
1069 /* Flush now if appending will result in overflow. */
1070 tcp_push_and_replace(lc, le, m);
1073 if (tcp_opt_len != 0 &&
1074 __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1075 *ts_ptr != TCP_LRO_TS_OPTION)) {
1077 * Maybe a sack in the new one? We need to
1078 * start all over after flushing the
1079 * current le. We will go up to the beginning
1080 * and flush it (calling the replace again possibly
1081 * or just returning).
1083 tcp_push_and_replace(lc, le, m);
1086 if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1087 tcp_push_and_replace(lc, le, m);
1090 if (tcp_opt_len != 0) {
1091 uint32_t tsval = ntohl(*(ts_ptr + 1));
1092 /* Make sure timestamp values are increasing. */
1093 if (TSTMP_GT(le->tsval, tsval)) {
1094 tcp_push_and_replace(lc, le, m);
1098 le->tsecr = *(ts_ptr + 2);
1100 /* Try to append the new segment. */
1101 if (__predict_false(ntohl(th->th_seq) != le->next_seq ||
1102 ((tcp_get_flags(th) & TH_ACK) !=
1103 (le->flags & TH_ACK)) ||
1104 (tcp_data_len == 0 &&
1105 le->ack_seq == th->th_ack &&
1106 le->window == th->th_win))) {
1107 /* Out of order packet, non-ACK + ACK or dup ACK. */
1108 tcp_push_and_replace(lc, le, m);
1111 if (tcp_data_len != 0 ||
1112 SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1113 le->next_seq += tcp_data_len;
1114 le->ack_seq = th->th_ack;
1115 le->window = th->th_win;
1116 le->needs_merge = 1;
1117 } else if (th->th_ack == le->ack_seq) {
1118 if (WIN_GT(th->th_win, le->window)) {
1119 le->window = th->th_win;
1120 le->needs_merge = 1;
1124 if (tcp_data_len == 0) {
1129 /* Merge TCP data checksum and length to head mbuf. */
1130 tcp_lro_mbuf_append_pkthdr(le, m);
1133 * Adjust the mbuf so that m_data points to the first byte of
1134 * the ULP payload. Adjust the mbuf to avoid complications and
1135 * append new segment to existing mbuf chain.
1137 m_adj(m, m->m_pkthdr.len - tcp_data_len);
1139 le->m_tail->m_next = m;
1140 le->m_tail = m_last(m);
1146 tcp_queue_pkts(struct inpcb *inp, struct tcpcb *tp, struct lro_entry *le)
1148 INP_WLOCK_ASSERT(inp);
1149 if (tp->t_in_pkt == NULL) {
1150 /* Nothing yet there */
1151 tp->t_in_pkt = le->m_head;
1152 tp->t_tail_pkt = le->m_last_mbuf;
1154 /* Already some there */
1155 tp->t_tail_pkt->m_nextpkt = le->m_head;
1156 tp->t_tail_pkt = le->m_last_mbuf;
1159 le->m_last_mbuf = NULL;
1162 static struct mbuf *
1163 tcp_lro_get_last_if_ackcmp(struct lro_ctrl *lc, struct lro_entry *le,
1164 struct inpcb *inp, int32_t *new_m)
1169 tp = intotcpcb(inp);
1170 if (__predict_false(tp == NULL))
1173 /* Look at the last mbuf if any in queue */
1175 if (m != NULL && (m->m_flags & M_ACKCMP) != 0) {
1176 if (M_TRAILINGSPACE(m) >= sizeof(struct tcp_ackent)) {
1177 tcp_lro_log(tp, lc, le, NULL, 23, 0, 0, 0, 0);
1179 counter_u64_add(tcp_extra_mbuf, 1);
1182 /* Mark we ran out of space */
1183 inp->inp_flags2 |= INP_MBUF_L_ACKS;
1186 /* Decide mbuf size. */
1187 if (inp->inp_flags2 & INP_MBUF_L_ACKS)
1188 m = m_getcl(M_NOWAIT, MT_DATA, M_ACKCMP | M_PKTHDR);
1190 m = m_gethdr(M_NOWAIT, MT_DATA);
1192 if (__predict_false(m == NULL)) {
1193 counter_u64_add(tcp_would_have_but, 1);
1196 counter_u64_add(tcp_comp_total, 1);
1197 m->m_flags |= M_ACKCMP;
1202 static struct inpcb *
1203 tcp_lro_lookup(struct ifnet *ifp, struct lro_parser *pa)
1207 switch (pa->data.lro_type) {
1209 case LRO_TYPE_IPV6_TCP:
1210 inp = in6_pcblookup(&V_tcbinfo,
1211 &pa->data.s_addr.v6,
1213 &pa->data.d_addr.v6,
1220 case LRO_TYPE_IPV4_TCP:
1221 inp = in_pcblookup(&V_tcbinfo,
1238 tcp_lro_ack_valid(struct mbuf *m, struct tcphdr *th, uint32_t **ppts, bool *other_opts)
1241 * This function returns two bits of valuable information.
1242 * a) Is what is present capable of being ack-compressed,
1243 * we can ack-compress if there is no options or just
1244 * a timestamp option, and of course the th_flags must
1245 * be correct as well.
1246 * b) Our other options present such as SACK. This is
1247 * used to determine if we want to wakeup or not.
1251 switch (th->th_off << 2) {
1252 case (sizeof(*th) + TCPOLEN_TSTAMP_APPA):
1253 *ppts = (uint32_t *)(th + 1);
1254 /* Check if we have only one timestamp option. */
1255 if (**ppts == TCP_LRO_TS_OPTION)
1256 *other_opts = false;
1265 *other_opts = false;
1273 /* For ACKCMP we only accept ACK, PUSH, ECE and CWR. */
1274 if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH | TH_ECE | TH_CWR)) != 0)
1276 /* If it has data on it we cannot compress it */
1277 if (m->m_pkthdr.lro_tcp_d_len)
1280 /* ACK flag must be set. */
1281 if (!(tcp_get_flags(th) & TH_ACK))
1287 tcp_lro_flush_tcphpts(struct lro_ctrl *lc, struct lro_entry *le)
1291 struct mbuf **pp, *cmp, *mv_to;
1292 struct ifnet *lagg_ifp;
1293 bool bpf_req, lagg_bpf_req, should_wake;
1295 /* Check if packet doesn't belongs to our network interface. */
1296 if ((tcplro_stacks_wanting_mbufq == 0) ||
1297 (le->outer.data.vlan_id != 0) ||
1298 (le->inner.data.lro_type != LRO_TYPE_NONE))
1299 return (TCP_LRO_CANNOT);
1303 * Be proactive about unspecified IPv6 address in source. As
1304 * we use all-zero to indicate unbounded/unconnected pcb,
1305 * unspecified IPv6 address can be used to confuse us.
1307 * Note that packets with unspecified IPv6 destination is
1308 * already dropped in ip6_input.
1310 if (__predict_false(le->outer.data.lro_type == LRO_TYPE_IPV6_TCP &&
1311 IN6_IS_ADDR_UNSPECIFIED(&le->outer.data.s_addr.v6)))
1312 return (TCP_LRO_CANNOT);
1314 if (__predict_false(le->inner.data.lro_type == LRO_TYPE_IPV6_TCP &&
1315 IN6_IS_ADDR_UNSPECIFIED(&le->inner.data.s_addr.v6)))
1316 return (TCP_LRO_CANNOT);
1318 /* Lookup inp, if any. */
1319 inp = tcp_lro_lookup(lc->ifp,
1320 (le->inner.data.lro_type == LRO_TYPE_NONE) ? &le->outer : &le->inner);
1322 return (TCP_LRO_CANNOT);
1324 counter_u64_add(tcp_inp_lro_locks_taken, 1);
1326 /* Get TCP control structure. */
1327 tp = intotcpcb(inp);
1329 /* Check if the inp is dead, Jim. */
1331 (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) {
1333 return (TCP_LRO_CANNOT);
1335 if ((inp->inp_irq_cpu_set == 0) && (lc->lro_cpu_is_set == 1)) {
1336 inp->inp_irq_cpu = lc->lro_last_cpu;
1337 inp->inp_irq_cpu_set = 1;
1339 /* Check if the transport doesn't support the needed optimizations. */
1340 if ((inp->inp_flags2 & (INP_SUPPORTS_MBUFQ | INP_MBUF_ACKCMP)) == 0) {
1342 return (TCP_LRO_CANNOT);
1345 if (inp->inp_flags2 & INP_MBUF_QUEUE_READY)
1346 should_wake = false;
1349 /* Check if packets should be tapped to BPF. */
1350 bpf_req = bpf_peers_present(lc->ifp->if_bpf);
1351 lagg_bpf_req = false;
1353 if (lc->ifp->if_type == IFT_IEEE8023ADLAG ||
1354 lc->ifp->if_type == IFT_INFINIBANDLAG) {
1355 struct lagg_port *lp = lc->ifp->if_lagg;
1356 struct lagg_softc *sc = lp->lp_softc;
1358 lagg_ifp = sc->sc_ifp;
1359 if (lagg_ifp != NULL)
1360 lagg_bpf_req = bpf_peers_present(lagg_ifp->if_bpf);
1363 /* Strip and compress all the incoming packets. */
1365 for (pp = &le->m_head; *pp != NULL; ) {
1367 if (do_bpf_strip_and_compress(inp, lc, le, pp,
1368 &cmp, &mv_to, &should_wake, bpf_req,
1369 lagg_bpf_req, lagg_ifp) == false) {
1370 /* Advance to next mbuf. */
1371 pp = &(*pp)->m_nextpkt;
1372 } else if (mv_to != NULL) {
1373 /* We are asked to move pp up */
1374 pp = &mv_to->m_nextpkt;
1377 /* Update "m_last_mbuf", if any. */
1378 if (pp == &le->m_head)
1379 le->m_last_mbuf = *pp;
1381 le->m_last_mbuf = __containerof(pp, struct mbuf, m_nextpkt);
1383 /* Check if any data mbufs left. */
1384 if (le->m_head != NULL) {
1385 counter_u64_add(tcp_inp_lro_direct_queue, 1);
1386 tcp_lro_log(tp, lc, le, NULL, 22, 1, inp->inp_flags2, 0, 1);
1387 tcp_queue_pkts(inp, tp, le);
1391 counter_u64_add(tcp_inp_lro_wokeup_queue, 1);
1392 if ((*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0))
1397 return (0); /* Success. */
1402 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
1404 /* Only optimise if there are multiple packets waiting. */
1411 CURVNET_SET(lc->ifp->if_vnet);
1412 error = tcp_lro_flush_tcphpts(lc, le);
1416 tcp_lro_condense(lc, le);
1417 tcp_flush_out_entry(lc, le);
1422 bzero(le, sizeof(*le));
1423 LIST_INSERT_HEAD(&lc->lro_free, le, next);
1426 #ifdef HAVE_INLINE_FLSLL
1427 #define tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
1429 static inline uint64_t
1430 tcp_lro_msb_64(uint64_t x)
1438 return (x & ~(x >> 1));
1443 * The tcp_lro_sort() routine is comparable to qsort(), except it has
1444 * a worst case complexity limit of O(MIN(N,64)*N), where N is the
1445 * number of elements to sort and 64 is the number of sequence bits
1446 * available. The algorithm is bit-slicing the 64-bit sequence number,
1447 * sorting one bit at a time from the most significant bit until the
1448 * least significant one, skipping the constant bits. This is
1449 * typically called a radix sort.
1452 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
1454 struct lro_mbuf_sort temp;
1461 /* for small arrays insertion sort is faster */
1463 for (x = 1; x < size; x++) {
1465 for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
1466 parray[y] = parray[y - 1];
1472 /* compute sequence bits which are constant */
1475 for (x = 0; x != size; x++) {
1476 ones |= parray[x].seq;
1477 zeros |= ~parray[x].seq;
1480 /* compute bits which are not constant into "ones" */
1485 /* pick the most significant bit which is not constant */
1486 ones = tcp_lro_msb_64(ones);
1489 * Move entries having cleared sequence bits to the beginning
1492 for (x = y = 0; y != size; y++) {
1494 if (parray[y].seq & ones)
1498 parray[x] = parray[y];
1503 KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
1506 tcp_lro_sort(parray, x);
1515 tcp_lro_flush_all(struct lro_ctrl *lc)
1522 /* check if no mbufs to flush */
1523 if (lc->lro_mbuf_count == 0)
1525 if (lc->lro_cpu_is_set == 0) {
1526 if (lc->lro_last_cpu == curcpu) {
1527 lc->lro_cnt_of_same_cpu++;
1528 /* Have we reached the threshold to declare a cpu? */
1529 if (lc->lro_cnt_of_same_cpu > tcp_lro_cpu_set_thresh)
1530 lc->lro_cpu_is_set = 1;
1532 lc->lro_last_cpu = curcpu;
1533 lc->lro_cnt_of_same_cpu = 0;
1536 CURVNET_SET(lc->ifp->if_vnet);
1538 /* get current time */
1539 binuptime(&lc->lro_last_queue_time);
1541 /* sort all mbufs according to stream */
1542 tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
1544 /* input data into LRO engine, stream by stream */
1546 for (x = 0; x != lc->lro_mbuf_count; x++) {
1550 mb = lc->lro_mbuf_data[x].mb;
1552 /* get sequence number, masking away the packet index */
1553 nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
1555 /* check for new stream */
1559 /* flush active streams */
1560 tcp_lro_rx_done(lc);
1563 /* add packet to LRO engine */
1564 if (tcp_lro_rx_common(lc, mb, 0, false) != 0) {
1565 /* input packet to network layer */
1566 (*lc->ifp->if_input)(lc->ifp, mb);
1573 /* flush active streams */
1574 tcp_lro_rx_done(lc);
1579 lc->lro_mbuf_count = 0;
1584 build_ack_entry(struct tcp_ackent *ae, struct tcphdr *th, struct mbuf *m,
1585 uint32_t *ts_ptr, uint16_t iptos)
1588 * Given a TCP ACK, summarize it down into the small TCP ACK
1591 ae->timestamp = m->m_pkthdr.rcv_tstmp;
1592 if (m->m_flags & M_TSTMP_LRO)
1593 ae->flags = TSTMP_LRO;
1594 else if (m->m_flags & M_TSTMP)
1595 ae->flags = TSTMP_HDWR;
1596 ae->seq = ntohl(th->th_seq);
1597 ae->ack = ntohl(th->th_ack);
1598 ae->flags |= tcp_get_flags(th);
1599 if (ts_ptr != NULL) {
1600 ae->ts_value = ntohl(ts_ptr[1]);
1601 ae->ts_echo = ntohl(ts_ptr[2]);
1602 ae->flags |= HAS_TSTMP;
1604 ae->win = ntohs(th->th_win);
1605 ae->codepoint = iptos;
1609 * Do BPF tap for either ACK_CMP packets or MBUF QUEUE type packets
1610 * and strip all, but the IPv4/IPv6 header.
1613 do_bpf_strip_and_compress(struct inpcb *inp, struct lro_ctrl *lc,
1614 struct lro_entry *le, struct mbuf **pp, struct mbuf **cmp, struct mbuf **mv_to,
1615 bool *should_wake, bool bpf_req, bool lagg_bpf_req, struct ifnet *lagg_ifp)
1620 struct ip6_hdr *ip6;
1625 struct tcp_ackent *ack_ent;
1628 bool other_opts, can_compress;
1634 /* Get current mbuf. */
1637 /* Let the BPF see the packet */
1638 if (__predict_false(bpf_req))
1639 ETHER_BPF_MTAP(lc->ifp, m);
1641 if (__predict_false(lagg_bpf_req))
1642 ETHER_BPF_MTAP(lagg_ifp, m);
1644 tcp_hdr_offset = m->m_pkthdr.lro_tcp_h_off;
1645 lro_type = le->inner.data.lro_type;
1648 lro_type = le->outer.data.lro_type;
1650 case LRO_TYPE_IPV4_TCP:
1651 tcp_hdr_offset -= sizeof(*le->outer.ip4);
1652 m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1654 case LRO_TYPE_IPV6_TCP:
1655 tcp_hdr_offset -= sizeof(*le->outer.ip6);
1656 m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1662 case LRO_TYPE_IPV4_TCP:
1663 tcp_hdr_offset -= sizeof(*le->outer.ip4);
1664 m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1666 case LRO_TYPE_IPV6_TCP:
1667 tcp_hdr_offset -= sizeof(*le->outer.ip6);
1668 m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1674 MPASS(tcp_hdr_offset >= 0);
1676 m_adj(m, tcp_hdr_offset);
1677 m->m_flags |= M_LRO_EHDRSTRP;
1678 m->m_flags &= ~M_ACKCMP;
1679 m->m_pkthdr.lro_tcp_h_off -= tcp_hdr_offset;
1681 th = tcp_lro_get_th(m);
1683 th->th_sum = 0; /* TCP checksum is valid. */
1685 /* Check if ACK can be compressed */
1686 can_compress = tcp_lro_ack_valid(m, th, &ts_ptr, &other_opts);
1688 /* Now lets look at the should wake states */
1689 if ((other_opts == true) &&
1690 ((inp->inp_flags2 & INP_DONT_SACK_QUEUE) == 0)) {
1692 * If there are other options (SACK?) and the
1693 * tcp endpoint has not expressly told us it does
1694 * not care about SACKS, then we should wake up.
1696 *should_wake = true;
1698 /* Is the ack compressable? */
1699 if (can_compress == false)
1701 /* Does the TCP endpoint support ACK compression? */
1702 if ((inp->inp_flags2 & INP_MBUF_ACKCMP) == 0)
1705 /* Lets get the TOS/traffic class field */
1706 l3.ptr = mtod(m, void *);
1708 case LRO_TYPE_IPV4_TCP:
1709 iptos = l3.ip4->ip_tos;
1711 case LRO_TYPE_IPV6_TCP:
1712 iptos = IPV6_TRAFFIC_CLASS(l3.ip6);
1715 iptos = 0; /* Keep compiler happy. */
1718 /* Now lets get space if we don't have some already */
1721 nm = tcp_lro_get_last_if_ackcmp(lc, le, inp, &n_mbuf);
1722 if (__predict_false(nm == NULL))
1727 * Link in the new cmp ack to our in-order place,
1728 * first set our cmp ack's next to where we are.
1733 * Set it up so mv_to is advanced to our
1734 * compressed ack. This way the caller can
1735 * advance pp to the right place.
1739 * Advance it here locally as well.
1741 pp = &nm->m_nextpkt;
1744 /* We have one already we are working on */
1746 if (M_TRAILINGSPACE(nm) < sizeof(struct tcp_ackent)) {
1747 /* We ran out of space */
1748 inp->inp_flags2 |= INP_MBUF_L_ACKS;
1752 MPASS(M_TRAILINGSPACE(nm) >= sizeof(struct tcp_ackent));
1753 counter_u64_add(tcp_inp_lro_compressed, 1);
1755 /* We can add in to the one on the tail */
1756 ack_ent = mtod(nm, struct tcp_ackent *);
1757 idx = (nm->m_len / sizeof(struct tcp_ackent));
1758 build_ack_entry(&ack_ent[idx], th, m, ts_ptr, iptos);
1760 /* Bump the size of both pkt-hdr and len */
1761 nm->m_len += sizeof(struct tcp_ackent);
1762 nm->m_pkthdr.len += sizeof(struct tcp_ackent);
1764 /* Advance to next mbuf before freeing. */
1766 m->m_nextpkt = NULL;
1770 counter_u64_add(tcp_uncomp_total, 1);
1776 static struct lro_head *
1777 tcp_lro_rx_get_bucket(struct lro_ctrl *lc, struct mbuf *m, struct lro_parser *parser)
1781 if (M_HASHTYPE_ISHASH(m)) {
1782 hash = m->m_pkthdr.flowid;
1784 for (unsigned i = hash = 0; i != LRO_RAW_ADDRESS_MAX; i++)
1785 hash += parser->data.raw[i];
1787 return (&lc->lro_hash[hash % lc->lro_hashsz]);
1791 tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash)
1793 struct lro_parser pi; /* inner address data */
1794 struct lro_parser po; /* outer address data */
1795 struct lro_parser *pa; /* current parser for TCP stream */
1796 struct lro_entry *le;
1797 struct lro_head *bucket;
1802 uint16_t tcp_data_sum;
1805 /* Quickly decide if packet cannot be LRO'ed */
1806 if (__predict_false(V_ipforwarding != 0))
1807 return (TCP_LRO_CANNOT);
1810 /* Quickly decide if packet cannot be LRO'ed */
1811 if (__predict_false(V_ip6_forwarding != 0))
1812 return (TCP_LRO_CANNOT);
1814 if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1815 ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1816 (m->m_pkthdr.csum_data != 0xffff)) {
1818 * The checksum either did not have hardware offload
1819 * or it was a bad checksum. We can't LRO such
1822 counter_u64_add(tcp_bad_csums, 1);
1823 return (TCP_LRO_CANNOT);
1825 /* We expect a contiguous header [eh, ip, tcp]. */
1826 pa = tcp_lro_parser(m, &po, &pi, true);
1827 if (__predict_false(pa == NULL))
1828 return (TCP_LRO_NOT_SUPPORTED);
1830 /* We don't expect any padding. */
1831 error = tcp_lro_trim_mbuf_chain(m, pa);
1832 if (__predict_false(error != 0))
1836 switch (pa->data.lro_type) {
1837 case LRO_TYPE_IPV4_TCP:
1838 error = tcp_lro_rx_ipv4(lc, m, pa->ip4);
1839 if (__predict_false(error != 0))
1846 /* If no hardware or arrival stamp on the packet add timestamp */
1847 if ((m->m_flags & (M_TSTMP_LRO | M_TSTMP)) == 0) {
1848 m->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
1849 m->m_flags |= M_TSTMP_LRO;
1852 /* Get pointer to TCP header. */
1855 /* Don't process SYN packets. */
1856 if (__predict_false(tcp_get_flags(th) & TH_SYN))
1857 return (TCP_LRO_CANNOT);
1859 /* Get total TCP header length and compute payload length. */
1860 tcp_opt_len = (th->th_off << 2);
1861 tcp_data_len = m->m_pkthdr.len - ((uint8_t *)th -
1862 (uint8_t *)m->m_data) - tcp_opt_len;
1863 tcp_opt_len -= sizeof(*th);
1865 /* Don't process invalid TCP headers. */
1866 if (__predict_false(tcp_opt_len < 0 || tcp_data_len < 0))
1867 return (TCP_LRO_CANNOT);
1869 /* Compute TCP data only checksum. */
1870 if (tcp_data_len == 0)
1871 tcp_data_sum = 0; /* no data, no checksum */
1872 else if (__predict_false(csum != 0))
1873 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~csum);
1875 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~th->th_sum);
1877 /* Save TCP info in mbuf. */
1878 m->m_nextpkt = NULL;
1879 m->m_pkthdr.rcvif = lc->ifp;
1880 m->m_pkthdr.lro_tcp_d_csum = tcp_data_sum;
1881 m->m_pkthdr.lro_tcp_d_len = tcp_data_len;
1882 m->m_pkthdr.lro_tcp_h_off = ((uint8_t *)th - (uint8_t *)m->m_data);
1883 m->m_pkthdr.lro_nsegs = 1;
1885 /* Get hash bucket. */
1887 bucket = &lc->lro_hash[0];
1889 bucket = tcp_lro_rx_get_bucket(lc, m, pa);
1892 /* Try to find a matching previous segment. */
1893 LIST_FOREACH(le, bucket, hash_next) {
1894 /* Compare addresses and ports. */
1895 if (lro_address_compare(&po.data, &le->outer.data) == false ||
1896 lro_address_compare(&pi.data, &le->inner.data) == false)
1899 /* Check if no data and old ACK. */
1900 if (tcp_data_len == 0 &&
1901 SEQ_LT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1906 /* Mark "m" in the last spot. */
1907 le->m_last_mbuf->m_nextpkt = m;
1908 /* Now set the tail to "m". */
1909 le->m_last_mbuf = m;
1913 /* Try to find an empty slot. */
1914 if (LIST_EMPTY(&lc->lro_free))
1915 return (TCP_LRO_NO_ENTRIES);
1917 /* Start a new segment chain. */
1918 le = LIST_FIRST(&lc->lro_free);
1919 LIST_REMOVE(le, next);
1920 tcp_lro_active_insert(lc, bucket, le);
1922 /* Make sure the headers are set. */
1926 /* Store time this entry was allocated. */
1927 le->alloc_time = lc->lro_last_queue_time;
1929 tcp_set_entry_to_mbuf(lc, le, m, th);
1931 /* Now set the tail to "m". */
1932 le->m_last_mbuf = m;
1938 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
1942 if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1943 ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1944 (m->m_pkthdr.csum_data != 0xffff)) {
1946 * The checksum either did not have hardware offload
1947 * or it was a bad checksum. We can't LRO such
1950 counter_u64_add(tcp_bad_csums, 1);
1951 return (TCP_LRO_CANNOT);
1953 /* get current time */
1954 binuptime(&lc->lro_last_queue_time);
1955 CURVNET_SET(lc->ifp->if_vnet);
1956 error = tcp_lro_rx_common(lc, m, csum, true);
1963 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
1967 if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
1968 lc->lro_mbuf_max == 0)) {
1974 /* check if packet is not LRO capable */
1975 if (__predict_false((lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
1976 /* input packet to network layer */
1977 (*lc->ifp->if_input) (lc->ifp, mb);
1981 /* create sequence number */
1982 lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
1983 (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
1984 (((uint64_t)mb->m_pkthdr.flowid) << 24) |
1985 ((uint64_t)lc->lro_mbuf_count);
1988 lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb;
1990 /* flush if array is full */
1991 if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max))
1992 tcp_lro_flush_all(lc);