]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/netinet/tcp_lro.c
tcp: Rack can crash with the new non-TSO fix..
[FreeBSD/FreeBSD.git] / sys / netinet / tcp_lro.c
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
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.
8  * All rights reserved.
9  *
10  * Portions of this software were developed by Bjoern Zeeb
11  * under sponsorship from the FreeBSD Foundation.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
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.
21  *
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
32  * SUCH DAMAGE.
33  */
34
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37
38 #include "opt_inet.h"
39 #include "opt_inet6.h"
40
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/malloc.h>
45 #include <sys/mbuf.h>
46 #include <sys/socket.h>
47 #include <sys/socketvar.h>
48 #include <sys/sockbuf.h>
49 #include <sys/sysctl.h>
50
51 #include <net/if.h>
52 #include <net/if_var.h>
53 #include <net/ethernet.h>
54 #include <net/bpf.h>
55 #include <net/vnet.h>
56 #include <net/if_dl.h>
57 #include <net/if_media.h>
58 #include <net/if_private.h>
59 #include <net/if_types.h>
60 #include <net/infiniband.h>
61 #include <net/if_lagg.h>
62
63 #include <netinet/in_systm.h>
64 #include <netinet/in.h>
65 #include <netinet/ip6.h>
66 #include <netinet/ip.h>
67 #include <netinet/ip_var.h>
68 #include <netinet/in_pcb.h>
69 #include <netinet6/in6_pcb.h>
70 #include <netinet/tcp.h>
71 #include <netinet/tcp_seq.h>
72 #include <netinet/tcp_lro.h>
73 #include <netinet/tcp_var.h>
74 #include <netinet/tcpip.h>
75 #include <netinet/tcp_hpts.h>
76 #include <netinet/tcp_log_buf.h>
77 #include <netinet/tcp_fsm.h>
78 #include <netinet/udp.h>
79 #include <netinet6/ip6_var.h>
80
81 #include <machine/in_cksum.h>
82
83 static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures");
84
85 #define TCP_LRO_TS_OPTION \
86     ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
87           (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
88
89 static void     tcp_lro_rx_done(struct lro_ctrl *lc);
90 static int      tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m,
91                     uint32_t csum, bool use_hash);
92
93 #ifdef TCPHPTS
94 static bool     do_bpf_strip_and_compress(struct inpcb *, struct lro_ctrl *,
95                 struct lro_entry *, struct mbuf **, struct mbuf **, struct mbuf **,
96                 bool *, bool, bool, struct ifnet *, bool);
97
98 #endif
99
100 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro,  CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
101     "TCP LRO");
102
103 static long tcplro_stacks_wanting_mbufq;
104 counter_u64_t tcp_inp_lro_direct_queue;
105 counter_u64_t tcp_inp_lro_wokeup_queue;
106 counter_u64_t tcp_inp_lro_compressed;
107 counter_u64_t tcp_inp_lro_locks_taken;
108 counter_u64_t tcp_extra_mbuf;
109 counter_u64_t tcp_would_have_but;
110 counter_u64_t tcp_comp_total;
111 counter_u64_t tcp_uncomp_total;
112 counter_u64_t tcp_bad_csums;
113
114 static unsigned tcp_lro_entries = TCP_LRO_ENTRIES;
115 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries,
116     CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_entries, 0,
117     "default number of LRO entries");
118
119 static uint32_t tcp_lro_cpu_set_thresh = TCP_LRO_CPU_DECLARATION_THRESH;
120 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_cpu_threshold,
121     CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_cpu_set_thresh, 0,
122     "Number of interrupts in a row on the same CPU that will make us declare an 'affinity' cpu?");
123
124 static uint32_t tcp_less_accurate_lro_ts = 0;
125 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_less_accurate,
126     CTLFLAG_MPSAFE, &tcp_less_accurate_lro_ts, 0,
127     "Do we trade off efficency by doing less timestamp operations for time accuracy?");
128
129 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, fullqueue, CTLFLAG_RD,
130     &tcp_inp_lro_direct_queue, "Number of lro's fully queued to transport");
131 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, wokeup, CTLFLAG_RD,
132     &tcp_inp_lro_wokeup_queue, "Number of lro's where we woke up transport via hpts");
133 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, compressed, CTLFLAG_RD,
134     &tcp_inp_lro_compressed, "Number of lro's compressed and sent to transport");
135 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lockcnt, CTLFLAG_RD,
136     &tcp_inp_lro_locks_taken, "Number of lro's inp_wlocks taken");
137 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, extra_mbuf, CTLFLAG_RD,
138     &tcp_extra_mbuf, "Number of times we had an extra compressed ack dropped into the tp");
139 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, would_have_but, CTLFLAG_RD,
140     &tcp_would_have_but, "Number of times we would have had an extra compressed, but mget failed");
141 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, with_m_ackcmp, CTLFLAG_RD,
142     &tcp_comp_total, "Number of mbufs queued with M_ACKCMP flags set");
143 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, without_m_ackcmp, CTLFLAG_RD,
144     &tcp_uncomp_total, "Number of mbufs queued without M_ACKCMP");
145 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lro_badcsum, CTLFLAG_RD,
146     &tcp_bad_csums, "Number of packets that the common code saw with bad csums");
147
148 void
149 tcp_lro_reg_mbufq(void)
150 {
151         atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, 1);
152 }
153
154 void
155 tcp_lro_dereg_mbufq(void)
156 {
157         atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, -1);
158 }
159
160 static __inline void
161 tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket,
162     struct lro_entry *le)
163 {
164
165         LIST_INSERT_HEAD(&lc->lro_active, le, next);
166         LIST_INSERT_HEAD(bucket, le, hash_next);
167 }
168
169 static __inline void
170 tcp_lro_active_remove(struct lro_entry *le)
171 {
172
173         LIST_REMOVE(le, next);          /* active list */
174         LIST_REMOVE(le, hash_next);     /* hash bucket */
175 }
176
177 int
178 tcp_lro_init(struct lro_ctrl *lc)
179 {
180         return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0));
181 }
182
183 int
184 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
185     unsigned lro_entries, unsigned lro_mbufs)
186 {
187         struct lro_entry *le;
188         size_t size;
189         unsigned i, elements;
190
191         lc->lro_bad_csum = 0;
192         lc->lro_queued = 0;
193         lc->lro_flushed = 0;
194         lc->lro_mbuf_count = 0;
195         lc->lro_mbuf_max = lro_mbufs;
196         lc->lro_cnt = lro_entries;
197         lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
198         lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
199         lc->ifp = ifp;
200         LIST_INIT(&lc->lro_free);
201         LIST_INIT(&lc->lro_active);
202
203         /* create hash table to accelerate entry lookup */
204         if (lro_entries > lro_mbufs)
205                 elements = lro_entries;
206         else
207                 elements = lro_mbufs;
208         lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz,
209             HASH_NOWAIT);
210         if (lc->lro_hash == NULL) {
211                 memset(lc, 0, sizeof(*lc));
212                 return (ENOMEM);
213         }
214
215         /* compute size to allocate */
216         size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
217             (lro_entries * sizeof(*le));
218         lc->lro_mbuf_data = (struct lro_mbuf_sort *)
219             malloc(size, M_LRO, M_NOWAIT | M_ZERO);
220
221         /* check for out of memory */
222         if (lc->lro_mbuf_data == NULL) {
223                 free(lc->lro_hash, M_LRO);
224                 memset(lc, 0, sizeof(*lc));
225                 return (ENOMEM);
226         }
227         /* compute offset for LRO entries */
228         le = (struct lro_entry *)
229             (lc->lro_mbuf_data + lro_mbufs);
230
231         /* setup linked list */
232         for (i = 0; i != lro_entries; i++)
233                 LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
234
235         return (0);
236 }
237
238 struct vxlan_header {
239         uint32_t        vxlh_flags;
240         uint32_t        vxlh_vni;
241 };
242
243 static inline void *
244 tcp_lro_low_level_parser(void *ptr, struct lro_parser *parser, bool update_data, bool is_vxlan, int mlen)
245 {
246         const struct ether_vlan_header *eh;
247         void *old;
248         uint16_t eth_type;
249
250         if (update_data)
251                 memset(parser, 0, sizeof(*parser));
252
253         old = ptr;
254
255         if (is_vxlan) {
256                 const struct vxlan_header *vxh;
257                 vxh = ptr;
258                 ptr = (uint8_t *)ptr + sizeof(*vxh);
259                 if (update_data) {
260                         parser->data.vxlan_vni =
261                             vxh->vxlh_vni & htonl(0xffffff00);
262                 }
263         }
264
265         eh = ptr;
266         if (__predict_false(eh->evl_encap_proto == htons(ETHERTYPE_VLAN))) {
267                 eth_type = eh->evl_proto;
268                 if (update_data) {
269                         /* strip priority and keep VLAN ID only */
270                         parser->data.vlan_id = eh->evl_tag & htons(EVL_VLID_MASK);
271                 }
272                 /* advance to next header */
273                 ptr = (uint8_t *)ptr + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
274                 mlen -= (ETHER_HDR_LEN  + ETHER_VLAN_ENCAP_LEN);
275         } else {
276                 eth_type = eh->evl_encap_proto;
277                 /* advance to next header */
278                 mlen -= ETHER_HDR_LEN;
279                 ptr = (uint8_t *)ptr + ETHER_HDR_LEN;
280         }
281         if (__predict_false(mlen <= 0))
282                 return (NULL);
283         switch (eth_type) {
284 #ifdef INET
285         case htons(ETHERTYPE_IP):
286                 parser->ip4 = ptr;
287                 if (__predict_false(mlen < sizeof(struct ip)))
288                         return (NULL);
289                 /* Ensure there are no IPv4 options. */
290                 if ((parser->ip4->ip_hl << 2) != sizeof (*parser->ip4))
291                         break;
292                 /* .. and the packet is not fragmented. */
293                 if (parser->ip4->ip_off & htons(IP_MF|IP_OFFMASK))
294                         break;
295                 /* .. and the packet has valid src/dst addrs */
296                 if (__predict_false(parser->ip4->ip_src.s_addr == INADDR_ANY ||
297                         parser->ip4->ip_dst.s_addr == INADDR_ANY))
298                         break;
299                 ptr = (uint8_t *)ptr + (parser->ip4->ip_hl << 2);
300                 mlen -= sizeof(struct ip);
301                 if (update_data) {
302                         parser->data.s_addr.v4 = parser->ip4->ip_src;
303                         parser->data.d_addr.v4 = parser->ip4->ip_dst;
304                 }
305                 switch (parser->ip4->ip_p) {
306                 case IPPROTO_UDP:
307                         if (__predict_false(mlen < sizeof(struct udphdr)))
308                                 return (NULL);
309                         parser->udp = ptr;
310                         if (update_data) {
311                                 parser->data.lro_type = LRO_TYPE_IPV4_UDP;
312                                 parser->data.s_port = parser->udp->uh_sport;
313                                 parser->data.d_port = parser->udp->uh_dport;
314                         } else {
315                                 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_UDP);
316                         }
317                         ptr = ((uint8_t *)ptr + sizeof(*parser->udp));
318                         parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
319                         return (ptr);
320                 case IPPROTO_TCP:
321                         parser->tcp = ptr;
322                         if (__predict_false(mlen < sizeof(struct tcphdr)))
323                                 return (NULL);
324                         if (update_data) {
325                                 parser->data.lro_type = LRO_TYPE_IPV4_TCP;
326                                 parser->data.s_port = parser->tcp->th_sport;
327                                 parser->data.d_port = parser->tcp->th_dport;
328                         } else {
329                                 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_TCP);
330                         }
331                         if (__predict_false(mlen < (parser->tcp->th_off << 2)))
332                                 return (NULL);
333                         ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
334                         parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
335                         return (ptr);
336                 default:
337                         break;
338                 }
339                 break;
340 #endif
341 #ifdef INET6
342         case htons(ETHERTYPE_IPV6):
343                 parser->ip6 = ptr;
344                 if (__predict_false(mlen < sizeof(struct ip6_hdr)))
345                         return (NULL);
346                 /* Ensure the packet has valid src/dst addrs */
347                 if (__predict_false(IN6_IS_ADDR_UNSPECIFIED(&parser->ip6->ip6_src) ||
348                         IN6_IS_ADDR_UNSPECIFIED(&parser->ip6->ip6_dst)))
349                         return (NULL);
350                 ptr = (uint8_t *)ptr + sizeof(*parser->ip6);
351                 if (update_data) {
352                         parser->data.s_addr.v6 = parser->ip6->ip6_src;
353                         parser->data.d_addr.v6 = parser->ip6->ip6_dst;
354                 }
355                 mlen -= sizeof(struct ip6_hdr);
356                 switch (parser->ip6->ip6_nxt) {
357                 case IPPROTO_UDP:
358                         if (__predict_false(mlen < sizeof(struct udphdr)))
359                                 return (NULL);
360                         parser->udp = ptr;
361                         if (update_data) {
362                                 parser->data.lro_type = LRO_TYPE_IPV6_UDP;
363                                 parser->data.s_port = parser->udp->uh_sport;
364                                 parser->data.d_port = parser->udp->uh_dport;
365                         } else {
366                                 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_UDP);
367                         }
368                         ptr = (uint8_t *)ptr + sizeof(*parser->udp);
369                         parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
370                         return (ptr);
371                 case IPPROTO_TCP:
372                         if (__predict_false(mlen < sizeof(struct tcphdr)))
373                                 return (NULL);
374                         parser->tcp = ptr;
375                         if (update_data) {
376                                 parser->data.lro_type = LRO_TYPE_IPV6_TCP;
377                                 parser->data.s_port = parser->tcp->th_sport;
378                                 parser->data.d_port = parser->tcp->th_dport;
379                         } else {
380                                 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_TCP);
381                         }
382                         if (__predict_false(mlen < (parser->tcp->th_off << 2)))
383                                 return (NULL);
384                         ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
385                         parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
386                         return (ptr);
387                 default:
388                         break;
389                 }
390                 break;
391 #endif
392         default:
393                 break;
394         }
395         /* Invalid packet - cannot parse */
396         return (NULL);
397 }
398
399 static const int vxlan_csum = CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
400     CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID;
401
402 static inline struct lro_parser *
403 tcp_lro_parser(struct mbuf *m, struct lro_parser *po, struct lro_parser *pi, bool update_data)
404 {
405         void *data_ptr;
406
407         /* Try to parse outer headers first. */
408         data_ptr = tcp_lro_low_level_parser(m->m_data, po, update_data, false, m->m_len);
409         if (data_ptr == NULL || po->total_hdr_len > m->m_len)
410                 return (NULL);
411
412         if (update_data) {
413                 /* Store VLAN ID, if any. */
414                 if (__predict_false(m->m_flags & M_VLANTAG)) {
415                         po->data.vlan_id =
416                             htons(m->m_pkthdr.ether_vtag) & htons(EVL_VLID_MASK);
417                 }
418                 /* Store decrypted flag, if any. */
419                 if (__predict_false((m->m_pkthdr.csum_flags &
420                     CSUM_TLS_MASK) == CSUM_TLS_DECRYPTED))
421                         po->data.lro_flags |= LRO_FLAG_DECRYPTED;
422         }
423
424         switch (po->data.lro_type) {
425         case LRO_TYPE_IPV4_UDP:
426         case LRO_TYPE_IPV6_UDP:
427                 /* Check for VXLAN headers. */
428                 if ((m->m_pkthdr.csum_flags & vxlan_csum) != vxlan_csum)
429                         break;
430
431                 /* Try to parse inner headers. */
432                 data_ptr = tcp_lro_low_level_parser(data_ptr, pi, update_data, true,
433                                                     (m->m_len - ((caddr_t)data_ptr - m->m_data)));
434                 if (data_ptr == NULL || (pi->total_hdr_len + po->total_hdr_len) > m->m_len)
435                         break;
436
437                 /* Verify supported header types. */
438                 switch (pi->data.lro_type) {
439                 case LRO_TYPE_IPV4_TCP:
440                 case LRO_TYPE_IPV6_TCP:
441                         return (pi);
442                 default:
443                         break;
444                 }
445                 break;
446         case LRO_TYPE_IPV4_TCP:
447         case LRO_TYPE_IPV6_TCP:
448                 if (update_data)
449                         memset(pi, 0, sizeof(*pi));
450                 return (po);
451         default:
452                 break;
453         }
454         return (NULL);
455 }
456
457 static inline int
458 tcp_lro_trim_mbuf_chain(struct mbuf *m, const struct lro_parser *po)
459 {
460         int len;
461
462         switch (po->data.lro_type) {
463 #ifdef INET
464         case LRO_TYPE_IPV4_TCP:
465                 len = ((uint8_t *)po->ip4 - (uint8_t *)m->m_data) +
466                     ntohs(po->ip4->ip_len);
467                 break;
468 #endif
469 #ifdef INET6
470         case LRO_TYPE_IPV6_TCP:
471                 len = ((uint8_t *)po->ip6 - (uint8_t *)m->m_data) +
472                     ntohs(po->ip6->ip6_plen) + sizeof(*po->ip6);
473                 break;
474 #endif
475         default:
476                 return (TCP_LRO_CANNOT);
477         }
478
479         /*
480          * If the frame is padded beyond the end of the IP packet,
481          * then trim the extra bytes off:
482          */
483         if (__predict_true(m->m_pkthdr.len == len)) {
484                 return (0);
485         } else if (m->m_pkthdr.len > len) {
486                 m_adj(m, len - m->m_pkthdr.len);
487                 return (0);
488         }
489         return (TCP_LRO_CANNOT);
490 }
491
492 static struct tcphdr *
493 tcp_lro_get_th(struct mbuf *m)
494 {
495         return ((struct tcphdr *)((uint8_t *)m->m_data + m->m_pkthdr.lro_tcp_h_off));
496 }
497
498 static void
499 lro_free_mbuf_chain(struct mbuf *m)
500 {
501         struct mbuf *save;
502
503         while (m) {
504                 save = m->m_nextpkt;
505                 m->m_nextpkt = NULL;
506                 m_freem(m);
507                 m = save;
508         }
509 }
510
511 void
512 tcp_lro_free(struct lro_ctrl *lc)
513 {
514         struct lro_entry *le;
515         unsigned x;
516
517         /* reset LRO free list */
518         LIST_INIT(&lc->lro_free);
519
520         /* free active mbufs, if any */
521         while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
522                 tcp_lro_active_remove(le);
523                 lro_free_mbuf_chain(le->m_head);
524         }
525
526         /* free hash table */
527         free(lc->lro_hash, M_LRO);
528         lc->lro_hash = NULL;
529         lc->lro_hashsz = 0;
530
531         /* free mbuf array, if any */
532         for (x = 0; x != lc->lro_mbuf_count; x++)
533                 m_freem(lc->lro_mbuf_data[x].mb);
534         lc->lro_mbuf_count = 0;
535
536         /* free allocated memory, if any */
537         free(lc->lro_mbuf_data, M_LRO);
538         lc->lro_mbuf_data = NULL;
539 }
540
541 static uint16_t
542 tcp_lro_rx_csum_tcphdr(const struct tcphdr *th)
543 {
544         const uint16_t *ptr;
545         uint32_t csum;
546         uint16_t len;
547
548         csum = -th->th_sum;     /* exclude checksum field */
549         len = th->th_off;
550         ptr = (const uint16_t *)th;
551         while (len--) {
552                 csum += *ptr;
553                 ptr++;
554                 csum += *ptr;
555                 ptr++;
556         }
557         while (csum > 0xffff)
558                 csum = (csum >> 16) + (csum & 0xffff);
559
560         return (csum);
561 }
562
563 static uint16_t
564 tcp_lro_rx_csum_data(const struct lro_parser *pa, uint16_t tcp_csum)
565 {
566         uint32_t c;
567         uint16_t cs;
568
569         c = tcp_csum;
570
571         switch (pa->data.lro_type) {
572 #ifdef INET6
573         case LRO_TYPE_IPV6_TCP:
574                 /* Compute full pseudo IPv6 header checksum. */
575                 cs = in6_cksum_pseudo(pa->ip6, ntohs(pa->ip6->ip6_plen), pa->ip6->ip6_nxt, 0);
576                 break;
577 #endif
578 #ifdef INET
579         case LRO_TYPE_IPV4_TCP:
580                 /* Compute full pseudo IPv4 header checsum. */
581                 cs = in_addword(ntohs(pa->ip4->ip_len) - sizeof(*pa->ip4), IPPROTO_TCP);
582                 cs = in_pseudo(pa->ip4->ip_src.s_addr, pa->ip4->ip_dst.s_addr, htons(cs));
583                 break;
584 #endif
585         default:
586                 cs = 0;         /* Keep compiler happy. */
587                 break;
588         }
589
590         /* Complement checksum. */
591         cs = ~cs;
592         c += cs;
593
594         /* Remove TCP header checksum. */
595         cs = ~tcp_lro_rx_csum_tcphdr(pa->tcp);
596         c += cs;
597
598         /* Compute checksum remainder. */
599         while (c > 0xffff)
600                 c = (c >> 16) + (c & 0xffff);
601
602         return (c);
603 }
604
605 static void
606 tcp_lro_rx_done(struct lro_ctrl *lc)
607 {
608         struct lro_entry *le;
609
610         while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
611                 tcp_lro_active_remove(le);
612                 tcp_lro_flush(lc, le);
613         }
614 }
615
616 static void
617 tcp_lro_flush_active(struct lro_ctrl *lc)
618 {
619         struct lro_entry *le;
620
621         /*
622          * Walk through the list of le entries, and
623          * any one that does have packets flush. This
624          * is called because we have an inbound packet
625          * (e.g. SYN) that has to have all others flushed
626          * in front of it. Note we have to do the remove
627          * because tcp_lro_flush() assumes that the entry
628          * is being freed. This is ok it will just get
629          * reallocated again like it was new.
630          */
631         LIST_FOREACH(le, &lc->lro_active, next) {
632                 if (le->m_head != NULL) {
633                         tcp_lro_active_remove(le);
634                         tcp_lro_flush(lc, le);
635                 }
636         }
637 }
638
639 void
640 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
641 {
642         struct lro_entry *le, *le_tmp;
643         uint64_t now, tov;
644         struct bintime bt;
645
646         NET_EPOCH_ASSERT();
647         if (LIST_EMPTY(&lc->lro_active))
648                 return;
649
650         /* get timeout time and current time in ns */
651         binuptime(&bt);
652         now = bintime2ns(&bt);
653         tov = ((timeout->tv_sec * 1000000000) + (timeout->tv_usec * 1000));
654         LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
655                 if (now >= (bintime2ns(&le->alloc_time) + tov)) {
656                         tcp_lro_active_remove(le);
657                         tcp_lro_flush(lc, le);
658                 }
659         }
660 }
661
662 #ifdef INET
663 static int
664 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4)
665 {
666         uint16_t csum;
667
668         /* Legacy IP has a header checksum that needs to be correct. */
669         if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) {
670                 if (__predict_false((m->m_pkthdr.csum_flags & CSUM_IP_VALID) == 0)) {
671                         lc->lro_bad_csum++;
672                         return (TCP_LRO_CANNOT);
673                 }
674         } else {
675                 csum = in_cksum_hdr(ip4);
676                 if (__predict_false(csum != 0)) {
677                         lc->lro_bad_csum++;
678                         return (TCP_LRO_CANNOT);
679                 }
680         }
681         return (0);
682 }
683 #endif
684
685 #ifdef TCPHPTS
686 static void
687 tcp_lro_log(struct tcpcb *tp, const struct lro_ctrl *lc,
688     const struct lro_entry *le, const struct mbuf *m,
689     int frm, int32_t tcp_data_len, uint32_t th_seq,
690     uint32_t th_ack, uint16_t th_win)
691 {
692         if (tcp_bblogging_on(tp)) {
693                 union tcp_log_stackspecific log;
694                 struct timeval tv, btv;
695                 uint32_t cts;
696
697                 cts = tcp_get_usecs(&tv);
698                 memset(&log, 0, sizeof(union tcp_log_stackspecific));
699                 log.u_bbr.flex8 = frm;
700                 log.u_bbr.flex1 = tcp_data_len;
701                 if (m)
702                         log.u_bbr.flex2 = m->m_pkthdr.len;
703                 else
704                         log.u_bbr.flex2 = 0;
705                 if (le->m_head) {
706                         log.u_bbr.flex3 = le->m_head->m_pkthdr.lro_nsegs;
707                         log.u_bbr.flex4 = le->m_head->m_pkthdr.lro_tcp_d_len;
708                         log.u_bbr.flex5 = le->m_head->m_pkthdr.len;
709                         log.u_bbr.delRate = le->m_head->m_flags;
710                         log.u_bbr.rttProp = le->m_head->m_pkthdr.rcv_tstmp;
711                 }
712                 log.u_bbr.inflight = th_seq;
713                 log.u_bbr.delivered = th_ack;
714                 log.u_bbr.timeStamp = cts;
715                 log.u_bbr.epoch = le->next_seq;
716                 log.u_bbr.lt_epoch = le->ack_seq;
717                 log.u_bbr.pacing_gain = th_win;
718                 log.u_bbr.cwnd_gain = le->window;
719                 log.u_bbr.lost = curcpu;
720                 log.u_bbr.cur_del_rate = (uintptr_t)m;
721                 log.u_bbr.bw_inuse = (uintptr_t)le->m_head;
722                 bintime2timeval(&lc->lro_last_queue_time, &btv);
723                 log.u_bbr.flex6 = tcp_tv_to_usectick(&btv);
724                 log.u_bbr.flex7 = le->compressed;
725                 log.u_bbr.pacing_gain = le->uncompressed;
726                 if (in_epoch(net_epoch_preempt))
727                         log.u_bbr.inhpts = 1;
728                 else
729                         log.u_bbr.inhpts = 0;
730                 TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
731                     &tptosocket(tp)->so_snd,
732                     TCP_LOG_LRO, 0, 0, &log, false, &tv);
733         }
734 }
735 #endif
736
737 static inline void
738 tcp_lro_assign_and_checksum_16(uint16_t *ptr, uint16_t value, uint16_t *psum)
739 {
740         uint32_t csum;
741
742         csum = 0xffff - *ptr + value;
743         while (csum > 0xffff)
744                 csum = (csum >> 16) + (csum & 0xffff);
745         *ptr = value;
746         *psum = csum;
747 }
748
749 static uint16_t
750 tcp_lro_update_checksum(const struct lro_parser *pa, const struct lro_entry *le,
751     uint16_t payload_len, uint16_t delta_sum)
752 {
753         uint32_t csum;
754         uint16_t tlen;
755         uint16_t temp[5] = {};
756
757         switch (pa->data.lro_type) {
758         case LRO_TYPE_IPV4_TCP:
759                 /* Compute new IPv4 length. */
760                 tlen = (pa->ip4->ip_hl << 2) + (pa->tcp->th_off << 2) + payload_len;
761                 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
762
763                 /* Subtract delta from current IPv4 checksum. */
764                 csum = pa->ip4->ip_sum + 0xffff - temp[0];
765                 while (csum > 0xffff)
766                         csum = (csum >> 16) + (csum & 0xffff);
767                 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
768                 goto update_tcp_header;
769
770         case LRO_TYPE_IPV6_TCP:
771                 /* Compute new IPv6 length. */
772                 tlen = (pa->tcp->th_off << 2) + payload_len;
773                 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
774                 goto update_tcp_header;
775
776         case LRO_TYPE_IPV4_UDP:
777                 /* Compute new IPv4 length. */
778                 tlen = (pa->ip4->ip_hl << 2) + sizeof(*pa->udp) + payload_len;
779                 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
780
781                 /* Subtract delta from current IPv4 checksum. */
782                 csum = pa->ip4->ip_sum + 0xffff - temp[0];
783                 while (csum > 0xffff)
784                         csum = (csum >> 16) + (csum & 0xffff);
785                 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
786                 goto update_udp_header;
787
788         case LRO_TYPE_IPV6_UDP:
789                 /* Compute new IPv6 length. */
790                 tlen = sizeof(*pa->udp) + payload_len;
791                 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
792                 goto update_udp_header;
793
794         default:
795                 return (0);
796         }
797
798 update_tcp_header:
799         /* Compute current TCP header checksum. */
800         temp[2] = tcp_lro_rx_csum_tcphdr(pa->tcp);
801
802         /* Incorporate the latest ACK into the TCP header. */
803         pa->tcp->th_ack = le->ack_seq;
804         pa->tcp->th_win = le->window;
805
806         /* Incorporate latest timestamp into the TCP header. */
807         if (le->timestamp != 0) {
808                 uint32_t *ts_ptr;
809
810                 ts_ptr = (uint32_t *)(pa->tcp + 1);
811                 ts_ptr[1] = htonl(le->tsval);
812                 ts_ptr[2] = le->tsecr;
813         }
814
815         /* Compute new TCP header checksum. */
816         temp[3] = tcp_lro_rx_csum_tcphdr(pa->tcp);
817
818         /* Compute new TCP checksum. */
819         csum = pa->tcp->th_sum + 0xffff - delta_sum +
820             0xffff - temp[0] + 0xffff - temp[3] + temp[2];
821         while (csum > 0xffff)
822                 csum = (csum >> 16) + (csum & 0xffff);
823
824         /* Assign new TCP checksum. */
825         tcp_lro_assign_and_checksum_16(&pa->tcp->th_sum, csum, &temp[4]);
826
827         /* Compute all modififications affecting next checksum. */
828         csum = temp[0] + temp[1] + 0xffff - temp[2] +
829             temp[3] + temp[4] + delta_sum;
830         while (csum > 0xffff)
831                 csum = (csum >> 16) + (csum & 0xffff);
832
833         /* Return delta checksum to next stage, if any. */
834         return (csum);
835
836 update_udp_header:
837         tlen = sizeof(*pa->udp) + payload_len;
838         /* Assign new UDP length and compute checksum delta. */
839         tcp_lro_assign_and_checksum_16(&pa->udp->uh_ulen, htons(tlen), &temp[2]);
840
841         /* Check if there is a UDP checksum. */
842         if (__predict_false(pa->udp->uh_sum != 0)) {
843                 /* Compute new UDP checksum. */
844                 csum = pa->udp->uh_sum + 0xffff - delta_sum +
845                     0xffff - temp[0] + 0xffff - temp[2];
846                 while (csum > 0xffff)
847                         csum = (csum >> 16) + (csum & 0xffff);
848                 /* Assign new UDP checksum. */
849                 tcp_lro_assign_and_checksum_16(&pa->udp->uh_sum, csum, &temp[3]);
850         }
851
852         /* Compute all modififications affecting next checksum. */
853         csum = temp[0] + temp[1] + temp[2] + temp[3] + delta_sum;
854         while (csum > 0xffff)
855                 csum = (csum >> 16) + (csum & 0xffff);
856
857         /* Return delta checksum to next stage, if any. */
858         return (csum);
859 }
860
861 static void
862 tcp_flush_out_entry(struct lro_ctrl *lc, struct lro_entry *le)
863 {
864         /* Check if we need to recompute any checksums. */
865         if (le->needs_merge) {
866                 uint16_t csum;
867
868                 switch (le->inner.data.lro_type) {
869                 case LRO_TYPE_IPV4_TCP:
870                         csum = tcp_lro_update_checksum(&le->inner, le,
871                             le->m_head->m_pkthdr.lro_tcp_d_len,
872                             le->m_head->m_pkthdr.lro_tcp_d_csum);
873                         csum = tcp_lro_update_checksum(&le->outer, NULL,
874                             le->m_head->m_pkthdr.lro_tcp_d_len +
875                             le->inner.total_hdr_len, csum);
876                         le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
877                             CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
878                         le->m_head->m_pkthdr.csum_data = 0xffff;
879                         if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
880                                 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
881                         break;
882                 case LRO_TYPE_IPV6_TCP:
883                         csum = tcp_lro_update_checksum(&le->inner, le,
884                             le->m_head->m_pkthdr.lro_tcp_d_len,
885                             le->m_head->m_pkthdr.lro_tcp_d_csum);
886                         csum = tcp_lro_update_checksum(&le->outer, NULL,
887                             le->m_head->m_pkthdr.lro_tcp_d_len +
888                             le->inner.total_hdr_len, csum);
889                         le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
890                             CSUM_PSEUDO_HDR;
891                         le->m_head->m_pkthdr.csum_data = 0xffff;
892                         if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
893                                 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
894                         break;
895                 case LRO_TYPE_NONE:
896                         switch (le->outer.data.lro_type) {
897                         case LRO_TYPE_IPV4_TCP:
898                                 csum = tcp_lro_update_checksum(&le->outer, le,
899                                     le->m_head->m_pkthdr.lro_tcp_d_len,
900                                     le->m_head->m_pkthdr.lro_tcp_d_csum);
901                                 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
902                                     CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
903                                 le->m_head->m_pkthdr.csum_data = 0xffff;
904                                 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
905                                         le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
906                                 break;
907                         case LRO_TYPE_IPV6_TCP:
908                                 csum = tcp_lro_update_checksum(&le->outer, le,
909                                     le->m_head->m_pkthdr.lro_tcp_d_len,
910                                     le->m_head->m_pkthdr.lro_tcp_d_csum);
911                                 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
912                                     CSUM_PSEUDO_HDR;
913                                 le->m_head->m_pkthdr.csum_data = 0xffff;
914                                 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
915                                         le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
916                                 break;
917                         default:
918                                 break;
919                         }
920                         break;
921                 default:
922                         break;
923                 }
924         }
925
926         /*
927          * Break any chain, this is not set to NULL on the singleton
928          * case m_nextpkt points to m_head. Other case set them
929          * m_nextpkt to NULL in push_and_replace.
930          */
931         le->m_head->m_nextpkt = NULL;
932         lc->lro_queued += le->m_head->m_pkthdr.lro_nsegs;
933         (*lc->ifp->if_input)(lc->ifp, le->m_head);
934 }
935
936 static void
937 tcp_set_entry_to_mbuf(struct lro_ctrl *lc, struct lro_entry *le,
938     struct mbuf *m, struct tcphdr *th)
939 {
940         uint32_t *ts_ptr;
941         uint16_t tcp_data_len;
942         uint16_t tcp_opt_len;
943
944         ts_ptr = (uint32_t *)(th + 1);
945         tcp_opt_len = (th->th_off << 2);
946         tcp_opt_len -= sizeof(*th);
947
948         /* Check if there is a timestamp option. */
949         if (tcp_opt_len == 0 ||
950             __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
951             *ts_ptr != TCP_LRO_TS_OPTION)) {
952                 /* We failed to find the timestamp option. */
953                 le->timestamp = 0;
954         } else {
955                 le->timestamp = 1;
956                 le->tsval = ntohl(*(ts_ptr + 1));
957                 le->tsecr = *(ts_ptr + 2);
958         }
959
960         tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
961
962         /* Pull out TCP sequence numbers and window size. */
963         le->next_seq = ntohl(th->th_seq) + tcp_data_len;
964         le->ack_seq = th->th_ack;
965         le->window = th->th_win;
966         le->flags = tcp_get_flags(th);
967         le->needs_merge = 0;
968
969         /* Setup new data pointers. */
970         le->m_head = m;
971         le->m_tail = m_last(m);
972 }
973
974 static void
975 tcp_push_and_replace(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m)
976 {
977         struct lro_parser *pa;
978
979         /*
980          * Push up the stack of the current entry
981          * and replace it with "m".
982          */
983         struct mbuf *msave;
984
985         /* Grab off the next and save it */
986         msave = le->m_head->m_nextpkt;
987         le->m_head->m_nextpkt = NULL;
988
989         /* Now push out the old entry */
990         tcp_flush_out_entry(lc, le);
991
992         /* Re-parse new header, should not fail. */
993         pa = tcp_lro_parser(m, &le->outer, &le->inner, false);
994         KASSERT(pa != NULL,
995             ("tcp_push_and_replace: LRO parser failed on m=%p\n", m));
996
997         /*
998          * Now to replace the data properly in the entry
999          * we have to reset the TCP header and
1000          * other fields.
1001          */
1002         tcp_set_entry_to_mbuf(lc, le, m, pa->tcp);
1003
1004         /* Restore the next list */
1005         m->m_nextpkt = msave;
1006 }
1007
1008 static void
1009 tcp_lro_mbuf_append_pkthdr(struct lro_entry *le, const struct mbuf *p)
1010 {
1011         struct mbuf *m;
1012         uint32_t csum;
1013
1014         m = le->m_head;
1015         if (m->m_pkthdr.lro_nsegs == 1) {
1016                 /* Compute relative checksum. */
1017                 csum = p->m_pkthdr.lro_tcp_d_csum;
1018         } else {
1019                 /* Merge TCP data checksums. */
1020                 csum = (uint32_t)m->m_pkthdr.lro_tcp_d_csum +
1021                     (uint32_t)p->m_pkthdr.lro_tcp_d_csum;
1022                 while (csum > 0xffff)
1023                         csum = (csum >> 16) + (csum & 0xffff);
1024         }
1025
1026         /* Update various counters. */
1027         m->m_pkthdr.len += p->m_pkthdr.lro_tcp_d_len;
1028         m->m_pkthdr.lro_tcp_d_csum = csum;
1029         m->m_pkthdr.lro_tcp_d_len += p->m_pkthdr.lro_tcp_d_len;
1030         m->m_pkthdr.lro_nsegs += p->m_pkthdr.lro_nsegs;
1031         le->needs_merge = 1;
1032 }
1033
1034 static void
1035 tcp_lro_condense(struct lro_ctrl *lc, struct lro_entry *le)
1036 {
1037         /*
1038          * Walk through the mbuf chain we
1039          * have on tap and compress/condense
1040          * as required.
1041          */
1042         uint32_t *ts_ptr;
1043         struct mbuf *m;
1044         struct tcphdr *th;
1045         uint32_t tcp_data_len_total;
1046         uint32_t tcp_data_seg_total;
1047         uint16_t tcp_data_len;
1048         uint16_t tcp_opt_len;
1049
1050         /*
1051          * First we must check the lead (m_head)
1052          * we must make sure that it is *not*
1053          * something that should be sent up
1054          * right away (sack etc).
1055          */
1056 again:
1057         m = le->m_head->m_nextpkt;
1058         if (m == NULL) {
1059                 /* Just one left. */
1060                 return;
1061         }
1062
1063         th = tcp_lro_get_th(m);
1064         tcp_opt_len = (th->th_off << 2);
1065         tcp_opt_len -= sizeof(*th);
1066         ts_ptr = (uint32_t *)(th + 1);
1067
1068         if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1069             *ts_ptr != TCP_LRO_TS_OPTION)) {
1070                 /*
1071                  * Its not the timestamp. We can't
1072                  * use this guy as the head.
1073                  */
1074                 le->m_head->m_nextpkt = m->m_nextpkt;
1075                 tcp_push_and_replace(lc, le, m);
1076                 goto again;
1077         }
1078         if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1079                 /*
1080                  * Make sure that previously seen segments/ACKs are delivered
1081                  * before this segment, e.g. FIN.
1082                  */
1083                 le->m_head->m_nextpkt = m->m_nextpkt;
1084                 tcp_push_and_replace(lc, le, m);
1085                 goto again;
1086         }
1087         while((m = le->m_head->m_nextpkt) != NULL) {
1088                 /*
1089                  * condense m into le, first
1090                  * pull m out of the list.
1091                  */
1092                 le->m_head->m_nextpkt = m->m_nextpkt;
1093                 m->m_nextpkt = NULL;
1094                 /* Setup my data */
1095                 tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
1096                 th = tcp_lro_get_th(m);
1097                 ts_ptr = (uint32_t *)(th + 1);
1098                 tcp_opt_len = (th->th_off << 2);
1099                 tcp_opt_len -= sizeof(*th);
1100                 tcp_data_len_total = le->m_head->m_pkthdr.lro_tcp_d_len + tcp_data_len;
1101                 tcp_data_seg_total = le->m_head->m_pkthdr.lro_nsegs + m->m_pkthdr.lro_nsegs;
1102
1103                 if (tcp_data_seg_total >= lc->lro_ackcnt_lim ||
1104                     tcp_data_len_total >= lc->lro_length_lim) {
1105                         /* Flush now if appending will result in overflow. */
1106                         tcp_push_and_replace(lc, le, m);
1107                         goto again;
1108                 }
1109                 if (tcp_opt_len != 0 &&
1110                     __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1111                     *ts_ptr != TCP_LRO_TS_OPTION)) {
1112                         /*
1113                          * Maybe a sack in the new one? We need to
1114                          * start all over after flushing the
1115                          * current le. We will go up to the beginning
1116                          * and flush it (calling the replace again possibly
1117                          * or just returning).
1118                          */
1119                         tcp_push_and_replace(lc, le, m);
1120                         goto again;
1121                 }
1122                 if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1123                         tcp_push_and_replace(lc, le, m);
1124                         goto again;
1125                 }
1126                 if (tcp_opt_len != 0) {
1127                         uint32_t tsval = ntohl(*(ts_ptr + 1));
1128                         /* Make sure timestamp values are increasing. */
1129                         if (TSTMP_GT(le->tsval, tsval))  {
1130                                 tcp_push_and_replace(lc, le, m);
1131                                 goto again;
1132                         }
1133                         le->tsval = tsval;
1134                         le->tsecr = *(ts_ptr + 2);
1135                 }
1136                 /* Try to append the new segment. */
1137                 if (__predict_false(ntohl(th->th_seq) != le->next_seq ||
1138                                     ((tcp_get_flags(th) & TH_ACK) !=
1139                                       (le->flags & TH_ACK)) ||
1140                                     (tcp_data_len == 0 &&
1141                                      le->ack_seq == th->th_ack &&
1142                                      le->window == th->th_win))) {
1143                         /* Out of order packet, non-ACK + ACK or dup ACK. */
1144                         tcp_push_and_replace(lc, le, m);
1145                         goto again;
1146                 }
1147                 if (tcp_data_len != 0 ||
1148                     SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1149                         le->next_seq += tcp_data_len;
1150                         le->ack_seq = th->th_ack;
1151                         le->window = th->th_win;
1152                         le->needs_merge = 1;
1153                 } else if (th->th_ack == le->ack_seq) {
1154                         if (WIN_GT(th->th_win, le->window)) {
1155                                 le->window = th->th_win;
1156                                 le->needs_merge = 1;
1157                         }
1158                 }
1159
1160                 if (tcp_data_len == 0) {
1161                         m_freem(m);
1162                         continue;
1163                 }
1164
1165                 /* Merge TCP data checksum and length to head mbuf. */
1166                 tcp_lro_mbuf_append_pkthdr(le, m);
1167
1168                 /*
1169                  * Adjust the mbuf so that m_data points to the first byte of
1170                  * the ULP payload.  Adjust the mbuf to avoid complications and
1171                  * append new segment to existing mbuf chain.
1172                  */
1173                 m_adj(m, m->m_pkthdr.len - tcp_data_len);
1174                 m_demote_pkthdr(m);
1175                 le->m_tail->m_next = m;
1176                 le->m_tail = m_last(m);
1177         }
1178 }
1179
1180 #ifdef TCPHPTS
1181 static void
1182 tcp_queue_pkts(struct tcpcb *tp, struct lro_entry *le)
1183 {
1184
1185         INP_WLOCK_ASSERT(tptoinpcb(tp));
1186
1187         STAILQ_HEAD(, mbuf) q = { le->m_head,
1188             &STAILQ_NEXT(le->m_last_mbuf, m_stailqpkt) };
1189         STAILQ_CONCAT(&tp->t_inqueue, &q);
1190         le->m_head = NULL;
1191         le->m_last_mbuf = NULL;
1192 }
1193
1194 static bool
1195 tcp_lro_check_wake_status(struct inpcb *inp)
1196 {
1197         struct tcpcb *tp;
1198
1199         tp = intotcpcb(inp);
1200         if (__predict_false(tp == NULL))
1201                 return (true);
1202         if (tp->t_fb->tfb_early_wake_check != NULL)
1203                 return ((tp->t_fb->tfb_early_wake_check)(tp));
1204         return (false);
1205 }
1206
1207 static struct mbuf *
1208 tcp_lro_get_last_if_ackcmp(struct lro_ctrl *lc, struct lro_entry *le,
1209     struct inpcb *inp, int32_t *new_m, bool can_append_old_cmp)
1210 {
1211         struct tcpcb *tp;
1212         struct mbuf *m;
1213
1214         tp = intotcpcb(inp);
1215         if (__predict_false(tp == NULL))
1216                 return (NULL);
1217
1218         /* Look at the last mbuf if any in queue */
1219         if (can_append_old_cmp) {
1220                 m = STAILQ_LAST(&tp->t_inqueue, mbuf, m_stailqpkt);
1221                 if (m != NULL && (m->m_flags & M_ACKCMP) != 0) {
1222                         if (M_TRAILINGSPACE(m) >= sizeof(struct tcp_ackent)) {
1223                                 tcp_lro_log(tp, lc, le, NULL, 23, 0, 0, 0, 0);
1224                                 *new_m = 0;
1225                                 counter_u64_add(tcp_extra_mbuf, 1);
1226                                 return (m);
1227                         } else {
1228                                 /* Mark we ran out of space */
1229                                 inp->inp_flags2 |= INP_MBUF_L_ACKS;
1230                         }
1231                 }
1232         }
1233         /* Decide mbuf size. */
1234         tcp_lro_log(tp, lc, le, NULL, 21, 0, 0, 0, 0);
1235         if (inp->inp_flags2 & INP_MBUF_L_ACKS)
1236                 m = m_getcl(M_NOWAIT, MT_DATA, M_ACKCMP | M_PKTHDR);
1237         else
1238                 m = m_gethdr(M_NOWAIT, MT_DATA);
1239
1240         if (__predict_false(m == NULL)) {
1241                 counter_u64_add(tcp_would_have_but, 1);
1242                 return (NULL);
1243         }
1244         counter_u64_add(tcp_comp_total, 1);
1245         m->m_pkthdr.rcvif = lc->ifp;
1246         m->m_flags |= M_ACKCMP;
1247         *new_m = 1;
1248         return (m);
1249 }
1250
1251 static struct inpcb *
1252 tcp_lro_lookup(struct ifnet *ifp, struct lro_parser *pa)
1253 {
1254         struct inpcb *inp;
1255
1256         switch (pa->data.lro_type) {
1257 #ifdef INET6
1258         case LRO_TYPE_IPV6_TCP:
1259                 inp = in6_pcblookup(&V_tcbinfo,
1260                     &pa->data.s_addr.v6,
1261                     pa->data.s_port,
1262                     &pa->data.d_addr.v6,
1263                     pa->data.d_port,
1264                     INPLOOKUP_WLOCKPCB,
1265                     ifp);
1266                 break;
1267 #endif
1268 #ifdef INET
1269         case LRO_TYPE_IPV4_TCP:
1270                 inp = in_pcblookup(&V_tcbinfo,
1271                     pa->data.s_addr.v4,
1272                     pa->data.s_port,
1273                     pa->data.d_addr.v4,
1274                     pa->data.d_port,
1275                     INPLOOKUP_WLOCKPCB,
1276                     ifp);
1277                 break;
1278 #endif
1279         default:
1280                 inp = NULL;
1281                 break;
1282         }
1283         return (inp);
1284 }
1285
1286 static inline bool
1287 tcp_lro_ack_valid(struct mbuf *m, struct tcphdr *th, uint32_t **ppts, bool *other_opts)
1288 {
1289         /*
1290          * This function returns two bits of valuable information.
1291          * a) Is what is present capable of being ack-compressed,
1292          *    we can ack-compress if there is no options or just
1293          *    a timestamp option, and of course the th_flags must
1294          *    be correct as well.
1295          * b) Our other options present such as SACK. This is
1296          *    used to determine if we want to wakeup or not.
1297          */
1298         bool ret = true;
1299
1300         switch (th->th_off << 2) {
1301         case (sizeof(*th) + TCPOLEN_TSTAMP_APPA):
1302                 *ppts = (uint32_t *)(th + 1);
1303                 /* Check if we have only one timestamp option. */
1304                 if (**ppts == TCP_LRO_TS_OPTION)
1305                         *other_opts = false;
1306                 else {
1307                         *other_opts = true;
1308                         ret = false;
1309                 }
1310                 break;
1311         case (sizeof(*th)):
1312                 /* No options. */
1313                 *ppts = NULL;
1314                 *other_opts = false;
1315                 break;
1316         default:
1317                 *ppts = NULL;
1318                 *other_opts = true;
1319                 ret = false;
1320                 break;
1321         }
1322         /* For ACKCMP we only accept ACK, PUSH, ECE and CWR. */
1323         if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH | TH_ECE | TH_CWR)) != 0)
1324                 ret = false;
1325         /* If it has data on it we cannot compress it */
1326         if (m->m_pkthdr.lro_tcp_d_len)
1327                 ret = false;
1328
1329         /* ACK flag must be set. */
1330         if (!(tcp_get_flags(th) & TH_ACK))
1331                 ret = false;
1332         return (ret);
1333 }
1334
1335 static int
1336 tcp_lro_flush_tcphpts(struct lro_ctrl *lc, struct lro_entry *le)
1337 {
1338         struct inpcb *inp;
1339         struct tcpcb *tp;
1340         struct mbuf **pp, *cmp, *mv_to;
1341         struct ifnet *lagg_ifp;
1342         bool bpf_req, lagg_bpf_req, should_wake, can_append_old_cmp;
1343
1344         /* Check if packet doesn't belongs to our network interface. */
1345         if ((tcplro_stacks_wanting_mbufq == 0) ||
1346             (le->outer.data.vlan_id != 0) ||
1347             (le->inner.data.lro_type != LRO_TYPE_NONE))
1348                 return (TCP_LRO_CANNOT);
1349
1350 #ifdef INET6
1351         /*
1352          * Be proactive about unspecified IPv6 address in source. As
1353          * we use all-zero to indicate unbounded/unconnected pcb,
1354          * unspecified IPv6 address can be used to confuse us.
1355          *
1356          * Note that packets with unspecified IPv6 destination is
1357          * already dropped in ip6_input.
1358          */
1359         if (__predict_false(le->outer.data.lro_type == LRO_TYPE_IPV6_TCP &&
1360             IN6_IS_ADDR_UNSPECIFIED(&le->outer.data.s_addr.v6)))
1361                 return (TCP_LRO_CANNOT);
1362
1363         if (__predict_false(le->inner.data.lro_type == LRO_TYPE_IPV6_TCP &&
1364             IN6_IS_ADDR_UNSPECIFIED(&le->inner.data.s_addr.v6)))
1365                 return (TCP_LRO_CANNOT);
1366 #endif
1367         /* Lookup inp, if any. */
1368         inp = tcp_lro_lookup(lc->ifp,
1369             (le->inner.data.lro_type == LRO_TYPE_NONE) ? &le->outer : &le->inner);
1370         if (inp == NULL)
1371                 return (TCP_LRO_CANNOT);
1372
1373         counter_u64_add(tcp_inp_lro_locks_taken, 1);
1374
1375         /* Get TCP control structure. */
1376         tp = intotcpcb(inp);
1377
1378         /* Check if the inp is dead, Jim. */
1379         if (tp->t_state == TCPS_TIME_WAIT) {
1380                 INP_WUNLOCK(inp);
1381                 return (TCP_LRO_CANNOT);
1382         }
1383         if ((inp->inp_irq_cpu_set == 0)  && (lc->lro_cpu_is_set == 1)) {
1384                 inp->inp_irq_cpu = lc->lro_last_cpu;
1385                 inp->inp_irq_cpu_set = 1;
1386         }
1387         /* Check if the transport doesn't support the needed optimizations. */
1388         if ((inp->inp_flags2 & (INP_SUPPORTS_MBUFQ | INP_MBUF_ACKCMP)) == 0) {
1389                 INP_WUNLOCK(inp);
1390                 return (TCP_LRO_CANNOT);
1391         }
1392
1393         if (inp->inp_flags2 & INP_MBUF_QUEUE_READY)
1394                 should_wake = false;
1395         else
1396                 should_wake = true;
1397         /* Check if packets should be tapped to BPF. */
1398         bpf_req = bpf_peers_present(lc->ifp->if_bpf);
1399         lagg_bpf_req = false;
1400         lagg_ifp = NULL;
1401         if (lc->ifp->if_type == IFT_IEEE8023ADLAG ||
1402             lc->ifp->if_type == IFT_INFINIBANDLAG) {
1403                 struct lagg_port *lp = lc->ifp->if_lagg;
1404                 struct lagg_softc *sc = lp->lp_softc;
1405
1406                 lagg_ifp = sc->sc_ifp;
1407                 if (lagg_ifp != NULL)
1408                         lagg_bpf_req = bpf_peers_present(lagg_ifp->if_bpf);
1409         }
1410
1411         /* Strip and compress all the incoming packets. */
1412         can_append_old_cmp = true;
1413         cmp = NULL;
1414         for (pp = &le->m_head; *pp != NULL; ) {
1415                 mv_to = NULL;
1416                 if (do_bpf_strip_and_compress(inp, lc, le, pp,
1417                         &cmp, &mv_to, &should_wake, bpf_req,
1418                         lagg_bpf_req, lagg_ifp, can_append_old_cmp) == false) {
1419                         /* Advance to next mbuf. */
1420                         pp = &(*pp)->m_nextpkt;
1421                         /*
1422                          * Once we have appended we can't look in the pending
1423                          * inbound packets for a compressed ack to append to.
1424                          */
1425                         can_append_old_cmp = false;
1426                         /*
1427                          * Once we append we also need to stop adding to any
1428                          * compressed ack we were remembering. A new cmp
1429                          * ack will be required.
1430                          */
1431                         cmp = NULL;
1432                         tcp_lro_log(tp, lc, le, NULL, 25, 0, 0, 0, 0);
1433                 } else if (mv_to != NULL) {
1434                         /* We are asked to move pp up */
1435                         pp = &mv_to->m_nextpkt;
1436                         tcp_lro_log(tp, lc, le, NULL, 24, 0, 0, 0, 0);
1437                 } else
1438                         tcp_lro_log(tp, lc, le, NULL, 26, 0, 0, 0, 0);
1439         }
1440         /* Update "m_last_mbuf", if any. */
1441         if (pp == &le->m_head)
1442                 le->m_last_mbuf = *pp;
1443         else
1444                 le->m_last_mbuf = __containerof(pp, struct mbuf, m_nextpkt);
1445
1446         /* Check if any data mbufs left. */
1447         if (le->m_head != NULL) {
1448                 counter_u64_add(tcp_inp_lro_direct_queue, 1);
1449                 tcp_lro_log(tp, lc, le, NULL, 22, 1, inp->inp_flags2, 0, 1);
1450                 tcp_queue_pkts(tp, le);
1451         }
1452         if (should_wake) {
1453                 /* Wakeup */
1454                 counter_u64_add(tcp_inp_lro_wokeup_queue, 1);
1455                 if ((*tp->t_fb->tfb_do_queued_segments)(tp, 0))
1456                         inp = NULL;
1457         }
1458         if (inp != NULL)
1459                 INP_WUNLOCK(inp);
1460         return (0);     /* Success. */
1461 }
1462 #endif
1463
1464 void
1465 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
1466 {
1467         /* Only optimise if there are multiple packets waiting. */
1468 #ifdef TCPHPTS
1469         int error;
1470 #endif
1471
1472         NET_EPOCH_ASSERT();
1473 #ifdef TCPHPTS
1474         CURVNET_SET(lc->ifp->if_vnet);
1475         error = tcp_lro_flush_tcphpts(lc, le);
1476         CURVNET_RESTORE();
1477         if (error != 0) {
1478 #endif
1479                 tcp_lro_condense(lc, le);
1480                 tcp_flush_out_entry(lc, le);
1481 #ifdef TCPHPTS
1482         }
1483 #endif
1484         lc->lro_flushed++;
1485         bzero(le, sizeof(*le));
1486         LIST_INSERT_HEAD(&lc->lro_free, le, next);
1487 }
1488
1489 #ifdef HAVE_INLINE_FLSLL
1490 #define tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
1491 #else
1492 static inline uint64_t
1493 tcp_lro_msb_64(uint64_t x)
1494 {
1495         x |= (x >> 1);
1496         x |= (x >> 2);
1497         x |= (x >> 4);
1498         x |= (x >> 8);
1499         x |= (x >> 16);
1500         x |= (x >> 32);
1501         return (x & ~(x >> 1));
1502 }
1503 #endif
1504
1505 /*
1506  * The tcp_lro_sort() routine is comparable to qsort(), except it has
1507  * a worst case complexity limit of O(MIN(N,64)*N), where N is the
1508  * number of elements to sort and 64 is the number of sequence bits
1509  * available. The algorithm is bit-slicing the 64-bit sequence number,
1510  * sorting one bit at a time from the most significant bit until the
1511  * least significant one, skipping the constant bits. This is
1512  * typically called a radix sort.
1513  */
1514 static void
1515 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
1516 {
1517         struct lro_mbuf_sort temp;
1518         uint64_t ones;
1519         uint64_t zeros;
1520         uint32_t x;
1521         uint32_t y;
1522
1523 repeat:
1524         /* for small arrays insertion sort is faster */
1525         if (size <= 12) {
1526                 for (x = 1; x < size; x++) {
1527                         temp = parray[x];
1528                         for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
1529                                 parray[y] = parray[y - 1];
1530                         parray[y] = temp;
1531                 }
1532                 return;
1533         }
1534
1535         /* compute sequence bits which are constant */
1536         ones = 0;
1537         zeros = 0;
1538         for (x = 0; x != size; x++) {
1539                 ones |= parray[x].seq;
1540                 zeros |= ~parray[x].seq;
1541         }
1542
1543         /* compute bits which are not constant into "ones" */
1544         ones &= zeros;
1545         if (ones == 0)
1546                 return;
1547
1548         /* pick the most significant bit which is not constant */
1549         ones = tcp_lro_msb_64(ones);
1550
1551         /*
1552          * Move entries having cleared sequence bits to the beginning
1553          * of the array:
1554          */
1555         for (x = y = 0; y != size; y++) {
1556                 /* skip set bits */
1557                 if (parray[y].seq & ones)
1558                         continue;
1559                 /* swap entries */
1560                 temp = parray[x];
1561                 parray[x] = parray[y];
1562                 parray[y] = temp;
1563                 x++;
1564         }
1565
1566         KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
1567
1568         /* sort zeros */
1569         tcp_lro_sort(parray, x);
1570
1571         /* sort ones */
1572         parray += x;
1573         size -= x;
1574         goto repeat;
1575 }
1576
1577 void
1578 tcp_lro_flush_all(struct lro_ctrl *lc)
1579 {
1580         uint64_t seq;
1581         uint64_t nseq;
1582         unsigned x;
1583
1584         NET_EPOCH_ASSERT();
1585         /* check if no mbufs to flush */
1586         if (lc->lro_mbuf_count == 0)
1587                 goto done;
1588         if (lc->lro_cpu_is_set == 0) {
1589                 if (lc->lro_last_cpu == curcpu) {
1590                         lc->lro_cnt_of_same_cpu++;
1591                         /* Have we reached the threshold to declare a cpu? */
1592                         if (lc->lro_cnt_of_same_cpu > tcp_lro_cpu_set_thresh)
1593                                 lc->lro_cpu_is_set = 1;
1594                 } else {
1595                         lc->lro_last_cpu = curcpu;
1596                         lc->lro_cnt_of_same_cpu = 0;
1597                 }
1598         }
1599         CURVNET_SET(lc->ifp->if_vnet);
1600
1601         /* get current time */
1602         binuptime(&lc->lro_last_queue_time);
1603
1604         /* sort all mbufs according to stream */
1605         tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
1606
1607         /* input data into LRO engine, stream by stream */
1608         seq = 0;
1609         for (x = 0; x != lc->lro_mbuf_count; x++) {
1610                 struct mbuf *mb;
1611
1612                 /* get mbuf */
1613                 mb = lc->lro_mbuf_data[x].mb;
1614
1615                 /* get sequence number, masking away the packet index */
1616                 nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
1617
1618                 /* check for new stream */
1619                 if (seq != nseq) {
1620                         seq = nseq;
1621
1622                         /* flush active streams */
1623                         tcp_lro_rx_done(lc);
1624                 }
1625
1626                 /* add packet to LRO engine */
1627                 if (tcp_lro_rx_common(lc, mb, 0, false) != 0) {
1628                         /* Flush anything we have acummulated */
1629                         tcp_lro_flush_active(lc);
1630                         /* input packet to network layer */
1631                         (*lc->ifp->if_input)(lc->ifp, mb);
1632                         lc->lro_queued++;
1633                         lc->lro_flushed++;
1634                 }
1635         }
1636         CURVNET_RESTORE();
1637 done:
1638         /* flush active streams */
1639         tcp_lro_rx_done(lc);
1640
1641 #ifdef TCPHPTS
1642         tcp_run_hpts();
1643 #endif
1644         lc->lro_mbuf_count = 0;
1645 }
1646
1647 #ifdef TCPHPTS
1648 static void
1649 build_ack_entry(struct tcp_ackent *ae, struct tcphdr *th, struct mbuf *m,
1650     uint32_t *ts_ptr, uint16_t iptos)
1651 {
1652         /*
1653          * Given a TCP ACK, summarize it down into the small TCP ACK
1654          * entry.
1655          */
1656         ae->timestamp = m->m_pkthdr.rcv_tstmp;
1657         ae->flags = 0;
1658         if (m->m_flags & M_TSTMP_LRO)
1659                 ae->flags |= TSTMP_LRO;
1660         else if (m->m_flags & M_TSTMP)
1661                 ae->flags |= TSTMP_HDWR;
1662         ae->seq = ntohl(th->th_seq);
1663         ae->ack = ntohl(th->th_ack);
1664         ae->flags |= tcp_get_flags(th);
1665         if (ts_ptr != NULL) {
1666                 ae->ts_value = ntohl(ts_ptr[1]);
1667                 ae->ts_echo = ntohl(ts_ptr[2]);
1668                 ae->flags |= HAS_TSTMP;
1669         }
1670         ae->win = ntohs(th->th_win);
1671         ae->codepoint = iptos;
1672 }
1673
1674 /*
1675  * Do BPF tap for either ACK_CMP packets or MBUF QUEUE type packets
1676  * and strip all, but the IPv4/IPv6 header.
1677  */
1678 static bool
1679 do_bpf_strip_and_compress(struct inpcb *inp, struct lro_ctrl *lc,
1680     struct lro_entry *le, struct mbuf **pp, struct mbuf **cmp, struct mbuf **mv_to,
1681     bool *should_wake, bool bpf_req, bool lagg_bpf_req, struct ifnet *lagg_ifp, bool can_append_old_cmp)
1682 {
1683         union {
1684                 void *ptr;
1685                 struct ip *ip4;
1686                 struct ip6_hdr *ip6;
1687         } l3;
1688         struct mbuf *m;
1689         struct mbuf *nm;
1690         struct tcphdr *th;
1691         struct tcp_ackent *ack_ent;
1692         uint32_t *ts_ptr;
1693         int32_t n_mbuf;
1694         bool other_opts, can_compress;
1695         uint8_t lro_type;
1696         uint16_t iptos;
1697         int tcp_hdr_offset;
1698         int idx;
1699
1700         /* Get current mbuf. */
1701         m = *pp;
1702
1703         /* Let the BPF see the packet */
1704         if (__predict_false(bpf_req))
1705                 ETHER_BPF_MTAP(lc->ifp, m);
1706
1707         if (__predict_false(lagg_bpf_req))
1708                 ETHER_BPF_MTAP(lagg_ifp, m);
1709
1710         tcp_hdr_offset = m->m_pkthdr.lro_tcp_h_off;
1711         lro_type = le->inner.data.lro_type;
1712         switch (lro_type) {
1713         case LRO_TYPE_NONE:
1714                 lro_type = le->outer.data.lro_type;
1715                 switch (lro_type) {
1716                 case LRO_TYPE_IPV4_TCP:
1717                         tcp_hdr_offset -= sizeof(*le->outer.ip4);
1718                         m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1719                         break;
1720                 case LRO_TYPE_IPV6_TCP:
1721                         tcp_hdr_offset -= sizeof(*le->outer.ip6);
1722                         m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1723                         break;
1724                 default:
1725                         goto compressed;
1726                 }
1727                 break;
1728         case LRO_TYPE_IPV4_TCP:
1729                 tcp_hdr_offset -= sizeof(*le->outer.ip4);
1730                 m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1731                 break;
1732         case LRO_TYPE_IPV6_TCP:
1733                 tcp_hdr_offset -= sizeof(*le->outer.ip6);
1734                 m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1735                 break;
1736         default:
1737                 goto compressed;
1738         }
1739
1740         MPASS(tcp_hdr_offset >= 0);
1741
1742         m_adj(m, tcp_hdr_offset);
1743         m->m_flags |= M_LRO_EHDRSTRP;
1744         m->m_flags &= ~M_ACKCMP;
1745         m->m_pkthdr.lro_tcp_h_off -= tcp_hdr_offset;
1746
1747         th = tcp_lro_get_th(m);
1748
1749         th->th_sum = 0;         /* TCP checksum is valid. */
1750
1751         /* Check if ACK can be compressed */
1752         can_compress = tcp_lro_ack_valid(m, th, &ts_ptr, &other_opts);
1753
1754         /* Now lets look at the should wake states */
1755         if ((other_opts == true) &&
1756             ((inp->inp_flags2 & INP_DONT_SACK_QUEUE) == 0)) {
1757                 /*
1758                  * If there are other options (SACK?) and the
1759                  * tcp endpoint has not expressly told us it does
1760                  * not care about SACKS, then we should wake up.
1761                  */
1762                 *should_wake = true;
1763         } else if (*should_wake == false) {
1764                 /* Wakeup override check if we are false here  */
1765                 *should_wake = tcp_lro_check_wake_status(inp);
1766         }
1767         /* Is the ack compressable? */
1768         if (can_compress == false)
1769                 goto done;
1770         /* Does the TCP endpoint support ACK compression? */
1771         if ((inp->inp_flags2 & INP_MBUF_ACKCMP) == 0)
1772                 goto done;
1773
1774         /* Lets get the TOS/traffic class field */
1775         l3.ptr = mtod(m, void *);
1776         switch (lro_type) {
1777         case LRO_TYPE_IPV4_TCP:
1778                 iptos = l3.ip4->ip_tos;
1779                 break;
1780         case LRO_TYPE_IPV6_TCP:
1781                 iptos = IPV6_TRAFFIC_CLASS(l3.ip6);
1782                 break;
1783         default:
1784                 iptos = 0;      /* Keep compiler happy. */
1785                 break;
1786         }
1787         /* Now lets get space if we don't have some already */
1788         if (*cmp == NULL) {
1789 new_one:
1790                 nm = tcp_lro_get_last_if_ackcmp(lc, le, inp, &n_mbuf, can_append_old_cmp);
1791                 if (__predict_false(nm == NULL))
1792                         goto done;
1793                 *cmp = nm;
1794                 if (n_mbuf) {
1795                         /*
1796                          *  Link in the new cmp ack to our in-order place,
1797                          * first set our cmp ack's next to where we are.
1798                          */
1799                         nm->m_nextpkt = m;
1800                         (*pp) = nm;
1801                         /*
1802                          * Set it up so mv_to is advanced to our
1803                          * compressed ack. This way the caller can
1804                          * advance pp to the right place.
1805                          */
1806                         *mv_to = nm;
1807                         /*
1808                          * Advance it here locally as well.
1809                          */
1810                         pp = &nm->m_nextpkt;
1811                 }
1812         } else {
1813                 /* We have one already we are working on */
1814                 nm = *cmp;
1815                 if (M_TRAILINGSPACE(nm) < sizeof(struct tcp_ackent)) {
1816                         /* We ran out of space */
1817                         inp->inp_flags2 |= INP_MBUF_L_ACKS;
1818                         goto new_one;
1819                 }
1820         }
1821         MPASS(M_TRAILINGSPACE(nm) >= sizeof(struct tcp_ackent));
1822         counter_u64_add(tcp_inp_lro_compressed, 1);
1823         le->compressed++;
1824         /* We can add in to the one on the tail */
1825         ack_ent = mtod(nm, struct tcp_ackent *);
1826         idx = (nm->m_len / sizeof(struct tcp_ackent));
1827         build_ack_entry(&ack_ent[idx], th, m, ts_ptr, iptos);
1828
1829         /* Bump the size of both pkt-hdr and len */
1830         nm->m_len += sizeof(struct tcp_ackent);
1831         nm->m_pkthdr.len += sizeof(struct tcp_ackent);
1832 compressed:
1833         /* Advance to next mbuf before freeing. */
1834         *pp = m->m_nextpkt;
1835         m->m_nextpkt = NULL;
1836         m_freem(m);
1837         return (true);
1838 done:
1839         counter_u64_add(tcp_uncomp_total, 1);
1840         le->uncompressed++;
1841         return (false);
1842 }
1843 #endif
1844
1845 static struct lro_head *
1846 tcp_lro_rx_get_bucket(struct lro_ctrl *lc, struct mbuf *m, struct lro_parser *parser)
1847 {
1848         u_long hash;
1849
1850         if (M_HASHTYPE_ISHASH(m)) {
1851                 hash = m->m_pkthdr.flowid;
1852         } else {
1853                 for (unsigned i = hash = 0; i != LRO_RAW_ADDRESS_MAX; i++)
1854                         hash += parser->data.raw[i];
1855         }
1856         return (&lc->lro_hash[hash % lc->lro_hashsz]);
1857 }
1858
1859 static int
1860 tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash)
1861 {
1862         struct lro_parser pi;   /* inner address data */
1863         struct lro_parser po;   /* outer address data */
1864         struct lro_parser *pa;  /* current parser for TCP stream */
1865         struct lro_entry *le;
1866         struct lro_head *bucket;
1867         struct tcphdr *th;
1868         int tcp_data_len;
1869         int tcp_opt_len;
1870         int error;
1871         uint16_t tcp_data_sum;
1872
1873 #ifdef INET
1874         /* Quickly decide if packet cannot be LRO'ed */
1875         if (__predict_false(V_ipforwarding != 0))
1876                 return (TCP_LRO_CANNOT);
1877 #endif
1878 #ifdef INET6
1879         /* Quickly decide if packet cannot be LRO'ed */
1880         if (__predict_false(V_ip6_forwarding != 0))
1881                 return (TCP_LRO_CANNOT);
1882 #endif
1883         if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1884              ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) || 
1885             (m->m_pkthdr.csum_data != 0xffff)) {
1886                 /* 
1887                  * The checksum either did not have hardware offload
1888                  * or it was a bad checksum. We can't LRO such
1889                  * a packet.
1890                  */
1891                 counter_u64_add(tcp_bad_csums, 1);
1892                 return (TCP_LRO_CANNOT);
1893         }
1894         /* We expect a contiguous header [eh, ip, tcp]. */
1895         pa = tcp_lro_parser(m, &po, &pi, true);
1896         if (__predict_false(pa == NULL))
1897                 return (TCP_LRO_NOT_SUPPORTED);
1898
1899         /* We don't expect any padding. */
1900         error = tcp_lro_trim_mbuf_chain(m, pa);
1901         if (__predict_false(error != 0))
1902                 return (error);
1903
1904 #ifdef INET
1905         switch (pa->data.lro_type) {
1906         case LRO_TYPE_IPV4_TCP:
1907                 error = tcp_lro_rx_ipv4(lc, m, pa->ip4);
1908                 if (__predict_false(error != 0))
1909                         return (error);
1910                 break;
1911         default:
1912                 break;
1913         }
1914 #endif
1915         /* If no hardware or arrival stamp on the packet add timestamp */
1916         if ((m->m_flags & (M_TSTMP_LRO | M_TSTMP)) == 0) {
1917                 m->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time); 
1918                 m->m_flags |= M_TSTMP_LRO;
1919         }
1920
1921         /* Get pointer to TCP header. */
1922         th = pa->tcp;
1923
1924         /* Don't process SYN packets. */
1925         if (__predict_false(tcp_get_flags(th) & TH_SYN))
1926                 return (TCP_LRO_CANNOT);
1927
1928         /* Get total TCP header length and compute payload length. */
1929         tcp_opt_len = (th->th_off << 2);
1930         tcp_data_len = m->m_pkthdr.len - ((uint8_t *)th -
1931             (uint8_t *)m->m_data) - tcp_opt_len;
1932         tcp_opt_len -= sizeof(*th);
1933
1934         /* Don't process invalid TCP headers. */
1935         if (__predict_false(tcp_opt_len < 0 || tcp_data_len < 0))
1936                 return (TCP_LRO_CANNOT);
1937
1938         /* Compute TCP data only checksum. */
1939         if (tcp_data_len == 0)
1940                 tcp_data_sum = 0;       /* no data, no checksum */
1941         else if (__predict_false(csum != 0))
1942                 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~csum);
1943         else
1944                 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~th->th_sum);
1945
1946         /* Save TCP info in mbuf. */
1947         m->m_nextpkt = NULL;
1948         m->m_pkthdr.rcvif = lc->ifp;
1949         m->m_pkthdr.lro_tcp_d_csum = tcp_data_sum;
1950         m->m_pkthdr.lro_tcp_d_len = tcp_data_len;
1951         m->m_pkthdr.lro_tcp_h_off = ((uint8_t *)th - (uint8_t *)m->m_data);
1952         m->m_pkthdr.lro_nsegs = 1;
1953
1954         /* Get hash bucket. */
1955         if (!use_hash) {
1956                 bucket = &lc->lro_hash[0];
1957         } else {
1958                 bucket = tcp_lro_rx_get_bucket(lc, m, pa);
1959         }
1960
1961         /* Try to find a matching previous segment. */
1962         LIST_FOREACH(le, bucket, hash_next) {
1963                 /* Compare addresses and ports. */
1964                 if (lro_address_compare(&po.data, &le->outer.data) == false ||
1965                     lro_address_compare(&pi.data, &le->inner.data) == false)
1966                         continue;
1967
1968                 /* Check if no data and old ACK. */
1969                 if (tcp_data_len == 0 &&
1970                     SEQ_LT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1971                         m_freem(m);
1972                         return (0);
1973                 }
1974
1975                 /* Mark "m" in the last spot. */
1976                 le->m_last_mbuf->m_nextpkt = m;
1977                 /* Now set the tail to "m". */
1978                 le->m_last_mbuf = m;
1979                 return (0);
1980         }
1981
1982         /* Try to find an empty slot. */
1983         if (LIST_EMPTY(&lc->lro_free))
1984                 return (TCP_LRO_NO_ENTRIES);
1985
1986         /* Start a new segment chain. */
1987         le = LIST_FIRST(&lc->lro_free);
1988         LIST_REMOVE(le, next);
1989         tcp_lro_active_insert(lc, bucket, le);
1990
1991         /* Make sure the headers are set. */
1992         le->inner = pi;
1993         le->outer = po;
1994
1995         /* Store time this entry was allocated. */
1996         le->alloc_time = lc->lro_last_queue_time;
1997
1998         tcp_set_entry_to_mbuf(lc, le, m, th);
1999
2000         /* Now set the tail to "m". */
2001         le->m_last_mbuf = m;
2002
2003         return (0);
2004 }
2005
2006 int
2007 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
2008 {
2009         int error;
2010
2011         if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
2012              ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) || 
2013             (m->m_pkthdr.csum_data != 0xffff)) {
2014                 /* 
2015                  * The checksum either did not have hardware offload
2016                  * or it was a bad checksum. We can't LRO such
2017                  * a packet.
2018                  */
2019                 counter_u64_add(tcp_bad_csums, 1);
2020                 return (TCP_LRO_CANNOT);
2021         }
2022         /* get current time */
2023         binuptime(&lc->lro_last_queue_time);
2024         CURVNET_SET(lc->ifp->if_vnet);
2025         error = tcp_lro_rx_common(lc, m, csum, true);
2026         if (__predict_false(error != 0)) {
2027                 /*
2028                  * Flush anything we have acummulated
2029                  * ahead of this packet that can't
2030                  * be LRO'd. This preserves order.
2031                  */
2032                 tcp_lro_flush_active(lc);
2033         }
2034         CURVNET_RESTORE();
2035
2036         return (error);
2037 }
2038
2039 void
2040 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
2041 {
2042         NET_EPOCH_ASSERT();
2043         /* sanity checks */
2044         if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
2045             lc->lro_mbuf_max == 0)) {
2046                 /* packet drop */
2047                 m_freem(mb);
2048                 return;
2049         }
2050
2051         /* check if packet is not LRO capable */
2052         if (__predict_false((lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
2053                 /* input packet to network layer */
2054                 (*lc->ifp->if_input) (lc->ifp, mb);
2055                 return;
2056         }
2057
2058         /* If no hardware or arrival stamp on the packet add timestamp */
2059         if ((tcplro_stacks_wanting_mbufq > 0) &&
2060             (tcp_less_accurate_lro_ts == 0) &&
2061             ((mb->m_flags & M_TSTMP) == 0)) {
2062                 /* Add in an LRO time since no hardware */
2063                 binuptime(&lc->lro_last_queue_time);
2064                 mb->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time); 
2065                 mb->m_flags |= M_TSTMP_LRO;
2066         }
2067
2068         /* create sequence number */
2069         lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
2070             (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
2071             (((uint64_t)mb->m_pkthdr.flowid) << 24) |
2072             ((uint64_t)lc->lro_mbuf_count);
2073
2074         /* enter mbuf */
2075         lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb;
2076
2077         /* flush if array is full */
2078         if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max))
2079                 tcp_lro_flush_all(lc);
2080 }
2081
2082 /* end */