xref: /freebsd/sys/netinet/tcp_lro.c (revision ce6a89e27cd190313be39bb479880aeda4778436)
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 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/vnet.h>
55 
56 #include <netinet/in_systm.h>
57 #include <netinet/in.h>
58 #include <netinet/ip6.h>
59 #include <netinet/ip.h>
60 #include <netinet/ip_var.h>
61 #include <netinet/in_pcb.h>
62 #include <netinet6/in6_pcb.h>
63 #include <netinet/tcp.h>
64 #include <netinet/tcp_seq.h>
65 #include <netinet/tcp_lro.h>
66 #include <netinet/tcp_var.h>
67 #include <netinet/tcp_hpts.h>
68 #include <netinet/tcp_log_buf.h>
69 #include <netinet6/ip6_var.h>
70 
71 #include <machine/in_cksum.h>
72 
73 static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures");
74 
75 #define	TCP_LRO_UPDATE_CSUM	1
76 #ifndef	TCP_LRO_UPDATE_CSUM
77 #define	TCP_LRO_INVALID_CSUM	0x0000
78 #endif
79 
80 static void	tcp_lro_rx_done(struct lro_ctrl *lc);
81 static int	tcp_lro_rx2(struct lro_ctrl *lc, struct mbuf *m,
82 		    uint32_t csum, int use_hash);
83 
84 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro,  CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
85     "TCP LRO");
86 
87 static long tcplro_stacks_wanting_mbufq = 0;
88 counter_u64_t tcp_inp_lro_direct_queue;
89 counter_u64_t tcp_inp_lro_wokeup_queue;
90 counter_u64_t tcp_inp_lro_compressed;
91 counter_u64_t tcp_inp_lro_single_push;
92 counter_u64_t tcp_inp_lro_locks_taken;
93 counter_u64_t tcp_inp_lro_sack_wake;
94 
95 static unsigned	tcp_lro_entries = TCP_LRO_ENTRIES;
96 static int32_t hold_lock_over_compress = 0;
97 SYSCTL_INT(_net_inet_tcp_lro, OID_AUTO, hold_lock, CTLFLAG_RW,
98     &hold_lock_over_compress, 0,
99     "Do we hold the lock over the compress of mbufs?");
100 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries,
101     CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_entries, 0,
102     "default number of LRO entries");
103 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, fullqueue, CTLFLAG_RD,
104     &tcp_inp_lro_direct_queue, "Number of lro's fully queued to transport");
105 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, wokeup, CTLFLAG_RD,
106     &tcp_inp_lro_wokeup_queue, "Number of lro's where we woke up transport via hpts");
107 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, compressed, CTLFLAG_RD,
108     &tcp_inp_lro_compressed, "Number of lro's compressed and sent to transport");
109 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, single, CTLFLAG_RD,
110     &tcp_inp_lro_single_push, "Number of lro's sent with single segment");
111 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lockcnt, CTLFLAG_RD,
112     &tcp_inp_lro_locks_taken, "Number of lro's inp_wlocks taken");
113 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, sackwakeups, CTLFLAG_RD,
114     &tcp_inp_lro_sack_wake, "Number of wakeups caused by sack/fin");
115 
116 void
117 tcp_lro_reg_mbufq(void)
118 {
119 	atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, 1);
120 }
121 
122 void
123 tcp_lro_dereg_mbufq(void)
124 {
125 	atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, -1);
126 }
127 
128 static __inline void
129 tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket,
130     struct lro_entry *le)
131 {
132 
133 	LIST_INSERT_HEAD(&lc->lro_active, le, next);
134 	LIST_INSERT_HEAD(bucket, le, hash_next);
135 }
136 
137 static __inline void
138 tcp_lro_active_remove(struct lro_entry *le)
139 {
140 
141 	LIST_REMOVE(le, next);		/* active list */
142 	LIST_REMOVE(le, hash_next);	/* hash bucket */
143 }
144 
145 int
146 tcp_lro_init(struct lro_ctrl *lc)
147 {
148 	return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0));
149 }
150 
151 int
152 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
153     unsigned lro_entries, unsigned lro_mbufs)
154 {
155 	struct lro_entry *le;
156 	size_t size;
157 	unsigned i, elements;
158 
159 	lc->lro_bad_csum = 0;
160 	lc->lro_queued = 0;
161 	lc->lro_flushed = 0;
162 	lc->lro_mbuf_count = 0;
163 	lc->lro_mbuf_max = lro_mbufs;
164 	lc->lro_cnt = lro_entries;
165 	lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
166 	lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
167 	lc->ifp = ifp;
168 	LIST_INIT(&lc->lro_free);
169 	LIST_INIT(&lc->lro_active);
170 
171 	/* create hash table to accelerate entry lookup */
172 	if (lro_entries > lro_mbufs)
173 		elements = lro_entries;
174 	else
175 		elements = lro_mbufs;
176 	lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz,
177 	    HASH_NOWAIT);
178 	if (lc->lro_hash == NULL) {
179 		memset(lc, 0, sizeof(*lc));
180 		return (ENOMEM);
181 	}
182 
183 	/* compute size to allocate */
184 	size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
185 	    (lro_entries * sizeof(*le));
186 	lc->lro_mbuf_data = (struct lro_mbuf_sort *)
187 	    malloc(size, M_LRO, M_NOWAIT | M_ZERO);
188 
189 	/* check for out of memory */
190 	if (lc->lro_mbuf_data == NULL) {
191 		free(lc->lro_hash, M_LRO);
192 		memset(lc, 0, sizeof(*lc));
193 		return (ENOMEM);
194 	}
195 	/* compute offset for LRO entries */
196 	le = (struct lro_entry *)
197 	    (lc->lro_mbuf_data + lro_mbufs);
198 
199 	/* setup linked list */
200 	for (i = 0; i != lro_entries; i++)
201 		LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
202 
203 	return (0);
204 }
205 
206 static struct tcphdr *
207 tcp_lro_get_th(struct lro_entry *le, struct mbuf *m)
208 {
209 	struct ether_header *eh;
210 	struct tcphdr *th = NULL;
211 #ifdef INET6
212 	struct ip6_hdr *ip6 = NULL;	/* Keep compiler happy. */
213 #endif
214 #ifdef INET
215 	struct ip *ip4 = NULL;		/* Keep compiler happy. */
216 #endif
217 
218 	eh = mtod(m, struct ether_header *);
219 	switch (le->eh_type) {
220 #ifdef INET6
221 	case ETHERTYPE_IPV6:
222 		ip6 = (struct ip6_hdr *)(eh + 1);
223 		th = (struct tcphdr *)(ip6 + 1);
224 		break;
225 #endif
226 #ifdef INET
227 	case ETHERTYPE_IP:
228 		ip4 = (struct ip *)(eh + 1);
229 		th = (struct tcphdr *)(ip4 + 1);
230 		break;
231 #endif
232 	}
233 	return (th);
234 }
235 
236 void
237 tcp_lro_free(struct lro_ctrl *lc)
238 {
239 	struct lro_entry *le;
240 	unsigned x;
241 
242 	/* reset LRO free list */
243 	LIST_INIT(&lc->lro_free);
244 
245 	/* free active mbufs, if any */
246 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
247 		tcp_lro_active_remove(le);
248 		m_freem(le->m_head);
249 	}
250 
251 	/* free hash table */
252 	free(lc->lro_hash, M_LRO);
253 	lc->lro_hash = NULL;
254 	lc->lro_hashsz = 0;
255 
256 	/* free mbuf array, if any */
257 	for (x = 0; x != lc->lro_mbuf_count; x++)
258 		m_freem(lc->lro_mbuf_data[x].mb);
259 	lc->lro_mbuf_count = 0;
260 
261 	/* free allocated memory, if any */
262 	free(lc->lro_mbuf_data, M_LRO);
263 	lc->lro_mbuf_data = NULL;
264 }
265 
266 static uint16_t
267 tcp_lro_csum_th(struct tcphdr *th)
268 {
269 	uint32_t ch;
270 	uint16_t *p, l;
271 
272 	ch = th->th_sum = 0x0000;
273 	l = th->th_off;
274 	p = (uint16_t *)th;
275 	while (l > 0) {
276 		ch += *p;
277 		p++;
278 		ch += *p;
279 		p++;
280 		l--;
281 	}
282 	while (ch > 0xffff)
283 		ch = (ch >> 16) + (ch & 0xffff);
284 
285 	return (ch & 0xffff);
286 }
287 
288 static uint16_t
289 tcp_lro_rx_csum_fixup(struct lro_entry *le, void *l3hdr, struct tcphdr *th,
290     uint16_t tcp_data_len, uint16_t csum)
291 {
292 	uint32_t c;
293 	uint16_t cs;
294 
295 	c = csum;
296 
297 	/* Remove length from checksum. */
298 	switch (le->eh_type) {
299 #ifdef INET6
300 	case ETHERTYPE_IPV6:
301 	{
302 		struct ip6_hdr *ip6;
303 
304 		ip6 = (struct ip6_hdr *)l3hdr;
305 		if (le->append_cnt == 0)
306 			cs = ip6->ip6_plen;
307 		else {
308 			uint32_t cx;
309 
310 			cx = ntohs(ip6->ip6_plen);
311 			cs = in6_cksum_pseudo(ip6, cx, ip6->ip6_nxt, 0);
312 		}
313 		break;
314 	}
315 #endif
316 #ifdef INET
317 	case ETHERTYPE_IP:
318 	{
319 		struct ip *ip4;
320 
321 		ip4 = (struct ip *)l3hdr;
322 		if (le->append_cnt == 0)
323 			cs = ip4->ip_len;
324 		else {
325 			cs = in_addword(ntohs(ip4->ip_len) - sizeof(*ip4),
326 			    IPPROTO_TCP);
327 			cs = in_pseudo(ip4->ip_src.s_addr, ip4->ip_dst.s_addr,
328 			    htons(cs));
329 		}
330 		break;
331 	}
332 #endif
333 	default:
334 		cs = 0;		/* Keep compiler happy. */
335 	}
336 
337 	cs = ~cs;
338 	c += cs;
339 
340 	/* Remove TCP header csum. */
341 	cs = ~tcp_lro_csum_th(th);
342 	c += cs;
343 	while (c > 0xffff)
344 		c = (c >> 16) + (c & 0xffff);
345 
346 	return (c & 0xffff);
347 }
348 
349 static void
350 tcp_lro_rx_done(struct lro_ctrl *lc)
351 {
352 	struct lro_entry *le;
353 
354 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
355 		tcp_lro_active_remove(le);
356 		tcp_lro_flush(lc, le);
357 	}
358 }
359 
360 void
361 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
362 {
363 	struct lro_entry *le, *le_tmp;
364 	struct timeval tv;
365 
366 	if (LIST_EMPTY(&lc->lro_active))
367 		return;
368 
369 	getmicrouptime(&tv);
370 	timevalsub(&tv, timeout);
371 	LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
372 		if (timevalcmp(&tv, &le->mtime, >=)) {
373 			tcp_lro_active_remove(le);
374 			tcp_lro_flush(lc, le);
375 		}
376 	}
377 }
378 
379 #ifdef INET6
380 static int
381 tcp_lro_rx_ipv6(struct lro_ctrl *lc, struct mbuf *m, struct ip6_hdr *ip6,
382     struct tcphdr **th)
383 {
384 
385 	/* XXX-BZ we should check the flow-label. */
386 
387 	/* XXX-BZ We do not yet support ext. hdrs. */
388 	if (ip6->ip6_nxt != IPPROTO_TCP)
389 		return (TCP_LRO_NOT_SUPPORTED);
390 
391 	/* Find the TCP header. */
392 	*th = (struct tcphdr *)(ip6 + 1);
393 
394 	return (0);
395 }
396 #endif
397 
398 #ifdef INET
399 static int
400 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4,
401     struct tcphdr **th)
402 {
403 	int csum_flags;
404 	uint16_t csum;
405 
406 	if (ip4->ip_p != IPPROTO_TCP)
407 		return (TCP_LRO_NOT_SUPPORTED);
408 
409 	/* Ensure there are no options. */
410 	if ((ip4->ip_hl << 2) != sizeof (*ip4))
411 		return (TCP_LRO_CANNOT);
412 
413 	/* .. and the packet is not fragmented. */
414 	if (ip4->ip_off & htons(IP_MF|IP_OFFMASK))
415 		return (TCP_LRO_CANNOT);
416 
417 	/* Legacy IP has a header checksum that needs to be correct. */
418 	csum_flags = m->m_pkthdr.csum_flags;
419 	if (csum_flags & CSUM_IP_CHECKED) {
420 		if (__predict_false((csum_flags & CSUM_IP_VALID) == 0)) {
421 			lc->lro_bad_csum++;
422 			return (TCP_LRO_CANNOT);
423 		}
424 	} else {
425 		csum = in_cksum_hdr(ip4);
426 		if (__predict_false((csum) != 0)) {
427 			lc->lro_bad_csum++;
428 			return (TCP_LRO_CANNOT);
429 		}
430 	}
431 	/* Find the TCP header (we assured there are no IP options). */
432 	*th = (struct tcphdr *)(ip4 + 1);
433 	return (0);
434 }
435 #endif
436 
437 static void
438 tcp_lro_log(struct tcpcb *tp, struct lro_ctrl *lc,
439 	    struct lro_entry *le, struct mbuf *m, int frm, int32_t tcp_data_len,
440 	    uint32_t th_seq , uint32_t th_ack, uint16_t th_win)
441 {
442 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
443 		union tcp_log_stackspecific log;
444 		struct timeval tv;
445 		uint32_t cts;
446 
447 		cts = tcp_get_usecs(&tv);
448 		memset(&log, 0, sizeof(union tcp_log_stackspecific));
449 		log.u_bbr.flex8 = frm;
450 		log.u_bbr.flex1 = tcp_data_len;
451 		if (m)
452 			log.u_bbr.flex2 = m->m_pkthdr.len;
453 		else
454 			log.u_bbr.flex2 = 0;
455 		log.u_bbr.flex3 = le->append_cnt;
456 		log.u_bbr.flex4 = le->p_len;
457 		log.u_bbr.flex5 = le->m_head->m_pkthdr.len;
458 		log.u_bbr.delRate = le->m_head->m_flags;
459 		log.u_bbr.rttProp = le->m_head->m_pkthdr.rcv_tstmp;
460 		log.u_bbr.flex6 = lc->lro_length_lim;
461 		log.u_bbr.flex7 = lc->lro_ackcnt_lim;
462 		log.u_bbr.inflight = th_seq;
463 		log.u_bbr.timeStamp = cts;
464 		log.u_bbr.epoch = le->next_seq;
465 		log.u_bbr.delivered = th_ack;
466 		log.u_bbr.lt_epoch = le->ack_seq;
467 		log.u_bbr.pacing_gain = th_win;
468 		log.u_bbr.cwnd_gain = le->window;
469 		log.u_bbr.cur_del_rate = (uintptr_t)m;
470 		log.u_bbr.bw_inuse = (uintptr_t)le->m_head;
471 		log.u_bbr.pkts_out = le->mbuf_cnt;	/* Total mbufs added */
472 		log.u_bbr.applimited = le->ulp_csum;
473 		log.u_bbr.lost = le->mbuf_appended;
474 		TCP_LOG_EVENTP(tp, NULL,
475 			       &tp->t_inpcb->inp_socket->so_rcv,
476 			       &tp->t_inpcb->inp_socket->so_snd,
477 			       TCP_LOG_LRO, 0,
478 			       0, &log, false, &tv);
479 	}
480 }
481 
482 static void
483 tcp_flush_out_le(struct tcpcb *tp, struct lro_ctrl *lc, struct lro_entry *le, int locked)
484 {
485 	if (le->append_cnt > 1) {
486 		struct tcphdr *th;
487 		uint16_t p_len;
488 
489 		p_len = htons(le->p_len);
490 		switch (le->eh_type) {
491 #ifdef INET6
492 		case ETHERTYPE_IPV6:
493 		{
494 			struct ip6_hdr *ip6;
495 
496 			ip6 = le->le_ip6;
497 			ip6->ip6_plen = p_len;
498 			th = (struct tcphdr *)(ip6 + 1);
499 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
500 			    CSUM_PSEUDO_HDR;
501 			le->p_len += ETHER_HDR_LEN + sizeof(*ip6);
502 			break;
503 		}
504 #endif
505 #ifdef INET
506 		case ETHERTYPE_IP:
507 		{
508 			struct ip *ip4;
509 			uint32_t cl;
510 			uint16_t c;
511 
512 			ip4 = le->le_ip4;
513 			/* Fix IP header checksum for new length. */
514 			c = ~ip4->ip_sum;
515 			cl = c;
516 			c = ~ip4->ip_len;
517 			cl += c + p_len;
518 			while (cl > 0xffff)
519 				cl = (cl >> 16) + (cl & 0xffff);
520 			c = cl;
521 			ip4->ip_sum = ~c;
522 			ip4->ip_len = p_len;
523 			th = (struct tcphdr *)(ip4 + 1);
524 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
525 			    CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
526 			le->p_len += ETHER_HDR_LEN;
527 			break;
528 		}
529 #endif
530 		default:
531 			th = NULL;	/* Keep compiler happy. */
532 		}
533 		le->m_head->m_pkthdr.csum_data = 0xffff;
534 		le->m_head->m_pkthdr.len = le->p_len;
535 
536 		/* Incorporate the latest ACK into the TCP header. */
537 		th->th_ack = le->ack_seq;
538 		th->th_win = le->window;
539 		/* Incorporate latest timestamp into the TCP header. */
540 		if (le->timestamp != 0) {
541 			uint32_t *ts_ptr;
542 
543 			ts_ptr = (uint32_t *)(th + 1);
544 			ts_ptr[1] = htonl(le->tsval);
545 			ts_ptr[2] = le->tsecr;
546 		}
547 		/* Update the TCP header checksum. */
548 		le->ulp_csum += p_len;
549 		le->ulp_csum += tcp_lro_csum_th(th);
550 		while (le->ulp_csum > 0xffff)
551 			le->ulp_csum = (le->ulp_csum >> 16) +
552 			    (le->ulp_csum & 0xffff);
553 		th->th_sum = (le->ulp_csum & 0xffff);
554 		th->th_sum = ~th->th_sum;
555 		if (tp && locked) {
556 			tcp_lro_log(tp, lc, le, NULL, 7, 0, 0, 0, 0);
557 		}
558 	}
559 	/*
560 	 * Break any chain, this is not set to NULL on the singleton
561 	 * case m_nextpkt points to m_head. Other case set them
562 	 * m_nextpkt to NULL in push_and_replace.
563 	 */
564 	le->m_head->m_nextpkt = NULL;
565 	le->m_head->m_pkthdr.lro_nsegs = le->append_cnt;
566 	if (tp && locked) {
567 		tcp_lro_log(tp, lc, le, le->m_head, 8, 0, 0, 0, 0);
568 	}
569 	(*lc->ifp->if_input)(lc->ifp, le->m_head);
570 	lc->lro_queued += le->append_cnt;
571 }
572 
573 static void
574 tcp_set_le_to_m(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m)
575 {
576 	struct ether_header *eh;
577 	void *l3hdr = NULL;		/* Keep compiler happy. */
578 	struct tcphdr *th;
579 #ifdef INET6
580 	struct ip6_hdr *ip6 = NULL;	/* Keep compiler happy. */
581 #endif
582 #ifdef INET
583 	struct ip *ip4 = NULL;		/* Keep compiler happy. */
584 #endif
585 	uint32_t *ts_ptr;
586 	int error, l, ts_failed = 0;
587 	uint16_t tcp_data_len;
588 	uint16_t csum;
589 
590 	error = -1;
591 	eh = mtod(m, struct ether_header *);
592 	/*
593 	 * We must reset the other pointers since the mbuf
594 	 * we were pointing too is about to go away.
595 	 */
596 	switch (le->eh_type) {
597 #ifdef INET6
598 	case ETHERTYPE_IPV6:
599 		l3hdr = ip6 = (struct ip6_hdr *)(eh + 1);
600 		error = tcp_lro_rx_ipv6(lc, m, ip6, &th);
601 		le->le_ip6 = ip6;
602 		le->source_ip6 = ip6->ip6_src;
603 		le->dest_ip6 = ip6->ip6_dst;
604 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN - sizeof(*ip6);
605 		break;
606 #endif
607 #ifdef INET
608 	case ETHERTYPE_IP:
609 		l3hdr = ip4 = (struct ip *)(eh + 1);
610 		error = tcp_lro_rx_ipv4(lc, m, ip4, &th);
611 		le->le_ip4 = ip4;
612 		le->source_ip4 = ip4->ip_src.s_addr;
613 		le->dest_ip4 = ip4->ip_dst.s_addr;
614 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN;
615 		break;
616 #endif
617 	}
618 	KASSERT(error == 0, ("%s: le=%p tcp_lro_rx_xxx failed\n",
619 				    __func__, le));
620 	ts_ptr = (uint32_t *)(th + 1);
621 	l = (th->th_off << 2);
622 	l -= sizeof(*th);
623 	if (l != 0 &&
624 	    (__predict_false(l != TCPOLEN_TSTAMP_APPA) ||
625 	     (*ts_ptr != ntohl(TCPOPT_NOP<<24|TCPOPT_NOP<<16|
626 			       TCPOPT_TIMESTAMP<<8|TCPOLEN_TIMESTAMP)))) {
627 		/* We have failed to find a timestamp some other option? */
628 		ts_failed = 1;
629 	}
630 	if ((l != 0) && (ts_failed == 0)) {
631 		le->timestamp = 1;
632 		le->tsval = ntohl(*(ts_ptr + 1));
633 		le->tsecr = *(ts_ptr + 2);
634 	} else
635 		le->timestamp = 0;
636 	le->source_port = th->th_sport;
637 	le->dest_port = th->th_dport;
638 	/* Pull out the csum */
639 	tcp_data_len = m->m_pkthdr.lro_len;
640 	le->next_seq = ntohl(th->th_seq) + tcp_data_len;
641 	le->ack_seq = th->th_ack;
642 	le->window = th->th_win;
643 	csum = th->th_sum;
644 	/* Setup the data pointers */
645 	le->m_head = m;
646 	le->m_tail = m_last(m);
647 	le->append_cnt = 0;
648 	le->ulp_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th, tcp_data_len,
649 					     ~csum);
650 	le->append_cnt++;
651 	th->th_sum = csum;	/* Restore checksum on first packet. */
652 }
653 
654 static void
655 tcp_push_and_replace(struct tcpcb *tp, struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m, int locked)
656 {
657 	/*
658 	 * Push up the stack the current le and replace
659 	 * it with m.
660 	 */
661 	struct mbuf *msave;
662 
663 	/* Grab off the next and save it */
664 	msave = le->m_head->m_nextpkt;
665 	le->m_head->m_nextpkt = NULL;
666 	/* Now push out the old le entry */
667 	tcp_flush_out_le(tp, lc, le, locked);
668 	/*
669 	 * Now to replace the data properly in the le
670 	 * we have to reset the tcp header and
671 	 * other fields.
672 	 */
673 	tcp_set_le_to_m(lc, le, m);
674 	/* Restore the next list */
675 	m->m_nextpkt = msave;
676 }
677 
678 static void
679 tcp_lro_condense(struct tcpcb *tp, struct lro_ctrl *lc, struct lro_entry *le, int locked)
680 {
681 	/*
682 	 * Walk through the mbuf chain we
683 	 * have on tap and compress/condense
684 	 * as required.
685 	 */
686 	uint32_t *ts_ptr;
687 	struct mbuf *m;
688 	struct tcphdr *th;
689 	uint16_t tcp_data_len, csum_upd;
690 	int l;
691 
692 	/*
693 	 * First we must check the lead (m_head)
694 	 * we must make sure that it is *not*
695 	 * something that should be sent up
696 	 * right away (sack etc).
697 	 */
698 again:
699 
700 	m = le->m_head->m_nextpkt;
701 	if (m == NULL) {
702 		/* Just the one left */
703 		return;
704 	}
705 	th = tcp_lro_get_th(le, le->m_head);
706 	KASSERT(th != NULL,
707 		("le:%p m:%p th comes back NULL?", le, le->m_head));
708 	l = (th->th_off << 2);
709 	l -= sizeof(*th);
710 	ts_ptr = (uint32_t *)(th + 1);
711 	if (l != 0 && (__predict_false(l != TCPOLEN_TSTAMP_APPA) ||
712 		       (*ts_ptr != ntohl(TCPOPT_NOP<<24|TCPOPT_NOP<<16|
713 					 TCPOPT_TIMESTAMP<<8|TCPOLEN_TIMESTAMP)))) {
714 		/*
715 		 * Its not the timestamp. We can't
716 		 * use this guy as the head.
717 		 */
718 		le->m_head->m_nextpkt = m->m_nextpkt;
719 		tcp_push_and_replace(tp, lc, le, m, locked);
720 		goto again;
721 	}
722 	if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0) {
723 		/*
724 		 * Make sure that previously seen segements/ACKs are delivered
725 		 * before this segment, e.g. FIN.
726 		 */
727 		le->m_head->m_nextpkt = m->m_nextpkt;
728 		tcp_push_and_replace(tp, lc, le, m, locked);
729 		goto again;
730 	}
731 	while((m = le->m_head->m_nextpkt) != NULL) {
732 		/*
733 		 * condense m into le, first
734 		 * pull m out of the list.
735 		 */
736 		le->m_head->m_nextpkt = m->m_nextpkt;
737 		m->m_nextpkt = NULL;
738 		/* Setup my data */
739 		tcp_data_len = m->m_pkthdr.lro_len;
740 		th = tcp_lro_get_th(le, m);
741 		KASSERT(th != NULL,
742 			("le:%p m:%p th comes back NULL?", le, m));
743 		ts_ptr = (uint32_t *)(th + 1);
744 		l = (th->th_off << 2);
745 		l -= sizeof(*th);
746 		if (tp && locked) {
747 			tcp_lro_log(tp, lc, le, m, 1, 0, 0, 0, 0);
748 		}
749 		if (le->append_cnt >= lc->lro_ackcnt_lim) {
750 			if (tp && locked) {
751 				tcp_lro_log(tp, lc, le, m, 2, 0, 0, 0, 0);
752 			}
753 			tcp_push_and_replace(tp, lc, le, m, locked);
754 			goto again;
755 		}
756 		if (le->p_len > (lc->lro_length_lim - tcp_data_len)) {
757 			/* Flush now if appending will result in overflow. */
758 			if (tp && locked) {
759 				tcp_lro_log(tp, lc, le, m, 3, tcp_data_len, 0, 0, 0);
760 			}
761 			tcp_push_and_replace(tp, lc, le, m, locked);
762 			goto again;
763 		}
764 		if (l != 0 && (__predict_false(l != TCPOLEN_TSTAMP_APPA) ||
765 			       (*ts_ptr != ntohl(TCPOPT_NOP<<24|TCPOPT_NOP<<16|
766 						 TCPOPT_TIMESTAMP<<8|TCPOLEN_TIMESTAMP)))) {
767 			/*
768 			 * Maybe a sack in the new one? We need to
769 			 * start all over after flushing the
770 			 * current le. We will go up to the beginning
771 			 * and flush it (calling the replace again possibly
772 			 * or just returning).
773 			 */
774 			tcp_push_and_replace(tp, lc, le, m, locked);
775 			goto again;
776 		}
777 		if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0) {
778 			tcp_push_and_replace(tp, lc, le, m, locked);
779 			goto again;
780 		}
781 		if (l != 0) {
782 			uint32_t tsval = ntohl(*(ts_ptr + 1));
783 			/* Make sure timestamp values are increasing. */
784 			if (TSTMP_GT(le->tsval, tsval))  {
785 				tcp_push_and_replace(tp, lc, le, m, locked);
786 				goto again;
787 			}
788 			le->tsval = tsval;
789 			le->tsecr = *(ts_ptr + 2);
790 		}
791 		/* Try to append the new segment. */
792 		if (__predict_false(ntohl(th->th_seq) != le->next_seq ||
793 				    (tcp_data_len == 0 &&
794 				     le->ack_seq == th->th_ack &&
795 				     le->window == th->th_win))) {
796 			/* Out of order packet or duplicate ACK. */
797 			if (tp && locked) {
798 				tcp_lro_log(tp, lc, le, m, 4, tcp_data_len,
799 					    ntohl(th->th_seq),
800 					    th->th_ack,
801 					    th->th_win);
802 			}
803 			tcp_push_and_replace(tp, lc, le, m, locked);
804 			goto again;
805 		}
806 		if (tcp_data_len || SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
807 			le->next_seq += tcp_data_len;
808 			le->ack_seq = th->th_ack;
809 			le->window = th->th_win;
810 		} else if (th->th_ack == le->ack_seq) {
811 			le->window = WIN_MAX(le->window, th->th_win);
812 		}
813 		csum_upd = m->m_pkthdr.lro_csum;
814 		le->ulp_csum += csum_upd;
815 		if (tcp_data_len == 0) {
816 			le->append_cnt++;
817 			le->mbuf_cnt--;
818 			if (tp && locked) {
819 				tcp_lro_log(tp, lc, le, m, 5, tcp_data_len,
820 					    ntohl(th->th_seq),
821 					    th->th_ack,
822 					    th->th_win);
823 			}
824 			m_freem(m);
825 			continue;
826 		}
827 		le->append_cnt++;
828 		le->mbuf_appended++;
829 		le->p_len += tcp_data_len;
830 		/*
831 		 * Adjust the mbuf so that m_data points to the first byte of
832 		 * the ULP payload.  Adjust the mbuf to avoid complications and
833 		 * append new segment to existing mbuf chain.
834 		 */
835 		m_adj(m, m->m_pkthdr.len - tcp_data_len);
836 		if (tp && locked) {
837 			tcp_lro_log(tp, lc, le, m, 6, tcp_data_len,
838 					    ntohl(th->th_seq),
839 					    th->th_ack,
840 					    th->th_win);
841 		}
842 		m_demote_pkthdr(m);
843 		le->m_tail->m_next = m;
844 		le->m_tail = m_last(m);
845 	}
846 }
847 
848 #ifdef TCPHPTS
849 static void
850 tcp_queue_pkts(struct tcpcb *tp, struct lro_entry *le)
851 {
852 	if (tp->t_in_pkt == NULL) {
853 		/* Nothing yet there */
854 		tp->t_in_pkt = le->m_head;
855 		tp->t_tail_pkt = le->m_last_mbuf;
856 	} else {
857 		/* Already some there */
858 		tp->t_tail_pkt->m_nextpkt = le->m_head;
859 		tp->t_tail_pkt = le->m_last_mbuf;
860 	}
861 	le->m_head = NULL;
862 	le->m_last_mbuf = NULL;
863 }
864 #endif
865 
866 void
867 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
868 {
869 	struct tcpcb *tp = NULL;
870 	int locked = 0;
871 #ifdef TCPHPTS
872 	struct inpcb *inp = NULL;
873 	int need_wakeup = 0, can_queue = 0;
874 	struct epoch_tracker et;
875 
876 	/* Now lets lookup the inp first */
877 	CURVNET_SET(lc->ifp->if_vnet);
878 	/*
879 	 * XXXRRS Currently the common input handler for
880 	 * mbuf queuing cannot handle VLAN Tagged. This needs
881 	 * to be fixed and the or condition removed (i.e. the
882 	 * common code should do the right lookup for the vlan
883 	 * tag and anything else that the vlan_input() does).
884 	 */
885 	if ((tcplro_stacks_wanting_mbufq == 0) || (le->m_head->m_flags & M_VLANTAG))
886 		goto skip_lookup;
887 	NET_EPOCH_ENTER(et);
888 	switch (le->eh_type) {
889 #ifdef INET6
890 	case ETHERTYPE_IPV6:
891 		inp = in6_pcblookup(&V_tcbinfo, &le->source_ip6,
892 				    le->source_port, &le->dest_ip6,le->dest_port,
893 				    INPLOOKUP_WLOCKPCB,
894 				    lc->ifp);
895 		break;
896 #endif
897 #ifdef INET
898 	case ETHERTYPE_IP:
899 		inp = in_pcblookup(&V_tcbinfo, le->le_ip4->ip_src,
900 				   le->source_port, le->le_ip4->ip_dst, le->dest_port,
901 				   INPLOOKUP_WLOCKPCB,
902 				   lc->ifp);
903 		break;
904 #endif
905 	}
906 	NET_EPOCH_EXIT(et);
907 	if (inp && ((inp->inp_flags & (INP_DROPPED|INP_TIMEWAIT)) ||
908 		    (inp->inp_flags2 & INP_FREED))) {
909 		/* We don't want this guy */
910 		INP_WUNLOCK(inp);
911 		inp = NULL;
912 	}
913 	if (inp && (inp->inp_flags2 & INP_SUPPORTS_MBUFQ)) {
914 		/* The transport supports mbuf queuing */
915 		can_queue = 1;
916 		if (le->need_wakeup ||
917 		    ((inp->inp_in_input == 0) &&
918 		     ((inp->inp_flags2 & INP_MBUF_QUEUE_READY) == 0))) {
919 			/*
920 			 * Either the transport is off on a keep-alive
921 			 * (it has the queue_ready flag clear and its
922 			 *  not already been woken) or the entry has
923 			 * some urgent thing (FIN or possibly SACK blocks).
924 			 * This means we need to wake the transport up by
925 			 * putting it on the input pacer.
926 			 */
927 			need_wakeup = 1;
928 			if ((inp->inp_flags2 & INP_DONT_SACK_QUEUE) &&
929 			    (le->need_wakeup != 1)) {
930 				/*
931 				 * Prohibited from a sack wakeup.
932 				 */
933 				need_wakeup = 0;
934 			}
935 		}
936 		/* Do we need to be awoken due to lots of data or acks? */
937 		if ((le->tcp_tot_p_len >= lc->lro_length_lim) ||
938 		    (le->mbuf_cnt >= lc->lro_ackcnt_lim))
939 			need_wakeup = 1;
940 	}
941 	if (inp) {
942 		tp = intotcpcb(inp);
943 		locked = 1;
944 	} else
945 		tp = NULL;
946 	if (can_queue) {
947 		counter_u64_add(tcp_inp_lro_direct_queue, 1);
948 		tcp_lro_log(tp, lc, le, NULL, 22, need_wakeup,
949 			    inp->inp_flags2, inp->inp_in_input, le->need_wakeup);
950 		tcp_queue_pkts(tp, le);
951 		if (need_wakeup) {
952 			/*
953 			 * We must get the guy to wakeup via
954 			 * hpts.
955 			 */
956 			counter_u64_add(tcp_inp_lro_wokeup_queue, 1);
957 			if (le->need_wakeup)
958 				counter_u64_add(tcp_inp_lro_sack_wake, 1);
959 			tcp_queue_to_input(inp);
960 		}
961 	}
962 	if (inp && (hold_lock_over_compress == 0)) {
963 		/* Unlock it */
964 		locked = 0;
965 		tp = NULL;
966 		counter_u64_add(tcp_inp_lro_locks_taken, 1);
967 		INP_WUNLOCK(inp);
968 	}
969 	if (can_queue == 0) {
970 skip_lookup:
971 #endif /* TCPHPTS */
972 		/* Old fashioned lro method */
973 		if (le->m_head != le->m_last_mbuf)  {
974 			counter_u64_add(tcp_inp_lro_compressed, 1);
975 			tcp_lro_condense(tp, lc, le, locked);
976 		} else
977 			counter_u64_add(tcp_inp_lro_single_push, 1);
978 		tcp_flush_out_le(tp, lc, le, locked);
979 #ifdef TCPHPTS
980 	}
981 	if (inp && locked) {
982 		counter_u64_add(tcp_inp_lro_locks_taken, 1);
983 		INP_WUNLOCK(inp);
984 	}
985 	CURVNET_RESTORE();
986 #endif
987 	lc->lro_flushed++;
988 	bzero(le, sizeof(*le));
989 	LIST_INSERT_HEAD(&lc->lro_free, le, next);
990 }
991 
992 #ifdef HAVE_INLINE_FLSLL
993 #define	tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
994 #else
995 static inline uint64_t
996 tcp_lro_msb_64(uint64_t x)
997 {
998 	x |= (x >> 1);
999 	x |= (x >> 2);
1000 	x |= (x >> 4);
1001 	x |= (x >> 8);
1002 	x |= (x >> 16);
1003 	x |= (x >> 32);
1004 	return (x & ~(x >> 1));
1005 }
1006 #endif
1007 
1008 /*
1009  * The tcp_lro_sort() routine is comparable to qsort(), except it has
1010  * a worst case complexity limit of O(MIN(N,64)*N), where N is the
1011  * number of elements to sort and 64 is the number of sequence bits
1012  * available. The algorithm is bit-slicing the 64-bit sequence number,
1013  * sorting one bit at a time from the most significant bit until the
1014  * least significant one, skipping the constant bits. This is
1015  * typically called a radix sort.
1016  */
1017 static void
1018 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
1019 {
1020 	struct lro_mbuf_sort temp;
1021 	uint64_t ones;
1022 	uint64_t zeros;
1023 	uint32_t x;
1024 	uint32_t y;
1025 
1026 repeat:
1027 	/* for small arrays insertion sort is faster */
1028 	if (size <= 12) {
1029 		for (x = 1; x < size; x++) {
1030 			temp = parray[x];
1031 			for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
1032 				parray[y] = parray[y - 1];
1033 			parray[y] = temp;
1034 		}
1035 		return;
1036 	}
1037 
1038 	/* compute sequence bits which are constant */
1039 	ones = 0;
1040 	zeros = 0;
1041 	for (x = 0; x != size; x++) {
1042 		ones |= parray[x].seq;
1043 		zeros |= ~parray[x].seq;
1044 	}
1045 
1046 	/* compute bits which are not constant into "ones" */
1047 	ones &= zeros;
1048 	if (ones == 0)
1049 		return;
1050 
1051 	/* pick the most significant bit which is not constant */
1052 	ones = tcp_lro_msb_64(ones);
1053 
1054 	/*
1055 	 * Move entries having cleared sequence bits to the beginning
1056 	 * of the array:
1057 	 */
1058 	for (x = y = 0; y != size; y++) {
1059 		/* skip set bits */
1060 		if (parray[y].seq & ones)
1061 			continue;
1062 		/* swap entries */
1063 		temp = parray[x];
1064 		parray[x] = parray[y];
1065 		parray[y] = temp;
1066 		x++;
1067 	}
1068 
1069 	KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
1070 
1071 	/* sort zeros */
1072 	tcp_lro_sort(parray, x);
1073 
1074 	/* sort ones */
1075 	parray += x;
1076 	size -= x;
1077 	goto repeat;
1078 }
1079 
1080 void
1081 tcp_lro_flush_all(struct lro_ctrl *lc)
1082 {
1083 	uint64_t seq;
1084 	uint64_t nseq;
1085 	unsigned x;
1086 
1087 	/* check if no mbufs to flush */
1088 	if (lc->lro_mbuf_count == 0)
1089 		goto done;
1090 
1091 	/* sort all mbufs according to stream */
1092 	tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
1093 
1094 	/* input data into LRO engine, stream by stream */
1095 	seq = 0;
1096 	for (x = 0; x != lc->lro_mbuf_count; x++) {
1097 		struct mbuf *mb;
1098 
1099 		/* get mbuf */
1100 		mb = lc->lro_mbuf_data[x].mb;
1101 
1102 		/* get sequence number, masking away the packet index */
1103 		nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
1104 
1105 		/* check for new stream */
1106 		if (seq != nseq) {
1107 			seq = nseq;
1108 
1109 			/* flush active streams */
1110 			tcp_lro_rx_done(lc);
1111 		}
1112 
1113 		/* add packet to LRO engine */
1114 		if (tcp_lro_rx2(lc, mb, 0, 0) != 0) {
1115 			/* input packet to network layer */
1116 			(*lc->ifp->if_input)(lc->ifp, mb);
1117 			lc->lro_queued++;
1118 			lc->lro_flushed++;
1119 		}
1120 	}
1121 done:
1122 	/* flush active streams */
1123 	tcp_lro_rx_done(lc);
1124 
1125 	lc->lro_mbuf_count = 0;
1126 }
1127 
1128 static void
1129 lro_set_mtime(struct timeval *tv, struct timespec *ts)
1130 {
1131 	tv->tv_sec = ts->tv_sec;
1132 	tv->tv_usec = ts->tv_nsec / 1000;
1133 }
1134 
1135 static int
1136 tcp_lro_rx2(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, int use_hash)
1137 {
1138 	struct lro_entry *le;
1139 	struct ether_header *eh;
1140 #ifdef INET6
1141 	struct ip6_hdr *ip6 = NULL;	/* Keep compiler happy. */
1142 #endif
1143 #ifdef INET
1144 	struct ip *ip4 = NULL;		/* Keep compiler happy. */
1145 #endif
1146 	struct tcphdr *th;
1147 	void *l3hdr = NULL;		/* Keep compiler happy. */
1148 	uint32_t *ts_ptr;
1149 	tcp_seq seq;
1150 	int error, ip_len, l;
1151 	uint16_t eh_type, tcp_data_len, need_flush;
1152 	struct lro_head *bucket;
1153 	struct timespec arrv;
1154 
1155 	/* We expect a contiguous header [eh, ip, tcp]. */
1156 	if ((m->m_flags & (M_TSTMP_LRO|M_TSTMP)) == 0) {
1157 		/* If no hardware or arrival stamp on the packet add arrival */
1158 		nanouptime(&arrv);
1159 		m->m_pkthdr.rcv_tstmp = (arrv.tv_sec * 1000000000) + arrv.tv_nsec;
1160 		m->m_flags |= M_TSTMP_LRO;
1161 	}
1162 	eh = mtod(m, struct ether_header *);
1163 	eh_type = ntohs(eh->ether_type);
1164 	switch (eh_type) {
1165 #ifdef INET6
1166 	case ETHERTYPE_IPV6:
1167 	{
1168 		CURVNET_SET(lc->ifp->if_vnet);
1169 		if (V_ip6_forwarding != 0) {
1170 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
1171 			CURVNET_RESTORE();
1172 			return (TCP_LRO_CANNOT);
1173 		}
1174 		CURVNET_RESTORE();
1175 		l3hdr = ip6 = (struct ip6_hdr *)(eh + 1);
1176 		error = tcp_lro_rx_ipv6(lc, m, ip6, &th);
1177 		if (error != 0)
1178 			return (error);
1179 		tcp_data_len = ntohs(ip6->ip6_plen);
1180 		ip_len = sizeof(*ip6) + tcp_data_len;
1181 		break;
1182 	}
1183 #endif
1184 #ifdef INET
1185 	case ETHERTYPE_IP:
1186 	{
1187 		CURVNET_SET(lc->ifp->if_vnet);
1188 		if (V_ipforwarding != 0) {
1189 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
1190 			CURVNET_RESTORE();
1191 			return (TCP_LRO_CANNOT);
1192 		}
1193 		CURVNET_RESTORE();
1194 		l3hdr = ip4 = (struct ip *)(eh + 1);
1195 		error = tcp_lro_rx_ipv4(lc, m, ip4, &th);
1196 		if (error != 0)
1197 			return (error);
1198 		ip_len = ntohs(ip4->ip_len);
1199 		tcp_data_len = ip_len - sizeof(*ip4);
1200 		break;
1201 	}
1202 #endif
1203 	/* XXX-BZ what happens in case of VLAN(s)? */
1204 	default:
1205 		return (TCP_LRO_NOT_SUPPORTED);
1206 	}
1207 
1208 	/*
1209 	 * If the frame is padded beyond the end of the IP packet, then we must
1210 	 * trim the extra bytes off.
1211 	 */
1212 	l = m->m_pkthdr.len - (ETHER_HDR_LEN + ip_len);
1213 	if (l != 0) {
1214 		if (l < 0)
1215 			/* Truncated packet. */
1216 			return (TCP_LRO_CANNOT);
1217 
1218 		m_adj(m, -l);
1219 	}
1220 	/*
1221 	 * Check TCP header constraints.
1222 	 */
1223 	if (th->th_flags & TH_SYN)
1224 		return (TCP_LRO_CANNOT);
1225 	if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0)
1226 		need_flush = 1;
1227 	else
1228 		need_flush = 0;
1229 	l = (th->th_off << 2);
1230 	ts_ptr = (uint32_t *)(th + 1);
1231 	tcp_data_len -= l;
1232 	l -= sizeof(*th);
1233 	if (l != 0 && (__predict_false(l != TCPOLEN_TSTAMP_APPA) ||
1234 		       (*ts_ptr != ntohl(TCPOPT_NOP<<24|TCPOPT_NOP<<16|
1235 					 TCPOPT_TIMESTAMP<<8|TCPOLEN_TIMESTAMP)))) {
1236 		/*
1237 		 * We have an option besides Timestamps, maybe
1238 		 * it is a sack (most likely) which means we
1239 		 * will probably need to wake up a sleeper (if
1240 		 * the guy does queueing).
1241 		 */
1242 		need_flush = 2;
1243 	}
1244 
1245 	/* If the driver did not pass in the checksum, set it now. */
1246 	if (csum == 0x0000)
1247 		csum = th->th_sum;
1248 	seq = ntohl(th->th_seq);
1249 	if (!use_hash) {
1250 		bucket = &lc->lro_hash[0];
1251 	} else if (M_HASHTYPE_ISHASH(m)) {
1252 		bucket = &lc->lro_hash[m->m_pkthdr.flowid % lc->lro_hashsz];
1253 	} else {
1254 		uint32_t hash;
1255 
1256 		switch (eh_type) {
1257 #ifdef INET
1258 		case ETHERTYPE_IP:
1259 			hash = ip4->ip_src.s_addr + ip4->ip_dst.s_addr;
1260 			break;
1261 #endif
1262 #ifdef INET6
1263 		case ETHERTYPE_IPV6:
1264 			hash = ip6->ip6_src.s6_addr32[0] +
1265 				ip6->ip6_dst.s6_addr32[0];
1266 			hash += ip6->ip6_src.s6_addr32[1] +
1267 				ip6->ip6_dst.s6_addr32[1];
1268 			hash += ip6->ip6_src.s6_addr32[2] +
1269 				ip6->ip6_dst.s6_addr32[2];
1270 			hash += ip6->ip6_src.s6_addr32[3] +
1271 				ip6->ip6_dst.s6_addr32[3];
1272 			break;
1273 #endif
1274 		default:
1275 			hash = 0;
1276 			break;
1277 		}
1278 		hash += th->th_sport + th->th_dport;
1279 		bucket = &lc->lro_hash[hash % lc->lro_hashsz];
1280 	}
1281 
1282 	/* Try to find a matching previous segment. */
1283 	LIST_FOREACH(le, bucket, hash_next) {
1284 		if (le->eh_type != eh_type)
1285 			continue;
1286 		if (le->source_port != th->th_sport ||
1287 		    le->dest_port != th->th_dport)
1288 			continue;
1289 		switch (eh_type) {
1290 #ifdef INET6
1291 		case ETHERTYPE_IPV6:
1292 			if (bcmp(&le->source_ip6, &ip6->ip6_src,
1293 				 sizeof(struct in6_addr)) != 0 ||
1294 			    bcmp(&le->dest_ip6, &ip6->ip6_dst,
1295 				 sizeof(struct in6_addr)) != 0)
1296 				continue;
1297 			break;
1298 #endif
1299 #ifdef INET
1300 		case ETHERTYPE_IP:
1301 			if (le->source_ip4 != ip4->ip_src.s_addr ||
1302 			    le->dest_ip4 != ip4->ip_dst.s_addr)
1303 				continue;
1304 			break;
1305 #endif
1306 		}
1307 		if (tcp_data_len || SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq)) ||
1308 		    (th->th_ack == le->ack_seq)) {
1309 			m->m_pkthdr.lro_len = tcp_data_len;
1310 		} else {
1311 			/* no data and old ack */
1312 			m_freem(m);
1313 			return (0);
1314 		}
1315 		if (need_flush)
1316 			le->need_wakeup = need_flush;
1317 		/* Save of the data only csum */
1318 		m->m_pkthdr.rcvif = lc->ifp;
1319 		m->m_pkthdr.lro_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th,
1320 						      tcp_data_len, ~csum);
1321 		th->th_sum = csum;	/* Restore checksum */
1322 		/* Save off the tail I am appending too (prev) */
1323 		le->m_prev_last = le->m_last_mbuf;
1324 		/* Mark me in the last spot */
1325 		le->m_last_mbuf->m_nextpkt = m;
1326 		/* Now set the tail to me  */
1327 		le->m_last_mbuf = m;
1328 		le->mbuf_cnt++;
1329 		m->m_nextpkt = NULL;
1330 		/* Add to the total size of data */
1331 		le->tcp_tot_p_len += tcp_data_len;
1332 		lro_set_mtime(&le->mtime, &arrv);
1333 		return (0);
1334 	}
1335 	/* Try to find an empty slot. */
1336 	if (LIST_EMPTY(&lc->lro_free))
1337 		return (TCP_LRO_NO_ENTRIES);
1338 
1339 	/* Start a new segment chain. */
1340 	le = LIST_FIRST(&lc->lro_free);
1341 	LIST_REMOVE(le, next);
1342 	tcp_lro_active_insert(lc, bucket, le);
1343 	lro_set_mtime(&le->mtime, &arrv);
1344 
1345 	/* Start filling in details. */
1346 	switch (eh_type) {
1347 #ifdef INET6
1348 	case ETHERTYPE_IPV6:
1349 		le->le_ip6 = ip6;
1350 		le->source_ip6 = ip6->ip6_src;
1351 		le->dest_ip6 = ip6->ip6_dst;
1352 		le->eh_type = eh_type;
1353 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN - sizeof(*ip6);
1354 		break;
1355 #endif
1356 #ifdef INET
1357 	case ETHERTYPE_IP:
1358 		le->le_ip4 = ip4;
1359 		le->source_ip4 = ip4->ip_src.s_addr;
1360 		le->dest_ip4 = ip4->ip_dst.s_addr;
1361 		le->eh_type = eh_type;
1362 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN;
1363 		break;
1364 #endif
1365 	}
1366 	le->source_port = th->th_sport;
1367 	le->dest_port = th->th_dport;
1368 	le->next_seq = seq + tcp_data_len;
1369 	le->ack_seq = th->th_ack;
1370 	le->window = th->th_win;
1371 	if (l != 0) {
1372 		le->timestamp = 1;
1373 		le->tsval = ntohl(*(ts_ptr + 1));
1374 		le->tsecr = *(ts_ptr + 2);
1375 	}
1376 	KASSERT(le->ulp_csum == 0, ("%s: le=%p le->ulp_csum=0x%04x\n",
1377 				    __func__, le, le->ulp_csum));
1378 
1379 	le->append_cnt = 0;
1380 	le->ulp_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th, tcp_data_len,
1381 					     ~csum);
1382 	le->append_cnt++;
1383 	th->th_sum = csum;	/* Restore checksum */
1384 	le->m_head = m;
1385 	m->m_pkthdr.rcvif = lc->ifp;
1386 	le->mbuf_cnt = 1;
1387 	if (need_flush)
1388 		le->need_wakeup = need_flush;
1389 	else
1390 		le->need_wakeup = 0;
1391 	le->m_tail = m_last(m);
1392 	le->m_last_mbuf = m;
1393 	m->m_nextpkt = NULL;
1394 	le->m_prev_last = NULL;
1395 	/*
1396 	 * We keep the total size here for cross checking when we may need
1397 	 * to flush/wakeup in the MBUF_QUEUE case.
1398 	 */
1399 	le->tcp_tot_p_len = tcp_data_len;
1400 	m->m_pkthdr.lro_len = tcp_data_len;
1401 	return (0);
1402 }
1403 
1404 int
1405 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
1406 {
1407 
1408 	return tcp_lro_rx2(lc, m, csum, 1);
1409 }
1410 
1411 void
1412 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
1413 {
1414 	struct timespec arrv;
1415 
1416 	/* sanity checks */
1417 	if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
1418 	    lc->lro_mbuf_max == 0)) {
1419 		/* packet drop */
1420 		m_freem(mb);
1421 		return;
1422 	}
1423 
1424 	/* check if packet is not LRO capable */
1425 	if (__predict_false(mb->m_pkthdr.csum_flags == 0 ||
1426 	    (lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
1427 
1428 		/* input packet to network layer */
1429 		(*lc->ifp->if_input) (lc->ifp, mb);
1430 		return;
1431 	}
1432 	/* Arrival Stamp the packet */
1433 
1434 	if ((mb->m_flags & M_TSTMP) == 0) {
1435 		/* If no hardware or arrival stamp on the packet add arrival */
1436 		nanouptime(&arrv);
1437 		mb->m_pkthdr.rcv_tstmp = ((arrv.tv_sec * 1000000000) +
1438 			                  arrv.tv_nsec);
1439 		mb->m_flags |= M_TSTMP_LRO;
1440 	}
1441 	/* create sequence number */
1442 	lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
1443 	    (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
1444 	    (((uint64_t)mb->m_pkthdr.flowid) << 24) |
1445 	    ((uint64_t)lc->lro_mbuf_count);
1446 
1447 	/* enter mbuf */
1448 	lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb;
1449 
1450 	/* flush if array is full */
1451 	if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max))
1452 		tcp_lro_flush_all(lc);
1453 }
1454 
1455 /* end */
1456