xref: /freebsd/sys/netinet/tcp_lro.c (revision a0ee8cc636cd5c2374ec44ca71226564ea0bca95)
1 /*-
2  * Copyright (c) 2007, Myricom Inc.
3  * Copyright (c) 2008, Intel Corporation.
4  * Copyright (c) 2012 The FreeBSD Foundation
5  * Copyright (c) 2016 Mellanox Technologies.
6  * All rights reserved.
7  *
8  * Portions of this software were developed by Bjoern Zeeb
9  * under sponsorship from the FreeBSD Foundation.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include "opt_inet.h"
37 #include "opt_inet6.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/mbuf.h>
42 #include <sys/kernel.h>
43 #include <sys/socket.h>
44 
45 #include <net/if.h>
46 #include <net/if_var.h>
47 #include <net/ethernet.h>
48 #include <net/vnet.h>
49 
50 #include <netinet/in_systm.h>
51 #include <netinet/in.h>
52 #include <netinet/ip6.h>
53 #include <netinet/ip.h>
54 #include <netinet/ip_var.h>
55 #include <netinet/tcp.h>
56 #include <netinet/tcp_lro.h>
57 
58 #include <netinet6/ip6_var.h>
59 
60 #include <machine/in_cksum.h>
61 
62 static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures");
63 
64 #define	TCP_LRO_UPDATE_CSUM	1
65 #ifndef	TCP_LRO_UPDATE_CSUM
66 #define	TCP_LRO_INVALID_CSUM	0x0000
67 #endif
68 
69 int
70 tcp_lro_init(struct lro_ctrl *lc)
71 {
72 	return (tcp_lro_init_args(lc, NULL, TCP_LRO_ENTRIES, 0));
73 }
74 
75 int
76 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
77     unsigned lro_entries, unsigned lro_mbufs)
78 {
79 	struct lro_entry *le;
80 	size_t size;
81 	unsigned i;
82 
83 	lc->lro_bad_csum = 0;
84 	lc->lro_queued = 0;
85 	lc->lro_flushed = 0;
86 	lc->lro_cnt = 0;
87 	lc->lro_mbuf_count = 0;
88 	lc->lro_mbuf_max = lro_mbufs;
89 	lc->lro_cnt = lro_entries;
90 	lc->ifp = ifp;
91 	SLIST_INIT(&lc->lro_free);
92 	SLIST_INIT(&lc->lro_active);
93 
94 	/* compute size to allocate */
95 	size = (lro_mbufs * sizeof(struct mbuf *)) +
96 	    (lro_entries * sizeof(*le));
97 	lc->lro_mbuf_data = (struct mbuf **)
98 	    malloc(size, M_LRO, M_NOWAIT | M_ZERO);
99 
100 	/* check for out of memory */
101 	if (lc->lro_mbuf_data == NULL) {
102 		memset(lc, 0, sizeof(*lc));
103 		return (ENOMEM);
104 	}
105 	/* compute offset for LRO entries */
106 	le = (struct lro_entry *)
107 	    (lc->lro_mbuf_data + lro_mbufs);
108 
109 	/* setup linked list */
110 	for (i = 0; i != lro_entries; i++)
111 		SLIST_INSERT_HEAD(&lc->lro_free, le + i, next);
112 
113 	return (0);
114 }
115 
116 void
117 tcp_lro_free(struct lro_ctrl *lc)
118 {
119 	struct lro_entry *le;
120 	unsigned x;
121 
122 	/* reset LRO free list */
123 	SLIST_INIT(&lc->lro_free);
124 
125 	/* free active mbufs, if any */
126 	while ((le = SLIST_FIRST(&lc->lro_active)) != NULL) {
127 		SLIST_REMOVE_HEAD(&lc->lro_active, next);
128 		m_freem(le->m_head);
129 	}
130 
131 	/* free mbuf array, if any */
132 	for (x = 0; x != lc->lro_mbuf_count; x++)
133 		m_freem(lc->lro_mbuf_data[x]);
134 	lc->lro_mbuf_count = 0;
135 
136 	/* free allocated memory, if any */
137 	free(lc->lro_mbuf_data, M_LRO);
138 	lc->lro_mbuf_data = NULL;
139 }
140 
141 #ifdef TCP_LRO_UPDATE_CSUM
142 static uint16_t
143 tcp_lro_csum_th(struct tcphdr *th)
144 {
145 	uint32_t ch;
146 	uint16_t *p, l;
147 
148 	ch = th->th_sum = 0x0000;
149 	l = th->th_off;
150 	p = (uint16_t *)th;
151 	while (l > 0) {
152 		ch += *p;
153 		p++;
154 		ch += *p;
155 		p++;
156 		l--;
157 	}
158 	while (ch > 0xffff)
159 		ch = (ch >> 16) + (ch & 0xffff);
160 
161 	return (ch & 0xffff);
162 }
163 
164 static uint16_t
165 tcp_lro_rx_csum_fixup(struct lro_entry *le, void *l3hdr, struct tcphdr *th,
166     uint16_t tcp_data_len, uint16_t csum)
167 {
168 	uint32_t c;
169 	uint16_t cs;
170 
171 	c = csum;
172 
173 	/* Remove length from checksum. */
174 	switch (le->eh_type) {
175 #ifdef INET6
176 	case ETHERTYPE_IPV6:
177 	{
178 		struct ip6_hdr *ip6;
179 
180 		ip6 = (struct ip6_hdr *)l3hdr;
181 		if (le->append_cnt == 0)
182 			cs = ip6->ip6_plen;
183 		else {
184 			uint32_t cx;
185 
186 			cx = ntohs(ip6->ip6_plen);
187 			cs = in6_cksum_pseudo(ip6, cx, ip6->ip6_nxt, 0);
188 		}
189 		break;
190 	}
191 #endif
192 #ifdef INET
193 	case ETHERTYPE_IP:
194 	{
195 		struct ip *ip4;
196 
197 		ip4 = (struct ip *)l3hdr;
198 		if (le->append_cnt == 0)
199 			cs = ip4->ip_len;
200 		else {
201 			cs = in_addword(ntohs(ip4->ip_len) - sizeof(*ip4),
202 			    IPPROTO_TCP);
203 			cs = in_pseudo(ip4->ip_src.s_addr, ip4->ip_dst.s_addr,
204 			    htons(cs));
205 		}
206 		break;
207 	}
208 #endif
209 	default:
210 		cs = 0;		/* Keep compiler happy. */
211 	}
212 
213 	cs = ~cs;
214 	c += cs;
215 
216 	/* Remove TCP header csum. */
217 	cs = ~tcp_lro_csum_th(th);
218 	c += cs;
219 	while (c > 0xffff)
220 		c = (c >> 16) + (c & 0xffff);
221 
222 	return (c & 0xffff);
223 }
224 #endif
225 
226 void
227 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
228 {
229 	struct lro_entry *le, *le_tmp;
230 	struct timeval tv;
231 
232 	if (SLIST_EMPTY(&lc->lro_active))
233 		return;
234 
235 	getmicrotime(&tv);
236 	timevalsub(&tv, timeout);
237 	SLIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
238 		if (timevalcmp(&tv, &le->mtime, >=)) {
239 			SLIST_REMOVE(&lc->lro_active, le, lro_entry, next);
240 			tcp_lro_flush(lc, le);
241 		}
242 	}
243 }
244 
245 void
246 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
247 {
248 
249 	if (le->append_cnt > 0) {
250 		struct tcphdr *th;
251 		uint16_t p_len;
252 
253 		p_len = htons(le->p_len);
254 		switch (le->eh_type) {
255 #ifdef INET6
256 		case ETHERTYPE_IPV6:
257 		{
258 			struct ip6_hdr *ip6;
259 
260 			ip6 = le->le_ip6;
261 			ip6->ip6_plen = p_len;
262 			th = (struct tcphdr *)(ip6 + 1);
263 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
264 			    CSUM_PSEUDO_HDR;
265 			le->p_len += ETHER_HDR_LEN + sizeof(*ip6);
266 			break;
267 		}
268 #endif
269 #ifdef INET
270 		case ETHERTYPE_IP:
271 		{
272 			struct ip *ip4;
273 #ifdef TCP_LRO_UPDATE_CSUM
274 			uint32_t cl;
275 			uint16_t c;
276 #endif
277 
278 			ip4 = le->le_ip4;
279 #ifdef TCP_LRO_UPDATE_CSUM
280 			/* Fix IP header checksum for new length. */
281 			c = ~ip4->ip_sum;
282 			cl = c;
283 			c = ~ip4->ip_len;
284 			cl += c + p_len;
285 			while (cl > 0xffff)
286 				cl = (cl >> 16) + (cl & 0xffff);
287 			c = cl;
288 			ip4->ip_sum = ~c;
289 #else
290 			ip4->ip_sum = TCP_LRO_INVALID_CSUM;
291 #endif
292 			ip4->ip_len = p_len;
293 			th = (struct tcphdr *)(ip4 + 1);
294 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
295 			    CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
296 			le->p_len += ETHER_HDR_LEN;
297 			break;
298 		}
299 #endif
300 		default:
301 			th = NULL;	/* Keep compiler happy. */
302 		}
303 		le->m_head->m_pkthdr.csum_data = 0xffff;
304 		le->m_head->m_pkthdr.len = le->p_len;
305 
306 		/* Incorporate the latest ACK into the TCP header. */
307 		th->th_ack = le->ack_seq;
308 		th->th_win = le->window;
309 		/* Incorporate latest timestamp into the TCP header. */
310 		if (le->timestamp != 0) {
311 			uint32_t *ts_ptr;
312 
313 			ts_ptr = (uint32_t *)(th + 1);
314 			ts_ptr[1] = htonl(le->tsval);
315 			ts_ptr[2] = le->tsecr;
316 		}
317 #ifdef TCP_LRO_UPDATE_CSUM
318 		/* Update the TCP header checksum. */
319 		le->ulp_csum += p_len;
320 		le->ulp_csum += tcp_lro_csum_th(th);
321 		while (le->ulp_csum > 0xffff)
322 			le->ulp_csum = (le->ulp_csum >> 16) +
323 			    (le->ulp_csum & 0xffff);
324 		th->th_sum = (le->ulp_csum & 0xffff);
325 		th->th_sum = ~th->th_sum;
326 #else
327 		th->th_sum = TCP_LRO_INVALID_CSUM;
328 #endif
329 	}
330 
331 	(*lc->ifp->if_input)(lc->ifp, le->m_head);
332 	lc->lro_queued += le->append_cnt + 1;
333 	lc->lro_flushed++;
334 	bzero(le, sizeof(*le));
335 	SLIST_INSERT_HEAD(&lc->lro_free, le, next);
336 }
337 
338 static int
339 tcp_lro_mbuf_compare_header(const void *ppa, const void *ppb)
340 {
341 	const struct mbuf *ma = *((const struct mbuf * const *)ppa);
342 	const struct mbuf *mb = *((const struct mbuf * const *)ppb);
343 	int ret;
344 
345 	ret = M_HASHTYPE_GET(ma) - M_HASHTYPE_GET(mb);
346 	if (ret != 0)
347 		goto done;
348 
349 	ret = ma->m_pkthdr.flowid - mb->m_pkthdr.flowid;
350 	if (ret != 0)
351 		goto done;
352 
353 	ret = TCP_LRO_SEQUENCE(ma) - TCP_LRO_SEQUENCE(mb);
354 done:
355 	return (ret);
356 }
357 
358 void
359 tcp_lro_flush_all(struct lro_ctrl *lc)
360 {
361 	struct lro_entry *le;
362 	uint32_t hashtype;
363 	uint32_t flowid;
364 	unsigned x;
365 
366 	/* check if no mbufs to flush */
367 	if (__predict_false(lc->lro_mbuf_count == 0))
368 		goto done;
369 
370 	/* sort all mbufs according to stream */
371 	qsort(lc->lro_mbuf_data, lc->lro_mbuf_count, sizeof(struct mbuf *),
372 	    &tcp_lro_mbuf_compare_header);
373 
374 	/* input data into LRO engine, stream by stream */
375 	flowid = 0;
376 	hashtype = M_HASHTYPE_NONE;
377 	for (x = 0; x != lc->lro_mbuf_count; x++) {
378 		struct mbuf *mb;
379 
380 		mb = lc->lro_mbuf_data[x];
381 
382 		/* check for new stream */
383 		if (mb->m_pkthdr.flowid != flowid ||
384 		    M_HASHTYPE_GET(mb) != hashtype) {
385 			flowid = mb->m_pkthdr.flowid;
386 			hashtype = M_HASHTYPE_GET(mb);
387 
388 			/* flush active streams */
389 			while ((le = SLIST_FIRST(&lc->lro_active)) != NULL) {
390 				SLIST_REMOVE_HEAD(&lc->lro_active, next);
391 				tcp_lro_flush(lc, le);
392 			}
393 		}
394 #ifdef TCP_LRO_RESET_SEQUENCE
395 		/* reset sequence number */
396 		TCP_LRO_SEQUENCE(mb) = 0;
397 #endif
398 		/* add packet to LRO engine */
399 		if (tcp_lro_rx(lc, mb, 0) != 0) {
400 			/* input packet to network layer */
401 			(*lc->ifp->if_input)(lc->ifp, mb);
402 			lc->lro_queued++;
403 			lc->lro_flushed++;
404 		}
405 	}
406 done:
407 	/* flush active streams */
408 	while ((le = SLIST_FIRST(&lc->lro_active)) != NULL) {
409 		SLIST_REMOVE_HEAD(&lc->lro_active, next);
410 		tcp_lro_flush(lc, le);
411 	}
412 	lc->lro_mbuf_count = 0;
413 }
414 
415 #ifdef INET6
416 static int
417 tcp_lro_rx_ipv6(struct lro_ctrl *lc, struct mbuf *m, struct ip6_hdr *ip6,
418     struct tcphdr **th)
419 {
420 
421 	/* XXX-BZ we should check the flow-label. */
422 
423 	/* XXX-BZ We do not yet support ext. hdrs. */
424 	if (ip6->ip6_nxt != IPPROTO_TCP)
425 		return (TCP_LRO_NOT_SUPPORTED);
426 
427 	/* Find the TCP header. */
428 	*th = (struct tcphdr *)(ip6 + 1);
429 
430 	return (0);
431 }
432 #endif
433 
434 #ifdef INET
435 static int
436 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4,
437     struct tcphdr **th)
438 {
439 	int csum_flags;
440 	uint16_t csum;
441 
442 	if (ip4->ip_p != IPPROTO_TCP)
443 		return (TCP_LRO_NOT_SUPPORTED);
444 
445 	/* Ensure there are no options. */
446 	if ((ip4->ip_hl << 2) != sizeof (*ip4))
447 		return (TCP_LRO_CANNOT);
448 
449 	/* .. and the packet is not fragmented. */
450 	if (ip4->ip_off & htons(IP_MF|IP_OFFMASK))
451 		return (TCP_LRO_CANNOT);
452 
453 	/* Legacy IP has a header checksum that needs to be correct. */
454 	csum_flags = m->m_pkthdr.csum_flags;
455 	if (csum_flags & CSUM_IP_CHECKED) {
456 		if (__predict_false((csum_flags & CSUM_IP_VALID) == 0)) {
457 			lc->lro_bad_csum++;
458 			return (TCP_LRO_CANNOT);
459 		}
460 	} else {
461 		csum = in_cksum_hdr(ip4);
462 		if (__predict_false((csum) != 0)) {
463 			lc->lro_bad_csum++;
464 			return (TCP_LRO_CANNOT);
465 		}
466 	}
467 
468 	/* Find the TCP header (we assured there are no IP options). */
469 	*th = (struct tcphdr *)(ip4 + 1);
470 
471 	return (0);
472 }
473 #endif
474 
475 int
476 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
477 {
478 	struct lro_entry *le;
479 	struct ether_header *eh;
480 #ifdef INET6
481 	struct ip6_hdr *ip6 = NULL;	/* Keep compiler happy. */
482 #endif
483 #ifdef INET
484 	struct ip *ip4 = NULL;		/* Keep compiler happy. */
485 #endif
486 	struct tcphdr *th;
487 	void *l3hdr = NULL;		/* Keep compiler happy. */
488 	uint32_t *ts_ptr;
489 	tcp_seq seq;
490 	int error, ip_len, l;
491 	uint16_t eh_type, tcp_data_len;
492 
493 	/* We expect a contiguous header [eh, ip, tcp]. */
494 
495 	eh = mtod(m, struct ether_header *);
496 	eh_type = ntohs(eh->ether_type);
497 	switch (eh_type) {
498 #ifdef INET6
499 	case ETHERTYPE_IPV6:
500 	{
501 		CURVNET_SET(lc->ifp->if_vnet);
502 		if (V_ip6_forwarding != 0) {
503 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
504 			CURVNET_RESTORE();
505 			return (TCP_LRO_CANNOT);
506 		}
507 		CURVNET_RESTORE();
508 		l3hdr = ip6 = (struct ip6_hdr *)(eh + 1);
509 		error = tcp_lro_rx_ipv6(lc, m, ip6, &th);
510 		if (error != 0)
511 			return (error);
512 		tcp_data_len = ntohs(ip6->ip6_plen);
513 		ip_len = sizeof(*ip6) + tcp_data_len;
514 		break;
515 	}
516 #endif
517 #ifdef INET
518 	case ETHERTYPE_IP:
519 	{
520 		CURVNET_SET(lc->ifp->if_vnet);
521 		if (V_ipforwarding != 0) {
522 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
523 			CURVNET_RESTORE();
524 			return (TCP_LRO_CANNOT);
525 		}
526 		CURVNET_RESTORE();
527 		l3hdr = ip4 = (struct ip *)(eh + 1);
528 		error = tcp_lro_rx_ipv4(lc, m, ip4, &th);
529 		if (error != 0)
530 			return (error);
531 		ip_len = ntohs(ip4->ip_len);
532 		tcp_data_len = ip_len - sizeof(*ip4);
533 		break;
534 	}
535 #endif
536 	/* XXX-BZ what happens in case of VLAN(s)? */
537 	default:
538 		return (TCP_LRO_NOT_SUPPORTED);
539 	}
540 
541 	/*
542 	 * If the frame is padded beyond the end of the IP packet, then we must
543 	 * trim the extra bytes off.
544 	 */
545 	l = m->m_pkthdr.len - (ETHER_HDR_LEN + ip_len);
546 	if (l != 0) {
547 		if (l < 0)
548 			/* Truncated packet. */
549 			return (TCP_LRO_CANNOT);
550 
551 		m_adj(m, -l);
552 	}
553 
554 	/*
555 	 * Check TCP header constraints.
556 	 */
557 	/* Ensure no bits set besides ACK or PSH. */
558 	if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0)
559 		return (TCP_LRO_CANNOT);
560 
561 	/* XXX-BZ We lose a AKC|PUSH flag concatinating multiple segments. */
562 	/* XXX-BZ Ideally we'd flush on PUSH? */
563 
564 	/*
565 	 * Check for timestamps.
566 	 * Since the only option we handle are timestamps, we only have to
567 	 * handle the simple case of aligned timestamps.
568 	 */
569 	l = (th->th_off << 2);
570 	tcp_data_len -= l;
571 	l -= sizeof(*th);
572 	ts_ptr = (uint32_t *)(th + 1);
573 	if (l != 0 && (__predict_false(l != TCPOLEN_TSTAMP_APPA) ||
574 	    (*ts_ptr != ntohl(TCPOPT_NOP<<24|TCPOPT_NOP<<16|
575 	    TCPOPT_TIMESTAMP<<8|TCPOLEN_TIMESTAMP))))
576 		return (TCP_LRO_CANNOT);
577 
578 	/* If the driver did not pass in the checksum, set it now. */
579 	if (csum == 0x0000)
580 		csum = th->th_sum;
581 
582 	seq = ntohl(th->th_seq);
583 
584 	/* Try to find a matching previous segment. */
585 	SLIST_FOREACH(le, &lc->lro_active, next) {
586 		if (le->eh_type != eh_type)
587 			continue;
588 		if (le->source_port != th->th_sport ||
589 		    le->dest_port != th->th_dport)
590 			continue;
591 		switch (eh_type) {
592 #ifdef INET6
593 		case ETHERTYPE_IPV6:
594 			if (bcmp(&le->source_ip6, &ip6->ip6_src,
595 			    sizeof(struct in6_addr)) != 0 ||
596 			    bcmp(&le->dest_ip6, &ip6->ip6_dst,
597 			    sizeof(struct in6_addr)) != 0)
598 				continue;
599 			break;
600 #endif
601 #ifdef INET
602 		case ETHERTYPE_IP:
603 			if (le->source_ip4 != ip4->ip_src.s_addr ||
604 			    le->dest_ip4 != ip4->ip_dst.s_addr)
605 				continue;
606 			break;
607 #endif
608 		}
609 
610 		/* Flush now if appending will result in overflow. */
611 		if (le->p_len > (65535 - tcp_data_len)) {
612 			SLIST_REMOVE(&lc->lro_active, le, lro_entry, next);
613 			tcp_lro_flush(lc, le);
614 			break;
615 		}
616 
617 		/* Try to append the new segment. */
618 		if (__predict_false(seq != le->next_seq ||
619 		    (tcp_data_len == 0 && le->ack_seq == th->th_ack))) {
620 			/* Out of order packet or duplicate ACK. */
621 			SLIST_REMOVE(&lc->lro_active, le, lro_entry, next);
622 			tcp_lro_flush(lc, le);
623 			return (TCP_LRO_CANNOT);
624 		}
625 
626 		if (l != 0) {
627 			uint32_t tsval = ntohl(*(ts_ptr + 1));
628 			/* Make sure timestamp values are increasing. */
629 			/* XXX-BZ flip and use TSTMP_GEQ macro for this? */
630 			if (__predict_false(le->tsval > tsval ||
631 			    *(ts_ptr + 2) == 0))
632 				return (TCP_LRO_CANNOT);
633 			le->tsval = tsval;
634 			le->tsecr = *(ts_ptr + 2);
635 		}
636 
637 		le->next_seq += tcp_data_len;
638 		le->ack_seq = th->th_ack;
639 		le->window = th->th_win;
640 		le->append_cnt++;
641 
642 #ifdef TCP_LRO_UPDATE_CSUM
643 		le->ulp_csum += tcp_lro_rx_csum_fixup(le, l3hdr, th,
644 		    tcp_data_len, ~csum);
645 #endif
646 
647 		if (tcp_data_len == 0) {
648 			m_freem(m);
649 			return (0);
650 		}
651 
652 		le->p_len += tcp_data_len;
653 
654 		/*
655 		 * Adjust the mbuf so that m_data points to the first byte of
656 		 * the ULP payload.  Adjust the mbuf to avoid complications and
657 		 * append new segment to existing mbuf chain.
658 		 */
659 		m_adj(m, m->m_pkthdr.len - tcp_data_len);
660 		m_demote_pkthdr(m);
661 
662 		le->m_tail->m_next = m;
663 		le->m_tail = m_last(m);
664 
665 		/*
666 		 * If a possible next full length packet would cause an
667 		 * overflow, pro-actively flush now.
668 		 */
669 		if (le->p_len > (65535 - lc->ifp->if_mtu)) {
670 			SLIST_REMOVE(&lc->lro_active, le, lro_entry, next);
671 			tcp_lro_flush(lc, le);
672 		} else
673 			getmicrotime(&le->mtime);
674 
675 		return (0);
676 	}
677 
678 	/* Try to find an empty slot. */
679 	if (SLIST_EMPTY(&lc->lro_free))
680 		return (TCP_LRO_CANNOT);
681 
682 	/* Start a new segment chain. */
683 	le = SLIST_FIRST(&lc->lro_free);
684 	SLIST_REMOVE_HEAD(&lc->lro_free, next);
685 	SLIST_INSERT_HEAD(&lc->lro_active, le, next);
686 	getmicrotime(&le->mtime);
687 
688 	/* Start filling in details. */
689 	switch (eh_type) {
690 #ifdef INET6
691 	case ETHERTYPE_IPV6:
692 		le->le_ip6 = ip6;
693 		le->source_ip6 = ip6->ip6_src;
694 		le->dest_ip6 = ip6->ip6_dst;
695 		le->eh_type = eh_type;
696 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN - sizeof(*ip6);
697 		break;
698 #endif
699 #ifdef INET
700 	case ETHERTYPE_IP:
701 		le->le_ip4 = ip4;
702 		le->source_ip4 = ip4->ip_src.s_addr;
703 		le->dest_ip4 = ip4->ip_dst.s_addr;
704 		le->eh_type = eh_type;
705 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN;
706 		break;
707 #endif
708 	}
709 	le->source_port = th->th_sport;
710 	le->dest_port = th->th_dport;
711 
712 	le->next_seq = seq + tcp_data_len;
713 	le->ack_seq = th->th_ack;
714 	le->window = th->th_win;
715 	if (l != 0) {
716 		le->timestamp = 1;
717 		le->tsval = ntohl(*(ts_ptr + 1));
718 		le->tsecr = *(ts_ptr + 2);
719 	}
720 
721 #ifdef TCP_LRO_UPDATE_CSUM
722 	/*
723 	 * Do not touch the csum of the first packet.  However save the
724 	 * "adjusted" checksum of just the source and destination addresses,
725 	 * the next header and the TCP payload.  The length and TCP header
726 	 * parts may change, so we remove those from the saved checksum and
727 	 * re-add with final values on tcp_lro_flush() if needed.
728 	 */
729 	KASSERT(le->ulp_csum == 0, ("%s: le=%p le->ulp_csum=0x%04x\n",
730 	    __func__, le, le->ulp_csum));
731 
732 	le->ulp_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th, tcp_data_len,
733 	    ~csum);
734 	th->th_sum = csum;	/* Restore checksum on first packet. */
735 #endif
736 
737 	le->m_head = m;
738 	le->m_tail = m_last(m);
739 
740 	return (0);
741 }
742 
743 void
744 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
745 {
746 	/* sanity checks */
747 	if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
748 	    lc->lro_mbuf_max == 0)) {
749 		/* packet drop */
750 		m_freem(mb);
751 		return;
752 	}
753 
754 	/* check if packet is not LRO capable */
755 	if (__predict_false(mb->m_pkthdr.csum_flags == 0 ||
756 	    (lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
757 		lc->lro_flushed++;
758 		lc->lro_queued++;
759 
760 		/* input packet to network layer */
761 		(*lc->ifp->if_input) (lc->ifp, mb);
762 		return;
763 	}
764 
765 	/* check if array is full */
766 	if (__predict_false(lc->lro_mbuf_count == lc->lro_mbuf_max))
767 		tcp_lro_flush_all(lc);
768 
769 	/* store sequence number */
770 	TCP_LRO_SEQUENCE(mb) = lc->lro_mbuf_count;
771 
772 	/* enter mbuf */
773 	lc->lro_mbuf_data[lc->lro_mbuf_count++] = mb;
774 }
775 
776 /* end */
777