xref: /freebsd/sys/netinet/tcp_lro.c (revision 839529caa9c35f92b638dbe074655598e7a6bb6f)
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/kernel.h>
42 #include <sys/malloc.h>
43 #include <sys/mbuf.h>
44 #include <sys/socket.h>
45 
46 #include <net/if.h>
47 #include <net/if_var.h>
48 #include <net/ethernet.h>
49 #include <net/vnet.h>
50 
51 #include <netinet/in_systm.h>
52 #include <netinet/in.h>
53 #include <netinet/ip6.h>
54 #include <netinet/ip.h>
55 #include <netinet/ip_var.h>
56 #include <netinet/tcp.h>
57 #include <netinet/tcp_lro.h>
58 
59 #include <netinet6/ip6_var.h>
60 
61 #include <machine/in_cksum.h>
62 
63 static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures");
64 
65 #define	TCP_LRO_UPDATE_CSUM	1
66 #ifndef	TCP_LRO_UPDATE_CSUM
67 #define	TCP_LRO_INVALID_CSUM	0x0000
68 #endif
69 
70 int
71 tcp_lro_init(struct lro_ctrl *lc)
72 {
73 	return (tcp_lro_init_args(lc, NULL, TCP_LRO_ENTRIES, 0));
74 }
75 
76 int
77 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
78     unsigned lro_entries, unsigned lro_mbufs)
79 {
80 	struct lro_entry *le;
81 	size_t size;
82 	unsigned i;
83 
84 	lc->lro_bad_csum = 0;
85 	lc->lro_queued = 0;
86 	lc->lro_flushed = 0;
87 	lc->lro_cnt = 0;
88 	lc->lro_mbuf_count = 0;
89 	lc->lro_mbuf_max = lro_mbufs;
90 	lc->lro_cnt = lro_entries;
91 	lc->ifp = ifp;
92 	SLIST_INIT(&lc->lro_free);
93 	SLIST_INIT(&lc->lro_active);
94 
95 	/* compute size to allocate */
96 	size = (lro_mbufs * sizeof(struct mbuf *)) +
97 	    (lro_entries * sizeof(*le));
98 	lc->lro_mbuf_data = (struct mbuf **)
99 	    malloc(size, M_LRO, M_NOWAIT | M_ZERO);
100 
101 	/* check for out of memory */
102 	if (lc->lro_mbuf_data == NULL) {
103 		memset(lc, 0, sizeof(*lc));
104 		return (ENOMEM);
105 	}
106 	/* compute offset for LRO entries */
107 	le = (struct lro_entry *)
108 	    (lc->lro_mbuf_data + lro_mbufs);
109 
110 	/* setup linked list */
111 	for (i = 0; i != lro_entries; i++)
112 		SLIST_INSERT_HEAD(&lc->lro_free, le + i, next);
113 
114 	return (0);
115 }
116 
117 void
118 tcp_lro_free(struct lro_ctrl *lc)
119 {
120 	struct lro_entry *le;
121 	unsigned x;
122 
123 	/* reset LRO free list */
124 	SLIST_INIT(&lc->lro_free);
125 
126 	/* free active mbufs, if any */
127 	while ((le = SLIST_FIRST(&lc->lro_active)) != NULL) {
128 		SLIST_REMOVE_HEAD(&lc->lro_active, next);
129 		m_freem(le->m_head);
130 	}
131 
132 	/* free mbuf array, if any */
133 	for (x = 0; x != lc->lro_mbuf_count; x++)
134 		m_freem(lc->lro_mbuf_data[x]);
135 	lc->lro_mbuf_count = 0;
136 
137 	/* free allocated memory, if any */
138 	free(lc->lro_mbuf_data, M_LRO);
139 	lc->lro_mbuf_data = NULL;
140 }
141 
142 #ifdef TCP_LRO_UPDATE_CSUM
143 static uint16_t
144 tcp_lro_csum_th(struct tcphdr *th)
145 {
146 	uint32_t ch;
147 	uint16_t *p, l;
148 
149 	ch = th->th_sum = 0x0000;
150 	l = th->th_off;
151 	p = (uint16_t *)th;
152 	while (l > 0) {
153 		ch += *p;
154 		p++;
155 		ch += *p;
156 		p++;
157 		l--;
158 	}
159 	while (ch > 0xffff)
160 		ch = (ch >> 16) + (ch & 0xffff);
161 
162 	return (ch & 0xffff);
163 }
164 
165 static uint16_t
166 tcp_lro_rx_csum_fixup(struct lro_entry *le, void *l3hdr, struct tcphdr *th,
167     uint16_t tcp_data_len, uint16_t csum)
168 {
169 	uint32_t c;
170 	uint16_t cs;
171 
172 	c = csum;
173 
174 	/* Remove length from checksum. */
175 	switch (le->eh_type) {
176 #ifdef INET6
177 	case ETHERTYPE_IPV6:
178 	{
179 		struct ip6_hdr *ip6;
180 
181 		ip6 = (struct ip6_hdr *)l3hdr;
182 		if (le->append_cnt == 0)
183 			cs = ip6->ip6_plen;
184 		else {
185 			uint32_t cx;
186 
187 			cx = ntohs(ip6->ip6_plen);
188 			cs = in6_cksum_pseudo(ip6, cx, ip6->ip6_nxt, 0);
189 		}
190 		break;
191 	}
192 #endif
193 #ifdef INET
194 	case ETHERTYPE_IP:
195 	{
196 		struct ip *ip4;
197 
198 		ip4 = (struct ip *)l3hdr;
199 		if (le->append_cnt == 0)
200 			cs = ip4->ip_len;
201 		else {
202 			cs = in_addword(ntohs(ip4->ip_len) - sizeof(*ip4),
203 			    IPPROTO_TCP);
204 			cs = in_pseudo(ip4->ip_src.s_addr, ip4->ip_dst.s_addr,
205 			    htons(cs));
206 		}
207 		break;
208 	}
209 #endif
210 	default:
211 		cs = 0;		/* Keep compiler happy. */
212 	}
213 
214 	cs = ~cs;
215 	c += cs;
216 
217 	/* Remove TCP header csum. */
218 	cs = ~tcp_lro_csum_th(th);
219 	c += cs;
220 	while (c > 0xffff)
221 		c = (c >> 16) + (c & 0xffff);
222 
223 	return (c & 0xffff);
224 }
225 #endif
226 
227 void
228 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
229 {
230 	struct lro_entry *le, *le_tmp;
231 	struct timeval tv;
232 
233 	if (SLIST_EMPTY(&lc->lro_active))
234 		return;
235 
236 	getmicrotime(&tv);
237 	timevalsub(&tv, timeout);
238 	SLIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
239 		if (timevalcmp(&tv, &le->mtime, >=)) {
240 			SLIST_REMOVE(&lc->lro_active, le, lro_entry, next);
241 			tcp_lro_flush(lc, le);
242 		}
243 	}
244 }
245 
246 void
247 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
248 {
249 
250 	if (le->append_cnt > 0) {
251 		struct tcphdr *th;
252 		uint16_t p_len;
253 
254 		p_len = htons(le->p_len);
255 		switch (le->eh_type) {
256 #ifdef INET6
257 		case ETHERTYPE_IPV6:
258 		{
259 			struct ip6_hdr *ip6;
260 
261 			ip6 = le->le_ip6;
262 			ip6->ip6_plen = p_len;
263 			th = (struct tcphdr *)(ip6 + 1);
264 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
265 			    CSUM_PSEUDO_HDR;
266 			le->p_len += ETHER_HDR_LEN + sizeof(*ip6);
267 			break;
268 		}
269 #endif
270 #ifdef INET
271 		case ETHERTYPE_IP:
272 		{
273 			struct ip *ip4;
274 #ifdef TCP_LRO_UPDATE_CSUM
275 			uint32_t cl;
276 			uint16_t c;
277 #endif
278 
279 			ip4 = le->le_ip4;
280 #ifdef TCP_LRO_UPDATE_CSUM
281 			/* Fix IP header checksum for new length. */
282 			c = ~ip4->ip_sum;
283 			cl = c;
284 			c = ~ip4->ip_len;
285 			cl += c + p_len;
286 			while (cl > 0xffff)
287 				cl = (cl >> 16) + (cl & 0xffff);
288 			c = cl;
289 			ip4->ip_sum = ~c;
290 #else
291 			ip4->ip_sum = TCP_LRO_INVALID_CSUM;
292 #endif
293 			ip4->ip_len = p_len;
294 			th = (struct tcphdr *)(ip4 + 1);
295 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
296 			    CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
297 			le->p_len += ETHER_HDR_LEN;
298 			break;
299 		}
300 #endif
301 		default:
302 			th = NULL;	/* Keep compiler happy. */
303 		}
304 		le->m_head->m_pkthdr.csum_data = 0xffff;
305 		le->m_head->m_pkthdr.len = le->p_len;
306 
307 		/* Incorporate the latest ACK into the TCP header. */
308 		th->th_ack = le->ack_seq;
309 		th->th_win = le->window;
310 		/* Incorporate latest timestamp into the TCP header. */
311 		if (le->timestamp != 0) {
312 			uint32_t *ts_ptr;
313 
314 			ts_ptr = (uint32_t *)(th + 1);
315 			ts_ptr[1] = htonl(le->tsval);
316 			ts_ptr[2] = le->tsecr;
317 		}
318 #ifdef TCP_LRO_UPDATE_CSUM
319 		/* Update the TCP header checksum. */
320 		le->ulp_csum += p_len;
321 		le->ulp_csum += tcp_lro_csum_th(th);
322 		while (le->ulp_csum > 0xffff)
323 			le->ulp_csum = (le->ulp_csum >> 16) +
324 			    (le->ulp_csum & 0xffff);
325 		th->th_sum = (le->ulp_csum & 0xffff);
326 		th->th_sum = ~th->th_sum;
327 #else
328 		th->th_sum = TCP_LRO_INVALID_CSUM;
329 #endif
330 	}
331 
332 	(*lc->ifp->if_input)(lc->ifp, le->m_head);
333 	lc->lro_queued += le->append_cnt + 1;
334 	lc->lro_flushed++;
335 	bzero(le, sizeof(*le));
336 	SLIST_INSERT_HEAD(&lc->lro_free, le, next);
337 }
338 
339 static int
340 tcp_lro_mbuf_compare_header(const void *ppa, const void *ppb)
341 {
342 	const struct mbuf *ma = *((const struct mbuf * const *)ppa);
343 	const struct mbuf *mb = *((const struct mbuf * const *)ppb);
344 	int ret;
345 
346 	ret = M_HASHTYPE_GET(ma) - M_HASHTYPE_GET(mb);
347 	if (ret != 0)
348 		goto done;
349 
350 	if (ma->m_pkthdr.flowid > mb->m_pkthdr.flowid)
351 		return (1);
352 	else if (ma->m_pkthdr.flowid < mb->m_pkthdr.flowid)
353 		return (-1);
354 
355 	ret = TCP_LRO_SEQUENCE(ma) - TCP_LRO_SEQUENCE(mb);
356 done:
357 	return (ret);
358 }
359 
360 void
361 tcp_lro_flush_all(struct lro_ctrl *lc)
362 {
363 	struct lro_entry *le;
364 	uint32_t hashtype;
365 	uint32_t flowid;
366 	unsigned x;
367 
368 	/* check if no mbufs to flush */
369 	if (__predict_false(lc->lro_mbuf_count == 0))
370 		goto done;
371 
372 	/* sort all mbufs according to stream */
373 	qsort(lc->lro_mbuf_data, lc->lro_mbuf_count, sizeof(struct mbuf *),
374 	    &tcp_lro_mbuf_compare_header);
375 
376 	/* input data into LRO engine, stream by stream */
377 	flowid = 0;
378 	hashtype = M_HASHTYPE_NONE;
379 	for (x = 0; x != lc->lro_mbuf_count; x++) {
380 		struct mbuf *mb;
381 
382 		mb = lc->lro_mbuf_data[x];
383 
384 		/* check for new stream */
385 		if (mb->m_pkthdr.flowid != flowid ||
386 		    M_HASHTYPE_GET(mb) != hashtype) {
387 			flowid = mb->m_pkthdr.flowid;
388 			hashtype = M_HASHTYPE_GET(mb);
389 
390 			/* flush active streams */
391 			while ((le = SLIST_FIRST(&lc->lro_active)) != NULL) {
392 				SLIST_REMOVE_HEAD(&lc->lro_active, next);
393 				tcp_lro_flush(lc, le);
394 			}
395 		}
396 #ifdef TCP_LRO_RESET_SEQUENCE
397 		/* reset sequence number */
398 		TCP_LRO_SEQUENCE(mb) = 0;
399 #endif
400 		/* add packet to LRO engine */
401 		if (tcp_lro_rx(lc, mb, 0) != 0) {
402 			/* input packet to network layer */
403 			(*lc->ifp->if_input)(lc->ifp, mb);
404 			lc->lro_queued++;
405 			lc->lro_flushed++;
406 		}
407 	}
408 done:
409 	/* flush active streams */
410 	while ((le = SLIST_FIRST(&lc->lro_active)) != NULL) {
411 		SLIST_REMOVE_HEAD(&lc->lro_active, next);
412 		tcp_lro_flush(lc, le);
413 	}
414 	lc->lro_mbuf_count = 0;
415 }
416 
417 #ifdef INET6
418 static int
419 tcp_lro_rx_ipv6(struct lro_ctrl *lc, struct mbuf *m, struct ip6_hdr *ip6,
420     struct tcphdr **th)
421 {
422 
423 	/* XXX-BZ we should check the flow-label. */
424 
425 	/* XXX-BZ We do not yet support ext. hdrs. */
426 	if (ip6->ip6_nxt != IPPROTO_TCP)
427 		return (TCP_LRO_NOT_SUPPORTED);
428 
429 	/* Find the TCP header. */
430 	*th = (struct tcphdr *)(ip6 + 1);
431 
432 	return (0);
433 }
434 #endif
435 
436 #ifdef INET
437 static int
438 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4,
439     struct tcphdr **th)
440 {
441 	int csum_flags;
442 	uint16_t csum;
443 
444 	if (ip4->ip_p != IPPROTO_TCP)
445 		return (TCP_LRO_NOT_SUPPORTED);
446 
447 	/* Ensure there are no options. */
448 	if ((ip4->ip_hl << 2) != sizeof (*ip4))
449 		return (TCP_LRO_CANNOT);
450 
451 	/* .. and the packet is not fragmented. */
452 	if (ip4->ip_off & htons(IP_MF|IP_OFFMASK))
453 		return (TCP_LRO_CANNOT);
454 
455 	/* Legacy IP has a header checksum that needs to be correct. */
456 	csum_flags = m->m_pkthdr.csum_flags;
457 	if (csum_flags & CSUM_IP_CHECKED) {
458 		if (__predict_false((csum_flags & CSUM_IP_VALID) == 0)) {
459 			lc->lro_bad_csum++;
460 			return (TCP_LRO_CANNOT);
461 		}
462 	} else {
463 		csum = in_cksum_hdr(ip4);
464 		if (__predict_false((csum) != 0)) {
465 			lc->lro_bad_csum++;
466 			return (TCP_LRO_CANNOT);
467 		}
468 	}
469 
470 	/* Find the TCP header (we assured there are no IP options). */
471 	*th = (struct tcphdr *)(ip4 + 1);
472 
473 	return (0);
474 }
475 #endif
476 
477 int
478 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
479 {
480 	struct lro_entry *le;
481 	struct ether_header *eh;
482 #ifdef INET6
483 	struct ip6_hdr *ip6 = NULL;	/* Keep compiler happy. */
484 #endif
485 #ifdef INET
486 	struct ip *ip4 = NULL;		/* Keep compiler happy. */
487 #endif
488 	struct tcphdr *th;
489 	void *l3hdr = NULL;		/* Keep compiler happy. */
490 	uint32_t *ts_ptr;
491 	tcp_seq seq;
492 	int error, ip_len, l;
493 	uint16_t eh_type, tcp_data_len;
494 
495 	/* We expect a contiguous header [eh, ip, tcp]. */
496 
497 	eh = mtod(m, struct ether_header *);
498 	eh_type = ntohs(eh->ether_type);
499 	switch (eh_type) {
500 #ifdef INET6
501 	case ETHERTYPE_IPV6:
502 	{
503 		CURVNET_SET(lc->ifp->if_vnet);
504 		if (V_ip6_forwarding != 0) {
505 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
506 			CURVNET_RESTORE();
507 			return (TCP_LRO_CANNOT);
508 		}
509 		CURVNET_RESTORE();
510 		l3hdr = ip6 = (struct ip6_hdr *)(eh + 1);
511 		error = tcp_lro_rx_ipv6(lc, m, ip6, &th);
512 		if (error != 0)
513 			return (error);
514 		tcp_data_len = ntohs(ip6->ip6_plen);
515 		ip_len = sizeof(*ip6) + tcp_data_len;
516 		break;
517 	}
518 #endif
519 #ifdef INET
520 	case ETHERTYPE_IP:
521 	{
522 		CURVNET_SET(lc->ifp->if_vnet);
523 		if (V_ipforwarding != 0) {
524 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
525 			CURVNET_RESTORE();
526 			return (TCP_LRO_CANNOT);
527 		}
528 		CURVNET_RESTORE();
529 		l3hdr = ip4 = (struct ip *)(eh + 1);
530 		error = tcp_lro_rx_ipv4(lc, m, ip4, &th);
531 		if (error != 0)
532 			return (error);
533 		ip_len = ntohs(ip4->ip_len);
534 		tcp_data_len = ip_len - sizeof(*ip4);
535 		break;
536 	}
537 #endif
538 	/* XXX-BZ what happens in case of VLAN(s)? */
539 	default:
540 		return (TCP_LRO_NOT_SUPPORTED);
541 	}
542 
543 	/*
544 	 * If the frame is padded beyond the end of the IP packet, then we must
545 	 * trim the extra bytes off.
546 	 */
547 	l = m->m_pkthdr.len - (ETHER_HDR_LEN + ip_len);
548 	if (l != 0) {
549 		if (l < 0)
550 			/* Truncated packet. */
551 			return (TCP_LRO_CANNOT);
552 
553 		m_adj(m, -l);
554 	}
555 
556 	/*
557 	 * Check TCP header constraints.
558 	 */
559 	/* Ensure no bits set besides ACK or PSH. */
560 	if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0)
561 		return (TCP_LRO_CANNOT);
562 
563 	/* XXX-BZ We lose a AKC|PUSH flag concatinating multiple segments. */
564 	/* XXX-BZ Ideally we'd flush on PUSH? */
565 
566 	/*
567 	 * Check for timestamps.
568 	 * Since the only option we handle are timestamps, we only have to
569 	 * handle the simple case of aligned timestamps.
570 	 */
571 	l = (th->th_off << 2);
572 	tcp_data_len -= l;
573 	l -= sizeof(*th);
574 	ts_ptr = (uint32_t *)(th + 1);
575 	if (l != 0 && (__predict_false(l != TCPOLEN_TSTAMP_APPA) ||
576 	    (*ts_ptr != ntohl(TCPOPT_NOP<<24|TCPOPT_NOP<<16|
577 	    TCPOPT_TIMESTAMP<<8|TCPOLEN_TIMESTAMP))))
578 		return (TCP_LRO_CANNOT);
579 
580 	/* If the driver did not pass in the checksum, set it now. */
581 	if (csum == 0x0000)
582 		csum = th->th_sum;
583 
584 	seq = ntohl(th->th_seq);
585 
586 	/* Try to find a matching previous segment. */
587 	SLIST_FOREACH(le, &lc->lro_active, next) {
588 		if (le->eh_type != eh_type)
589 			continue;
590 		if (le->source_port != th->th_sport ||
591 		    le->dest_port != th->th_dport)
592 			continue;
593 		switch (eh_type) {
594 #ifdef INET6
595 		case ETHERTYPE_IPV6:
596 			if (bcmp(&le->source_ip6, &ip6->ip6_src,
597 			    sizeof(struct in6_addr)) != 0 ||
598 			    bcmp(&le->dest_ip6, &ip6->ip6_dst,
599 			    sizeof(struct in6_addr)) != 0)
600 				continue;
601 			break;
602 #endif
603 #ifdef INET
604 		case ETHERTYPE_IP:
605 			if (le->source_ip4 != ip4->ip_src.s_addr ||
606 			    le->dest_ip4 != ip4->ip_dst.s_addr)
607 				continue;
608 			break;
609 #endif
610 		}
611 
612 		/* Flush now if appending will result in overflow. */
613 		if (le->p_len > (65535 - tcp_data_len)) {
614 			SLIST_REMOVE(&lc->lro_active, le, lro_entry, next);
615 			tcp_lro_flush(lc, le);
616 			break;
617 		}
618 
619 		/* Try to append the new segment. */
620 		if (__predict_false(seq != le->next_seq ||
621 		    (tcp_data_len == 0 && le->ack_seq == th->th_ack))) {
622 			/* Out of order packet or duplicate ACK. */
623 			SLIST_REMOVE(&lc->lro_active, le, lro_entry, next);
624 			tcp_lro_flush(lc, le);
625 			return (TCP_LRO_CANNOT);
626 		}
627 
628 		if (l != 0) {
629 			uint32_t tsval = ntohl(*(ts_ptr + 1));
630 			/* Make sure timestamp values are increasing. */
631 			/* XXX-BZ flip and use TSTMP_GEQ macro for this? */
632 			if (__predict_false(le->tsval > tsval ||
633 			    *(ts_ptr + 2) == 0))
634 				return (TCP_LRO_CANNOT);
635 			le->tsval = tsval;
636 			le->tsecr = *(ts_ptr + 2);
637 		}
638 
639 		le->next_seq += tcp_data_len;
640 		le->ack_seq = th->th_ack;
641 		le->window = th->th_win;
642 		le->append_cnt++;
643 
644 #ifdef TCP_LRO_UPDATE_CSUM
645 		le->ulp_csum += tcp_lro_rx_csum_fixup(le, l3hdr, th,
646 		    tcp_data_len, ~csum);
647 #endif
648 
649 		if (tcp_data_len == 0) {
650 			m_freem(m);
651 			return (0);
652 		}
653 
654 		le->p_len += tcp_data_len;
655 
656 		/*
657 		 * Adjust the mbuf so that m_data points to the first byte of
658 		 * the ULP payload.  Adjust the mbuf to avoid complications and
659 		 * append new segment to existing mbuf chain.
660 		 */
661 		m_adj(m, m->m_pkthdr.len - tcp_data_len);
662 		m_demote_pkthdr(m);
663 
664 		le->m_tail->m_next = m;
665 		le->m_tail = m_last(m);
666 
667 		/*
668 		 * If a possible next full length packet would cause an
669 		 * overflow, pro-actively flush now.
670 		 */
671 		if (le->p_len > (65535 - lc->ifp->if_mtu)) {
672 			SLIST_REMOVE(&lc->lro_active, le, lro_entry, next);
673 			tcp_lro_flush(lc, le);
674 		} else
675 			getmicrotime(&le->mtime);
676 
677 		return (0);
678 	}
679 
680 	/* Try to find an empty slot. */
681 	if (SLIST_EMPTY(&lc->lro_free))
682 		return (TCP_LRO_CANNOT);
683 
684 	/* Start a new segment chain. */
685 	le = SLIST_FIRST(&lc->lro_free);
686 	SLIST_REMOVE_HEAD(&lc->lro_free, next);
687 	SLIST_INSERT_HEAD(&lc->lro_active, le, next);
688 	getmicrotime(&le->mtime);
689 
690 	/* Start filling in details. */
691 	switch (eh_type) {
692 #ifdef INET6
693 	case ETHERTYPE_IPV6:
694 		le->le_ip6 = ip6;
695 		le->source_ip6 = ip6->ip6_src;
696 		le->dest_ip6 = ip6->ip6_dst;
697 		le->eh_type = eh_type;
698 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN - sizeof(*ip6);
699 		break;
700 #endif
701 #ifdef INET
702 	case ETHERTYPE_IP:
703 		le->le_ip4 = ip4;
704 		le->source_ip4 = ip4->ip_src.s_addr;
705 		le->dest_ip4 = ip4->ip_dst.s_addr;
706 		le->eh_type = eh_type;
707 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN;
708 		break;
709 #endif
710 	}
711 	le->source_port = th->th_sport;
712 	le->dest_port = th->th_dport;
713 
714 	le->next_seq = seq + tcp_data_len;
715 	le->ack_seq = th->th_ack;
716 	le->window = th->th_win;
717 	if (l != 0) {
718 		le->timestamp = 1;
719 		le->tsval = ntohl(*(ts_ptr + 1));
720 		le->tsecr = *(ts_ptr + 2);
721 	}
722 
723 #ifdef TCP_LRO_UPDATE_CSUM
724 	/*
725 	 * Do not touch the csum of the first packet.  However save the
726 	 * "adjusted" checksum of just the source and destination addresses,
727 	 * the next header and the TCP payload.  The length and TCP header
728 	 * parts may change, so we remove those from the saved checksum and
729 	 * re-add with final values on tcp_lro_flush() if needed.
730 	 */
731 	KASSERT(le->ulp_csum == 0, ("%s: le=%p le->ulp_csum=0x%04x\n",
732 	    __func__, le, le->ulp_csum));
733 
734 	le->ulp_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th, tcp_data_len,
735 	    ~csum);
736 	th->th_sum = csum;	/* Restore checksum on first packet. */
737 #endif
738 
739 	le->m_head = m;
740 	le->m_tail = m_last(m);
741 
742 	return (0);
743 }
744 
745 void
746 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
747 {
748 	/* sanity checks */
749 	if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
750 	    lc->lro_mbuf_max == 0)) {
751 		/* packet drop */
752 		m_freem(mb);
753 		return;
754 	}
755 
756 	/* check if packet is not LRO capable */
757 	if (__predict_false(mb->m_pkthdr.csum_flags == 0 ||
758 	    (lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
759 		lc->lro_flushed++;
760 		lc->lro_queued++;
761 
762 		/* input packet to network layer */
763 		(*lc->ifp->if_input) (lc->ifp, mb);
764 		return;
765 	}
766 
767 	/* check if array is full */
768 	if (__predict_false(lc->lro_mbuf_count == lc->lro_mbuf_max))
769 		tcp_lro_flush_all(lc);
770 
771 	/* store sequence number */
772 	TCP_LRO_SEQUENCE(mb) = lc->lro_mbuf_count;
773 
774 	/* enter mbuf */
775 	lc->lro_mbuf_data[lc->lro_mbuf_count++] = mb;
776 }
777 
778 /* end */
779