xref: /freebsd/sys/netinet/tcp_lro.c (revision 8ef24a0d4b28fe230e20637f56869cc4148cd2ca)
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 static void	tcp_lro_rx_done(struct lro_ctrl *lc);
71 
72 static __inline void
73 tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_entry *le)
74 {
75 
76 	LIST_INSERT_HEAD(&lc->lro_active, le, next);
77 }
78 
79 static __inline void
80 tcp_lro_active_remove(struct lro_entry *le)
81 {
82 
83 	LIST_REMOVE(le, next);
84 }
85 
86 int
87 tcp_lro_init(struct lro_ctrl *lc)
88 {
89 	return (tcp_lro_init_args(lc, NULL, TCP_LRO_ENTRIES, 0));
90 }
91 
92 int
93 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
94     unsigned lro_entries, unsigned lro_mbufs)
95 {
96 	struct lro_entry *le;
97 	size_t size;
98 	unsigned i;
99 
100 	lc->lro_bad_csum = 0;
101 	lc->lro_queued = 0;
102 	lc->lro_flushed = 0;
103 	lc->lro_cnt = 0;
104 	lc->lro_mbuf_count = 0;
105 	lc->lro_mbuf_max = lro_mbufs;
106 	lc->lro_cnt = lro_entries;
107 	lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
108 	lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
109 	lc->ifp = ifp;
110 	LIST_INIT(&lc->lro_free);
111 	LIST_INIT(&lc->lro_active);
112 
113 	/* compute size to allocate */
114 	size = (lro_mbufs * sizeof(struct mbuf *)) +
115 	    (lro_entries * sizeof(*le));
116 	lc->lro_mbuf_data = (struct mbuf **)
117 	    malloc(size, M_LRO, M_NOWAIT | M_ZERO);
118 
119 	/* check for out of memory */
120 	if (lc->lro_mbuf_data == NULL) {
121 		memset(lc, 0, sizeof(*lc));
122 		return (ENOMEM);
123 	}
124 	/* compute offset for LRO entries */
125 	le = (struct lro_entry *)
126 	    (lc->lro_mbuf_data + lro_mbufs);
127 
128 	/* setup linked list */
129 	for (i = 0; i != lro_entries; i++)
130 		LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
131 
132 	return (0);
133 }
134 
135 void
136 tcp_lro_free(struct lro_ctrl *lc)
137 {
138 	struct lro_entry *le;
139 	unsigned x;
140 
141 	/* reset LRO free list */
142 	LIST_INIT(&lc->lro_free);
143 
144 	/* free active mbufs, if any */
145 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
146 		tcp_lro_active_remove(le);
147 		m_freem(le->m_head);
148 	}
149 
150 	/* free mbuf array, if any */
151 	for (x = 0; x != lc->lro_mbuf_count; x++)
152 		m_freem(lc->lro_mbuf_data[x]);
153 	lc->lro_mbuf_count = 0;
154 
155 	/* free allocated memory, if any */
156 	free(lc->lro_mbuf_data, M_LRO);
157 	lc->lro_mbuf_data = NULL;
158 }
159 
160 #ifdef TCP_LRO_UPDATE_CSUM
161 static uint16_t
162 tcp_lro_csum_th(struct tcphdr *th)
163 {
164 	uint32_t ch;
165 	uint16_t *p, l;
166 
167 	ch = th->th_sum = 0x0000;
168 	l = th->th_off;
169 	p = (uint16_t *)th;
170 	while (l > 0) {
171 		ch += *p;
172 		p++;
173 		ch += *p;
174 		p++;
175 		l--;
176 	}
177 	while (ch > 0xffff)
178 		ch = (ch >> 16) + (ch & 0xffff);
179 
180 	return (ch & 0xffff);
181 }
182 
183 static uint16_t
184 tcp_lro_rx_csum_fixup(struct lro_entry *le, void *l3hdr, struct tcphdr *th,
185     uint16_t tcp_data_len, uint16_t csum)
186 {
187 	uint32_t c;
188 	uint16_t cs;
189 
190 	c = csum;
191 
192 	/* Remove length from checksum. */
193 	switch (le->eh_type) {
194 #ifdef INET6
195 	case ETHERTYPE_IPV6:
196 	{
197 		struct ip6_hdr *ip6;
198 
199 		ip6 = (struct ip6_hdr *)l3hdr;
200 		if (le->append_cnt == 0)
201 			cs = ip6->ip6_plen;
202 		else {
203 			uint32_t cx;
204 
205 			cx = ntohs(ip6->ip6_plen);
206 			cs = in6_cksum_pseudo(ip6, cx, ip6->ip6_nxt, 0);
207 		}
208 		break;
209 	}
210 #endif
211 #ifdef INET
212 	case ETHERTYPE_IP:
213 	{
214 		struct ip *ip4;
215 
216 		ip4 = (struct ip *)l3hdr;
217 		if (le->append_cnt == 0)
218 			cs = ip4->ip_len;
219 		else {
220 			cs = in_addword(ntohs(ip4->ip_len) - sizeof(*ip4),
221 			    IPPROTO_TCP);
222 			cs = in_pseudo(ip4->ip_src.s_addr, ip4->ip_dst.s_addr,
223 			    htons(cs));
224 		}
225 		break;
226 	}
227 #endif
228 	default:
229 		cs = 0;		/* Keep compiler happy. */
230 	}
231 
232 	cs = ~cs;
233 	c += cs;
234 
235 	/* Remove TCP header csum. */
236 	cs = ~tcp_lro_csum_th(th);
237 	c += cs;
238 	while (c > 0xffff)
239 		c = (c >> 16) + (c & 0xffff);
240 
241 	return (c & 0xffff);
242 }
243 #endif
244 
245 static void
246 tcp_lro_rx_done(struct lro_ctrl *lc)
247 {
248 	struct lro_entry *le;
249 
250 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
251 		tcp_lro_active_remove(le);
252 		tcp_lro_flush(lc, le);
253 	}
254 }
255 
256 void
257 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
258 {
259 	struct lro_entry *le, *le_tmp;
260 	struct timeval tv;
261 
262 	if (LIST_EMPTY(&lc->lro_active))
263 		return;
264 
265 	getmicrotime(&tv);
266 	timevalsub(&tv, timeout);
267 	LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
268 		if (timevalcmp(&tv, &le->mtime, >=)) {
269 			tcp_lro_active_remove(le);
270 			tcp_lro_flush(lc, le);
271 		}
272 	}
273 }
274 
275 void
276 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
277 {
278 
279 	if (le->append_cnt > 0) {
280 		struct tcphdr *th;
281 		uint16_t p_len;
282 
283 		p_len = htons(le->p_len);
284 		switch (le->eh_type) {
285 #ifdef INET6
286 		case ETHERTYPE_IPV6:
287 		{
288 			struct ip6_hdr *ip6;
289 
290 			ip6 = le->le_ip6;
291 			ip6->ip6_plen = p_len;
292 			th = (struct tcphdr *)(ip6 + 1);
293 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
294 			    CSUM_PSEUDO_HDR;
295 			le->p_len += ETHER_HDR_LEN + sizeof(*ip6);
296 			break;
297 		}
298 #endif
299 #ifdef INET
300 		case ETHERTYPE_IP:
301 		{
302 			struct ip *ip4;
303 #ifdef TCP_LRO_UPDATE_CSUM
304 			uint32_t cl;
305 			uint16_t c;
306 #endif
307 
308 			ip4 = le->le_ip4;
309 #ifdef TCP_LRO_UPDATE_CSUM
310 			/* Fix IP header checksum for new length. */
311 			c = ~ip4->ip_sum;
312 			cl = c;
313 			c = ~ip4->ip_len;
314 			cl += c + p_len;
315 			while (cl > 0xffff)
316 				cl = (cl >> 16) + (cl & 0xffff);
317 			c = cl;
318 			ip4->ip_sum = ~c;
319 #else
320 			ip4->ip_sum = TCP_LRO_INVALID_CSUM;
321 #endif
322 			ip4->ip_len = p_len;
323 			th = (struct tcphdr *)(ip4 + 1);
324 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
325 			    CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
326 			le->p_len += ETHER_HDR_LEN;
327 			break;
328 		}
329 #endif
330 		default:
331 			th = NULL;	/* Keep compiler happy. */
332 		}
333 		le->m_head->m_pkthdr.csum_data = 0xffff;
334 		le->m_head->m_pkthdr.len = le->p_len;
335 
336 		/* Incorporate the latest ACK into the TCP header. */
337 		th->th_ack = le->ack_seq;
338 		th->th_win = le->window;
339 		/* Incorporate latest timestamp into the TCP header. */
340 		if (le->timestamp != 0) {
341 			uint32_t *ts_ptr;
342 
343 			ts_ptr = (uint32_t *)(th + 1);
344 			ts_ptr[1] = htonl(le->tsval);
345 			ts_ptr[2] = le->tsecr;
346 		}
347 #ifdef TCP_LRO_UPDATE_CSUM
348 		/* Update the TCP header checksum. */
349 		le->ulp_csum += p_len;
350 		le->ulp_csum += tcp_lro_csum_th(th);
351 		while (le->ulp_csum > 0xffff)
352 			le->ulp_csum = (le->ulp_csum >> 16) +
353 			    (le->ulp_csum & 0xffff);
354 		th->th_sum = (le->ulp_csum & 0xffff);
355 		th->th_sum = ~th->th_sum;
356 #else
357 		th->th_sum = TCP_LRO_INVALID_CSUM;
358 #endif
359 	}
360 
361 	(*lc->ifp->if_input)(lc->ifp, le->m_head);
362 	lc->lro_queued += le->append_cnt + 1;
363 	lc->lro_flushed++;
364 	bzero(le, sizeof(*le));
365 	LIST_INSERT_HEAD(&lc->lro_free, le, next);
366 }
367 
368 static int
369 tcp_lro_mbuf_compare_header(const void *ppa, const void *ppb)
370 {
371 	const struct mbuf *ma = *((const struct mbuf * const *)ppa);
372 	const struct mbuf *mb = *((const struct mbuf * const *)ppb);
373 	int ret;
374 
375 	ret = M_HASHTYPE_GET(ma) - M_HASHTYPE_GET(mb);
376 	if (ret != 0)
377 		goto done;
378 
379 	if (ma->m_pkthdr.flowid > mb->m_pkthdr.flowid)
380 		return (1);
381 	else if (ma->m_pkthdr.flowid < mb->m_pkthdr.flowid)
382 		return (-1);
383 
384 	ret = TCP_LRO_SEQUENCE(ma) - TCP_LRO_SEQUENCE(mb);
385 done:
386 	return (ret);
387 }
388 
389 void
390 tcp_lro_flush_all(struct lro_ctrl *lc)
391 {
392 	uint32_t hashtype;
393 	uint32_t flowid;
394 	unsigned x;
395 
396 	/* check if no mbufs to flush */
397 	if (lc->lro_mbuf_count == 0)
398 		goto done;
399 
400 	/* sort all mbufs according to stream */
401 	qsort(lc->lro_mbuf_data, lc->lro_mbuf_count, sizeof(struct mbuf *),
402 	    &tcp_lro_mbuf_compare_header);
403 
404 	/* input data into LRO engine, stream by stream */
405 	flowid = 0;
406 	hashtype = M_HASHTYPE_NONE;
407 	for (x = 0; x != lc->lro_mbuf_count; x++) {
408 		struct mbuf *mb;
409 
410 		mb = lc->lro_mbuf_data[x];
411 
412 		/* check for new stream */
413 		if (mb->m_pkthdr.flowid != flowid ||
414 		    M_HASHTYPE_GET(mb) != hashtype) {
415 			flowid = mb->m_pkthdr.flowid;
416 			hashtype = M_HASHTYPE_GET(mb);
417 
418 			/* flush active streams */
419 			tcp_lro_rx_done(lc);
420 		}
421 #ifdef TCP_LRO_RESET_SEQUENCE
422 		/* reset sequence number */
423 		TCP_LRO_SEQUENCE(mb) = 0;
424 #endif
425 		/* add packet to LRO engine */
426 		if (tcp_lro_rx(lc, mb, 0) != 0) {
427 			/* input packet to network layer */
428 			(*lc->ifp->if_input)(lc->ifp, mb);
429 			lc->lro_queued++;
430 			lc->lro_flushed++;
431 		}
432 	}
433 done:
434 	/* flush active streams */
435 	tcp_lro_rx_done(lc);
436 
437 	lc->lro_mbuf_count = 0;
438 }
439 
440 #ifdef INET6
441 static int
442 tcp_lro_rx_ipv6(struct lro_ctrl *lc, struct mbuf *m, struct ip6_hdr *ip6,
443     struct tcphdr **th)
444 {
445 
446 	/* XXX-BZ we should check the flow-label. */
447 
448 	/* XXX-BZ We do not yet support ext. hdrs. */
449 	if (ip6->ip6_nxt != IPPROTO_TCP)
450 		return (TCP_LRO_NOT_SUPPORTED);
451 
452 	/* Find the TCP header. */
453 	*th = (struct tcphdr *)(ip6 + 1);
454 
455 	return (0);
456 }
457 #endif
458 
459 #ifdef INET
460 static int
461 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4,
462     struct tcphdr **th)
463 {
464 	int csum_flags;
465 	uint16_t csum;
466 
467 	if (ip4->ip_p != IPPROTO_TCP)
468 		return (TCP_LRO_NOT_SUPPORTED);
469 
470 	/* Ensure there are no options. */
471 	if ((ip4->ip_hl << 2) != sizeof (*ip4))
472 		return (TCP_LRO_CANNOT);
473 
474 	/* .. and the packet is not fragmented. */
475 	if (ip4->ip_off & htons(IP_MF|IP_OFFMASK))
476 		return (TCP_LRO_CANNOT);
477 
478 	/* Legacy IP has a header checksum that needs to be correct. */
479 	csum_flags = m->m_pkthdr.csum_flags;
480 	if (csum_flags & CSUM_IP_CHECKED) {
481 		if (__predict_false((csum_flags & CSUM_IP_VALID) == 0)) {
482 			lc->lro_bad_csum++;
483 			return (TCP_LRO_CANNOT);
484 		}
485 	} else {
486 		csum = in_cksum_hdr(ip4);
487 		if (__predict_false((csum) != 0)) {
488 			lc->lro_bad_csum++;
489 			return (TCP_LRO_CANNOT);
490 		}
491 	}
492 
493 	/* Find the TCP header (we assured there are no IP options). */
494 	*th = (struct tcphdr *)(ip4 + 1);
495 
496 	return (0);
497 }
498 #endif
499 
500 int
501 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
502 {
503 	struct lro_entry *le;
504 	struct ether_header *eh;
505 #ifdef INET6
506 	struct ip6_hdr *ip6 = NULL;	/* Keep compiler happy. */
507 #endif
508 #ifdef INET
509 	struct ip *ip4 = NULL;		/* Keep compiler happy. */
510 #endif
511 	struct tcphdr *th;
512 	void *l3hdr = NULL;		/* Keep compiler happy. */
513 	uint32_t *ts_ptr;
514 	tcp_seq seq;
515 	int error, ip_len, l;
516 	uint16_t eh_type, tcp_data_len;
517 
518 	/* We expect a contiguous header [eh, ip, tcp]. */
519 
520 	eh = mtod(m, struct ether_header *);
521 	eh_type = ntohs(eh->ether_type);
522 	switch (eh_type) {
523 #ifdef INET6
524 	case ETHERTYPE_IPV6:
525 	{
526 		CURVNET_SET(lc->ifp->if_vnet);
527 		if (V_ip6_forwarding != 0) {
528 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
529 			CURVNET_RESTORE();
530 			return (TCP_LRO_CANNOT);
531 		}
532 		CURVNET_RESTORE();
533 		l3hdr = ip6 = (struct ip6_hdr *)(eh + 1);
534 		error = tcp_lro_rx_ipv6(lc, m, ip6, &th);
535 		if (error != 0)
536 			return (error);
537 		tcp_data_len = ntohs(ip6->ip6_plen);
538 		ip_len = sizeof(*ip6) + tcp_data_len;
539 		break;
540 	}
541 #endif
542 #ifdef INET
543 	case ETHERTYPE_IP:
544 	{
545 		CURVNET_SET(lc->ifp->if_vnet);
546 		if (V_ipforwarding != 0) {
547 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
548 			CURVNET_RESTORE();
549 			return (TCP_LRO_CANNOT);
550 		}
551 		CURVNET_RESTORE();
552 		l3hdr = ip4 = (struct ip *)(eh + 1);
553 		error = tcp_lro_rx_ipv4(lc, m, ip4, &th);
554 		if (error != 0)
555 			return (error);
556 		ip_len = ntohs(ip4->ip_len);
557 		tcp_data_len = ip_len - sizeof(*ip4);
558 		break;
559 	}
560 #endif
561 	/* XXX-BZ what happens in case of VLAN(s)? */
562 	default:
563 		return (TCP_LRO_NOT_SUPPORTED);
564 	}
565 
566 	/*
567 	 * If the frame is padded beyond the end of the IP packet, then we must
568 	 * trim the extra bytes off.
569 	 */
570 	l = m->m_pkthdr.len - (ETHER_HDR_LEN + ip_len);
571 	if (l != 0) {
572 		if (l < 0)
573 			/* Truncated packet. */
574 			return (TCP_LRO_CANNOT);
575 
576 		m_adj(m, -l);
577 	}
578 
579 	/*
580 	 * Check TCP header constraints.
581 	 */
582 	/* Ensure no bits set besides ACK or PSH. */
583 	if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0)
584 		return (TCP_LRO_CANNOT);
585 
586 	/* XXX-BZ We lose a ACK|PUSH flag concatenating multiple segments. */
587 	/* XXX-BZ Ideally we'd flush on PUSH? */
588 
589 	/*
590 	 * Check for timestamps.
591 	 * Since the only option we handle are timestamps, we only have to
592 	 * handle the simple case of aligned timestamps.
593 	 */
594 	l = (th->th_off << 2);
595 	tcp_data_len -= l;
596 	l -= sizeof(*th);
597 	ts_ptr = (uint32_t *)(th + 1);
598 	if (l != 0 && (__predict_false(l != TCPOLEN_TSTAMP_APPA) ||
599 	    (*ts_ptr != ntohl(TCPOPT_NOP<<24|TCPOPT_NOP<<16|
600 	    TCPOPT_TIMESTAMP<<8|TCPOLEN_TIMESTAMP))))
601 		return (TCP_LRO_CANNOT);
602 
603 	/* If the driver did not pass in the checksum, set it now. */
604 	if (csum == 0x0000)
605 		csum = th->th_sum;
606 
607 	seq = ntohl(th->th_seq);
608 
609 	/* Try to find a matching previous segment. */
610 	LIST_FOREACH(le, &lc->lro_active, next) {
611 		if (le->eh_type != eh_type)
612 			continue;
613 		if (le->source_port != th->th_sport ||
614 		    le->dest_port != th->th_dport)
615 			continue;
616 		switch (eh_type) {
617 #ifdef INET6
618 		case ETHERTYPE_IPV6:
619 			if (bcmp(&le->source_ip6, &ip6->ip6_src,
620 			    sizeof(struct in6_addr)) != 0 ||
621 			    bcmp(&le->dest_ip6, &ip6->ip6_dst,
622 			    sizeof(struct in6_addr)) != 0)
623 				continue;
624 			break;
625 #endif
626 #ifdef INET
627 		case ETHERTYPE_IP:
628 			if (le->source_ip4 != ip4->ip_src.s_addr ||
629 			    le->dest_ip4 != ip4->ip_dst.s_addr)
630 				continue;
631 			break;
632 #endif
633 		}
634 
635 		/* Flush now if appending will result in overflow. */
636 		if (le->p_len > (lc->lro_length_lim - tcp_data_len)) {
637 			tcp_lro_active_remove(le);
638 			tcp_lro_flush(lc, le);
639 			break;
640 		}
641 
642 		/* Try to append the new segment. */
643 		if (__predict_false(seq != le->next_seq ||
644 		    (tcp_data_len == 0 && le->ack_seq == th->th_ack))) {
645 			/* Out of order packet or duplicate ACK. */
646 			tcp_lro_active_remove(le);
647 			tcp_lro_flush(lc, le);
648 			return (TCP_LRO_CANNOT);
649 		}
650 
651 		if (l != 0) {
652 			uint32_t tsval = ntohl(*(ts_ptr + 1));
653 			/* Make sure timestamp values are increasing. */
654 			/* XXX-BZ flip and use TSTMP_GEQ macro for this? */
655 			if (__predict_false(le->tsval > tsval ||
656 			    *(ts_ptr + 2) == 0))
657 				return (TCP_LRO_CANNOT);
658 			le->tsval = tsval;
659 			le->tsecr = *(ts_ptr + 2);
660 		}
661 
662 		le->next_seq += tcp_data_len;
663 		le->ack_seq = th->th_ack;
664 		le->window = th->th_win;
665 		le->append_cnt++;
666 
667 #ifdef TCP_LRO_UPDATE_CSUM
668 		le->ulp_csum += tcp_lro_rx_csum_fixup(le, l3hdr, th,
669 		    tcp_data_len, ~csum);
670 #endif
671 
672 		if (tcp_data_len == 0) {
673 			m_freem(m);
674 			/*
675 			 * Flush this LRO entry, if this ACK should not
676 			 * be further delayed.
677 			 */
678 			if (le->append_cnt >= lc->lro_ackcnt_lim) {
679 				tcp_lro_active_remove(le);
680 				tcp_lro_flush(lc, le);
681 			}
682 			return (0);
683 		}
684 
685 		le->p_len += tcp_data_len;
686 
687 		/*
688 		 * Adjust the mbuf so that m_data points to the first byte of
689 		 * the ULP payload.  Adjust the mbuf to avoid complications and
690 		 * append new segment to existing mbuf chain.
691 		 */
692 		m_adj(m, m->m_pkthdr.len - tcp_data_len);
693 		m_demote_pkthdr(m);
694 
695 		le->m_tail->m_next = m;
696 		le->m_tail = m_last(m);
697 
698 		/*
699 		 * If a possible next full length packet would cause an
700 		 * overflow, pro-actively flush now.
701 		 */
702 		if (le->p_len > (lc->lro_length_lim - lc->ifp->if_mtu)) {
703 			tcp_lro_active_remove(le);
704 			tcp_lro_flush(lc, le);
705 		} else
706 			getmicrotime(&le->mtime);
707 
708 		return (0);
709 	}
710 
711 	/* Try to find an empty slot. */
712 	if (LIST_EMPTY(&lc->lro_free))
713 		return (TCP_LRO_NO_ENTRIES);
714 
715 	/* Start a new segment chain. */
716 	le = LIST_FIRST(&lc->lro_free);
717 	LIST_REMOVE(le, next);
718 	tcp_lro_active_insert(lc, le);
719 	getmicrotime(&le->mtime);
720 
721 	/* Start filling in details. */
722 	switch (eh_type) {
723 #ifdef INET6
724 	case ETHERTYPE_IPV6:
725 		le->le_ip6 = ip6;
726 		le->source_ip6 = ip6->ip6_src;
727 		le->dest_ip6 = ip6->ip6_dst;
728 		le->eh_type = eh_type;
729 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN - sizeof(*ip6);
730 		break;
731 #endif
732 #ifdef INET
733 	case ETHERTYPE_IP:
734 		le->le_ip4 = ip4;
735 		le->source_ip4 = ip4->ip_src.s_addr;
736 		le->dest_ip4 = ip4->ip_dst.s_addr;
737 		le->eh_type = eh_type;
738 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN;
739 		break;
740 #endif
741 	}
742 	le->source_port = th->th_sport;
743 	le->dest_port = th->th_dport;
744 
745 	le->next_seq = seq + tcp_data_len;
746 	le->ack_seq = th->th_ack;
747 	le->window = th->th_win;
748 	if (l != 0) {
749 		le->timestamp = 1;
750 		le->tsval = ntohl(*(ts_ptr + 1));
751 		le->tsecr = *(ts_ptr + 2);
752 	}
753 
754 #ifdef TCP_LRO_UPDATE_CSUM
755 	/*
756 	 * Do not touch the csum of the first packet.  However save the
757 	 * "adjusted" checksum of just the source and destination addresses,
758 	 * the next header and the TCP payload.  The length and TCP header
759 	 * parts may change, so we remove those from the saved checksum and
760 	 * re-add with final values on tcp_lro_flush() if needed.
761 	 */
762 	KASSERT(le->ulp_csum == 0, ("%s: le=%p le->ulp_csum=0x%04x\n",
763 	    __func__, le, le->ulp_csum));
764 
765 	le->ulp_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th, tcp_data_len,
766 	    ~csum);
767 	th->th_sum = csum;	/* Restore checksum on first packet. */
768 #endif
769 
770 	le->m_head = m;
771 	le->m_tail = m_last(m);
772 
773 	return (0);
774 }
775 
776 void
777 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
778 {
779 	/* sanity checks */
780 	if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
781 	    lc->lro_mbuf_max == 0)) {
782 		/* packet drop */
783 		m_freem(mb);
784 		return;
785 	}
786 
787 	/* check if packet is not LRO capable */
788 	if (__predict_false(mb->m_pkthdr.csum_flags == 0 ||
789 	    (lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
790 		lc->lro_flushed++;
791 		lc->lro_queued++;
792 
793 		/* input packet to network layer */
794 		(*lc->ifp->if_input) (lc->ifp, mb);
795 		return;
796 	}
797 
798 	/* check if array is full */
799 	if (__predict_false(lc->lro_mbuf_count == lc->lro_mbuf_max))
800 		tcp_lro_flush_all(lc);
801 
802 	/* store sequence number */
803 	TCP_LRO_SEQUENCE(mb) = lc->lro_mbuf_count;
804 
805 	/* enter mbuf */
806 	lc->lro_mbuf_data[lc->lro_mbuf_count++] = mb;
807 }
808 
809 /* end */
810