xref: /freebsd/sys/netinet/tcp_lro.c (revision 7447ca0eb235974642312b9555caec00b57d8fc1)
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 #include <sys/sysctl.h>
46 
47 #include <net/if.h>
48 #include <net/if_var.h>
49 #include <net/ethernet.h>
50 #include <net/vnet.h>
51 
52 #include <netinet/in_systm.h>
53 #include <netinet/in.h>
54 #include <netinet/ip6.h>
55 #include <netinet/ip.h>
56 #include <netinet/ip_var.h>
57 #include <netinet/tcp.h>
58 #include <netinet/tcp_lro.h>
59 #include <netinet/tcp_var.h>
60 
61 #include <netinet6/ip6_var.h>
62 
63 #include <machine/in_cksum.h>
64 
65 static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures");
66 
67 #define	TCP_LRO_UPDATE_CSUM	1
68 #ifndef	TCP_LRO_UPDATE_CSUM
69 #define	TCP_LRO_INVALID_CSUM	0x0000
70 #endif
71 
72 static void	tcp_lro_rx_done(struct lro_ctrl *lc);
73 static int	tcp_lro_rx2(struct lro_ctrl *lc, struct mbuf *m,
74 		    uint32_t csum, int use_hash);
75 
76 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro,  CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
77     "TCP LRO");
78 
79 static unsigned	tcp_lro_entries = TCP_LRO_ENTRIES;
80 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries,
81     CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_entries, 0,
82     "default number of LRO entries");
83 
84 static __inline void
85 tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket,
86     struct lro_entry *le)
87 {
88 
89 	LIST_INSERT_HEAD(&lc->lro_active, le, next);
90 	LIST_INSERT_HEAD(bucket, le, hash_next);
91 }
92 
93 static __inline void
94 tcp_lro_active_remove(struct lro_entry *le)
95 {
96 
97 	LIST_REMOVE(le, next);		/* active list */
98 	LIST_REMOVE(le, hash_next);	/* hash bucket */
99 }
100 
101 int
102 tcp_lro_init(struct lro_ctrl *lc)
103 {
104 	return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0));
105 }
106 
107 int
108 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
109     unsigned lro_entries, unsigned lro_mbufs)
110 {
111 	struct lro_entry *le;
112 	size_t size;
113 	unsigned i, elements;
114 
115 	lc->lro_bad_csum = 0;
116 	lc->lro_queued = 0;
117 	lc->lro_flushed = 0;
118 	lc->lro_cnt = 0;
119 	lc->lro_mbuf_count = 0;
120 	lc->lro_mbuf_max = lro_mbufs;
121 	lc->lro_cnt = lro_entries;
122 	lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
123 	lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
124 	lc->ifp = ifp;
125 	LIST_INIT(&lc->lro_free);
126 	LIST_INIT(&lc->lro_active);
127 
128 	/* create hash table to accelerate entry lookup */
129 	if (lro_entries > lro_mbufs)
130 		elements = lro_entries;
131 	else
132 		elements = lro_mbufs;
133 	lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz,
134 	    HASH_NOWAIT);
135 	if (lc->lro_hash == NULL) {
136 		memset(lc, 0, sizeof(*lc));
137 		return (ENOMEM);
138 	}
139 
140 	/* compute size to allocate */
141 	size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
142 	    (lro_entries * sizeof(*le));
143 	lc->lro_mbuf_data = (struct lro_mbuf_sort *)
144 	    malloc(size, M_LRO, M_NOWAIT | M_ZERO);
145 
146 	/* check for out of memory */
147 	if (lc->lro_mbuf_data == NULL) {
148 		memset(lc, 0, sizeof(*lc));
149 		return (ENOMEM);
150 	}
151 	/* compute offset for LRO entries */
152 	le = (struct lro_entry *)
153 	    (lc->lro_mbuf_data + lro_mbufs);
154 
155 	/* setup linked list */
156 	for (i = 0; i != lro_entries; i++)
157 		LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
158 
159 	return (0);
160 }
161 
162 void
163 tcp_lro_free(struct lro_ctrl *lc)
164 {
165 	struct lro_entry *le;
166 	unsigned x;
167 
168 	/* reset LRO free list */
169 	LIST_INIT(&lc->lro_free);
170 
171 	/* free active mbufs, if any */
172 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
173 		tcp_lro_active_remove(le);
174 		m_freem(le->m_head);
175 	}
176 
177 	/* free hash table */
178 	if (lc->lro_hash != NULL) {
179 		free(lc->lro_hash, M_LRO);
180 		lc->lro_hash = NULL;
181 	}
182 	lc->lro_hashsz = 0;
183 
184 	/* free mbuf array, if any */
185 	for (x = 0; x != lc->lro_mbuf_count; x++)
186 		m_freem(lc->lro_mbuf_data[x].mb);
187 	lc->lro_mbuf_count = 0;
188 
189 	/* free allocated memory, if any */
190 	free(lc->lro_mbuf_data, M_LRO);
191 	lc->lro_mbuf_data = NULL;
192 }
193 
194 #ifdef TCP_LRO_UPDATE_CSUM
195 static uint16_t
196 tcp_lro_csum_th(struct tcphdr *th)
197 {
198 	uint32_t ch;
199 	uint16_t *p, l;
200 
201 	ch = th->th_sum = 0x0000;
202 	l = th->th_off;
203 	p = (uint16_t *)th;
204 	while (l > 0) {
205 		ch += *p;
206 		p++;
207 		ch += *p;
208 		p++;
209 		l--;
210 	}
211 	while (ch > 0xffff)
212 		ch = (ch >> 16) + (ch & 0xffff);
213 
214 	return (ch & 0xffff);
215 }
216 
217 static uint16_t
218 tcp_lro_rx_csum_fixup(struct lro_entry *le, void *l3hdr, struct tcphdr *th,
219     uint16_t tcp_data_len, uint16_t csum)
220 {
221 	uint32_t c;
222 	uint16_t cs;
223 
224 	c = csum;
225 
226 	/* Remove length from checksum. */
227 	switch (le->eh_type) {
228 #ifdef INET6
229 	case ETHERTYPE_IPV6:
230 	{
231 		struct ip6_hdr *ip6;
232 
233 		ip6 = (struct ip6_hdr *)l3hdr;
234 		if (le->append_cnt == 0)
235 			cs = ip6->ip6_plen;
236 		else {
237 			uint32_t cx;
238 
239 			cx = ntohs(ip6->ip6_plen);
240 			cs = in6_cksum_pseudo(ip6, cx, ip6->ip6_nxt, 0);
241 		}
242 		break;
243 	}
244 #endif
245 #ifdef INET
246 	case ETHERTYPE_IP:
247 	{
248 		struct ip *ip4;
249 
250 		ip4 = (struct ip *)l3hdr;
251 		if (le->append_cnt == 0)
252 			cs = ip4->ip_len;
253 		else {
254 			cs = in_addword(ntohs(ip4->ip_len) - sizeof(*ip4),
255 			    IPPROTO_TCP);
256 			cs = in_pseudo(ip4->ip_src.s_addr, ip4->ip_dst.s_addr,
257 			    htons(cs));
258 		}
259 		break;
260 	}
261 #endif
262 	default:
263 		cs = 0;		/* Keep compiler happy. */
264 	}
265 
266 	cs = ~cs;
267 	c += cs;
268 
269 	/* Remove TCP header csum. */
270 	cs = ~tcp_lro_csum_th(th);
271 	c += cs;
272 	while (c > 0xffff)
273 		c = (c >> 16) + (c & 0xffff);
274 
275 	return (c & 0xffff);
276 }
277 #endif
278 
279 static void
280 tcp_lro_rx_done(struct lro_ctrl *lc)
281 {
282 	struct lro_entry *le;
283 
284 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
285 		tcp_lro_active_remove(le);
286 		tcp_lro_flush(lc, le);
287 	}
288 }
289 
290 void
291 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
292 {
293 	struct lro_entry *le, *le_tmp;
294 	struct timeval tv;
295 
296 	if (LIST_EMPTY(&lc->lro_active))
297 		return;
298 
299 	getmicrotime(&tv);
300 	timevalsub(&tv, timeout);
301 	LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
302 		if (timevalcmp(&tv, &le->mtime, >=)) {
303 			tcp_lro_active_remove(le);
304 			tcp_lro_flush(lc, le);
305 		}
306 	}
307 }
308 
309 void
310 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
311 {
312 
313 	if (le->append_cnt > 0) {
314 		struct tcphdr *th;
315 		uint16_t p_len;
316 
317 		p_len = htons(le->p_len);
318 		switch (le->eh_type) {
319 #ifdef INET6
320 		case ETHERTYPE_IPV6:
321 		{
322 			struct ip6_hdr *ip6;
323 
324 			ip6 = le->le_ip6;
325 			ip6->ip6_plen = p_len;
326 			th = (struct tcphdr *)(ip6 + 1);
327 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
328 			    CSUM_PSEUDO_HDR;
329 			le->p_len += ETHER_HDR_LEN + sizeof(*ip6);
330 			break;
331 		}
332 #endif
333 #ifdef INET
334 		case ETHERTYPE_IP:
335 		{
336 			struct ip *ip4;
337 #ifdef TCP_LRO_UPDATE_CSUM
338 			uint32_t cl;
339 			uint16_t c;
340 #endif
341 
342 			ip4 = le->le_ip4;
343 #ifdef TCP_LRO_UPDATE_CSUM
344 			/* Fix IP header checksum for new length. */
345 			c = ~ip4->ip_sum;
346 			cl = c;
347 			c = ~ip4->ip_len;
348 			cl += c + p_len;
349 			while (cl > 0xffff)
350 				cl = (cl >> 16) + (cl & 0xffff);
351 			c = cl;
352 			ip4->ip_sum = ~c;
353 #else
354 			ip4->ip_sum = TCP_LRO_INVALID_CSUM;
355 #endif
356 			ip4->ip_len = p_len;
357 			th = (struct tcphdr *)(ip4 + 1);
358 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
359 			    CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
360 			le->p_len += ETHER_HDR_LEN;
361 			break;
362 		}
363 #endif
364 		default:
365 			th = NULL;	/* Keep compiler happy. */
366 		}
367 		le->m_head->m_pkthdr.csum_data = 0xffff;
368 		le->m_head->m_pkthdr.len = le->p_len;
369 
370 		/* Incorporate the latest ACK into the TCP header. */
371 		th->th_ack = le->ack_seq;
372 		th->th_win = le->window;
373 		/* Incorporate latest timestamp into the TCP header. */
374 		if (le->timestamp != 0) {
375 			uint32_t *ts_ptr;
376 
377 			ts_ptr = (uint32_t *)(th + 1);
378 			ts_ptr[1] = htonl(le->tsval);
379 			ts_ptr[2] = le->tsecr;
380 		}
381 #ifdef TCP_LRO_UPDATE_CSUM
382 		/* Update the TCP header checksum. */
383 		le->ulp_csum += p_len;
384 		le->ulp_csum += tcp_lro_csum_th(th);
385 		while (le->ulp_csum > 0xffff)
386 			le->ulp_csum = (le->ulp_csum >> 16) +
387 			    (le->ulp_csum & 0xffff);
388 		th->th_sum = (le->ulp_csum & 0xffff);
389 		th->th_sum = ~th->th_sum;
390 #else
391 		th->th_sum = TCP_LRO_INVALID_CSUM;
392 #endif
393 	}
394 
395 	le->m_head->m_pkthdr.lro_nsegs = le->append_cnt + 1;
396 	(*lc->ifp->if_input)(lc->ifp, le->m_head);
397 	lc->lro_queued += le->append_cnt + 1;
398 	lc->lro_flushed++;
399 	bzero(le, sizeof(*le));
400 	LIST_INSERT_HEAD(&lc->lro_free, le, next);
401 }
402 
403 #ifdef HAVE_INLINE_FLSLL
404 #define	tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
405 #else
406 static inline uint64_t
407 tcp_lro_msb_64(uint64_t x)
408 {
409 	x |= (x >> 1);
410 	x |= (x >> 2);
411 	x |= (x >> 4);
412 	x |= (x >> 8);
413 	x |= (x >> 16);
414 	x |= (x >> 32);
415 	return (x & ~(x >> 1));
416 }
417 #endif
418 
419 /*
420  * The tcp_lro_sort() routine is comparable to qsort(), except it has
421  * a worst case complexity limit of O(MIN(N,64)*N), where N is the
422  * number of elements to sort and 64 is the number of sequence bits
423  * available. The algorithm is bit-slicing the 64-bit sequence number,
424  * sorting one bit at a time from the most significant bit until the
425  * least significant one, skipping the constant bits. This is
426  * typically called a radix sort.
427  */
428 static void
429 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
430 {
431 	struct lro_mbuf_sort temp;
432 	uint64_t ones;
433 	uint64_t zeros;
434 	uint32_t x;
435 	uint32_t y;
436 
437 repeat:
438 	/* for small arrays insertion sort is faster */
439 	if (size <= 12) {
440 		for (x = 1; x < size; x++) {
441 			temp = parray[x];
442 			for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
443 				parray[y] = parray[y - 1];
444 			parray[y] = temp;
445 		}
446 		return;
447 	}
448 
449 	/* compute sequence bits which are constant */
450 	ones = 0;
451 	zeros = 0;
452 	for (x = 0; x != size; x++) {
453 		ones |= parray[x].seq;
454 		zeros |= ~parray[x].seq;
455 	}
456 
457 	/* compute bits which are not constant into "ones" */
458 	ones &= zeros;
459 	if (ones == 0)
460 		return;
461 
462 	/* pick the most significant bit which is not constant */
463 	ones = tcp_lro_msb_64(ones);
464 
465 	/*
466 	 * Move entries having cleared sequence bits to the beginning
467 	 * of the array:
468 	 */
469 	for (x = y = 0; y != size; y++) {
470 		/* skip set bits */
471 		if (parray[y].seq & ones)
472 			continue;
473 		/* swap entries */
474 		temp = parray[x];
475 		parray[x] = parray[y];
476 		parray[y] = temp;
477 		x++;
478 	}
479 
480 	KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
481 
482 	/* sort zeros */
483 	tcp_lro_sort(parray, x);
484 
485 	/* sort ones */
486 	parray += x;
487 	size -= x;
488 	goto repeat;
489 }
490 
491 void
492 tcp_lro_flush_all(struct lro_ctrl *lc)
493 {
494 	uint64_t seq;
495 	uint64_t nseq;
496 	unsigned x;
497 
498 	/* check if no mbufs to flush */
499 	if (lc->lro_mbuf_count == 0)
500 		goto done;
501 
502 	/* sort all mbufs according to stream */
503 	tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
504 
505 	/* input data into LRO engine, stream by stream */
506 	seq = 0;
507 	for (x = 0; x != lc->lro_mbuf_count; x++) {
508 		struct mbuf *mb;
509 
510 		/* get mbuf */
511 		mb = lc->lro_mbuf_data[x].mb;
512 
513 		/* get sequence number, masking away the packet index */
514 		nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
515 
516 		/* check for new stream */
517 		if (seq != nseq) {
518 			seq = nseq;
519 
520 			/* flush active streams */
521 			tcp_lro_rx_done(lc);
522 		}
523 
524 		/* add packet to LRO engine */
525 		if (tcp_lro_rx2(lc, mb, 0, 0) != 0) {
526 			/* input packet to network layer */
527 			(*lc->ifp->if_input)(lc->ifp, mb);
528 			lc->lro_queued++;
529 			lc->lro_flushed++;
530 		}
531 	}
532 done:
533 	/* flush active streams */
534 	tcp_lro_rx_done(lc);
535 
536 	lc->lro_mbuf_count = 0;
537 }
538 
539 #ifdef INET6
540 static int
541 tcp_lro_rx_ipv6(struct lro_ctrl *lc, struct mbuf *m, struct ip6_hdr *ip6,
542     struct tcphdr **th)
543 {
544 
545 	/* XXX-BZ we should check the flow-label. */
546 
547 	/* XXX-BZ We do not yet support ext. hdrs. */
548 	if (ip6->ip6_nxt != IPPROTO_TCP)
549 		return (TCP_LRO_NOT_SUPPORTED);
550 
551 	/* Find the TCP header. */
552 	*th = (struct tcphdr *)(ip6 + 1);
553 
554 	return (0);
555 }
556 #endif
557 
558 #ifdef INET
559 static int
560 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4,
561     struct tcphdr **th)
562 {
563 	int csum_flags;
564 	uint16_t csum;
565 
566 	if (ip4->ip_p != IPPROTO_TCP)
567 		return (TCP_LRO_NOT_SUPPORTED);
568 
569 	/* Ensure there are no options. */
570 	if ((ip4->ip_hl << 2) != sizeof (*ip4))
571 		return (TCP_LRO_CANNOT);
572 
573 	/* .. and the packet is not fragmented. */
574 	if (ip4->ip_off & htons(IP_MF|IP_OFFMASK))
575 		return (TCP_LRO_CANNOT);
576 
577 	/* Legacy IP has a header checksum that needs to be correct. */
578 	csum_flags = m->m_pkthdr.csum_flags;
579 	if (csum_flags & CSUM_IP_CHECKED) {
580 		if (__predict_false((csum_flags & CSUM_IP_VALID) == 0)) {
581 			lc->lro_bad_csum++;
582 			return (TCP_LRO_CANNOT);
583 		}
584 	} else {
585 		csum = in_cksum_hdr(ip4);
586 		if (__predict_false((csum) != 0)) {
587 			lc->lro_bad_csum++;
588 			return (TCP_LRO_CANNOT);
589 		}
590 	}
591 
592 	/* Find the TCP header (we assured there are no IP options). */
593 	*th = (struct tcphdr *)(ip4 + 1);
594 
595 	return (0);
596 }
597 #endif
598 
599 static int
600 tcp_lro_rx2(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, int use_hash)
601 {
602 	struct lro_entry *le;
603 	struct ether_header *eh;
604 #ifdef INET6
605 	struct ip6_hdr *ip6 = NULL;	/* Keep compiler happy. */
606 #endif
607 #ifdef INET
608 	struct ip *ip4 = NULL;		/* Keep compiler happy. */
609 #endif
610 	struct tcphdr *th;
611 	void *l3hdr = NULL;		/* Keep compiler happy. */
612 	uint32_t *ts_ptr;
613 	tcp_seq seq;
614 	int error, ip_len, l;
615 	uint16_t eh_type, tcp_data_len;
616 	struct lro_head *bucket;
617 	int force_flush = 0;
618 
619 	/* We expect a contiguous header [eh, ip, tcp]. */
620 
621 	eh = mtod(m, struct ether_header *);
622 	eh_type = ntohs(eh->ether_type);
623 	switch (eh_type) {
624 #ifdef INET6
625 	case ETHERTYPE_IPV6:
626 	{
627 		CURVNET_SET(lc->ifp->if_vnet);
628 		if (V_ip6_forwarding != 0) {
629 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
630 			CURVNET_RESTORE();
631 			return (TCP_LRO_CANNOT);
632 		}
633 		CURVNET_RESTORE();
634 		l3hdr = ip6 = (struct ip6_hdr *)(eh + 1);
635 		error = tcp_lro_rx_ipv6(lc, m, ip6, &th);
636 		if (error != 0)
637 			return (error);
638 		tcp_data_len = ntohs(ip6->ip6_plen);
639 		ip_len = sizeof(*ip6) + tcp_data_len;
640 		break;
641 	}
642 #endif
643 #ifdef INET
644 	case ETHERTYPE_IP:
645 	{
646 		CURVNET_SET(lc->ifp->if_vnet);
647 		if (V_ipforwarding != 0) {
648 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
649 			CURVNET_RESTORE();
650 			return (TCP_LRO_CANNOT);
651 		}
652 		CURVNET_RESTORE();
653 		l3hdr = ip4 = (struct ip *)(eh + 1);
654 		error = tcp_lro_rx_ipv4(lc, m, ip4, &th);
655 		if (error != 0)
656 			return (error);
657 		ip_len = ntohs(ip4->ip_len);
658 		tcp_data_len = ip_len - sizeof(*ip4);
659 		break;
660 	}
661 #endif
662 	/* XXX-BZ what happens in case of VLAN(s)? */
663 	default:
664 		return (TCP_LRO_NOT_SUPPORTED);
665 	}
666 
667 	/*
668 	 * If the frame is padded beyond the end of the IP packet, then we must
669 	 * trim the extra bytes off.
670 	 */
671 	l = m->m_pkthdr.len - (ETHER_HDR_LEN + ip_len);
672 	if (l != 0) {
673 		if (l < 0)
674 			/* Truncated packet. */
675 			return (TCP_LRO_CANNOT);
676 
677 		m_adj(m, -l);
678 	}
679 
680 	/*
681 	 * Check TCP header constraints.
682 	 */
683 	/* Ensure no bits set besides ACK or PSH. */
684 	if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0) {
685 		if (th->th_flags & TH_SYN)
686 			return (TCP_LRO_CANNOT);
687 		/*
688 		 * Make sure that previously seen segements/ACKs are delivered
689 		 * before this segement, e.g. FIN.
690 		 */
691 		force_flush = 1;
692 	}
693 
694 	/* XXX-BZ We lose a ACK|PUSH flag concatenating multiple segments. */
695 	/* XXX-BZ Ideally we'd flush on PUSH? */
696 
697 	/*
698 	 * Check for timestamps.
699 	 * Since the only option we handle are timestamps, we only have to
700 	 * handle the simple case of aligned timestamps.
701 	 */
702 	l = (th->th_off << 2);
703 	tcp_data_len -= l;
704 	l -= sizeof(*th);
705 	ts_ptr = (uint32_t *)(th + 1);
706 	if (l != 0 && (__predict_false(l != TCPOLEN_TSTAMP_APPA) ||
707 	    (*ts_ptr != ntohl(TCPOPT_NOP<<24|TCPOPT_NOP<<16|
708 	    TCPOPT_TIMESTAMP<<8|TCPOLEN_TIMESTAMP)))) {
709 		/*
710 		 * Make sure that previously seen segements/ACKs are delivered
711 		 * before this segement.
712 		 */
713 		force_flush = 1;
714 	}
715 
716 	/* If the driver did not pass in the checksum, set it now. */
717 	if (csum == 0x0000)
718 		csum = th->th_sum;
719 
720 	seq = ntohl(th->th_seq);
721 
722 	if (!use_hash) {
723 		bucket = &lc->lro_hash[0];
724 	} else if (M_HASHTYPE_ISHASH(m)) {
725 		bucket = &lc->lro_hash[m->m_pkthdr.flowid % lc->lro_hashsz];
726 	} else {
727 		uint32_t hash;
728 
729 		switch (eh_type) {
730 #ifdef INET
731 		case ETHERTYPE_IP:
732 			hash = ip4->ip_src.s_addr + ip4->ip_dst.s_addr;
733 			break;
734 #endif
735 #ifdef INET6
736 		case ETHERTYPE_IPV6:
737 			hash = ip6->ip6_src.s6_addr32[0] +
738 			    ip6->ip6_dst.s6_addr32[0];
739 			hash += ip6->ip6_src.s6_addr32[1] +
740 			    ip6->ip6_dst.s6_addr32[1];
741 			hash += ip6->ip6_src.s6_addr32[2] +
742 			    ip6->ip6_dst.s6_addr32[2];
743 			hash += ip6->ip6_src.s6_addr32[3] +
744 			    ip6->ip6_dst.s6_addr32[3];
745 			break;
746 #endif
747 		default:
748 			hash = 0;
749 			break;
750 		}
751 		hash += th->th_sport + th->th_dport;
752 		bucket = &lc->lro_hash[hash % lc->lro_hashsz];
753 	}
754 
755 	/* Try to find a matching previous segment. */
756 	LIST_FOREACH(le, bucket, hash_next) {
757 		if (le->eh_type != eh_type)
758 			continue;
759 		if (le->source_port != th->th_sport ||
760 		    le->dest_port != th->th_dport)
761 			continue;
762 		switch (eh_type) {
763 #ifdef INET6
764 		case ETHERTYPE_IPV6:
765 			if (bcmp(&le->source_ip6, &ip6->ip6_src,
766 			    sizeof(struct in6_addr)) != 0 ||
767 			    bcmp(&le->dest_ip6, &ip6->ip6_dst,
768 			    sizeof(struct in6_addr)) != 0)
769 				continue;
770 			break;
771 #endif
772 #ifdef INET
773 		case ETHERTYPE_IP:
774 			if (le->source_ip4 != ip4->ip_src.s_addr ||
775 			    le->dest_ip4 != ip4->ip_dst.s_addr)
776 				continue;
777 			break;
778 #endif
779 		}
780 
781 		if (force_flush) {
782 			/* Timestamps mismatch; this is a FIN, etc */
783 			tcp_lro_active_remove(le);
784 			tcp_lro_flush(lc, le);
785 			return (TCP_LRO_CANNOT);
786 		}
787 
788 		/* Flush now if appending will result in overflow. */
789 		if (le->p_len > (lc->lro_length_lim - tcp_data_len)) {
790 			tcp_lro_active_remove(le);
791 			tcp_lro_flush(lc, le);
792 			break;
793 		}
794 
795 		/* Try to append the new segment. */
796 		if (__predict_false(seq != le->next_seq ||
797 		    (tcp_data_len == 0 && le->ack_seq == th->th_ack))) {
798 			/* Out of order packet or duplicate ACK. */
799 			tcp_lro_active_remove(le);
800 			tcp_lro_flush(lc, le);
801 			return (TCP_LRO_CANNOT);
802 		}
803 
804 		if (l != 0) {
805 			uint32_t tsval = ntohl(*(ts_ptr + 1));
806 			/* Make sure timestamp values are increasing. */
807 			/* XXX-BZ flip and use TSTMP_GEQ macro for this? */
808 			if (__predict_false(le->tsval > tsval ||
809 			    *(ts_ptr + 2) == 0))
810 				return (TCP_LRO_CANNOT);
811 			le->tsval = tsval;
812 			le->tsecr = *(ts_ptr + 2);
813 		}
814 
815 		le->next_seq += tcp_data_len;
816 		le->ack_seq = th->th_ack;
817 		le->window = th->th_win;
818 		le->append_cnt++;
819 
820 #ifdef TCP_LRO_UPDATE_CSUM
821 		le->ulp_csum += tcp_lro_rx_csum_fixup(le, l3hdr, th,
822 		    tcp_data_len, ~csum);
823 #endif
824 
825 		if (tcp_data_len == 0) {
826 			m_freem(m);
827 			/*
828 			 * Flush this LRO entry, if this ACK should not
829 			 * be further delayed.
830 			 */
831 			if (le->append_cnt >= lc->lro_ackcnt_lim) {
832 				tcp_lro_active_remove(le);
833 				tcp_lro_flush(lc, le);
834 			}
835 			return (0);
836 		}
837 
838 		le->p_len += tcp_data_len;
839 
840 		/*
841 		 * Adjust the mbuf so that m_data points to the first byte of
842 		 * the ULP payload.  Adjust the mbuf to avoid complications and
843 		 * append new segment to existing mbuf chain.
844 		 */
845 		m_adj(m, m->m_pkthdr.len - tcp_data_len);
846 		m_demote_pkthdr(m);
847 
848 		le->m_tail->m_next = m;
849 		le->m_tail = m_last(m);
850 
851 		/*
852 		 * If a possible next full length packet would cause an
853 		 * overflow, pro-actively flush now.
854 		 */
855 		if (le->p_len > (lc->lro_length_lim - lc->ifp->if_mtu)) {
856 			tcp_lro_active_remove(le);
857 			tcp_lro_flush(lc, le);
858 		} else
859 			getmicrotime(&le->mtime);
860 
861 		return (0);
862 	}
863 
864 	if (force_flush) {
865 		/*
866 		 * Nothing to flush, but this segment can not be further
867 		 * aggregated/delayed.
868 		 */
869 		return (TCP_LRO_CANNOT);
870 	}
871 
872 	/* Try to find an empty slot. */
873 	if (LIST_EMPTY(&lc->lro_free))
874 		return (TCP_LRO_NO_ENTRIES);
875 
876 	/* Start a new segment chain. */
877 	le = LIST_FIRST(&lc->lro_free);
878 	LIST_REMOVE(le, next);
879 	tcp_lro_active_insert(lc, bucket, le);
880 	getmicrotime(&le->mtime);
881 
882 	/* Start filling in details. */
883 	switch (eh_type) {
884 #ifdef INET6
885 	case ETHERTYPE_IPV6:
886 		le->le_ip6 = ip6;
887 		le->source_ip6 = ip6->ip6_src;
888 		le->dest_ip6 = ip6->ip6_dst;
889 		le->eh_type = eh_type;
890 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN - sizeof(*ip6);
891 		break;
892 #endif
893 #ifdef INET
894 	case ETHERTYPE_IP:
895 		le->le_ip4 = ip4;
896 		le->source_ip4 = ip4->ip_src.s_addr;
897 		le->dest_ip4 = ip4->ip_dst.s_addr;
898 		le->eh_type = eh_type;
899 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN;
900 		break;
901 #endif
902 	}
903 	le->source_port = th->th_sport;
904 	le->dest_port = th->th_dport;
905 
906 	le->next_seq = seq + tcp_data_len;
907 	le->ack_seq = th->th_ack;
908 	le->window = th->th_win;
909 	if (l != 0) {
910 		le->timestamp = 1;
911 		le->tsval = ntohl(*(ts_ptr + 1));
912 		le->tsecr = *(ts_ptr + 2);
913 	}
914 
915 #ifdef TCP_LRO_UPDATE_CSUM
916 	/*
917 	 * Do not touch the csum of the first packet.  However save the
918 	 * "adjusted" checksum of just the source and destination addresses,
919 	 * the next header and the TCP payload.  The length and TCP header
920 	 * parts may change, so we remove those from the saved checksum and
921 	 * re-add with final values on tcp_lro_flush() if needed.
922 	 */
923 	KASSERT(le->ulp_csum == 0, ("%s: le=%p le->ulp_csum=0x%04x\n",
924 	    __func__, le, le->ulp_csum));
925 
926 	le->ulp_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th, tcp_data_len,
927 	    ~csum);
928 	th->th_sum = csum;	/* Restore checksum on first packet. */
929 #endif
930 
931 	le->m_head = m;
932 	le->m_tail = m_last(m);
933 
934 	return (0);
935 }
936 
937 int
938 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
939 {
940 
941 	return tcp_lro_rx2(lc, m, csum, 1);
942 }
943 
944 void
945 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
946 {
947 	/* sanity checks */
948 	if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
949 	    lc->lro_mbuf_max == 0)) {
950 		/* packet drop */
951 		m_freem(mb);
952 		return;
953 	}
954 
955 	/* check if packet is not LRO capable */
956 	if (__predict_false(mb->m_pkthdr.csum_flags == 0 ||
957 	    (lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
958 		lc->lro_flushed++;
959 		lc->lro_queued++;
960 
961 		/* input packet to network layer */
962 		(*lc->ifp->if_input) (lc->ifp, mb);
963 		return;
964 	}
965 
966 	/* check if array is full */
967 	if (__predict_false(lc->lro_mbuf_count == lc->lro_mbuf_max))
968 		tcp_lro_flush_all(lc);
969 
970 	/* create sequence number */
971 	lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
972 	    (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
973 	    (((uint64_t)mb->m_pkthdr.flowid) << 24) |
974 	    ((uint64_t)lc->lro_mbuf_count);
975 
976 	/* enter mbuf */
977 	lc->lro_mbuf_data[lc->lro_mbuf_count++].mb = mb;
978 }
979 
980 /* end */
981