xref: /freebsd/sys/netinet/tcp_lro.c (revision ce437beff1f521b5ac35ea67d0669c46b3bef27f)
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 	(*lc->ifp->if_input)(lc->ifp, le->m_head);
396 	lc->lro_queued += le->append_cnt + 1;
397 	lc->lro_flushed++;
398 	bzero(le, sizeof(*le));
399 	LIST_INSERT_HEAD(&lc->lro_free, le, next);
400 }
401 
402 #ifdef HAVE_INLINE_FLSLL
403 #define	tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
404 #else
405 static inline uint64_t
406 tcp_lro_msb_64(uint64_t x)
407 {
408 	x |= (x >> 1);
409 	x |= (x >> 2);
410 	x |= (x >> 4);
411 	x |= (x >> 8);
412 	x |= (x >> 16);
413 	x |= (x >> 32);
414 	return (x & ~(x >> 1));
415 }
416 #endif
417 
418 /*
419  * The tcp_lro_sort() routine is comparable to qsort(), except it has
420  * a worst case complexity limit of O(MIN(N,64)*N), where N is the
421  * number of elements to sort and 64 is the number of sequence bits
422  * available. The algorithm is bit-slicing the 64-bit sequence number,
423  * sorting one bit at a time from the most significant bit until the
424  * least significant one, skipping the constant bits. This is
425  * typically called a radix sort.
426  */
427 static void
428 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
429 {
430 	struct lro_mbuf_sort temp;
431 	uint64_t ones;
432 	uint64_t zeros;
433 	uint32_t x;
434 	uint32_t y;
435 
436 repeat:
437 	/* for small arrays insertion sort is faster */
438 	if (size <= 12) {
439 		for (x = 1; x < size; x++) {
440 			temp = parray[x];
441 			for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
442 				parray[y] = parray[y - 1];
443 			parray[y] = temp;
444 		}
445 		return;
446 	}
447 
448 	/* compute sequence bits which are constant */
449 	ones = 0;
450 	zeros = 0;
451 	for (x = 0; x != size; x++) {
452 		ones |= parray[x].seq;
453 		zeros |= ~parray[x].seq;
454 	}
455 
456 	/* compute bits which are not constant into "ones" */
457 	ones &= zeros;
458 	if (ones == 0)
459 		return;
460 
461 	/* pick the most significant bit which is not constant */
462 	ones = tcp_lro_msb_64(ones);
463 
464 	/*
465 	 * Move entries having cleared sequence bits to the beginning
466 	 * of the array:
467 	 */
468 	for (x = y = 0; y != size; y++) {
469 		/* skip set bits */
470 		if (parray[y].seq & ones)
471 			continue;
472 		/* swap entries */
473 		temp = parray[x];
474 		parray[x] = parray[y];
475 		parray[y] = temp;
476 		x++;
477 	}
478 
479 	KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
480 
481 	/* sort zeros */
482 	tcp_lro_sort(parray, x);
483 
484 	/* sort ones */
485 	parray += x;
486 	size -= x;
487 	goto repeat;
488 }
489 
490 void
491 tcp_lro_flush_all(struct lro_ctrl *lc)
492 {
493 	uint64_t seq;
494 	uint64_t nseq;
495 	unsigned x;
496 
497 	/* check if no mbufs to flush */
498 	if (lc->lro_mbuf_count == 0)
499 		goto done;
500 
501 	/* sort all mbufs according to stream */
502 	tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
503 
504 	/* input data into LRO engine, stream by stream */
505 	seq = 0;
506 	for (x = 0; x != lc->lro_mbuf_count; x++) {
507 		struct mbuf *mb;
508 
509 		/* get mbuf */
510 		mb = lc->lro_mbuf_data[x].mb;
511 
512 		/* get sequence number, masking away the packet index */
513 		nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
514 
515 		/* check for new stream */
516 		if (seq != nseq) {
517 			seq = nseq;
518 
519 			/* flush active streams */
520 			tcp_lro_rx_done(lc);
521 		}
522 
523 		/* add packet to LRO engine */
524 		if (tcp_lro_rx2(lc, mb, 0, 0) != 0) {
525 			/* input packet to network layer */
526 			(*lc->ifp->if_input)(lc->ifp, mb);
527 			lc->lro_queued++;
528 			lc->lro_flushed++;
529 		}
530 	}
531 done:
532 	/* flush active streams */
533 	tcp_lro_rx_done(lc);
534 
535 	lc->lro_mbuf_count = 0;
536 }
537 
538 #ifdef INET6
539 static int
540 tcp_lro_rx_ipv6(struct lro_ctrl *lc, struct mbuf *m, struct ip6_hdr *ip6,
541     struct tcphdr **th)
542 {
543 
544 	/* XXX-BZ we should check the flow-label. */
545 
546 	/* XXX-BZ We do not yet support ext. hdrs. */
547 	if (ip6->ip6_nxt != IPPROTO_TCP)
548 		return (TCP_LRO_NOT_SUPPORTED);
549 
550 	/* Find the TCP header. */
551 	*th = (struct tcphdr *)(ip6 + 1);
552 
553 	return (0);
554 }
555 #endif
556 
557 #ifdef INET
558 static int
559 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4,
560     struct tcphdr **th)
561 {
562 	int csum_flags;
563 	uint16_t csum;
564 
565 	if (ip4->ip_p != IPPROTO_TCP)
566 		return (TCP_LRO_NOT_SUPPORTED);
567 
568 	/* Ensure there are no options. */
569 	if ((ip4->ip_hl << 2) != sizeof (*ip4))
570 		return (TCP_LRO_CANNOT);
571 
572 	/* .. and the packet is not fragmented. */
573 	if (ip4->ip_off & htons(IP_MF|IP_OFFMASK))
574 		return (TCP_LRO_CANNOT);
575 
576 	/* Legacy IP has a header checksum that needs to be correct. */
577 	csum_flags = m->m_pkthdr.csum_flags;
578 	if (csum_flags & CSUM_IP_CHECKED) {
579 		if (__predict_false((csum_flags & CSUM_IP_VALID) == 0)) {
580 			lc->lro_bad_csum++;
581 			return (TCP_LRO_CANNOT);
582 		}
583 	} else {
584 		csum = in_cksum_hdr(ip4);
585 		if (__predict_false((csum) != 0)) {
586 			lc->lro_bad_csum++;
587 			return (TCP_LRO_CANNOT);
588 		}
589 	}
590 
591 	/* Find the TCP header (we assured there are no IP options). */
592 	*th = (struct tcphdr *)(ip4 + 1);
593 
594 	return (0);
595 }
596 #endif
597 
598 static int
599 tcp_lro_rx2(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, int use_hash)
600 {
601 	struct lro_entry *le;
602 	struct ether_header *eh;
603 #ifdef INET6
604 	struct ip6_hdr *ip6 = NULL;	/* Keep compiler happy. */
605 #endif
606 #ifdef INET
607 	struct ip *ip4 = NULL;		/* Keep compiler happy. */
608 #endif
609 	struct tcphdr *th;
610 	void *l3hdr = NULL;		/* Keep compiler happy. */
611 	uint32_t *ts_ptr;
612 	tcp_seq seq;
613 	int error, ip_len, l;
614 	uint16_t eh_type, tcp_data_len;
615 	struct lro_head *bucket;
616 	int force_flush = 0;
617 
618 	/* We expect a contiguous header [eh, ip, tcp]. */
619 
620 	eh = mtod(m, struct ether_header *);
621 	eh_type = ntohs(eh->ether_type);
622 	switch (eh_type) {
623 #ifdef INET6
624 	case ETHERTYPE_IPV6:
625 	{
626 		CURVNET_SET(lc->ifp->if_vnet);
627 		if (V_ip6_forwarding != 0) {
628 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
629 			CURVNET_RESTORE();
630 			return (TCP_LRO_CANNOT);
631 		}
632 		CURVNET_RESTORE();
633 		l3hdr = ip6 = (struct ip6_hdr *)(eh + 1);
634 		error = tcp_lro_rx_ipv6(lc, m, ip6, &th);
635 		if (error != 0)
636 			return (error);
637 		tcp_data_len = ntohs(ip6->ip6_plen);
638 		ip_len = sizeof(*ip6) + tcp_data_len;
639 		break;
640 	}
641 #endif
642 #ifdef INET
643 	case ETHERTYPE_IP:
644 	{
645 		CURVNET_SET(lc->ifp->if_vnet);
646 		if (V_ipforwarding != 0) {
647 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
648 			CURVNET_RESTORE();
649 			return (TCP_LRO_CANNOT);
650 		}
651 		CURVNET_RESTORE();
652 		l3hdr = ip4 = (struct ip *)(eh + 1);
653 		error = tcp_lro_rx_ipv4(lc, m, ip4, &th);
654 		if (error != 0)
655 			return (error);
656 		ip_len = ntohs(ip4->ip_len);
657 		tcp_data_len = ip_len - sizeof(*ip4);
658 		break;
659 	}
660 #endif
661 	/* XXX-BZ what happens in case of VLAN(s)? */
662 	default:
663 		return (TCP_LRO_NOT_SUPPORTED);
664 	}
665 
666 	/*
667 	 * If the frame is padded beyond the end of the IP packet, then we must
668 	 * trim the extra bytes off.
669 	 */
670 	l = m->m_pkthdr.len - (ETHER_HDR_LEN + ip_len);
671 	if (l != 0) {
672 		if (l < 0)
673 			/* Truncated packet. */
674 			return (TCP_LRO_CANNOT);
675 
676 		m_adj(m, -l);
677 	}
678 
679 	/*
680 	 * Check TCP header constraints.
681 	 */
682 	/* Ensure no bits set besides ACK or PSH. */
683 	if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0) {
684 		if (th->th_flags & TH_SYN)
685 			return (TCP_LRO_CANNOT);
686 		/*
687 		 * Make sure that previously seen segements/ACKs are delivered
688 		 * before this segement, e.g. FIN.
689 		 */
690 		force_flush = 1;
691 	}
692 
693 	/* XXX-BZ We lose a ACK|PUSH flag concatenating multiple segments. */
694 	/* XXX-BZ Ideally we'd flush on PUSH? */
695 
696 	/*
697 	 * Check for timestamps.
698 	 * Since the only option we handle are timestamps, we only have to
699 	 * handle the simple case of aligned timestamps.
700 	 */
701 	l = (th->th_off << 2);
702 	tcp_data_len -= l;
703 	l -= sizeof(*th);
704 	ts_ptr = (uint32_t *)(th + 1);
705 	if (l != 0 && (__predict_false(l != TCPOLEN_TSTAMP_APPA) ||
706 	    (*ts_ptr != ntohl(TCPOPT_NOP<<24|TCPOPT_NOP<<16|
707 	    TCPOPT_TIMESTAMP<<8|TCPOLEN_TIMESTAMP)))) {
708 		/*
709 		 * Make sure that previously seen segements/ACKs are delivered
710 		 * before this segement.
711 		 */
712 		force_flush = 1;
713 	}
714 
715 	/* If the driver did not pass in the checksum, set it now. */
716 	if (csum == 0x0000)
717 		csum = th->th_sum;
718 
719 	seq = ntohl(th->th_seq);
720 
721 	if (!use_hash) {
722 		bucket = &lc->lro_hash[0];
723 	} else if (M_HASHTYPE_ISHASH(m)) {
724 		bucket = &lc->lro_hash[m->m_pkthdr.flowid % lc->lro_hashsz];
725 	} else {
726 		uint32_t hash;
727 
728 		switch (eh_type) {
729 #ifdef INET
730 		case ETHERTYPE_IP:
731 			hash = ip4->ip_src.s_addr + ip4->ip_dst.s_addr;
732 			break;
733 #endif
734 #ifdef INET6
735 		case ETHERTYPE_IPV6:
736 			hash = ip6->ip6_src.s6_addr32[0] +
737 			    ip6->ip6_dst.s6_addr32[0];
738 			hash += ip6->ip6_src.s6_addr32[1] +
739 			    ip6->ip6_dst.s6_addr32[1];
740 			hash += ip6->ip6_src.s6_addr32[2] +
741 			    ip6->ip6_dst.s6_addr32[2];
742 			hash += ip6->ip6_src.s6_addr32[3] +
743 			    ip6->ip6_dst.s6_addr32[3];
744 			break;
745 #endif
746 		default:
747 			hash = 0;
748 			break;
749 		}
750 		hash += th->th_sport + th->th_dport;
751 		bucket = &lc->lro_hash[hash % lc->lro_hashsz];
752 	}
753 
754 	/* Try to find a matching previous segment. */
755 	LIST_FOREACH(le, bucket, hash_next) {
756 		if (le->eh_type != eh_type)
757 			continue;
758 		if (le->source_port != th->th_sport ||
759 		    le->dest_port != th->th_dport)
760 			continue;
761 		switch (eh_type) {
762 #ifdef INET6
763 		case ETHERTYPE_IPV6:
764 			if (bcmp(&le->source_ip6, &ip6->ip6_src,
765 			    sizeof(struct in6_addr)) != 0 ||
766 			    bcmp(&le->dest_ip6, &ip6->ip6_dst,
767 			    sizeof(struct in6_addr)) != 0)
768 				continue;
769 			break;
770 #endif
771 #ifdef INET
772 		case ETHERTYPE_IP:
773 			if (le->source_ip4 != ip4->ip_src.s_addr ||
774 			    le->dest_ip4 != ip4->ip_dst.s_addr)
775 				continue;
776 			break;
777 #endif
778 		}
779 
780 		if (force_flush) {
781 			/* Timestamps mismatch; this is a FIN, etc */
782 			tcp_lro_active_remove(le);
783 			tcp_lro_flush(lc, le);
784 			return (TCP_LRO_CANNOT);
785 		}
786 
787 		/* Flush now if appending will result in overflow. */
788 		if (le->p_len > (lc->lro_length_lim - tcp_data_len)) {
789 			tcp_lro_active_remove(le);
790 			tcp_lro_flush(lc, le);
791 			break;
792 		}
793 
794 		/* Try to append the new segment. */
795 		if (__predict_false(seq != le->next_seq ||
796 		    (tcp_data_len == 0 && le->ack_seq == th->th_ack))) {
797 			/* Out of order packet or duplicate ACK. */
798 			tcp_lro_active_remove(le);
799 			tcp_lro_flush(lc, le);
800 			return (TCP_LRO_CANNOT);
801 		}
802 
803 		if (l != 0) {
804 			uint32_t tsval = ntohl(*(ts_ptr + 1));
805 			/* Make sure timestamp values are increasing. */
806 			/* XXX-BZ flip and use TSTMP_GEQ macro for this? */
807 			if (__predict_false(le->tsval > tsval ||
808 			    *(ts_ptr + 2) == 0))
809 				return (TCP_LRO_CANNOT);
810 			le->tsval = tsval;
811 			le->tsecr = *(ts_ptr + 2);
812 		}
813 
814 		le->next_seq += tcp_data_len;
815 		le->ack_seq = th->th_ack;
816 		le->window = th->th_win;
817 		le->append_cnt++;
818 
819 #ifdef TCP_LRO_UPDATE_CSUM
820 		le->ulp_csum += tcp_lro_rx_csum_fixup(le, l3hdr, th,
821 		    tcp_data_len, ~csum);
822 #endif
823 
824 		if (tcp_data_len == 0) {
825 			m_freem(m);
826 			/*
827 			 * Flush this LRO entry, if this ACK should not
828 			 * be further delayed.
829 			 */
830 			if (le->append_cnt >= lc->lro_ackcnt_lim) {
831 				tcp_lro_active_remove(le);
832 				tcp_lro_flush(lc, le);
833 			}
834 			return (0);
835 		}
836 
837 		le->p_len += tcp_data_len;
838 
839 		/*
840 		 * Adjust the mbuf so that m_data points to the first byte of
841 		 * the ULP payload.  Adjust the mbuf to avoid complications and
842 		 * append new segment to existing mbuf chain.
843 		 */
844 		m_adj(m, m->m_pkthdr.len - tcp_data_len);
845 		m_demote_pkthdr(m);
846 
847 		le->m_tail->m_next = m;
848 		le->m_tail = m_last(m);
849 
850 		/*
851 		 * If a possible next full length packet would cause an
852 		 * overflow, pro-actively flush now.
853 		 */
854 		if (le->p_len > (lc->lro_length_lim - lc->ifp->if_mtu)) {
855 			tcp_lro_active_remove(le);
856 			tcp_lro_flush(lc, le);
857 		} else
858 			getmicrotime(&le->mtime);
859 
860 		return (0);
861 	}
862 
863 	if (force_flush) {
864 		/*
865 		 * Nothing to flush, but this segment can not be further
866 		 * aggregated/delayed.
867 		 */
868 		return (TCP_LRO_CANNOT);
869 	}
870 
871 	/* Try to find an empty slot. */
872 	if (LIST_EMPTY(&lc->lro_free))
873 		return (TCP_LRO_NO_ENTRIES);
874 
875 	/* Start a new segment chain. */
876 	le = LIST_FIRST(&lc->lro_free);
877 	LIST_REMOVE(le, next);
878 	tcp_lro_active_insert(lc, bucket, le);
879 	getmicrotime(&le->mtime);
880 
881 	/* Start filling in details. */
882 	switch (eh_type) {
883 #ifdef INET6
884 	case ETHERTYPE_IPV6:
885 		le->le_ip6 = ip6;
886 		le->source_ip6 = ip6->ip6_src;
887 		le->dest_ip6 = ip6->ip6_dst;
888 		le->eh_type = eh_type;
889 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN - sizeof(*ip6);
890 		break;
891 #endif
892 #ifdef INET
893 	case ETHERTYPE_IP:
894 		le->le_ip4 = ip4;
895 		le->source_ip4 = ip4->ip_src.s_addr;
896 		le->dest_ip4 = ip4->ip_dst.s_addr;
897 		le->eh_type = eh_type;
898 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN;
899 		break;
900 #endif
901 	}
902 	le->source_port = th->th_sport;
903 	le->dest_port = th->th_dport;
904 
905 	le->next_seq = seq + tcp_data_len;
906 	le->ack_seq = th->th_ack;
907 	le->window = th->th_win;
908 	if (l != 0) {
909 		le->timestamp = 1;
910 		le->tsval = ntohl(*(ts_ptr + 1));
911 		le->tsecr = *(ts_ptr + 2);
912 	}
913 
914 #ifdef TCP_LRO_UPDATE_CSUM
915 	/*
916 	 * Do not touch the csum of the first packet.  However save the
917 	 * "adjusted" checksum of just the source and destination addresses,
918 	 * the next header and the TCP payload.  The length and TCP header
919 	 * parts may change, so we remove those from the saved checksum and
920 	 * re-add with final values on tcp_lro_flush() if needed.
921 	 */
922 	KASSERT(le->ulp_csum == 0, ("%s: le=%p le->ulp_csum=0x%04x\n",
923 	    __func__, le, le->ulp_csum));
924 
925 	le->ulp_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th, tcp_data_len,
926 	    ~csum);
927 	th->th_sum = csum;	/* Restore checksum on first packet. */
928 #endif
929 
930 	le->m_head = m;
931 	le->m_tail = m_last(m);
932 
933 	return (0);
934 }
935 
936 int
937 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
938 {
939 
940 	return tcp_lro_rx2(lc, m, csum, 1);
941 }
942 
943 void
944 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
945 {
946 	/* sanity checks */
947 	if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
948 	    lc->lro_mbuf_max == 0)) {
949 		/* packet drop */
950 		m_freem(mb);
951 		return;
952 	}
953 
954 	/* check if packet is not LRO capable */
955 	if (__predict_false(mb->m_pkthdr.csum_flags == 0 ||
956 	    (lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
957 		lc->lro_flushed++;
958 		lc->lro_queued++;
959 
960 		/* input packet to network layer */
961 		(*lc->ifp->if_input) (lc->ifp, mb);
962 		return;
963 	}
964 
965 	/* check if array is full */
966 	if (__predict_false(lc->lro_mbuf_count == lc->lro_mbuf_max))
967 		tcp_lro_flush_all(lc);
968 
969 	/* create sequence number */
970 	lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
971 	    (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
972 	    (((uint64_t)mb->m_pkthdr.flowid) << 24) |
973 	    ((uint64_t)lc->lro_mbuf_count);
974 
975 	/* enter mbuf */
976 	lc->lro_mbuf_data[lc->lro_mbuf_count++].mb = mb;
977 }
978 
979 /* end */
980