xref: /freebsd/sys/netinet/tcp_lro.c (revision 6829dae12bb055451fa467da4589c43bd03b1e64)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2007, Myricom Inc.
5  * Copyright (c) 2008, Intel Corporation.
6  * Copyright (c) 2012 The FreeBSD Foundation
7  * Copyright (c) 2016 Mellanox Technologies.
8  * All rights reserved.
9  *
10  * Portions of this software were developed by Bjoern Zeeb
11  * under sponsorship from the FreeBSD Foundation.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include "opt_inet.h"
39 #include "opt_inet6.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/malloc.h>
45 #include <sys/mbuf.h>
46 #include <sys/socket.h>
47 #include <sys/sysctl.h>
48 
49 #include <net/if.h>
50 #include <net/if_var.h>
51 #include <net/ethernet.h>
52 #include <net/vnet.h>
53 
54 #include <netinet/in_systm.h>
55 #include <netinet/in.h>
56 #include <netinet/ip6.h>
57 #include <netinet/ip.h>
58 #include <netinet/ip_var.h>
59 #include <netinet/tcp.h>
60 #include <netinet/tcp_seq.h>
61 #include <netinet/tcp_lro.h>
62 #include <netinet/tcp_var.h>
63 
64 #include <netinet6/ip6_var.h>
65 
66 #include <machine/in_cksum.h>
67 
68 static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures");
69 
70 #define	TCP_LRO_UPDATE_CSUM	1
71 #ifndef	TCP_LRO_UPDATE_CSUM
72 #define	TCP_LRO_INVALID_CSUM	0x0000
73 #endif
74 
75 static void	tcp_lro_rx_done(struct lro_ctrl *lc);
76 static int	tcp_lro_rx2(struct lro_ctrl *lc, struct mbuf *m,
77 		    uint32_t csum, int use_hash);
78 
79 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro,  CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
80     "TCP LRO");
81 
82 static unsigned	tcp_lro_entries = TCP_LRO_ENTRIES;
83 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries,
84     CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_entries, 0,
85     "default number of LRO entries");
86 
87 static __inline void
88 tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket,
89     struct lro_entry *le)
90 {
91 
92 	LIST_INSERT_HEAD(&lc->lro_active, le, next);
93 	LIST_INSERT_HEAD(bucket, le, hash_next);
94 }
95 
96 static __inline void
97 tcp_lro_active_remove(struct lro_entry *le)
98 {
99 
100 	LIST_REMOVE(le, next);		/* active list */
101 	LIST_REMOVE(le, hash_next);	/* hash bucket */
102 }
103 
104 int
105 tcp_lro_init(struct lro_ctrl *lc)
106 {
107 	return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0));
108 }
109 
110 int
111 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
112     unsigned lro_entries, unsigned lro_mbufs)
113 {
114 	struct lro_entry *le;
115 	size_t size;
116 	unsigned i, elements;
117 
118 	lc->lro_bad_csum = 0;
119 	lc->lro_queued = 0;
120 	lc->lro_flushed = 0;
121 	lc->lro_mbuf_count = 0;
122 	lc->lro_mbuf_max = lro_mbufs;
123 	lc->lro_cnt = lro_entries;
124 	lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
125 	lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
126 	lc->ifp = ifp;
127 	LIST_INIT(&lc->lro_free);
128 	LIST_INIT(&lc->lro_active);
129 
130 	/* create hash table to accelerate entry lookup */
131 	if (lro_entries > lro_mbufs)
132 		elements = lro_entries;
133 	else
134 		elements = lro_mbufs;
135 	lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz,
136 	    HASH_NOWAIT);
137 	if (lc->lro_hash == NULL) {
138 		memset(lc, 0, sizeof(*lc));
139 		return (ENOMEM);
140 	}
141 
142 	/* compute size to allocate */
143 	size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
144 	    (lro_entries * sizeof(*le));
145 	lc->lro_mbuf_data = (struct lro_mbuf_sort *)
146 	    malloc(size, M_LRO, M_NOWAIT | M_ZERO);
147 
148 	/* check for out of memory */
149 	if (lc->lro_mbuf_data == NULL) {
150 		free(lc->lro_hash, M_LRO);
151 		memset(lc, 0, sizeof(*lc));
152 		return (ENOMEM);
153 	}
154 	/* compute offset for LRO entries */
155 	le = (struct lro_entry *)
156 	    (lc->lro_mbuf_data + lro_mbufs);
157 
158 	/* setup linked list */
159 	for (i = 0; i != lro_entries; i++)
160 		LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
161 
162 	return (0);
163 }
164 
165 void
166 tcp_lro_free(struct lro_ctrl *lc)
167 {
168 	struct lro_entry *le;
169 	unsigned x;
170 
171 	/* reset LRO free list */
172 	LIST_INIT(&lc->lro_free);
173 
174 	/* free active mbufs, if any */
175 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
176 		tcp_lro_active_remove(le);
177 		m_freem(le->m_head);
178 	}
179 
180 	/* free hash table */
181 	free(lc->lro_hash, M_LRO);
182 	lc->lro_hash = NULL;
183 	lc->lro_hashsz = 0;
184 
185 	/* free mbuf array, if any */
186 	for (x = 0; x != lc->lro_mbuf_count; x++)
187 		m_freem(lc->lro_mbuf_data[x].mb);
188 	lc->lro_mbuf_count = 0;
189 
190 	/* free allocated memory, if any */
191 	free(lc->lro_mbuf_data, M_LRO);
192 	lc->lro_mbuf_data = NULL;
193 }
194 
195 #ifdef TCP_LRO_UPDATE_CSUM
196 static uint16_t
197 tcp_lro_csum_th(struct tcphdr *th)
198 {
199 	uint32_t ch;
200 	uint16_t *p, l;
201 
202 	ch = th->th_sum = 0x0000;
203 	l = th->th_off;
204 	p = (uint16_t *)th;
205 	while (l > 0) {
206 		ch += *p;
207 		p++;
208 		ch += *p;
209 		p++;
210 		l--;
211 	}
212 	while (ch > 0xffff)
213 		ch = (ch >> 16) + (ch & 0xffff);
214 
215 	return (ch & 0xffff);
216 }
217 
218 static uint16_t
219 tcp_lro_rx_csum_fixup(struct lro_entry *le, void *l3hdr, struct tcphdr *th,
220     uint16_t tcp_data_len, uint16_t csum)
221 {
222 	uint32_t c;
223 	uint16_t cs;
224 
225 	c = csum;
226 
227 	/* Remove length from checksum. */
228 	switch (le->eh_type) {
229 #ifdef INET6
230 	case ETHERTYPE_IPV6:
231 	{
232 		struct ip6_hdr *ip6;
233 
234 		ip6 = (struct ip6_hdr *)l3hdr;
235 		if (le->append_cnt == 0)
236 			cs = ip6->ip6_plen;
237 		else {
238 			uint32_t cx;
239 
240 			cx = ntohs(ip6->ip6_plen);
241 			cs = in6_cksum_pseudo(ip6, cx, ip6->ip6_nxt, 0);
242 		}
243 		break;
244 	}
245 #endif
246 #ifdef INET
247 	case ETHERTYPE_IP:
248 	{
249 		struct ip *ip4;
250 
251 		ip4 = (struct ip *)l3hdr;
252 		if (le->append_cnt == 0)
253 			cs = ip4->ip_len;
254 		else {
255 			cs = in_addword(ntohs(ip4->ip_len) - sizeof(*ip4),
256 			    IPPROTO_TCP);
257 			cs = in_pseudo(ip4->ip_src.s_addr, ip4->ip_dst.s_addr,
258 			    htons(cs));
259 		}
260 		break;
261 	}
262 #endif
263 	default:
264 		cs = 0;		/* Keep compiler happy. */
265 	}
266 
267 	cs = ~cs;
268 	c += cs;
269 
270 	/* Remove TCP header csum. */
271 	cs = ~tcp_lro_csum_th(th);
272 	c += cs;
273 	while (c > 0xffff)
274 		c = (c >> 16) + (c & 0xffff);
275 
276 	return (c & 0xffff);
277 }
278 #endif
279 
280 static void
281 tcp_lro_rx_done(struct lro_ctrl *lc)
282 {
283 	struct lro_entry *le;
284 
285 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
286 		tcp_lro_active_remove(le);
287 		tcp_lro_flush(lc, le);
288 	}
289 }
290 
291 void
292 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
293 {
294 	struct lro_entry *le, *le_tmp;
295 	struct timeval tv;
296 
297 	if (LIST_EMPTY(&lc->lro_active))
298 		return;
299 
300 	getmicrotime(&tv);
301 	timevalsub(&tv, timeout);
302 	LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
303 		if (timevalcmp(&tv, &le->mtime, >=)) {
304 			tcp_lro_active_remove(le);
305 			tcp_lro_flush(lc, le);
306 		}
307 	}
308 }
309 
310 void
311 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
312 {
313 
314 	if (le->append_cnt > 0) {
315 		struct tcphdr *th;
316 		uint16_t p_len;
317 
318 		p_len = htons(le->p_len);
319 		switch (le->eh_type) {
320 #ifdef INET6
321 		case ETHERTYPE_IPV6:
322 		{
323 			struct ip6_hdr *ip6;
324 
325 			ip6 = le->le_ip6;
326 			ip6->ip6_plen = p_len;
327 			th = (struct tcphdr *)(ip6 + 1);
328 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
329 			    CSUM_PSEUDO_HDR;
330 			le->p_len += ETHER_HDR_LEN + sizeof(*ip6);
331 			break;
332 		}
333 #endif
334 #ifdef INET
335 		case ETHERTYPE_IP:
336 		{
337 			struct ip *ip4;
338 #ifdef TCP_LRO_UPDATE_CSUM
339 			uint32_t cl;
340 			uint16_t c;
341 #endif
342 
343 			ip4 = le->le_ip4;
344 #ifdef TCP_LRO_UPDATE_CSUM
345 			/* Fix IP header checksum for new length. */
346 			c = ~ip4->ip_sum;
347 			cl = c;
348 			c = ~ip4->ip_len;
349 			cl += c + p_len;
350 			while (cl > 0xffff)
351 				cl = (cl >> 16) + (cl & 0xffff);
352 			c = cl;
353 			ip4->ip_sum = ~c;
354 #else
355 			ip4->ip_sum = TCP_LRO_INVALID_CSUM;
356 #endif
357 			ip4->ip_len = p_len;
358 			th = (struct tcphdr *)(ip4 + 1);
359 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
360 			    CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
361 			le->p_len += ETHER_HDR_LEN;
362 			break;
363 		}
364 #endif
365 		default:
366 			th = NULL;	/* Keep compiler happy. */
367 		}
368 		le->m_head->m_pkthdr.csum_data = 0xffff;
369 		le->m_head->m_pkthdr.len = le->p_len;
370 
371 		/* Incorporate the latest ACK into the TCP header. */
372 		th->th_ack = le->ack_seq;
373 		th->th_win = le->window;
374 		/* Incorporate latest timestamp into the TCP header. */
375 		if (le->timestamp != 0) {
376 			uint32_t *ts_ptr;
377 
378 			ts_ptr = (uint32_t *)(th + 1);
379 			ts_ptr[1] = htonl(le->tsval);
380 			ts_ptr[2] = le->tsecr;
381 		}
382 #ifdef TCP_LRO_UPDATE_CSUM
383 		/* Update the TCP header checksum. */
384 		le->ulp_csum += p_len;
385 		le->ulp_csum += tcp_lro_csum_th(th);
386 		while (le->ulp_csum > 0xffff)
387 			le->ulp_csum = (le->ulp_csum >> 16) +
388 			    (le->ulp_csum & 0xffff);
389 		th->th_sum = (le->ulp_csum & 0xffff);
390 		th->th_sum = ~th->th_sum;
391 #else
392 		th->th_sum = TCP_LRO_INVALID_CSUM;
393 #endif
394 	}
395 
396 	le->m_head->m_pkthdr.lro_nsegs = le->append_cnt + 1;
397 	(*lc->ifp->if_input)(lc->ifp, le->m_head);
398 	lc->lro_queued += le->append_cnt + 1;
399 	lc->lro_flushed++;
400 	bzero(le, sizeof(*le));
401 	LIST_INSERT_HEAD(&lc->lro_free, le, next);
402 }
403 
404 #ifdef HAVE_INLINE_FLSLL
405 #define	tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
406 #else
407 static inline uint64_t
408 tcp_lro_msb_64(uint64_t x)
409 {
410 	x |= (x >> 1);
411 	x |= (x >> 2);
412 	x |= (x >> 4);
413 	x |= (x >> 8);
414 	x |= (x >> 16);
415 	x |= (x >> 32);
416 	return (x & ~(x >> 1));
417 }
418 #endif
419 
420 /*
421  * The tcp_lro_sort() routine is comparable to qsort(), except it has
422  * a worst case complexity limit of O(MIN(N,64)*N), where N is the
423  * number of elements to sort and 64 is the number of sequence bits
424  * available. The algorithm is bit-slicing the 64-bit sequence number,
425  * sorting one bit at a time from the most significant bit until the
426  * least significant one, skipping the constant bits. This is
427  * typically called a radix sort.
428  */
429 static void
430 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
431 {
432 	struct lro_mbuf_sort temp;
433 	uint64_t ones;
434 	uint64_t zeros;
435 	uint32_t x;
436 	uint32_t y;
437 
438 repeat:
439 	/* for small arrays insertion sort is faster */
440 	if (size <= 12) {
441 		for (x = 1; x < size; x++) {
442 			temp = parray[x];
443 			for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
444 				parray[y] = parray[y - 1];
445 			parray[y] = temp;
446 		}
447 		return;
448 	}
449 
450 	/* compute sequence bits which are constant */
451 	ones = 0;
452 	zeros = 0;
453 	for (x = 0; x != size; x++) {
454 		ones |= parray[x].seq;
455 		zeros |= ~parray[x].seq;
456 	}
457 
458 	/* compute bits which are not constant into "ones" */
459 	ones &= zeros;
460 	if (ones == 0)
461 		return;
462 
463 	/* pick the most significant bit which is not constant */
464 	ones = tcp_lro_msb_64(ones);
465 
466 	/*
467 	 * Move entries having cleared sequence bits to the beginning
468 	 * of the array:
469 	 */
470 	for (x = y = 0; y != size; y++) {
471 		/* skip set bits */
472 		if (parray[y].seq & ones)
473 			continue;
474 		/* swap entries */
475 		temp = parray[x];
476 		parray[x] = parray[y];
477 		parray[y] = temp;
478 		x++;
479 	}
480 
481 	KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
482 
483 	/* sort zeros */
484 	tcp_lro_sort(parray, x);
485 
486 	/* sort ones */
487 	parray += x;
488 	size -= x;
489 	goto repeat;
490 }
491 
492 void
493 tcp_lro_flush_all(struct lro_ctrl *lc)
494 {
495 	uint64_t seq;
496 	uint64_t nseq;
497 	unsigned x;
498 
499 	/* check if no mbufs to flush */
500 	if (lc->lro_mbuf_count == 0)
501 		goto done;
502 
503 	/* sort all mbufs according to stream */
504 	tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
505 
506 	/* input data into LRO engine, stream by stream */
507 	seq = 0;
508 	for (x = 0; x != lc->lro_mbuf_count; x++) {
509 		struct mbuf *mb;
510 
511 		/* get mbuf */
512 		mb = lc->lro_mbuf_data[x].mb;
513 
514 		/* get sequence number, masking away the packet index */
515 		nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
516 
517 		/* check for new stream */
518 		if (seq != nseq) {
519 			seq = nseq;
520 
521 			/* flush active streams */
522 			tcp_lro_rx_done(lc);
523 		}
524 
525 		/* add packet to LRO engine */
526 		if (tcp_lro_rx2(lc, mb, 0, 0) != 0) {
527 			/* input packet to network layer */
528 			(*lc->ifp->if_input)(lc->ifp, mb);
529 			lc->lro_queued++;
530 			lc->lro_flushed++;
531 		}
532 	}
533 done:
534 	/* flush active streams */
535 	tcp_lro_rx_done(lc);
536 
537 	lc->lro_mbuf_count = 0;
538 }
539 
540 #ifdef INET6
541 static int
542 tcp_lro_rx_ipv6(struct lro_ctrl *lc, struct mbuf *m, struct ip6_hdr *ip6,
543     struct tcphdr **th)
544 {
545 
546 	/* XXX-BZ we should check the flow-label. */
547 
548 	/* XXX-BZ We do not yet support ext. hdrs. */
549 	if (ip6->ip6_nxt != IPPROTO_TCP)
550 		return (TCP_LRO_NOT_SUPPORTED);
551 
552 	/* Find the TCP header. */
553 	*th = (struct tcphdr *)(ip6 + 1);
554 
555 	return (0);
556 }
557 #endif
558 
559 #ifdef INET
560 static int
561 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4,
562     struct tcphdr **th)
563 {
564 	int csum_flags;
565 	uint16_t csum;
566 
567 	if (ip4->ip_p != IPPROTO_TCP)
568 		return (TCP_LRO_NOT_SUPPORTED);
569 
570 	/* Ensure there are no options. */
571 	if ((ip4->ip_hl << 2) != sizeof (*ip4))
572 		return (TCP_LRO_CANNOT);
573 
574 	/* .. and the packet is not fragmented. */
575 	if (ip4->ip_off & htons(IP_MF|IP_OFFMASK))
576 		return (TCP_LRO_CANNOT);
577 
578 	/* Legacy IP has a header checksum that needs to be correct. */
579 	csum_flags = m->m_pkthdr.csum_flags;
580 	if (csum_flags & CSUM_IP_CHECKED) {
581 		if (__predict_false((csum_flags & CSUM_IP_VALID) == 0)) {
582 			lc->lro_bad_csum++;
583 			return (TCP_LRO_CANNOT);
584 		}
585 	} else {
586 		csum = in_cksum_hdr(ip4);
587 		if (__predict_false((csum) != 0)) {
588 			lc->lro_bad_csum++;
589 			return (TCP_LRO_CANNOT);
590 		}
591 	}
592 
593 	/* Find the TCP header (we assured there are no IP options). */
594 	*th = (struct tcphdr *)(ip4 + 1);
595 
596 	return (0);
597 }
598 #endif
599 
600 static int
601 tcp_lro_rx2(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, int use_hash)
602 {
603 	struct lro_entry *le;
604 	struct ether_header *eh;
605 #ifdef INET6
606 	struct ip6_hdr *ip6 = NULL;	/* Keep compiler happy. */
607 #endif
608 #ifdef INET
609 	struct ip *ip4 = NULL;		/* Keep compiler happy. */
610 #endif
611 	struct tcphdr *th;
612 	void *l3hdr = NULL;		/* Keep compiler happy. */
613 	uint32_t *ts_ptr;
614 	tcp_seq seq;
615 	int error, ip_len, l;
616 	uint16_t eh_type, tcp_data_len;
617 	struct lro_head *bucket;
618 	int force_flush = 0;
619 
620 	/* We expect a contiguous header [eh, ip, tcp]. */
621 
622 	eh = mtod(m, struct ether_header *);
623 	eh_type = ntohs(eh->ether_type);
624 	switch (eh_type) {
625 #ifdef INET6
626 	case ETHERTYPE_IPV6:
627 	{
628 		CURVNET_SET(lc->ifp->if_vnet);
629 		if (V_ip6_forwarding != 0) {
630 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
631 			CURVNET_RESTORE();
632 			return (TCP_LRO_CANNOT);
633 		}
634 		CURVNET_RESTORE();
635 		l3hdr = ip6 = (struct ip6_hdr *)(eh + 1);
636 		error = tcp_lro_rx_ipv6(lc, m, ip6, &th);
637 		if (error != 0)
638 			return (error);
639 		tcp_data_len = ntohs(ip6->ip6_plen);
640 		ip_len = sizeof(*ip6) + tcp_data_len;
641 		break;
642 	}
643 #endif
644 #ifdef INET
645 	case ETHERTYPE_IP:
646 	{
647 		CURVNET_SET(lc->ifp->if_vnet);
648 		if (V_ipforwarding != 0) {
649 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
650 			CURVNET_RESTORE();
651 			return (TCP_LRO_CANNOT);
652 		}
653 		CURVNET_RESTORE();
654 		l3hdr = ip4 = (struct ip *)(eh + 1);
655 		error = tcp_lro_rx_ipv4(lc, m, ip4, &th);
656 		if (error != 0)
657 			return (error);
658 		ip_len = ntohs(ip4->ip_len);
659 		tcp_data_len = ip_len - sizeof(*ip4);
660 		break;
661 	}
662 #endif
663 	/* XXX-BZ what happens in case of VLAN(s)? */
664 	default:
665 		return (TCP_LRO_NOT_SUPPORTED);
666 	}
667 
668 	/*
669 	 * If the frame is padded beyond the end of the IP packet, then we must
670 	 * trim the extra bytes off.
671 	 */
672 	l = m->m_pkthdr.len - (ETHER_HDR_LEN + ip_len);
673 	if (l != 0) {
674 		if (l < 0)
675 			/* Truncated packet. */
676 			return (TCP_LRO_CANNOT);
677 
678 		m_adj(m, -l);
679 	}
680 
681 	/*
682 	 * Check TCP header constraints.
683 	 */
684 	/* Ensure no bits set besides ACK or PSH. */
685 	if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0) {
686 		if (th->th_flags & TH_SYN)
687 			return (TCP_LRO_CANNOT);
688 		/*
689 		 * Make sure that previously seen segements/ACKs are delivered
690 		 * before this segement, e.g. FIN.
691 		 */
692 		force_flush = 1;
693 	}
694 
695 	/* XXX-BZ We lose a ACK|PUSH flag concatenating multiple segments. */
696 	/* XXX-BZ Ideally we'd flush on PUSH? */
697 
698 	/*
699 	 * Check for timestamps.
700 	 * Since the only option we handle are timestamps, we only have to
701 	 * handle the simple case of aligned timestamps.
702 	 */
703 	l = (th->th_off << 2);
704 	tcp_data_len -= l;
705 	l -= sizeof(*th);
706 	ts_ptr = (uint32_t *)(th + 1);
707 	if (l != 0 && (__predict_false(l != TCPOLEN_TSTAMP_APPA) ||
708 	    (*ts_ptr != ntohl(TCPOPT_NOP<<24|TCPOPT_NOP<<16|
709 	    TCPOPT_TIMESTAMP<<8|TCPOLEN_TIMESTAMP)))) {
710 		/*
711 		 * Make sure that previously seen segements/ACKs are delivered
712 		 * before this segement.
713 		 */
714 		force_flush = 1;
715 	}
716 
717 	/* If the driver did not pass in the checksum, set it now. */
718 	if (csum == 0x0000)
719 		csum = th->th_sum;
720 
721 	seq = ntohl(th->th_seq);
722 
723 	if (!use_hash) {
724 		bucket = &lc->lro_hash[0];
725 	} else if (M_HASHTYPE_ISHASH(m)) {
726 		bucket = &lc->lro_hash[m->m_pkthdr.flowid % lc->lro_hashsz];
727 	} else {
728 		uint32_t hash;
729 
730 		switch (eh_type) {
731 #ifdef INET
732 		case ETHERTYPE_IP:
733 			hash = ip4->ip_src.s_addr + ip4->ip_dst.s_addr;
734 			break;
735 #endif
736 #ifdef INET6
737 		case ETHERTYPE_IPV6:
738 			hash = ip6->ip6_src.s6_addr32[0] +
739 			    ip6->ip6_dst.s6_addr32[0];
740 			hash += ip6->ip6_src.s6_addr32[1] +
741 			    ip6->ip6_dst.s6_addr32[1];
742 			hash += ip6->ip6_src.s6_addr32[2] +
743 			    ip6->ip6_dst.s6_addr32[2];
744 			hash += ip6->ip6_src.s6_addr32[3] +
745 			    ip6->ip6_dst.s6_addr32[3];
746 			break;
747 #endif
748 		default:
749 			hash = 0;
750 			break;
751 		}
752 		hash += th->th_sport + th->th_dport;
753 		bucket = &lc->lro_hash[hash % lc->lro_hashsz];
754 	}
755 
756 	/* Try to find a matching previous segment. */
757 	LIST_FOREACH(le, bucket, hash_next) {
758 		if (le->eh_type != eh_type)
759 			continue;
760 		if (le->source_port != th->th_sport ||
761 		    le->dest_port != th->th_dport)
762 			continue;
763 		switch (eh_type) {
764 #ifdef INET6
765 		case ETHERTYPE_IPV6:
766 			if (bcmp(&le->source_ip6, &ip6->ip6_src,
767 			    sizeof(struct in6_addr)) != 0 ||
768 			    bcmp(&le->dest_ip6, &ip6->ip6_dst,
769 			    sizeof(struct in6_addr)) != 0)
770 				continue;
771 			break;
772 #endif
773 #ifdef INET
774 		case ETHERTYPE_IP:
775 			if (le->source_ip4 != ip4->ip_src.s_addr ||
776 			    le->dest_ip4 != ip4->ip_dst.s_addr)
777 				continue;
778 			break;
779 #endif
780 		}
781 
782 		if (force_flush) {
783 			/* Timestamps mismatch; this is a FIN, etc */
784 			tcp_lro_active_remove(le);
785 			tcp_lro_flush(lc, le);
786 			return (TCP_LRO_CANNOT);
787 		}
788 
789 		/* Flush now if appending will result in overflow. */
790 		if (le->p_len > (lc->lro_length_lim - tcp_data_len)) {
791 			tcp_lro_active_remove(le);
792 			tcp_lro_flush(lc, le);
793 			break;
794 		}
795 
796 		/* Try to append the new segment. */
797 		if (__predict_false(seq != le->next_seq ||
798 		    (tcp_data_len == 0 &&
799 		    le->ack_seq == th->th_ack &&
800 		    le->window == th->th_win))) {
801 			/* Out of order packet or duplicate ACK. */
802 			tcp_lro_active_remove(le);
803 			tcp_lro_flush(lc, le);
804 			return (TCP_LRO_CANNOT);
805 		}
806 
807 		if (l != 0) {
808 			uint32_t tsval = ntohl(*(ts_ptr + 1));
809 			/* Make sure timestamp values are increasing. */
810 			/* XXX-BZ flip and use TSTMP_GEQ macro for this? */
811 			if (__predict_false(le->tsval > tsval ||
812 			    *(ts_ptr + 2) == 0))
813 				return (TCP_LRO_CANNOT);
814 			le->tsval = tsval;
815 			le->tsecr = *(ts_ptr + 2);
816 		}
817 		if (tcp_data_len || SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
818 			le->next_seq += tcp_data_len;
819 			le->ack_seq = th->th_ack;
820 			le->window = th->th_win;
821 			le->append_cnt++;
822 		} else if (th->th_ack == le->ack_seq) {
823 			le->window = WIN_MAX(le->window, th->th_win);
824 			le->append_cnt++;
825 		} else {
826 			/* no data and old ack */
827 			le->append_cnt++;
828 			m_freem(m);
829 			return (0);
830 		}
831 #ifdef TCP_LRO_UPDATE_CSUM
832 		le->ulp_csum += tcp_lro_rx_csum_fixup(le, l3hdr, th,
833 		    tcp_data_len, ~csum);
834 #endif
835 
836 		if (tcp_data_len == 0) {
837 			m_freem(m);
838 			/*
839 			 * Flush this LRO entry, if this ACK should not
840 			 * be further delayed.
841 			 */
842 			if (le->append_cnt >= lc->lro_ackcnt_lim) {
843 				tcp_lro_active_remove(le);
844 				tcp_lro_flush(lc, le);
845 			}
846 			return (0);
847 		}
848 
849 		le->p_len += tcp_data_len;
850 
851 		/*
852 		 * Adjust the mbuf so that m_data points to the first byte of
853 		 * the ULP payload.  Adjust the mbuf to avoid complications and
854 		 * append new segment to existing mbuf chain.
855 		 */
856 		m_adj(m, m->m_pkthdr.len - tcp_data_len);
857 		m_demote_pkthdr(m);
858 
859 		le->m_tail->m_next = m;
860 		le->m_tail = m_last(m);
861 
862 		/*
863 		 * If a possible next full length packet would cause an
864 		 * overflow, pro-actively flush now.
865 		 */
866 		if (le->p_len > (lc->lro_length_lim - lc->ifp->if_mtu)) {
867 			tcp_lro_active_remove(le);
868 			tcp_lro_flush(lc, le);
869 		} else
870 			getmicrotime(&le->mtime);
871 
872 		return (0);
873 	}
874 
875 	if (force_flush) {
876 		/*
877 		 * Nothing to flush, but this segment can not be further
878 		 * aggregated/delayed.
879 		 */
880 		return (TCP_LRO_CANNOT);
881 	}
882 
883 	/* Try to find an empty slot. */
884 	if (LIST_EMPTY(&lc->lro_free))
885 		return (TCP_LRO_NO_ENTRIES);
886 
887 	/* Start a new segment chain. */
888 	le = LIST_FIRST(&lc->lro_free);
889 	LIST_REMOVE(le, next);
890 	tcp_lro_active_insert(lc, bucket, le);
891 	getmicrotime(&le->mtime);
892 
893 	/* Start filling in details. */
894 	switch (eh_type) {
895 #ifdef INET6
896 	case ETHERTYPE_IPV6:
897 		le->le_ip6 = ip6;
898 		le->source_ip6 = ip6->ip6_src;
899 		le->dest_ip6 = ip6->ip6_dst;
900 		le->eh_type = eh_type;
901 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN - sizeof(*ip6);
902 		break;
903 #endif
904 #ifdef INET
905 	case ETHERTYPE_IP:
906 		le->le_ip4 = ip4;
907 		le->source_ip4 = ip4->ip_src.s_addr;
908 		le->dest_ip4 = ip4->ip_dst.s_addr;
909 		le->eh_type = eh_type;
910 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN;
911 		break;
912 #endif
913 	}
914 	le->source_port = th->th_sport;
915 	le->dest_port = th->th_dport;
916 
917 	le->next_seq = seq + tcp_data_len;
918 	le->ack_seq = th->th_ack;
919 	le->window = th->th_win;
920 	if (l != 0) {
921 		le->timestamp = 1;
922 		le->tsval = ntohl(*(ts_ptr + 1));
923 		le->tsecr = *(ts_ptr + 2);
924 	}
925 
926 #ifdef TCP_LRO_UPDATE_CSUM
927 	/*
928 	 * Do not touch the csum of the first packet.  However save the
929 	 * "adjusted" checksum of just the source and destination addresses,
930 	 * the next header and the TCP payload.  The length and TCP header
931 	 * parts may change, so we remove those from the saved checksum and
932 	 * re-add with final values on tcp_lro_flush() if needed.
933 	 */
934 	KASSERT(le->ulp_csum == 0, ("%s: le=%p le->ulp_csum=0x%04x\n",
935 	    __func__, le, le->ulp_csum));
936 
937 	le->ulp_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th, tcp_data_len,
938 	    ~csum);
939 	th->th_sum = csum;	/* Restore checksum on first packet. */
940 #endif
941 
942 	le->m_head = m;
943 	le->m_tail = m_last(m);
944 
945 	return (0);
946 }
947 
948 int
949 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
950 {
951 
952 	return tcp_lro_rx2(lc, m, csum, 1);
953 }
954 
955 void
956 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
957 {
958 	/* sanity checks */
959 	if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
960 	    lc->lro_mbuf_max == 0)) {
961 		/* packet drop */
962 		m_freem(mb);
963 		return;
964 	}
965 
966 	/* check if packet is not LRO capable */
967 	if (__predict_false(mb->m_pkthdr.csum_flags == 0 ||
968 	    (lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
969 
970 		/* input packet to network layer */
971 		(*lc->ifp->if_input) (lc->ifp, mb);
972 		return;
973 	}
974 
975 	/* create sequence number */
976 	lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
977 	    (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
978 	    (((uint64_t)mb->m_pkthdr.flowid) << 24) |
979 	    ((uint64_t)lc->lro_mbuf_count);
980 
981 	/* enter mbuf */
982 	lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb;
983 
984 	/* flush if array is full */
985 	if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max))
986 		tcp_lro_flush_all(lc);
987 }
988 
989 /* end */
990