xref: /freebsd/sys/netinet/tcp_lro.c (revision eb9da1ada8b6b2c74378a5c17029ec5a7fb199e6)
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 static int	tcp_lro_rx2(struct lro_ctrl *lc, struct mbuf *m,
72 		    uint32_t csum, int use_hash);
73 
74 static __inline void
75 tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket,
76     struct lro_entry *le)
77 {
78 
79 	LIST_INSERT_HEAD(&lc->lro_active, le, next);
80 	LIST_INSERT_HEAD(bucket, le, hash_next);
81 }
82 
83 static __inline void
84 tcp_lro_active_remove(struct lro_entry *le)
85 {
86 
87 	LIST_REMOVE(le, next);		/* active list */
88 	LIST_REMOVE(le, hash_next);	/* hash bucket */
89 }
90 
91 int
92 tcp_lro_init(struct lro_ctrl *lc)
93 {
94 	return (tcp_lro_init_args(lc, NULL, TCP_LRO_ENTRIES, 0));
95 }
96 
97 int
98 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
99     unsigned lro_entries, unsigned lro_mbufs)
100 {
101 	struct lro_entry *le;
102 	size_t size;
103 	unsigned i, elements;
104 
105 	lc->lro_bad_csum = 0;
106 	lc->lro_queued = 0;
107 	lc->lro_flushed = 0;
108 	lc->lro_cnt = 0;
109 	lc->lro_mbuf_count = 0;
110 	lc->lro_mbuf_max = lro_mbufs;
111 	lc->lro_cnt = lro_entries;
112 	lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
113 	lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
114 	lc->ifp = ifp;
115 	LIST_INIT(&lc->lro_free);
116 	LIST_INIT(&lc->lro_active);
117 
118 	/* create hash table to accelerate entry lookup */
119 	if (lro_entries > lro_mbufs)
120 		elements = lro_entries;
121 	else
122 		elements = lro_mbufs;
123 	lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz,
124 	    HASH_NOWAIT);
125 	if (lc->lro_hash == NULL) {
126 		memset(lc, 0, sizeof(*lc));
127 		return (ENOMEM);
128 	}
129 
130 	/* compute size to allocate */
131 	size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
132 	    (lro_entries * sizeof(*le));
133 	lc->lro_mbuf_data = (struct lro_mbuf_sort *)
134 	    malloc(size, M_LRO, M_NOWAIT | M_ZERO);
135 
136 	/* check for out of memory */
137 	if (lc->lro_mbuf_data == NULL) {
138 		memset(lc, 0, sizeof(*lc));
139 		return (ENOMEM);
140 	}
141 	/* compute offset for LRO entries */
142 	le = (struct lro_entry *)
143 	    (lc->lro_mbuf_data + lro_mbufs);
144 
145 	/* setup linked list */
146 	for (i = 0; i != lro_entries; i++)
147 		LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
148 
149 	return (0);
150 }
151 
152 void
153 tcp_lro_free(struct lro_ctrl *lc)
154 {
155 	struct lro_entry *le;
156 	unsigned x;
157 
158 	/* reset LRO free list */
159 	LIST_INIT(&lc->lro_free);
160 
161 	/* free active mbufs, if any */
162 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
163 		tcp_lro_active_remove(le);
164 		m_freem(le->m_head);
165 	}
166 
167 	/* free hash table */
168 	if (lc->lro_hash != NULL) {
169 		free(lc->lro_hash, M_LRO);
170 		lc->lro_hash = NULL;
171 	}
172 	lc->lro_hashsz = 0;
173 
174 	/* free mbuf array, if any */
175 	for (x = 0; x != lc->lro_mbuf_count; x++)
176 		m_freem(lc->lro_mbuf_data[x].mb);
177 	lc->lro_mbuf_count = 0;
178 
179 	/* free allocated memory, if any */
180 	free(lc->lro_mbuf_data, M_LRO);
181 	lc->lro_mbuf_data = NULL;
182 }
183 
184 #ifdef TCP_LRO_UPDATE_CSUM
185 static uint16_t
186 tcp_lro_csum_th(struct tcphdr *th)
187 {
188 	uint32_t ch;
189 	uint16_t *p, l;
190 
191 	ch = th->th_sum = 0x0000;
192 	l = th->th_off;
193 	p = (uint16_t *)th;
194 	while (l > 0) {
195 		ch += *p;
196 		p++;
197 		ch += *p;
198 		p++;
199 		l--;
200 	}
201 	while (ch > 0xffff)
202 		ch = (ch >> 16) + (ch & 0xffff);
203 
204 	return (ch & 0xffff);
205 }
206 
207 static uint16_t
208 tcp_lro_rx_csum_fixup(struct lro_entry *le, void *l3hdr, struct tcphdr *th,
209     uint16_t tcp_data_len, uint16_t csum)
210 {
211 	uint32_t c;
212 	uint16_t cs;
213 
214 	c = csum;
215 
216 	/* Remove length from checksum. */
217 	switch (le->eh_type) {
218 #ifdef INET6
219 	case ETHERTYPE_IPV6:
220 	{
221 		struct ip6_hdr *ip6;
222 
223 		ip6 = (struct ip6_hdr *)l3hdr;
224 		if (le->append_cnt == 0)
225 			cs = ip6->ip6_plen;
226 		else {
227 			uint32_t cx;
228 
229 			cx = ntohs(ip6->ip6_plen);
230 			cs = in6_cksum_pseudo(ip6, cx, ip6->ip6_nxt, 0);
231 		}
232 		break;
233 	}
234 #endif
235 #ifdef INET
236 	case ETHERTYPE_IP:
237 	{
238 		struct ip *ip4;
239 
240 		ip4 = (struct ip *)l3hdr;
241 		if (le->append_cnt == 0)
242 			cs = ip4->ip_len;
243 		else {
244 			cs = in_addword(ntohs(ip4->ip_len) - sizeof(*ip4),
245 			    IPPROTO_TCP);
246 			cs = in_pseudo(ip4->ip_src.s_addr, ip4->ip_dst.s_addr,
247 			    htons(cs));
248 		}
249 		break;
250 	}
251 #endif
252 	default:
253 		cs = 0;		/* Keep compiler happy. */
254 	}
255 
256 	cs = ~cs;
257 	c += cs;
258 
259 	/* Remove TCP header csum. */
260 	cs = ~tcp_lro_csum_th(th);
261 	c += cs;
262 	while (c > 0xffff)
263 		c = (c >> 16) + (c & 0xffff);
264 
265 	return (c & 0xffff);
266 }
267 #endif
268 
269 static void
270 tcp_lro_rx_done(struct lro_ctrl *lc)
271 {
272 	struct lro_entry *le;
273 
274 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
275 		tcp_lro_active_remove(le);
276 		tcp_lro_flush(lc, le);
277 	}
278 }
279 
280 void
281 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
282 {
283 	struct lro_entry *le, *le_tmp;
284 	struct timeval tv;
285 
286 	if (LIST_EMPTY(&lc->lro_active))
287 		return;
288 
289 	getmicrotime(&tv);
290 	timevalsub(&tv, timeout);
291 	LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
292 		if (timevalcmp(&tv, &le->mtime, >=)) {
293 			tcp_lro_active_remove(le);
294 			tcp_lro_flush(lc, le);
295 		}
296 	}
297 }
298 
299 void
300 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
301 {
302 
303 	if (le->append_cnt > 0) {
304 		struct tcphdr *th;
305 		uint16_t p_len;
306 
307 		p_len = htons(le->p_len);
308 		switch (le->eh_type) {
309 #ifdef INET6
310 		case ETHERTYPE_IPV6:
311 		{
312 			struct ip6_hdr *ip6;
313 
314 			ip6 = le->le_ip6;
315 			ip6->ip6_plen = p_len;
316 			th = (struct tcphdr *)(ip6 + 1);
317 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
318 			    CSUM_PSEUDO_HDR;
319 			le->p_len += ETHER_HDR_LEN + sizeof(*ip6);
320 			break;
321 		}
322 #endif
323 #ifdef INET
324 		case ETHERTYPE_IP:
325 		{
326 			struct ip *ip4;
327 #ifdef TCP_LRO_UPDATE_CSUM
328 			uint32_t cl;
329 			uint16_t c;
330 #endif
331 
332 			ip4 = le->le_ip4;
333 #ifdef TCP_LRO_UPDATE_CSUM
334 			/* Fix IP header checksum for new length. */
335 			c = ~ip4->ip_sum;
336 			cl = c;
337 			c = ~ip4->ip_len;
338 			cl += c + p_len;
339 			while (cl > 0xffff)
340 				cl = (cl >> 16) + (cl & 0xffff);
341 			c = cl;
342 			ip4->ip_sum = ~c;
343 #else
344 			ip4->ip_sum = TCP_LRO_INVALID_CSUM;
345 #endif
346 			ip4->ip_len = p_len;
347 			th = (struct tcphdr *)(ip4 + 1);
348 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
349 			    CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
350 			le->p_len += ETHER_HDR_LEN;
351 			break;
352 		}
353 #endif
354 		default:
355 			th = NULL;	/* Keep compiler happy. */
356 		}
357 		le->m_head->m_pkthdr.csum_data = 0xffff;
358 		le->m_head->m_pkthdr.len = le->p_len;
359 
360 		/* Incorporate the latest ACK into the TCP header. */
361 		th->th_ack = le->ack_seq;
362 		th->th_win = le->window;
363 		/* Incorporate latest timestamp into the TCP header. */
364 		if (le->timestamp != 0) {
365 			uint32_t *ts_ptr;
366 
367 			ts_ptr = (uint32_t *)(th + 1);
368 			ts_ptr[1] = htonl(le->tsval);
369 			ts_ptr[2] = le->tsecr;
370 		}
371 #ifdef TCP_LRO_UPDATE_CSUM
372 		/* Update the TCP header checksum. */
373 		le->ulp_csum += p_len;
374 		le->ulp_csum += tcp_lro_csum_th(th);
375 		while (le->ulp_csum > 0xffff)
376 			le->ulp_csum = (le->ulp_csum >> 16) +
377 			    (le->ulp_csum & 0xffff);
378 		th->th_sum = (le->ulp_csum & 0xffff);
379 		th->th_sum = ~th->th_sum;
380 #else
381 		th->th_sum = TCP_LRO_INVALID_CSUM;
382 #endif
383 	}
384 
385 	(*lc->ifp->if_input)(lc->ifp, le->m_head);
386 	lc->lro_queued += le->append_cnt + 1;
387 	lc->lro_flushed++;
388 	bzero(le, sizeof(*le));
389 	LIST_INSERT_HEAD(&lc->lro_free, le, next);
390 }
391 
392 #ifdef HAVE_INLINE_FLSLL
393 #define	tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
394 #else
395 static inline uint64_t
396 tcp_lro_msb_64(uint64_t x)
397 {
398 	x |= (x >> 1);
399 	x |= (x >> 2);
400 	x |= (x >> 4);
401 	x |= (x >> 8);
402 	x |= (x >> 16);
403 	x |= (x >> 32);
404 	return (x & ~(x >> 1));
405 }
406 #endif
407 
408 /*
409  * The tcp_lro_sort() routine is comparable to qsort(), except it has
410  * a worst case complexity limit of O(MIN(N,64)*N), where N is the
411  * number of elements to sort and 64 is the number of sequence bits
412  * available. The algorithm is bit-slicing the 64-bit sequence number,
413  * sorting one bit at a time from the most significant bit until the
414  * least significant one, skipping the constant bits. This is
415  * typically called a radix sort.
416  */
417 static void
418 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
419 {
420 	struct lro_mbuf_sort temp;
421 	uint64_t ones;
422 	uint64_t zeros;
423 	uint32_t x;
424 	uint32_t y;
425 
426 repeat:
427 	/* for small arrays insertion sort is faster */
428 	if (size <= 12) {
429 		for (x = 1; x < size; x++) {
430 			temp = parray[x];
431 			for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
432 				parray[y] = parray[y - 1];
433 			parray[y] = temp;
434 		}
435 		return;
436 	}
437 
438 	/* compute sequence bits which are constant */
439 	ones = 0;
440 	zeros = 0;
441 	for (x = 0; x != size; x++) {
442 		ones |= parray[x].seq;
443 		zeros |= ~parray[x].seq;
444 	}
445 
446 	/* compute bits which are not constant into "ones" */
447 	ones &= zeros;
448 	if (ones == 0)
449 		return;
450 
451 	/* pick the most significant bit which is not constant */
452 	ones = tcp_lro_msb_64(ones);
453 
454 	/*
455 	 * Move entries having cleared sequence bits to the beginning
456 	 * of the array:
457 	 */
458 	for (x = y = 0; y != size; y++) {
459 		/* skip set bits */
460 		if (parray[y].seq & ones)
461 			continue;
462 		/* swap entries */
463 		temp = parray[x];
464 		parray[x] = parray[y];
465 		parray[y] = temp;
466 		x++;
467 	}
468 
469 	KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
470 
471 	/* sort zeros */
472 	tcp_lro_sort(parray, x);
473 
474 	/* sort ones */
475 	parray += x;
476 	size -= x;
477 	goto repeat;
478 }
479 
480 void
481 tcp_lro_flush_all(struct lro_ctrl *lc)
482 {
483 	uint64_t seq;
484 	uint64_t nseq;
485 	unsigned x;
486 
487 	/* check if no mbufs to flush */
488 	if (lc->lro_mbuf_count == 0)
489 		goto done;
490 
491 	/* sort all mbufs according to stream */
492 	tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
493 
494 	/* input data into LRO engine, stream by stream */
495 	seq = 0;
496 	for (x = 0; x != lc->lro_mbuf_count; x++) {
497 		struct mbuf *mb;
498 
499 		/* get mbuf */
500 		mb = lc->lro_mbuf_data[x].mb;
501 
502 		/* get sequence number, masking away the packet index */
503 		nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
504 
505 		/* check for new stream */
506 		if (seq != nseq) {
507 			seq = nseq;
508 
509 			/* flush active streams */
510 			tcp_lro_rx_done(lc);
511 		}
512 
513 		/* add packet to LRO engine */
514 		if (tcp_lro_rx2(lc, mb, 0, 0) != 0) {
515 			/* input packet to network layer */
516 			(*lc->ifp->if_input)(lc->ifp, mb);
517 			lc->lro_queued++;
518 			lc->lro_flushed++;
519 		}
520 	}
521 done:
522 	/* flush active streams */
523 	tcp_lro_rx_done(lc);
524 
525 	lc->lro_mbuf_count = 0;
526 }
527 
528 #ifdef INET6
529 static int
530 tcp_lro_rx_ipv6(struct lro_ctrl *lc, struct mbuf *m, struct ip6_hdr *ip6,
531     struct tcphdr **th)
532 {
533 
534 	/* XXX-BZ we should check the flow-label. */
535 
536 	/* XXX-BZ We do not yet support ext. hdrs. */
537 	if (ip6->ip6_nxt != IPPROTO_TCP)
538 		return (TCP_LRO_NOT_SUPPORTED);
539 
540 	/* Find the TCP header. */
541 	*th = (struct tcphdr *)(ip6 + 1);
542 
543 	return (0);
544 }
545 #endif
546 
547 #ifdef INET
548 static int
549 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4,
550     struct tcphdr **th)
551 {
552 	int csum_flags;
553 	uint16_t csum;
554 
555 	if (ip4->ip_p != IPPROTO_TCP)
556 		return (TCP_LRO_NOT_SUPPORTED);
557 
558 	/* Ensure there are no options. */
559 	if ((ip4->ip_hl << 2) != sizeof (*ip4))
560 		return (TCP_LRO_CANNOT);
561 
562 	/* .. and the packet is not fragmented. */
563 	if (ip4->ip_off & htons(IP_MF|IP_OFFMASK))
564 		return (TCP_LRO_CANNOT);
565 
566 	/* Legacy IP has a header checksum that needs to be correct. */
567 	csum_flags = m->m_pkthdr.csum_flags;
568 	if (csum_flags & CSUM_IP_CHECKED) {
569 		if (__predict_false((csum_flags & CSUM_IP_VALID) == 0)) {
570 			lc->lro_bad_csum++;
571 			return (TCP_LRO_CANNOT);
572 		}
573 	} else {
574 		csum = in_cksum_hdr(ip4);
575 		if (__predict_false((csum) != 0)) {
576 			lc->lro_bad_csum++;
577 			return (TCP_LRO_CANNOT);
578 		}
579 	}
580 
581 	/* Find the TCP header (we assured there are no IP options). */
582 	*th = (struct tcphdr *)(ip4 + 1);
583 
584 	return (0);
585 }
586 #endif
587 
588 static int
589 tcp_lro_rx2(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, int use_hash)
590 {
591 	struct lro_entry *le;
592 	struct ether_header *eh;
593 #ifdef INET6
594 	struct ip6_hdr *ip6 = NULL;	/* Keep compiler happy. */
595 #endif
596 #ifdef INET
597 	struct ip *ip4 = NULL;		/* Keep compiler happy. */
598 #endif
599 	struct tcphdr *th;
600 	void *l3hdr = NULL;		/* Keep compiler happy. */
601 	uint32_t *ts_ptr;
602 	tcp_seq seq;
603 	int error, ip_len, l;
604 	uint16_t eh_type, tcp_data_len;
605 	struct lro_head *bucket;
606 	int force_flush = 0;
607 
608 	/* We expect a contiguous header [eh, ip, tcp]. */
609 
610 	eh = mtod(m, struct ether_header *);
611 	eh_type = ntohs(eh->ether_type);
612 	switch (eh_type) {
613 #ifdef INET6
614 	case ETHERTYPE_IPV6:
615 	{
616 		CURVNET_SET(lc->ifp->if_vnet);
617 		if (V_ip6_forwarding != 0) {
618 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
619 			CURVNET_RESTORE();
620 			return (TCP_LRO_CANNOT);
621 		}
622 		CURVNET_RESTORE();
623 		l3hdr = ip6 = (struct ip6_hdr *)(eh + 1);
624 		error = tcp_lro_rx_ipv6(lc, m, ip6, &th);
625 		if (error != 0)
626 			return (error);
627 		tcp_data_len = ntohs(ip6->ip6_plen);
628 		ip_len = sizeof(*ip6) + tcp_data_len;
629 		break;
630 	}
631 #endif
632 #ifdef INET
633 	case ETHERTYPE_IP:
634 	{
635 		CURVNET_SET(lc->ifp->if_vnet);
636 		if (V_ipforwarding != 0) {
637 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
638 			CURVNET_RESTORE();
639 			return (TCP_LRO_CANNOT);
640 		}
641 		CURVNET_RESTORE();
642 		l3hdr = ip4 = (struct ip *)(eh + 1);
643 		error = tcp_lro_rx_ipv4(lc, m, ip4, &th);
644 		if (error != 0)
645 			return (error);
646 		ip_len = ntohs(ip4->ip_len);
647 		tcp_data_len = ip_len - sizeof(*ip4);
648 		break;
649 	}
650 #endif
651 	/* XXX-BZ what happens in case of VLAN(s)? */
652 	default:
653 		return (TCP_LRO_NOT_SUPPORTED);
654 	}
655 
656 	/*
657 	 * If the frame is padded beyond the end of the IP packet, then we must
658 	 * trim the extra bytes off.
659 	 */
660 	l = m->m_pkthdr.len - (ETHER_HDR_LEN + ip_len);
661 	if (l != 0) {
662 		if (l < 0)
663 			/* Truncated packet. */
664 			return (TCP_LRO_CANNOT);
665 
666 		m_adj(m, -l);
667 	}
668 
669 	/*
670 	 * Check TCP header constraints.
671 	 */
672 	/* Ensure no bits set besides ACK or PSH. */
673 	if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0) {
674 		if (th->th_flags & TH_SYN)
675 			return (TCP_LRO_CANNOT);
676 		/*
677 		 * Make sure that previously seen segements/ACKs are delivered
678 		 * before this segement, e.g. FIN.
679 		 */
680 		force_flush = 1;
681 	}
682 
683 	/* XXX-BZ We lose a ACK|PUSH flag concatenating multiple segments. */
684 	/* XXX-BZ Ideally we'd flush on PUSH? */
685 
686 	/*
687 	 * Check for timestamps.
688 	 * Since the only option we handle are timestamps, we only have to
689 	 * handle the simple case of aligned timestamps.
690 	 */
691 	l = (th->th_off << 2);
692 	tcp_data_len -= l;
693 	l -= sizeof(*th);
694 	ts_ptr = (uint32_t *)(th + 1);
695 	if (l != 0 && (__predict_false(l != TCPOLEN_TSTAMP_APPA) ||
696 	    (*ts_ptr != ntohl(TCPOPT_NOP<<24|TCPOPT_NOP<<16|
697 	    TCPOPT_TIMESTAMP<<8|TCPOLEN_TIMESTAMP)))) {
698 		/*
699 		 * Make sure that previously seen segements/ACKs are delivered
700 		 * before this segement.
701 		 */
702 		force_flush = 1;
703 	}
704 
705 	/* If the driver did not pass in the checksum, set it now. */
706 	if (csum == 0x0000)
707 		csum = th->th_sum;
708 
709 	seq = ntohl(th->th_seq);
710 
711 	if (!use_hash) {
712 		bucket = &lc->lro_hash[0];
713 	} else if (M_HASHTYPE_ISHASH(m)) {
714 		bucket = &lc->lro_hash[m->m_pkthdr.flowid % lc->lro_hashsz];
715 	} else {
716 		uint32_t hash;
717 
718 		switch (eh_type) {
719 #ifdef INET
720 		case ETHERTYPE_IP:
721 			hash = ip4->ip_src.s_addr + ip4->ip_dst.s_addr;
722 			break;
723 #endif
724 #ifdef INET6
725 		case ETHERTYPE_IPV6:
726 			hash = ip6->ip6_src.s6_addr32[0] +
727 			    ip6->ip6_dst.s6_addr32[0];
728 			hash += ip6->ip6_src.s6_addr32[1] +
729 			    ip6->ip6_dst.s6_addr32[1];
730 			hash += ip6->ip6_src.s6_addr32[2] +
731 			    ip6->ip6_dst.s6_addr32[2];
732 			hash += ip6->ip6_src.s6_addr32[3] +
733 			    ip6->ip6_dst.s6_addr32[3];
734 			break;
735 #endif
736 		default:
737 			hash = 0;
738 			break;
739 		}
740 		hash += th->th_sport + th->th_dport;
741 		bucket = &lc->lro_hash[hash % lc->lro_hashsz];
742 	}
743 
744 	/* Try to find a matching previous segment. */
745 	LIST_FOREACH(le, bucket, hash_next) {
746 		if (le->eh_type != eh_type)
747 			continue;
748 		if (le->source_port != th->th_sport ||
749 		    le->dest_port != th->th_dport)
750 			continue;
751 		switch (eh_type) {
752 #ifdef INET6
753 		case ETHERTYPE_IPV6:
754 			if (bcmp(&le->source_ip6, &ip6->ip6_src,
755 			    sizeof(struct in6_addr)) != 0 ||
756 			    bcmp(&le->dest_ip6, &ip6->ip6_dst,
757 			    sizeof(struct in6_addr)) != 0)
758 				continue;
759 			break;
760 #endif
761 #ifdef INET
762 		case ETHERTYPE_IP:
763 			if (le->source_ip4 != ip4->ip_src.s_addr ||
764 			    le->dest_ip4 != ip4->ip_dst.s_addr)
765 				continue;
766 			break;
767 #endif
768 		}
769 
770 		if (force_flush) {
771 			/* Timestamps mismatch; this is a FIN, etc */
772 			tcp_lro_active_remove(le);
773 			tcp_lro_flush(lc, le);
774 			return (TCP_LRO_CANNOT);
775 		}
776 
777 		/* Flush now if appending will result in overflow. */
778 		if (le->p_len > (lc->lro_length_lim - tcp_data_len)) {
779 			tcp_lro_active_remove(le);
780 			tcp_lro_flush(lc, le);
781 			break;
782 		}
783 
784 		/* Try to append the new segment. */
785 		if (__predict_false(seq != le->next_seq ||
786 		    (tcp_data_len == 0 && le->ack_seq == th->th_ack))) {
787 			/* Out of order packet or duplicate ACK. */
788 			tcp_lro_active_remove(le);
789 			tcp_lro_flush(lc, le);
790 			return (TCP_LRO_CANNOT);
791 		}
792 
793 		if (l != 0) {
794 			uint32_t tsval = ntohl(*(ts_ptr + 1));
795 			/* Make sure timestamp values are increasing. */
796 			/* XXX-BZ flip and use TSTMP_GEQ macro for this? */
797 			if (__predict_false(le->tsval > tsval ||
798 			    *(ts_ptr + 2) == 0))
799 				return (TCP_LRO_CANNOT);
800 			le->tsval = tsval;
801 			le->tsecr = *(ts_ptr + 2);
802 		}
803 
804 		le->next_seq += tcp_data_len;
805 		le->ack_seq = th->th_ack;
806 		le->window = th->th_win;
807 		le->append_cnt++;
808 
809 #ifdef TCP_LRO_UPDATE_CSUM
810 		le->ulp_csum += tcp_lro_rx_csum_fixup(le, l3hdr, th,
811 		    tcp_data_len, ~csum);
812 #endif
813 
814 		if (tcp_data_len == 0) {
815 			m_freem(m);
816 			/*
817 			 * Flush this LRO entry, if this ACK should not
818 			 * be further delayed.
819 			 */
820 			if (le->append_cnt >= lc->lro_ackcnt_lim) {
821 				tcp_lro_active_remove(le);
822 				tcp_lro_flush(lc, le);
823 			}
824 			return (0);
825 		}
826 
827 		le->p_len += tcp_data_len;
828 
829 		/*
830 		 * Adjust the mbuf so that m_data points to the first byte of
831 		 * the ULP payload.  Adjust the mbuf to avoid complications and
832 		 * append new segment to existing mbuf chain.
833 		 */
834 		m_adj(m, m->m_pkthdr.len - tcp_data_len);
835 		m_demote_pkthdr(m);
836 
837 		le->m_tail->m_next = m;
838 		le->m_tail = m_last(m);
839 
840 		/*
841 		 * If a possible next full length packet would cause an
842 		 * overflow, pro-actively flush now.
843 		 */
844 		if (le->p_len > (lc->lro_length_lim - lc->ifp->if_mtu)) {
845 			tcp_lro_active_remove(le);
846 			tcp_lro_flush(lc, le);
847 		} else
848 			getmicrotime(&le->mtime);
849 
850 		return (0);
851 	}
852 
853 	if (force_flush) {
854 		/*
855 		 * Nothing to flush, but this segment can not be further
856 		 * aggregated/delayed.
857 		 */
858 		return (TCP_LRO_CANNOT);
859 	}
860 
861 	/* Try to find an empty slot. */
862 	if (LIST_EMPTY(&lc->lro_free))
863 		return (TCP_LRO_NO_ENTRIES);
864 
865 	/* Start a new segment chain. */
866 	le = LIST_FIRST(&lc->lro_free);
867 	LIST_REMOVE(le, next);
868 	tcp_lro_active_insert(lc, bucket, le);
869 	getmicrotime(&le->mtime);
870 
871 	/* Start filling in details. */
872 	switch (eh_type) {
873 #ifdef INET6
874 	case ETHERTYPE_IPV6:
875 		le->le_ip6 = ip6;
876 		le->source_ip6 = ip6->ip6_src;
877 		le->dest_ip6 = ip6->ip6_dst;
878 		le->eh_type = eh_type;
879 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN - sizeof(*ip6);
880 		break;
881 #endif
882 #ifdef INET
883 	case ETHERTYPE_IP:
884 		le->le_ip4 = ip4;
885 		le->source_ip4 = ip4->ip_src.s_addr;
886 		le->dest_ip4 = ip4->ip_dst.s_addr;
887 		le->eh_type = eh_type;
888 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN;
889 		break;
890 #endif
891 	}
892 	le->source_port = th->th_sport;
893 	le->dest_port = th->th_dport;
894 
895 	le->next_seq = seq + tcp_data_len;
896 	le->ack_seq = th->th_ack;
897 	le->window = th->th_win;
898 	if (l != 0) {
899 		le->timestamp = 1;
900 		le->tsval = ntohl(*(ts_ptr + 1));
901 		le->tsecr = *(ts_ptr + 2);
902 	}
903 
904 #ifdef TCP_LRO_UPDATE_CSUM
905 	/*
906 	 * Do not touch the csum of the first packet.  However save the
907 	 * "adjusted" checksum of just the source and destination addresses,
908 	 * the next header and the TCP payload.  The length and TCP header
909 	 * parts may change, so we remove those from the saved checksum and
910 	 * re-add with final values on tcp_lro_flush() if needed.
911 	 */
912 	KASSERT(le->ulp_csum == 0, ("%s: le=%p le->ulp_csum=0x%04x\n",
913 	    __func__, le, le->ulp_csum));
914 
915 	le->ulp_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th, tcp_data_len,
916 	    ~csum);
917 	th->th_sum = csum;	/* Restore checksum on first packet. */
918 #endif
919 
920 	le->m_head = m;
921 	le->m_tail = m_last(m);
922 
923 	return (0);
924 }
925 
926 int
927 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
928 {
929 
930 	return tcp_lro_rx2(lc, m, csum, 1);
931 }
932 
933 void
934 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
935 {
936 	/* sanity checks */
937 	if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
938 	    lc->lro_mbuf_max == 0)) {
939 		/* packet drop */
940 		m_freem(mb);
941 		return;
942 	}
943 
944 	/* check if packet is not LRO capable */
945 	if (__predict_false(mb->m_pkthdr.csum_flags == 0 ||
946 	    (lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
947 		lc->lro_flushed++;
948 		lc->lro_queued++;
949 
950 		/* input packet to network layer */
951 		(*lc->ifp->if_input) (lc->ifp, mb);
952 		return;
953 	}
954 
955 	/* check if array is full */
956 	if (__predict_false(lc->lro_mbuf_count == lc->lro_mbuf_max))
957 		tcp_lro_flush_all(lc);
958 
959 	/* create sequence number */
960 	lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
961 	    (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
962 	    (((uint64_t)mb->m_pkthdr.flowid) << 24) |
963 	    ((uint64_t)lc->lro_mbuf_count);
964 
965 	/* enter mbuf */
966 	lc->lro_mbuf_data[lc->lro_mbuf_count++].mb = mb;
967 }
968 
969 /* end */
970