xref: /freebsd/sys/netinet/tcp_lro.c (revision 41059135ce931c0f1014a999ffabc6bc470ce856)
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_mbuf_count = 0;
119 	lc->lro_mbuf_max = lro_mbufs;
120 	lc->lro_cnt = lro_entries;
121 	lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
122 	lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
123 	lc->ifp = ifp;
124 	LIST_INIT(&lc->lro_free);
125 	LIST_INIT(&lc->lro_active);
126 
127 	/* create hash table to accelerate entry lookup */
128 	if (lro_entries > lro_mbufs)
129 		elements = lro_entries;
130 	else
131 		elements = lro_mbufs;
132 	lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz,
133 	    HASH_NOWAIT);
134 	if (lc->lro_hash == NULL) {
135 		memset(lc, 0, sizeof(*lc));
136 		return (ENOMEM);
137 	}
138 
139 	/* compute size to allocate */
140 	size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
141 	    (lro_entries * sizeof(*le));
142 	lc->lro_mbuf_data = (struct lro_mbuf_sort *)
143 	    malloc(size, M_LRO, M_NOWAIT | M_ZERO);
144 
145 	/* check for out of memory */
146 	if (lc->lro_mbuf_data == NULL) {
147 		free(lc->lro_hash, M_LRO);
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 	free(lc->lro_hash, M_LRO);
179 	lc->lro_hash = NULL;
180 	lc->lro_hashsz = 0;
181 
182 	/* free mbuf array, if any */
183 	for (x = 0; x != lc->lro_mbuf_count; x++)
184 		m_freem(lc->lro_mbuf_data[x].mb);
185 	lc->lro_mbuf_count = 0;
186 
187 	/* free allocated memory, if any */
188 	free(lc->lro_mbuf_data, M_LRO);
189 	lc->lro_mbuf_data = NULL;
190 }
191 
192 #ifdef TCP_LRO_UPDATE_CSUM
193 static uint16_t
194 tcp_lro_csum_th(struct tcphdr *th)
195 {
196 	uint32_t ch;
197 	uint16_t *p, l;
198 
199 	ch = th->th_sum = 0x0000;
200 	l = th->th_off;
201 	p = (uint16_t *)th;
202 	while (l > 0) {
203 		ch += *p;
204 		p++;
205 		ch += *p;
206 		p++;
207 		l--;
208 	}
209 	while (ch > 0xffff)
210 		ch = (ch >> 16) + (ch & 0xffff);
211 
212 	return (ch & 0xffff);
213 }
214 
215 static uint16_t
216 tcp_lro_rx_csum_fixup(struct lro_entry *le, void *l3hdr, struct tcphdr *th,
217     uint16_t tcp_data_len, uint16_t csum)
218 {
219 	uint32_t c;
220 	uint16_t cs;
221 
222 	c = csum;
223 
224 	/* Remove length from checksum. */
225 	switch (le->eh_type) {
226 #ifdef INET6
227 	case ETHERTYPE_IPV6:
228 	{
229 		struct ip6_hdr *ip6;
230 
231 		ip6 = (struct ip6_hdr *)l3hdr;
232 		if (le->append_cnt == 0)
233 			cs = ip6->ip6_plen;
234 		else {
235 			uint32_t cx;
236 
237 			cx = ntohs(ip6->ip6_plen);
238 			cs = in6_cksum_pseudo(ip6, cx, ip6->ip6_nxt, 0);
239 		}
240 		break;
241 	}
242 #endif
243 #ifdef INET
244 	case ETHERTYPE_IP:
245 	{
246 		struct ip *ip4;
247 
248 		ip4 = (struct ip *)l3hdr;
249 		if (le->append_cnt == 0)
250 			cs = ip4->ip_len;
251 		else {
252 			cs = in_addword(ntohs(ip4->ip_len) - sizeof(*ip4),
253 			    IPPROTO_TCP);
254 			cs = in_pseudo(ip4->ip_src.s_addr, ip4->ip_dst.s_addr,
255 			    htons(cs));
256 		}
257 		break;
258 	}
259 #endif
260 	default:
261 		cs = 0;		/* Keep compiler happy. */
262 	}
263 
264 	cs = ~cs;
265 	c += cs;
266 
267 	/* Remove TCP header csum. */
268 	cs = ~tcp_lro_csum_th(th);
269 	c += cs;
270 	while (c > 0xffff)
271 		c = (c >> 16) + (c & 0xffff);
272 
273 	return (c & 0xffff);
274 }
275 #endif
276 
277 static void
278 tcp_lro_rx_done(struct lro_ctrl *lc)
279 {
280 	struct lro_entry *le;
281 
282 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
283 		tcp_lro_active_remove(le);
284 		tcp_lro_flush(lc, le);
285 	}
286 }
287 
288 void
289 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
290 {
291 	struct lro_entry *le, *le_tmp;
292 	struct timeval tv;
293 
294 	if (LIST_EMPTY(&lc->lro_active))
295 		return;
296 
297 	getmicrotime(&tv);
298 	timevalsub(&tv, timeout);
299 	LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
300 		if (timevalcmp(&tv, &le->mtime, >=)) {
301 			tcp_lro_active_remove(le);
302 			tcp_lro_flush(lc, le);
303 		}
304 	}
305 }
306 
307 void
308 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
309 {
310 
311 	if (le->append_cnt > 0) {
312 		struct tcphdr *th;
313 		uint16_t p_len;
314 
315 		p_len = htons(le->p_len);
316 		switch (le->eh_type) {
317 #ifdef INET6
318 		case ETHERTYPE_IPV6:
319 		{
320 			struct ip6_hdr *ip6;
321 
322 			ip6 = le->le_ip6;
323 			ip6->ip6_plen = p_len;
324 			th = (struct tcphdr *)(ip6 + 1);
325 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
326 			    CSUM_PSEUDO_HDR;
327 			le->p_len += ETHER_HDR_LEN + sizeof(*ip6);
328 			break;
329 		}
330 #endif
331 #ifdef INET
332 		case ETHERTYPE_IP:
333 		{
334 			struct ip *ip4;
335 #ifdef TCP_LRO_UPDATE_CSUM
336 			uint32_t cl;
337 			uint16_t c;
338 #endif
339 
340 			ip4 = le->le_ip4;
341 #ifdef TCP_LRO_UPDATE_CSUM
342 			/* Fix IP header checksum for new length. */
343 			c = ~ip4->ip_sum;
344 			cl = c;
345 			c = ~ip4->ip_len;
346 			cl += c + p_len;
347 			while (cl > 0xffff)
348 				cl = (cl >> 16) + (cl & 0xffff);
349 			c = cl;
350 			ip4->ip_sum = ~c;
351 #else
352 			ip4->ip_sum = TCP_LRO_INVALID_CSUM;
353 #endif
354 			ip4->ip_len = p_len;
355 			th = (struct tcphdr *)(ip4 + 1);
356 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
357 			    CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
358 			le->p_len += ETHER_HDR_LEN;
359 			break;
360 		}
361 #endif
362 		default:
363 			th = NULL;	/* Keep compiler happy. */
364 		}
365 		le->m_head->m_pkthdr.csum_data = 0xffff;
366 		le->m_head->m_pkthdr.len = le->p_len;
367 
368 		/* Incorporate the latest ACK into the TCP header. */
369 		th->th_ack = le->ack_seq;
370 		th->th_win = le->window;
371 		/* Incorporate latest timestamp into the TCP header. */
372 		if (le->timestamp != 0) {
373 			uint32_t *ts_ptr;
374 
375 			ts_ptr = (uint32_t *)(th + 1);
376 			ts_ptr[1] = htonl(le->tsval);
377 			ts_ptr[2] = le->tsecr;
378 		}
379 #ifdef TCP_LRO_UPDATE_CSUM
380 		/* Update the TCP header checksum. */
381 		le->ulp_csum += p_len;
382 		le->ulp_csum += tcp_lro_csum_th(th);
383 		while (le->ulp_csum > 0xffff)
384 			le->ulp_csum = (le->ulp_csum >> 16) +
385 			    (le->ulp_csum & 0xffff);
386 		th->th_sum = (le->ulp_csum & 0xffff);
387 		th->th_sum = ~th->th_sum;
388 #else
389 		th->th_sum = TCP_LRO_INVALID_CSUM;
390 #endif
391 	}
392 
393 	le->m_head->m_pkthdr.lro_nsegs = le->append_cnt + 1;
394 	(*lc->ifp->if_input)(lc->ifp, le->m_head);
395 	lc->lro_queued += le->append_cnt + 1;
396 	lc->lro_flushed++;
397 	bzero(le, sizeof(*le));
398 	LIST_INSERT_HEAD(&lc->lro_free, le, next);
399 }
400 
401 #ifdef HAVE_INLINE_FLSLL
402 #define	tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
403 #else
404 static inline uint64_t
405 tcp_lro_msb_64(uint64_t x)
406 {
407 	x |= (x >> 1);
408 	x |= (x >> 2);
409 	x |= (x >> 4);
410 	x |= (x >> 8);
411 	x |= (x >> 16);
412 	x |= (x >> 32);
413 	return (x & ~(x >> 1));
414 }
415 #endif
416 
417 /*
418  * The tcp_lro_sort() routine is comparable to qsort(), except it has
419  * a worst case complexity limit of O(MIN(N,64)*N), where N is the
420  * number of elements to sort and 64 is the number of sequence bits
421  * available. The algorithm is bit-slicing the 64-bit sequence number,
422  * sorting one bit at a time from the most significant bit until the
423  * least significant one, skipping the constant bits. This is
424  * typically called a radix sort.
425  */
426 static void
427 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
428 {
429 	struct lro_mbuf_sort temp;
430 	uint64_t ones;
431 	uint64_t zeros;
432 	uint32_t x;
433 	uint32_t y;
434 
435 repeat:
436 	/* for small arrays insertion sort is faster */
437 	if (size <= 12) {
438 		for (x = 1; x < size; x++) {
439 			temp = parray[x];
440 			for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
441 				parray[y] = parray[y - 1];
442 			parray[y] = temp;
443 		}
444 		return;
445 	}
446 
447 	/* compute sequence bits which are constant */
448 	ones = 0;
449 	zeros = 0;
450 	for (x = 0; x != size; x++) {
451 		ones |= parray[x].seq;
452 		zeros |= ~parray[x].seq;
453 	}
454 
455 	/* compute bits which are not constant into "ones" */
456 	ones &= zeros;
457 	if (ones == 0)
458 		return;
459 
460 	/* pick the most significant bit which is not constant */
461 	ones = tcp_lro_msb_64(ones);
462 
463 	/*
464 	 * Move entries having cleared sequence bits to the beginning
465 	 * of the array:
466 	 */
467 	for (x = y = 0; y != size; y++) {
468 		/* skip set bits */
469 		if (parray[y].seq & ones)
470 			continue;
471 		/* swap entries */
472 		temp = parray[x];
473 		parray[x] = parray[y];
474 		parray[y] = temp;
475 		x++;
476 	}
477 
478 	KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
479 
480 	/* sort zeros */
481 	tcp_lro_sort(parray, x);
482 
483 	/* sort ones */
484 	parray += x;
485 	size -= x;
486 	goto repeat;
487 }
488 
489 void
490 tcp_lro_flush_all(struct lro_ctrl *lc)
491 {
492 	uint64_t seq;
493 	uint64_t nseq;
494 	unsigned x;
495 
496 	/* check if no mbufs to flush */
497 	if (lc->lro_mbuf_count == 0)
498 		goto done;
499 
500 	/* sort all mbufs according to stream */
501 	tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
502 
503 	/* input data into LRO engine, stream by stream */
504 	seq = 0;
505 	for (x = 0; x != lc->lro_mbuf_count; x++) {
506 		struct mbuf *mb;
507 
508 		/* get mbuf */
509 		mb = lc->lro_mbuf_data[x].mb;
510 
511 		/* get sequence number, masking away the packet index */
512 		nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
513 
514 		/* check for new stream */
515 		if (seq != nseq) {
516 			seq = nseq;
517 
518 			/* flush active streams */
519 			tcp_lro_rx_done(lc);
520 		}
521 
522 		/* add packet to LRO engine */
523 		if (tcp_lro_rx2(lc, mb, 0, 0) != 0) {
524 			/* input packet to network layer */
525 			(*lc->ifp->if_input)(lc->ifp, mb);
526 			lc->lro_queued++;
527 			lc->lro_flushed++;
528 		}
529 	}
530 done:
531 	/* flush active streams */
532 	tcp_lro_rx_done(lc);
533 
534 	lc->lro_mbuf_count = 0;
535 }
536 
537 #ifdef INET6
538 static int
539 tcp_lro_rx_ipv6(struct lro_ctrl *lc, struct mbuf *m, struct ip6_hdr *ip6,
540     struct tcphdr **th)
541 {
542 
543 	/* XXX-BZ we should check the flow-label. */
544 
545 	/* XXX-BZ We do not yet support ext. hdrs. */
546 	if (ip6->ip6_nxt != IPPROTO_TCP)
547 		return (TCP_LRO_NOT_SUPPORTED);
548 
549 	/* Find the TCP header. */
550 	*th = (struct tcphdr *)(ip6 + 1);
551 
552 	return (0);
553 }
554 #endif
555 
556 #ifdef INET
557 static int
558 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4,
559     struct tcphdr **th)
560 {
561 	int csum_flags;
562 	uint16_t csum;
563 
564 	if (ip4->ip_p != IPPROTO_TCP)
565 		return (TCP_LRO_NOT_SUPPORTED);
566 
567 	/* Ensure there are no options. */
568 	if ((ip4->ip_hl << 2) != sizeof (*ip4))
569 		return (TCP_LRO_CANNOT);
570 
571 	/* .. and the packet is not fragmented. */
572 	if (ip4->ip_off & htons(IP_MF|IP_OFFMASK))
573 		return (TCP_LRO_CANNOT);
574 
575 	/* Legacy IP has a header checksum that needs to be correct. */
576 	csum_flags = m->m_pkthdr.csum_flags;
577 	if (csum_flags & CSUM_IP_CHECKED) {
578 		if (__predict_false((csum_flags & CSUM_IP_VALID) == 0)) {
579 			lc->lro_bad_csum++;
580 			return (TCP_LRO_CANNOT);
581 		}
582 	} else {
583 		csum = in_cksum_hdr(ip4);
584 		if (__predict_false((csum) != 0)) {
585 			lc->lro_bad_csum++;
586 			return (TCP_LRO_CANNOT);
587 		}
588 	}
589 
590 	/* Find the TCP header (we assured there are no IP options). */
591 	*th = (struct tcphdr *)(ip4 + 1);
592 
593 	return (0);
594 }
595 #endif
596 
597 static int
598 tcp_lro_rx2(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, int use_hash)
599 {
600 	struct lro_entry *le;
601 	struct ether_header *eh;
602 #ifdef INET6
603 	struct ip6_hdr *ip6 = NULL;	/* Keep compiler happy. */
604 #endif
605 #ifdef INET
606 	struct ip *ip4 = NULL;		/* Keep compiler happy. */
607 #endif
608 	struct tcphdr *th;
609 	void *l3hdr = NULL;		/* Keep compiler happy. */
610 	uint32_t *ts_ptr;
611 	tcp_seq seq;
612 	int error, ip_len, l;
613 	uint16_t eh_type, tcp_data_len;
614 	struct lro_head *bucket;
615 	int force_flush = 0;
616 
617 	/* We expect a contiguous header [eh, ip, tcp]. */
618 
619 	eh = mtod(m, struct ether_header *);
620 	eh_type = ntohs(eh->ether_type);
621 	switch (eh_type) {
622 #ifdef INET6
623 	case ETHERTYPE_IPV6:
624 	{
625 		CURVNET_SET(lc->ifp->if_vnet);
626 		if (V_ip6_forwarding != 0) {
627 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
628 			CURVNET_RESTORE();
629 			return (TCP_LRO_CANNOT);
630 		}
631 		CURVNET_RESTORE();
632 		l3hdr = ip6 = (struct ip6_hdr *)(eh + 1);
633 		error = tcp_lro_rx_ipv6(lc, m, ip6, &th);
634 		if (error != 0)
635 			return (error);
636 		tcp_data_len = ntohs(ip6->ip6_plen);
637 		ip_len = sizeof(*ip6) + tcp_data_len;
638 		break;
639 	}
640 #endif
641 #ifdef INET
642 	case ETHERTYPE_IP:
643 	{
644 		CURVNET_SET(lc->ifp->if_vnet);
645 		if (V_ipforwarding != 0) {
646 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
647 			CURVNET_RESTORE();
648 			return (TCP_LRO_CANNOT);
649 		}
650 		CURVNET_RESTORE();
651 		l3hdr = ip4 = (struct ip *)(eh + 1);
652 		error = tcp_lro_rx_ipv4(lc, m, ip4, &th);
653 		if (error != 0)
654 			return (error);
655 		ip_len = ntohs(ip4->ip_len);
656 		tcp_data_len = ip_len - sizeof(*ip4);
657 		break;
658 	}
659 #endif
660 	/* XXX-BZ what happens in case of VLAN(s)? */
661 	default:
662 		return (TCP_LRO_NOT_SUPPORTED);
663 	}
664 
665 	/*
666 	 * If the frame is padded beyond the end of the IP packet, then we must
667 	 * trim the extra bytes off.
668 	 */
669 	l = m->m_pkthdr.len - (ETHER_HDR_LEN + ip_len);
670 	if (l != 0) {
671 		if (l < 0)
672 			/* Truncated packet. */
673 			return (TCP_LRO_CANNOT);
674 
675 		m_adj(m, -l);
676 	}
677 
678 	/*
679 	 * Check TCP header constraints.
680 	 */
681 	/* Ensure no bits set besides ACK or PSH. */
682 	if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0) {
683 		if (th->th_flags & TH_SYN)
684 			return (TCP_LRO_CANNOT);
685 		/*
686 		 * Make sure that previously seen segements/ACKs are delivered
687 		 * before this segement, e.g. FIN.
688 		 */
689 		force_flush = 1;
690 	}
691 
692 	/* XXX-BZ We lose a ACK|PUSH flag concatenating multiple segments. */
693 	/* XXX-BZ Ideally we'd flush on PUSH? */
694 
695 	/*
696 	 * Check for timestamps.
697 	 * Since the only option we handle are timestamps, we only have to
698 	 * handle the simple case of aligned timestamps.
699 	 */
700 	l = (th->th_off << 2);
701 	tcp_data_len -= l;
702 	l -= sizeof(*th);
703 	ts_ptr = (uint32_t *)(th + 1);
704 	if (l != 0 && (__predict_false(l != TCPOLEN_TSTAMP_APPA) ||
705 	    (*ts_ptr != ntohl(TCPOPT_NOP<<24|TCPOPT_NOP<<16|
706 	    TCPOPT_TIMESTAMP<<8|TCPOLEN_TIMESTAMP)))) {
707 		/*
708 		 * Make sure that previously seen segements/ACKs are delivered
709 		 * before this segement.
710 		 */
711 		force_flush = 1;
712 	}
713 
714 	/* If the driver did not pass in the checksum, set it now. */
715 	if (csum == 0x0000)
716 		csum = th->th_sum;
717 
718 	seq = ntohl(th->th_seq);
719 
720 	if (!use_hash) {
721 		bucket = &lc->lro_hash[0];
722 	} else if (M_HASHTYPE_ISHASH(m)) {
723 		bucket = &lc->lro_hash[m->m_pkthdr.flowid % lc->lro_hashsz];
724 	} else {
725 		uint32_t hash;
726 
727 		switch (eh_type) {
728 #ifdef INET
729 		case ETHERTYPE_IP:
730 			hash = ip4->ip_src.s_addr + ip4->ip_dst.s_addr;
731 			break;
732 #endif
733 #ifdef INET6
734 		case ETHERTYPE_IPV6:
735 			hash = ip6->ip6_src.s6_addr32[0] +
736 			    ip6->ip6_dst.s6_addr32[0];
737 			hash += ip6->ip6_src.s6_addr32[1] +
738 			    ip6->ip6_dst.s6_addr32[1];
739 			hash += ip6->ip6_src.s6_addr32[2] +
740 			    ip6->ip6_dst.s6_addr32[2];
741 			hash += ip6->ip6_src.s6_addr32[3] +
742 			    ip6->ip6_dst.s6_addr32[3];
743 			break;
744 #endif
745 		default:
746 			hash = 0;
747 			break;
748 		}
749 		hash += th->th_sport + th->th_dport;
750 		bucket = &lc->lro_hash[hash % lc->lro_hashsz];
751 	}
752 
753 	/* Try to find a matching previous segment. */
754 	LIST_FOREACH(le, bucket, hash_next) {
755 		if (le->eh_type != eh_type)
756 			continue;
757 		if (le->source_port != th->th_sport ||
758 		    le->dest_port != th->th_dport)
759 			continue;
760 		switch (eh_type) {
761 #ifdef INET6
762 		case ETHERTYPE_IPV6:
763 			if (bcmp(&le->source_ip6, &ip6->ip6_src,
764 			    sizeof(struct in6_addr)) != 0 ||
765 			    bcmp(&le->dest_ip6, &ip6->ip6_dst,
766 			    sizeof(struct in6_addr)) != 0)
767 				continue;
768 			break;
769 #endif
770 #ifdef INET
771 		case ETHERTYPE_IP:
772 			if (le->source_ip4 != ip4->ip_src.s_addr ||
773 			    le->dest_ip4 != ip4->ip_dst.s_addr)
774 				continue;
775 			break;
776 #endif
777 		}
778 
779 		if (force_flush) {
780 			/* Timestamps mismatch; this is a FIN, etc */
781 			tcp_lro_active_remove(le);
782 			tcp_lro_flush(lc, le);
783 			return (TCP_LRO_CANNOT);
784 		}
785 
786 		/* Flush now if appending will result in overflow. */
787 		if (le->p_len > (lc->lro_length_lim - tcp_data_len)) {
788 			tcp_lro_active_remove(le);
789 			tcp_lro_flush(lc, le);
790 			break;
791 		}
792 
793 		/* Try to append the new segment. */
794 		if (__predict_false(seq != le->next_seq ||
795 		    (tcp_data_len == 0 && le->ack_seq == th->th_ack))) {
796 			/* Out of order packet or duplicate ACK. */
797 			tcp_lro_active_remove(le);
798 			tcp_lro_flush(lc, le);
799 			return (TCP_LRO_CANNOT);
800 		}
801 
802 		if (l != 0) {
803 			uint32_t tsval = ntohl(*(ts_ptr + 1));
804 			/* Make sure timestamp values are increasing. */
805 			/* XXX-BZ flip and use TSTMP_GEQ macro for this? */
806 			if (__predict_false(le->tsval > tsval ||
807 			    *(ts_ptr + 2) == 0))
808 				return (TCP_LRO_CANNOT);
809 			le->tsval = tsval;
810 			le->tsecr = *(ts_ptr + 2);
811 		}
812 
813 		le->next_seq += tcp_data_len;
814 		le->ack_seq = th->th_ack;
815 		le->window = th->th_win;
816 		le->append_cnt++;
817 
818 #ifdef TCP_LRO_UPDATE_CSUM
819 		le->ulp_csum += tcp_lro_rx_csum_fixup(le, l3hdr, th,
820 		    tcp_data_len, ~csum);
821 #endif
822 
823 		if (tcp_data_len == 0) {
824 			m_freem(m);
825 			/*
826 			 * Flush this LRO entry, if this ACK should not
827 			 * be further delayed.
828 			 */
829 			if (le->append_cnt >= lc->lro_ackcnt_lim) {
830 				tcp_lro_active_remove(le);
831 				tcp_lro_flush(lc, le);
832 			}
833 			return (0);
834 		}
835 
836 		le->p_len += tcp_data_len;
837 
838 		/*
839 		 * Adjust the mbuf so that m_data points to the first byte of
840 		 * the ULP payload.  Adjust the mbuf to avoid complications and
841 		 * append new segment to existing mbuf chain.
842 		 */
843 		m_adj(m, m->m_pkthdr.len - tcp_data_len);
844 		m_demote_pkthdr(m);
845 
846 		le->m_tail->m_next = m;
847 		le->m_tail = m_last(m);
848 
849 		/*
850 		 * If a possible next full length packet would cause an
851 		 * overflow, pro-actively flush now.
852 		 */
853 		if (le->p_len > (lc->lro_length_lim - lc->ifp->if_mtu)) {
854 			tcp_lro_active_remove(le);
855 			tcp_lro_flush(lc, le);
856 		} else
857 			getmicrotime(&le->mtime);
858 
859 		return (0);
860 	}
861 
862 	if (force_flush) {
863 		/*
864 		 * Nothing to flush, but this segment can not be further
865 		 * aggregated/delayed.
866 		 */
867 		return (TCP_LRO_CANNOT);
868 	}
869 
870 	/* Try to find an empty slot. */
871 	if (LIST_EMPTY(&lc->lro_free))
872 		return (TCP_LRO_NO_ENTRIES);
873 
874 	/* Start a new segment chain. */
875 	le = LIST_FIRST(&lc->lro_free);
876 	LIST_REMOVE(le, next);
877 	tcp_lro_active_insert(lc, bucket, le);
878 	getmicrotime(&le->mtime);
879 
880 	/* Start filling in details. */
881 	switch (eh_type) {
882 #ifdef INET6
883 	case ETHERTYPE_IPV6:
884 		le->le_ip6 = ip6;
885 		le->source_ip6 = ip6->ip6_src;
886 		le->dest_ip6 = ip6->ip6_dst;
887 		le->eh_type = eh_type;
888 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN - sizeof(*ip6);
889 		break;
890 #endif
891 #ifdef INET
892 	case ETHERTYPE_IP:
893 		le->le_ip4 = ip4;
894 		le->source_ip4 = ip4->ip_src.s_addr;
895 		le->dest_ip4 = ip4->ip_dst.s_addr;
896 		le->eh_type = eh_type;
897 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN;
898 		break;
899 #endif
900 	}
901 	le->source_port = th->th_sport;
902 	le->dest_port = th->th_dport;
903 
904 	le->next_seq = seq + tcp_data_len;
905 	le->ack_seq = th->th_ack;
906 	le->window = th->th_win;
907 	if (l != 0) {
908 		le->timestamp = 1;
909 		le->tsval = ntohl(*(ts_ptr + 1));
910 		le->tsecr = *(ts_ptr + 2);
911 	}
912 
913 #ifdef TCP_LRO_UPDATE_CSUM
914 	/*
915 	 * Do not touch the csum of the first packet.  However save the
916 	 * "adjusted" checksum of just the source and destination addresses,
917 	 * the next header and the TCP payload.  The length and TCP header
918 	 * parts may change, so we remove those from the saved checksum and
919 	 * re-add with final values on tcp_lro_flush() if needed.
920 	 */
921 	KASSERT(le->ulp_csum == 0, ("%s: le=%p le->ulp_csum=0x%04x\n",
922 	    __func__, le, le->ulp_csum));
923 
924 	le->ulp_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th, tcp_data_len,
925 	    ~csum);
926 	th->th_sum = csum;	/* Restore checksum on first packet. */
927 #endif
928 
929 	le->m_head = m;
930 	le->m_tail = m_last(m);
931 
932 	return (0);
933 }
934 
935 int
936 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
937 {
938 
939 	return tcp_lro_rx2(lc, m, csum, 1);
940 }
941 
942 void
943 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
944 {
945 	/* sanity checks */
946 	if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
947 	    lc->lro_mbuf_max == 0)) {
948 		/* packet drop */
949 		m_freem(mb);
950 		return;
951 	}
952 
953 	/* check if packet is not LRO capable */
954 	if (__predict_false(mb->m_pkthdr.csum_flags == 0 ||
955 	    (lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
956 
957 		/* input packet to network layer */
958 		(*lc->ifp->if_input) (lc->ifp, mb);
959 		return;
960 	}
961 
962 	/* create sequence number */
963 	lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
964 	    (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
965 	    (((uint64_t)mb->m_pkthdr.flowid) << 24) |
966 	    ((uint64_t)lc->lro_mbuf_count);
967 
968 	/* enter mbuf */
969 	lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb;
970 
971 	/* flush if array is full */
972 	if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max))
973 		tcp_lro_flush_all(lc);
974 }
975 
976 /* end */
977