xref: /freebsd/sys/netinet/tcp_lro.c (revision 7431dfd4580e850375fe5478d92ec770344db098)
1 /*-
2  * Copyright (c) 2007, Myricom Inc.
3  * Copyright (c) 2008, Intel Corporation.
4  * Copyright (c) 2012 The FreeBSD Foundation
5  * All rights reserved.
6  *
7  * Portions of this software were developed by Bjoern Zeeb
8  * under sponsorship from the FreeBSD Foundation.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include "opt_inet.h"
36 #include "opt_inet6.h"
37 
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/mbuf.h>
41 #include <sys/kernel.h>
42 #include <sys/socket.h>
43 
44 #include <net/if.h>
45 #include <net/if_var.h>
46 #include <net/ethernet.h>
47 #include <net/vnet.h>
48 
49 #include <netinet/in_systm.h>
50 #include <netinet/in.h>
51 #include <netinet/ip6.h>
52 #include <netinet/ip.h>
53 #include <netinet/ip_var.h>
54 #include <netinet/tcp.h>
55 #include <netinet/tcp_lro.h>
56 
57 #include <netinet6/ip6_var.h>
58 
59 #include <machine/in_cksum.h>
60 
61 #ifndef LRO_ENTRIES
62 #define	LRO_ENTRIES	8	/* # of LRO entries per RX queue. */
63 #endif
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 int
71 tcp_lro_init(struct lro_ctrl *lc)
72 {
73 	struct lro_entry *le;
74 	int error, i;
75 
76 	lc->lro_bad_csum = 0;
77 	lc->lro_queued = 0;
78 	lc->lro_flushed = 0;
79 	lc->lro_cnt = 0;
80 	SLIST_INIT(&lc->lro_free);
81 	SLIST_INIT(&lc->lro_active);
82 
83 	error = 0;
84 	for (i = 0; i < LRO_ENTRIES; i++) {
85 		le = (struct lro_entry *)malloc(sizeof(*le), M_DEVBUF,
86 		    M_NOWAIT | M_ZERO);
87                 if (le == NULL) {
88 			if (i == 0)
89 				error = ENOMEM;
90                         break;
91                 }
92 		lc->lro_cnt = i + 1;
93 		SLIST_INSERT_HEAD(&lc->lro_free, le, next);
94         }
95 
96 	return (error);
97 }
98 
99 void
100 tcp_lro_free(struct lro_ctrl *lc)
101 {
102 	struct lro_entry *le;
103 
104 	while (!SLIST_EMPTY(&lc->lro_free)) {
105 		le = SLIST_FIRST(&lc->lro_free);
106 		SLIST_REMOVE_HEAD(&lc->lro_free, next);
107 		free(le, M_DEVBUF);
108 	}
109 }
110 
111 #ifdef TCP_LRO_UPDATE_CSUM
112 static uint16_t
113 tcp_lro_csum_th(struct tcphdr *th)
114 {
115 	uint32_t ch;
116 	uint16_t *p, l;
117 
118 	ch = th->th_sum = 0x0000;
119 	l = th->th_off;
120 	p = (uint16_t *)th;
121 	while (l > 0) {
122 		ch += *p;
123 		p++;
124 		ch += *p;
125 		p++;
126 		l--;
127 	}
128 	while (ch > 0xffff)
129 		ch = (ch >> 16) + (ch & 0xffff);
130 
131 	return (ch & 0xffff);
132 }
133 
134 static uint16_t
135 tcp_lro_rx_csum_fixup(struct lro_entry *le, void *l3hdr, struct tcphdr *th,
136     uint16_t tcp_data_len, uint16_t csum)
137 {
138 	uint32_t c;
139 	uint16_t cs;
140 
141 	c = csum;
142 
143 	/* Remove length from checksum. */
144 	switch (le->eh_type) {
145 #ifdef INET6
146 	case ETHERTYPE_IPV6:
147 	{
148 		struct ip6_hdr *ip6;
149 
150 		ip6 = (struct ip6_hdr *)l3hdr;
151 		if (le->append_cnt == 0)
152 			cs = ip6->ip6_plen;
153 		else {
154 			uint32_t cx;
155 
156 			cx = ntohs(ip6->ip6_plen);
157 			cs = in6_cksum_pseudo(ip6, cx, ip6->ip6_nxt, 0);
158 		}
159 		break;
160 	}
161 #endif
162 #ifdef INET
163 	case ETHERTYPE_IP:
164 	{
165 		struct ip *ip4;
166 
167 		ip4 = (struct ip *)l3hdr;
168 		if (le->append_cnt == 0)
169 			cs = ip4->ip_len;
170 		else {
171 			cs = in_addword(ntohs(ip4->ip_len) - sizeof(*ip4),
172 			    IPPROTO_TCP);
173 			cs = in_pseudo(ip4->ip_src.s_addr, ip4->ip_dst.s_addr,
174 			    htons(cs));
175 		}
176 		break;
177 	}
178 #endif
179 	default:
180 		cs = 0;		/* Keep compiler happy. */
181 	}
182 
183 	cs = ~cs;
184 	c += cs;
185 
186 	/* Remove TCP header csum. */
187 	cs = ~tcp_lro_csum_th(th);
188 	c += cs;
189 	while (c > 0xffff)
190 		c = (c >> 16) + (c & 0xffff);
191 
192 	return (c & 0xffff);
193 }
194 #endif
195 
196 void
197 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
198 {
199 	struct lro_entry *le, *le_tmp;
200 	struct timeval tv;
201 
202 	if (SLIST_EMPTY(&lc->lro_active))
203 		return;
204 
205 	getmicrotime(&tv);
206 	timevalsub(&tv, timeout);
207 	SLIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
208 		if (timevalcmp(&tv, &le->mtime, >=)) {
209 			SLIST_REMOVE(&lc->lro_active, le, lro_entry, next);
210 			tcp_lro_flush(lc, le);
211 		}
212 	}
213 }
214 
215 void
216 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
217 {
218 
219 	if (le->append_cnt > 0) {
220 		struct tcphdr *th;
221 		uint16_t p_len;
222 
223 		p_len = htons(le->p_len);
224 		switch (le->eh_type) {
225 #ifdef INET6
226 		case ETHERTYPE_IPV6:
227 		{
228 			struct ip6_hdr *ip6;
229 
230 			ip6 = le->le_ip6;
231 			ip6->ip6_plen = p_len;
232 			th = (struct tcphdr *)(ip6 + 1);
233 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
234 			    CSUM_PSEUDO_HDR;
235 			le->p_len += ETHER_HDR_LEN + sizeof(*ip6);
236 			break;
237 		}
238 #endif
239 #ifdef INET
240 		case ETHERTYPE_IP:
241 		{
242 			struct ip *ip4;
243 #ifdef TCP_LRO_UPDATE_CSUM
244 			uint32_t cl;
245 			uint16_t c;
246 #endif
247 
248 			ip4 = le->le_ip4;
249 #ifdef TCP_LRO_UPDATE_CSUM
250 			/* Fix IP header checksum for new length. */
251 			c = ~ip4->ip_sum;
252 			cl = c;
253 			c = ~ip4->ip_len;
254 			cl += c + p_len;
255 			while (cl > 0xffff)
256 				cl = (cl >> 16) + (cl & 0xffff);
257 			c = cl;
258 			ip4->ip_sum = ~c;
259 #else
260 			ip4->ip_sum = TCP_LRO_INVALID_CSUM;
261 #endif
262 			ip4->ip_len = p_len;
263 			th = (struct tcphdr *)(ip4 + 1);
264 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
265 			    CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
266 			le->p_len += ETHER_HDR_LEN;
267 			break;
268 		}
269 #endif
270 		default:
271 			th = NULL;	/* Keep compiler happy. */
272 		}
273 		le->m_head->m_pkthdr.csum_data = 0xffff;
274 		le->m_head->m_pkthdr.len = le->p_len;
275 
276 		/* Incorporate the latest ACK into the TCP header. */
277 		th->th_ack = le->ack_seq;
278 		th->th_win = le->window;
279 		/* Incorporate latest timestamp into the TCP header. */
280 		if (le->timestamp != 0) {
281 			uint32_t *ts_ptr;
282 
283 			ts_ptr = (uint32_t *)(th + 1);
284 			ts_ptr[1] = htonl(le->tsval);
285 			ts_ptr[2] = le->tsecr;
286 		}
287 #ifdef TCP_LRO_UPDATE_CSUM
288 		/* Update the TCP header checksum. */
289 		le->ulp_csum += p_len;
290 		le->ulp_csum += tcp_lro_csum_th(th);
291 		while (le->ulp_csum > 0xffff)
292 			le->ulp_csum = (le->ulp_csum >> 16) +
293 			    (le->ulp_csum & 0xffff);
294 		th->th_sum = (le->ulp_csum & 0xffff);
295 		th->th_sum = ~th->th_sum;
296 #else
297 		th->th_sum = TCP_LRO_INVALID_CSUM;
298 #endif
299 	}
300 
301 	(*lc->ifp->if_input)(lc->ifp, le->m_head);
302 	lc->lro_queued += le->append_cnt + 1;
303 	lc->lro_flushed++;
304 	bzero(le, sizeof(*le));
305 	SLIST_INSERT_HEAD(&lc->lro_free, le, next);
306 }
307 
308 #ifdef INET6
309 static int
310 tcp_lro_rx_ipv6(struct lro_ctrl *lc, struct mbuf *m, struct ip6_hdr *ip6,
311     struct tcphdr **th)
312 {
313 
314 	/* XXX-BZ we should check the flow-label. */
315 
316 	/* XXX-BZ We do not yet support ext. hdrs. */
317 	if (ip6->ip6_nxt != IPPROTO_TCP)
318 		return (TCP_LRO_NOT_SUPPORTED);
319 
320 	/* Find the TCP header. */
321 	*th = (struct tcphdr *)(ip6 + 1);
322 
323 	return (0);
324 }
325 #endif
326 
327 #ifdef INET
328 static int
329 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4,
330     struct tcphdr **th)
331 {
332 	int csum_flags;
333 	uint16_t csum;
334 
335 	if (ip4->ip_p != IPPROTO_TCP)
336 		return (TCP_LRO_NOT_SUPPORTED);
337 
338 	/* Ensure there are no options. */
339 	if ((ip4->ip_hl << 2) != sizeof (*ip4))
340 		return (TCP_LRO_CANNOT);
341 
342 	/* .. and the packet is not fragmented. */
343 	if (ip4->ip_off & htons(IP_MF|IP_OFFMASK))
344 		return (TCP_LRO_CANNOT);
345 
346 	/* Legacy IP has a header checksum that needs to be correct. */
347 	csum_flags = m->m_pkthdr.csum_flags;
348 	if (csum_flags & CSUM_IP_CHECKED) {
349 		if (__predict_false((csum_flags & CSUM_IP_VALID) == 0)) {
350 			lc->lro_bad_csum++;
351 			return (TCP_LRO_CANNOT);
352 		}
353 	} else {
354 		csum = in_cksum_hdr(ip4);
355 		if (__predict_false((csum) != 0)) {
356 			lc->lro_bad_csum++;
357 			return (TCP_LRO_CANNOT);
358 		}
359 	}
360 
361 	/* Find the TCP header (we assured there are no IP options). */
362 	*th = (struct tcphdr *)(ip4 + 1);
363 
364 	return (0);
365 }
366 #endif
367 
368 int
369 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
370 {
371 	struct lro_entry *le;
372 	struct ether_header *eh;
373 #ifdef INET6
374 	struct ip6_hdr *ip6 = NULL;	/* Keep compiler happy. */
375 #endif
376 #ifdef INET
377 	struct ip *ip4 = NULL;		/* Keep compiler happy. */
378 #endif
379 	struct tcphdr *th;
380 	void *l3hdr = NULL;		/* Keep compiler happy. */
381 	uint32_t *ts_ptr;
382 	tcp_seq seq;
383 	int error, ip_len, l;
384 	uint16_t eh_type, tcp_data_len;
385 
386 	/* We expect a contiguous header [eh, ip, tcp]. */
387 
388 	eh = mtod(m, struct ether_header *);
389 	eh_type = ntohs(eh->ether_type);
390 	switch (eh_type) {
391 #ifdef INET6
392 	case ETHERTYPE_IPV6:
393 	{
394 		CURVNET_SET(lc->ifp->if_vnet);
395 		if (V_ip6_forwarding != 0) {
396 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
397 			CURVNET_RESTORE();
398 			return (TCP_LRO_CANNOT);
399 		}
400 		CURVNET_RESTORE();
401 		l3hdr = ip6 = (struct ip6_hdr *)(eh + 1);
402 		error = tcp_lro_rx_ipv6(lc, m, ip6, &th);
403 		if (error != 0)
404 			return (error);
405 		tcp_data_len = ntohs(ip6->ip6_plen);
406 		ip_len = sizeof(*ip6) + tcp_data_len;
407 		break;
408 	}
409 #endif
410 #ifdef INET
411 	case ETHERTYPE_IP:
412 	{
413 		CURVNET_SET(lc->ifp->if_vnet);
414 		if (V_ipforwarding != 0) {
415 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
416 			CURVNET_RESTORE();
417 			return (TCP_LRO_CANNOT);
418 		}
419 		CURVNET_RESTORE();
420 		l3hdr = ip4 = (struct ip *)(eh + 1);
421 		error = tcp_lro_rx_ipv4(lc, m, ip4, &th);
422 		if (error != 0)
423 			return (error);
424 		ip_len = ntohs(ip4->ip_len);
425 		tcp_data_len = ip_len - sizeof(*ip4);
426 		break;
427 	}
428 #endif
429 	/* XXX-BZ what happens in case of VLAN(s)? */
430 	default:
431 		return (TCP_LRO_NOT_SUPPORTED);
432 	}
433 
434 	/*
435 	 * If the frame is padded beyond the end of the IP packet, then we must
436 	 * trim the extra bytes off.
437 	 */
438 	l = m->m_pkthdr.len - (ETHER_HDR_LEN + ip_len);
439 	if (l != 0) {
440 		if (l < 0)
441 			/* Truncated packet. */
442 			return (TCP_LRO_CANNOT);
443 
444 		m_adj(m, -l);
445 	}
446 
447 	/*
448 	 * Check TCP header constraints.
449 	 */
450 	/* Ensure no bits set besides ACK or PSH. */
451 	if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0)
452 		return (TCP_LRO_CANNOT);
453 
454 	/* XXX-BZ We lose a AKC|PUSH flag concatinating multiple segments. */
455 	/* XXX-BZ Ideally we'd flush on PUSH? */
456 
457 	/*
458 	 * Check for timestamps.
459 	 * Since the only option we handle are timestamps, we only have to
460 	 * handle the simple case of aligned timestamps.
461 	 */
462 	l = (th->th_off << 2);
463 	tcp_data_len -= l;
464 	l -= sizeof(*th);
465 	ts_ptr = (uint32_t *)(th + 1);
466 	if (l != 0 && (__predict_false(l != TCPOLEN_TSTAMP_APPA) ||
467 	    (*ts_ptr != ntohl(TCPOPT_NOP<<24|TCPOPT_NOP<<16|
468 	    TCPOPT_TIMESTAMP<<8|TCPOLEN_TIMESTAMP))))
469 		return (TCP_LRO_CANNOT);
470 
471 	/* If the driver did not pass in the checksum, set it now. */
472 	if (csum == 0x0000)
473 		csum = th->th_sum;
474 
475 	seq = ntohl(th->th_seq);
476 
477 	/* Try to find a matching previous segment. */
478 	SLIST_FOREACH(le, &lc->lro_active, next) {
479 		if (le->eh_type != eh_type)
480 			continue;
481 		if (le->source_port != th->th_sport ||
482 		    le->dest_port != th->th_dport)
483 			continue;
484 		switch (eh_type) {
485 #ifdef INET6
486 		case ETHERTYPE_IPV6:
487 			if (bcmp(&le->source_ip6, &ip6->ip6_src,
488 			    sizeof(struct in6_addr)) != 0 ||
489 			    bcmp(&le->dest_ip6, &ip6->ip6_dst,
490 			    sizeof(struct in6_addr)) != 0)
491 				continue;
492 			break;
493 #endif
494 #ifdef INET
495 		case ETHERTYPE_IP:
496 			if (le->source_ip4 != ip4->ip_src.s_addr ||
497 			    le->dest_ip4 != ip4->ip_dst.s_addr)
498 				continue;
499 			break;
500 #endif
501 		}
502 
503 		/* Flush now if appending will result in overflow. */
504 		if (le->p_len > (65535 - tcp_data_len)) {
505 			SLIST_REMOVE(&lc->lro_active, le, lro_entry, next);
506 			tcp_lro_flush(lc, le);
507 			break;
508 		}
509 
510 		/* Try to append the new segment. */
511 		if (__predict_false(seq != le->next_seq ||
512 		    (tcp_data_len == 0 && le->ack_seq == th->th_ack))) {
513 			/* Out of order packet or duplicate ACK. */
514 			SLIST_REMOVE(&lc->lro_active, le, lro_entry, next);
515 			tcp_lro_flush(lc, le);
516 			return (TCP_LRO_CANNOT);
517 		}
518 
519 		if (l != 0) {
520 			uint32_t tsval = ntohl(*(ts_ptr + 1));
521 			/* Make sure timestamp values are increasing. */
522 			/* XXX-BZ flip and use TSTMP_GEQ macro for this? */
523 			if (__predict_false(le->tsval > tsval ||
524 			    *(ts_ptr + 2) == 0))
525 				return (TCP_LRO_CANNOT);
526 			le->tsval = tsval;
527 			le->tsecr = *(ts_ptr + 2);
528 		}
529 
530 		le->next_seq += tcp_data_len;
531 		le->ack_seq = th->th_ack;
532 		le->window = th->th_win;
533 		le->append_cnt++;
534 
535 #ifdef TCP_LRO_UPDATE_CSUM
536 		le->ulp_csum += tcp_lro_rx_csum_fixup(le, l3hdr, th,
537 		    tcp_data_len, ~csum);
538 #endif
539 
540 		if (tcp_data_len == 0) {
541 			m_freem(m);
542 			return (0);
543 		}
544 
545 		le->p_len += tcp_data_len;
546 
547 		/*
548 		 * Adjust the mbuf so that m_data points to the first byte of
549 		 * the ULP payload.  Adjust the mbuf to avoid complications and
550 		 * append new segment to existing mbuf chain.
551 		 */
552 		m_adj(m, m->m_pkthdr.len - tcp_data_len);
553 		m->m_flags &= ~M_PKTHDR;
554 
555 		le->m_tail->m_next = m;
556 		le->m_tail = m_last(m);
557 
558 		/*
559 		 * If a possible next full length packet would cause an
560 		 * overflow, pro-actively flush now.
561 		 */
562 		if (le->p_len > (65535 - lc->ifp->if_mtu)) {
563 			SLIST_REMOVE(&lc->lro_active, le, lro_entry, next);
564 			tcp_lro_flush(lc, le);
565 		} else
566 			getmicrotime(&le->mtime);
567 
568 		return (0);
569 	}
570 
571 	/* Try to find an empty slot. */
572 	if (SLIST_EMPTY(&lc->lro_free))
573 		return (TCP_LRO_CANNOT);
574 
575 	/* Start a new segment chain. */
576 	le = SLIST_FIRST(&lc->lro_free);
577 	SLIST_REMOVE_HEAD(&lc->lro_free, next);
578 	SLIST_INSERT_HEAD(&lc->lro_active, le, next);
579 	getmicrotime(&le->mtime);
580 
581 	/* Start filling in details. */
582 	switch (eh_type) {
583 #ifdef INET6
584 	case ETHERTYPE_IPV6:
585 		le->le_ip6 = ip6;
586 		le->source_ip6 = ip6->ip6_src;
587 		le->dest_ip6 = ip6->ip6_dst;
588 		le->eh_type = eh_type;
589 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN - sizeof(*ip6);
590 		break;
591 #endif
592 #ifdef INET
593 	case ETHERTYPE_IP:
594 		le->le_ip4 = ip4;
595 		le->source_ip4 = ip4->ip_src.s_addr;
596 		le->dest_ip4 = ip4->ip_dst.s_addr;
597 		le->eh_type = eh_type;
598 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN;
599 		break;
600 #endif
601 	}
602 	le->source_port = th->th_sport;
603 	le->dest_port = th->th_dport;
604 
605 	le->next_seq = seq + tcp_data_len;
606 	le->ack_seq = th->th_ack;
607 	le->window = th->th_win;
608 	if (l != 0) {
609 		le->timestamp = 1;
610 		le->tsval = ntohl(*(ts_ptr + 1));
611 		le->tsecr = *(ts_ptr + 2);
612 	}
613 
614 #ifdef TCP_LRO_UPDATE_CSUM
615 	/*
616 	 * Do not touch the csum of the first packet.  However save the
617 	 * "adjusted" checksum of just the source and destination addresses,
618 	 * the next header and the TCP payload.  The length and TCP header
619 	 * parts may change, so we remove those from the saved checksum and
620 	 * re-add with final values on tcp_lro_flush() if needed.
621 	 */
622 	KASSERT(le->ulp_csum == 0, ("%s: le=%p le->ulp_csum=0x%04x\n",
623 	    __func__, le, le->ulp_csum));
624 
625 	le->ulp_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th, tcp_data_len,
626 	    ~csum);
627 	th->th_sum = csum;	/* Restore checksum on first packet. */
628 #endif
629 
630 	le->m_head = m;
631 	le->m_tail = m_last(m);
632 
633 	return (0);
634 }
635 
636 /* end */
637