xref: /freebsd/sys/netinet/tcp_lro.c (revision ddd5b8e9b4d8957fce018c520657cdfa4ecffad3)
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(struct lro_ctrl *lc, struct lro_entry *le)
198 {
199 
200 	if (le->append_cnt > 0) {
201 		struct tcphdr *th;
202 		uint16_t p_len;
203 
204 		p_len = htons(le->p_len);
205 		switch (le->eh_type) {
206 #ifdef INET6
207 		case ETHERTYPE_IPV6:
208 		{
209 			struct ip6_hdr *ip6;
210 
211 			ip6 = le->le_ip6;
212 			ip6->ip6_plen = p_len;
213 			th = (struct tcphdr *)(ip6 + 1);
214 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
215 			    CSUM_PSEUDO_HDR;
216 			le->p_len += ETHER_HDR_LEN + sizeof(*ip6);
217 			break;
218 		}
219 #endif
220 #ifdef INET
221 		case ETHERTYPE_IP:
222 		{
223 			struct ip *ip4;
224 #ifdef TCP_LRO_UPDATE_CSUM
225 			uint32_t cl;
226 			uint16_t c;
227 #endif
228 
229 			ip4 = le->le_ip4;
230 #ifdef TCP_LRO_UPDATE_CSUM
231 			/* Fix IP header checksum for new length. */
232 			c = ~ip4->ip_sum;
233 			cl = c;
234 			c = ~ip4->ip_len;
235 			cl += c + p_len;
236 			while (cl > 0xffff)
237 				cl = (cl >> 16) + (cl & 0xffff);
238 			c = cl;
239 			ip4->ip_sum = ~c;
240 #else
241 			ip4->ip_sum = TCP_LRO_INVALID_CSUM;
242 #endif
243 			ip4->ip_len = p_len;
244 			th = (struct tcphdr *)(ip4 + 1);
245 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
246 			    CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
247 			le->p_len += ETHER_HDR_LEN;
248 			break;
249 		}
250 #endif
251 		default:
252 			th = NULL;	/* Keep compiler happy. */
253 		}
254 		le->m_head->m_pkthdr.csum_data = 0xffff;
255 		le->m_head->m_pkthdr.len = le->p_len;
256 
257 		/* Incorporate the latest ACK into the TCP header. */
258 		th->th_ack = le->ack_seq;
259 		th->th_win = le->window;
260 		/* Incorporate latest timestamp into the TCP header. */
261 		if (le->timestamp != 0) {
262 			uint32_t *ts_ptr;
263 
264 			ts_ptr = (uint32_t *)(th + 1);
265 			ts_ptr[1] = htonl(le->tsval);
266 			ts_ptr[2] = le->tsecr;
267 		}
268 #ifdef TCP_LRO_UPDATE_CSUM
269 		/* Update the TCP header checksum. */
270 		le->ulp_csum += p_len;
271 		le->ulp_csum += tcp_lro_csum_th(th);
272 		while (le->ulp_csum > 0xffff)
273 			le->ulp_csum = (le->ulp_csum >> 16) +
274 			    (le->ulp_csum & 0xffff);
275 		th->th_sum = (le->ulp_csum & 0xffff);
276 		th->th_sum = ~th->th_sum;
277 #else
278 		th->th_sum = TCP_LRO_INVALID_CSUM;
279 #endif
280 	}
281 
282 	(*lc->ifp->if_input)(lc->ifp, le->m_head);
283 	lc->lro_queued += le->append_cnt + 1;
284 	lc->lro_flushed++;
285 	bzero(le, sizeof(*le));
286 	SLIST_INSERT_HEAD(&lc->lro_free, le, next);
287 }
288 
289 #ifdef INET6
290 static int
291 tcp_lro_rx_ipv6(struct lro_ctrl *lc, struct mbuf *m, struct ip6_hdr *ip6,
292     struct tcphdr **th)
293 {
294 
295 	/* XXX-BZ we should check the flow-label. */
296 
297 	/* XXX-BZ We do not yet support ext. hdrs. */
298 	if (ip6->ip6_nxt != IPPROTO_TCP)
299 		return (TCP_LRO_NOT_SUPPORTED);
300 
301 	/* Find the TCP header. */
302 	*th = (struct tcphdr *)(ip6 + 1);
303 
304 	return (0);
305 }
306 #endif
307 
308 #ifdef INET
309 static int
310 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4,
311     struct tcphdr **th)
312 {
313 	int csum_flags;
314 	uint16_t csum;
315 
316 	if (ip4->ip_p != IPPROTO_TCP)
317 		return (TCP_LRO_NOT_SUPPORTED);
318 
319 	/* Ensure there are no options. */
320 	if ((ip4->ip_hl << 2) != sizeof (*ip4))
321 		return (TCP_LRO_CANNOT);
322 
323 	/* .. and the packet is not fragmented. */
324 	if (ip4->ip_off & htons(IP_MF|IP_OFFMASK))
325 		return (TCP_LRO_CANNOT);
326 
327 	/* Legacy IP has a header checksum that needs to be correct. */
328 	csum_flags = m->m_pkthdr.csum_flags;
329 	if (csum_flags & CSUM_IP_CHECKED) {
330 		if (__predict_false((csum_flags & CSUM_IP_VALID) == 0)) {
331 			lc->lro_bad_csum++;
332 			return (TCP_LRO_CANNOT);
333 		}
334 	} else {
335 		csum = in_cksum_hdr(ip4);
336 		if (__predict_false((csum) != 0)) {
337 			lc->lro_bad_csum++;
338 			return (TCP_LRO_CANNOT);
339 		}
340 	}
341 
342 	/* Find the TCP header (we assured there are no IP options). */
343 	*th = (struct tcphdr *)(ip4 + 1);
344 
345 	return (0);
346 }
347 #endif
348 
349 int
350 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
351 {
352 	struct lro_entry *le;
353 	struct ether_header *eh;
354 #ifdef INET6
355 	struct ip6_hdr *ip6 = NULL;	/* Keep compiler happy. */
356 #endif
357 #ifdef INET
358 	struct ip *ip4 = NULL;		/* Keep compiler happy. */
359 #endif
360 	struct tcphdr *th;
361 	void *l3hdr = NULL;		/* Keep compiler happy. */
362 	uint32_t *ts_ptr;
363 	tcp_seq seq;
364 	int error, ip_len, l;
365 	uint16_t eh_type, tcp_data_len;
366 
367 	/* We expect a contiguous header [eh, ip, tcp]. */
368 
369 	eh = mtod(m, struct ether_header *);
370 	eh_type = ntohs(eh->ether_type);
371 	switch (eh_type) {
372 #ifdef INET6
373 	case ETHERTYPE_IPV6:
374 	{
375 		CURVNET_SET(lc->ifp->if_vnet);
376 		if (V_ip6_forwarding != 0) {
377 			/* XXX-BZ stats but changing lro_ctrl is a problem. */
378 			CURVNET_RESTORE();
379 			return (TCP_LRO_CANNOT);
380 		}
381 		CURVNET_RESTORE();
382 		l3hdr = ip6 = (struct ip6_hdr *)(eh + 1);
383 		error = tcp_lro_rx_ipv6(lc, m, ip6, &th);
384 		if (error != 0)
385 			return (error);
386 		tcp_data_len = ntohs(ip6->ip6_plen);
387 		ip_len = sizeof(*ip6) + tcp_data_len;
388 		break;
389 	}
390 #endif
391 #ifdef INET
392 	case ETHERTYPE_IP:
393 	{
394 		CURVNET_SET(lc->ifp->if_vnet);
395 		if (V_ipforwarding != 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 = ip4 = (struct ip *)(eh + 1);
402 		error = tcp_lro_rx_ipv4(lc, m, ip4, &th);
403 		if (error != 0)
404 			return (error);
405 		ip_len = ntohs(ip4->ip_len);
406 		tcp_data_len = ip_len - sizeof(*ip4);
407 		break;
408 	}
409 #endif
410 	/* XXX-BZ what happens in case of VLAN(s)? */
411 	default:
412 		return (TCP_LRO_NOT_SUPPORTED);
413 	}
414 
415 	/*
416 	 * If the frame is padded beyond the end of the IP packet, then we must
417 	 * trim the extra bytes off.
418 	 */
419 	l = m->m_pkthdr.len - (ETHER_HDR_LEN + ip_len);
420 	if (l != 0) {
421 		if (l < 0)
422 			/* Truncated packet. */
423 			return (TCP_LRO_CANNOT);
424 
425 		m_adj(m, -l);
426 	}
427 
428 	/*
429 	 * Check TCP header constraints.
430 	 */
431 	/* Ensure no bits set besides ACK or PSH. */
432 	if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0)
433 		return (TCP_LRO_CANNOT);
434 
435 	/* XXX-BZ We lose a AKC|PUSH flag concatinating multiple segments. */
436 	/* XXX-BZ Ideally we'd flush on PUSH? */
437 
438 	/*
439 	 * Check for timestamps.
440 	 * Since the only option we handle are timestamps, we only have to
441 	 * handle the simple case of aligned timestamps.
442 	 */
443 	l = (th->th_off << 2);
444 	tcp_data_len -= l;
445 	l -= sizeof(*th);
446 	ts_ptr = (uint32_t *)(th + 1);
447 	if (l != 0 && (__predict_false(l != TCPOLEN_TSTAMP_APPA) ||
448 	    (*ts_ptr != ntohl(TCPOPT_NOP<<24|TCPOPT_NOP<<16|
449 	    TCPOPT_TIMESTAMP<<8|TCPOLEN_TIMESTAMP))))
450 		return (TCP_LRO_CANNOT);
451 
452 	/* If the driver did not pass in the checksum, set it now. */
453 	if (csum == 0x0000)
454 		csum = th->th_sum;
455 
456 	seq = ntohl(th->th_seq);
457 
458 	/* Try to find a matching previous segment. */
459 	SLIST_FOREACH(le, &lc->lro_active, next) {
460 		if (le->eh_type != eh_type)
461 			continue;
462 		if (le->source_port != th->th_sport ||
463 		    le->dest_port != th->th_dport)
464 			continue;
465 		switch (eh_type) {
466 #ifdef INET6
467 		case ETHERTYPE_IPV6:
468 			if (bcmp(&le->source_ip6, &ip6->ip6_src,
469 			    sizeof(struct in6_addr)) != 0 ||
470 			    bcmp(&le->dest_ip6, &ip6->ip6_dst,
471 			    sizeof(struct in6_addr)) != 0)
472 				continue;
473 			break;
474 #endif
475 #ifdef INET
476 		case ETHERTYPE_IP:
477 			if (le->source_ip4 != ip4->ip_src.s_addr ||
478 			    le->dest_ip4 != ip4->ip_dst.s_addr)
479 				continue;
480 			break;
481 #endif
482 		}
483 
484 		/* Flush now if appending will result in overflow. */
485 		if (le->p_len > (65535 - tcp_data_len)) {
486 			SLIST_REMOVE(&lc->lro_active, le, lro_entry, next);
487 			tcp_lro_flush(lc, le);
488 			break;
489 		}
490 
491 		/* Try to append the new segment. */
492 		if (__predict_false(seq != le->next_seq ||
493 		    (tcp_data_len == 0 && le->ack_seq == th->th_ack))) {
494 			/* Out of order packet or duplicate ACK. */
495 			SLIST_REMOVE(&lc->lro_active, le, lro_entry, next);
496 			tcp_lro_flush(lc, le);
497 			return (TCP_LRO_CANNOT);
498 		}
499 
500 		if (l != 0) {
501 			uint32_t tsval = ntohl(*(ts_ptr + 1));
502 			/* Make sure timestamp values are increasing. */
503 			/* XXX-BZ flip and use TSTMP_GEQ macro for this? */
504 			if (__predict_false(le->tsval > tsval ||
505 			    *(ts_ptr + 2) == 0))
506 				return (TCP_LRO_CANNOT);
507 			le->tsval = tsval;
508 			le->tsecr = *(ts_ptr + 2);
509 		}
510 
511 		le->next_seq += tcp_data_len;
512 		le->ack_seq = th->th_ack;
513 		le->window = th->th_win;
514 		le->append_cnt++;
515 
516 #ifdef TCP_LRO_UPDATE_CSUM
517 		le->ulp_csum += tcp_lro_rx_csum_fixup(le, l3hdr, th,
518 		    tcp_data_len, ~csum);
519 #endif
520 
521 		if (tcp_data_len == 0) {
522 			m_freem(m);
523 			return (0);
524 		}
525 
526 		le->p_len += tcp_data_len;
527 
528 		/*
529 		 * Adjust the mbuf so that m_data points to the first byte of
530 		 * the ULP payload.  Adjust the mbuf to avoid complications and
531 		 * append new segment to existing mbuf chain.
532 		 */
533 		m_adj(m, m->m_pkthdr.len - tcp_data_len);
534 		m->m_flags &= ~M_PKTHDR;
535 
536 		le->m_tail->m_next = m;
537 		le->m_tail = m_last(m);
538 
539 		/*
540 		 * If a possible next full length packet would cause an
541 		 * overflow, pro-actively flush now.
542 		 */
543 		if (le->p_len > (65535 - lc->ifp->if_mtu)) {
544 			SLIST_REMOVE(&lc->lro_active, le, lro_entry, next);
545 			tcp_lro_flush(lc, le);
546 		}
547 
548 		return (0);
549 	}
550 
551 	/* Try to find an empty slot. */
552 	if (SLIST_EMPTY(&lc->lro_free))
553 		return (TCP_LRO_CANNOT);
554 
555 	/* Start a new segment chain. */
556 	le = SLIST_FIRST(&lc->lro_free);
557 	SLIST_REMOVE_HEAD(&lc->lro_free, next);
558 	SLIST_INSERT_HEAD(&lc->lro_active, le, next);
559 
560 	/* Start filling in details. */
561 	switch (eh_type) {
562 #ifdef INET6
563 	case ETHERTYPE_IPV6:
564 		le->le_ip6 = ip6;
565 		le->source_ip6 = ip6->ip6_src;
566 		le->dest_ip6 = ip6->ip6_dst;
567 		le->eh_type = eh_type;
568 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN - sizeof(*ip6);
569 		break;
570 #endif
571 #ifdef INET
572 	case ETHERTYPE_IP:
573 		le->le_ip4 = ip4;
574 		le->source_ip4 = ip4->ip_src.s_addr;
575 		le->dest_ip4 = ip4->ip_dst.s_addr;
576 		le->eh_type = eh_type;
577 		le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN;
578 		break;
579 #endif
580 	}
581 	le->source_port = th->th_sport;
582 	le->dest_port = th->th_dport;
583 
584 	le->next_seq = seq + tcp_data_len;
585 	le->ack_seq = th->th_ack;
586 	le->window = th->th_win;
587 	if (l != 0) {
588 		le->timestamp = 1;
589 		le->tsval = ntohl(*(ts_ptr + 1));
590 		le->tsecr = *(ts_ptr + 2);
591 	}
592 
593 #ifdef TCP_LRO_UPDATE_CSUM
594 	/*
595 	 * Do not touch the csum of the first packet.  However save the
596 	 * "adjusted" checksum of just the source and destination addresses,
597 	 * the next header and the TCP payload.  The length and TCP header
598 	 * parts may change, so we remove those from the saved checksum and
599 	 * re-add with final values on tcp_lro_flush() if needed.
600 	 */
601 	KASSERT(le->ulp_csum == 0, ("%s: le=%p le->ulp_csum=0x%04x\n",
602 	    __func__, le, le->ulp_csum));
603 
604 	le->ulp_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th, tcp_data_len,
605 	    ~csum);
606 	th->th_sum = csum;	/* Restore checksum on first packet. */
607 #endif
608 
609 	le->m_head = m;
610 	le->m_tail = m_last(m);
611 
612 	return (0);
613 }
614 
615 /* end */
616