xref: /freebsd/sys/netinet/tcp_lro.c (revision 849f9ac370bd66993ce5cc6fca0d2ef9bd03c2c9)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2007, Myricom Inc.
5  * Copyright (c) 2008, Intel Corporation.
6  * Copyright (c) 2012 The FreeBSD Foundation
7  * Copyright (c) 2016-2021 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 #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/socketvar.h>
46 #include <sys/sockbuf.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/bpf.h>
53 #include <net/vnet.h>
54 #include <net/if_dl.h>
55 #include <net/if_media.h>
56 #include <net/if_private.h>
57 #include <net/if_types.h>
58 #include <net/infiniband.h>
59 #include <net/if_lagg.h>
60 
61 #include <netinet/in_systm.h>
62 #include <netinet/in.h>
63 #include <netinet/ip6.h>
64 #include <netinet/ip.h>
65 #include <netinet/ip_var.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet6/in6_pcb.h>
68 #include <netinet/tcp.h>
69 #include <netinet/tcp_seq.h>
70 #include <netinet/tcp_lro.h>
71 #include <netinet/tcp_var.h>
72 #include <netinet/tcpip.h>
73 #include <netinet/tcp_hpts.h>
74 #include <netinet/tcp_log_buf.h>
75 #include <netinet/tcp_fsm.h>
76 #include <netinet/udp.h>
77 #include <netinet6/ip6_var.h>
78 
79 #include <machine/in_cksum.h>
80 
81 static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures");
82 
83 static void	tcp_lro_rx_done(struct lro_ctrl *lc);
84 static int	tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m,
85 		    uint32_t csum, bool use_hash);
86 
87 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro,  CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
88     "TCP LRO");
89 
90 long tcplro_stacks_wanting_mbufq;
91 int	(*tcp_lro_flush_tcphpts)(struct lro_ctrl *lc, struct lro_entry *le);
92 
93 counter_u64_t tcp_inp_lro_direct_queue;
94 counter_u64_t tcp_inp_lro_wokeup_queue;
95 counter_u64_t tcp_inp_lro_compressed;
96 counter_u64_t tcp_inp_lro_locks_taken;
97 counter_u64_t tcp_extra_mbuf;
98 counter_u64_t tcp_would_have_but;
99 counter_u64_t tcp_comp_total;
100 counter_u64_t tcp_uncomp_total;
101 counter_u64_t tcp_bad_csums;
102 
103 static unsigned	tcp_lro_entries = TCP_LRO_ENTRIES;
104 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries,
105     CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_entries, 0,
106     "default number of LRO entries");
107 
108 static uint32_t tcp_lro_cpu_set_thresh = TCP_LRO_CPU_DECLARATION_THRESH;
109 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_cpu_threshold,
110     CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_cpu_set_thresh, 0,
111     "Number of interrupts in a row on the same CPU that will make us declare an 'affinity' cpu?");
112 
113 static uint32_t tcp_less_accurate_lro_ts = 0;
114 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_less_accurate,
115     CTLFLAG_MPSAFE, &tcp_less_accurate_lro_ts, 0,
116     "Do we trade off efficency by doing less timestamp operations for time accuracy?");
117 
118 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, fullqueue, CTLFLAG_RD,
119     &tcp_inp_lro_direct_queue, "Number of lro's fully queued to transport");
120 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, wokeup, CTLFLAG_RD,
121     &tcp_inp_lro_wokeup_queue, "Number of lro's where we woke up transport via hpts");
122 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, compressed, CTLFLAG_RD,
123     &tcp_inp_lro_compressed, "Number of lro's compressed and sent to transport");
124 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lockcnt, CTLFLAG_RD,
125     &tcp_inp_lro_locks_taken, "Number of lro's inp_wlocks taken");
126 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, extra_mbuf, CTLFLAG_RD,
127     &tcp_extra_mbuf, "Number of times we had an extra compressed ack dropped into the tp");
128 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, would_have_but, CTLFLAG_RD,
129     &tcp_would_have_but, "Number of times we would have had an extra compressed, but mget failed");
130 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, with_m_ackcmp, CTLFLAG_RD,
131     &tcp_comp_total, "Number of mbufs queued with M_ACKCMP flags set");
132 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, without_m_ackcmp, CTLFLAG_RD,
133     &tcp_uncomp_total, "Number of mbufs queued without M_ACKCMP");
134 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lro_badcsum, CTLFLAG_RD,
135     &tcp_bad_csums, "Number of packets that the common code saw with bad csums");
136 
137 void
138 tcp_lro_reg_mbufq(void)
139 {
140 	atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, 1);
141 }
142 
143 void
144 tcp_lro_dereg_mbufq(void)
145 {
146 	atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, -1);
147 }
148 
149 static __inline void
150 tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket,
151     struct lro_entry *le)
152 {
153 
154 	LIST_INSERT_HEAD(&lc->lro_active, le, next);
155 	LIST_INSERT_HEAD(bucket, le, hash_next);
156 }
157 
158 static __inline void
159 tcp_lro_active_remove(struct lro_entry *le)
160 {
161 
162 	LIST_REMOVE(le, next);		/* active list */
163 	LIST_REMOVE(le, hash_next);	/* hash bucket */
164 }
165 
166 int
167 tcp_lro_init(struct lro_ctrl *lc)
168 {
169 	return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0));
170 }
171 
172 int
173 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
174     unsigned lro_entries, unsigned lro_mbufs)
175 {
176 	struct lro_entry *le;
177 	size_t size;
178 	unsigned i;
179 
180 	lc->lro_bad_csum = 0;
181 	lc->lro_queued = 0;
182 	lc->lro_flushed = 0;
183 	lc->lro_mbuf_count = 0;
184 	lc->lro_mbuf_max = lro_mbufs;
185 	lc->lro_cnt = lro_entries;
186 	lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
187 	lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
188 	lc->ifp = ifp;
189 	LIST_INIT(&lc->lro_free);
190 	LIST_INIT(&lc->lro_active);
191 
192 	/* create hash table to accelerate entry lookup */
193 	lc->lro_hash = phashinit_flags(lro_entries, M_LRO, &lc->lro_hashsz,
194 	    HASH_NOWAIT);
195 	if (lc->lro_hash == NULL) {
196 		memset(lc, 0, sizeof(*lc));
197 		return (ENOMEM);
198 	}
199 
200 	/* compute size to allocate */
201 	size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
202 	    (lro_entries * sizeof(*le));
203 	lc->lro_mbuf_data = (struct lro_mbuf_sort *)
204 	    malloc(size, M_LRO, M_NOWAIT | M_ZERO);
205 
206 	/* check for out of memory */
207 	if (lc->lro_mbuf_data == NULL) {
208 		free(lc->lro_hash, M_LRO);
209 		memset(lc, 0, sizeof(*lc));
210 		return (ENOMEM);
211 	}
212 	/* compute offset for LRO entries */
213 	le = (struct lro_entry *)
214 	    (lc->lro_mbuf_data + lro_mbufs);
215 
216 	/* setup linked list */
217 	for (i = 0; i != lro_entries; i++)
218 		LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
219 
220 	return (0);
221 }
222 
223 struct vxlan_header {
224 	uint32_t	vxlh_flags;
225 	uint32_t	vxlh_vni;
226 };
227 
228 static inline void *
229 tcp_lro_low_level_parser(void *ptr, struct lro_parser *parser, bool update_data, bool is_vxlan, int mlen)
230 {
231 	const struct ether_vlan_header *eh;
232 	void *old;
233 	uint16_t eth_type;
234 
235 	if (update_data)
236 		memset(parser, 0, sizeof(*parser));
237 
238 	old = ptr;
239 
240 	if (is_vxlan) {
241 		const struct vxlan_header *vxh;
242 		vxh = ptr;
243 		ptr = (uint8_t *)ptr + sizeof(*vxh);
244 		if (update_data) {
245 			parser->data.vxlan_vni =
246 			    vxh->vxlh_vni & htonl(0xffffff00);
247 		}
248 	}
249 
250 	eh = ptr;
251 	if (__predict_false(eh->evl_encap_proto == htons(ETHERTYPE_VLAN))) {
252 		eth_type = eh->evl_proto;
253 		if (update_data) {
254 			/* strip priority and keep VLAN ID only */
255 			parser->data.vlan_id = eh->evl_tag & htons(EVL_VLID_MASK);
256 		}
257 		/* advance to next header */
258 		ptr = (uint8_t *)ptr + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
259 		mlen -= (ETHER_HDR_LEN  + ETHER_VLAN_ENCAP_LEN);
260 	} else {
261 		eth_type = eh->evl_encap_proto;
262 		/* advance to next header */
263 		mlen -= ETHER_HDR_LEN;
264 		ptr = (uint8_t *)ptr + ETHER_HDR_LEN;
265 	}
266 	if (__predict_false(mlen <= 0))
267 		return (NULL);
268 	switch (eth_type) {
269 #ifdef INET
270 	case htons(ETHERTYPE_IP):
271 		parser->ip4 = ptr;
272 		if (__predict_false(mlen < sizeof(struct ip)))
273 			return (NULL);
274 		/* Ensure there are no IPv4 options. */
275 		if ((parser->ip4->ip_hl << 2) != sizeof (*parser->ip4))
276 			break;
277 		/* .. and the packet is not fragmented. */
278 		if (parser->ip4->ip_off & htons(IP_MF|IP_OFFMASK))
279 			break;
280 		/* .. and the packet has valid src/dst addrs */
281 		if (__predict_false(parser->ip4->ip_src.s_addr == INADDR_ANY ||
282 			parser->ip4->ip_dst.s_addr == INADDR_ANY))
283 			break;
284 		ptr = (uint8_t *)ptr + (parser->ip4->ip_hl << 2);
285 		mlen -= sizeof(struct ip);
286 		if (update_data) {
287 			parser->data.s_addr.v4 = parser->ip4->ip_src;
288 			parser->data.d_addr.v4 = parser->ip4->ip_dst;
289 		}
290 		switch (parser->ip4->ip_p) {
291 		case IPPROTO_UDP:
292 			if (__predict_false(mlen < sizeof(struct udphdr)))
293 				return (NULL);
294 			parser->udp = ptr;
295 			if (update_data) {
296 				parser->data.lro_type = LRO_TYPE_IPV4_UDP;
297 				parser->data.s_port = parser->udp->uh_sport;
298 				parser->data.d_port = parser->udp->uh_dport;
299 			} else {
300 				MPASS(parser->data.lro_type == LRO_TYPE_IPV4_UDP);
301 			}
302 			ptr = ((uint8_t *)ptr + sizeof(*parser->udp));
303 			parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
304 			return (ptr);
305 		case IPPROTO_TCP:
306 			parser->tcp = ptr;
307 			if (__predict_false(mlen < sizeof(struct tcphdr)))
308 				return (NULL);
309 			if (update_data) {
310 				parser->data.lro_type = LRO_TYPE_IPV4_TCP;
311 				parser->data.s_port = parser->tcp->th_sport;
312 				parser->data.d_port = parser->tcp->th_dport;
313 			} else {
314 				MPASS(parser->data.lro_type == LRO_TYPE_IPV4_TCP);
315 			}
316 			if (__predict_false(mlen < (parser->tcp->th_off << 2)))
317 				return (NULL);
318 			ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
319 			parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
320 			return (ptr);
321 		default:
322 			break;
323 		}
324 		break;
325 #endif
326 #ifdef INET6
327 	case htons(ETHERTYPE_IPV6):
328 		parser->ip6 = ptr;
329 		if (__predict_false(mlen < sizeof(struct ip6_hdr)))
330 			return (NULL);
331 		/* Ensure the packet has valid src/dst addrs */
332 		if (__predict_false(IN6_IS_ADDR_UNSPECIFIED(&parser->ip6->ip6_src) ||
333 			IN6_IS_ADDR_UNSPECIFIED(&parser->ip6->ip6_dst)))
334 			return (NULL);
335 		ptr = (uint8_t *)ptr + sizeof(*parser->ip6);
336 		if (update_data) {
337 			parser->data.s_addr.v6 = parser->ip6->ip6_src;
338 			parser->data.d_addr.v6 = parser->ip6->ip6_dst;
339 		}
340 		mlen -= sizeof(struct ip6_hdr);
341 		switch (parser->ip6->ip6_nxt) {
342 		case IPPROTO_UDP:
343 			if (__predict_false(mlen < sizeof(struct udphdr)))
344 				return (NULL);
345 			parser->udp = ptr;
346 			if (update_data) {
347 				parser->data.lro_type = LRO_TYPE_IPV6_UDP;
348 				parser->data.s_port = parser->udp->uh_sport;
349 				parser->data.d_port = parser->udp->uh_dport;
350 			} else {
351 				MPASS(parser->data.lro_type == LRO_TYPE_IPV6_UDP);
352 			}
353 			ptr = (uint8_t *)ptr + sizeof(*parser->udp);
354 			parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
355 			return (ptr);
356 		case IPPROTO_TCP:
357 			if (__predict_false(mlen < sizeof(struct tcphdr)))
358 				return (NULL);
359 			parser->tcp = ptr;
360 			if (update_data) {
361 				parser->data.lro_type = LRO_TYPE_IPV6_TCP;
362 				parser->data.s_port = parser->tcp->th_sport;
363 				parser->data.d_port = parser->tcp->th_dport;
364 			} else {
365 				MPASS(parser->data.lro_type == LRO_TYPE_IPV6_TCP);
366 			}
367 			if (__predict_false(mlen < (parser->tcp->th_off << 2)))
368 				return (NULL);
369 			ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
370 			parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
371 			return (ptr);
372 		default:
373 			break;
374 		}
375 		break;
376 #endif
377 	default:
378 		break;
379 	}
380 	/* Invalid packet - cannot parse */
381 	return (NULL);
382 }
383 
384 static const int vxlan_csum = CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
385     CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID;
386 
387 static inline struct lro_parser *
388 tcp_lro_parser(struct mbuf *m, struct lro_parser *po, struct lro_parser *pi, bool update_data)
389 {
390 	void *data_ptr;
391 
392 	/* Try to parse outer headers first. */
393 	data_ptr = tcp_lro_low_level_parser(m->m_data, po, update_data, false, m->m_len);
394 	if (data_ptr == NULL || po->total_hdr_len > m->m_len)
395 		return (NULL);
396 
397 	if (update_data) {
398 		/* Store VLAN ID, if any. */
399 		if (__predict_false(m->m_flags & M_VLANTAG)) {
400 			po->data.vlan_id =
401 			    htons(m->m_pkthdr.ether_vtag) & htons(EVL_VLID_MASK);
402 		}
403 		/* Store decrypted flag, if any. */
404 		if (__predict_false((m->m_pkthdr.csum_flags &
405 		    CSUM_TLS_MASK) == CSUM_TLS_DECRYPTED))
406 			po->data.lro_flags |= LRO_FLAG_DECRYPTED;
407 	}
408 
409 	switch (po->data.lro_type) {
410 	case LRO_TYPE_IPV4_UDP:
411 	case LRO_TYPE_IPV6_UDP:
412 		/* Check for VXLAN headers. */
413 		if ((m->m_pkthdr.csum_flags & vxlan_csum) != vxlan_csum)
414 			break;
415 
416 		/* Try to parse inner headers. */
417 		data_ptr = tcp_lro_low_level_parser(data_ptr, pi, update_data, true,
418 						    (m->m_len - ((caddr_t)data_ptr - m->m_data)));
419 		if (data_ptr == NULL || (pi->total_hdr_len + po->total_hdr_len) > m->m_len)
420 			break;
421 
422 		/* Verify supported header types. */
423 		switch (pi->data.lro_type) {
424 		case LRO_TYPE_IPV4_TCP:
425 		case LRO_TYPE_IPV6_TCP:
426 			return (pi);
427 		default:
428 			break;
429 		}
430 		break;
431 	case LRO_TYPE_IPV4_TCP:
432 	case LRO_TYPE_IPV6_TCP:
433 		if (update_data)
434 			memset(pi, 0, sizeof(*pi));
435 		return (po);
436 	default:
437 		break;
438 	}
439 	return (NULL);
440 }
441 
442 static inline int
443 tcp_lro_trim_mbuf_chain(struct mbuf *m, const struct lro_parser *po)
444 {
445 	int len;
446 
447 	switch (po->data.lro_type) {
448 #ifdef INET
449 	case LRO_TYPE_IPV4_TCP:
450 		len = ((uint8_t *)po->ip4 - (uint8_t *)m->m_data) +
451 		    ntohs(po->ip4->ip_len);
452 		break;
453 #endif
454 #ifdef INET6
455 	case LRO_TYPE_IPV6_TCP:
456 		len = ((uint8_t *)po->ip6 - (uint8_t *)m->m_data) +
457 		    ntohs(po->ip6->ip6_plen) + sizeof(*po->ip6);
458 		break;
459 #endif
460 	default:
461 		return (TCP_LRO_CANNOT);
462 	}
463 
464 	/*
465 	 * If the frame is padded beyond the end of the IP packet,
466 	 * then trim the extra bytes off:
467 	 */
468 	if (__predict_true(m->m_pkthdr.len == len)) {
469 		return (0);
470 	} else if (m->m_pkthdr.len > len) {
471 		m_adj(m, len - m->m_pkthdr.len);
472 		return (0);
473 	}
474 	return (TCP_LRO_CANNOT);
475 }
476 
477 static void
478 lro_free_mbuf_chain(struct mbuf *m)
479 {
480 	struct mbuf *save;
481 
482 	while (m) {
483 		save = m->m_nextpkt;
484 		m->m_nextpkt = NULL;
485 		m_freem(m);
486 		m = save;
487 	}
488 }
489 
490 void
491 tcp_lro_free(struct lro_ctrl *lc)
492 {
493 	struct lro_entry *le;
494 	unsigned x;
495 
496 	/* reset LRO free list */
497 	LIST_INIT(&lc->lro_free);
498 
499 	/* free active mbufs, if any */
500 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
501 		tcp_lro_active_remove(le);
502 		lro_free_mbuf_chain(le->m_head);
503 	}
504 
505 	/* free hash table */
506 	free(lc->lro_hash, M_LRO);
507 	lc->lro_hash = NULL;
508 	lc->lro_hashsz = 0;
509 
510 	/* free mbuf array, if any */
511 	for (x = 0; x != lc->lro_mbuf_count; x++)
512 		m_freem(lc->lro_mbuf_data[x].mb);
513 	lc->lro_mbuf_count = 0;
514 
515 	/* free allocated memory, if any */
516 	free(lc->lro_mbuf_data, M_LRO);
517 	lc->lro_mbuf_data = NULL;
518 }
519 
520 static uint16_t
521 tcp_lro_rx_csum_tcphdr(const struct tcphdr *th)
522 {
523 	const uint16_t *ptr;
524 	uint32_t csum;
525 	uint16_t len;
526 
527 	csum = -th->th_sum;	/* exclude checksum field */
528 	len = th->th_off;
529 	ptr = (const uint16_t *)th;
530 	while (len--) {
531 		csum += *ptr;
532 		ptr++;
533 		csum += *ptr;
534 		ptr++;
535 	}
536 	while (csum > 0xffff)
537 		csum = (csum >> 16) + (csum & 0xffff);
538 
539 	return (csum);
540 }
541 
542 static uint16_t
543 tcp_lro_rx_csum_data(const struct lro_parser *pa, uint16_t tcp_csum)
544 {
545 	uint32_t c;
546 	uint16_t cs;
547 
548 	c = tcp_csum;
549 
550 	switch (pa->data.lro_type) {
551 #ifdef INET6
552 	case LRO_TYPE_IPV6_TCP:
553 		/* Compute full pseudo IPv6 header checksum. */
554 		cs = in6_cksum_pseudo(pa->ip6, ntohs(pa->ip6->ip6_plen), pa->ip6->ip6_nxt, 0);
555 		break;
556 #endif
557 #ifdef INET
558 	case LRO_TYPE_IPV4_TCP:
559 		/* Compute full pseudo IPv4 header checsum. */
560 		cs = in_addword(ntohs(pa->ip4->ip_len) - sizeof(*pa->ip4), IPPROTO_TCP);
561 		cs = in_pseudo(pa->ip4->ip_src.s_addr, pa->ip4->ip_dst.s_addr, htons(cs));
562 		break;
563 #endif
564 	default:
565 		cs = 0;		/* Keep compiler happy. */
566 		break;
567 	}
568 
569 	/* Complement checksum. */
570 	cs = ~cs;
571 	c += cs;
572 
573 	/* Remove TCP header checksum. */
574 	cs = ~tcp_lro_rx_csum_tcphdr(pa->tcp);
575 	c += cs;
576 
577 	/* Compute checksum remainder. */
578 	while (c > 0xffff)
579 		c = (c >> 16) + (c & 0xffff);
580 
581 	return (c);
582 }
583 
584 static void
585 tcp_lro_rx_done(struct lro_ctrl *lc)
586 {
587 	struct lro_entry *le;
588 
589 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
590 		tcp_lro_active_remove(le);
591 		tcp_lro_flush(lc, le);
592 	}
593 }
594 
595 static void
596 tcp_lro_flush_active(struct lro_ctrl *lc)
597 {
598 	struct lro_entry *le, *le_tmp;
599 
600 	/*
601 	 * Walk through the list of le entries, and
602 	 * any one that does have packets flush. This
603 	 * is called because we have an inbound packet
604 	 * (e.g. SYN) that has to have all others flushed
605 	 * in front of it. Note we have to do the remove
606 	 * because tcp_lro_flush() assumes that the entry
607 	 * is being freed. This is ok it will just get
608 	 * reallocated again like it was new.
609 	 */
610 	LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
611 		if (le->m_head != NULL) {
612 			tcp_lro_active_remove(le);
613 			tcp_lro_flush(lc, le);
614 		}
615 	}
616 }
617 
618 void
619 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
620 {
621 	struct lro_entry *le, *le_tmp;
622 	uint64_t now, tov;
623 	struct bintime bt;
624 
625 	NET_EPOCH_ASSERT();
626 	if (LIST_EMPTY(&lc->lro_active))
627 		return;
628 
629 	/* get timeout time and current time in ns */
630 	binuptime(&bt);
631 	now = bintime2ns(&bt);
632 	tov = ((timeout->tv_sec * 1000000000) + (timeout->tv_usec * 1000));
633 	LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
634 		if (now >= (bintime2ns(&le->alloc_time) + tov)) {
635 			tcp_lro_active_remove(le);
636 			tcp_lro_flush(lc, le);
637 		}
638 	}
639 }
640 
641 #ifdef INET
642 static int
643 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4)
644 {
645 	uint16_t csum;
646 
647 	/* Legacy IP has a header checksum that needs to be correct. */
648 	if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) {
649 		if (__predict_false((m->m_pkthdr.csum_flags & CSUM_IP_VALID) == 0)) {
650 			lc->lro_bad_csum++;
651 			return (TCP_LRO_CANNOT);
652 		}
653 	} else {
654 		csum = in_cksum_hdr(ip4);
655 		if (__predict_false(csum != 0)) {
656 			lc->lro_bad_csum++;
657 			return (TCP_LRO_CANNOT);
658 		}
659 	}
660 	return (0);
661 }
662 #endif
663 
664 static inline void
665 tcp_lro_assign_and_checksum_16(uint16_t *ptr, uint16_t value, uint16_t *psum)
666 {
667 	uint32_t csum;
668 
669 	csum = 0xffff - *ptr + value;
670 	while (csum > 0xffff)
671 		csum = (csum >> 16) + (csum & 0xffff);
672 	*ptr = value;
673 	*psum = csum;
674 }
675 
676 static uint16_t
677 tcp_lro_update_checksum(const struct lro_parser *pa, const struct lro_entry *le,
678     uint16_t payload_len, uint16_t delta_sum)
679 {
680 	uint32_t csum;
681 	uint16_t tlen;
682 	uint16_t temp[5] = {};
683 
684 	switch (pa->data.lro_type) {
685 	case LRO_TYPE_IPV4_TCP:
686 		/* Compute new IPv4 length. */
687 		tlen = (pa->ip4->ip_hl << 2) + (pa->tcp->th_off << 2) + payload_len;
688 		tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
689 
690 		/* Subtract delta from current IPv4 checksum. */
691 		csum = pa->ip4->ip_sum + 0xffff - temp[0];
692 		while (csum > 0xffff)
693 			csum = (csum >> 16) + (csum & 0xffff);
694 		tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
695 		goto update_tcp_header;
696 
697 	case LRO_TYPE_IPV6_TCP:
698 		/* Compute new IPv6 length. */
699 		tlen = (pa->tcp->th_off << 2) + payload_len;
700 		tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
701 		goto update_tcp_header;
702 
703 	case LRO_TYPE_IPV4_UDP:
704 		/* Compute new IPv4 length. */
705 		tlen = (pa->ip4->ip_hl << 2) + sizeof(*pa->udp) + payload_len;
706 		tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
707 
708 		/* Subtract delta from current IPv4 checksum. */
709 		csum = pa->ip4->ip_sum + 0xffff - temp[0];
710 		while (csum > 0xffff)
711 			csum = (csum >> 16) + (csum & 0xffff);
712 		tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
713 		goto update_udp_header;
714 
715 	case LRO_TYPE_IPV6_UDP:
716 		/* Compute new IPv6 length. */
717 		tlen = sizeof(*pa->udp) + payload_len;
718 		tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
719 		goto update_udp_header;
720 
721 	default:
722 		return (0);
723 	}
724 
725 update_tcp_header:
726 	/* Compute current TCP header checksum. */
727 	temp[2] = tcp_lro_rx_csum_tcphdr(pa->tcp);
728 
729 	/* Incorporate the latest ACK into the TCP header. */
730 	pa->tcp->th_ack = le->ack_seq;
731 	pa->tcp->th_win = le->window;
732 
733 	/* Incorporate latest timestamp into the TCP header. */
734 	if (le->timestamp != 0) {
735 		uint32_t *ts_ptr;
736 
737 		ts_ptr = (uint32_t *)(pa->tcp + 1);
738 		ts_ptr[1] = htonl(le->tsval);
739 		ts_ptr[2] = le->tsecr;
740 	}
741 
742 	/* Compute new TCP header checksum. */
743 	temp[3] = tcp_lro_rx_csum_tcphdr(pa->tcp);
744 
745 	/* Compute new TCP checksum. */
746 	csum = pa->tcp->th_sum + 0xffff - delta_sum +
747 	    0xffff - temp[0] + 0xffff - temp[3] + temp[2];
748 	while (csum > 0xffff)
749 		csum = (csum >> 16) + (csum & 0xffff);
750 
751 	/* Assign new TCP checksum. */
752 	tcp_lro_assign_and_checksum_16(&pa->tcp->th_sum, csum, &temp[4]);
753 
754 	/* Compute all modififications affecting next checksum. */
755 	csum = temp[0] + temp[1] + 0xffff - temp[2] +
756 	    temp[3] + temp[4] + delta_sum;
757 	while (csum > 0xffff)
758 		csum = (csum >> 16) + (csum & 0xffff);
759 
760 	/* Return delta checksum to next stage, if any. */
761 	return (csum);
762 
763 update_udp_header:
764 	tlen = sizeof(*pa->udp) + payload_len;
765 	/* Assign new UDP length and compute checksum delta. */
766 	tcp_lro_assign_and_checksum_16(&pa->udp->uh_ulen, htons(tlen), &temp[2]);
767 
768 	/* Check if there is a UDP checksum. */
769 	if (__predict_false(pa->udp->uh_sum != 0)) {
770 		/* Compute new UDP checksum. */
771 		csum = pa->udp->uh_sum + 0xffff - delta_sum +
772 		    0xffff - temp[0] + 0xffff - temp[2];
773 		while (csum > 0xffff)
774 			csum = (csum >> 16) + (csum & 0xffff);
775 		/* Assign new UDP checksum. */
776 		tcp_lro_assign_and_checksum_16(&pa->udp->uh_sum, csum, &temp[3]);
777 	}
778 
779 	/* Compute all modififications affecting next checksum. */
780 	csum = temp[0] + temp[1] + temp[2] + temp[3] + delta_sum;
781 	while (csum > 0xffff)
782 		csum = (csum >> 16) + (csum & 0xffff);
783 
784 	/* Return delta checksum to next stage, if any. */
785 	return (csum);
786 }
787 
788 static void
789 tcp_flush_out_entry(struct lro_ctrl *lc, struct lro_entry *le)
790 {
791 	/* Check if we need to recompute any checksums. */
792 	if (le->needs_merge) {
793 		uint16_t csum;
794 
795 		switch (le->inner.data.lro_type) {
796 		case LRO_TYPE_IPV4_TCP:
797 			csum = tcp_lro_update_checksum(&le->inner, le,
798 			    le->m_head->m_pkthdr.lro_tcp_d_len,
799 			    le->m_head->m_pkthdr.lro_tcp_d_csum);
800 			csum = tcp_lro_update_checksum(&le->outer, NULL,
801 			    le->m_head->m_pkthdr.lro_tcp_d_len +
802 			    le->inner.total_hdr_len, csum);
803 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
804 			    CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
805 			le->m_head->m_pkthdr.csum_data = 0xffff;
806 			if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
807 				le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
808 			break;
809 		case LRO_TYPE_IPV6_TCP:
810 			csum = tcp_lro_update_checksum(&le->inner, le,
811 			    le->m_head->m_pkthdr.lro_tcp_d_len,
812 			    le->m_head->m_pkthdr.lro_tcp_d_csum);
813 			csum = tcp_lro_update_checksum(&le->outer, NULL,
814 			    le->m_head->m_pkthdr.lro_tcp_d_len +
815 			    le->inner.total_hdr_len, csum);
816 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
817 			    CSUM_PSEUDO_HDR;
818 			le->m_head->m_pkthdr.csum_data = 0xffff;
819 			if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
820 				le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
821 			break;
822 		case LRO_TYPE_NONE:
823 			switch (le->outer.data.lro_type) {
824 			case LRO_TYPE_IPV4_TCP:
825 				csum = tcp_lro_update_checksum(&le->outer, le,
826 				    le->m_head->m_pkthdr.lro_tcp_d_len,
827 				    le->m_head->m_pkthdr.lro_tcp_d_csum);
828 				le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
829 				    CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
830 				le->m_head->m_pkthdr.csum_data = 0xffff;
831 				if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
832 					le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
833 				break;
834 			case LRO_TYPE_IPV6_TCP:
835 				csum = tcp_lro_update_checksum(&le->outer, le,
836 				    le->m_head->m_pkthdr.lro_tcp_d_len,
837 				    le->m_head->m_pkthdr.lro_tcp_d_csum);
838 				le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
839 				    CSUM_PSEUDO_HDR;
840 				le->m_head->m_pkthdr.csum_data = 0xffff;
841 				if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
842 					le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
843 				break;
844 			default:
845 				break;
846 			}
847 			break;
848 		default:
849 			break;
850 		}
851 	}
852 
853 	/*
854 	 * Break any chain, this is not set to NULL on the singleton
855 	 * case m_nextpkt points to m_head. Other case set them
856 	 * m_nextpkt to NULL in push_and_replace.
857 	 */
858 	le->m_head->m_nextpkt = NULL;
859 	lc->lro_queued += le->m_head->m_pkthdr.lro_nsegs;
860 	(*lc->ifp->if_input)(lc->ifp, le->m_head);
861 }
862 
863 static void
864 tcp_set_entry_to_mbuf(struct lro_ctrl *lc, struct lro_entry *le,
865     struct mbuf *m, struct tcphdr *th)
866 {
867 	uint32_t *ts_ptr;
868 	uint16_t tcp_data_len;
869 	uint16_t tcp_opt_len;
870 
871 	ts_ptr = (uint32_t *)(th + 1);
872 	tcp_opt_len = (th->th_off << 2);
873 	tcp_opt_len -= sizeof(*th);
874 
875 	/* Check if there is a timestamp option. */
876 	if (tcp_opt_len == 0 ||
877 	    __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
878 	    *ts_ptr != TCP_LRO_TS_OPTION)) {
879 		/* We failed to find the timestamp option. */
880 		le->timestamp = 0;
881 	} else {
882 		le->timestamp = 1;
883 		le->tsval = ntohl(*(ts_ptr + 1));
884 		le->tsecr = *(ts_ptr + 2);
885 	}
886 
887 	tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
888 
889 	/* Pull out TCP sequence numbers and window size. */
890 	le->next_seq = ntohl(th->th_seq) + tcp_data_len;
891 	le->ack_seq = th->th_ack;
892 	le->window = th->th_win;
893 	le->flags = tcp_get_flags(th);
894 	le->needs_merge = 0;
895 
896 	/* Setup new data pointers. */
897 	le->m_head = m;
898 	le->m_tail = m_last(m);
899 }
900 
901 static void
902 tcp_push_and_replace(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m)
903 {
904 	struct lro_parser *pa;
905 
906 	/*
907 	 * Push up the stack of the current entry
908 	 * and replace it with "m".
909 	 */
910 	struct mbuf *msave;
911 
912 	/* Grab off the next and save it */
913 	msave = le->m_head->m_nextpkt;
914 	le->m_head->m_nextpkt = NULL;
915 
916 	/* Now push out the old entry */
917 	tcp_flush_out_entry(lc, le);
918 
919 	/* Re-parse new header, should not fail. */
920 	pa = tcp_lro_parser(m, &le->outer, &le->inner, false);
921 	KASSERT(pa != NULL,
922 	    ("tcp_push_and_replace: LRO parser failed on m=%p\n", m));
923 
924 	/*
925 	 * Now to replace the data properly in the entry
926 	 * we have to reset the TCP header and
927 	 * other fields.
928 	 */
929 	tcp_set_entry_to_mbuf(lc, le, m, pa->tcp);
930 
931 	/* Restore the next list */
932 	m->m_nextpkt = msave;
933 }
934 
935 static void
936 tcp_lro_mbuf_append_pkthdr(struct lro_entry *le, const struct mbuf *p)
937 {
938 	struct mbuf *m;
939 	uint32_t csum;
940 
941 	m = le->m_head;
942 	if (m->m_pkthdr.lro_nsegs == 1) {
943 		/* Compute relative checksum. */
944 		csum = p->m_pkthdr.lro_tcp_d_csum;
945 	} else {
946 		/* Merge TCP data checksums. */
947 		csum = (uint32_t)m->m_pkthdr.lro_tcp_d_csum +
948 		    (uint32_t)p->m_pkthdr.lro_tcp_d_csum;
949 		while (csum > 0xffff)
950 			csum = (csum >> 16) + (csum & 0xffff);
951 	}
952 
953 	/* Update various counters. */
954 	m->m_pkthdr.len += p->m_pkthdr.lro_tcp_d_len;
955 	m->m_pkthdr.lro_tcp_d_csum = csum;
956 	m->m_pkthdr.lro_tcp_d_len += p->m_pkthdr.lro_tcp_d_len;
957 	m->m_pkthdr.lro_nsegs += p->m_pkthdr.lro_nsegs;
958 	le->needs_merge = 1;
959 }
960 
961 static void
962 tcp_lro_condense(struct lro_ctrl *lc, struct lro_entry *le)
963 {
964 	/*
965 	 * Walk through the mbuf chain we
966 	 * have on tap and compress/condense
967 	 * as required.
968 	 */
969 	uint32_t *ts_ptr;
970 	struct mbuf *m;
971 	struct tcphdr *th;
972 	uint32_t tcp_data_len_total;
973 	uint32_t tcp_data_seg_total;
974 	uint16_t tcp_data_len;
975 	uint16_t tcp_opt_len;
976 
977 	/*
978 	 * First we must check the lead (m_head)
979 	 * we must make sure that it is *not*
980 	 * something that should be sent up
981 	 * right away (sack etc).
982 	 */
983 again:
984 	m = le->m_head->m_nextpkt;
985 	if (m == NULL) {
986 		/* Just one left. */
987 		return;
988 	}
989 
990 	th = tcp_lro_get_th(m);
991 	tcp_opt_len = (th->th_off << 2);
992 	tcp_opt_len -= sizeof(*th);
993 	ts_ptr = (uint32_t *)(th + 1);
994 
995 	if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
996 	    *ts_ptr != TCP_LRO_TS_OPTION)) {
997 		/*
998 		 * Its not the timestamp. We can't
999 		 * use this guy as the head.
1000 		 */
1001 		le->m_head->m_nextpkt = m->m_nextpkt;
1002 		tcp_push_and_replace(lc, le, m);
1003 		goto again;
1004 	}
1005 	if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1006 		/*
1007 		 * Make sure that previously seen segments/ACKs are delivered
1008 		 * before this segment, e.g. FIN.
1009 		 */
1010 		le->m_head->m_nextpkt = m->m_nextpkt;
1011 		tcp_push_and_replace(lc, le, m);
1012 		goto again;
1013 	}
1014 	while((m = le->m_head->m_nextpkt) != NULL) {
1015 		/*
1016 		 * condense m into le, first
1017 		 * pull m out of the list.
1018 		 */
1019 		le->m_head->m_nextpkt = m->m_nextpkt;
1020 		m->m_nextpkt = NULL;
1021 		/* Setup my data */
1022 		tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
1023 		th = tcp_lro_get_th(m);
1024 		ts_ptr = (uint32_t *)(th + 1);
1025 		tcp_opt_len = (th->th_off << 2);
1026 		tcp_opt_len -= sizeof(*th);
1027 		tcp_data_len_total = le->m_head->m_pkthdr.lro_tcp_d_len + tcp_data_len;
1028 		tcp_data_seg_total = le->m_head->m_pkthdr.lro_nsegs + m->m_pkthdr.lro_nsegs;
1029 
1030 		if (tcp_data_seg_total >= lc->lro_ackcnt_lim ||
1031 		    tcp_data_len_total >= lc->lro_length_lim) {
1032 			/* Flush now if appending will result in overflow. */
1033 			tcp_push_and_replace(lc, le, m);
1034 			goto again;
1035 		}
1036 		if (tcp_opt_len != 0 &&
1037 		    __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1038 		    *ts_ptr != TCP_LRO_TS_OPTION)) {
1039 			/*
1040 			 * Maybe a sack in the new one? We need to
1041 			 * start all over after flushing the
1042 			 * current le. We will go up to the beginning
1043 			 * and flush it (calling the replace again possibly
1044 			 * or just returning).
1045 			 */
1046 			tcp_push_and_replace(lc, le, m);
1047 			goto again;
1048 		}
1049 		if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1050 			tcp_push_and_replace(lc, le, m);
1051 			goto again;
1052 		}
1053 		if (tcp_opt_len != 0) {
1054 			uint32_t tsval = ntohl(*(ts_ptr + 1));
1055 			/* Make sure timestamp values are increasing. */
1056 			if (TSTMP_GT(le->tsval, tsval))  {
1057 				tcp_push_and_replace(lc, le, m);
1058 				goto again;
1059 			}
1060 			le->tsval = tsval;
1061 			le->tsecr = *(ts_ptr + 2);
1062 		}
1063 		/* Try to append the new segment. */
1064 		if (__predict_false(ntohl(th->th_seq) != le->next_seq ||
1065 				    ((tcp_get_flags(th) & TH_ACK) !=
1066 				      (le->flags & TH_ACK)) ||
1067 				    (tcp_data_len == 0 &&
1068 				     le->ack_seq == th->th_ack &&
1069 				     le->window == th->th_win))) {
1070 			/* Out of order packet, non-ACK + ACK or dup ACK. */
1071 			tcp_push_and_replace(lc, le, m);
1072 			goto again;
1073 		}
1074 		if (tcp_data_len != 0 ||
1075 		    SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1076 			le->next_seq += tcp_data_len;
1077 			le->ack_seq = th->th_ack;
1078 			le->window = th->th_win;
1079 			le->needs_merge = 1;
1080 		} else if (th->th_ack == le->ack_seq) {
1081 			if (WIN_GT(th->th_win, le->window)) {
1082 				le->window = th->th_win;
1083 				le->needs_merge = 1;
1084 			}
1085 		}
1086 
1087 		if (tcp_data_len == 0) {
1088 			m_freem(m);
1089 			continue;
1090 		}
1091 
1092 		/* Merge TCP data checksum and length to head mbuf. */
1093 		tcp_lro_mbuf_append_pkthdr(le, m);
1094 
1095 		/*
1096 		 * Adjust the mbuf so that m_data points to the first byte of
1097 		 * the ULP payload.  Adjust the mbuf to avoid complications and
1098 		 * append new segment to existing mbuf chain.
1099 		 */
1100 		m_adj(m, m->m_pkthdr.len - tcp_data_len);
1101 		m_demote_pkthdr(m);
1102 		le->m_tail->m_next = m;
1103 		le->m_tail = m_last(m);
1104 	}
1105 }
1106 
1107 void
1108 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
1109 {
1110 
1111 	/* Only optimise if there are multiple packets waiting. */
1112 	NET_EPOCH_ASSERT();
1113 	if (tcp_lro_flush_tcphpts == NULL ||
1114 	    tcp_lro_flush_tcphpts(lc, le) != 0) {
1115 		tcp_lro_condense(lc, le);
1116 		tcp_flush_out_entry(lc, le);
1117 	}
1118 	lc->lro_flushed++;
1119 	bzero(le, sizeof(*le));
1120 	LIST_INSERT_HEAD(&lc->lro_free, le, next);
1121 }
1122 
1123 #define	tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
1124 
1125 /*
1126  * The tcp_lro_sort() routine is comparable to qsort(), except it has
1127  * a worst case complexity limit of O(MIN(N,64)*N), where N is the
1128  * number of elements to sort and 64 is the number of sequence bits
1129  * available. The algorithm is bit-slicing the 64-bit sequence number,
1130  * sorting one bit at a time from the most significant bit until the
1131  * least significant one, skipping the constant bits. This is
1132  * typically called a radix sort.
1133  */
1134 static void
1135 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
1136 {
1137 	struct lro_mbuf_sort temp;
1138 	uint64_t ones;
1139 	uint64_t zeros;
1140 	uint32_t x;
1141 	uint32_t y;
1142 
1143 repeat:
1144 	/* for small arrays insertion sort is faster */
1145 	if (size <= 12) {
1146 		for (x = 1; x < size; x++) {
1147 			temp = parray[x];
1148 			for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
1149 				parray[y] = parray[y - 1];
1150 			parray[y] = temp;
1151 		}
1152 		return;
1153 	}
1154 
1155 	/* compute sequence bits which are constant */
1156 	ones = 0;
1157 	zeros = 0;
1158 	for (x = 0; x != size; x++) {
1159 		ones |= parray[x].seq;
1160 		zeros |= ~parray[x].seq;
1161 	}
1162 
1163 	/* compute bits which are not constant into "ones" */
1164 	ones &= zeros;
1165 	if (ones == 0)
1166 		return;
1167 
1168 	/* pick the most significant bit which is not constant */
1169 	ones = tcp_lro_msb_64(ones);
1170 
1171 	/*
1172 	 * Move entries having cleared sequence bits to the beginning
1173 	 * of the array:
1174 	 */
1175 	for (x = y = 0; y != size; y++) {
1176 		/* skip set bits */
1177 		if (parray[y].seq & ones)
1178 			continue;
1179 		/* swap entries */
1180 		temp = parray[x];
1181 		parray[x] = parray[y];
1182 		parray[y] = temp;
1183 		x++;
1184 	}
1185 
1186 	KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
1187 
1188 	/* sort zeros */
1189 	tcp_lro_sort(parray, x);
1190 
1191 	/* sort ones */
1192 	parray += x;
1193 	size -= x;
1194 	goto repeat;
1195 }
1196 
1197 void
1198 tcp_lro_flush_all(struct lro_ctrl *lc)
1199 {
1200 	uint64_t seq;
1201 	uint64_t nseq;
1202 	unsigned x;
1203 
1204 	NET_EPOCH_ASSERT();
1205 	/* check if no mbufs to flush */
1206 	if (lc->lro_mbuf_count == 0)
1207 		goto done;
1208 	if (lc->lro_cpu_is_set == 0) {
1209 		if (lc->lro_last_cpu == curcpu) {
1210 			lc->lro_cnt_of_same_cpu++;
1211 			/* Have we reached the threshold to declare a cpu? */
1212 			if (lc->lro_cnt_of_same_cpu > tcp_lro_cpu_set_thresh)
1213 				lc->lro_cpu_is_set = 1;
1214 		} else {
1215 			lc->lro_last_cpu = curcpu;
1216 			lc->lro_cnt_of_same_cpu = 0;
1217 		}
1218 	}
1219 	CURVNET_SET(lc->ifp->if_vnet);
1220 
1221 	/* get current time */
1222 	binuptime(&lc->lro_last_queue_time);
1223 
1224 	/* sort all mbufs according to stream */
1225 	tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
1226 
1227 	/* input data into LRO engine, stream by stream */
1228 	seq = 0;
1229 	for (x = 0; x != lc->lro_mbuf_count; x++) {
1230 		struct mbuf *mb;
1231 
1232 		/* get mbuf */
1233 		mb = lc->lro_mbuf_data[x].mb;
1234 
1235 		/* get sequence number, masking away the packet index */
1236 		nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
1237 
1238 		/* check for new stream */
1239 		if (seq != nseq) {
1240 			seq = nseq;
1241 
1242 			/* flush active streams */
1243 			tcp_lro_rx_done(lc);
1244 		}
1245 
1246 		/* add packet to LRO engine */
1247 		if (tcp_lro_rx_common(lc, mb, 0, false) != 0) {
1248  			/* Flush anything we have acummulated */
1249  			tcp_lro_flush_active(lc);
1250 			/* input packet to network layer */
1251 			(*lc->ifp->if_input)(lc->ifp, mb);
1252 			lc->lro_queued++;
1253 			lc->lro_flushed++;
1254 		}
1255 	}
1256 	CURVNET_RESTORE();
1257 done:
1258 	/* flush active streams */
1259 	tcp_lro_rx_done(lc);
1260 	tcp_hpts_softclock();
1261 	lc->lro_mbuf_count = 0;
1262 }
1263 
1264 static struct lro_head *
1265 tcp_lro_rx_get_bucket(struct lro_ctrl *lc, struct mbuf *m, struct lro_parser *parser)
1266 {
1267 	u_long hash;
1268 
1269 	if (M_HASHTYPE_ISHASH(m)) {
1270 		hash = m->m_pkthdr.flowid;
1271 	} else {
1272 		for (unsigned i = hash = 0; i != LRO_RAW_ADDRESS_MAX; i++)
1273 			hash += parser->data.raw[i];
1274 	}
1275 	return (&lc->lro_hash[hash % lc->lro_hashsz]);
1276 }
1277 
1278 static int
1279 tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash)
1280 {
1281 	struct lro_parser pi;	/* inner address data */
1282 	struct lro_parser po;	/* outer address data */
1283 	struct lro_parser *pa;	/* current parser for TCP stream */
1284 	struct lro_entry *le;
1285 	struct lro_head *bucket;
1286 	struct tcphdr *th;
1287 	int tcp_data_len;
1288 	int tcp_opt_len;
1289 	int error;
1290 	uint16_t tcp_data_sum;
1291 
1292 #ifdef INET
1293 	/* Quickly decide if packet cannot be LRO'ed */
1294 	if (__predict_false(V_ipforwarding != 0))
1295 		return (TCP_LRO_CANNOT);
1296 #endif
1297 #ifdef INET6
1298 	/* Quickly decide if packet cannot be LRO'ed */
1299 	if (__predict_false(V_ip6_forwarding != 0))
1300 		return (TCP_LRO_CANNOT);
1301 #endif
1302 	if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1303 	     ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1304 	    (m->m_pkthdr.csum_data != 0xffff)) {
1305 		/*
1306 		 * The checksum either did not have hardware offload
1307 		 * or it was a bad checksum. We can't LRO such
1308 		 * a packet.
1309 		 */
1310 		counter_u64_add(tcp_bad_csums, 1);
1311 		return (TCP_LRO_CANNOT);
1312 	}
1313 	/* We expect a contiguous header [eh, ip, tcp]. */
1314 	pa = tcp_lro_parser(m, &po, &pi, true);
1315 	if (__predict_false(pa == NULL))
1316 		return (TCP_LRO_NOT_SUPPORTED);
1317 
1318 	/* We don't expect any padding. */
1319 	error = tcp_lro_trim_mbuf_chain(m, pa);
1320 	if (__predict_false(error != 0))
1321 		return (error);
1322 
1323 #ifdef INET
1324 	switch (pa->data.lro_type) {
1325 	case LRO_TYPE_IPV4_TCP:
1326 		error = tcp_lro_rx_ipv4(lc, m, pa->ip4);
1327 		if (__predict_false(error != 0))
1328 			return (error);
1329 		break;
1330 	default:
1331 		break;
1332 	}
1333 #endif
1334 	/* If no hardware or arrival stamp on the packet add timestamp */
1335 	if ((m->m_flags & (M_TSTMP_LRO | M_TSTMP)) == 0) {
1336 		m->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
1337 		m->m_flags |= M_TSTMP_LRO;
1338 	}
1339 
1340 	/* Get pointer to TCP header. */
1341 	th = pa->tcp;
1342 
1343 	/* Don't process SYN packets. */
1344 	if (__predict_false(tcp_get_flags(th) & TH_SYN))
1345 		return (TCP_LRO_CANNOT);
1346 
1347 	/* Get total TCP header length and compute payload length. */
1348 	tcp_opt_len = (th->th_off << 2);
1349 	tcp_data_len = m->m_pkthdr.len - ((uint8_t *)th -
1350 	    (uint8_t *)m->m_data) - tcp_opt_len;
1351 	tcp_opt_len -= sizeof(*th);
1352 
1353 	/* Don't process invalid TCP headers. */
1354 	if (__predict_false(tcp_opt_len < 0 || tcp_data_len < 0))
1355 		return (TCP_LRO_CANNOT);
1356 
1357 	/* Compute TCP data only checksum. */
1358 	if (tcp_data_len == 0)
1359 		tcp_data_sum = 0;	/* no data, no checksum */
1360 	else if (__predict_false(csum != 0))
1361 		tcp_data_sum = tcp_lro_rx_csum_data(pa, ~csum);
1362 	else
1363 		tcp_data_sum = tcp_lro_rx_csum_data(pa, ~th->th_sum);
1364 
1365 	/* Save TCP info in mbuf. */
1366 	m->m_nextpkt = NULL;
1367 	m->m_pkthdr.rcvif = lc->ifp;
1368 	m->m_pkthdr.lro_tcp_d_csum = tcp_data_sum;
1369 	m->m_pkthdr.lro_tcp_d_len = tcp_data_len;
1370 	m->m_pkthdr.lro_tcp_h_off = ((uint8_t *)th - (uint8_t *)m->m_data);
1371 	m->m_pkthdr.lro_nsegs = 1;
1372 
1373 	/* Get hash bucket. */
1374 	if (!use_hash) {
1375 		bucket = &lc->lro_hash[0];
1376 	} else {
1377 		bucket = tcp_lro_rx_get_bucket(lc, m, pa);
1378 	}
1379 
1380 	/* Try to find a matching previous segment. */
1381 	LIST_FOREACH(le, bucket, hash_next) {
1382 		/* Compare addresses and ports. */
1383 		if (lro_address_compare(&po.data, &le->outer.data) == false ||
1384 		    lro_address_compare(&pi.data, &le->inner.data) == false)
1385 			continue;
1386 
1387 		/* Check if no data and old ACK. */
1388 		if (tcp_data_len == 0 &&
1389 		    SEQ_LT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1390 			m_freem(m);
1391 			return (0);
1392 		}
1393 
1394 		/* Mark "m" in the last spot. */
1395 		le->m_last_mbuf->m_nextpkt = m;
1396 		/* Now set the tail to "m". */
1397 		le->m_last_mbuf = m;
1398 		return (0);
1399 	}
1400 
1401 	/* Try to find an empty slot. */
1402 	if (LIST_EMPTY(&lc->lro_free))
1403 		return (TCP_LRO_NO_ENTRIES);
1404 
1405 	/* Start a new segment chain. */
1406 	le = LIST_FIRST(&lc->lro_free);
1407 	LIST_REMOVE(le, next);
1408 	tcp_lro_active_insert(lc, bucket, le);
1409 
1410 	/* Make sure the headers are set. */
1411 	le->inner = pi;
1412 	le->outer = po;
1413 
1414 	/* Store time this entry was allocated. */
1415 	le->alloc_time = lc->lro_last_queue_time;
1416 
1417 	tcp_set_entry_to_mbuf(lc, le, m, th);
1418 
1419 	/* Now set the tail to "m". */
1420 	le->m_last_mbuf = m;
1421 
1422 	return (0);
1423 }
1424 
1425 int
1426 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
1427 {
1428 	int error;
1429 
1430 	if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1431 	     ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1432 	    (m->m_pkthdr.csum_data != 0xffff)) {
1433 		/*
1434 		 * The checksum either did not have hardware offload
1435 		 * or it was a bad checksum. We can't LRO such
1436 		 * a packet.
1437 		 */
1438 		counter_u64_add(tcp_bad_csums, 1);
1439 		return (TCP_LRO_CANNOT);
1440 	}
1441 	/* get current time */
1442 	binuptime(&lc->lro_last_queue_time);
1443 	CURVNET_SET(lc->ifp->if_vnet);
1444 	error = tcp_lro_rx_common(lc, m, csum, true);
1445 	if (__predict_false(error != 0)) {
1446 		/*
1447 		 * Flush anything we have acummulated
1448 		 * ahead of this packet that can't
1449 		 * be LRO'd. This preserves order.
1450 		 */
1451 		tcp_lro_flush_active(lc);
1452 	}
1453 	CURVNET_RESTORE();
1454 
1455 	return (error);
1456 }
1457 
1458 void
1459 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
1460 {
1461 	NET_EPOCH_ASSERT();
1462 	/* sanity checks */
1463 	if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
1464 	    lc->lro_mbuf_max == 0)) {
1465 		/* packet drop */
1466 		m_freem(mb);
1467 		return;
1468 	}
1469 
1470 	/* check if packet is not LRO capable */
1471 	if (__predict_false((lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
1472 		/* input packet to network layer */
1473 		(*lc->ifp->if_input) (lc->ifp, mb);
1474 		return;
1475 	}
1476 
1477  	/* If no hardware or arrival stamp on the packet add timestamp */
1478  	if ((tcplro_stacks_wanting_mbufq > 0) &&
1479  	    (tcp_less_accurate_lro_ts == 0) &&
1480  	    ((mb->m_flags & M_TSTMP) == 0)) {
1481  		/* Add in an LRO time since no hardware */
1482  		binuptime(&lc->lro_last_queue_time);
1483  		mb->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
1484  		mb->m_flags |= M_TSTMP_LRO;
1485  	}
1486 
1487 	/* create sequence number */
1488 	lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
1489 	    (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
1490 	    (((uint64_t)mb->m_pkthdr.flowid) << 24) |
1491 	    ((uint64_t)lc->lro_mbuf_count);
1492 
1493 	/* enter mbuf */
1494 	lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb;
1495 
1496 	/* flush if array is full */
1497 	if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max))
1498 		tcp_lro_flush_all(lc);
1499 }
1500 
1501 /* end */
1502