xref: /freebsd/sys/netinet/tcp_lro.c (revision e47161e5f1f01ef300c6e7efdb9c92e3a6c497ff)
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 static void	tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le);
87 
88 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro,  CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
89     "TCP LRO");
90 
91 long tcplro_stacks_wanting_mbufq;
92 int	(*tcp_lro_flush_tcphpts)(struct lro_ctrl *lc, struct lro_entry *le);
93 
94 counter_u64_t tcp_inp_lro_direct_queue;
95 counter_u64_t tcp_inp_lro_wokeup_queue;
96 counter_u64_t tcp_inp_lro_compressed;
97 counter_u64_t tcp_inp_lro_locks_taken;
98 counter_u64_t tcp_extra_mbuf;
99 counter_u64_t tcp_would_have_but;
100 counter_u64_t tcp_comp_total;
101 counter_u64_t tcp_uncomp_total;
102 counter_u64_t tcp_bad_csums;
103 
104 static unsigned	tcp_lro_entries = TCP_LRO_ENTRIES;
105 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries,
106     CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_entries, 0,
107     "default number of LRO entries");
108 
109 static uint32_t tcp_lro_cpu_set_thresh = TCP_LRO_CPU_DECLARATION_THRESH;
110 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_cpu_threshold,
111     CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_cpu_set_thresh, 0,
112     "Number of interrupts in a row on the same CPU that will make us declare an 'affinity' cpu?");
113 
114 static uint32_t tcp_less_accurate_lro_ts = 0;
115 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_less_accurate,
116     CTLFLAG_MPSAFE, &tcp_less_accurate_lro_ts, 0,
117     "Do we trade off efficency by doing less timestamp operations for time accuracy?");
118 
119 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, fullqueue, CTLFLAG_RD,
120     &tcp_inp_lro_direct_queue, "Number of lro's fully queued to transport");
121 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, wokeup, CTLFLAG_RD,
122     &tcp_inp_lro_wokeup_queue, "Number of lro's where we woke up transport via hpts");
123 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, compressed, CTLFLAG_RD,
124     &tcp_inp_lro_compressed, "Number of lro's compressed and sent to transport");
125 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lockcnt, CTLFLAG_RD,
126     &tcp_inp_lro_locks_taken, "Number of lro's inp_wlocks taken");
127 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, extra_mbuf, CTLFLAG_RD,
128     &tcp_extra_mbuf, "Number of times we had an extra compressed ack dropped into the tp");
129 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, would_have_but, CTLFLAG_RD,
130     &tcp_would_have_but, "Number of times we would have had an extra compressed, but mget failed");
131 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, with_m_ackcmp, CTLFLAG_RD,
132     &tcp_comp_total, "Number of mbufs queued with M_ACKCMP flags set");
133 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, without_m_ackcmp, CTLFLAG_RD,
134     &tcp_uncomp_total, "Number of mbufs queued without M_ACKCMP");
135 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lro_badcsum, CTLFLAG_RD,
136     &tcp_bad_csums, "Number of packets that the common code saw with bad csums");
137 
138 void
139 tcp_lro_reg_mbufq(void)
140 {
141 	atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, 1);
142 }
143 
144 void
145 tcp_lro_dereg_mbufq(void)
146 {
147 	atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, -1);
148 }
149 
150 static __inline void
151 tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket,
152     struct lro_entry *le)
153 {
154 
155 	LIST_INSERT_HEAD(&lc->lro_active, le, next);
156 	LIST_INSERT_HEAD(bucket, le, hash_next);
157 }
158 
159 static __inline void
160 tcp_lro_active_remove(struct lro_entry *le)
161 {
162 
163 	LIST_REMOVE(le, next);		/* active list */
164 	LIST_REMOVE(le, hash_next);	/* hash bucket */
165 }
166 
167 int
168 tcp_lro_init(struct lro_ctrl *lc)
169 {
170 	return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0));
171 }
172 
173 int
174 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
175     unsigned lro_entries, unsigned lro_mbufs)
176 {
177 	struct lro_entry *le;
178 	size_t size;
179 	unsigned i;
180 
181 	lc->lro_bad_csum = 0;
182 	lc->lro_queued = 0;
183 	lc->lro_flushed = 0;
184 	lc->lro_mbuf_count = 0;
185 	lc->lro_mbuf_max = lro_mbufs;
186 	lc->lro_cnt = lro_entries;
187 	lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
188 	lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
189 	lc->ifp = ifp;
190 	LIST_INIT(&lc->lro_free);
191 	LIST_INIT(&lc->lro_active);
192 
193 	/* create hash table to accelerate entry lookup */
194 	lc->lro_hash = phashinit_flags(lro_entries, M_LRO, &lc->lro_hashsz,
195 	    HASH_NOWAIT);
196 	if (lc->lro_hash == NULL) {
197 		memset(lc, 0, sizeof(*lc));
198 		return (ENOMEM);
199 	}
200 
201 	/* compute size to allocate */
202 	size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
203 	    (lro_entries * sizeof(*le));
204 	lc->lro_mbuf_data = (struct lro_mbuf_sort *)
205 	    malloc(size, M_LRO, M_NOWAIT | M_ZERO);
206 
207 	/* check for out of memory */
208 	if (lc->lro_mbuf_data == NULL) {
209 		free(lc->lro_hash, M_LRO);
210 		memset(lc, 0, sizeof(*lc));
211 		return (ENOMEM);
212 	}
213 	/* compute offset for LRO entries */
214 	le = (struct lro_entry *)
215 	    (lc->lro_mbuf_data + lro_mbufs);
216 
217 	/* setup linked list */
218 	for (i = 0; i != lro_entries; i++)
219 		LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
220 
221 	return (0);
222 }
223 
224 struct vxlan_header {
225 	uint32_t	vxlh_flags;
226 	uint32_t	vxlh_vni;
227 };
228 
229 static inline void *
230 tcp_lro_low_level_parser(void *ptr, struct lro_parser *parser, bool update_data, bool is_vxlan, int mlen)
231 {
232 	const struct ether_vlan_header *eh;
233 	void *old;
234 	uint16_t eth_type;
235 
236 	if (update_data)
237 		memset(parser, 0, sizeof(*parser));
238 
239 	old = ptr;
240 
241 	if (is_vxlan) {
242 		const struct vxlan_header *vxh;
243 		vxh = ptr;
244 		ptr = (uint8_t *)ptr + sizeof(*vxh);
245 		if (update_data) {
246 			parser->data.vxlan_vni =
247 			    vxh->vxlh_vni & htonl(0xffffff00);
248 		}
249 	}
250 
251 	eh = ptr;
252 	if (__predict_false(eh->evl_encap_proto == htons(ETHERTYPE_VLAN))) {
253 		eth_type = eh->evl_proto;
254 		if (update_data) {
255 			/* strip priority and keep VLAN ID only */
256 			parser->data.vlan_id = eh->evl_tag & htons(EVL_VLID_MASK);
257 		}
258 		/* advance to next header */
259 		ptr = (uint8_t *)ptr + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
260 		mlen -= (ETHER_HDR_LEN  + ETHER_VLAN_ENCAP_LEN);
261 	} else {
262 		eth_type = eh->evl_encap_proto;
263 		/* advance to next header */
264 		mlen -= ETHER_HDR_LEN;
265 		ptr = (uint8_t *)ptr + ETHER_HDR_LEN;
266 	}
267 	if (__predict_false(mlen <= 0))
268 		return (NULL);
269 	switch (eth_type) {
270 #ifdef INET
271 	case htons(ETHERTYPE_IP):
272 		parser->ip4 = ptr;
273 		if (__predict_false(mlen < sizeof(struct ip)))
274 			return (NULL);
275 		/* Ensure there are no IPv4 options. */
276 		if ((parser->ip4->ip_hl << 2) != sizeof (*parser->ip4))
277 			break;
278 		/* .. and the packet is not fragmented. */
279 		if (parser->ip4->ip_off & htons(IP_MF|IP_OFFMASK))
280 			break;
281 		/* .. and the packet has valid src/dst addrs */
282 		if (__predict_false(parser->ip4->ip_src.s_addr == INADDR_ANY ||
283 			parser->ip4->ip_dst.s_addr == INADDR_ANY))
284 			break;
285 		ptr = (uint8_t *)ptr + (parser->ip4->ip_hl << 2);
286 		mlen -= sizeof(struct ip);
287 		if (update_data) {
288 			parser->data.s_addr.v4 = parser->ip4->ip_src;
289 			parser->data.d_addr.v4 = parser->ip4->ip_dst;
290 		}
291 		switch (parser->ip4->ip_p) {
292 		case IPPROTO_UDP:
293 			if (__predict_false(mlen < sizeof(struct udphdr)))
294 				return (NULL);
295 			parser->udp = ptr;
296 			if (update_data) {
297 				parser->data.lro_type = LRO_TYPE_IPV4_UDP;
298 				parser->data.s_port = parser->udp->uh_sport;
299 				parser->data.d_port = parser->udp->uh_dport;
300 			} else {
301 				MPASS(parser->data.lro_type == LRO_TYPE_IPV4_UDP);
302 			}
303 			ptr = ((uint8_t *)ptr + sizeof(*parser->udp));
304 			parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
305 			return (ptr);
306 		case IPPROTO_TCP:
307 			parser->tcp = ptr;
308 			if (__predict_false(mlen < sizeof(struct tcphdr)))
309 				return (NULL);
310 			if (update_data) {
311 				parser->data.lro_type = LRO_TYPE_IPV4_TCP;
312 				parser->data.s_port = parser->tcp->th_sport;
313 				parser->data.d_port = parser->tcp->th_dport;
314 			} else {
315 				MPASS(parser->data.lro_type == LRO_TYPE_IPV4_TCP);
316 			}
317 			if (__predict_false(mlen < (parser->tcp->th_off << 2)))
318 				return (NULL);
319 			ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
320 			parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
321 			return (ptr);
322 		default:
323 			break;
324 		}
325 		break;
326 #endif
327 #ifdef INET6
328 	case htons(ETHERTYPE_IPV6):
329 		parser->ip6 = ptr;
330 		if (__predict_false(mlen < sizeof(struct ip6_hdr)))
331 			return (NULL);
332 		/* Ensure the packet has valid src/dst addrs */
333 		if (__predict_false(IN6_IS_ADDR_UNSPECIFIED(&parser->ip6->ip6_src) ||
334 			IN6_IS_ADDR_UNSPECIFIED(&parser->ip6->ip6_dst)))
335 			return (NULL);
336 		ptr = (uint8_t *)ptr + sizeof(*parser->ip6);
337 		if (update_data) {
338 			parser->data.s_addr.v6 = parser->ip6->ip6_src;
339 			parser->data.d_addr.v6 = parser->ip6->ip6_dst;
340 		}
341 		mlen -= sizeof(struct ip6_hdr);
342 		switch (parser->ip6->ip6_nxt) {
343 		case IPPROTO_UDP:
344 			if (__predict_false(mlen < sizeof(struct udphdr)))
345 				return (NULL);
346 			parser->udp = ptr;
347 			if (update_data) {
348 				parser->data.lro_type = LRO_TYPE_IPV6_UDP;
349 				parser->data.s_port = parser->udp->uh_sport;
350 				parser->data.d_port = parser->udp->uh_dport;
351 			} else {
352 				MPASS(parser->data.lro_type == LRO_TYPE_IPV6_UDP);
353 			}
354 			ptr = (uint8_t *)ptr + sizeof(*parser->udp);
355 			parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
356 			return (ptr);
357 		case IPPROTO_TCP:
358 			if (__predict_false(mlen < sizeof(struct tcphdr)))
359 				return (NULL);
360 			parser->tcp = ptr;
361 			if (update_data) {
362 				parser->data.lro_type = LRO_TYPE_IPV6_TCP;
363 				parser->data.s_port = parser->tcp->th_sport;
364 				parser->data.d_port = parser->tcp->th_dport;
365 			} else {
366 				MPASS(parser->data.lro_type == LRO_TYPE_IPV6_TCP);
367 			}
368 			if (__predict_false(mlen < (parser->tcp->th_off << 2)))
369 				return (NULL);
370 			ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
371 			parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
372 			return (ptr);
373 		default:
374 			break;
375 		}
376 		break;
377 #endif
378 	default:
379 		break;
380 	}
381 	/* Invalid packet - cannot parse */
382 	return (NULL);
383 }
384 
385 static const int vxlan_csum = CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
386     CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID;
387 
388 static inline struct lro_parser *
389 tcp_lro_parser(struct mbuf *m, struct lro_parser *po, struct lro_parser *pi, bool update_data)
390 {
391 	void *data_ptr;
392 
393 	/* Try to parse outer headers first. */
394 	data_ptr = tcp_lro_low_level_parser(m->m_data, po, update_data, false, m->m_len);
395 	if (data_ptr == NULL || po->total_hdr_len > m->m_len)
396 		return (NULL);
397 
398 	if (update_data) {
399 		/* Store VLAN ID, if any. */
400 		if (__predict_false(m->m_flags & M_VLANTAG)) {
401 			po->data.vlan_id =
402 			    htons(m->m_pkthdr.ether_vtag) & htons(EVL_VLID_MASK);
403 		}
404 		/* Store decrypted flag, if any. */
405 		if (__predict_false((m->m_pkthdr.csum_flags &
406 		    CSUM_TLS_MASK) == CSUM_TLS_DECRYPTED))
407 			po->data.lro_flags |= LRO_FLAG_DECRYPTED;
408 	}
409 
410 	switch (po->data.lro_type) {
411 	case LRO_TYPE_IPV4_UDP:
412 	case LRO_TYPE_IPV6_UDP:
413 		/* Check for VXLAN headers. */
414 		if ((m->m_pkthdr.csum_flags & vxlan_csum) != vxlan_csum)
415 			break;
416 
417 		/* Try to parse inner headers. */
418 		data_ptr = tcp_lro_low_level_parser(data_ptr, pi, update_data, true,
419 						    (m->m_len - ((caddr_t)data_ptr - m->m_data)));
420 		if (data_ptr == NULL || (pi->total_hdr_len + po->total_hdr_len) > m->m_len)
421 			break;
422 
423 		/* Verify supported header types. */
424 		switch (pi->data.lro_type) {
425 		case LRO_TYPE_IPV4_TCP:
426 		case LRO_TYPE_IPV6_TCP:
427 			return (pi);
428 		default:
429 			break;
430 		}
431 		break;
432 	case LRO_TYPE_IPV4_TCP:
433 	case LRO_TYPE_IPV6_TCP:
434 		if (update_data)
435 			memset(pi, 0, sizeof(*pi));
436 		return (po);
437 	default:
438 		break;
439 	}
440 	return (NULL);
441 }
442 
443 static inline int
444 tcp_lro_trim_mbuf_chain(struct mbuf *m, const struct lro_parser *po)
445 {
446 	int len;
447 
448 	switch (po->data.lro_type) {
449 #ifdef INET
450 	case LRO_TYPE_IPV4_TCP:
451 		len = ((uint8_t *)po->ip4 - (uint8_t *)m->m_data) +
452 		    ntohs(po->ip4->ip_len);
453 		break;
454 #endif
455 #ifdef INET6
456 	case LRO_TYPE_IPV6_TCP:
457 		len = ((uint8_t *)po->ip6 - (uint8_t *)m->m_data) +
458 		    ntohs(po->ip6->ip6_plen) + sizeof(*po->ip6);
459 		break;
460 #endif
461 	default:
462 		return (TCP_LRO_CANNOT);
463 	}
464 
465 	/*
466 	 * If the frame is padded beyond the end of the IP packet,
467 	 * then trim the extra bytes off:
468 	 */
469 	if (__predict_true(m->m_pkthdr.len == len)) {
470 		return (0);
471 	} else if (m->m_pkthdr.len > len) {
472 		m_adj(m, len - m->m_pkthdr.len);
473 		return (0);
474 	}
475 	return (TCP_LRO_CANNOT);
476 }
477 
478 static void
479 lro_free_mbuf_chain(struct mbuf *m)
480 {
481 	struct mbuf *save;
482 
483 	while (m) {
484 		save = m->m_nextpkt;
485 		m->m_nextpkt = NULL;
486 		m_freem(m);
487 		m = save;
488 	}
489 }
490 
491 void
492 tcp_lro_free(struct lro_ctrl *lc)
493 {
494 	struct lro_entry *le;
495 	unsigned x;
496 
497 	/* reset LRO free list */
498 	LIST_INIT(&lc->lro_free);
499 
500 	/* free active mbufs, if any */
501 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
502 		tcp_lro_active_remove(le);
503 		lro_free_mbuf_chain(le->m_head);
504 	}
505 
506 	/* free hash table */
507 	free(lc->lro_hash, M_LRO);
508 	lc->lro_hash = NULL;
509 	lc->lro_hashsz = 0;
510 
511 	/* free mbuf array, if any */
512 	for (x = 0; x != lc->lro_mbuf_count; x++)
513 		m_freem(lc->lro_mbuf_data[x].mb);
514 	lc->lro_mbuf_count = 0;
515 
516 	/* free allocated memory, if any */
517 	free(lc->lro_mbuf_data, M_LRO);
518 	lc->lro_mbuf_data = NULL;
519 }
520 
521 static uint16_t
522 tcp_lro_rx_csum_tcphdr(const struct tcphdr *th)
523 {
524 	const uint16_t *ptr;
525 	uint32_t csum;
526 	uint16_t len;
527 
528 	csum = -th->th_sum;	/* exclude checksum field */
529 	len = th->th_off;
530 	ptr = (const uint16_t *)th;
531 	while (len--) {
532 		csum += *ptr;
533 		ptr++;
534 		csum += *ptr;
535 		ptr++;
536 	}
537 	while (csum > 0xffff)
538 		csum = (csum >> 16) + (csum & 0xffff);
539 
540 	return (csum);
541 }
542 
543 static uint16_t
544 tcp_lro_rx_csum_data(const struct lro_parser *pa, uint16_t tcp_csum)
545 {
546 	uint32_t c;
547 	uint16_t cs;
548 
549 	c = tcp_csum;
550 
551 	switch (pa->data.lro_type) {
552 #ifdef INET6
553 	case LRO_TYPE_IPV6_TCP:
554 		/* Compute full pseudo IPv6 header checksum. */
555 		cs = in6_cksum_pseudo(pa->ip6, ntohs(pa->ip6->ip6_plen), pa->ip6->ip6_nxt, 0);
556 		break;
557 #endif
558 #ifdef INET
559 	case LRO_TYPE_IPV4_TCP:
560 		/* Compute full pseudo IPv4 header checsum. */
561 		cs = in_addword(ntohs(pa->ip4->ip_len) - sizeof(*pa->ip4), IPPROTO_TCP);
562 		cs = in_pseudo(pa->ip4->ip_src.s_addr, pa->ip4->ip_dst.s_addr, htons(cs));
563 		break;
564 #endif
565 	default:
566 		cs = 0;		/* Keep compiler happy. */
567 		break;
568 	}
569 
570 	/* Complement checksum. */
571 	cs = ~cs;
572 	c += cs;
573 
574 	/* Remove TCP header checksum. */
575 	cs = ~tcp_lro_rx_csum_tcphdr(pa->tcp);
576 	c += cs;
577 
578 	/* Compute checksum remainder. */
579 	while (c > 0xffff)
580 		c = (c >> 16) + (c & 0xffff);
581 
582 	return (c);
583 }
584 
585 static void
586 tcp_lro_rx_done(struct lro_ctrl *lc)
587 {
588 	struct lro_entry *le;
589 
590 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
591 		tcp_lro_active_remove(le);
592 		tcp_lro_flush(lc, le);
593 	}
594 }
595 
596 static void
597 tcp_lro_flush_active(struct lro_ctrl *lc)
598 {
599 	struct lro_entry *le, *le_tmp;
600 
601 	/*
602 	 * Walk through the list of le entries, and
603 	 * any one that does have packets flush. This
604 	 * is called because we have an inbound packet
605 	 * (e.g. SYN) that has to have all others flushed
606 	 * in front of it. Note we have to do the remove
607 	 * because tcp_lro_flush() assumes that the entry
608 	 * is being freed. This is ok it will just get
609 	 * reallocated again like it was new.
610 	 */
611 	LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
612 		if (le->m_head != NULL) {
613 			tcp_lro_active_remove(le);
614 			tcp_lro_flush(lc, le);
615 		}
616 	}
617 }
618 
619 void
620 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
621 {
622 	struct lro_entry *le, *le_tmp;
623 	uint64_t now, tov;
624 	struct bintime bt;
625 
626 	NET_EPOCH_ASSERT();
627 	if (LIST_EMPTY(&lc->lro_active))
628 		return;
629 
630 	/* get timeout time and current time in ns */
631 	binuptime(&bt);
632 	now = bintime2ns(&bt);
633 	tov = ((timeout->tv_sec * 1000000000) + (timeout->tv_usec * 1000));
634 	LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
635 		if (now >= (bintime2ns(&le->alloc_time) + tov)) {
636 			tcp_lro_active_remove(le);
637 			tcp_lro_flush(lc, le);
638 		}
639 	}
640 }
641 
642 #ifdef INET
643 static int
644 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4)
645 {
646 	uint16_t csum;
647 
648 	/* Legacy IP has a header checksum that needs to be correct. */
649 	if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) {
650 		if (__predict_false((m->m_pkthdr.csum_flags & CSUM_IP_VALID) == 0)) {
651 			lc->lro_bad_csum++;
652 			return (TCP_LRO_CANNOT);
653 		}
654 	} else {
655 		csum = in_cksum_hdr(ip4);
656 		if (__predict_false(csum != 0)) {
657 			lc->lro_bad_csum++;
658 			return (TCP_LRO_CANNOT);
659 		}
660 	}
661 	return (0);
662 }
663 #endif
664 
665 static inline void
666 tcp_lro_assign_and_checksum_16(uint16_t *ptr, uint16_t value, uint16_t *psum)
667 {
668 	uint32_t csum;
669 
670 	csum = 0xffff - *ptr + value;
671 	while (csum > 0xffff)
672 		csum = (csum >> 16) + (csum & 0xffff);
673 	*ptr = value;
674 	*psum = csum;
675 }
676 
677 static uint16_t
678 tcp_lro_update_checksum(const struct lro_parser *pa, const struct lro_entry *le,
679     uint16_t payload_len, uint16_t delta_sum)
680 {
681 	uint32_t csum;
682 	uint16_t tlen;
683 	uint16_t temp[5] = {};
684 
685 	switch (pa->data.lro_type) {
686 	case LRO_TYPE_IPV4_TCP:
687 		/* Compute new IPv4 length. */
688 		tlen = (pa->ip4->ip_hl << 2) + (pa->tcp->th_off << 2) + payload_len;
689 		tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
690 
691 		/* Subtract delta from current IPv4 checksum. */
692 		csum = pa->ip4->ip_sum + 0xffff - temp[0];
693 		while (csum > 0xffff)
694 			csum = (csum >> 16) + (csum & 0xffff);
695 		tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
696 		goto update_tcp_header;
697 
698 	case LRO_TYPE_IPV6_TCP:
699 		/* Compute new IPv6 length. */
700 		tlen = (pa->tcp->th_off << 2) + payload_len;
701 		tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
702 		goto update_tcp_header;
703 
704 	case LRO_TYPE_IPV4_UDP:
705 		/* Compute new IPv4 length. */
706 		tlen = (pa->ip4->ip_hl << 2) + sizeof(*pa->udp) + payload_len;
707 		tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
708 
709 		/* Subtract delta from current IPv4 checksum. */
710 		csum = pa->ip4->ip_sum + 0xffff - temp[0];
711 		while (csum > 0xffff)
712 			csum = (csum >> 16) + (csum & 0xffff);
713 		tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
714 		goto update_udp_header;
715 
716 	case LRO_TYPE_IPV6_UDP:
717 		/* Compute new IPv6 length. */
718 		tlen = sizeof(*pa->udp) + payload_len;
719 		tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
720 		goto update_udp_header;
721 
722 	default:
723 		return (0);
724 	}
725 
726 update_tcp_header:
727 	/* Compute current TCP header checksum. */
728 	temp[2] = tcp_lro_rx_csum_tcphdr(pa->tcp);
729 
730 	/* Incorporate the latest ACK into the TCP header. */
731 	pa->tcp->th_ack = le->ack_seq;
732 	pa->tcp->th_win = le->window;
733 
734 	/* Incorporate latest timestamp into the TCP header. */
735 	if (le->timestamp != 0) {
736 		uint32_t *ts_ptr;
737 
738 		ts_ptr = (uint32_t *)(pa->tcp + 1);
739 		ts_ptr[1] = htonl(le->tsval);
740 		ts_ptr[2] = le->tsecr;
741 	}
742 
743 	/* Compute new TCP header checksum. */
744 	temp[3] = tcp_lro_rx_csum_tcphdr(pa->tcp);
745 
746 	/* Compute new TCP checksum. */
747 	csum = pa->tcp->th_sum + 0xffff - delta_sum +
748 	    0xffff - temp[0] + 0xffff - temp[3] + temp[2];
749 	while (csum > 0xffff)
750 		csum = (csum >> 16) + (csum & 0xffff);
751 
752 	/* Assign new TCP checksum. */
753 	tcp_lro_assign_and_checksum_16(&pa->tcp->th_sum, csum, &temp[4]);
754 
755 	/* Compute all modififications affecting next checksum. */
756 	csum = temp[0] + temp[1] + 0xffff - temp[2] +
757 	    temp[3] + temp[4] + delta_sum;
758 	while (csum > 0xffff)
759 		csum = (csum >> 16) + (csum & 0xffff);
760 
761 	/* Return delta checksum to next stage, if any. */
762 	return (csum);
763 
764 update_udp_header:
765 	tlen = sizeof(*pa->udp) + payload_len;
766 	/* Assign new UDP length and compute checksum delta. */
767 	tcp_lro_assign_and_checksum_16(&pa->udp->uh_ulen, htons(tlen), &temp[2]);
768 
769 	/* Check if there is a UDP checksum. */
770 	if (__predict_false(pa->udp->uh_sum != 0)) {
771 		/* Compute new UDP checksum. */
772 		csum = pa->udp->uh_sum + 0xffff - delta_sum +
773 		    0xffff - temp[0] + 0xffff - temp[2];
774 		while (csum > 0xffff)
775 			csum = (csum >> 16) + (csum & 0xffff);
776 		/* Assign new UDP checksum. */
777 		tcp_lro_assign_and_checksum_16(&pa->udp->uh_sum, csum, &temp[3]);
778 	}
779 
780 	/* Compute all modififications affecting next checksum. */
781 	csum = temp[0] + temp[1] + temp[2] + temp[3] + delta_sum;
782 	while (csum > 0xffff)
783 		csum = (csum >> 16) + (csum & 0xffff);
784 
785 	/* Return delta checksum to next stage, if any. */
786 	return (csum);
787 }
788 
789 static void
790 tcp_flush_out_entry(struct lro_ctrl *lc, struct lro_entry *le)
791 {
792 	/* Check if we need to recompute any checksums. */
793 	if (le->needs_merge) {
794 		uint16_t csum;
795 
796 		switch (le->inner.data.lro_type) {
797 		case LRO_TYPE_IPV4_TCP:
798 			csum = tcp_lro_update_checksum(&le->inner, le,
799 			    le->m_head->m_pkthdr.lro_tcp_d_len,
800 			    le->m_head->m_pkthdr.lro_tcp_d_csum);
801 			csum = tcp_lro_update_checksum(&le->outer, NULL,
802 			    le->m_head->m_pkthdr.lro_tcp_d_len +
803 			    le->inner.total_hdr_len, csum);
804 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
805 			    CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
806 			le->m_head->m_pkthdr.csum_data = 0xffff;
807 			if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
808 				le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
809 			break;
810 		case LRO_TYPE_IPV6_TCP:
811 			csum = tcp_lro_update_checksum(&le->inner, le,
812 			    le->m_head->m_pkthdr.lro_tcp_d_len,
813 			    le->m_head->m_pkthdr.lro_tcp_d_csum);
814 			csum = tcp_lro_update_checksum(&le->outer, NULL,
815 			    le->m_head->m_pkthdr.lro_tcp_d_len +
816 			    le->inner.total_hdr_len, csum);
817 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
818 			    CSUM_PSEUDO_HDR;
819 			le->m_head->m_pkthdr.csum_data = 0xffff;
820 			if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
821 				le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
822 			break;
823 		case LRO_TYPE_NONE:
824 			switch (le->outer.data.lro_type) {
825 			case LRO_TYPE_IPV4_TCP:
826 				csum = tcp_lro_update_checksum(&le->outer, le,
827 				    le->m_head->m_pkthdr.lro_tcp_d_len,
828 				    le->m_head->m_pkthdr.lro_tcp_d_csum);
829 				le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
830 				    CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
831 				le->m_head->m_pkthdr.csum_data = 0xffff;
832 				if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
833 					le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
834 				break;
835 			case LRO_TYPE_IPV6_TCP:
836 				csum = tcp_lro_update_checksum(&le->outer, le,
837 				    le->m_head->m_pkthdr.lro_tcp_d_len,
838 				    le->m_head->m_pkthdr.lro_tcp_d_csum);
839 				le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
840 				    CSUM_PSEUDO_HDR;
841 				le->m_head->m_pkthdr.csum_data = 0xffff;
842 				if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
843 					le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
844 				break;
845 			default:
846 				break;
847 			}
848 			break;
849 		default:
850 			break;
851 		}
852 	}
853 
854 	/*
855 	 * Break any chain, this is not set to NULL on the singleton
856 	 * case m_nextpkt points to m_head. Other case set them
857 	 * m_nextpkt to NULL in push_and_replace.
858 	 */
859 	le->m_head->m_nextpkt = NULL;
860 	lc->lro_queued += le->m_head->m_pkthdr.lro_nsegs;
861 	(*lc->ifp->if_input)(lc->ifp, le->m_head);
862 }
863 
864 static void
865 tcp_set_entry_to_mbuf(struct lro_ctrl *lc, struct lro_entry *le,
866     struct mbuf *m, struct tcphdr *th)
867 {
868 	uint32_t *ts_ptr;
869 	uint16_t tcp_data_len;
870 	uint16_t tcp_opt_len;
871 
872 	ts_ptr = (uint32_t *)(th + 1);
873 	tcp_opt_len = (th->th_off << 2);
874 	tcp_opt_len -= sizeof(*th);
875 
876 	/* Check if there is a timestamp option. */
877 	if (tcp_opt_len == 0 ||
878 	    __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
879 	    *ts_ptr != TCP_LRO_TS_OPTION)) {
880 		/* We failed to find the timestamp option. */
881 		le->timestamp = 0;
882 	} else {
883 		le->timestamp = 1;
884 		le->tsval = ntohl(*(ts_ptr + 1));
885 		le->tsecr = *(ts_ptr + 2);
886 	}
887 
888 	tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
889 
890 	/* Pull out TCP sequence numbers and window size. */
891 	le->next_seq = ntohl(th->th_seq) + tcp_data_len;
892 	le->ack_seq = th->th_ack;
893 	le->window = th->th_win;
894 	le->flags = tcp_get_flags(th);
895 	le->needs_merge = 0;
896 
897 	/* Setup new data pointers. */
898 	le->m_head = m;
899 	le->m_tail = m_last(m);
900 }
901 
902 static void
903 tcp_push_and_replace(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m)
904 {
905 	struct lro_parser *pa;
906 
907 	/*
908 	 * Push up the stack of the current entry
909 	 * and replace it with "m".
910 	 */
911 	struct mbuf *msave;
912 
913 	/* Grab off the next and save it */
914 	msave = le->m_head->m_nextpkt;
915 	le->m_head->m_nextpkt = NULL;
916 
917 	/* Now push out the old entry */
918 	tcp_flush_out_entry(lc, le);
919 
920 	/* Re-parse new header, should not fail. */
921 	pa = tcp_lro_parser(m, &le->outer, &le->inner, false);
922 	KASSERT(pa != NULL,
923 	    ("tcp_push_and_replace: LRO parser failed on m=%p\n", m));
924 
925 	/*
926 	 * Now to replace the data properly in the entry
927 	 * we have to reset the TCP header and
928 	 * other fields.
929 	 */
930 	tcp_set_entry_to_mbuf(lc, le, m, pa->tcp);
931 
932 	/* Restore the next list */
933 	m->m_nextpkt = msave;
934 }
935 
936 static void
937 tcp_lro_mbuf_append_pkthdr(struct lro_entry *le, const struct mbuf *p)
938 {
939 	struct mbuf *m;
940 	uint32_t csum;
941 
942 	m = le->m_head;
943 	if (m->m_pkthdr.lro_nsegs == 1) {
944 		/* Compute relative checksum. */
945 		csum = p->m_pkthdr.lro_tcp_d_csum;
946 	} else {
947 		/* Merge TCP data checksums. */
948 		csum = (uint32_t)m->m_pkthdr.lro_tcp_d_csum +
949 		    (uint32_t)p->m_pkthdr.lro_tcp_d_csum;
950 		while (csum > 0xffff)
951 			csum = (csum >> 16) + (csum & 0xffff);
952 	}
953 
954 	/* Update various counters. */
955 	m->m_pkthdr.len += p->m_pkthdr.lro_tcp_d_len;
956 	m->m_pkthdr.lro_tcp_d_csum = csum;
957 	m->m_pkthdr.lro_tcp_d_len += p->m_pkthdr.lro_tcp_d_len;
958 	m->m_pkthdr.lro_nsegs += p->m_pkthdr.lro_nsegs;
959 	le->needs_merge = 1;
960 }
961 
962 static void
963 tcp_lro_condense(struct lro_ctrl *lc, struct lro_entry *le)
964 {
965 	/*
966 	 * Walk through the mbuf chain we
967 	 * have on tap and compress/condense
968 	 * as required.
969 	 */
970 	uint32_t *ts_ptr;
971 	struct mbuf *m;
972 	struct tcphdr *th;
973 	uint32_t tcp_data_len_total;
974 	uint32_t tcp_data_seg_total;
975 	uint16_t tcp_data_len;
976 	uint16_t tcp_opt_len;
977 
978 	/*
979 	 * First we must check the lead (m_head)
980 	 * we must make sure that it is *not*
981 	 * something that should be sent up
982 	 * right away (sack etc).
983 	 */
984 again:
985 	m = le->m_head->m_nextpkt;
986 	if (m == NULL) {
987 		/* Just one left. */
988 		return;
989 	}
990 
991 	th = tcp_lro_get_th(m);
992 	tcp_opt_len = (th->th_off << 2);
993 	tcp_opt_len -= sizeof(*th);
994 	ts_ptr = (uint32_t *)(th + 1);
995 
996 	if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
997 	    *ts_ptr != TCP_LRO_TS_OPTION)) {
998 		/*
999 		 * Its not the timestamp. We can't
1000 		 * use this guy as the head.
1001 		 */
1002 		le->m_head->m_nextpkt = m->m_nextpkt;
1003 		tcp_push_and_replace(lc, le, m);
1004 		goto again;
1005 	}
1006 	if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1007 		/*
1008 		 * Make sure that previously seen segments/ACKs are delivered
1009 		 * before this segment, e.g. FIN.
1010 		 */
1011 		le->m_head->m_nextpkt = m->m_nextpkt;
1012 		tcp_push_and_replace(lc, le, m);
1013 		goto again;
1014 	}
1015 	while((m = le->m_head->m_nextpkt) != NULL) {
1016 		/*
1017 		 * condense m into le, first
1018 		 * pull m out of the list.
1019 		 */
1020 		le->m_head->m_nextpkt = m->m_nextpkt;
1021 		m->m_nextpkt = NULL;
1022 		/* Setup my data */
1023 		tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
1024 		th = tcp_lro_get_th(m);
1025 		ts_ptr = (uint32_t *)(th + 1);
1026 		tcp_opt_len = (th->th_off << 2);
1027 		tcp_opt_len -= sizeof(*th);
1028 		tcp_data_len_total = le->m_head->m_pkthdr.lro_tcp_d_len + tcp_data_len;
1029 		tcp_data_seg_total = le->m_head->m_pkthdr.lro_nsegs + m->m_pkthdr.lro_nsegs;
1030 
1031 		if (tcp_data_seg_total >= lc->lro_ackcnt_lim ||
1032 		    tcp_data_len_total >= lc->lro_length_lim) {
1033 			/* Flush now if appending will result in overflow. */
1034 			tcp_push_and_replace(lc, le, m);
1035 			goto again;
1036 		}
1037 		if (tcp_opt_len != 0 &&
1038 		    __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1039 		    *ts_ptr != TCP_LRO_TS_OPTION)) {
1040 			/*
1041 			 * Maybe a sack in the new one? We need to
1042 			 * start all over after flushing the
1043 			 * current le. We will go up to the beginning
1044 			 * and flush it (calling the replace again possibly
1045 			 * or just returning).
1046 			 */
1047 			tcp_push_and_replace(lc, le, m);
1048 			goto again;
1049 		}
1050 		if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1051 			tcp_push_and_replace(lc, le, m);
1052 			goto again;
1053 		}
1054 		if (tcp_opt_len != 0) {
1055 			uint32_t tsval = ntohl(*(ts_ptr + 1));
1056 			/* Make sure timestamp values are increasing. */
1057 			if (TSTMP_GT(le->tsval, tsval))  {
1058 				tcp_push_and_replace(lc, le, m);
1059 				goto again;
1060 			}
1061 			le->tsval = tsval;
1062 			le->tsecr = *(ts_ptr + 2);
1063 		}
1064 		/* Try to append the new segment. */
1065 		if (__predict_false(ntohl(th->th_seq) != le->next_seq ||
1066 				    ((tcp_get_flags(th) & TH_ACK) !=
1067 				      (le->flags & TH_ACK)) ||
1068 				    (tcp_data_len == 0 &&
1069 				     le->ack_seq == th->th_ack &&
1070 				     le->window == th->th_win))) {
1071 			/* Out of order packet, non-ACK + ACK or dup ACK. */
1072 			tcp_push_and_replace(lc, le, m);
1073 			goto again;
1074 		}
1075 		if (tcp_data_len != 0 ||
1076 		    SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1077 			le->next_seq += tcp_data_len;
1078 			le->ack_seq = th->th_ack;
1079 			le->window = th->th_win;
1080 			le->needs_merge = 1;
1081 		} else if (th->th_ack == le->ack_seq) {
1082 			if (WIN_GT(th->th_win, le->window)) {
1083 				le->window = th->th_win;
1084 				le->needs_merge = 1;
1085 			}
1086 		}
1087 
1088 		if (tcp_data_len == 0) {
1089 			m_freem(m);
1090 			continue;
1091 		}
1092 
1093 		/* Merge TCP data checksum and length to head mbuf. */
1094 		tcp_lro_mbuf_append_pkthdr(le, m);
1095 
1096 		/*
1097 		 * Adjust the mbuf so that m_data points to the first byte of
1098 		 * the ULP payload.  Adjust the mbuf to avoid complications and
1099 		 * append new segment to existing mbuf chain.
1100 		 */
1101 		m_adj(m, m->m_pkthdr.len - tcp_data_len);
1102 		m_demote_pkthdr(m);
1103 		le->m_tail->m_next = m;
1104 		le->m_tail = m_last(m);
1105 	}
1106 }
1107 
1108 static void
1109 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
1110 {
1111 
1112 	/* Only optimise if there are multiple packets waiting. */
1113 	NET_EPOCH_ASSERT();
1114 	if (tcp_lro_flush_tcphpts == NULL ||
1115 	    tcp_lro_flush_tcphpts(lc, le) != 0) {
1116 		tcp_lro_condense(lc, le);
1117 		tcp_flush_out_entry(lc, le);
1118 	}
1119 	lc->lro_flushed++;
1120 	bzero(le, sizeof(*le));
1121 	LIST_INSERT_HEAD(&lc->lro_free, le, next);
1122 }
1123 
1124 #define	tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
1125 
1126 /*
1127  * The tcp_lro_sort() routine is comparable to qsort(), except it has
1128  * a worst case complexity limit of O(MIN(N,64)*N), where N is the
1129  * number of elements to sort and 64 is the number of sequence bits
1130  * available. The algorithm is bit-slicing the 64-bit sequence number,
1131  * sorting one bit at a time from the most significant bit until the
1132  * least significant one, skipping the constant bits. This is
1133  * typically called a radix sort.
1134  */
1135 static void
1136 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
1137 {
1138 	struct lro_mbuf_sort temp;
1139 	uint64_t ones;
1140 	uint64_t zeros;
1141 	uint32_t x;
1142 	uint32_t y;
1143 
1144 repeat:
1145 	/* for small arrays insertion sort is faster */
1146 	if (size <= 12) {
1147 		for (x = 1; x < size; x++) {
1148 			temp = parray[x];
1149 			for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
1150 				parray[y] = parray[y - 1];
1151 			parray[y] = temp;
1152 		}
1153 		return;
1154 	}
1155 
1156 	/* compute sequence bits which are constant */
1157 	ones = 0;
1158 	zeros = 0;
1159 	for (x = 0; x != size; x++) {
1160 		ones |= parray[x].seq;
1161 		zeros |= ~parray[x].seq;
1162 	}
1163 
1164 	/* compute bits which are not constant into "ones" */
1165 	ones &= zeros;
1166 	if (ones == 0)
1167 		return;
1168 
1169 	/* pick the most significant bit which is not constant */
1170 	ones = tcp_lro_msb_64(ones);
1171 
1172 	/*
1173 	 * Move entries having cleared sequence bits to the beginning
1174 	 * of the array:
1175 	 */
1176 	for (x = y = 0; y != size; y++) {
1177 		/* skip set bits */
1178 		if (parray[y].seq & ones)
1179 			continue;
1180 		/* swap entries */
1181 		temp = parray[x];
1182 		parray[x] = parray[y];
1183 		parray[y] = temp;
1184 		x++;
1185 	}
1186 
1187 	KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
1188 
1189 	/* sort zeros */
1190 	tcp_lro_sort(parray, x);
1191 
1192 	/* sort ones */
1193 	parray += x;
1194 	size -= x;
1195 	goto repeat;
1196 }
1197 
1198 void
1199 tcp_lro_flush_all(struct lro_ctrl *lc)
1200 {
1201 	uint64_t seq;
1202 	uint64_t nseq;
1203 	unsigned x;
1204 
1205 	NET_EPOCH_ASSERT();
1206 	/* check if no mbufs to flush */
1207 	if (lc->lro_mbuf_count == 0)
1208 		goto done;
1209 	if (lc->lro_cpu_is_set == 0) {
1210 		if (lc->lro_last_cpu == curcpu) {
1211 			lc->lro_cnt_of_same_cpu++;
1212 			/* Have we reached the threshold to declare a cpu? */
1213 			if (lc->lro_cnt_of_same_cpu > tcp_lro_cpu_set_thresh)
1214 				lc->lro_cpu_is_set = 1;
1215 		} else {
1216 			lc->lro_last_cpu = curcpu;
1217 			lc->lro_cnt_of_same_cpu = 0;
1218 		}
1219 	}
1220 	CURVNET_SET(lc->ifp->if_vnet);
1221 
1222 	/* get current time */
1223 	binuptime(&lc->lro_last_queue_time);
1224 
1225 	/* sort all mbufs according to stream */
1226 	tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
1227 
1228 	/* input data into LRO engine, stream by stream */
1229 	seq = 0;
1230 	for (x = 0; x != lc->lro_mbuf_count; x++) {
1231 		struct mbuf *mb;
1232 
1233 		/* get mbuf */
1234 		mb = lc->lro_mbuf_data[x].mb;
1235 
1236 		/* get sequence number, masking away the packet index */
1237 		nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
1238 
1239 		/* check for new stream */
1240 		if (seq != nseq) {
1241 			seq = nseq;
1242 
1243 			/* flush active streams */
1244 			tcp_lro_rx_done(lc);
1245 		}
1246 
1247 		/* add packet to LRO engine */
1248 		if (tcp_lro_rx_common(lc, mb, 0, false) != 0) {
1249  			/* Flush anything we have acummulated */
1250  			tcp_lro_flush_active(lc);
1251 			/* input packet to network layer */
1252 			(*lc->ifp->if_input)(lc->ifp, mb);
1253 			lc->lro_queued++;
1254 			lc->lro_flushed++;
1255 		}
1256 	}
1257 	CURVNET_RESTORE();
1258 done:
1259 	/* flush active streams */
1260 	tcp_lro_rx_done(lc);
1261 	tcp_hpts_softclock();
1262 	lc->lro_mbuf_count = 0;
1263 }
1264 
1265 static struct lro_head *
1266 tcp_lro_rx_get_bucket(struct lro_ctrl *lc, struct mbuf *m, struct lro_parser *parser)
1267 {
1268 	u_long hash;
1269 
1270 	if (M_HASHTYPE_ISHASH(m)) {
1271 		hash = m->m_pkthdr.flowid;
1272 	} else {
1273 		for (unsigned i = hash = 0; i != LRO_RAW_ADDRESS_MAX; i++)
1274 			hash += parser->data.raw[i];
1275 	}
1276 	return (&lc->lro_hash[hash % lc->lro_hashsz]);
1277 }
1278 
1279 static int
1280 tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash)
1281 {
1282 	struct lro_parser pi;	/* inner address data */
1283 	struct lro_parser po;	/* outer address data */
1284 	struct lro_parser *pa;	/* current parser for TCP stream */
1285 	struct lro_entry *le;
1286 	struct lro_head *bucket;
1287 	struct tcphdr *th;
1288 	int tcp_data_len;
1289 	int tcp_opt_len;
1290 	int error;
1291 	uint16_t tcp_data_sum;
1292 
1293 #ifdef INET
1294 	/* Quickly decide if packet cannot be LRO'ed */
1295 	if (__predict_false(V_ipforwarding != 0))
1296 		return (TCP_LRO_CANNOT);
1297 #endif
1298 #ifdef INET6
1299 	/* Quickly decide if packet cannot be LRO'ed */
1300 	if (__predict_false(V_ip6_forwarding != 0))
1301 		return (TCP_LRO_CANNOT);
1302 #endif
1303 	if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1304 	     ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1305 	    (m->m_pkthdr.csum_data != 0xffff)) {
1306 		/*
1307 		 * The checksum either did not have hardware offload
1308 		 * or it was a bad checksum. We can't LRO such
1309 		 * a packet.
1310 		 */
1311 		counter_u64_add(tcp_bad_csums, 1);
1312 		return (TCP_LRO_CANNOT);
1313 	}
1314 	/* We expect a contiguous header [eh, ip, tcp]. */
1315 	pa = tcp_lro_parser(m, &po, &pi, true);
1316 	if (__predict_false(pa == NULL))
1317 		return (TCP_LRO_NOT_SUPPORTED);
1318 
1319 	/* We don't expect any padding. */
1320 	error = tcp_lro_trim_mbuf_chain(m, pa);
1321 	if (__predict_false(error != 0))
1322 		return (error);
1323 
1324 #ifdef INET
1325 	switch (pa->data.lro_type) {
1326 	case LRO_TYPE_IPV4_TCP:
1327 		error = tcp_lro_rx_ipv4(lc, m, pa->ip4);
1328 		if (__predict_false(error != 0))
1329 			return (error);
1330 		break;
1331 	default:
1332 		break;
1333 	}
1334 #endif
1335 	/* If no hardware or arrival stamp on the packet add timestamp */
1336 	if ((m->m_flags & (M_TSTMP_LRO | M_TSTMP)) == 0) {
1337 		m->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
1338 		m->m_flags |= M_TSTMP_LRO;
1339 	}
1340 
1341 	/* Get pointer to TCP header. */
1342 	th = pa->tcp;
1343 
1344 	/* Don't process SYN packets. */
1345 	if (__predict_false(tcp_get_flags(th) & TH_SYN))
1346 		return (TCP_LRO_CANNOT);
1347 
1348 	/* Get total TCP header length and compute payload length. */
1349 	tcp_opt_len = (th->th_off << 2);
1350 	tcp_data_len = m->m_pkthdr.len - ((uint8_t *)th -
1351 	    (uint8_t *)m->m_data) - tcp_opt_len;
1352 	tcp_opt_len -= sizeof(*th);
1353 
1354 	/* Don't process invalid TCP headers. */
1355 	if (__predict_false(tcp_opt_len < 0 || tcp_data_len < 0))
1356 		return (TCP_LRO_CANNOT);
1357 
1358 	/* Compute TCP data only checksum. */
1359 	if (tcp_data_len == 0)
1360 		tcp_data_sum = 0;	/* no data, no checksum */
1361 	else if (__predict_false(csum != 0))
1362 		tcp_data_sum = tcp_lro_rx_csum_data(pa, ~csum);
1363 	else
1364 		tcp_data_sum = tcp_lro_rx_csum_data(pa, ~th->th_sum);
1365 
1366 	/* Save TCP info in mbuf. */
1367 	m->m_nextpkt = NULL;
1368 	m->m_pkthdr.rcvif = lc->ifp;
1369 	m->m_pkthdr.lro_tcp_d_csum = tcp_data_sum;
1370 	m->m_pkthdr.lro_tcp_d_len = tcp_data_len;
1371 	m->m_pkthdr.lro_tcp_h_off = ((uint8_t *)th - (uint8_t *)m->m_data);
1372 	m->m_pkthdr.lro_nsegs = 1;
1373 
1374 	/* Get hash bucket. */
1375 	if (!use_hash) {
1376 		bucket = &lc->lro_hash[0];
1377 	} else {
1378 		bucket = tcp_lro_rx_get_bucket(lc, m, pa);
1379 	}
1380 
1381 	/* Try to find a matching previous segment. */
1382 	LIST_FOREACH(le, bucket, hash_next) {
1383 		/* Compare addresses and ports. */
1384 		if (lro_address_compare(&po.data, &le->outer.data) == false ||
1385 		    lro_address_compare(&pi.data, &le->inner.data) == false)
1386 			continue;
1387 
1388 		/* Check if no data and old ACK. */
1389 		if (tcp_data_len == 0 &&
1390 		    SEQ_LT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1391 			m_freem(m);
1392 			return (0);
1393 		}
1394 
1395 		/* Mark "m" in the last spot. */
1396 		le->m_last_mbuf->m_nextpkt = m;
1397 		/* Now set the tail to "m". */
1398 		le->m_last_mbuf = m;
1399 		return (0);
1400 	}
1401 
1402 	/* Try to find an empty slot. */
1403 	if (LIST_EMPTY(&lc->lro_free))
1404 		return (TCP_LRO_NO_ENTRIES);
1405 
1406 	/* Start a new segment chain. */
1407 	le = LIST_FIRST(&lc->lro_free);
1408 	LIST_REMOVE(le, next);
1409 	tcp_lro_active_insert(lc, bucket, le);
1410 
1411 	/* Make sure the headers are set. */
1412 	le->inner = pi;
1413 	le->outer = po;
1414 
1415 	/* Store time this entry was allocated. */
1416 	le->alloc_time = lc->lro_last_queue_time;
1417 
1418 	tcp_set_entry_to_mbuf(lc, le, m, th);
1419 
1420 	/* Now set the tail to "m". */
1421 	le->m_last_mbuf = m;
1422 
1423 	return (0);
1424 }
1425 
1426 int
1427 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
1428 {
1429 	int error;
1430 
1431 	if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1432 	     ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1433 	    (m->m_pkthdr.csum_data != 0xffff)) {
1434 		/*
1435 		 * The checksum either did not have hardware offload
1436 		 * or it was a bad checksum. We can't LRO such
1437 		 * a packet.
1438 		 */
1439 		counter_u64_add(tcp_bad_csums, 1);
1440 		return (TCP_LRO_CANNOT);
1441 	}
1442 	/* get current time */
1443 	binuptime(&lc->lro_last_queue_time);
1444 	CURVNET_SET(lc->ifp->if_vnet);
1445 	error = tcp_lro_rx_common(lc, m, csum, true);
1446 	if (__predict_false(error != 0)) {
1447 		/*
1448 		 * Flush anything we have acummulated
1449 		 * ahead of this packet that can't
1450 		 * be LRO'd. This preserves order.
1451 		 */
1452 		tcp_lro_flush_active(lc);
1453 	}
1454 	CURVNET_RESTORE();
1455 
1456 	return (error);
1457 }
1458 
1459 void
1460 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
1461 {
1462 	NET_EPOCH_ASSERT();
1463 	/* sanity checks */
1464 	if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
1465 	    lc->lro_mbuf_max == 0)) {
1466 		/* packet drop */
1467 		m_freem(mb);
1468 		return;
1469 	}
1470 
1471 	/* check if packet is not LRO capable */
1472 	if (__predict_false((lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
1473 		/* input packet to network layer */
1474 		(*lc->ifp->if_input) (lc->ifp, mb);
1475 		return;
1476 	}
1477 
1478  	/* If no hardware or arrival stamp on the packet add timestamp */
1479  	if ((tcplro_stacks_wanting_mbufq > 0) &&
1480  	    (tcp_less_accurate_lro_ts == 0) &&
1481  	    ((mb->m_flags & M_TSTMP) == 0)) {
1482  		/* Add in an LRO time since no hardware */
1483  		binuptime(&lc->lro_last_queue_time);
1484  		mb->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
1485  		mb->m_flags |= M_TSTMP_LRO;
1486  	}
1487 
1488 	/* create sequence number */
1489 	lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
1490 	    (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
1491 	    (((uint64_t)mb->m_pkthdr.flowid) << 24) |
1492 	    ((uint64_t)lc->lro_mbuf_count);
1493 
1494 	/* enter mbuf */
1495 	lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb;
1496 
1497 	/* flush if array is full */
1498 	if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max))
1499 		tcp_lro_flush_all(lc);
1500 }
1501 
1502 /* end */
1503