xref: /freebsd/sys/netinet/tcp_lro.c (revision e0c4386e7e71d93b0edc0c8fa156263fc4a8b0b6)
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, elements;
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 	if (lro_entries > lro_mbufs)
194 		elements = lro_entries;
195 	else
196 		elements = lro_mbufs;
197 	lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz,
198 	    HASH_NOWAIT);
199 	if (lc->lro_hash == NULL) {
200 		memset(lc, 0, sizeof(*lc));
201 		return (ENOMEM);
202 	}
203 
204 	/* compute size to allocate */
205 	size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
206 	    (lro_entries * sizeof(*le));
207 	lc->lro_mbuf_data = (struct lro_mbuf_sort *)
208 	    malloc(size, M_LRO, M_NOWAIT | M_ZERO);
209 
210 	/* check for out of memory */
211 	if (lc->lro_mbuf_data == NULL) {
212 		free(lc->lro_hash, M_LRO);
213 		memset(lc, 0, sizeof(*lc));
214 		return (ENOMEM);
215 	}
216 	/* compute offset for LRO entries */
217 	le = (struct lro_entry *)
218 	    (lc->lro_mbuf_data + lro_mbufs);
219 
220 	/* setup linked list */
221 	for (i = 0; i != lro_entries; i++)
222 		LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
223 
224 	return (0);
225 }
226 
227 struct vxlan_header {
228 	uint32_t	vxlh_flags;
229 	uint32_t	vxlh_vni;
230 };
231 
232 static inline void *
233 tcp_lro_low_level_parser(void *ptr, struct lro_parser *parser, bool update_data, bool is_vxlan, int mlen)
234 {
235 	const struct ether_vlan_header *eh;
236 	void *old;
237 	uint16_t eth_type;
238 
239 	if (update_data)
240 		memset(parser, 0, sizeof(*parser));
241 
242 	old = ptr;
243 
244 	if (is_vxlan) {
245 		const struct vxlan_header *vxh;
246 		vxh = ptr;
247 		ptr = (uint8_t *)ptr + sizeof(*vxh);
248 		if (update_data) {
249 			parser->data.vxlan_vni =
250 			    vxh->vxlh_vni & htonl(0xffffff00);
251 		}
252 	}
253 
254 	eh = ptr;
255 	if (__predict_false(eh->evl_encap_proto == htons(ETHERTYPE_VLAN))) {
256 		eth_type = eh->evl_proto;
257 		if (update_data) {
258 			/* strip priority and keep VLAN ID only */
259 			parser->data.vlan_id = eh->evl_tag & htons(EVL_VLID_MASK);
260 		}
261 		/* advance to next header */
262 		ptr = (uint8_t *)ptr + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
263 		mlen -= (ETHER_HDR_LEN  + ETHER_VLAN_ENCAP_LEN);
264 	} else {
265 		eth_type = eh->evl_encap_proto;
266 		/* advance to next header */
267 		mlen -= ETHER_HDR_LEN;
268 		ptr = (uint8_t *)ptr + ETHER_HDR_LEN;
269 	}
270 	if (__predict_false(mlen <= 0))
271 		return (NULL);
272 	switch (eth_type) {
273 #ifdef INET
274 	case htons(ETHERTYPE_IP):
275 		parser->ip4 = ptr;
276 		if (__predict_false(mlen < sizeof(struct ip)))
277 			return (NULL);
278 		/* Ensure there are no IPv4 options. */
279 		if ((parser->ip4->ip_hl << 2) != sizeof (*parser->ip4))
280 			break;
281 		/* .. and the packet is not fragmented. */
282 		if (parser->ip4->ip_off & htons(IP_MF|IP_OFFMASK))
283 			break;
284 		/* .. and the packet has valid src/dst addrs */
285 		if (__predict_false(parser->ip4->ip_src.s_addr == INADDR_ANY ||
286 			parser->ip4->ip_dst.s_addr == INADDR_ANY))
287 			break;
288 		ptr = (uint8_t *)ptr + (parser->ip4->ip_hl << 2);
289 		mlen -= sizeof(struct ip);
290 		if (update_data) {
291 			parser->data.s_addr.v4 = parser->ip4->ip_src;
292 			parser->data.d_addr.v4 = parser->ip4->ip_dst;
293 		}
294 		switch (parser->ip4->ip_p) {
295 		case IPPROTO_UDP:
296 			if (__predict_false(mlen < sizeof(struct udphdr)))
297 				return (NULL);
298 			parser->udp = ptr;
299 			if (update_data) {
300 				parser->data.lro_type = LRO_TYPE_IPV4_UDP;
301 				parser->data.s_port = parser->udp->uh_sport;
302 				parser->data.d_port = parser->udp->uh_dport;
303 			} else {
304 				MPASS(parser->data.lro_type == LRO_TYPE_IPV4_UDP);
305 			}
306 			ptr = ((uint8_t *)ptr + sizeof(*parser->udp));
307 			parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
308 			return (ptr);
309 		case IPPROTO_TCP:
310 			parser->tcp = ptr;
311 			if (__predict_false(mlen < sizeof(struct tcphdr)))
312 				return (NULL);
313 			if (update_data) {
314 				parser->data.lro_type = LRO_TYPE_IPV4_TCP;
315 				parser->data.s_port = parser->tcp->th_sport;
316 				parser->data.d_port = parser->tcp->th_dport;
317 			} else {
318 				MPASS(parser->data.lro_type == LRO_TYPE_IPV4_TCP);
319 			}
320 			if (__predict_false(mlen < (parser->tcp->th_off << 2)))
321 				return (NULL);
322 			ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
323 			parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
324 			return (ptr);
325 		default:
326 			break;
327 		}
328 		break;
329 #endif
330 #ifdef INET6
331 	case htons(ETHERTYPE_IPV6):
332 		parser->ip6 = ptr;
333 		if (__predict_false(mlen < sizeof(struct ip6_hdr)))
334 			return (NULL);
335 		/* Ensure the packet has valid src/dst addrs */
336 		if (__predict_false(IN6_IS_ADDR_UNSPECIFIED(&parser->ip6->ip6_src) ||
337 			IN6_IS_ADDR_UNSPECIFIED(&parser->ip6->ip6_dst)))
338 			return (NULL);
339 		ptr = (uint8_t *)ptr + sizeof(*parser->ip6);
340 		if (update_data) {
341 			parser->data.s_addr.v6 = parser->ip6->ip6_src;
342 			parser->data.d_addr.v6 = parser->ip6->ip6_dst;
343 		}
344 		mlen -= sizeof(struct ip6_hdr);
345 		switch (parser->ip6->ip6_nxt) {
346 		case IPPROTO_UDP:
347 			if (__predict_false(mlen < sizeof(struct udphdr)))
348 				return (NULL);
349 			parser->udp = ptr;
350 			if (update_data) {
351 				parser->data.lro_type = LRO_TYPE_IPV6_UDP;
352 				parser->data.s_port = parser->udp->uh_sport;
353 				parser->data.d_port = parser->udp->uh_dport;
354 			} else {
355 				MPASS(parser->data.lro_type == LRO_TYPE_IPV6_UDP);
356 			}
357 			ptr = (uint8_t *)ptr + sizeof(*parser->udp);
358 			parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
359 			return (ptr);
360 		case IPPROTO_TCP:
361 			if (__predict_false(mlen < sizeof(struct tcphdr)))
362 				return (NULL);
363 			parser->tcp = ptr;
364 			if (update_data) {
365 				parser->data.lro_type = LRO_TYPE_IPV6_TCP;
366 				parser->data.s_port = parser->tcp->th_sport;
367 				parser->data.d_port = parser->tcp->th_dport;
368 			} else {
369 				MPASS(parser->data.lro_type == LRO_TYPE_IPV6_TCP);
370 			}
371 			if (__predict_false(mlen < (parser->tcp->th_off << 2)))
372 				return (NULL);
373 			ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
374 			parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
375 			return (ptr);
376 		default:
377 			break;
378 		}
379 		break;
380 #endif
381 	default:
382 		break;
383 	}
384 	/* Invalid packet - cannot parse */
385 	return (NULL);
386 }
387 
388 static const int vxlan_csum = CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
389     CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID;
390 
391 static inline struct lro_parser *
392 tcp_lro_parser(struct mbuf *m, struct lro_parser *po, struct lro_parser *pi, bool update_data)
393 {
394 	void *data_ptr;
395 
396 	/* Try to parse outer headers first. */
397 	data_ptr = tcp_lro_low_level_parser(m->m_data, po, update_data, false, m->m_len);
398 	if (data_ptr == NULL || po->total_hdr_len > m->m_len)
399 		return (NULL);
400 
401 	if (update_data) {
402 		/* Store VLAN ID, if any. */
403 		if (__predict_false(m->m_flags & M_VLANTAG)) {
404 			po->data.vlan_id =
405 			    htons(m->m_pkthdr.ether_vtag) & htons(EVL_VLID_MASK);
406 		}
407 		/* Store decrypted flag, if any. */
408 		if (__predict_false((m->m_pkthdr.csum_flags &
409 		    CSUM_TLS_MASK) == CSUM_TLS_DECRYPTED))
410 			po->data.lro_flags |= LRO_FLAG_DECRYPTED;
411 	}
412 
413 	switch (po->data.lro_type) {
414 	case LRO_TYPE_IPV4_UDP:
415 	case LRO_TYPE_IPV6_UDP:
416 		/* Check for VXLAN headers. */
417 		if ((m->m_pkthdr.csum_flags & vxlan_csum) != vxlan_csum)
418 			break;
419 
420 		/* Try to parse inner headers. */
421 		data_ptr = tcp_lro_low_level_parser(data_ptr, pi, update_data, true,
422 						    (m->m_len - ((caddr_t)data_ptr - m->m_data)));
423 		if (data_ptr == NULL || (pi->total_hdr_len + po->total_hdr_len) > m->m_len)
424 			break;
425 
426 		/* Verify supported header types. */
427 		switch (pi->data.lro_type) {
428 		case LRO_TYPE_IPV4_TCP:
429 		case LRO_TYPE_IPV6_TCP:
430 			return (pi);
431 		default:
432 			break;
433 		}
434 		break;
435 	case LRO_TYPE_IPV4_TCP:
436 	case LRO_TYPE_IPV6_TCP:
437 		if (update_data)
438 			memset(pi, 0, sizeof(*pi));
439 		return (po);
440 	default:
441 		break;
442 	}
443 	return (NULL);
444 }
445 
446 static inline int
447 tcp_lro_trim_mbuf_chain(struct mbuf *m, const struct lro_parser *po)
448 {
449 	int len;
450 
451 	switch (po->data.lro_type) {
452 #ifdef INET
453 	case LRO_TYPE_IPV4_TCP:
454 		len = ((uint8_t *)po->ip4 - (uint8_t *)m->m_data) +
455 		    ntohs(po->ip4->ip_len);
456 		break;
457 #endif
458 #ifdef INET6
459 	case LRO_TYPE_IPV6_TCP:
460 		len = ((uint8_t *)po->ip6 - (uint8_t *)m->m_data) +
461 		    ntohs(po->ip6->ip6_plen) + sizeof(*po->ip6);
462 		break;
463 #endif
464 	default:
465 		return (TCP_LRO_CANNOT);
466 	}
467 
468 	/*
469 	 * If the frame is padded beyond the end of the IP packet,
470 	 * then trim the extra bytes off:
471 	 */
472 	if (__predict_true(m->m_pkthdr.len == len)) {
473 		return (0);
474 	} else if (m->m_pkthdr.len > len) {
475 		m_adj(m, len - m->m_pkthdr.len);
476 		return (0);
477 	}
478 	return (TCP_LRO_CANNOT);
479 }
480 
481 static void
482 lro_free_mbuf_chain(struct mbuf *m)
483 {
484 	struct mbuf *save;
485 
486 	while (m) {
487 		save = m->m_nextpkt;
488 		m->m_nextpkt = NULL;
489 		m_freem(m);
490 		m = save;
491 	}
492 }
493 
494 void
495 tcp_lro_free(struct lro_ctrl *lc)
496 {
497 	struct lro_entry *le;
498 	unsigned x;
499 
500 	/* reset LRO free list */
501 	LIST_INIT(&lc->lro_free);
502 
503 	/* free active mbufs, if any */
504 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
505 		tcp_lro_active_remove(le);
506 		lro_free_mbuf_chain(le->m_head);
507 	}
508 
509 	/* free hash table */
510 	free(lc->lro_hash, M_LRO);
511 	lc->lro_hash = NULL;
512 	lc->lro_hashsz = 0;
513 
514 	/* free mbuf array, if any */
515 	for (x = 0; x != lc->lro_mbuf_count; x++)
516 		m_freem(lc->lro_mbuf_data[x].mb);
517 	lc->lro_mbuf_count = 0;
518 
519 	/* free allocated memory, if any */
520 	free(lc->lro_mbuf_data, M_LRO);
521 	lc->lro_mbuf_data = NULL;
522 }
523 
524 static uint16_t
525 tcp_lro_rx_csum_tcphdr(const struct tcphdr *th)
526 {
527 	const uint16_t *ptr;
528 	uint32_t csum;
529 	uint16_t len;
530 
531 	csum = -th->th_sum;	/* exclude checksum field */
532 	len = th->th_off;
533 	ptr = (const uint16_t *)th;
534 	while (len--) {
535 		csum += *ptr;
536 		ptr++;
537 		csum += *ptr;
538 		ptr++;
539 	}
540 	while (csum > 0xffff)
541 		csum = (csum >> 16) + (csum & 0xffff);
542 
543 	return (csum);
544 }
545 
546 static uint16_t
547 tcp_lro_rx_csum_data(const struct lro_parser *pa, uint16_t tcp_csum)
548 {
549 	uint32_t c;
550 	uint16_t cs;
551 
552 	c = tcp_csum;
553 
554 	switch (pa->data.lro_type) {
555 #ifdef INET6
556 	case LRO_TYPE_IPV6_TCP:
557 		/* Compute full pseudo IPv6 header checksum. */
558 		cs = in6_cksum_pseudo(pa->ip6, ntohs(pa->ip6->ip6_plen), pa->ip6->ip6_nxt, 0);
559 		break;
560 #endif
561 #ifdef INET
562 	case LRO_TYPE_IPV4_TCP:
563 		/* Compute full pseudo IPv4 header checsum. */
564 		cs = in_addword(ntohs(pa->ip4->ip_len) - sizeof(*pa->ip4), IPPROTO_TCP);
565 		cs = in_pseudo(pa->ip4->ip_src.s_addr, pa->ip4->ip_dst.s_addr, htons(cs));
566 		break;
567 #endif
568 	default:
569 		cs = 0;		/* Keep compiler happy. */
570 		break;
571 	}
572 
573 	/* Complement checksum. */
574 	cs = ~cs;
575 	c += cs;
576 
577 	/* Remove TCP header checksum. */
578 	cs = ~tcp_lro_rx_csum_tcphdr(pa->tcp);
579 	c += cs;
580 
581 	/* Compute checksum remainder. */
582 	while (c > 0xffff)
583 		c = (c >> 16) + (c & 0xffff);
584 
585 	return (c);
586 }
587 
588 static void
589 tcp_lro_rx_done(struct lro_ctrl *lc)
590 {
591 	struct lro_entry *le;
592 
593 	while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
594 		tcp_lro_active_remove(le);
595 		tcp_lro_flush(lc, le);
596 	}
597 }
598 
599 static void
600 tcp_lro_flush_active(struct lro_ctrl *lc)
601 {
602 	struct lro_entry *le;
603 
604 	/*
605 	 * Walk through the list of le entries, and
606 	 * any one that does have packets flush. This
607 	 * is called because we have an inbound packet
608 	 * (e.g. SYN) that has to have all others flushed
609 	 * in front of it. Note we have to do the remove
610 	 * because tcp_lro_flush() assumes that the entry
611 	 * is being freed. This is ok it will just get
612 	 * reallocated again like it was new.
613 	 */
614 	LIST_FOREACH(le, &lc->lro_active, next) {
615 		if (le->m_head != NULL) {
616 			tcp_lro_active_remove(le);
617 			tcp_lro_flush(lc, le);
618 		}
619 	}
620 }
621 
622 void
623 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
624 {
625 	struct lro_entry *le, *le_tmp;
626 	uint64_t now, tov;
627 	struct bintime bt;
628 
629 	NET_EPOCH_ASSERT();
630 	if (LIST_EMPTY(&lc->lro_active))
631 		return;
632 
633 	/* get timeout time and current time in ns */
634 	binuptime(&bt);
635 	now = bintime2ns(&bt);
636 	tov = ((timeout->tv_sec * 1000000000) + (timeout->tv_usec * 1000));
637 	LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
638 		if (now >= (bintime2ns(&le->alloc_time) + tov)) {
639 			tcp_lro_active_remove(le);
640 			tcp_lro_flush(lc, le);
641 		}
642 	}
643 }
644 
645 #ifdef INET
646 static int
647 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4)
648 {
649 	uint16_t csum;
650 
651 	/* Legacy IP has a header checksum that needs to be correct. */
652 	if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) {
653 		if (__predict_false((m->m_pkthdr.csum_flags & CSUM_IP_VALID) == 0)) {
654 			lc->lro_bad_csum++;
655 			return (TCP_LRO_CANNOT);
656 		}
657 	} else {
658 		csum = in_cksum_hdr(ip4);
659 		if (__predict_false(csum != 0)) {
660 			lc->lro_bad_csum++;
661 			return (TCP_LRO_CANNOT);
662 		}
663 	}
664 	return (0);
665 }
666 #endif
667 
668 static inline void
669 tcp_lro_assign_and_checksum_16(uint16_t *ptr, uint16_t value, uint16_t *psum)
670 {
671 	uint32_t csum;
672 
673 	csum = 0xffff - *ptr + value;
674 	while (csum > 0xffff)
675 		csum = (csum >> 16) + (csum & 0xffff);
676 	*ptr = value;
677 	*psum = csum;
678 }
679 
680 static uint16_t
681 tcp_lro_update_checksum(const struct lro_parser *pa, const struct lro_entry *le,
682     uint16_t payload_len, uint16_t delta_sum)
683 {
684 	uint32_t csum;
685 	uint16_t tlen;
686 	uint16_t temp[5] = {};
687 
688 	switch (pa->data.lro_type) {
689 	case LRO_TYPE_IPV4_TCP:
690 		/* Compute new IPv4 length. */
691 		tlen = (pa->ip4->ip_hl << 2) + (pa->tcp->th_off << 2) + payload_len;
692 		tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
693 
694 		/* Subtract delta from current IPv4 checksum. */
695 		csum = pa->ip4->ip_sum + 0xffff - temp[0];
696 		while (csum > 0xffff)
697 			csum = (csum >> 16) + (csum & 0xffff);
698 		tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
699 		goto update_tcp_header;
700 
701 	case LRO_TYPE_IPV6_TCP:
702 		/* Compute new IPv6 length. */
703 		tlen = (pa->tcp->th_off << 2) + payload_len;
704 		tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
705 		goto update_tcp_header;
706 
707 	case LRO_TYPE_IPV4_UDP:
708 		/* Compute new IPv4 length. */
709 		tlen = (pa->ip4->ip_hl << 2) + sizeof(*pa->udp) + payload_len;
710 		tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
711 
712 		/* Subtract delta from current IPv4 checksum. */
713 		csum = pa->ip4->ip_sum + 0xffff - temp[0];
714 		while (csum > 0xffff)
715 			csum = (csum >> 16) + (csum & 0xffff);
716 		tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
717 		goto update_udp_header;
718 
719 	case LRO_TYPE_IPV6_UDP:
720 		/* Compute new IPv6 length. */
721 		tlen = sizeof(*pa->udp) + payload_len;
722 		tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
723 		goto update_udp_header;
724 
725 	default:
726 		return (0);
727 	}
728 
729 update_tcp_header:
730 	/* Compute current TCP header checksum. */
731 	temp[2] = tcp_lro_rx_csum_tcphdr(pa->tcp);
732 
733 	/* Incorporate the latest ACK into the TCP header. */
734 	pa->tcp->th_ack = le->ack_seq;
735 	pa->tcp->th_win = le->window;
736 
737 	/* Incorporate latest timestamp into the TCP header. */
738 	if (le->timestamp != 0) {
739 		uint32_t *ts_ptr;
740 
741 		ts_ptr = (uint32_t *)(pa->tcp + 1);
742 		ts_ptr[1] = htonl(le->tsval);
743 		ts_ptr[2] = le->tsecr;
744 	}
745 
746 	/* Compute new TCP header checksum. */
747 	temp[3] = tcp_lro_rx_csum_tcphdr(pa->tcp);
748 
749 	/* Compute new TCP checksum. */
750 	csum = pa->tcp->th_sum + 0xffff - delta_sum +
751 	    0xffff - temp[0] + 0xffff - temp[3] + temp[2];
752 	while (csum > 0xffff)
753 		csum = (csum >> 16) + (csum & 0xffff);
754 
755 	/* Assign new TCP checksum. */
756 	tcp_lro_assign_and_checksum_16(&pa->tcp->th_sum, csum, &temp[4]);
757 
758 	/* Compute all modififications affecting next checksum. */
759 	csum = temp[0] + temp[1] + 0xffff - temp[2] +
760 	    temp[3] + temp[4] + delta_sum;
761 	while (csum > 0xffff)
762 		csum = (csum >> 16) + (csum & 0xffff);
763 
764 	/* Return delta checksum to next stage, if any. */
765 	return (csum);
766 
767 update_udp_header:
768 	tlen = sizeof(*pa->udp) + payload_len;
769 	/* Assign new UDP length and compute checksum delta. */
770 	tcp_lro_assign_and_checksum_16(&pa->udp->uh_ulen, htons(tlen), &temp[2]);
771 
772 	/* Check if there is a UDP checksum. */
773 	if (__predict_false(pa->udp->uh_sum != 0)) {
774 		/* Compute new UDP checksum. */
775 		csum = pa->udp->uh_sum + 0xffff - delta_sum +
776 		    0xffff - temp[0] + 0xffff - temp[2];
777 		while (csum > 0xffff)
778 			csum = (csum >> 16) + (csum & 0xffff);
779 		/* Assign new UDP checksum. */
780 		tcp_lro_assign_and_checksum_16(&pa->udp->uh_sum, csum, &temp[3]);
781 	}
782 
783 	/* Compute all modififications affecting next checksum. */
784 	csum = temp[0] + temp[1] + temp[2] + temp[3] + delta_sum;
785 	while (csum > 0xffff)
786 		csum = (csum >> 16) + (csum & 0xffff);
787 
788 	/* Return delta checksum to next stage, if any. */
789 	return (csum);
790 }
791 
792 static void
793 tcp_flush_out_entry(struct lro_ctrl *lc, struct lro_entry *le)
794 {
795 	/* Check if we need to recompute any checksums. */
796 	if (le->needs_merge) {
797 		uint16_t csum;
798 
799 		switch (le->inner.data.lro_type) {
800 		case LRO_TYPE_IPV4_TCP:
801 			csum = tcp_lro_update_checksum(&le->inner, le,
802 			    le->m_head->m_pkthdr.lro_tcp_d_len,
803 			    le->m_head->m_pkthdr.lro_tcp_d_csum);
804 			csum = tcp_lro_update_checksum(&le->outer, NULL,
805 			    le->m_head->m_pkthdr.lro_tcp_d_len +
806 			    le->inner.total_hdr_len, csum);
807 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
808 			    CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
809 			le->m_head->m_pkthdr.csum_data = 0xffff;
810 			if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
811 				le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
812 			break;
813 		case LRO_TYPE_IPV6_TCP:
814 			csum = tcp_lro_update_checksum(&le->inner, le,
815 			    le->m_head->m_pkthdr.lro_tcp_d_len,
816 			    le->m_head->m_pkthdr.lro_tcp_d_csum);
817 			csum = tcp_lro_update_checksum(&le->outer, NULL,
818 			    le->m_head->m_pkthdr.lro_tcp_d_len +
819 			    le->inner.total_hdr_len, csum);
820 			le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
821 			    CSUM_PSEUDO_HDR;
822 			le->m_head->m_pkthdr.csum_data = 0xffff;
823 			if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
824 				le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
825 			break;
826 		case LRO_TYPE_NONE:
827 			switch (le->outer.data.lro_type) {
828 			case LRO_TYPE_IPV4_TCP:
829 				csum = tcp_lro_update_checksum(&le->outer, le,
830 				    le->m_head->m_pkthdr.lro_tcp_d_len,
831 				    le->m_head->m_pkthdr.lro_tcp_d_csum);
832 				le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
833 				    CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
834 				le->m_head->m_pkthdr.csum_data = 0xffff;
835 				if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
836 					le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
837 				break;
838 			case LRO_TYPE_IPV6_TCP:
839 				csum = tcp_lro_update_checksum(&le->outer, le,
840 				    le->m_head->m_pkthdr.lro_tcp_d_len,
841 				    le->m_head->m_pkthdr.lro_tcp_d_csum);
842 				le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
843 				    CSUM_PSEUDO_HDR;
844 				le->m_head->m_pkthdr.csum_data = 0xffff;
845 				if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
846 					le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
847 				break;
848 			default:
849 				break;
850 			}
851 			break;
852 		default:
853 			break;
854 		}
855 	}
856 
857 	/*
858 	 * Break any chain, this is not set to NULL on the singleton
859 	 * case m_nextpkt points to m_head. Other case set them
860 	 * m_nextpkt to NULL in push_and_replace.
861 	 */
862 	le->m_head->m_nextpkt = NULL;
863 	lc->lro_queued += le->m_head->m_pkthdr.lro_nsegs;
864 	(*lc->ifp->if_input)(lc->ifp, le->m_head);
865 }
866 
867 static void
868 tcp_set_entry_to_mbuf(struct lro_ctrl *lc, struct lro_entry *le,
869     struct mbuf *m, struct tcphdr *th)
870 {
871 	uint32_t *ts_ptr;
872 	uint16_t tcp_data_len;
873 	uint16_t tcp_opt_len;
874 
875 	ts_ptr = (uint32_t *)(th + 1);
876 	tcp_opt_len = (th->th_off << 2);
877 	tcp_opt_len -= sizeof(*th);
878 
879 	/* Check if there is a timestamp option. */
880 	if (tcp_opt_len == 0 ||
881 	    __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
882 	    *ts_ptr != TCP_LRO_TS_OPTION)) {
883 		/* We failed to find the timestamp option. */
884 		le->timestamp = 0;
885 	} else {
886 		le->timestamp = 1;
887 		le->tsval = ntohl(*(ts_ptr + 1));
888 		le->tsecr = *(ts_ptr + 2);
889 	}
890 
891 	tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
892 
893 	/* Pull out TCP sequence numbers and window size. */
894 	le->next_seq = ntohl(th->th_seq) + tcp_data_len;
895 	le->ack_seq = th->th_ack;
896 	le->window = th->th_win;
897 	le->flags = tcp_get_flags(th);
898 	le->needs_merge = 0;
899 
900 	/* Setup new data pointers. */
901 	le->m_head = m;
902 	le->m_tail = m_last(m);
903 }
904 
905 static void
906 tcp_push_and_replace(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m)
907 {
908 	struct lro_parser *pa;
909 
910 	/*
911 	 * Push up the stack of the current entry
912 	 * and replace it with "m".
913 	 */
914 	struct mbuf *msave;
915 
916 	/* Grab off the next and save it */
917 	msave = le->m_head->m_nextpkt;
918 	le->m_head->m_nextpkt = NULL;
919 
920 	/* Now push out the old entry */
921 	tcp_flush_out_entry(lc, le);
922 
923 	/* Re-parse new header, should not fail. */
924 	pa = tcp_lro_parser(m, &le->outer, &le->inner, false);
925 	KASSERT(pa != NULL,
926 	    ("tcp_push_and_replace: LRO parser failed on m=%p\n", m));
927 
928 	/*
929 	 * Now to replace the data properly in the entry
930 	 * we have to reset the TCP header and
931 	 * other fields.
932 	 */
933 	tcp_set_entry_to_mbuf(lc, le, m, pa->tcp);
934 
935 	/* Restore the next list */
936 	m->m_nextpkt = msave;
937 }
938 
939 static void
940 tcp_lro_mbuf_append_pkthdr(struct lro_entry *le, const struct mbuf *p)
941 {
942 	struct mbuf *m;
943 	uint32_t csum;
944 
945 	m = le->m_head;
946 	if (m->m_pkthdr.lro_nsegs == 1) {
947 		/* Compute relative checksum. */
948 		csum = p->m_pkthdr.lro_tcp_d_csum;
949 	} else {
950 		/* Merge TCP data checksums. */
951 		csum = (uint32_t)m->m_pkthdr.lro_tcp_d_csum +
952 		    (uint32_t)p->m_pkthdr.lro_tcp_d_csum;
953 		while (csum > 0xffff)
954 			csum = (csum >> 16) + (csum & 0xffff);
955 	}
956 
957 	/* Update various counters. */
958 	m->m_pkthdr.len += p->m_pkthdr.lro_tcp_d_len;
959 	m->m_pkthdr.lro_tcp_d_csum = csum;
960 	m->m_pkthdr.lro_tcp_d_len += p->m_pkthdr.lro_tcp_d_len;
961 	m->m_pkthdr.lro_nsegs += p->m_pkthdr.lro_nsegs;
962 	le->needs_merge = 1;
963 }
964 
965 static void
966 tcp_lro_condense(struct lro_ctrl *lc, struct lro_entry *le)
967 {
968 	/*
969 	 * Walk through the mbuf chain we
970 	 * have on tap and compress/condense
971 	 * as required.
972 	 */
973 	uint32_t *ts_ptr;
974 	struct mbuf *m;
975 	struct tcphdr *th;
976 	uint32_t tcp_data_len_total;
977 	uint32_t tcp_data_seg_total;
978 	uint16_t tcp_data_len;
979 	uint16_t tcp_opt_len;
980 
981 	/*
982 	 * First we must check the lead (m_head)
983 	 * we must make sure that it is *not*
984 	 * something that should be sent up
985 	 * right away (sack etc).
986 	 */
987 again:
988 	m = le->m_head->m_nextpkt;
989 	if (m == NULL) {
990 		/* Just one left. */
991 		return;
992 	}
993 
994 	th = tcp_lro_get_th(m);
995 	tcp_opt_len = (th->th_off << 2);
996 	tcp_opt_len -= sizeof(*th);
997 	ts_ptr = (uint32_t *)(th + 1);
998 
999 	if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1000 	    *ts_ptr != TCP_LRO_TS_OPTION)) {
1001 		/*
1002 		 * Its not the timestamp. We can't
1003 		 * use this guy as the head.
1004 		 */
1005 		le->m_head->m_nextpkt = m->m_nextpkt;
1006 		tcp_push_and_replace(lc, le, m);
1007 		goto again;
1008 	}
1009 	if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1010 		/*
1011 		 * Make sure that previously seen segments/ACKs are delivered
1012 		 * before this segment, e.g. FIN.
1013 		 */
1014 		le->m_head->m_nextpkt = m->m_nextpkt;
1015 		tcp_push_and_replace(lc, le, m);
1016 		goto again;
1017 	}
1018 	while((m = le->m_head->m_nextpkt) != NULL) {
1019 		/*
1020 		 * condense m into le, first
1021 		 * pull m out of the list.
1022 		 */
1023 		le->m_head->m_nextpkt = m->m_nextpkt;
1024 		m->m_nextpkt = NULL;
1025 		/* Setup my data */
1026 		tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
1027 		th = tcp_lro_get_th(m);
1028 		ts_ptr = (uint32_t *)(th + 1);
1029 		tcp_opt_len = (th->th_off << 2);
1030 		tcp_opt_len -= sizeof(*th);
1031 		tcp_data_len_total = le->m_head->m_pkthdr.lro_tcp_d_len + tcp_data_len;
1032 		tcp_data_seg_total = le->m_head->m_pkthdr.lro_nsegs + m->m_pkthdr.lro_nsegs;
1033 
1034 		if (tcp_data_seg_total >= lc->lro_ackcnt_lim ||
1035 		    tcp_data_len_total >= lc->lro_length_lim) {
1036 			/* Flush now if appending will result in overflow. */
1037 			tcp_push_and_replace(lc, le, m);
1038 			goto again;
1039 		}
1040 		if (tcp_opt_len != 0 &&
1041 		    __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1042 		    *ts_ptr != TCP_LRO_TS_OPTION)) {
1043 			/*
1044 			 * Maybe a sack in the new one? We need to
1045 			 * start all over after flushing the
1046 			 * current le. We will go up to the beginning
1047 			 * and flush it (calling the replace again possibly
1048 			 * or just returning).
1049 			 */
1050 			tcp_push_and_replace(lc, le, m);
1051 			goto again;
1052 		}
1053 		if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1054 			tcp_push_and_replace(lc, le, m);
1055 			goto again;
1056 		}
1057 		if (tcp_opt_len != 0) {
1058 			uint32_t tsval = ntohl(*(ts_ptr + 1));
1059 			/* Make sure timestamp values are increasing. */
1060 			if (TSTMP_GT(le->tsval, tsval))  {
1061 				tcp_push_and_replace(lc, le, m);
1062 				goto again;
1063 			}
1064 			le->tsval = tsval;
1065 			le->tsecr = *(ts_ptr + 2);
1066 		}
1067 		/* Try to append the new segment. */
1068 		if (__predict_false(ntohl(th->th_seq) != le->next_seq ||
1069 				    ((tcp_get_flags(th) & TH_ACK) !=
1070 				      (le->flags & TH_ACK)) ||
1071 				    (tcp_data_len == 0 &&
1072 				     le->ack_seq == th->th_ack &&
1073 				     le->window == th->th_win))) {
1074 			/* Out of order packet, non-ACK + ACK or dup ACK. */
1075 			tcp_push_and_replace(lc, le, m);
1076 			goto again;
1077 		}
1078 		if (tcp_data_len != 0 ||
1079 		    SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1080 			le->next_seq += tcp_data_len;
1081 			le->ack_seq = th->th_ack;
1082 			le->window = th->th_win;
1083 			le->needs_merge = 1;
1084 		} else if (th->th_ack == le->ack_seq) {
1085 			if (WIN_GT(th->th_win, le->window)) {
1086 				le->window = th->th_win;
1087 				le->needs_merge = 1;
1088 			}
1089 		}
1090 
1091 		if (tcp_data_len == 0) {
1092 			m_freem(m);
1093 			continue;
1094 		}
1095 
1096 		/* Merge TCP data checksum and length to head mbuf. */
1097 		tcp_lro_mbuf_append_pkthdr(le, m);
1098 
1099 		/*
1100 		 * Adjust the mbuf so that m_data points to the first byte of
1101 		 * the ULP payload.  Adjust the mbuf to avoid complications and
1102 		 * append new segment to existing mbuf chain.
1103 		 */
1104 		m_adj(m, m->m_pkthdr.len - tcp_data_len);
1105 		m_demote_pkthdr(m);
1106 		le->m_tail->m_next = m;
1107 		le->m_tail = m_last(m);
1108 	}
1109 }
1110 
1111 void
1112 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
1113 {
1114 
1115 	/* Only optimise if there are multiple packets waiting. */
1116 	NET_EPOCH_ASSERT();
1117 	if (tcp_lro_flush_tcphpts == NULL ||
1118 	    tcp_lro_flush_tcphpts(lc, le) != 0) {
1119 		tcp_lro_condense(lc, le);
1120 		tcp_flush_out_entry(lc, le);
1121 	}
1122 	lc->lro_flushed++;
1123 	bzero(le, sizeof(*le));
1124 	LIST_INSERT_HEAD(&lc->lro_free, le, next);
1125 }
1126 
1127 #define	tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
1128 
1129 /*
1130  * The tcp_lro_sort() routine is comparable to qsort(), except it has
1131  * a worst case complexity limit of O(MIN(N,64)*N), where N is the
1132  * number of elements to sort and 64 is the number of sequence bits
1133  * available. The algorithm is bit-slicing the 64-bit sequence number,
1134  * sorting one bit at a time from the most significant bit until the
1135  * least significant one, skipping the constant bits. This is
1136  * typically called a radix sort.
1137  */
1138 static void
1139 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
1140 {
1141 	struct lro_mbuf_sort temp;
1142 	uint64_t ones;
1143 	uint64_t zeros;
1144 	uint32_t x;
1145 	uint32_t y;
1146 
1147 repeat:
1148 	/* for small arrays insertion sort is faster */
1149 	if (size <= 12) {
1150 		for (x = 1; x < size; x++) {
1151 			temp = parray[x];
1152 			for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
1153 				parray[y] = parray[y - 1];
1154 			parray[y] = temp;
1155 		}
1156 		return;
1157 	}
1158 
1159 	/* compute sequence bits which are constant */
1160 	ones = 0;
1161 	zeros = 0;
1162 	for (x = 0; x != size; x++) {
1163 		ones |= parray[x].seq;
1164 		zeros |= ~parray[x].seq;
1165 	}
1166 
1167 	/* compute bits which are not constant into "ones" */
1168 	ones &= zeros;
1169 	if (ones == 0)
1170 		return;
1171 
1172 	/* pick the most significant bit which is not constant */
1173 	ones = tcp_lro_msb_64(ones);
1174 
1175 	/*
1176 	 * Move entries having cleared sequence bits to the beginning
1177 	 * of the array:
1178 	 */
1179 	for (x = y = 0; y != size; y++) {
1180 		/* skip set bits */
1181 		if (parray[y].seq & ones)
1182 			continue;
1183 		/* swap entries */
1184 		temp = parray[x];
1185 		parray[x] = parray[y];
1186 		parray[y] = temp;
1187 		x++;
1188 	}
1189 
1190 	KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
1191 
1192 	/* sort zeros */
1193 	tcp_lro_sort(parray, x);
1194 
1195 	/* sort ones */
1196 	parray += x;
1197 	size -= x;
1198 	goto repeat;
1199 }
1200 
1201 void
1202 tcp_lro_flush_all(struct lro_ctrl *lc)
1203 {
1204 	uint64_t seq;
1205 	uint64_t nseq;
1206 	unsigned x;
1207 
1208 	NET_EPOCH_ASSERT();
1209 	/* check if no mbufs to flush */
1210 	if (lc->lro_mbuf_count == 0)
1211 		goto done;
1212 	if (lc->lro_cpu_is_set == 0) {
1213 		if (lc->lro_last_cpu == curcpu) {
1214 			lc->lro_cnt_of_same_cpu++;
1215 			/* Have we reached the threshold to declare a cpu? */
1216 			if (lc->lro_cnt_of_same_cpu > tcp_lro_cpu_set_thresh)
1217 				lc->lro_cpu_is_set = 1;
1218 		} else {
1219 			lc->lro_last_cpu = curcpu;
1220 			lc->lro_cnt_of_same_cpu = 0;
1221 		}
1222 	}
1223 	CURVNET_SET(lc->ifp->if_vnet);
1224 
1225 	/* get current time */
1226 	binuptime(&lc->lro_last_queue_time);
1227 
1228 	/* sort all mbufs according to stream */
1229 	tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
1230 
1231 	/* input data into LRO engine, stream by stream */
1232 	seq = 0;
1233 	for (x = 0; x != lc->lro_mbuf_count; x++) {
1234 		struct mbuf *mb;
1235 
1236 		/* get mbuf */
1237 		mb = lc->lro_mbuf_data[x].mb;
1238 
1239 		/* get sequence number, masking away the packet index */
1240 		nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
1241 
1242 		/* check for new stream */
1243 		if (seq != nseq) {
1244 			seq = nseq;
1245 
1246 			/* flush active streams */
1247 			tcp_lro_rx_done(lc);
1248 		}
1249 
1250 		/* add packet to LRO engine */
1251 		if (tcp_lro_rx_common(lc, mb, 0, false) != 0) {
1252  			/* Flush anything we have acummulated */
1253  			tcp_lro_flush_active(lc);
1254 			/* input packet to network layer */
1255 			(*lc->ifp->if_input)(lc->ifp, mb);
1256 			lc->lro_queued++;
1257 			lc->lro_flushed++;
1258 		}
1259 	}
1260 	CURVNET_RESTORE();
1261 done:
1262 	/* flush active streams */
1263 	tcp_lro_rx_done(lc);
1264 	tcp_hpts_softclock();
1265 	lc->lro_mbuf_count = 0;
1266 }
1267 
1268 static struct lro_head *
1269 tcp_lro_rx_get_bucket(struct lro_ctrl *lc, struct mbuf *m, struct lro_parser *parser)
1270 {
1271 	u_long hash;
1272 
1273 	if (M_HASHTYPE_ISHASH(m)) {
1274 		hash = m->m_pkthdr.flowid;
1275 	} else {
1276 		for (unsigned i = hash = 0; i != LRO_RAW_ADDRESS_MAX; i++)
1277 			hash += parser->data.raw[i];
1278 	}
1279 	return (&lc->lro_hash[hash % lc->lro_hashsz]);
1280 }
1281 
1282 static int
1283 tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash)
1284 {
1285 	struct lro_parser pi;	/* inner address data */
1286 	struct lro_parser po;	/* outer address data */
1287 	struct lro_parser *pa;	/* current parser for TCP stream */
1288 	struct lro_entry *le;
1289 	struct lro_head *bucket;
1290 	struct tcphdr *th;
1291 	int tcp_data_len;
1292 	int tcp_opt_len;
1293 	int error;
1294 	uint16_t tcp_data_sum;
1295 
1296 #ifdef INET
1297 	/* Quickly decide if packet cannot be LRO'ed */
1298 	if (__predict_false(V_ipforwarding != 0))
1299 		return (TCP_LRO_CANNOT);
1300 #endif
1301 #ifdef INET6
1302 	/* Quickly decide if packet cannot be LRO'ed */
1303 	if (__predict_false(V_ip6_forwarding != 0))
1304 		return (TCP_LRO_CANNOT);
1305 #endif
1306 	if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1307 	     ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1308 	    (m->m_pkthdr.csum_data != 0xffff)) {
1309 		/*
1310 		 * The checksum either did not have hardware offload
1311 		 * or it was a bad checksum. We can't LRO such
1312 		 * a packet.
1313 		 */
1314 		counter_u64_add(tcp_bad_csums, 1);
1315 		return (TCP_LRO_CANNOT);
1316 	}
1317 	/* We expect a contiguous header [eh, ip, tcp]. */
1318 	pa = tcp_lro_parser(m, &po, &pi, true);
1319 	if (__predict_false(pa == NULL))
1320 		return (TCP_LRO_NOT_SUPPORTED);
1321 
1322 	/* We don't expect any padding. */
1323 	error = tcp_lro_trim_mbuf_chain(m, pa);
1324 	if (__predict_false(error != 0))
1325 		return (error);
1326 
1327 #ifdef INET
1328 	switch (pa->data.lro_type) {
1329 	case LRO_TYPE_IPV4_TCP:
1330 		error = tcp_lro_rx_ipv4(lc, m, pa->ip4);
1331 		if (__predict_false(error != 0))
1332 			return (error);
1333 		break;
1334 	default:
1335 		break;
1336 	}
1337 #endif
1338 	/* If no hardware or arrival stamp on the packet add timestamp */
1339 	if ((m->m_flags & (M_TSTMP_LRO | M_TSTMP)) == 0) {
1340 		m->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
1341 		m->m_flags |= M_TSTMP_LRO;
1342 	}
1343 
1344 	/* Get pointer to TCP header. */
1345 	th = pa->tcp;
1346 
1347 	/* Don't process SYN packets. */
1348 	if (__predict_false(tcp_get_flags(th) & TH_SYN))
1349 		return (TCP_LRO_CANNOT);
1350 
1351 	/* Get total TCP header length and compute payload length. */
1352 	tcp_opt_len = (th->th_off << 2);
1353 	tcp_data_len = m->m_pkthdr.len - ((uint8_t *)th -
1354 	    (uint8_t *)m->m_data) - tcp_opt_len;
1355 	tcp_opt_len -= sizeof(*th);
1356 
1357 	/* Don't process invalid TCP headers. */
1358 	if (__predict_false(tcp_opt_len < 0 || tcp_data_len < 0))
1359 		return (TCP_LRO_CANNOT);
1360 
1361 	/* Compute TCP data only checksum. */
1362 	if (tcp_data_len == 0)
1363 		tcp_data_sum = 0;	/* no data, no checksum */
1364 	else if (__predict_false(csum != 0))
1365 		tcp_data_sum = tcp_lro_rx_csum_data(pa, ~csum);
1366 	else
1367 		tcp_data_sum = tcp_lro_rx_csum_data(pa, ~th->th_sum);
1368 
1369 	/* Save TCP info in mbuf. */
1370 	m->m_nextpkt = NULL;
1371 	m->m_pkthdr.rcvif = lc->ifp;
1372 	m->m_pkthdr.lro_tcp_d_csum = tcp_data_sum;
1373 	m->m_pkthdr.lro_tcp_d_len = tcp_data_len;
1374 	m->m_pkthdr.lro_tcp_h_off = ((uint8_t *)th - (uint8_t *)m->m_data);
1375 	m->m_pkthdr.lro_nsegs = 1;
1376 
1377 	/* Get hash bucket. */
1378 	if (!use_hash) {
1379 		bucket = &lc->lro_hash[0];
1380 	} else {
1381 		bucket = tcp_lro_rx_get_bucket(lc, m, pa);
1382 	}
1383 
1384 	/* Try to find a matching previous segment. */
1385 	LIST_FOREACH(le, bucket, hash_next) {
1386 		/* Compare addresses and ports. */
1387 		if (lro_address_compare(&po.data, &le->outer.data) == false ||
1388 		    lro_address_compare(&pi.data, &le->inner.data) == false)
1389 			continue;
1390 
1391 		/* Check if no data and old ACK. */
1392 		if (tcp_data_len == 0 &&
1393 		    SEQ_LT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1394 			m_freem(m);
1395 			return (0);
1396 		}
1397 
1398 		/* Mark "m" in the last spot. */
1399 		le->m_last_mbuf->m_nextpkt = m;
1400 		/* Now set the tail to "m". */
1401 		le->m_last_mbuf = m;
1402 		return (0);
1403 	}
1404 
1405 	/* Try to find an empty slot. */
1406 	if (LIST_EMPTY(&lc->lro_free))
1407 		return (TCP_LRO_NO_ENTRIES);
1408 
1409 	/* Start a new segment chain. */
1410 	le = LIST_FIRST(&lc->lro_free);
1411 	LIST_REMOVE(le, next);
1412 	tcp_lro_active_insert(lc, bucket, le);
1413 
1414 	/* Make sure the headers are set. */
1415 	le->inner = pi;
1416 	le->outer = po;
1417 
1418 	/* Store time this entry was allocated. */
1419 	le->alloc_time = lc->lro_last_queue_time;
1420 
1421 	tcp_set_entry_to_mbuf(lc, le, m, th);
1422 
1423 	/* Now set the tail to "m". */
1424 	le->m_last_mbuf = m;
1425 
1426 	return (0);
1427 }
1428 
1429 int
1430 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
1431 {
1432 	int error;
1433 
1434 	if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1435 	     ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1436 	    (m->m_pkthdr.csum_data != 0xffff)) {
1437 		/*
1438 		 * The checksum either did not have hardware offload
1439 		 * or it was a bad checksum. We can't LRO such
1440 		 * a packet.
1441 		 */
1442 		counter_u64_add(tcp_bad_csums, 1);
1443 		return (TCP_LRO_CANNOT);
1444 	}
1445 	/* get current time */
1446 	binuptime(&lc->lro_last_queue_time);
1447 	CURVNET_SET(lc->ifp->if_vnet);
1448 	error = tcp_lro_rx_common(lc, m, csum, true);
1449 	if (__predict_false(error != 0)) {
1450 		/*
1451 		 * Flush anything we have acummulated
1452 		 * ahead of this packet that can't
1453 		 * be LRO'd. This preserves order.
1454 		 */
1455 		tcp_lro_flush_active(lc);
1456 	}
1457 	CURVNET_RESTORE();
1458 
1459 	return (error);
1460 }
1461 
1462 void
1463 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
1464 {
1465 	NET_EPOCH_ASSERT();
1466 	/* sanity checks */
1467 	if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
1468 	    lc->lro_mbuf_max == 0)) {
1469 		/* packet drop */
1470 		m_freem(mb);
1471 		return;
1472 	}
1473 
1474 	/* check if packet is not LRO capable */
1475 	if (__predict_false((lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
1476 		/* input packet to network layer */
1477 		(*lc->ifp->if_input) (lc->ifp, mb);
1478 		return;
1479 	}
1480 
1481  	/* If no hardware or arrival stamp on the packet add timestamp */
1482  	if ((tcplro_stacks_wanting_mbufq > 0) &&
1483  	    (tcp_less_accurate_lro_ts == 0) &&
1484  	    ((mb->m_flags & M_TSTMP) == 0)) {
1485  		/* Add in an LRO time since no hardware */
1486  		binuptime(&lc->lro_last_queue_time);
1487  		mb->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
1488  		mb->m_flags |= M_TSTMP_LRO;
1489  	}
1490 
1491 	/* create sequence number */
1492 	lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
1493 	    (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
1494 	    (((uint64_t)mb->m_pkthdr.flowid) << 24) |
1495 	    ((uint64_t)lc->lro_mbuf_count);
1496 
1497 	/* enter mbuf */
1498 	lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb;
1499 
1500 	/* flush if array is full */
1501 	if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max))
1502 		tcp_lro_flush_all(lc);
1503 }
1504 
1505 /* end */
1506