tcp_rxtx.c revision 71ba97fe
1/*
2 * Copyright (c) 2016-2017  Intel Corporation.
3 * Licensed under the Apache License, Version 2.0 (the "License");
4 * you may not use this file except in compliance with the License.
5 * You may obtain a copy of the License at:
6 *
7 *     http://www.apache.org/licenses/LICENSE-2.0
8 *
9 * Unless required by applicable law or agreed to in writing, software
10 * distributed under the License is distributed on an "AS IS" BASIS,
11 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 * See the License for the specific language governing permissions and
13 * limitations under the License.
14 */
15
16#include <rte_errno.h>
17#include <rte_ethdev.h>
18#include <rte_ip.h>
19#include <rte_ip_frag.h>
20#include <rte_tcp.h>
21
22#include "tcp_stream.h"
23#include "tcp_timer.h"
24#include "stream_table.h"
25#include "syncookie.h"
26#include "misc.h"
27#include "tcp_ctl.h"
28#include "tcp_rxq.h"
29#include "tcp_txq.h"
30#include "tcp_tx_seg.h"
31
32#define	TCP_MAX_PKT_SEG	0x20
33
34/*
35 * checks if input TCP ports and IP addresses match given stream.
36 * returns zero on success.
37 */
38static inline int
39rx_check_stream(const struct tle_tcp_stream *s, const union pkt_info *pi)
40{
41	int32_t rc;
42
43	if (pi->tf.type == TLE_V4)
44		rc = (pi->port.raw & s->s.pmsk.raw) != s->s.port.raw ||
45			(pi->addr4.raw & s->s.ipv4.mask.raw) !=
46			s->s.ipv4.addr.raw;
47	else
48		rc = (pi->port.raw & s->s.pmsk.raw) != s->s.port.raw ||
49			ymm_mask_cmp(&pi->addr6->raw, &s->s.ipv6.addr.raw,
50			&s->s.ipv6.mask.raw) != 0;
51
52	return rc;
53}
54
55static inline struct tle_tcp_stream *
56rx_obtain_listen_stream(const struct tle_dev *dev, const union pkt_info *pi,
57	uint32_t type)
58{
59	struct tle_tcp_stream *s;
60
61	s = (struct tle_tcp_stream *)dev->dp[type]->streams[pi->port.dst];
62	if (s == NULL || tcp_stream_acquire(s) < 0)
63		return NULL;
64
65	/* check that we have a proper stream. */
66	if (s->tcb.state != TCP_ST_LISTEN) {
67		tcp_stream_release(s);
68		s = NULL;
69	}
70
71	return s;
72}
73
74static inline struct tle_tcp_stream *
75rx_obtain_stream(const struct tle_dev *dev, struct stbl *st,
76	const union pkt_info *pi, uint32_t type)
77{
78	struct tle_tcp_stream *s;
79
80	s = stbl_find_data(st, pi);
81	if (s == NULL) {
82		if (pi->tf.flags == TCP_FLAG_ACK)
83			return rx_obtain_listen_stream(dev, pi, type);
84		return NULL;
85	}
86
87	if (tcp_stream_acquire(s) < 0)
88		return NULL;
89	/* check that we have a proper stream. */
90	else if (s->tcb.state == TCP_ST_CLOSED) {
91		tcp_stream_release(s);
92		s = NULL;
93	}
94
95	return s;
96}
97
98/*
99 * Consider 2 pkt_info *equal* if their:
100 * - types (IPv4/IPv6)
101 * - TCP flags
102 * - checksum flags
103 * - TCP src and dst ports
104 * - IP src and dst addresses
105 * are equal.
106 */
107static inline int
108pkt_info_bulk_eq(const union pkt_info pi[], uint32_t num)
109{
110	uint32_t i;
111
112	i = 1;
113
114	if (pi[0].tf.type == TLE_V4) {
115		while (i != num && xmm_cmp(&pi[0].raw, &pi[i].raw) == 0)
116			i++;
117
118	} else if (pi[0].tf.type == TLE_V6) {
119		while (i != num &&
120				pi[0].raw.u64[0] == pi[i].raw.u64[0] &&
121				ymm_cmp(&pi[0].addr6->raw,
122				&pi[i].addr6->raw) == 0)
123			i++;
124	}
125
126	return i;
127}
128
129static inline int
130pkt_info_bulk_syneq(const union pkt_info pi[], uint32_t num)
131{
132	uint32_t i;
133
134	i = 1;
135
136	if (pi[0].tf.type == TLE_V4) {
137		while (i != num && pi[0].tf.raw == pi[i].tf.raw &&
138				pi[0].port.dst == pi[i].port.dst &&
139				pi[0].addr4.dst == pi[i].addr4.dst)
140			i++;
141
142	} else if (pi[0].tf.type == TLE_V6) {
143		while (i != num && pi[0].tf.raw == pi[i].tf.raw &&
144				pi[0].port.dst == pi[i].port.dst &&
145				xmm_cmp(&pi[0].addr6->dst,
146				&pi[i].addr6->dst) == 0)
147			i++;
148	}
149
150	return i;
151}
152
153static inline void
154stream_drb_free(struct tle_tcp_stream *s, struct tle_drb *drbs[],
155	uint32_t nb_drb)
156{
157	_rte_ring_enqueue_burst(s->tx.drb.r, (void **)drbs, nb_drb);
158}
159
160static inline uint32_t
161stream_drb_alloc(struct tle_tcp_stream *s, struct tle_drb *drbs[],
162	uint32_t nb_drb)
163{
164	return _rte_ring_dequeue_burst(s->tx.drb.r, (void **)drbs, nb_drb);
165}
166
167static inline uint32_t
168get_ip_pid(struct tle_dev *dev, uint32_t num, uint32_t type, uint32_t st)
169{
170	uint32_t pid;
171	rte_atomic32_t *pa;
172
173	pa = &dev->tx.packet_id[type];
174
175	if (st == 0) {
176		pid = rte_atomic32_add_return(pa, num);
177		return pid - num;
178	} else {
179		pid = rte_atomic32_read(pa);
180		rte_atomic32_set(pa, pid + num);
181		return pid;
182	}
183}
184
185static inline void
186fill_tcph(struct tcp_hdr *l4h, const struct tcb *tcb, union l4_ports port,
187	uint32_t seq, uint8_t hlen, uint8_t flags)
188{
189	uint16_t wnd;
190
191	l4h->src_port = port.dst;
192	l4h->dst_port = port.src;
193
194	wnd = (flags & TCP_FLAG_SYN) ?
195		RTE_MIN(tcb->rcv.wnd, (uint32_t)UINT16_MAX) :
196		tcb->rcv.wnd >> tcb->rcv.wscale;
197
198	/* ??? use sse shuffle to hton all remaining 16 bytes at once. ??? */
199	l4h->sent_seq = rte_cpu_to_be_32(seq);
200	l4h->recv_ack = rte_cpu_to_be_32(tcb->rcv.nxt);
201	l4h->data_off = hlen / TCP_DATA_ALIGN << TCP_DATA_OFFSET;
202	l4h->tcp_flags = flags;
203	l4h->rx_win = rte_cpu_to_be_16(wnd);
204	l4h->cksum = 0;
205	l4h->tcp_urp = 0;
206
207	if (flags & TCP_FLAG_SYN)
208		fill_syn_opts(l4h + 1, &tcb->so);
209	else if ((flags & TCP_FLAG_RST) == 0 && tcb->so.ts.raw != 0)
210		fill_tms_opts(l4h + 1, tcb->snd.ts, tcb->rcv.ts);
211}
212
213static inline int
214tcp_fill_mbuf(struct rte_mbuf *m, const struct tle_tcp_stream *s,
215	const struct tle_dest *dst, uint64_t ol_flags,
216	union l4_ports port, uint32_t seq, uint32_t flags,
217	uint32_t pid, uint32_t swcsm)
218{
219	uint32_t l4, len, plen;
220	struct tcp_hdr *l4h;
221	char *l2h;
222
223	len = dst->l2_len + dst->l3_len;
224	plen = m->pkt_len;
225
226	if (flags & TCP_FLAG_SYN)
227		l4 = sizeof(*l4h) + TCP_TX_OPT_LEN_MAX;
228	else if ((flags & TCP_FLAG_RST) == 0 && s->tcb.rcv.ts != 0)
229		l4 = sizeof(*l4h) + TCP_TX_OPT_LEN_TMS;
230	else
231		l4 = sizeof(*l4h);
232
233	/* adjust mbuf to put L2/L3/L4 headers into it. */
234	l2h = rte_pktmbuf_prepend(m, len + l4);
235	if (l2h == NULL)
236		return -EINVAL;
237
238	/* copy L2/L3 header */
239	rte_memcpy(l2h, dst->hdr, len);
240
241	/* setup TCP header & options */
242	l4h = (struct tcp_hdr *)(l2h + len);
243	fill_tcph(l4h, &s->tcb, port, seq, l4, flags);
244
245	/* setup mbuf TX offload related fields. */
246	m->tx_offload = _mbuf_tx_offload(dst->l2_len, dst->l3_len, l4, 0, 0, 0);
247	m->ol_flags |= ol_flags;
248
249	/* update proto specific fields. */
250
251	if (s->s.type == TLE_V4) {
252		struct ipv4_hdr *l3h;
253		l3h = (struct ipv4_hdr *)(l2h + dst->l2_len);
254		l3h->packet_id = rte_cpu_to_be_16(pid);
255		l3h->total_length = rte_cpu_to_be_16(plen + dst->l3_len + l4);
256
257		if ((ol_flags & PKT_TX_TCP_CKSUM) != 0)
258			l4h->cksum = _ipv4x_phdr_cksum(l3h, m->l3_len,
259				ol_flags);
260		else if (swcsm != 0)
261			l4h->cksum = _ipv4_udptcp_mbuf_cksum(m, len, l3h);
262
263		if ((ol_flags & PKT_TX_IP_CKSUM) == 0 && swcsm != 0)
264			l3h->hdr_checksum = _ipv4x_cksum(l3h, m->l3_len);
265	} else {
266		struct ipv6_hdr *l3h;
267		l3h = (struct ipv6_hdr *)(l2h + dst->l2_len);
268		l3h->payload_len = rte_cpu_to_be_16(plen + l4);
269		if ((ol_flags & PKT_TX_TCP_CKSUM) != 0)
270			l4h->cksum = rte_ipv6_phdr_cksum(l3h, ol_flags);
271		else if (swcsm != 0)
272			l4h->cksum = _ipv6_udptcp_mbuf_cksum(m, len, l3h);
273	}
274
275	return 0;
276}
277
278/*
279 * That function supposed to be used only for data packets.
280 * Assumes that L2/L3/L4 headers and mbuf fields already setup properly.
281 *  - updates tcp SEG.SEQ, SEG.ACK, TS.VAL, TS.ECR.
282 *  - if no HW cksum offloads are enabled, calculates TCP checksum.
283 */
284static inline void
285tcp_update_mbuf(struct rte_mbuf *m, uint32_t type, const struct tcb *tcb,
286	uint32_t seq, uint32_t pid)
287{
288	struct tcp_hdr *l4h;
289	uint32_t len;
290
291	len = m->l2_len + m->l3_len;
292	l4h = rte_pktmbuf_mtod_offset(m, struct tcp_hdr *, len);
293
294	l4h->sent_seq = rte_cpu_to_be_32(seq);
295	l4h->recv_ack = rte_cpu_to_be_32(tcb->rcv.nxt);
296
297	if (tcb->so.ts.raw != 0)
298		fill_tms_opts(l4h + 1, tcb->snd.ts, tcb->rcv.ts);
299
300	if (type == TLE_V4) {
301		struct ipv4_hdr *l3h;
302		l3h = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *, m->l2_len);
303		l3h->hdr_checksum = 0;
304		l3h->packet_id = rte_cpu_to_be_16(pid);
305		if ((m->ol_flags & PKT_TX_IP_CKSUM) == 0)
306			l3h->hdr_checksum = _ipv4x_cksum(l3h, m->l3_len);
307	}
308
309	/* have to calculate TCP checksum in SW */
310	if ((m->ol_flags & PKT_TX_TCP_CKSUM) == 0) {
311
312		l4h->cksum = 0;
313
314		if (type == TLE_V4) {
315			struct ipv4_hdr *l3h;
316			l3h = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *,
317				m->l2_len);
318			l4h->cksum = _ipv4_udptcp_mbuf_cksum(m, len, l3h);
319
320		} else {
321			struct ipv6_hdr *l3h;
322			l3h = rte_pktmbuf_mtod_offset(m, struct ipv6_hdr *,
323				m->l2_len);
324			l4h->cksum = _ipv6_udptcp_mbuf_cksum(m, len, l3h);
325		}
326	}
327}
328
329/* Send data packets that need to be ACK-ed by peer */
330static inline uint32_t
331tx_data_pkts(struct tle_tcp_stream *s, struct rte_mbuf *const m[], uint32_t num)
332{
333	uint32_t bsz, i, nb, nbm;
334	struct tle_dev *dev;
335	struct tle_drb *drb[num];
336
337	/* calculate how many drbs are needed.*/
338	bsz = s->tx.drb.nb_elem;
339	nbm = (num + bsz - 1) / bsz;
340
341	/* allocate drbs, adjust number of packets. */
342	nb = stream_drb_alloc(s, drb, nbm);
343
344	/* drb ring is empty. */
345	if (nb == 0)
346		return 0;
347
348	else if (nb != nbm)
349		num = nb * bsz;
350
351	dev = s->tx.dst.dev;
352
353	/* enqueue pkts for TX. */
354	nbm = nb;
355	i = tle_dring_mp_enqueue(&dev->tx.dr, (const void * const*)m,
356		num, drb, &nb);
357
358	/* free unused drbs. */
359	if (nb != 0)
360		stream_drb_free(s, drb + nbm - nb, nb);
361
362	return i;
363}
364
365static inline uint32_t
366tx_data_bulk(struct tle_tcp_stream *s, union seqlen *sl, struct rte_mbuf *mi[],
367	uint32_t num)
368{
369	uint32_t fail, i, k, n, mss, pid, plen, sz, tn, type;
370	struct tle_dev *dev;
371	struct rte_mbuf *mb;
372	struct rte_mbuf *mo[MAX_PKT_BURST + TCP_MAX_PKT_SEG];
373
374	mss = s->tcb.snd.mss;
375	type = s->s.type;
376
377	dev = s->tx.dst.dev;
378	pid = get_ip_pid(dev, num, type, (s->flags & TLE_CTX_FLAG_ST) != 0);
379
380	k = 0;
381	tn = 0;
382	fail = 0;
383	for (i = 0; i != num && sl->len != 0 && fail == 0; i++) {
384
385		mb = mi[i];
386		sz = RTE_MIN(sl->len, mss);
387		plen = PKT_L4_PLEN(mb);
388
389		/*fast path, no need to use indirect mbufs. */
390		if (plen <= sz) {
391
392			/* update pkt TCP header */
393			tcp_update_mbuf(mb, type, &s->tcb, sl->seq, pid + i);
394
395			/* keep mbuf till ACK is received. */
396			rte_pktmbuf_refcnt_update(mb, 1);
397			sl->len -= plen;
398			sl->seq += plen;
399			mo[k++] = mb;
400		/* remaining snd.wnd is less them MSS, send nothing */
401		} else if (sz < mss)
402			break;
403		/* packet indirection needed */
404		else
405			RTE_VERIFY(0);
406
407		if (k >= MAX_PKT_BURST) {
408			n = tx_data_pkts(s, mo, k);
409			fail = k - n;
410			tn += n;
411			k = 0;
412		}
413	}
414
415	if (k != 0) {
416		n = tx_data_pkts(s, mo, k);
417		fail = k - n;
418		tn += n;
419	}
420
421	if (fail != 0) {
422		sz = tcp_mbuf_seq_free(mo + n, fail);
423		sl->seq -= sz;
424		sl->len += sz;
425	}
426
427	return tn;
428}
429
430/*
431 * gets data from stream send buffer, updates it and
432 * queues it into TX device queue.
433 * Note that this function and is not MT safe.
434 */
435static inline uint32_t
436tx_nxt_data(struct tle_tcp_stream *s, uint32_t tms)
437{
438	uint32_t n, num, tn, wnd;
439	struct rte_mbuf **mi;
440	union seqlen sl;
441
442	tn = 0;
443	wnd = s->tcb.snd.wnd - (uint32_t)(s->tcb.snd.nxt - s->tcb.snd.una);
444	sl.seq = s->tcb.snd.nxt;
445	sl.len = RTE_MIN(wnd, s->tcb.snd.cwnd);
446
447	if (sl.len == 0)
448		return tn;
449
450	/* update send timestamp */
451	s->tcb.snd.ts = tms;
452
453	do {
454		/* get group of packets */
455		mi = tcp_txq_get_nxt_objs(s, &num);
456
457		/* stream send buffer is empty */
458		if (num == 0)
459			break;
460
461		/* queue data packets for TX */
462		n = tx_data_bulk(s, &sl, mi, num);
463		tn += n;
464
465		/* update consumer head */
466		tcp_txq_set_nxt_head(s, n);
467	} while (n == num);
468
469	s->tcb.snd.nxt += sl.seq - (uint32_t)s->tcb.snd.nxt;
470	return tn;
471}
472
473static inline void
474free_una_data(struct tle_tcp_stream *s, uint32_t len)
475{
476	uint32_t i, num, plen;
477	struct rte_mbuf **mi;
478
479	plen = 0;
480
481	do {
482		/* get group of packets */
483		mi = tcp_txq_get_una_objs(s, &num);
484
485		if (num == 0)
486			break;
487
488		/* free acked data */
489		for (i = 0; i != num && plen != len; i++) {
490			uint32_t next_pkt_len = PKT_L4_PLEN(mi[i]);
491			if (plen + next_pkt_len > len) {
492				/* keep SND.UNA at the start of the packet */
493				len = plen;
494				break;
495			} else {
496				plen += next_pkt_len;
497			}
498			rte_pktmbuf_free(mi[i]);
499		}
500
501		/* update consumer tail */
502		tcp_txq_set_una_tail(s, i);
503	} while (plen < len);
504
505	s->tcb.snd.una += len;
506
507	/*
508	 * that could happen in case of retransmit,
509	 * adjust SND.NXT with SND.UNA.
510	 */
511	if (s->tcb.snd.una > s->tcb.snd.nxt) {
512		tcp_txq_rst_nxt_head(s);
513		s->tcb.snd.nxt = s->tcb.snd.una;
514	}
515}
516
517static inline uint16_t
518calc_smss(uint16_t mss, const struct tle_dest *dst)
519{
520	uint16_t n;
521
522	n = dst->mtu - dst->l2_len - dst->l3_len - TCP_TX_HDR_DACK;
523	mss = RTE_MIN(n, mss);
524	return mss;
525}
526
527/*
528 * RFC 6928 2
529 * min (10*MSS, max (2*MSS, 14600))
530 *
531 * or using user provided initial congestion window (icw)
532 * min (10*MSS, max (2*MSS, icw))
533 */
534static inline uint32_t
535initial_cwnd(uint32_t smss, uint32_t icw)
536{
537	return RTE_MIN(10 * smss, RTE_MAX(2 * smss, icw));
538}
539
540/*
541 * queue standalone packet to he particular output device
542 * It assumes that:
543 * - L2/L3/L4 headers should be already set.
544 * - packet fits into one segment.
545 */
546static inline int
547send_pkt(struct tle_tcp_stream *s, struct tle_dev *dev, struct rte_mbuf *m)
548{
549	uint32_t n, nb;
550	struct tle_drb *drb;
551
552	if (stream_drb_alloc(s, &drb, 1) == 0)
553		return -ENOBUFS;
554
555	/* enqueue pkt for TX. */
556	nb = 1;
557	n = tle_dring_mp_enqueue(&dev->tx.dr, (const void * const*)&m, 1,
558		&drb, &nb);
559
560	/* free unused drbs. */
561	if (nb != 0)
562		stream_drb_free(s, &drb, 1);
563
564	return (n == 1) ? 0 : -ENOBUFS;
565}
566
567static inline int
568send_ctrl_pkt(struct tle_tcp_stream *s, struct rte_mbuf *m, uint32_t seq,
569	uint32_t flags)
570{
571	const struct tle_dest *dst;
572	uint32_t pid, type;
573	int32_t rc;
574
575	dst = &s->tx.dst;
576	type = s->s.type;
577	pid = get_ip_pid(dst->dev, 1, type, (s->flags & TLE_CTX_FLAG_ST) != 0);
578
579	rc = tcp_fill_mbuf(m, s, dst, 0, s->s.port, seq, flags, pid, 1);
580	if (rc == 0)
581		rc = send_pkt(s, dst->dev, m);
582
583	return rc;
584}
585
586static inline int
587send_rst(struct tle_tcp_stream *s, uint32_t seq)
588{
589	struct rte_mbuf *m;
590	int32_t rc;
591
592	m = rte_pktmbuf_alloc(s->tx.dst.head_mp);
593	if (m == NULL)
594		return -ENOMEM;
595
596	rc = send_ctrl_pkt(s, m, seq, TCP_FLAG_RST);
597	if (rc != 0)
598		rte_pktmbuf_free(m);
599
600	return rc;
601}
602
603static inline int
604send_ack(struct tle_tcp_stream *s, uint32_t tms, uint32_t flags)
605{
606	struct rte_mbuf *m;
607	uint32_t seq;
608	int32_t rc;
609
610	m = rte_pktmbuf_alloc(s->tx.dst.head_mp);
611	if (m == NULL)
612		return -ENOMEM;
613
614	seq = s->tcb.snd.nxt - ((flags & (TCP_FLAG_FIN | TCP_FLAG_SYN)) != 0);
615	s->tcb.snd.ts = tms;
616
617	rc = send_ctrl_pkt(s, m, seq, flags);
618	if (rc != 0) {
619		rte_pktmbuf_free(m);
620		return rc;
621	}
622
623	s->tcb.snd.ack = s->tcb.rcv.nxt;
624	return 0;
625}
626
627
628static int
629sync_ack(struct tle_tcp_stream *s, const union pkt_info *pi,
630	const union seg_info *si, uint32_t ts, struct rte_mbuf *m)
631{
632	uint16_t len;
633	int32_t rc;
634	uint32_t pid, seq, type;
635	struct tle_dev *dev;
636	const void *da;
637	struct tle_dest dst;
638	const struct tcp_hdr *th;
639
640	type = s->s.type;
641
642	/* get destination information. */
643	if (type == TLE_V4)
644		da = &pi->addr4.src;
645	else
646		da = &pi->addr6->src;
647
648	rc = stream_get_dest(&s->s, da, &dst);
649	if (rc < 0)
650		return rc;
651
652	th = rte_pktmbuf_mtod_offset(m, const struct tcp_hdr *,
653		m->l2_len + m->l3_len);
654	get_syn_opts(&s->tcb.so, (uintptr_t)(th + 1), m->l4_len - sizeof(*th));
655
656	/* reset wscale option if timestamp is not present */
657	if (s->tcb.so.ts.val == 0)
658		s->tcb.so.wscale = 0;
659
660	s->tcb.rcv.nxt = si->seq + 1;
661	seq = sync_gen_seq(pi, s->tcb.rcv.nxt, ts, s->tcb.so.mss,
662				s->s.ctx->prm.hash_alg,
663				&s->s.ctx->prm.secret_key);
664	s->tcb.so.ts.ecr = s->tcb.so.ts.val;
665	s->tcb.so.ts.val = sync_gen_ts(ts, s->tcb.so.wscale);
666	s->tcb.so.wscale = (s->tcb.so.wscale == TCP_WSCALE_NONE) ?
667		TCP_WSCALE_NONE : TCP_WSCALE_DEFAULT;
668	s->tcb.so.mss = calc_smss(dst.mtu, &dst);
669
670	/* reset mbuf's data contents. */
671	len = m->l2_len + m->l3_len + m->l4_len;
672	m->tx_offload = 0;
673	if (rte_pktmbuf_adj(m, len) == NULL)
674		return -EINVAL;
675
676	dev = dst.dev;
677	pid = get_ip_pid(dev, 1, type, (s->flags & TLE_CTX_FLAG_ST) != 0);
678
679	rc = tcp_fill_mbuf(m, s, &dst, 0, pi->port, seq,
680		TCP_FLAG_SYN | TCP_FLAG_ACK, pid, 1);
681	if (rc == 0)
682		rc = send_pkt(s, dev, m);
683
684	return rc;
685}
686
687/*
688 * RFC 793:
689 * There are four cases for the acceptability test for an incoming segment:
690 * Segment Receive  Test
691 * Length  Window
692 * ------- -------  -------------------------------------------
693 *    0       0     SEG.SEQ = RCV.NXT
694 *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
695 *   >0       0     not acceptable
696 *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
697 *                  or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
698 */
699static inline int
700check_seqn(const struct tcb *tcb, uint32_t seqn, uint32_t len)
701{
702	uint32_t n;
703
704	n = seqn + len;
705	if (seqn - tcb->rcv.nxt >= tcb->rcv.wnd &&
706			n - tcb->rcv.nxt > tcb->rcv.wnd)
707		return -ERANGE;
708
709	return 0;
710}
711
712static inline union tsopt
713rx_tms_opt(const struct tcb *tcb, const struct rte_mbuf *mb)
714{
715	union tsopt ts;
716	uintptr_t opt;
717	const struct tcp_hdr *th;
718
719	if (tcb->so.ts.val != 0) {
720		opt = rte_pktmbuf_mtod_offset(mb, uintptr_t,
721			mb->l2_len + mb->l3_len + sizeof(*th));
722		ts = get_tms_opts(opt, mb->l4_len - sizeof(*th));
723	} else
724		ts.raw = 0;
725
726	return ts;
727}
728
729/*
730 * PAWS and sequence check.
731 * RFC 1323 4.2.1
732 */
733static inline int
734rx_check_seq(struct tcb *tcb, uint32_t seq, uint32_t len, const union tsopt ts)
735{
736	int32_t rc;
737
738	/* RFC 1323 4.2.1 R2 */
739	rc = check_seqn(tcb, seq, len);
740	if (rc < 0)
741		return rc;
742
743	if (ts.raw != 0) {
744
745		/* RFC 1323 4.2.1 R1 */
746		if (tcp_seq_lt(ts.val, tcb->rcv.ts))
747			return -ERANGE;
748
749		/* RFC 1323 4.2.1 R3 */
750		if (tcp_seq_leq(seq, tcb->snd.ack) &&
751				tcp_seq_lt(tcb->snd.ack, seq + len))
752			tcb->rcv.ts = ts.val;
753	}
754
755	return rc;
756}
757
758static inline int
759rx_check_ack(const struct tcb *tcb, uint32_t ack)
760{
761	uint32_t max;
762
763	max = (uint32_t)RTE_MAX(tcb->snd.nxt, tcb->snd.rcvr);
764
765	if (tcp_seq_leq(tcb->snd.una, ack) && tcp_seq_leq(ack, max))
766		return 0;
767
768	return -ERANGE;
769}
770
771static inline int
772rx_check_seqack(struct tcb *tcb, uint32_t seq, uint32_t ack, uint32_t len,
773	const union tsopt ts)
774{
775	int32_t rc;
776
777	rc = rx_check_seq(tcb, seq, len, ts);
778	rc |= rx_check_ack(tcb, ack);
779	return rc;
780}
781
782static inline int
783restore_syn_opt(union seg_info *si, union tsopt *to,
784	const union pkt_info *pi, uint32_t ts, const struct rte_mbuf *mb,
785	uint32_t hash_alg, rte_xmm_t *secret_key)
786{
787	int32_t rc;
788	uint32_t len;
789	const struct tcp_hdr *th;
790
791	/* check that ACK, etc fields are what we expected. */
792	rc = sync_check_ack(pi, si->seq, si->ack - 1, ts,
793				hash_alg,
794				secret_key);
795	if (rc < 0)
796		return rc;
797
798	si->mss = rc;
799
800	th = rte_pktmbuf_mtod_offset(mb, const struct tcp_hdr *,
801		mb->l2_len + mb->l3_len);
802	len = mb->l4_len - sizeof(*th);
803	to[0] = get_tms_opts((uintptr_t)(th + 1), len);
804	return 0;
805}
806
807static inline void
808stream_term(struct tle_tcp_stream *s)
809{
810	struct sdr *dr;
811
812	s->tcb.state = TCP_ST_CLOSED;
813	rte_smp_wmb();
814
815	timer_stop(s);
816
817	/* close() was already invoked, schedule final cleanup */
818	if ((s->tcb.uop & TCP_OP_CLOSE) != 0) {
819
820		dr = CTX_TCP_SDR(s->s.ctx);
821		STAILQ_INSERT_TAIL(&dr->be, &s->s, link);
822
823	/* notify user that stream need to be closed */
824	} else if (s->err.ev != NULL)
825		tle_event_raise(s->err.ev);
826	else if (s->err.cb.func != NULL)
827		s->err.cb.func(s->err.cb.data, &s->s);
828}
829
830static inline int
831stream_fill_dest(struct tle_tcp_stream *s)
832{
833	int32_t rc;
834	uint32_t type;
835	const void *da;
836
837        type = s->s.type;
838	if (type == TLE_V4)
839		da = &s->s.ipv4.addr.src;
840	else
841		da = &s->s.ipv6.addr.src;
842
843	rc = stream_get_dest(&s->s, da, &s->tx.dst);
844	return (rc < 0) ? rc : 0;
845}
846
847/*
848 * helper function, prepares a new accept stream.
849 */
850static inline int
851accept_prep_stream(struct tle_tcp_stream *ps, struct stbl *st,
852	struct tle_tcp_stream *cs, const union tsopt *to,
853	uint32_t tms, const union pkt_info *pi, const union seg_info *si)
854{
855	int32_t rc;
856	uint32_t rtt;
857
858	/* some TX still pending for that stream. */
859	if (TCP_STREAM_TX_PENDING(cs))
860		return -EAGAIN;
861
862	/* setup L4 ports and L3 addresses fields. */
863	cs->s.port.raw = pi->port.raw;
864	cs->s.pmsk.raw = UINT32_MAX;
865
866	if (pi->tf.type == TLE_V4) {
867		cs->s.ipv4.addr = pi->addr4;
868		cs->s.ipv4.mask.src = INADDR_NONE;
869		cs->s.ipv4.mask.dst = INADDR_NONE;
870	} else if (pi->tf.type == TLE_V6) {
871		cs->s.ipv6.addr = *pi->addr6;
872		rte_memcpy(&cs->s.ipv6.mask.src, &tle_ipv6_none,
873			sizeof(cs->s.ipv6.mask.src));
874		rte_memcpy(&cs->s.ipv6.mask.dst, &tle_ipv6_none,
875			sizeof(cs->s.ipv6.mask.dst));
876	}
877
878	/* setup TCB */
879	sync_fill_tcb(&cs->tcb, si, to);
880	cs->tcb.rcv.wnd = calc_rx_wnd(cs, cs->tcb.rcv.wscale);
881
882	/*
883	 * estimate the rto
884	 * for now rtt is calculated based on the tcp TMS option,
885	 * later add real-time one
886	 */
887	if (cs->tcb.so.ts.ecr) {
888		rtt = tms - cs->tcb.so.ts.ecr;
889		rto_estimate(&cs->tcb, rtt);
890	} else
891		cs->tcb.snd.rto = TCP_RTO_DEFAULT;
892
893	/* copy streams type & flags. */
894	cs->s.type = ps->s.type;
895	cs->flags = ps->flags;
896
897	/* retrive and cache destination information. */
898	rc = stream_fill_dest(cs);
899	if (rc != 0)
900		return rc;
901
902	/* update snd.mss with SMSS value */
903	cs->tcb.snd.mss = calc_smss(cs->tcb.snd.mss, &cs->tx.dst);
904
905	/* setup congestion variables */
906	cs->tcb.snd.cwnd = initial_cwnd(cs->tcb.snd.mss, ps->tcb.snd.cwnd);
907	cs->tcb.snd.ssthresh = cs->tcb.snd.wnd;
908	cs->tcb.snd.rto_tw = ps->tcb.snd.rto_tw;
909
910	cs->tcb.state = TCP_ST_ESTABLISHED;
911
912	/* add stream to the table */
913	cs->ste = stbl_add_stream(st, pi, cs);
914	if (cs->ste == NULL)
915		return -ENOBUFS;
916
917	cs->tcb.uop |= TCP_OP_ACCEPT;
918	tcp_stream_up(cs);
919	return 0;
920}
921
922
923/*
924 * ACK for new connection request arrived.
925 * Check that the packet meets all conditions and try to open a new stream.
926 * returns:
927 * < 0  - invalid packet
928 * == 0 - packet is valid and new stream was opened for it.
929 * > 0  - packet is valid, but failed to open new stream.
930 */
931static inline int
932rx_ack_listen(struct tle_tcp_stream *s, struct stbl *st,
933	const union pkt_info *pi, union seg_info *si,
934	uint32_t tms, struct rte_mbuf *mb, struct tle_tcp_stream **csp)
935{
936	int32_t rc;
937	struct tle_ctx *ctx;
938	struct tle_stream *ts;
939	struct tle_tcp_stream *cs;
940	union tsopt to;
941
942	*csp = NULL;
943
944	if (pi->tf.flags != TCP_FLAG_ACK || rx_check_stream(s, pi) != 0)
945		return -EINVAL;
946
947	ctx = s->s.ctx;
948	rc = restore_syn_opt(si, &to, pi, tms, mb, ctx->prm.hash_alg,
949				&ctx->prm.secret_key);
950	if (rc < 0)
951		return rc;
952
953	/* allocate new stream */
954	cs = tcp_stream_get(ctx, 0);
955	if (cs == NULL)
956		return ENFILE;
957
958	/* prepare stream to handle new connection */
959	if (accept_prep_stream(s, st, cs, &to, tms, pi, si) == 0) {
960
961		/* put new stream in the accept queue */
962		ts = &cs->s;
963		if (_rte_ring_enqueue_burst(s->rx.q,
964				(void * const *)&ts, 1) == 1) {
965			*csp = cs;
966			return 0;
967		}
968
969		/* cleanup on failure */
970		tcp_stream_down(cs);
971		stbl_del_stream(st, cs->ste, cs, 0);
972		cs->ste = NULL;
973	}
974
975	tcp_stream_reset(ctx, cs);
976	return ENOBUFS;
977}
978
979static inline int
980data_pkt_adjust(const struct tcb *tcb, struct rte_mbuf **mb, uint32_t hlen,
981	uint32_t *seqn, uint32_t *plen)
982{
983	uint32_t len, n, seq;
984
985	seq = *seqn;
986	len = *plen;
987
988	rte_pktmbuf_adj(*mb, hlen);
989	if (len == 0)
990		return -ENODATA;
991	/* cut off the start of the packet */
992	else if (tcp_seq_lt(seq, tcb->rcv.nxt)) {
993		n = tcb->rcv.nxt - seq;
994		if (n >= len)
995			return -ENODATA;
996
997		*mb = _rte_pktmbuf_adj(*mb, n);
998		*seqn = seq + n;
999		*plen = len - n;
1000	}
1001
1002	return 0;
1003}
1004
1005static inline uint32_t
1006rx_ackdata(struct tle_tcp_stream *s, uint32_t ack)
1007{
1008	uint32_t k, n;
1009
1010	n = ack - (uint32_t)s->tcb.snd.una;
1011
1012	/* some more data was acked. */
1013	if (n != 0) {
1014
1015		/* advance SND.UNA and free related packets. */
1016		k = rte_ring_free_count(s->tx.q);
1017		free_una_data(s, n);
1018
1019		/* mark the stream as available for writing */
1020		if (rte_ring_free_count(s->tx.q) != 0) {
1021			if (s->tx.ev != NULL)
1022				tle_event_raise(s->tx.ev);
1023			else if (k == 0 && s->tx.cb.func != NULL)
1024				s->tx.cb.func(s->tx.cb.data, &s->s);
1025		}
1026	}
1027
1028	return n;
1029}
1030
1031static void
1032stream_timewait(struct tle_tcp_stream *s, uint32_t rto)
1033{
1034	if (rto != 0) {
1035		s->tcb.state = TCP_ST_TIME_WAIT;
1036		s->tcb.snd.rto = rto;
1037		timer_reset(s);
1038	} else
1039		stream_term(s);
1040}
1041
1042static void
1043rx_fin_state(struct tle_tcp_stream *s, struct resp_info *rsp)
1044{
1045	uint32_t state;
1046	int32_t ackfin;
1047
1048	s->tcb.rcv.nxt += 1;
1049
1050	ackfin = (s->tcb.snd.una == s->tcb.snd.fss);
1051	state = s->tcb.state;
1052
1053	if (state == TCP_ST_ESTABLISHED) {
1054		s->tcb.state = TCP_ST_CLOSE_WAIT;
1055		/* raise err.ev & err.cb */
1056		if (s->err.ev != NULL)
1057			tle_event_raise(s->err.ev);
1058		else if (s->err.cb.func != NULL)
1059			s->err.cb.func(s->err.cb.data, &s->s);
1060	} else if (state == TCP_ST_FIN_WAIT_1 || state == TCP_ST_CLOSING) {
1061		rsp->flags |= TCP_FLAG_ACK;
1062		if (ackfin != 0)
1063			stream_timewait(s, s->tcb.snd.rto_tw);
1064		else
1065			s->tcb.state = TCP_ST_CLOSING;
1066	} else if (state == TCP_ST_FIN_WAIT_2) {
1067		rsp->flags |= TCP_FLAG_ACK;
1068		stream_timewait(s, s->tcb.snd.rto_tw);
1069	} else if (state == TCP_ST_LAST_ACK && ackfin != 0) {
1070		stream_term(s);
1071	}
1072}
1073
1074/*
1075 * FIN process for ESTABLISHED state
1076 * returns:
1077 * 0 < - error occurred
1078 * 0 - FIN was processed OK, and mbuf can be free/reused.
1079 * 0 > - FIN was processed OK and mbuf can't be free/reused.
1080 */
1081static inline int
1082rx_fin(struct tle_tcp_stream *s, uint32_t state,
1083	const union seg_info *si, struct rte_mbuf *mb,
1084	struct resp_info *rsp)
1085{
1086	uint32_t hlen, plen, seq;
1087	int32_t ret;
1088	union tsopt ts;
1089
1090	hlen = PKT_L234_HLEN(mb);
1091	plen = mb->pkt_len - hlen;
1092	seq = si->seq;
1093
1094	ts = rx_tms_opt(&s->tcb, mb);
1095	ret = rx_check_seqack(&s->tcb, seq, si->ack, plen, ts);
1096	if (ret != 0)
1097		return ret;
1098
1099	if (state < TCP_ST_ESTABLISHED)
1100		return -EINVAL;
1101
1102	if (plen != 0) {
1103
1104		ret = data_pkt_adjust(&s->tcb, &mb, hlen, &seq, &plen);
1105		if (ret != 0)
1106			return ret;
1107		if (rx_data_enqueue(s, seq, plen, &mb, 1) != 1)
1108			return -ENOBUFS;
1109	}
1110
1111	/*
1112	 * fast-path: all data & FIN was already sent out
1113	 * and now is acknowledged.
1114	 */
1115	if (s->tcb.snd.fss == s->tcb.snd.nxt &&
1116			si->ack == (uint32_t)s->tcb.snd.nxt) {
1117		s->tcb.snd.una = s->tcb.snd.fss;
1118		empty_tq(s);
1119	/* conventional ACK processiing */
1120	} else
1121		rx_ackdata(s, si->ack);
1122
1123	/* some fragments still missing */
1124	if (seq + plen != s->tcb.rcv.nxt) {
1125		s->tcb.rcv.frs.seq = seq + plen;
1126		s->tcb.rcv.frs.on = 1;
1127	} else
1128		rx_fin_state(s, rsp);
1129
1130	return plen;
1131}
1132
1133static inline int
1134rx_rst(struct tle_tcp_stream *s, uint32_t state, uint32_t flags,
1135	const union seg_info *si)
1136{
1137	int32_t rc;
1138
1139	/*
1140	 * RFC 793: In all states except SYN-SENT, all reset (RST) segments
1141	 * are validated by checking their SEQ-fields.
1142	 * A reset is valid if its sequence number is in the window.
1143	 * In the SYN-SENT state (a RST received in response to an initial SYN),
1144	 * the RST is acceptable if the ACK field acknowledges the SYN.
1145	 */
1146	if (state == TCP_ST_SYN_SENT) {
1147		rc = ((flags & TCP_FLAG_ACK) == 0 ||
1148				si->ack != s->tcb.snd.nxt) ?
1149			-ERANGE : 0;
1150	}
1151
1152	else
1153		rc = check_seqn(&s->tcb, si->seq, 0);
1154
1155	if (rc == 0)
1156		stream_term(s);
1157
1158	return rc;
1159}
1160
1161/*
1162 *  check do we have FIN  that was received out-of-order.
1163 *  if yes, try to process it now.
1164 */
1165static inline void
1166rx_ofo_fin(struct tle_tcp_stream *s, struct resp_info *rsp)
1167{
1168	if (s->tcb.rcv.frs.on != 0 && s->tcb.rcv.nxt == s->tcb.rcv.frs.seq)
1169		rx_fin_state(s, rsp);
1170}
1171
1172static inline void
1173dack_info_init(struct dack_info *tack, const struct tcb *tcb)
1174{
1175	static const struct dack_info zero_dack;
1176
1177	tack[0] = zero_dack;
1178	tack->ack = tcb->snd.una;
1179	tack->segs.dup = tcb->rcv.dupack;
1180	tack->wu.raw = tcb->snd.wu.raw;
1181	tack->wnd = tcb->snd.wnd >> tcb->snd.wscale;
1182}
1183
1184static inline void
1185ack_window_update(struct tcb *tcb, const struct dack_info *tack)
1186{
1187	tcb->snd.wu.raw = tack->wu.raw;
1188	tcb->snd.wnd = tack->wnd << tcb->snd.wscale;
1189}
1190
1191static inline void
1192ack_cwnd_update(struct tcb *tcb, uint32_t acked, const struct dack_info *tack)
1193{
1194	uint32_t n;
1195
1196	n = tack->segs.ack * tcb->snd.mss;
1197
1198	/* slow start phase, RFC 5681 3.1 (2)  */
1199	if (tcb->snd.cwnd < tcb->snd.ssthresh)
1200		tcb->snd.cwnd += RTE_MIN(acked, n);
1201	/* congestion avoidance phase, RFC 5681 3.1 (3) */
1202	else
1203		tcb->snd.cwnd += RTE_MAX(1U, n * tcb->snd.mss / tcb->snd.cwnd);
1204}
1205
1206static inline void
1207rto_ssthresh_update(struct tcb *tcb)
1208{
1209	uint32_t k, n;
1210
1211	/* RFC 5681 3.1 (4)  */
1212	n = (tcb->snd.nxt - tcb->snd.una) / 2;
1213	k = 2 * tcb->snd.mss;
1214	tcb->snd.ssthresh = RTE_MAX(n, k);
1215}
1216
1217static inline void
1218rto_cwnd_update(struct tcb *tcb)
1219{
1220
1221	if (tcb->snd.nb_retx == 0)
1222		rto_ssthresh_update(tcb);
1223
1224	/*
1225	 * RFC 5681 3.1: upon a timeout cwnd MUST be set to
1226	 * no more than 1 full-sized segment.
1227	 */
1228	tcb->snd.cwnd = tcb->snd.mss;
1229}
1230
1231static inline void
1232ack_info_update(struct dack_info *tack, const union seg_info *si,
1233	int32_t badseq, uint32_t dlen, const union tsopt ts)
1234{
1235	if (badseq != 0) {
1236		tack->segs.badseq++;
1237		return;
1238	}
1239
1240	/* segnt with incoming data */
1241	tack->segs.data += (dlen != 0);
1242
1243	/* segment with newly acked data */
1244	if (tcp_seq_lt(tack->ack, si->ack)) {
1245		tack->segs.dup = 0;
1246		tack->segs.ack++;
1247		tack->ack = si->ack;
1248		tack->ts = ts;
1249
1250	/*
1251	 * RFC 5681: An acknowledgment is considered a "duplicate" when:
1252	 * (a) the receiver of the ACK has outstanding data
1253	 * (b) the incoming acknowledgment carries no data
1254	 * (c) the SYN and FIN bits are both off
1255	 * (d) the acknowledgment number is equal to the TCP.UNA
1256	 * (e) the advertised window in the incoming acknowledgment equals the
1257	 * advertised window in the last incoming acknowledgment.
1258	 *
1259	 * Here will have only to check only for (b),(d),(e).
1260	 * (a) will be checked later for the whole bulk of packets,
1261	 * (c) should never happen here.
1262	 */
1263	} else if (dlen == 0 && si->wnd == tack->wnd && ++tack->segs.dup == 3) {
1264		tack->dup3.seg = tack->segs.ack + 1;
1265		tack->dup3.ack = tack->ack;
1266	}
1267
1268	/*
1269	 * RFC 793:
1270	 * If SND.UNA < SEG.ACK =< SND.NXT, the send window should be
1271	 * updated.  If (SND.WL1 < SEG.SEQ or (SND.WL1 = SEG.SEQ and
1272	 * SND.WL2 =< SEG.ACK)), set SND.WND <- SEG.WND, set
1273	 * SND.WL1 <- SEG.SEQ, and set SND.WL2 <- SEG.ACK.
1274	 */
1275	if (tcp_seq_lt(tack->wu.wl1, si->seq) ||
1276			(si->seq == tack->wu.wl1 &&
1277			tcp_seq_leq(tack->wu.wl2, si->ack))) {
1278
1279		tack->wu.wl1 = si->seq;
1280		tack->wu.wl2 = si->ack;
1281		tack->wnd = si->wnd;
1282	}
1283}
1284
1285static inline uint32_t
1286rx_data_ack(struct tle_tcp_stream *s, struct dack_info *tack,
1287	const union seg_info si[], struct rte_mbuf *mb[], struct rte_mbuf *rp[],
1288	int32_t rc[], uint32_t num)
1289{
1290	uint32_t i, j, k, n, t;
1291	uint32_t hlen, plen, seq, tlen;
1292	int32_t ret;
1293	union tsopt ts;
1294
1295	k = 0;
1296	for (i = 0; i != num; i = j) {
1297
1298		hlen = PKT_L234_HLEN(mb[i]);
1299		plen = mb[i]->pkt_len - hlen;
1300		seq = si[i].seq;
1301
1302		ts = rx_tms_opt(&s->tcb, mb[i]);
1303		ret = rx_check_seqack(&s->tcb, seq, si[i].ack, plen, ts);
1304
1305		/* account segment received */
1306		ack_info_update(tack, &si[i], ret != 0, plen, ts);
1307
1308		if (ret == 0) {
1309			/* skip duplicate data, if any */
1310			ret = data_pkt_adjust(&s->tcb, &mb[i], hlen,
1311				&seq, &plen);
1312		}
1313
1314		j = i + 1;
1315		if (ret != 0) {
1316			rp[k] = mb[i];
1317			rc[k] = -ret;
1318			k++;
1319			continue;
1320		}
1321
1322		/* group sequential packets together. */
1323		for (tlen = plen; j != num; tlen += plen, j++) {
1324
1325			hlen = PKT_L234_HLEN(mb[j]);
1326			plen = mb[j]->pkt_len - hlen;
1327
1328			/* not consecutive packet */
1329			if (plen == 0 || seq + tlen != si[j].seq)
1330				break;
1331
1332			/* check SEQ/ACK */
1333			ts = rx_tms_opt(&s->tcb, mb[j]);
1334			ret = rx_check_seqack(&s->tcb, si[j].seq, si[j].ack,
1335				plen, ts);
1336
1337			if (ret != 0)
1338				break;
1339
1340			/* account for segment received */
1341			ack_info_update(tack, &si[j], ret != 0, plen, ts);
1342
1343			rte_pktmbuf_adj(mb[j], hlen);
1344		}
1345
1346		n = j - i;
1347
1348		/* account for OFO data */
1349		if (seq != s->tcb.rcv.nxt)
1350			tack->segs.ofo += n;
1351
1352		/* enqueue packets */
1353		t = rx_data_enqueue(s, seq, tlen, mb + i, n);
1354
1355		/* if we are out of space in stream recv buffer. */
1356		for (; t != n; t++) {
1357			rp[k] = mb[i + t];
1358			rc[k] = -ENOBUFS;
1359			k++;
1360		}
1361	}
1362
1363	return num - k;
1364}
1365
1366static inline void
1367start_fast_retransmit(struct tle_tcp_stream *s)
1368{
1369	struct tcb *tcb;
1370
1371	tcb = &s->tcb;
1372
1373	/* RFC 6582 3.2.2 */
1374	tcb->snd.rcvr = tcb->snd.nxt;
1375	tcb->snd.fastack = 1;
1376
1377	/* RFC 5681 3.2.2 */
1378	rto_ssthresh_update(tcb);
1379
1380	/* RFC 5681 3.2.3 */
1381	tcp_txq_rst_nxt_head(s);
1382	tcb->snd.nxt = tcb->snd.una;
1383	tcb->snd.cwnd = tcb->snd.ssthresh + 3 * tcb->snd.mss;
1384}
1385
1386static inline void
1387stop_fast_retransmit(struct tle_tcp_stream *s)
1388{
1389	struct tcb *tcb;
1390	uint32_t n;
1391
1392	tcb = &s->tcb;
1393	n = tcb->snd.nxt - tcb->snd.una;
1394	tcb->snd.cwnd = RTE_MIN(tcb->snd.ssthresh,
1395		RTE_MAX(n, tcb->snd.mss) + tcb->snd.mss);
1396	tcb->snd.fastack = 0;
1397}
1398
1399static inline int
1400in_fast_retransmit(struct tle_tcp_stream *s, uint32_t ack_len, uint32_t ack_num,
1401	uint32_t dup_num)
1402{
1403	uint32_t n;
1404	struct tcb *tcb;
1405
1406	tcb = &s->tcb;
1407
1408	/* RFC 5682 3.2.3 partial ACK */
1409	if (ack_len != 0) {
1410
1411		n = ack_num * tcb->snd.mss;
1412		if (ack_len >= n)
1413			tcb->snd.cwnd -= ack_len - n;
1414		else
1415			tcb->snd.cwnd -= ack_len % tcb->snd.mss;
1416
1417		/*
1418		 * For the first partial ACK that arrives
1419		 * during fast recovery, also reset the
1420		 * retransmit timer.
1421		 */
1422		if (tcb->snd.fastack == 1)
1423			timer_reset(s);
1424
1425		tcb->snd.fastack += ack_num;
1426		return 1;
1427
1428	/* RFC 5681 3.2.4 */
1429	} else if (dup_num > 3) {
1430		s->tcb.snd.cwnd += (dup_num - 3) * tcb->snd.mss;
1431		return 1;
1432	}
1433
1434	return 0;
1435}
1436
1437static inline int
1438process_ack(struct tle_tcp_stream *s, uint32_t acked,
1439	const struct dack_info *tack)
1440{
1441	int32_t send;
1442
1443	send = 0;
1444
1445	/* normal mode */
1446	if (s->tcb.snd.fastack == 0) {
1447
1448		send = 1;
1449
1450		/* RFC 6582 3.2.2 switch to fast retransmit mode */
1451		if (tack->dup3.seg != 0 && s->tcb.snd.una != s->tcb.snd.nxt &&
1452				s->tcb.snd.una >= s->tcb.snd.rcvr) {
1453
1454			start_fast_retransmit(s);
1455			in_fast_retransmit(s,
1456				tack->ack - tack->dup3.ack,
1457				tack->segs.ack - tack->dup3.seg - 1,
1458				tack->segs.dup);
1459
1460		/* remain in normal mode */
1461		} else if (acked != 0) {
1462			ack_cwnd_update(&s->tcb, acked, tack);
1463			timer_stop(s);
1464		}
1465
1466	/* fast retransmit mode */
1467	} else {
1468
1469		/* remain in fast retransmit mode */
1470		if (s->tcb.snd.una < s->tcb.snd.rcvr) {
1471
1472			send = in_fast_retransmit(s, acked, tack->segs.ack,
1473				tack->segs.dup);
1474		} else {
1475			/* RFC 5682 3.2.3 full ACK */
1476			stop_fast_retransmit(s);
1477			timer_stop(s);
1478
1479			/* if we have another series of dup ACKs */
1480			if (tack->dup3.seg != 0 &&
1481					s->tcb.snd.una != s->tcb.snd.nxt &&
1482					tcp_seq_leq((uint32_t)s->tcb.snd.rcvr,
1483					tack->dup3.ack)) {
1484
1485				/* restart fast retransmit again. */
1486				start_fast_retransmit(s);
1487				send = in_fast_retransmit(s,
1488					tack->ack - tack->dup3.ack,
1489					tack->segs.ack - tack->dup3.seg - 1,
1490					tack->segs.dup);
1491			}
1492		}
1493	}
1494
1495	return send;
1496}
1497
1498/*
1499 * our FIN was acked, stop rto timer, change stream state,
1500 * and possibly close the stream.
1501 */
1502static inline void
1503rx_ackfin(struct tle_tcp_stream *s)
1504{
1505	uint32_t state;
1506
1507	s->tcb.snd.una = s->tcb.snd.fss;
1508	empty_tq(s);
1509
1510	state = s->tcb.state;
1511	if (state == TCP_ST_LAST_ACK)
1512		stream_term(s);
1513	else if (state == TCP_ST_FIN_WAIT_1) {
1514		timer_stop(s);
1515		s->tcb.state = TCP_ST_FIN_WAIT_2;
1516	} else if (state == TCP_ST_CLOSING) {
1517		stream_timewait(s, s->tcb.snd.rto_tw);
1518	}
1519}
1520
1521static inline void
1522rx_process_ack(struct tle_tcp_stream *s, uint32_t ts,
1523	const struct dack_info *tack)
1524{
1525	int32_t send;
1526	uint32_t n;
1527
1528	s->tcb.rcv.dupack = tack->segs.dup;
1529
1530	n = rx_ackdata(s, tack->ack);
1531	send = process_ack(s, n, tack);
1532
1533	/* try to send more data. */
1534	if ((n != 0 || send != 0) && tcp_txq_nxt_cnt(s) != 0)
1535		txs_enqueue(s->s.ctx, s);
1536
1537	/* restart RTO timer. */
1538	if (s->tcb.snd.nxt != s->tcb.snd.una)
1539		timer_start(s);
1540
1541	/* update rto, if fresh packet is here then calculate rtt */
1542	if (tack->ts.ecr != 0)
1543		rto_estimate(&s->tcb, ts - tack->ts.ecr);
1544}
1545
1546/*
1547 * process <SYN,ACK>
1548 * returns negative value on failure, or zero on success.
1549 */
1550static inline int
1551rx_synack(struct tle_tcp_stream *s, uint32_t ts, uint32_t state,
1552	const union seg_info *si, struct rte_mbuf *mb,
1553	struct resp_info *rsp)
1554{
1555	struct syn_opts so;
1556	struct tcp_hdr *th;
1557
1558	if (state != TCP_ST_SYN_SENT)
1559		return -EINVAL;
1560
1561	/*
1562	 * RFC 793 3.9: in the SYN-SENT state
1563	 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send a reset
1564	 * <SEQ=SEG.ACK><CTL=RST>
1565	 * and discard the segment.
1566	 * The connection remains in the same state.
1567	 */
1568	if (si->ack != (uint32_t)s->tcb.snd.nxt) {
1569		send_rst(s, si->ack);
1570		return 0;
1571	}
1572
1573	th = rte_pktmbuf_mtod_offset(mb, struct tcp_hdr *,
1574		mb->l2_len + mb->l3_len);
1575	get_syn_opts(&so, (uintptr_t)(th + 1), mb->l4_len - sizeof(*th));
1576
1577	s->tcb.so = so;
1578
1579	s->tcb.snd.una = s->tcb.snd.nxt;
1580	s->tcb.snd.mss = calc_smss(so.mss, &s->tx.dst);
1581	s->tcb.snd.wnd = si->wnd << so.wscale;
1582	s->tcb.snd.wu.wl1 = si->seq;
1583	s->tcb.snd.wu.wl2 = si->ack;
1584	s->tcb.snd.wscale = so.wscale;
1585
1586	/* setup congestion variables */
1587	s->tcb.snd.cwnd = initial_cwnd(s->tcb.snd.mss, s->tcb.snd.cwnd);
1588	s->tcb.snd.ssthresh = s->tcb.snd.wnd;
1589
1590	s->tcb.rcv.ts = so.ts.val;
1591	s->tcb.rcv.irs = si->seq;
1592	s->tcb.rcv.nxt = si->seq + 1;
1593
1594	/* if peer doesn't support WSCALE opt, recalculate RCV.WND */
1595	s->tcb.rcv.wscale = (so.wscale == TCP_WSCALE_NONE) ?
1596		TCP_WSCALE_NONE : TCP_WSCALE_DEFAULT;
1597	s->tcb.rcv.wnd = calc_rx_wnd(s, s->tcb.rcv.wscale);
1598
1599	/* calculate initial rto */
1600	rto_estimate(&s->tcb, ts - s->tcb.snd.ts);
1601
1602	rsp->flags |= TCP_FLAG_ACK;
1603
1604	timer_stop(s);
1605	s->tcb.state = TCP_ST_ESTABLISHED;
1606	rte_smp_wmb();
1607
1608	if (s->tx.ev != NULL)
1609		tle_event_raise(s->tx.ev);
1610	else if (s->tx.cb.func != NULL)
1611		s->tx.cb.func(s->tx.cb.data, &s->s);
1612
1613	return 0;
1614}
1615
1616static inline uint32_t
1617rx_stream(struct tle_tcp_stream *s, uint32_t ts,
1618	const union pkt_info *pi, const union seg_info si[],
1619	struct rte_mbuf *mb[], struct rte_mbuf *rp[], int32_t rc[],
1620	uint32_t num)
1621{
1622	uint32_t i, k, n, state;
1623	int32_t ret;
1624	struct resp_info rsp;
1625	struct dack_info tack;
1626
1627	k = 0;
1628	rsp.flags = 0;
1629
1630	state = s->tcb.state;
1631
1632	/*
1633	 * first check for the states/flags where we don't
1634	 * expect groups of packets.
1635	 */
1636
1637	/* process RST */
1638	if ((pi->tf.flags & TCP_FLAG_RST) != 0) {
1639		for (i = 0;
1640				i != num &&
1641				rx_rst(s, state, pi->tf.flags, &si[i]);
1642				i++)
1643			;
1644		i = 0;
1645
1646	/* RFC 793: if the ACK bit is off drop the segment and return */
1647	} else if ((pi->tf.flags & TCP_FLAG_ACK) == 0) {
1648		i = 0;
1649	/*
1650	 * first check for the states/flags where we don't
1651	 * expect groups of packets.
1652	 */
1653
1654	/* process <SYN,ACK> */
1655	} else if ((pi->tf.flags & TCP_FLAG_SYN) != 0) {
1656		for (i = 0; i != num; i++) {
1657			ret = rx_synack(s, ts, state, &si[i], mb[i], &rsp);
1658			if (ret == 0)
1659				break;
1660
1661			rc[k] = -ret;
1662			rp[k] = mb[i];
1663			k++;
1664		}
1665
1666	/* process FIN */
1667	} else if ((pi->tf.flags & TCP_FLAG_FIN) != 0) {
1668		ret = 0;
1669		for (i = 0; i != num; i++) {
1670			ret = rx_fin(s, state, &si[i], mb[i], &rsp);
1671			if (ret >= 0)
1672				break;
1673
1674			rc[k] = -ret;
1675			rp[k] = mb[i];
1676			k++;
1677		}
1678		i += (ret > 0);
1679
1680	/* normal data/ack packets */
1681	} else if (state >= TCP_ST_ESTABLISHED && state <= TCP_ST_LAST_ACK) {
1682
1683		/* process incoming data packets. */
1684		dack_info_init(&tack, &s->tcb);
1685		n = rx_data_ack(s, &tack, si, mb, rp, rc, num);
1686
1687		/* follow up actions based on aggregated information */
1688
1689		/* update SND.WND */
1690		ack_window_update(&s->tcb, &tack);
1691
1692		/*
1693		 * fast-path: all data & FIN was already sent out
1694		 * and now is acknowledged.
1695		 */
1696		if (s->tcb.snd.fss == s->tcb.snd.nxt &&
1697				tack.ack == (uint32_t)s->tcb.snd.nxt)
1698			rx_ackfin(s);
1699		else
1700			rx_process_ack(s, ts, &tack);
1701
1702		/*
1703		 * send an immediate ACK if either:
1704		 * - received segment with invalid seq/ack number
1705		 * - received segment with OFO data
1706		 * - received segment with INO data and no TX is scheduled
1707		 *   for that stream.
1708		 */
1709		if (tack.segs.badseq != 0 || tack.segs.ofo != 0 ||
1710				(tack.segs.data != 0 &&
1711				rte_atomic32_read(&s->tx.arm) == 0))
1712			rsp.flags |= TCP_FLAG_ACK;
1713
1714		rx_ofo_fin(s, &rsp);
1715
1716		k += num - n;
1717		i = num;
1718
1719	/* unhandled state, drop all packets. */
1720	} else
1721		i = 0;
1722
1723	/* we have a response packet to send. */
1724	if (rsp.flags != 0) {
1725		send_ack(s, ts, rsp.flags);
1726
1727		/* start the timer for FIN packet */
1728		if ((rsp.flags & TCP_FLAG_FIN) != 0)
1729			timer_reset(s);
1730	}
1731
1732	/* unprocessed packets */
1733	for (; i != num; i++, k++) {
1734		rc[k] = ENODATA;
1735		rp[k] = mb[i];
1736	}
1737
1738	return num - k;
1739}
1740
1741static inline uint32_t
1742rx_new_stream(struct tle_tcp_stream *s, uint32_t ts,
1743	const union pkt_info *pi, const union seg_info si[],
1744	struct rte_mbuf *mb[], struct rte_mbuf *rp[], int32_t rc[],
1745	uint32_t num)
1746{
1747	uint32_t i;
1748
1749	if (tcp_stream_acquire(s) > 0) {
1750		i = rx_stream(s, ts, pi, si, mb, rp, rc, num);
1751		tcp_stream_release(s);
1752		return i;
1753	}
1754
1755	for (i = 0; i != num; i++) {
1756		rc[i] = ENOENT;
1757		rp[i] = mb[i];
1758	}
1759	return 0;
1760}
1761
1762static inline uint32_t
1763rx_postsyn(struct tle_dev *dev, struct stbl *st, uint32_t type, uint32_t ts,
1764	const union pkt_info pi[], union seg_info si[],
1765	struct rte_mbuf *mb[], struct rte_mbuf *rp[], int32_t rc[],
1766	uint32_t num)
1767{
1768	struct tle_tcp_stream *cs, *s;
1769	uint32_t i, k, n, state;
1770	int32_t ret;
1771
1772	s = rx_obtain_stream(dev, st, &pi[0], type);
1773	if (s == NULL) {
1774		for (i = 0; i != num; i++) {
1775			rc[i] = ENOENT;
1776			rp[i] = mb[i];
1777		}
1778		return 0;
1779	}
1780
1781	k = 0;
1782	state = s->tcb.state;
1783
1784	if (state == TCP_ST_LISTEN) {
1785
1786		/* one connection per flow */
1787		cs = NULL;
1788		ret = -EINVAL;
1789		for (i = 0; i != num; i++) {
1790
1791			ret = rx_ack_listen(s, st, pi, &si[i], ts, mb[i], &cs);
1792
1793			/* valid packet encountered */
1794			if (ret >= 0)
1795				break;
1796
1797			/* invalid packet, keep trying to find a proper one */
1798			rc[k] = -ret;
1799			rp[k] = mb[i];
1800			k++;
1801		}
1802
1803		/* packet is valid, but we are out of streams to serve it */
1804		if (ret > 0) {
1805			for (; i != num; i++, k++) {
1806				rc[k] = ret;
1807				rp[k] = mb[i];
1808			}
1809		/* new stream is accepted */
1810		} else if (ret == 0) {
1811
1812			/* inform listen stream about new connections */
1813			if (s->rx.ev != NULL)
1814				tle_event_raise(s->rx.ev);
1815			else if (s->rx.cb.func != NULL &&
1816					rte_ring_count(s->rx.q) == 1)
1817				s->rx.cb.func(s->rx.cb.data, &s->s);
1818
1819			/* if there is no data, drop current packet */
1820			if (PKT_L4_PLEN(mb[i]) == 0) {
1821				rc[k] = ENODATA;
1822				rp[k++] = mb[i++];
1823			}
1824
1825			/*  process remaining packets for that stream */
1826			if (num != i) {
1827				n = rx_new_stream(cs, ts, pi + i, si + i,
1828					mb + i, rp + k, rc + k, num - i);
1829				k += num - n - i;
1830			}
1831		}
1832
1833	} else {
1834		i = rx_stream(s, ts, pi, si, mb, rp, rc, num);
1835		k = num - i;
1836	}
1837
1838	tcp_stream_release(s);
1839	return num - k;
1840}
1841
1842
1843static inline uint32_t
1844rx_syn(struct tle_dev *dev, uint32_t type, uint32_t ts,
1845	const union pkt_info pi[], const union seg_info si[],
1846	struct rte_mbuf *mb[], struct rte_mbuf *rp[], int32_t rc[],
1847	uint32_t num)
1848{
1849	struct tle_tcp_stream *s;
1850	uint32_t i, k;
1851	int32_t ret;
1852
1853	s = rx_obtain_listen_stream(dev, &pi[0], type);
1854	if (s == NULL) {
1855		for (i = 0; i != num; i++) {
1856			rc[i] = ENOENT;
1857			rp[i] = mb[i];
1858		}
1859		return 0;
1860	}
1861
1862	k = 0;
1863	for (i = 0; i != num; i++) {
1864
1865		/* check that this remote is allowed to connect */
1866		if (rx_check_stream(s, &pi[i]) != 0)
1867			ret = -ENOENT;
1868		else
1869			/* syncokie: reply with <SYN,ACK> */
1870			ret = sync_ack(s, &pi[i], &si[i], ts, mb[i]);
1871
1872		if (ret != 0) {
1873			rc[k] = -ret;
1874			rp[k] = mb[i];
1875			k++;
1876		}
1877	}
1878
1879	tcp_stream_release(s);
1880	return num - k;
1881}
1882
1883uint16_t
1884tle_tcp_rx_bulk(struct tle_dev *dev, struct rte_mbuf *pkt[],
1885	struct rte_mbuf *rp[], int32_t rc[], uint16_t num)
1886{
1887	struct stbl *st;
1888	struct tle_ctx *ctx;
1889	uint32_t i, j, k, mt, n, t, ts;
1890	union pkt_info pi[num];
1891	union seg_info si[num];
1892	union {
1893		uint8_t t[TLE_VNUM];
1894		uint32_t raw;
1895	} stu;
1896
1897	ctx = dev->ctx;
1898	ts = tcp_get_tms(ctx->cycles_ms_shift);
1899	st = CTX_TCP_STLB(ctx);
1900	mt = ((ctx->prm.flags & TLE_CTX_FLAG_ST) == 0);
1901
1902	stu.raw = 0;
1903
1904	/* extract packet info and check the L3/L4 csums */
1905	for (i = 0; i != num; i++) {
1906
1907		get_pkt_info(pkt[i], &pi[i], &si[i]);
1908
1909		t = pi[i].tf.type;
1910		pi[i].csf = check_pkt_csum(pkt[i], pi[i].csf, t, IPPROTO_TCP);
1911		stu.t[t] = mt;
1912	}
1913
1914	if (stu.t[TLE_V4] != 0)
1915		stbl_lock(st, TLE_V4);
1916	if (stu.t[TLE_V6] != 0)
1917		stbl_lock(st, TLE_V6);
1918
1919	k = 0;
1920	for (i = 0; i != num; i += j) {
1921
1922		t = pi[i].tf.type;
1923
1924		/*basic checks for incoming packet */
1925		if (t >= TLE_VNUM || pi[i].csf != 0 || dev->dp[t] == NULL) {
1926			rc[k] = EINVAL;
1927			rp[k] = pkt[i];
1928			j = 1;
1929			k++;
1930		/* process input SYN packets */
1931		} else if (pi[i].tf.flags == TCP_FLAG_SYN) {
1932			j = pkt_info_bulk_syneq(pi + i, num - i);
1933			n = rx_syn(dev, t, ts, pi + i, si + i, pkt + i,
1934				rp + k, rc + k, j);
1935			k += j - n;
1936		} else {
1937			j = pkt_info_bulk_eq(pi + i, num - i);
1938			n = rx_postsyn(dev, st, t, ts, pi + i, si + i, pkt + i,
1939				rp + k, rc + k, j);
1940			k += j - n;
1941		}
1942	}
1943
1944	if (stu.t[TLE_V4] != 0)
1945		stbl_unlock(st, TLE_V4);
1946	if (stu.t[TLE_V6] != 0)
1947		stbl_unlock(st, TLE_V6);
1948
1949	return num - k;
1950}
1951
1952uint16_t
1953tle_tcp_stream_accept(struct tle_stream *ts, struct tle_stream *rs[],
1954	uint32_t num)
1955{
1956	uint32_t n;
1957	struct tle_tcp_stream *s;
1958	struct tle_memtank *mts;
1959
1960	s = TCP_STREAM(ts);
1961	n = _rte_ring_dequeue_burst(s->rx.q, (void **)rs, num);
1962	if (n == 0)
1963		return 0;
1964
1965	mts = CTX_TCP_MTS(ts->ctx);
1966
1967	/*
1968	 * if we still have packets to read,
1969	 * then rearm stream RX event.
1970	 */
1971	if (n == num && rte_ring_count(s->rx.q) != 0) {
1972		if (tcp_stream_try_acquire(s) > 0 && s->rx.ev != NULL)
1973			tle_event_raise(s->rx.ev);
1974		tcp_stream_release(s);
1975	}
1976
1977	tle_memtank_grow(mts);
1978	return n;
1979}
1980
1981uint16_t
1982tle_tcp_tx_bulk(struct tle_dev *dev, struct rte_mbuf *pkt[], uint16_t num)
1983{
1984	uint32_t i, j, k, n;
1985	struct tle_drb *drb[num];
1986	struct tle_tcp_stream *s;
1987
1988	/* extract packets from device TX queue. */
1989
1990	k = num;
1991	n = tle_dring_sc_dequeue(&dev->tx.dr, (const void **)(uintptr_t)pkt,
1992		num, drb, &k);
1993
1994	if (n == 0)
1995		return 0;
1996
1997	/* free empty drbs and notify related streams. */
1998
1999	for (i = 0; i != k; i = j) {
2000		s = drb[i]->udata;
2001		for (j = i + 1; j != k && s == drb[j]->udata; j++)
2002			;
2003		stream_drb_free(s, drb + i, j - i);
2004	}
2005
2006	return n;
2007}
2008
2009static inline void
2010stream_fill_pkt_info(const struct tle_tcp_stream *s, union pkt_info *pi)
2011{
2012	if (s->s.type == TLE_V4)
2013		pi->addr4 = s->s.ipv4.addr;
2014	else
2015		pi->addr6 = &s->s.ipv6.addr;
2016
2017	pi->port = s->s.port;
2018	pi->tf.type = s->s.type;
2019}
2020
2021static int
2022stream_fill_addr(struct tle_tcp_stream *s, const struct sockaddr *addr)
2023{
2024	const struct sockaddr_in *in4;
2025	const struct sockaddr_in6 *in6;
2026	const struct tle_dev_param *prm;
2027	int32_t rc;
2028
2029	rc = 0;
2030	s->s.pmsk.raw = UINT32_MAX;
2031
2032	/* setup L4 src ports and src address fields. */
2033	if (s->s.type == TLE_V4) {
2034		in4 = (const struct sockaddr_in *)addr;
2035		if (in4->sin_addr.s_addr == INADDR_ANY || in4->sin_port == 0)
2036			return -EINVAL;
2037
2038		s->s.port.src = in4->sin_port;
2039		s->s.ipv4.addr.src = in4->sin_addr.s_addr;
2040		s->s.ipv4.mask.src = INADDR_NONE;
2041		s->s.ipv4.mask.dst = INADDR_NONE;
2042
2043	} else if (s->s.type == TLE_V6) {
2044		in6 = (const struct sockaddr_in6 *)addr;
2045		if (memcmp(&in6->sin6_addr, &tle_ipv6_any,
2046				sizeof(tle_ipv6_any)) == 0 ||
2047				in6->sin6_port == 0)
2048			return -EINVAL;
2049
2050		s->s.port.src = in6->sin6_port;
2051		rte_memcpy(&s->s.ipv6.addr.src, &in6->sin6_addr,
2052			sizeof(s->s.ipv6.addr.src));
2053		rte_memcpy(&s->s.ipv6.mask.src, &tle_ipv6_none,
2054			sizeof(s->s.ipv6.mask.src));
2055		rte_memcpy(&s->s.ipv6.mask.dst, &tle_ipv6_none,
2056			sizeof(s->s.ipv6.mask.dst));
2057	}
2058
2059	/* setup the destination device. */
2060	rc = stream_fill_dest(s);
2061	if (rc != 0)
2062		return rc;
2063
2064	/* setup L4 dst address from device param */
2065	prm = &s->tx.dst.dev->prm;
2066	if (s->s.type == TLE_V4) {
2067		if (s->s.ipv4.addr.dst == INADDR_ANY)
2068			s->s.ipv4.addr.dst = prm->local_addr4.s_addr;
2069	} else if (memcmp(&s->s.ipv6.addr.dst, &tle_ipv6_any,
2070			sizeof(tle_ipv6_any)) == 0)
2071		memcpy(&s->s.ipv6.addr.dst, &prm->local_addr6,
2072			sizeof(s->s.ipv6.addr.dst));
2073
2074	return rc;
2075}
2076
2077static inline int
2078tx_syn(struct tle_tcp_stream *s, const struct sockaddr *addr)
2079{
2080	int32_t rc;
2081	uint32_t tms, seq;
2082	union pkt_info pi;
2083	struct stbl *st;
2084	struct stbl_entry *se;
2085
2086	/* fill stream address */
2087	rc = stream_fill_addr(s, addr);
2088	if (rc != 0)
2089		return rc;
2090
2091	/* fill pkt info to generate seq.*/
2092	stream_fill_pkt_info(s, &pi);
2093
2094	tms = tcp_get_tms(s->s.ctx->cycles_ms_shift);
2095	s->tcb.so.ts.val = tms;
2096	s->tcb.so.ts.ecr = 0;
2097	s->tcb.so.wscale = TCP_WSCALE_DEFAULT;
2098	s->tcb.so.mss = calc_smss(s->tx.dst.mtu, &s->tx.dst);
2099
2100	/* note that rcv.nxt is 0 here for sync_gen_seq.*/
2101	seq = sync_gen_seq(&pi, s->tcb.rcv.nxt, tms, s->tcb.so.mss,
2102				s->s.ctx->prm.hash_alg,
2103				&s->s.ctx->prm.secret_key);
2104	s->tcb.snd.iss = seq;
2105	s->tcb.snd.rcvr = seq;
2106	s->tcb.snd.una = seq;
2107	s->tcb.snd.nxt = seq + 1;
2108	s->tcb.snd.rto = TCP_RTO_DEFAULT;
2109	s->tcb.snd.ts = tms;
2110
2111	s->tcb.rcv.mss = s->tcb.so.mss;
2112	s->tcb.rcv.wscale = TCP_WSCALE_DEFAULT;
2113	s->tcb.rcv.wnd = calc_rx_wnd(s, s->tcb.rcv.wscale);
2114	s->tcb.rcv.ts = 0;
2115
2116	/* add the stream in stream table */
2117	st = CTX_TCP_STLB(s->s.ctx);
2118	se = stbl_add_stream_lock(st, s);
2119	if (se == NULL)
2120		return -ENOBUFS;
2121	s->ste = se;
2122
2123	/* put stream into the to-send queue */
2124	txs_enqueue(s->s.ctx, s);
2125
2126	return 0;
2127}
2128
2129int
2130tle_tcp_stream_connect(struct tle_stream *ts, const struct sockaddr *addr)
2131{
2132	struct tle_tcp_stream *s;
2133	uint32_t type;
2134	int32_t rc;
2135
2136	if (ts == NULL || addr == NULL)
2137		return -EINVAL;
2138
2139	s = TCP_STREAM(ts);
2140	type = s->s.type;
2141	if (type >= TLE_VNUM)
2142		return -EINVAL;
2143
2144	if (tcp_stream_try_acquire(s) > 0) {
2145		rc = rte_atomic16_cmpset(&s->tcb.state, TCP_ST_CLOSED,
2146			TCP_ST_SYN_SENT);
2147		rc = (rc == 0) ? -EDEADLK : 0;
2148	} else
2149		rc = -EINVAL;
2150
2151	if (rc != 0) {
2152		tcp_stream_release(s);
2153		return rc;
2154	}
2155
2156	/* fill stream, prepare and transmit syn pkt */
2157	s->tcb.uop |= TCP_OP_CONNECT;
2158	rc = tx_syn(s, addr);
2159	tcp_stream_release(s);
2160
2161	/* error happened, do a cleanup */
2162	if (rc != 0)
2163		tle_tcp_stream_close(ts);
2164
2165	return rc;
2166}
2167
2168uint16_t
2169tle_tcp_stream_recv(struct tle_stream *ts, struct rte_mbuf *pkt[], uint16_t num)
2170{
2171	uint32_t n;
2172	struct tle_tcp_stream *s;
2173
2174	s = TCP_STREAM(ts);
2175	n = _rte_ring_mcs_dequeue_burst(s->rx.q, (void **)pkt, num);
2176	if (n == 0)
2177		return 0;
2178
2179	/*
2180	 * if we still have packets to read,
2181	 * then rearm stream RX event.
2182	 */
2183	if (n == num && rte_ring_count(s->rx.q) != 0) {
2184		if (tcp_stream_try_acquire(s) > 0 && s->rx.ev != NULL)
2185			tle_event_raise(s->rx.ev);
2186		tcp_stream_release(s);
2187	}
2188
2189	return n;
2190}
2191
2192ssize_t
2193tle_tcp_stream_readv(struct tle_stream *ts, const struct iovec *iov,
2194	int iovcnt)
2195{
2196	int32_t i;
2197	uint32_t mn, n, tn;
2198	size_t sz;
2199	struct tle_tcp_stream *s;
2200	struct iovec iv;
2201	struct rxq_objs mo[2];
2202
2203	s = TCP_STREAM(ts);
2204
2205	/* get group of packets */
2206	mn = tcp_rxq_get_objs(s, mo);
2207	if (mn == 0)
2208		return 0;
2209
2210	sz = 0;
2211	n = 0;
2212	for (i = 0; i != iovcnt; i++) {
2213		iv = iov[i];
2214		sz += iv.iov_len;
2215		n += _mbus_to_iovec(&iv, mo[0].mb + n, mo[0].num - n);
2216		if (iv.iov_len != 0) {
2217			sz -= iv.iov_len;
2218			break;
2219		}
2220	}
2221
2222	tn = n;
2223
2224	if (i != iovcnt && mn != 1) {
2225		n = 0;
2226		do {
2227			sz += iv.iov_len;
2228			n += _mbus_to_iovec(&iv, mo[1].mb + n, mo[1].num - n);
2229			if (iv.iov_len != 0) {
2230				sz -= iv.iov_len;
2231				break;
2232			}
2233			if (i + 1 != iovcnt)
2234				iv = iov[i + 1];
2235		} while (++i != iovcnt);
2236		tn += n;
2237	}
2238
2239	tcp_rxq_consume(s, tn);
2240
2241	/*
2242	 * if we still have packets to read,
2243	 * then rearm stream RX event.
2244	 */
2245	if (i == iovcnt && rte_ring_count(s->rx.q) != 0) {
2246		if (tcp_stream_try_acquire(s) > 0 && s->rx.ev != NULL)
2247			tle_event_raise(s->rx.ev);
2248		tcp_stream_release(s);
2249	}
2250
2251	return sz;
2252}
2253
2254static inline int32_t
2255tx_segments(struct tle_tcp_stream *s, uint64_t ol_flags,
2256	struct rte_mbuf *segs[], uint32_t num)
2257{
2258	uint32_t i;
2259	int32_t rc;
2260
2261	for (i = 0; i != num; i++) {
2262		/* Build L2/L3/L4 header */
2263		rc = tcp_fill_mbuf(segs[i], s, &s->tx.dst, ol_flags, s->s.port,
2264			0, TCP_FLAG_ACK, 0, 0);
2265		if (rc != 0) {
2266			free_mbufs(segs, num);
2267			break;
2268		}
2269	}
2270
2271	if (i == num) {
2272		/* queue packets for further transmission. */
2273		rc = _rte_ring_enqueue_bulk(s->tx.q, (void **)segs, num);
2274		if (rc != 0)
2275			free_mbufs(segs, num);
2276	}
2277
2278	return rc;
2279}
2280
2281uint16_t
2282tle_tcp_stream_send(struct tle_stream *ts, struct rte_mbuf *pkt[], uint16_t num)
2283{
2284	uint32_t i, j, k, mss, n, state;
2285	int32_t rc;
2286	uint64_t ol_flags;
2287	struct tle_tcp_stream *s;
2288	struct rte_mbuf *segs[TCP_MAX_PKT_SEG];
2289
2290	s = TCP_STREAM(ts);
2291
2292	/* mark stream as not closable. */
2293	if (tcp_stream_acquire(s) < 0) {
2294		rte_errno = EAGAIN;
2295		return 0;
2296	}
2297
2298	state = s->tcb.state;
2299	if (state != TCP_ST_ESTABLISHED && state != TCP_ST_CLOSE_WAIT) {
2300		rte_errno = ENOTCONN;
2301		tcp_stream_release(s);
2302		return 0;
2303	}
2304
2305	mss = s->tcb.snd.mss;
2306	ol_flags = s->tx.dst.ol_flags;
2307
2308	k = 0;
2309	rc = 0;
2310	while (k != num) {
2311		/* prepare and check for TX */
2312		for (i = k; i != num; i++) {
2313			if (pkt[i]->pkt_len > mss ||
2314					pkt[i]->nb_segs > TCP_MAX_PKT_SEG)
2315				break;
2316			rc = tcp_fill_mbuf(pkt[i], s, &s->tx.dst, ol_flags,
2317				s->s.port, 0, TCP_FLAG_ACK, 0, 0);
2318			if (rc != 0)
2319				break;
2320		}
2321
2322		if (i != k) {
2323			/* queue packets for further transmission. */
2324			n = _rte_ring_enqueue_burst(s->tx.q,
2325				(void **)pkt + k, (i - k));
2326			k += n;
2327
2328			/*
2329			 * for unsent, but already modified packets:
2330			 * remove pkt l2/l3 headers, restore ol_flags
2331			 */
2332			if (i != k) {
2333				ol_flags = ~s->tx.dst.ol_flags;
2334				for (j = k; j != i; j++) {
2335					rte_pktmbuf_adj(pkt[j], pkt[j]->l2_len +
2336						pkt[j]->l3_len +
2337						pkt[j]->l4_len);
2338					pkt[j]->ol_flags &= ol_flags;
2339				}
2340				break;
2341			}
2342		}
2343
2344		if (rc != 0) {
2345			rte_errno = -rc;
2346			break;
2347
2348		/* segment large packet and enqueue for sending */
2349		} else if (i != num) {
2350			/* segment the packet. */
2351			rc = tcp_segmentation(pkt[i], segs, RTE_DIM(segs),
2352				&s->tx.dst, mss);
2353			if (rc < 0) {
2354				rte_errno = -rc;
2355				break;
2356			}
2357
2358			rc = tx_segments(s, ol_flags, segs, rc);
2359			if (rc == 0) {
2360				/* free the large mbuf */
2361				rte_pktmbuf_free(pkt[i]);
2362				/* set the mbuf as consumed */
2363				k++;
2364			} else
2365				/* no space left in tx queue */
2366				break;
2367		}
2368	}
2369
2370	/* notify BE about more data to send */
2371	if (k != 0)
2372		txs_enqueue(s->s.ctx, s);
2373	/* if possible, re-arm stream write event. */
2374	if (rte_ring_free_count(s->tx.q) != 0 && s->tx.ev != NULL)
2375		tle_event_raise(s->tx.ev);
2376
2377	tcp_stream_release(s);
2378
2379	return k;
2380}
2381
2382ssize_t
2383tle_tcp_stream_writev(struct tle_stream *ts, struct rte_mempool *mp,
2384	const struct iovec *iov, int iovcnt)
2385{
2386	int32_t i, rc;
2387	uint32_t j, k, n, num, slen, state;
2388	uint64_t ol_flags;
2389	size_t sz, tsz;
2390	struct tle_tcp_stream *s;
2391	struct iovec iv;
2392	struct rte_mbuf *mb[2 * MAX_PKT_BURST];
2393
2394	s = TCP_STREAM(ts);
2395
2396	/* mark stream as not closable. */
2397	if (tcp_stream_acquire(s) < 0) {
2398		rte_errno = EAGAIN;
2399		return -1;
2400	}
2401
2402	state = s->tcb.state;
2403	if (state != TCP_ST_ESTABLISHED && state != TCP_ST_CLOSE_WAIT) {
2404		rte_errno = ENOTCONN;
2405		tcp_stream_release(s);
2406		return -1;
2407	}
2408
2409	/* figure out how many mbufs do we need */
2410	tsz = 0;
2411	for (i = 0; i != iovcnt; i++)
2412		tsz += iov[i].iov_len;
2413
2414	slen = rte_pktmbuf_data_room_size(mp);
2415	slen = RTE_MIN(slen, s->tcb.snd.mss);
2416
2417	num = (tsz + slen - 1) / slen;
2418	n = rte_ring_free_count(s->tx.q);
2419	num = RTE_MIN(num, n);
2420	n = RTE_MIN(num, RTE_DIM(mb));
2421
2422	/* allocate mbufs */
2423	if (rte_pktmbuf_alloc_bulk(mp, mb, n) != 0) {
2424		rte_errno = ENOMEM;
2425		tcp_stream_release(s);
2426		return -1;
2427	}
2428
2429	/* copy data into the mbufs */
2430	k = 0;
2431	sz = 0;
2432	for (i = 0; i != iovcnt; i++) {
2433		iv = iov[i];
2434		sz += iv.iov_len;
2435		k += _iovec_to_mbsegs(&iv, slen, mb + k, n - k);
2436		if (iv.iov_len != 0) {
2437			sz -= iv.iov_len;
2438			break;
2439		}
2440	}
2441
2442	/* partially filled segment */
2443	k += (k != n && mb[k]->data_len != 0);
2444
2445	/* fill pkt headers */
2446	ol_flags = s->tx.dst.ol_flags;
2447
2448	for (j = 0; j != k; j++) {
2449		rc = tcp_fill_mbuf(mb[j], s, &s->tx.dst, ol_flags,
2450			s->s.port, 0, TCP_FLAG_ACK, 0, 0);
2451		if (rc != 0)
2452			break;
2453	}
2454
2455	/* if no error encountered, then enqueue pkts for transmission */
2456	if (k == j)
2457		k = _rte_ring_enqueue_burst(s->tx.q, (void **)mb, j);
2458	else
2459		k = 0;
2460
2461	if (k != j) {
2462
2463		/* free pkts that were not enqueued */
2464		free_mbufs(mb + k, j - k);
2465
2466		/* our last segment can be partially filled */
2467		sz += slen - sz % slen;
2468		sz -= (j - k) * slen;
2469
2470		/* report an error */
2471		if (rc != 0) {
2472			rte_errno = -rc;
2473			sz = -1;
2474		}
2475	}
2476
2477        if (k != 0) {
2478
2479		/* notify BE about more data to send */
2480		txs_enqueue(s->s.ctx, s);
2481
2482		/* if possible, re-arm stream write event. */
2483		if (rte_ring_free_count(s->tx.q) != 0 && s->tx.ev != NULL)
2484			tle_event_raise(s->tx.ev);
2485	}
2486
2487	tcp_stream_release(s);
2488	return sz;
2489}
2490
2491/* send data and FIN (if needed) */
2492static inline void
2493tx_data_fin(struct tle_tcp_stream *s, uint32_t tms, uint32_t state)
2494{
2495	/* try to send some data */
2496	tx_nxt_data(s, tms);
2497
2498	/* we also have to send a FIN */
2499	if (state != TCP_ST_ESTABLISHED &&
2500			state != TCP_ST_CLOSE_WAIT &&
2501			tcp_txq_nxt_cnt(s) == 0 &&
2502			s->tcb.snd.fss != s->tcb.snd.nxt) {
2503		s->tcb.snd.fss = ++s->tcb.snd.nxt;
2504		send_ack(s, tms, TCP_FLAG_FIN | TCP_FLAG_ACK);
2505	}
2506}
2507
2508static inline void
2509tx_stream(struct tle_tcp_stream *s, uint32_t tms)
2510{
2511	uint32_t state;
2512
2513	state = s->tcb.state;
2514
2515	if (state == TCP_ST_SYN_SENT) {
2516		/* send the SYN, start the rto timer */
2517		send_ack(s, tms, TCP_FLAG_SYN);
2518		timer_start(s);
2519
2520	} else if (state >= TCP_ST_ESTABLISHED && state <= TCP_ST_LAST_ACK) {
2521
2522		tx_data_fin(s, tms, state);
2523
2524		/* start RTO timer. */
2525		if (s->tcb.snd.nxt != s->tcb.snd.una)
2526			timer_start(s);
2527	}
2528}
2529
2530static inline void
2531rto_stream(struct tle_tcp_stream *s, uint32_t tms)
2532{
2533	uint32_t state;
2534
2535	state = s->tcb.state;
2536
2537	TCP_LOG(DEBUG, "%s(%p, tms=%u): state=%u, "
2538		"retx=%u, retm=%u, "
2539		"rto=%u, snd.ts=%u, tmo=%u, "
2540		"snd.nxt=%lu, snd.una=%lu, flight_size=%lu, "
2541		"snd.rcvr=%lu, snd.fastack=%u, "
2542		"wnd=%u, cwnd=%u, ssthresh=%u, "
2543		"bytes sent=%lu, pkt remain=%u;\n",
2544		__func__, s, tms, s->tcb.state,
2545		s->tcb.snd.nb_retx, s->tcb.snd.nb_retm,
2546		s->tcb.snd.rto, s->tcb.snd.ts, tms - s->tcb.snd.ts,
2547		s->tcb.snd.nxt, s->tcb.snd.una, s->tcb.snd.nxt - s->tcb.snd.una,
2548		s->tcb.snd.rcvr, s->tcb.snd.fastack,
2549		s->tcb.snd.wnd, s->tcb.snd.cwnd, s->tcb.snd.ssthresh,
2550		s->tcb.snd.nxt - s->tcb.snd.iss, tcp_txq_nxt_cnt(s));
2551
2552	if (s->tcb.snd.nb_retx < s->tcb.snd.nb_retm) {
2553
2554		if (state >= TCP_ST_ESTABLISHED && state <= TCP_ST_LAST_ACK) {
2555
2556			/* update SND.CWD and SND.SSTHRESH */
2557			rto_cwnd_update(&s->tcb);
2558
2559			/* RFC 6582 3.2.4 */
2560			s->tcb.snd.rcvr = s->tcb.snd.nxt;
2561			s->tcb.snd.fastack = 0;
2562
2563			/* restart from last acked data */
2564			tcp_txq_rst_nxt_head(s);
2565			s->tcb.snd.nxt = s->tcb.snd.una;
2566
2567			tx_data_fin(s, tms, state);
2568
2569		} else if (state == TCP_ST_SYN_SENT) {
2570			/* resending SYN */
2571			s->tcb.so.ts.val = tms;
2572
2573			/* According to RFC 6928 2:
2574			 * To reduce the chance for spurious SYN or SYN/ACK
2575			 * retransmission, it is RECOMMENDED that
2576			 * implementations refrain from resetting the initial
2577			 * window to 1 segment, unless there have been more
2578			 * than one SYN or SYN/ACK retransmissions or true loss
2579			 * detection has been made.
2580			 */
2581			if (s->tcb.snd.nb_retx != 0)
2582				s->tcb.snd.cwnd = s->tcb.snd.mss;
2583
2584			send_ack(s, tms, TCP_FLAG_SYN);
2585
2586		} else if (state == TCP_ST_TIME_WAIT) {
2587			stream_term(s);
2588		}
2589
2590		/* RFC6298:5.5 back off the timer */
2591		s->tcb.snd.rto = rto_roundup(2 * s->tcb.snd.rto);
2592		s->tcb.snd.nb_retx++;
2593		timer_restart(s);
2594
2595	} else {
2596		send_rst(s, s->tcb.snd.nxt);
2597		stream_term(s);
2598	}
2599}
2600
2601int
2602tle_tcp_process(struct tle_ctx *ctx, uint32_t num)
2603{
2604	uint32_t i, k, tms;
2605	struct sdr *dr;
2606	struct tle_timer_wheel *tw;
2607	struct tle_stream *p;
2608	struct tle_tcp_stream *s, *rs[num];
2609
2610	/* process streams with RTO exipred */
2611
2612	tw = CTX_TCP_TMWHL(ctx);
2613	tms = tcp_get_tms(ctx->cycles_ms_shift);
2614	tle_timer_expire(tw, tms);
2615
2616	k = tle_timer_get_expired_bulk(tw, (void **)rs, RTE_DIM(rs));
2617
2618	for (i = 0; i != k; i++) {
2619
2620		s = rs[i];
2621		s->timer.handle = NULL;
2622		if (tcp_stream_try_acquire(s) > 0)
2623			rto_stream(s, tms);
2624		tcp_stream_release(s);
2625	}
2626
2627	/* process streams from to-send queue */
2628
2629	k = txs_dequeue_bulk(ctx, rs, RTE_DIM(rs));
2630
2631	for (i = 0; i != k; i++) {
2632
2633		s = rs[i];
2634		rte_atomic32_set(&s->tx.arm, 0);
2635
2636		if (tcp_stream_try_acquire(s) > 0)
2637			tx_stream(s, tms);
2638		else
2639			txs_enqueue(s->s.ctx, s);
2640		tcp_stream_release(s);
2641	}
2642
2643	/* collect streams to close from the death row */
2644
2645	dr = CTX_TCP_SDR(ctx);
2646	for (k = 0, p = STAILQ_FIRST(&dr->be);
2647			k != num && p != NULL;
2648			k++, p = STAILQ_NEXT(p, link))
2649		rs[k] = TCP_STREAM(p);
2650
2651	if (p == NULL)
2652		STAILQ_INIT(&dr->be);
2653	else
2654		STAILQ_FIRST(&dr->be) = p;
2655
2656	/* cleanup closed streams */
2657	for (i = 0; i != k; i++) {
2658		s = rs[i];
2659		tcp_stream_down(s);
2660		tcp_stream_reset(ctx, s);
2661	}
2662
2663	return 0;
2664}
2665