1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 */
21
22 /*
23 * Changes:
24 * Pedro Roque : Fast Retransmit/Recovery.
25 * Two receive queues.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
29 * Header prediction.
30 * Variable renaming.
31 *
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
44 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * timestamps.
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
48 * data segments.
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
52 * Andi Kleen: Add tcp_measure_rcv_mss to make
53 * connections with MSS<min(MTU,ann. MSS)
54 * work without delayed acks.
55 * Andi Kleen: Process packets with PSH set in the
56 * fast path.
57 * J Hadi Salim: ECN support
58 * Andrei Gurtov,
59 * Pasi Sarolahti,
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
63 */
64
65 #define pr_fmt(fmt) "TCP: " fmt
66
67 #include <linux/mm.h>
68 #include <linux/slab.h>
69 #include <linux/module.h>
70 #include <linux/sysctl.h>
71 #include <linux/kernel.h>
72 #include <linux/prefetch.h>
73 #include <net/dst.h>
74 #include <net/tcp.h>
75 #include <net/proto_memory.h>
76 #include <net/inet_common.h>
77 #include <linux/ipsec.h>
78 #include <linux/unaligned.h>
79 #include <linux/errqueue.h>
80 #include <trace/events/tcp.h>
81 #include <linux/jump_label_ratelimit.h>
82 #include <net/busy_poll.h>
83 #include <net/mptcp.h>
84
85 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
86
87 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
88 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
89 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
90 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
91 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
92 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
93 #define FLAG_ECE 0x40 /* ECE in this ACK */
94 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
95 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
96 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
97 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
98 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
99 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
100 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
101 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
102 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
103 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
104 #define FLAG_DSACK_TLP 0x20000 /* DSACK for tail loss probe */
105 #define FLAG_TS_PROGRESS 0x40000 /* Positive timestamp delta */
106
107 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
108 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
109 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
110 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
111
112 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
113 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
114
115 #define REXMIT_NONE 0 /* no loss recovery to do */
116 #define REXMIT_LOST 1 /* retransmit packets marked lost */
117 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
118
119 #if IS_ENABLED(CONFIG_TLS_DEVICE)
120 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ);
121
clean_acked_data_enable(struct tcp_sock * tp,void (* cad)(struct sock * sk,u32 ack_seq))122 void clean_acked_data_enable(struct tcp_sock *tp,
123 void (*cad)(struct sock *sk, u32 ack_seq))
124 {
125 tp->tcp_clean_acked = cad;
126 static_branch_deferred_inc(&clean_acked_data_enabled);
127 }
128 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
129
clean_acked_data_disable(struct tcp_sock * tp)130 void clean_acked_data_disable(struct tcp_sock *tp)
131 {
132 static_branch_slow_dec_deferred(&clean_acked_data_enabled);
133 tp->tcp_clean_acked = NULL;
134 }
135 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
136
clean_acked_data_flush(void)137 void clean_acked_data_flush(void)
138 {
139 static_key_deferred_flush(&clean_acked_data_enabled);
140 }
141 EXPORT_SYMBOL_GPL(clean_acked_data_flush);
142 #endif
143
144 #ifdef CONFIG_CGROUP_BPF
bpf_skops_parse_hdr(struct sock * sk,struct sk_buff * skb)145 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
146 {
147 bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown &&
148 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
149 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG);
150 bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
151 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG);
152 struct bpf_sock_ops_kern sock_ops;
153
154 if (likely(!unknown_opt && !parse_all_opt))
155 return;
156
157 /* The skb will be handled in the
158 * bpf_skops_established() or
159 * bpf_skops_write_hdr_opt().
160 */
161 switch (sk->sk_state) {
162 case TCP_SYN_RECV:
163 case TCP_SYN_SENT:
164 case TCP_LISTEN:
165 return;
166 }
167
168 sock_owned_by_me(sk);
169
170 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
171 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB;
172 sock_ops.is_fullsock = 1;
173 sock_ops.is_locked_tcp_sock = 1;
174 sock_ops.sk = sk;
175 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
176
177 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
178 }
179
bpf_skops_established(struct sock * sk,int bpf_op,struct sk_buff * skb)180 static void bpf_skops_established(struct sock *sk, int bpf_op,
181 struct sk_buff *skb)
182 {
183 struct bpf_sock_ops_kern sock_ops;
184
185 sock_owned_by_me(sk);
186
187 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
188 sock_ops.op = bpf_op;
189 sock_ops.is_fullsock = 1;
190 sock_ops.is_locked_tcp_sock = 1;
191 sock_ops.sk = sk;
192 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */
193 if (skb)
194 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
195
196 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
197 }
198 #else
bpf_skops_parse_hdr(struct sock * sk,struct sk_buff * skb)199 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
200 {
201 }
202
bpf_skops_established(struct sock * sk,int bpf_op,struct sk_buff * skb)203 static void bpf_skops_established(struct sock *sk, int bpf_op,
204 struct sk_buff *skb)
205 {
206 }
207 #endif
208
tcp_gro_dev_warn(const struct sock * sk,const struct sk_buff * skb,unsigned int len)209 static __cold void tcp_gro_dev_warn(const struct sock *sk, const struct sk_buff *skb,
210 unsigned int len)
211 {
212 struct net_device *dev;
213
214 rcu_read_lock();
215 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
216 if (!dev || len >= READ_ONCE(dev->mtu))
217 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
218 dev ? dev->name : "Unknown driver");
219 rcu_read_unlock();
220 }
221
222 /* Adapt the MSS value used to make delayed ack decision to the
223 * real world.
224 */
tcp_measure_rcv_mss(struct sock * sk,const struct sk_buff * skb)225 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
226 {
227 struct inet_connection_sock *icsk = inet_csk(sk);
228 const unsigned int lss = icsk->icsk_ack.last_seg_size;
229 unsigned int len;
230
231 icsk->icsk_ack.last_seg_size = 0;
232
233 /* skb->len may jitter because of SACKs, even if peer
234 * sends good full-sized frames.
235 */
236 len = skb_shinfo(skb)->gso_size ? : skb->len;
237 if (len >= icsk->icsk_ack.rcv_mss) {
238 /* Note: divides are still a bit expensive.
239 * For the moment, only adjust scaling_ratio
240 * when we update icsk_ack.rcv_mss.
241 */
242 if (unlikely(len != icsk->icsk_ack.rcv_mss)) {
243 u64 val = (u64)skb->len << TCP_RMEM_TO_WIN_SCALE;
244 u8 old_ratio = tcp_sk(sk)->scaling_ratio;
245
246 do_div(val, skb->truesize);
247 tcp_sk(sk)->scaling_ratio = val ? val : 1;
248
249 if (old_ratio != tcp_sk(sk)->scaling_ratio) {
250 struct tcp_sock *tp = tcp_sk(sk);
251
252 val = tcp_win_from_space(sk, sk->sk_rcvbuf);
253 tcp_set_window_clamp(sk, val);
254
255 if (tp->window_clamp < tp->rcvq_space.space)
256 tp->rcvq_space.space = tp->window_clamp;
257 }
258 }
259 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
260 tcp_sk(sk)->advmss);
261 /* Account for possibly-removed options */
262 DO_ONCE_LITE_IF(len > icsk->icsk_ack.rcv_mss + MAX_TCP_OPTION_SPACE,
263 tcp_gro_dev_warn, sk, skb, len);
264 /* If the skb has a len of exactly 1*MSS and has the PSH bit
265 * set then it is likely the end of an application write. So
266 * more data may not be arriving soon, and yet the data sender
267 * may be waiting for an ACK if cwnd-bound or using TX zero
268 * copy. So we set ICSK_ACK_PUSHED here so that
269 * tcp_cleanup_rbuf() will send an ACK immediately if the app
270 * reads all of the data and is not ping-pong. If len > MSS
271 * then this logic does not matter (and does not hurt) because
272 * tcp_cleanup_rbuf() will always ACK immediately if the app
273 * reads data and there is more than an MSS of unACKed data.
274 */
275 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH)
276 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
277 } else {
278 /* Otherwise, we make more careful check taking into account,
279 * that SACKs block is variable.
280 *
281 * "len" is invariant segment length, including TCP header.
282 */
283 len += skb->data - skb_transport_header(skb);
284 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
285 /* If PSH is not set, packet should be
286 * full sized, provided peer TCP is not badly broken.
287 * This observation (if it is correct 8)) allows
288 * to handle super-low mtu links fairly.
289 */
290 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
291 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
292 /* Subtract also invariant (if peer is RFC compliant),
293 * tcp header plus fixed timestamp option length.
294 * Resulting "len" is MSS free of SACK jitter.
295 */
296 len -= tcp_sk(sk)->tcp_header_len;
297 icsk->icsk_ack.last_seg_size = len;
298 if (len == lss) {
299 icsk->icsk_ack.rcv_mss = len;
300 return;
301 }
302 }
303 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
304 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
305 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
306 }
307 }
308
tcp_incr_quickack(struct sock * sk,unsigned int max_quickacks)309 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
310 {
311 struct inet_connection_sock *icsk = inet_csk(sk);
312 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
313
314 if (quickacks == 0)
315 quickacks = 2;
316 quickacks = min(quickacks, max_quickacks);
317 if (quickacks > icsk->icsk_ack.quick)
318 icsk->icsk_ack.quick = quickacks;
319 }
320
tcp_enter_quickack_mode(struct sock * sk,unsigned int max_quickacks)321 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
322 {
323 struct inet_connection_sock *icsk = inet_csk(sk);
324
325 tcp_incr_quickack(sk, max_quickacks);
326 inet_csk_exit_pingpong_mode(sk);
327 icsk->icsk_ack.ato = TCP_ATO_MIN;
328 }
329
330 /* Send ACKs quickly, if "quick" count is not exhausted
331 * and the session is not interactive.
332 */
333
tcp_in_quickack_mode(struct sock * sk)334 static bool tcp_in_quickack_mode(struct sock *sk)
335 {
336 const struct inet_connection_sock *icsk = inet_csk(sk);
337
338 return icsk->icsk_ack.dst_quick_ack ||
339 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
340 }
341
tcp_ecn_queue_cwr(struct tcp_sock * tp)342 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
343 {
344 if (tcp_ecn_mode_rfc3168(tp))
345 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
346 }
347
tcp_ecn_accept_cwr(struct sock * sk,const struct sk_buff * skb)348 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
349 {
350 if (tcp_hdr(skb)->cwr) {
351 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
352
353 /* If the sender is telling us it has entered CWR, then its
354 * cwnd may be very low (even just 1 packet), so we should ACK
355 * immediately.
356 */
357 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
358 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
359 }
360 }
361
tcp_ecn_withdraw_cwr(struct tcp_sock * tp)362 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
363 {
364 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
365 }
366
tcp_data_ecn_check(struct sock * sk,const struct sk_buff * skb)367 static void tcp_data_ecn_check(struct sock *sk, const struct sk_buff *skb)
368 {
369 struct tcp_sock *tp = tcp_sk(sk);
370
371 if (tcp_ecn_disabled(tp))
372 return;
373
374 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
375 case INET_ECN_NOT_ECT:
376 /* Funny extension: if ECT is not set on a segment,
377 * and we already seen ECT on a previous segment,
378 * it is probably a retransmit.
379 */
380 if (tp->ecn_flags & TCP_ECN_SEEN)
381 tcp_enter_quickack_mode(sk, 2);
382 break;
383 case INET_ECN_CE:
384 if (tcp_ca_needs_ecn(sk))
385 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
386
387 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
388 /* Better not delay acks, sender can have a very low cwnd */
389 tcp_enter_quickack_mode(sk, 2);
390 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
391 }
392 tp->ecn_flags |= TCP_ECN_SEEN;
393 break;
394 default:
395 if (tcp_ca_needs_ecn(sk))
396 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
397 tp->ecn_flags |= TCP_ECN_SEEN;
398 break;
399 }
400 }
401
tcp_ecn_rcv_synack(struct tcp_sock * tp,const struct tcphdr * th)402 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
403 {
404 if (tcp_ecn_mode_rfc3168(tp) && (!th->ece || th->cwr))
405 tcp_ecn_mode_set(tp, TCP_ECN_DISABLED);
406 }
407
tcp_ecn_rcv_syn(struct tcp_sock * tp,const struct tcphdr * th)408 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
409 {
410 if (tcp_ecn_mode_rfc3168(tp) && (!th->ece || !th->cwr))
411 tcp_ecn_mode_set(tp, TCP_ECN_DISABLED);
412 }
413
tcp_ecn_rcv_ecn_echo(const struct tcp_sock * tp,const struct tcphdr * th)414 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
415 {
416 if (th->ece && !th->syn && tcp_ecn_mode_rfc3168(tp))
417 return true;
418 return false;
419 }
420
tcp_count_delivered_ce(struct tcp_sock * tp,u32 ecn_count)421 static void tcp_count_delivered_ce(struct tcp_sock *tp, u32 ecn_count)
422 {
423 tp->delivered_ce += ecn_count;
424 }
425
426 /* Updates the delivered and delivered_ce counts */
tcp_count_delivered(struct tcp_sock * tp,u32 delivered,bool ece_ack)427 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
428 bool ece_ack)
429 {
430 tp->delivered += delivered;
431 if (ece_ack)
432 tcp_count_delivered_ce(tp, delivered);
433 }
434
435 /* Buffer size and advertised window tuning.
436 *
437 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
438 */
439
tcp_sndbuf_expand(struct sock * sk)440 static void tcp_sndbuf_expand(struct sock *sk)
441 {
442 const struct tcp_sock *tp = tcp_sk(sk);
443 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
444 int sndmem, per_mss;
445 u32 nr_segs;
446
447 /* Worst case is non GSO/TSO : each frame consumes one skb
448 * and skb->head is kmalloced using power of two area of memory
449 */
450 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
451 MAX_TCP_HEADER +
452 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
453
454 per_mss = roundup_pow_of_two(per_mss) +
455 SKB_DATA_ALIGN(sizeof(struct sk_buff));
456
457 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp));
458 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
459
460 /* Fast Recovery (RFC 5681 3.2) :
461 * Cubic needs 1.7 factor, rounded to 2 to include
462 * extra cushion (application might react slowly to EPOLLOUT)
463 */
464 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
465 sndmem *= nr_segs * per_mss;
466
467 if (sk->sk_sndbuf < sndmem)
468 WRITE_ONCE(sk->sk_sndbuf,
469 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2])));
470 }
471
472 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
473 *
474 * All tcp_full_space() is split to two parts: "network" buffer, allocated
475 * forward and advertised in receiver window (tp->rcv_wnd) and
476 * "application buffer", required to isolate scheduling/application
477 * latencies from network.
478 * window_clamp is maximal advertised window. It can be less than
479 * tcp_full_space(), in this case tcp_full_space() - window_clamp
480 * is reserved for "application" buffer. The less window_clamp is
481 * the smoother our behaviour from viewpoint of network, but the lower
482 * throughput and the higher sensitivity of the connection to losses. 8)
483 *
484 * rcv_ssthresh is more strict window_clamp used at "slow start"
485 * phase to predict further behaviour of this connection.
486 * It is used for two goals:
487 * - to enforce header prediction at sender, even when application
488 * requires some significant "application buffer". It is check #1.
489 * - to prevent pruning of receive queue because of misprediction
490 * of receiver window. Check #2.
491 *
492 * The scheme does not work when sender sends good segments opening
493 * window and then starts to feed us spaghetti. But it should work
494 * in common situations. Otherwise, we have to rely on queue collapsing.
495 */
496
497 /* Slow part of check#2. */
__tcp_grow_window(const struct sock * sk,const struct sk_buff * skb,unsigned int skbtruesize)498 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb,
499 unsigned int skbtruesize)
500 {
501 const struct tcp_sock *tp = tcp_sk(sk);
502 /* Optimize this! */
503 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1;
504 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1;
505
506 while (tp->rcv_ssthresh <= window) {
507 if (truesize <= skb->len)
508 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
509
510 truesize >>= 1;
511 window >>= 1;
512 }
513 return 0;
514 }
515
516 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing
517 * can play nice with us, as sk_buff and skb->head might be either
518 * freed or shared with up to MAX_SKB_FRAGS segments.
519 * Only give a boost to drivers using page frag(s) to hold the frame(s),
520 * and if no payload was pulled in skb->head before reaching us.
521 */
truesize_adjust(bool adjust,const struct sk_buff * skb)522 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb)
523 {
524 u32 truesize = skb->truesize;
525
526 if (adjust && !skb_headlen(skb)) {
527 truesize -= SKB_TRUESIZE(skb_end_offset(skb));
528 /* paranoid check, some drivers might be buggy */
529 if (unlikely((int)truesize < (int)skb->len))
530 truesize = skb->truesize;
531 }
532 return truesize;
533 }
534
tcp_grow_window(struct sock * sk,const struct sk_buff * skb,bool adjust)535 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb,
536 bool adjust)
537 {
538 struct tcp_sock *tp = tcp_sk(sk);
539 int room;
540
541 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
542
543 if (room <= 0)
544 return;
545
546 /* Check #1 */
547 if (!tcp_under_memory_pressure(sk)) {
548 unsigned int truesize = truesize_adjust(adjust, skb);
549 int incr;
550
551 /* Check #2. Increase window, if skb with such overhead
552 * will fit to rcvbuf in future.
553 */
554 if (tcp_win_from_space(sk, truesize) <= skb->len)
555 incr = 2 * tp->advmss;
556 else
557 incr = __tcp_grow_window(sk, skb, truesize);
558
559 if (incr) {
560 incr = max_t(int, incr, 2 * skb->len);
561 tp->rcv_ssthresh += min(room, incr);
562 inet_csk(sk)->icsk_ack.quick |= 1;
563 }
564 } else {
565 /* Under pressure:
566 * Adjust rcv_ssthresh according to reserved mem
567 */
568 tcp_adjust_rcv_ssthresh(sk);
569 }
570 }
571
572 /* 3. Try to fixup all. It is made immediately after connection enters
573 * established state.
574 */
tcp_init_buffer_space(struct sock * sk)575 static void tcp_init_buffer_space(struct sock *sk)
576 {
577 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win);
578 struct tcp_sock *tp = tcp_sk(sk);
579 int maxwin;
580
581 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
582 tcp_sndbuf_expand(sk);
583
584 tcp_mstamp_refresh(tp);
585 tp->rcvq_space.time = tp->tcp_mstamp;
586 tp->rcvq_space.seq = tp->copied_seq;
587
588 maxwin = tcp_full_space(sk);
589
590 if (tp->window_clamp >= maxwin) {
591 WRITE_ONCE(tp->window_clamp, maxwin);
592
593 if (tcp_app_win && maxwin > 4 * tp->advmss)
594 WRITE_ONCE(tp->window_clamp,
595 max(maxwin - (maxwin >> tcp_app_win),
596 4 * tp->advmss));
597 }
598
599 /* Force reservation of one segment. */
600 if (tcp_app_win &&
601 tp->window_clamp > 2 * tp->advmss &&
602 tp->window_clamp + tp->advmss > maxwin)
603 WRITE_ONCE(tp->window_clamp,
604 max(2 * tp->advmss, maxwin - tp->advmss));
605
606 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
607 tp->snd_cwnd_stamp = tcp_jiffies32;
608 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
609 (u32)TCP_INIT_CWND * tp->advmss);
610 }
611
612 /* 4. Recalculate window clamp after socket hit its memory bounds. */
tcp_clamp_window(struct sock * sk)613 static void tcp_clamp_window(struct sock *sk)
614 {
615 struct tcp_sock *tp = tcp_sk(sk);
616 struct inet_connection_sock *icsk = inet_csk(sk);
617 struct net *net = sock_net(sk);
618 int rmem2;
619
620 icsk->icsk_ack.quick = 0;
621 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
622
623 if (sk->sk_rcvbuf < rmem2 &&
624 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
625 !tcp_under_memory_pressure(sk) &&
626 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
627 WRITE_ONCE(sk->sk_rcvbuf,
628 min(atomic_read(&sk->sk_rmem_alloc), rmem2));
629 }
630 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
631 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
632 }
633
634 /* Initialize RCV_MSS value.
635 * RCV_MSS is an our guess about MSS used by the peer.
636 * We haven't any direct information about the MSS.
637 * It's better to underestimate the RCV_MSS rather than overestimate.
638 * Overestimations make us ACKing less frequently than needed.
639 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
640 */
tcp_initialize_rcv_mss(struct sock * sk)641 void tcp_initialize_rcv_mss(struct sock *sk)
642 {
643 const struct tcp_sock *tp = tcp_sk(sk);
644 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
645
646 hint = min(hint, tp->rcv_wnd / 2);
647 hint = min(hint, TCP_MSS_DEFAULT);
648 hint = max(hint, TCP_MIN_MSS);
649
650 inet_csk(sk)->icsk_ack.rcv_mss = hint;
651 }
652 EXPORT_IPV6_MOD(tcp_initialize_rcv_mss);
653
654 /* Receiver "autotuning" code.
655 *
656 * The algorithm for RTT estimation w/o timestamps is based on
657 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
658 * <https://public.lanl.gov/radiant/pubs.html#DRS>
659 *
660 * More detail on this code can be found at
661 * <http://staff.psc.edu/jheffner/>,
662 * though this reference is out of date. A new paper
663 * is pending.
664 */
tcp_rcv_rtt_update(struct tcp_sock * tp,u32 sample,int win_dep)665 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
666 {
667 u32 new_sample = tp->rcv_rtt_est.rtt_us;
668 long m = sample;
669
670 if (new_sample != 0) {
671 /* If we sample in larger samples in the non-timestamp
672 * case, we could grossly overestimate the RTT especially
673 * with chatty applications or bulk transfer apps which
674 * are stalled on filesystem I/O.
675 *
676 * Also, since we are only going for a minimum in the
677 * non-timestamp case, we do not smooth things out
678 * else with timestamps disabled convergence takes too
679 * long.
680 */
681 if (!win_dep) {
682 m -= (new_sample >> 3);
683 new_sample += m;
684 } else {
685 m <<= 3;
686 if (m < new_sample)
687 new_sample = m;
688 }
689 } else {
690 /* No previous measure. */
691 new_sample = m << 3;
692 }
693
694 tp->rcv_rtt_est.rtt_us = new_sample;
695 }
696
tcp_rcv_rtt_measure(struct tcp_sock * tp)697 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
698 {
699 u32 delta_us;
700
701 if (tp->rcv_rtt_est.time == 0)
702 goto new_measure;
703 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
704 return;
705 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
706 if (!delta_us)
707 delta_us = 1;
708 tcp_rcv_rtt_update(tp, delta_us, 1);
709
710 new_measure:
711 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
712 tp->rcv_rtt_est.time = tp->tcp_mstamp;
713 }
714
tcp_rtt_tsopt_us(const struct tcp_sock * tp)715 static s32 tcp_rtt_tsopt_us(const struct tcp_sock *tp)
716 {
717 u32 delta, delta_us;
718
719 delta = tcp_time_stamp_ts(tp) - tp->rx_opt.rcv_tsecr;
720 if (tp->tcp_usec_ts)
721 return delta;
722
723 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
724 if (!delta)
725 delta = 1;
726 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
727 return delta_us;
728 }
729 return -1;
730 }
731
tcp_rcv_rtt_measure_ts(struct sock * sk,const struct sk_buff * skb)732 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
733 const struct sk_buff *skb)
734 {
735 struct tcp_sock *tp = tcp_sk(sk);
736
737 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
738 return;
739 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
740
741 if (TCP_SKB_CB(skb)->end_seq -
742 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
743 s32 delta = tcp_rtt_tsopt_us(tp);
744
745 if (delta >= 0)
746 tcp_rcv_rtt_update(tp, delta, 0);
747 }
748 }
749
750 /*
751 * This function should be called every time data is copied to user space.
752 * It calculates the appropriate TCP receive buffer space.
753 */
tcp_rcv_space_adjust(struct sock * sk)754 void tcp_rcv_space_adjust(struct sock *sk)
755 {
756 struct tcp_sock *tp = tcp_sk(sk);
757 u32 copied;
758 int time;
759
760 trace_tcp_rcv_space_adjust(sk);
761
762 tcp_mstamp_refresh(tp);
763 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
764 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
765 return;
766
767 /* Number of bytes copied to user in last RTT */
768 copied = tp->copied_seq - tp->rcvq_space.seq;
769 if (copied <= tp->rcvq_space.space)
770 goto new_measure;
771
772 /* A bit of theory :
773 * copied = bytes received in previous RTT, our base window
774 * To cope with packet losses, we need a 2x factor
775 * To cope with slow start, and sender growing its cwin by 100 %
776 * every RTT, we need a 4x factor, because the ACK we are sending
777 * now is for the next RTT, not the current one :
778 * <prev RTT . ><current RTT .. ><next RTT .... >
779 */
780
781 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
782 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
783 u64 rcvwin, grow;
784 int rcvbuf;
785
786 /* minimal window to cope with packet losses, assuming
787 * steady state. Add some cushion because of small variations.
788 */
789 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
790
791 /* Accommodate for sender rate increase (eg. slow start) */
792 grow = rcvwin * (copied - tp->rcvq_space.space);
793 do_div(grow, tp->rcvq_space.space);
794 rcvwin += (grow << 1);
795
796 rcvbuf = min_t(u64, tcp_space_from_win(sk, rcvwin),
797 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
798 if (rcvbuf > sk->sk_rcvbuf) {
799 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
800
801 /* Make the window clamp follow along. */
802 WRITE_ONCE(tp->window_clamp,
803 tcp_win_from_space(sk, rcvbuf));
804 }
805 }
806 tp->rcvq_space.space = copied;
807
808 new_measure:
809 tp->rcvq_space.seq = tp->copied_seq;
810 tp->rcvq_space.time = tp->tcp_mstamp;
811 }
812
tcp_save_lrcv_flowlabel(struct sock * sk,const struct sk_buff * skb)813 static void tcp_save_lrcv_flowlabel(struct sock *sk, const struct sk_buff *skb)
814 {
815 #if IS_ENABLED(CONFIG_IPV6)
816 struct inet_connection_sock *icsk = inet_csk(sk);
817
818 if (skb->protocol == htons(ETH_P_IPV6))
819 icsk->icsk_ack.lrcv_flowlabel = ntohl(ip6_flowlabel(ipv6_hdr(skb)));
820 #endif
821 }
822
823 /* There is something which you must keep in mind when you analyze the
824 * behavior of the tp->ato delayed ack timeout interval. When a
825 * connection starts up, we want to ack as quickly as possible. The
826 * problem is that "good" TCP's do slow start at the beginning of data
827 * transmission. The means that until we send the first few ACK's the
828 * sender will sit on his end and only queue most of his data, because
829 * he can only send snd_cwnd unacked packets at any given time. For
830 * each ACK we send, he increments snd_cwnd and transmits more of his
831 * queue. -DaveM
832 */
tcp_event_data_recv(struct sock * sk,struct sk_buff * skb)833 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
834 {
835 struct tcp_sock *tp = tcp_sk(sk);
836 struct inet_connection_sock *icsk = inet_csk(sk);
837 u32 now;
838
839 inet_csk_schedule_ack(sk);
840
841 tcp_measure_rcv_mss(sk, skb);
842
843 tcp_rcv_rtt_measure(tp);
844
845 now = tcp_jiffies32;
846
847 if (!icsk->icsk_ack.ato) {
848 /* The _first_ data packet received, initialize
849 * delayed ACK engine.
850 */
851 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
852 icsk->icsk_ack.ato = TCP_ATO_MIN;
853 } else {
854 int m = now - icsk->icsk_ack.lrcvtime;
855
856 if (m <= TCP_ATO_MIN / 2) {
857 /* The fastest case is the first. */
858 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
859 } else if (m < icsk->icsk_ack.ato) {
860 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
861 if (icsk->icsk_ack.ato > icsk->icsk_rto)
862 icsk->icsk_ack.ato = icsk->icsk_rto;
863 } else if (m > icsk->icsk_rto) {
864 /* Too long gap. Apparently sender failed to
865 * restart window, so that we send ACKs quickly.
866 */
867 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
868 }
869 }
870 icsk->icsk_ack.lrcvtime = now;
871 tcp_save_lrcv_flowlabel(sk, skb);
872
873 tcp_data_ecn_check(sk, skb);
874
875 if (skb->len >= 128)
876 tcp_grow_window(sk, skb, true);
877 }
878
879 /* Called to compute a smoothed rtt estimate. The data fed to this
880 * routine either comes from timestamps, or from segments that were
881 * known _not_ to have been retransmitted [see Karn/Partridge
882 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
883 * piece by Van Jacobson.
884 * NOTE: the next three routines used to be one big routine.
885 * To save cycles in the RFC 1323 implementation it was better to break
886 * it up into three procedures. -- erics
887 */
tcp_rtt_estimator(struct sock * sk,long mrtt_us)888 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
889 {
890 struct tcp_sock *tp = tcp_sk(sk);
891 long m = mrtt_us; /* RTT */
892 u32 srtt = tp->srtt_us;
893
894 /* The following amusing code comes from Jacobson's
895 * article in SIGCOMM '88. Note that rtt and mdev
896 * are scaled versions of rtt and mean deviation.
897 * This is designed to be as fast as possible
898 * m stands for "measurement".
899 *
900 * On a 1990 paper the rto value is changed to:
901 * RTO = rtt + 4 * mdev
902 *
903 * Funny. This algorithm seems to be very broken.
904 * These formulae increase RTO, when it should be decreased, increase
905 * too slowly, when it should be increased quickly, decrease too quickly
906 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
907 * does not matter how to _calculate_ it. Seems, it was trap
908 * that VJ failed to avoid. 8)
909 */
910 if (srtt != 0) {
911 m -= (srtt >> 3); /* m is now error in rtt est */
912 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
913 if (m < 0) {
914 m = -m; /* m is now abs(error) */
915 m -= (tp->mdev_us >> 2); /* similar update on mdev */
916 /* This is similar to one of Eifel findings.
917 * Eifel blocks mdev updates when rtt decreases.
918 * This solution is a bit different: we use finer gain
919 * for mdev in this case (alpha*beta).
920 * Like Eifel it also prevents growth of rto,
921 * but also it limits too fast rto decreases,
922 * happening in pure Eifel.
923 */
924 if (m > 0)
925 m >>= 3;
926 } else {
927 m -= (tp->mdev_us >> 2); /* similar update on mdev */
928 }
929 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
930 if (tp->mdev_us > tp->mdev_max_us) {
931 tp->mdev_max_us = tp->mdev_us;
932 if (tp->mdev_max_us > tp->rttvar_us)
933 tp->rttvar_us = tp->mdev_max_us;
934 }
935 if (after(tp->snd_una, tp->rtt_seq)) {
936 if (tp->mdev_max_us < tp->rttvar_us)
937 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
938 tp->rtt_seq = tp->snd_nxt;
939 tp->mdev_max_us = tcp_rto_min_us(sk);
940
941 tcp_bpf_rtt(sk, mrtt_us, srtt);
942 }
943 } else {
944 /* no previous measure. */
945 srtt = m << 3; /* take the measured time to be rtt */
946 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
947 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
948 tp->mdev_max_us = tp->rttvar_us;
949 tp->rtt_seq = tp->snd_nxt;
950
951 tcp_bpf_rtt(sk, mrtt_us, srtt);
952 }
953 tp->srtt_us = max(1U, srtt);
954 }
955
tcp_update_pacing_rate(struct sock * sk)956 static void tcp_update_pacing_rate(struct sock *sk)
957 {
958 const struct tcp_sock *tp = tcp_sk(sk);
959 u64 rate;
960
961 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
962 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
963
964 /* current rate is (cwnd * mss) / srtt
965 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
966 * In Congestion Avoidance phase, set it to 120 % the current rate.
967 *
968 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
969 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
970 * end of slow start and should slow down.
971 */
972 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2)
973 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio);
974 else
975 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio);
976
977 rate *= max(tcp_snd_cwnd(tp), tp->packets_out);
978
979 if (likely(tp->srtt_us))
980 do_div(rate, tp->srtt_us);
981
982 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
983 * without any lock. We want to make sure compiler wont store
984 * intermediate values in this location.
985 */
986 WRITE_ONCE(sk->sk_pacing_rate,
987 min_t(u64, rate, READ_ONCE(sk->sk_max_pacing_rate)));
988 }
989
990 /* Calculate rto without backoff. This is the second half of Van Jacobson's
991 * routine referred to above.
992 */
tcp_set_rto(struct sock * sk)993 static void tcp_set_rto(struct sock *sk)
994 {
995 const struct tcp_sock *tp = tcp_sk(sk);
996 /* Old crap is replaced with new one. 8)
997 *
998 * More seriously:
999 * 1. If rtt variance happened to be less 50msec, it is hallucination.
1000 * It cannot be less due to utterly erratic ACK generation made
1001 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
1002 * to do with delayed acks, because at cwnd>2 true delack timeout
1003 * is invisible. Actually, Linux-2.4 also generates erratic
1004 * ACKs in some circumstances.
1005 */
1006 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
1007
1008 /* 2. Fixups made earlier cannot be right.
1009 * If we do not estimate RTO correctly without them,
1010 * all the algo is pure shit and should be replaced
1011 * with correct one. It is exactly, which we pretend to do.
1012 */
1013
1014 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
1015 * guarantees that rto is higher.
1016 */
1017 tcp_bound_rto(sk);
1018 }
1019
tcp_init_cwnd(const struct tcp_sock * tp,const struct dst_entry * dst)1020 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
1021 {
1022 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
1023
1024 if (!cwnd)
1025 cwnd = TCP_INIT_CWND;
1026 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
1027 }
1028
1029 struct tcp_sacktag_state {
1030 /* Timestamps for earliest and latest never-retransmitted segment
1031 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1032 * but congestion control should still get an accurate delay signal.
1033 */
1034 u64 first_sackt;
1035 u64 last_sackt;
1036 u32 reord;
1037 u32 sack_delivered;
1038 int flag;
1039 unsigned int mss_now;
1040 struct rate_sample *rate;
1041 };
1042
1043 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
1044 * and spurious retransmission information if this DSACK is unlikely caused by
1045 * sender's action:
1046 * - DSACKed sequence range is larger than maximum receiver's window.
1047 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
1048 */
tcp_dsack_seen(struct tcp_sock * tp,u32 start_seq,u32 end_seq,struct tcp_sacktag_state * state)1049 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
1050 u32 end_seq, struct tcp_sacktag_state *state)
1051 {
1052 u32 seq_len, dup_segs = 1;
1053
1054 if (!before(start_seq, end_seq))
1055 return 0;
1056
1057 seq_len = end_seq - start_seq;
1058 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
1059 if (seq_len > tp->max_window)
1060 return 0;
1061 if (seq_len > tp->mss_cache)
1062 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
1063 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq)
1064 state->flag |= FLAG_DSACK_TLP;
1065
1066 tp->dsack_dups += dup_segs;
1067 /* Skip the DSACK if dup segs weren't retransmitted by sender */
1068 if (tp->dsack_dups > tp->total_retrans)
1069 return 0;
1070
1071 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
1072 /* We increase the RACK ordering window in rounds where we receive
1073 * DSACKs that may have been due to reordering causing RACK to trigger
1074 * a spurious fast recovery. Thus RACK ignores DSACKs that happen
1075 * without having seen reordering, or that match TLP probes (TLP
1076 * is timer-driven, not triggered by RACK).
1077 */
1078 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP))
1079 tp->rack.dsack_seen = 1;
1080
1081 state->flag |= FLAG_DSACKING_ACK;
1082 /* A spurious retransmission is delivered */
1083 state->sack_delivered += dup_segs;
1084
1085 return dup_segs;
1086 }
1087
1088 /* It's reordering when higher sequence was delivered (i.e. sacked) before
1089 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1090 * distance is approximated in full-mss packet distance ("reordering").
1091 */
tcp_check_sack_reordering(struct sock * sk,const u32 low_seq,const int ts)1092 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1093 const int ts)
1094 {
1095 struct tcp_sock *tp = tcp_sk(sk);
1096 const u32 mss = tp->mss_cache;
1097 u32 fack, metric;
1098
1099 fack = tcp_highest_sack_seq(tp);
1100 if (!before(low_seq, fack))
1101 return;
1102
1103 metric = fack - low_seq;
1104 if ((metric > tp->reordering * mss) && mss) {
1105 #if FASTRETRANS_DEBUG > 1
1106 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1107 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1108 tp->reordering,
1109 0,
1110 tp->sacked_out,
1111 tp->undo_marker ? tp->undo_retrans : 0);
1112 #endif
1113 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1114 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1115 }
1116
1117 /* This exciting event is worth to be remembered. 8) */
1118 tp->reord_seen++;
1119 NET_INC_STATS(sock_net(sk),
1120 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1121 }
1122
1123 /* This must be called before lost_out or retrans_out are updated
1124 * on a new loss, because we want to know if all skbs previously
1125 * known to be lost have already been retransmitted, indicating
1126 * that this newly lost skb is our next skb to retransmit.
1127 */
tcp_verify_retransmit_hint(struct tcp_sock * tp,struct sk_buff * skb)1128 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1129 {
1130 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1131 (tp->retransmit_skb_hint &&
1132 before(TCP_SKB_CB(skb)->seq,
1133 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1134 tp->retransmit_skb_hint = skb;
1135 }
1136
1137 /* Sum the number of packets on the wire we have marked as lost, and
1138 * notify the congestion control module that the given skb was marked lost.
1139 */
tcp_notify_skb_loss_event(struct tcp_sock * tp,const struct sk_buff * skb)1140 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1141 {
1142 tp->lost += tcp_skb_pcount(skb);
1143 }
1144
tcp_mark_skb_lost(struct sock * sk,struct sk_buff * skb)1145 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1146 {
1147 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1148 struct tcp_sock *tp = tcp_sk(sk);
1149
1150 if (sacked & TCPCB_SACKED_ACKED)
1151 return;
1152
1153 tcp_verify_retransmit_hint(tp, skb);
1154 if (sacked & TCPCB_LOST) {
1155 if (sacked & TCPCB_SACKED_RETRANS) {
1156 /* Account for retransmits that are lost again */
1157 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1158 tp->retrans_out -= tcp_skb_pcount(skb);
1159 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1160 tcp_skb_pcount(skb));
1161 tcp_notify_skb_loss_event(tp, skb);
1162 }
1163 } else {
1164 tp->lost_out += tcp_skb_pcount(skb);
1165 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1166 tcp_notify_skb_loss_event(tp, skb);
1167 }
1168 }
1169
1170 /* This procedure tags the retransmission queue when SACKs arrive.
1171 *
1172 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1173 * Packets in queue with these bits set are counted in variables
1174 * sacked_out, retrans_out and lost_out, correspondingly.
1175 *
1176 * Valid combinations are:
1177 * Tag InFlight Description
1178 * 0 1 - orig segment is in flight.
1179 * S 0 - nothing flies, orig reached receiver.
1180 * L 0 - nothing flies, orig lost by net.
1181 * R 2 - both orig and retransmit are in flight.
1182 * L|R 1 - orig is lost, retransmit is in flight.
1183 * S|R 1 - orig reached receiver, retrans is still in flight.
1184 * (L|S|R is logically valid, it could occur when L|R is sacked,
1185 * but it is equivalent to plain S and code short-circuits it to S.
1186 * L|S is logically invalid, it would mean -1 packet in flight 8))
1187 *
1188 * These 6 states form finite state machine, controlled by the following events:
1189 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1190 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1191 * 3. Loss detection event of two flavors:
1192 * A. Scoreboard estimator decided the packet is lost.
1193 * A'. Reno "three dupacks" marks head of queue lost.
1194 * B. SACK arrives sacking SND.NXT at the moment, when the
1195 * segment was retransmitted.
1196 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1197 *
1198 * It is pleasant to note, that state diagram turns out to be commutative,
1199 * so that we are allowed not to be bothered by order of our actions,
1200 * when multiple events arrive simultaneously. (see the function below).
1201 *
1202 * Reordering detection.
1203 * --------------------
1204 * Reordering metric is maximal distance, which a packet can be displaced
1205 * in packet stream. With SACKs we can estimate it:
1206 *
1207 * 1. SACK fills old hole and the corresponding segment was not
1208 * ever retransmitted -> reordering. Alas, we cannot use it
1209 * when segment was retransmitted.
1210 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1211 * for retransmitted and already SACKed segment -> reordering..
1212 * Both of these heuristics are not used in Loss state, when we cannot
1213 * account for retransmits accurately.
1214 *
1215 * SACK block validation.
1216 * ----------------------
1217 *
1218 * SACK block range validation checks that the received SACK block fits to
1219 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1220 * Note that SND.UNA is not included to the range though being valid because
1221 * it means that the receiver is rather inconsistent with itself reporting
1222 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1223 * perfectly valid, however, in light of RFC2018 which explicitly states
1224 * that "SACK block MUST reflect the newest segment. Even if the newest
1225 * segment is going to be discarded ...", not that it looks very clever
1226 * in case of head skb. Due to potentional receiver driven attacks, we
1227 * choose to avoid immediate execution of a walk in write queue due to
1228 * reneging and defer head skb's loss recovery to standard loss recovery
1229 * procedure that will eventually trigger (nothing forbids us doing this).
1230 *
1231 * Implements also blockage to start_seq wrap-around. Problem lies in the
1232 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1233 * there's no guarantee that it will be before snd_nxt (n). The problem
1234 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1235 * wrap (s_w):
1236 *
1237 * <- outs wnd -> <- wrapzone ->
1238 * u e n u_w e_w s n_w
1239 * | | | | | | |
1240 * |<------------+------+----- TCP seqno space --------------+---------->|
1241 * ...-- <2^31 ->| |<--------...
1242 * ...---- >2^31 ------>| |<--------...
1243 *
1244 * Current code wouldn't be vulnerable but it's better still to discard such
1245 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1246 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1247 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1248 * equal to the ideal case (infinite seqno space without wrap caused issues).
1249 *
1250 * With D-SACK the lower bound is extended to cover sequence space below
1251 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1252 * again, D-SACK block must not to go across snd_una (for the same reason as
1253 * for the normal SACK blocks, explained above). But there all simplicity
1254 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1255 * fully below undo_marker they do not affect behavior in anyway and can
1256 * therefore be safely ignored. In rare cases (which are more or less
1257 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1258 * fragmentation and packet reordering past skb's retransmission. To consider
1259 * them correctly, the acceptable range must be extended even more though
1260 * the exact amount is rather hard to quantify. However, tp->max_window can
1261 * be used as an exaggerated estimate.
1262 */
tcp_is_sackblock_valid(struct tcp_sock * tp,bool is_dsack,u32 start_seq,u32 end_seq)1263 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1264 u32 start_seq, u32 end_seq)
1265 {
1266 /* Too far in future, or reversed (interpretation is ambiguous) */
1267 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1268 return false;
1269
1270 /* Nasty start_seq wrap-around check (see comments above) */
1271 if (!before(start_seq, tp->snd_nxt))
1272 return false;
1273
1274 /* In outstanding window? ...This is valid exit for D-SACKs too.
1275 * start_seq == snd_una is non-sensical (see comments above)
1276 */
1277 if (after(start_seq, tp->snd_una))
1278 return true;
1279
1280 if (!is_dsack || !tp->undo_marker)
1281 return false;
1282
1283 /* ...Then it's D-SACK, and must reside below snd_una completely */
1284 if (after(end_seq, tp->snd_una))
1285 return false;
1286
1287 if (!before(start_seq, tp->undo_marker))
1288 return true;
1289
1290 /* Too old */
1291 if (!after(end_seq, tp->undo_marker))
1292 return false;
1293
1294 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1295 * start_seq < undo_marker and end_seq >= undo_marker.
1296 */
1297 return !before(start_seq, end_seq - tp->max_window);
1298 }
1299
tcp_check_dsack(struct sock * sk,const struct sk_buff * ack_skb,struct tcp_sack_block_wire * sp,int num_sacks,u32 prior_snd_una,struct tcp_sacktag_state * state)1300 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1301 struct tcp_sack_block_wire *sp, int num_sacks,
1302 u32 prior_snd_una, struct tcp_sacktag_state *state)
1303 {
1304 struct tcp_sock *tp = tcp_sk(sk);
1305 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1306 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1307 u32 dup_segs;
1308
1309 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1310 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1311 } else if (num_sacks > 1) {
1312 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1313 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1314
1315 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1316 return false;
1317 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1318 } else {
1319 return false;
1320 }
1321
1322 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1323 if (!dup_segs) { /* Skip dubious DSACK */
1324 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1325 return false;
1326 }
1327
1328 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1329
1330 /* D-SACK for already forgotten data... Do dumb counting. */
1331 if (tp->undo_marker && tp->undo_retrans > 0 &&
1332 !after(end_seq_0, prior_snd_una) &&
1333 after(end_seq_0, tp->undo_marker))
1334 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1335
1336 return true;
1337 }
1338
1339 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1340 * the incoming SACK may not exactly match but we can find smaller MSS
1341 * aligned portion of it that matches. Therefore we might need to fragment
1342 * which may fail and creates some hassle (caller must handle error case
1343 * returns).
1344 *
1345 * FIXME: this could be merged to shift decision code
1346 */
tcp_match_skb_to_sack(struct sock * sk,struct sk_buff * skb,u32 start_seq,u32 end_seq)1347 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1348 u32 start_seq, u32 end_seq)
1349 {
1350 int err;
1351 bool in_sack;
1352 unsigned int pkt_len;
1353 unsigned int mss;
1354
1355 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1356 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1357
1358 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1359 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1360 mss = tcp_skb_mss(skb);
1361 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1362
1363 if (!in_sack) {
1364 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1365 if (pkt_len < mss)
1366 pkt_len = mss;
1367 } else {
1368 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1369 if (pkt_len < mss)
1370 return -EINVAL;
1371 }
1372
1373 /* Round if necessary so that SACKs cover only full MSSes
1374 * and/or the remaining small portion (if present)
1375 */
1376 if (pkt_len > mss) {
1377 unsigned int new_len = (pkt_len / mss) * mss;
1378 if (!in_sack && new_len < pkt_len)
1379 new_len += mss;
1380 pkt_len = new_len;
1381 }
1382
1383 if (pkt_len >= skb->len && !in_sack)
1384 return 0;
1385
1386 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1387 pkt_len, mss, GFP_ATOMIC);
1388 if (err < 0)
1389 return err;
1390 }
1391
1392 return in_sack;
1393 }
1394
1395 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
tcp_sacktag_one(struct sock * sk,struct tcp_sacktag_state * state,u8 sacked,u32 start_seq,u32 end_seq,int dup_sack,int pcount,u64 xmit_time)1396 static u8 tcp_sacktag_one(struct sock *sk,
1397 struct tcp_sacktag_state *state, u8 sacked,
1398 u32 start_seq, u32 end_seq,
1399 int dup_sack, int pcount,
1400 u64 xmit_time)
1401 {
1402 struct tcp_sock *tp = tcp_sk(sk);
1403
1404 /* Account D-SACK for retransmitted packet. */
1405 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1406 if (tp->undo_marker && tp->undo_retrans > 0 &&
1407 after(end_seq, tp->undo_marker))
1408 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1409 if ((sacked & TCPCB_SACKED_ACKED) &&
1410 before(start_seq, state->reord))
1411 state->reord = start_seq;
1412 }
1413
1414 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1415 if (!after(end_seq, tp->snd_una))
1416 return sacked;
1417
1418 if (!(sacked & TCPCB_SACKED_ACKED)) {
1419 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1420
1421 if (sacked & TCPCB_SACKED_RETRANS) {
1422 /* If the segment is not tagged as lost,
1423 * we do not clear RETRANS, believing
1424 * that retransmission is still in flight.
1425 */
1426 if (sacked & TCPCB_LOST) {
1427 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1428 tp->lost_out -= pcount;
1429 tp->retrans_out -= pcount;
1430 }
1431 } else {
1432 if (!(sacked & TCPCB_RETRANS)) {
1433 /* New sack for not retransmitted frame,
1434 * which was in hole. It is reordering.
1435 */
1436 if (before(start_seq,
1437 tcp_highest_sack_seq(tp)) &&
1438 before(start_seq, state->reord))
1439 state->reord = start_seq;
1440
1441 if (!after(end_seq, tp->high_seq))
1442 state->flag |= FLAG_ORIG_SACK_ACKED;
1443 if (state->first_sackt == 0)
1444 state->first_sackt = xmit_time;
1445 state->last_sackt = xmit_time;
1446 }
1447
1448 if (sacked & TCPCB_LOST) {
1449 sacked &= ~TCPCB_LOST;
1450 tp->lost_out -= pcount;
1451 }
1452 }
1453
1454 sacked |= TCPCB_SACKED_ACKED;
1455 state->flag |= FLAG_DATA_SACKED;
1456 tp->sacked_out += pcount;
1457 /* Out-of-order packets delivered */
1458 state->sack_delivered += pcount;
1459
1460 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1461 if (tp->lost_skb_hint &&
1462 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1463 tp->lost_cnt_hint += pcount;
1464 }
1465
1466 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1467 * frames and clear it. undo_retrans is decreased above, L|R frames
1468 * are accounted above as well.
1469 */
1470 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1471 sacked &= ~TCPCB_SACKED_RETRANS;
1472 tp->retrans_out -= pcount;
1473 }
1474
1475 return sacked;
1476 }
1477
1478 /* Shift newly-SACKed bytes from this skb to the immediately previous
1479 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1480 */
tcp_shifted_skb(struct sock * sk,struct sk_buff * prev,struct sk_buff * skb,struct tcp_sacktag_state * state,unsigned int pcount,int shifted,int mss,bool dup_sack)1481 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1482 struct sk_buff *skb,
1483 struct tcp_sacktag_state *state,
1484 unsigned int pcount, int shifted, int mss,
1485 bool dup_sack)
1486 {
1487 struct tcp_sock *tp = tcp_sk(sk);
1488 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1489 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1490
1491 BUG_ON(!pcount);
1492
1493 /* Adjust counters and hints for the newly sacked sequence
1494 * range but discard the return value since prev is already
1495 * marked. We must tag the range first because the seq
1496 * advancement below implicitly advances
1497 * tcp_highest_sack_seq() when skb is highest_sack.
1498 */
1499 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1500 start_seq, end_seq, dup_sack, pcount,
1501 tcp_skb_timestamp_us(skb));
1502 tcp_rate_skb_delivered(sk, skb, state->rate);
1503
1504 if (skb == tp->lost_skb_hint)
1505 tp->lost_cnt_hint += pcount;
1506
1507 TCP_SKB_CB(prev)->end_seq += shifted;
1508 TCP_SKB_CB(skb)->seq += shifted;
1509
1510 tcp_skb_pcount_add(prev, pcount);
1511 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1512 tcp_skb_pcount_add(skb, -pcount);
1513
1514 /* When we're adding to gso_segs == 1, gso_size will be zero,
1515 * in theory this shouldn't be necessary but as long as DSACK
1516 * code can come after this skb later on it's better to keep
1517 * setting gso_size to something.
1518 */
1519 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1520 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1521
1522 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1523 if (tcp_skb_pcount(skb) <= 1)
1524 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1525
1526 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1527 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1528
1529 if (skb->len > 0) {
1530 BUG_ON(!tcp_skb_pcount(skb));
1531 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1532 return false;
1533 }
1534
1535 /* Whole SKB was eaten :-) */
1536
1537 if (skb == tp->retransmit_skb_hint)
1538 tp->retransmit_skb_hint = prev;
1539 if (skb == tp->lost_skb_hint) {
1540 tp->lost_skb_hint = prev;
1541 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1542 }
1543
1544 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1545 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1546 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1547 TCP_SKB_CB(prev)->end_seq++;
1548
1549 if (skb == tcp_highest_sack(sk))
1550 tcp_advance_highest_sack(sk, skb);
1551
1552 tcp_skb_collapse_tstamp(prev, skb);
1553 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1554 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1555
1556 tcp_rtx_queue_unlink_and_free(skb, sk);
1557
1558 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1559
1560 return true;
1561 }
1562
1563 /* I wish gso_size would have a bit more sane initialization than
1564 * something-or-zero which complicates things
1565 */
tcp_skb_seglen(const struct sk_buff * skb)1566 static int tcp_skb_seglen(const struct sk_buff *skb)
1567 {
1568 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1569 }
1570
1571 /* Shifting pages past head area doesn't work */
skb_can_shift(const struct sk_buff * skb)1572 static int skb_can_shift(const struct sk_buff *skb)
1573 {
1574 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1575 }
1576
tcp_skb_shift(struct sk_buff * to,struct sk_buff * from,int pcount,int shiftlen)1577 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1578 int pcount, int shiftlen)
1579 {
1580 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1581 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1582 * to make sure not storing more than 65535 * 8 bytes per skb,
1583 * even if current MSS is bigger.
1584 */
1585 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1586 return 0;
1587 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1588 return 0;
1589 return skb_shift(to, from, shiftlen);
1590 }
1591
1592 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1593 * skb.
1594 */
tcp_shift_skb_data(struct sock * sk,struct sk_buff * skb,struct tcp_sacktag_state * state,u32 start_seq,u32 end_seq,bool dup_sack)1595 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1596 struct tcp_sacktag_state *state,
1597 u32 start_seq, u32 end_seq,
1598 bool dup_sack)
1599 {
1600 struct tcp_sock *tp = tcp_sk(sk);
1601 struct sk_buff *prev;
1602 int mss;
1603 int pcount = 0;
1604 int len;
1605 int in_sack;
1606
1607 /* Normally R but no L won't result in plain S */
1608 if (!dup_sack &&
1609 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1610 goto fallback;
1611 if (!skb_can_shift(skb))
1612 goto fallback;
1613 /* This frame is about to be dropped (was ACKed). */
1614 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1615 goto fallback;
1616
1617 /* Can only happen with delayed DSACK + discard craziness */
1618 prev = skb_rb_prev(skb);
1619 if (!prev)
1620 goto fallback;
1621
1622 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1623 goto fallback;
1624
1625 if (!tcp_skb_can_collapse(prev, skb))
1626 goto fallback;
1627
1628 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1629 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1630
1631 if (in_sack) {
1632 len = skb->len;
1633 pcount = tcp_skb_pcount(skb);
1634 mss = tcp_skb_seglen(skb);
1635
1636 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1637 * drop this restriction as unnecessary
1638 */
1639 if (mss != tcp_skb_seglen(prev))
1640 goto fallback;
1641 } else {
1642 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1643 goto noop;
1644 /* CHECKME: This is non-MSS split case only?, this will
1645 * cause skipped skbs due to advancing loop btw, original
1646 * has that feature too
1647 */
1648 if (tcp_skb_pcount(skb) <= 1)
1649 goto noop;
1650
1651 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1652 if (!in_sack) {
1653 /* TODO: head merge to next could be attempted here
1654 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1655 * though it might not be worth of the additional hassle
1656 *
1657 * ...we can probably just fallback to what was done
1658 * previously. We could try merging non-SACKed ones
1659 * as well but it probably isn't going to buy off
1660 * because later SACKs might again split them, and
1661 * it would make skb timestamp tracking considerably
1662 * harder problem.
1663 */
1664 goto fallback;
1665 }
1666
1667 len = end_seq - TCP_SKB_CB(skb)->seq;
1668 BUG_ON(len < 0);
1669 BUG_ON(len > skb->len);
1670
1671 /* MSS boundaries should be honoured or else pcount will
1672 * severely break even though it makes things bit trickier.
1673 * Optimize common case to avoid most of the divides
1674 */
1675 mss = tcp_skb_mss(skb);
1676
1677 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1678 * drop this restriction as unnecessary
1679 */
1680 if (mss != tcp_skb_seglen(prev))
1681 goto fallback;
1682
1683 if (len == mss) {
1684 pcount = 1;
1685 } else if (len < mss) {
1686 goto noop;
1687 } else {
1688 pcount = len / mss;
1689 len = pcount * mss;
1690 }
1691 }
1692
1693 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1694 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1695 goto fallback;
1696
1697 if (!tcp_skb_shift(prev, skb, pcount, len))
1698 goto fallback;
1699 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1700 goto out;
1701
1702 /* Hole filled allows collapsing with the next as well, this is very
1703 * useful when hole on every nth skb pattern happens
1704 */
1705 skb = skb_rb_next(prev);
1706 if (!skb)
1707 goto out;
1708
1709 if (!skb_can_shift(skb) ||
1710 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1711 (mss != tcp_skb_seglen(skb)))
1712 goto out;
1713
1714 if (!tcp_skb_can_collapse(prev, skb))
1715 goto out;
1716 len = skb->len;
1717 pcount = tcp_skb_pcount(skb);
1718 if (tcp_skb_shift(prev, skb, pcount, len))
1719 tcp_shifted_skb(sk, prev, skb, state, pcount,
1720 len, mss, 0);
1721
1722 out:
1723 return prev;
1724
1725 noop:
1726 return skb;
1727
1728 fallback:
1729 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1730 return NULL;
1731 }
1732
tcp_sacktag_walk(struct sk_buff * skb,struct sock * sk,struct tcp_sack_block * next_dup,struct tcp_sacktag_state * state,u32 start_seq,u32 end_seq,bool dup_sack_in)1733 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1734 struct tcp_sack_block *next_dup,
1735 struct tcp_sacktag_state *state,
1736 u32 start_seq, u32 end_seq,
1737 bool dup_sack_in)
1738 {
1739 struct tcp_sock *tp = tcp_sk(sk);
1740 struct sk_buff *tmp;
1741
1742 skb_rbtree_walk_from(skb) {
1743 int in_sack = 0;
1744 bool dup_sack = dup_sack_in;
1745
1746 /* queue is in-order => we can short-circuit the walk early */
1747 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1748 break;
1749
1750 if (next_dup &&
1751 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1752 in_sack = tcp_match_skb_to_sack(sk, skb,
1753 next_dup->start_seq,
1754 next_dup->end_seq);
1755 if (in_sack > 0)
1756 dup_sack = true;
1757 }
1758
1759 /* skb reference here is a bit tricky to get right, since
1760 * shifting can eat and free both this skb and the next,
1761 * so not even _safe variant of the loop is enough.
1762 */
1763 if (in_sack <= 0) {
1764 tmp = tcp_shift_skb_data(sk, skb, state,
1765 start_seq, end_seq, dup_sack);
1766 if (tmp) {
1767 if (tmp != skb) {
1768 skb = tmp;
1769 continue;
1770 }
1771
1772 in_sack = 0;
1773 } else {
1774 in_sack = tcp_match_skb_to_sack(sk, skb,
1775 start_seq,
1776 end_seq);
1777 }
1778 }
1779
1780 if (unlikely(in_sack < 0))
1781 break;
1782
1783 if (in_sack) {
1784 TCP_SKB_CB(skb)->sacked =
1785 tcp_sacktag_one(sk,
1786 state,
1787 TCP_SKB_CB(skb)->sacked,
1788 TCP_SKB_CB(skb)->seq,
1789 TCP_SKB_CB(skb)->end_seq,
1790 dup_sack,
1791 tcp_skb_pcount(skb),
1792 tcp_skb_timestamp_us(skb));
1793 tcp_rate_skb_delivered(sk, skb, state->rate);
1794 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1795 list_del_init(&skb->tcp_tsorted_anchor);
1796
1797 if (!before(TCP_SKB_CB(skb)->seq,
1798 tcp_highest_sack_seq(tp)))
1799 tcp_advance_highest_sack(sk, skb);
1800 }
1801 }
1802 return skb;
1803 }
1804
tcp_sacktag_bsearch(struct sock * sk,u32 seq)1805 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1806 {
1807 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1808 struct sk_buff *skb;
1809
1810 while (*p) {
1811 parent = *p;
1812 skb = rb_to_skb(parent);
1813 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1814 p = &parent->rb_left;
1815 continue;
1816 }
1817 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1818 p = &parent->rb_right;
1819 continue;
1820 }
1821 return skb;
1822 }
1823 return NULL;
1824 }
1825
tcp_sacktag_skip(struct sk_buff * skb,struct sock * sk,u32 skip_to_seq)1826 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1827 u32 skip_to_seq)
1828 {
1829 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1830 return skb;
1831
1832 return tcp_sacktag_bsearch(sk, skip_to_seq);
1833 }
1834
tcp_maybe_skipping_dsack(struct sk_buff * skb,struct sock * sk,struct tcp_sack_block * next_dup,struct tcp_sacktag_state * state,u32 skip_to_seq)1835 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1836 struct sock *sk,
1837 struct tcp_sack_block *next_dup,
1838 struct tcp_sacktag_state *state,
1839 u32 skip_to_seq)
1840 {
1841 if (!next_dup)
1842 return skb;
1843
1844 if (before(next_dup->start_seq, skip_to_seq)) {
1845 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1846 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1847 next_dup->start_seq, next_dup->end_seq,
1848 1);
1849 }
1850
1851 return skb;
1852 }
1853
tcp_sack_cache_ok(const struct tcp_sock * tp,const struct tcp_sack_block * cache)1854 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1855 {
1856 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1857 }
1858
1859 static int
tcp_sacktag_write_queue(struct sock * sk,const struct sk_buff * ack_skb,u32 prior_snd_una,struct tcp_sacktag_state * state)1860 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1861 u32 prior_snd_una, struct tcp_sacktag_state *state)
1862 {
1863 struct tcp_sock *tp = tcp_sk(sk);
1864 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1865 TCP_SKB_CB(ack_skb)->sacked);
1866 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1867 struct tcp_sack_block sp[TCP_NUM_SACKS];
1868 struct tcp_sack_block *cache;
1869 struct sk_buff *skb;
1870 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1871 int used_sacks;
1872 bool found_dup_sack = false;
1873 int i, j;
1874 int first_sack_index;
1875
1876 state->flag = 0;
1877 state->reord = tp->snd_nxt;
1878
1879 if (!tp->sacked_out)
1880 tcp_highest_sack_reset(sk);
1881
1882 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1883 num_sacks, prior_snd_una, state);
1884
1885 /* Eliminate too old ACKs, but take into
1886 * account more or less fresh ones, they can
1887 * contain valid SACK info.
1888 */
1889 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1890 return 0;
1891
1892 if (!tp->packets_out)
1893 goto out;
1894
1895 used_sacks = 0;
1896 first_sack_index = 0;
1897 for (i = 0; i < num_sacks; i++) {
1898 bool dup_sack = !i && found_dup_sack;
1899
1900 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1901 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1902
1903 if (!tcp_is_sackblock_valid(tp, dup_sack,
1904 sp[used_sacks].start_seq,
1905 sp[used_sacks].end_seq)) {
1906 int mib_idx;
1907
1908 if (dup_sack) {
1909 if (!tp->undo_marker)
1910 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1911 else
1912 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1913 } else {
1914 /* Don't count olds caused by ACK reordering */
1915 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1916 !after(sp[used_sacks].end_seq, tp->snd_una))
1917 continue;
1918 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1919 }
1920
1921 NET_INC_STATS(sock_net(sk), mib_idx);
1922 if (i == 0)
1923 first_sack_index = -1;
1924 continue;
1925 }
1926
1927 /* Ignore very old stuff early */
1928 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1929 if (i == 0)
1930 first_sack_index = -1;
1931 continue;
1932 }
1933
1934 used_sacks++;
1935 }
1936
1937 /* order SACK blocks to allow in order walk of the retrans queue */
1938 for (i = used_sacks - 1; i > 0; i--) {
1939 for (j = 0; j < i; j++) {
1940 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1941 swap(sp[j], sp[j + 1]);
1942
1943 /* Track where the first SACK block goes to */
1944 if (j == first_sack_index)
1945 first_sack_index = j + 1;
1946 }
1947 }
1948 }
1949
1950 state->mss_now = tcp_current_mss(sk);
1951 skb = NULL;
1952 i = 0;
1953
1954 if (!tp->sacked_out) {
1955 /* It's already past, so skip checking against it */
1956 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1957 } else {
1958 cache = tp->recv_sack_cache;
1959 /* Skip empty blocks in at head of the cache */
1960 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1961 !cache->end_seq)
1962 cache++;
1963 }
1964
1965 while (i < used_sacks) {
1966 u32 start_seq = sp[i].start_seq;
1967 u32 end_seq = sp[i].end_seq;
1968 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1969 struct tcp_sack_block *next_dup = NULL;
1970
1971 if (found_dup_sack && ((i + 1) == first_sack_index))
1972 next_dup = &sp[i + 1];
1973
1974 /* Skip too early cached blocks */
1975 while (tcp_sack_cache_ok(tp, cache) &&
1976 !before(start_seq, cache->end_seq))
1977 cache++;
1978
1979 /* Can skip some work by looking recv_sack_cache? */
1980 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1981 after(end_seq, cache->start_seq)) {
1982
1983 /* Head todo? */
1984 if (before(start_seq, cache->start_seq)) {
1985 skb = tcp_sacktag_skip(skb, sk, start_seq);
1986 skb = tcp_sacktag_walk(skb, sk, next_dup,
1987 state,
1988 start_seq,
1989 cache->start_seq,
1990 dup_sack);
1991 }
1992
1993 /* Rest of the block already fully processed? */
1994 if (!after(end_seq, cache->end_seq))
1995 goto advance_sp;
1996
1997 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1998 state,
1999 cache->end_seq);
2000
2001 /* ...tail remains todo... */
2002 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
2003 /* ...but better entrypoint exists! */
2004 skb = tcp_highest_sack(sk);
2005 if (!skb)
2006 break;
2007 cache++;
2008 goto walk;
2009 }
2010
2011 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
2012 /* Check overlap against next cached too (past this one already) */
2013 cache++;
2014 continue;
2015 }
2016
2017 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
2018 skb = tcp_highest_sack(sk);
2019 if (!skb)
2020 break;
2021 }
2022 skb = tcp_sacktag_skip(skb, sk, start_seq);
2023
2024 walk:
2025 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
2026 start_seq, end_seq, dup_sack);
2027
2028 advance_sp:
2029 i++;
2030 }
2031
2032 /* Clear the head of the cache sack blocks so we can skip it next time */
2033 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
2034 tp->recv_sack_cache[i].start_seq = 0;
2035 tp->recv_sack_cache[i].end_seq = 0;
2036 }
2037 for (j = 0; j < used_sacks; j++)
2038 tp->recv_sack_cache[i++] = sp[j];
2039
2040 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
2041 tcp_check_sack_reordering(sk, state->reord, 0);
2042
2043 tcp_verify_left_out(tp);
2044 out:
2045
2046 #if FASTRETRANS_DEBUG > 0
2047 WARN_ON((int)tp->sacked_out < 0);
2048 WARN_ON((int)tp->lost_out < 0);
2049 WARN_ON((int)tp->retrans_out < 0);
2050 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
2051 #endif
2052 return state->flag;
2053 }
2054
2055 /* Limits sacked_out so that sum with lost_out isn't ever larger than
2056 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
2057 */
tcp_limit_reno_sacked(struct tcp_sock * tp)2058 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
2059 {
2060 u32 holes;
2061
2062 holes = max(tp->lost_out, 1U);
2063 holes = min(holes, tp->packets_out);
2064
2065 if ((tp->sacked_out + holes) > tp->packets_out) {
2066 tp->sacked_out = tp->packets_out - holes;
2067 return true;
2068 }
2069 return false;
2070 }
2071
2072 /* If we receive more dupacks than we expected counting segments
2073 * in assumption of absent reordering, interpret this as reordering.
2074 * The only another reason could be bug in receiver TCP.
2075 */
tcp_check_reno_reordering(struct sock * sk,const int addend)2076 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2077 {
2078 struct tcp_sock *tp = tcp_sk(sk);
2079
2080 if (!tcp_limit_reno_sacked(tp))
2081 return;
2082
2083 tp->reordering = min_t(u32, tp->packets_out + addend,
2084 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
2085 tp->reord_seen++;
2086 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
2087 }
2088
2089 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2090
tcp_add_reno_sack(struct sock * sk,int num_dupack,bool ece_ack)2091 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2092 {
2093 if (num_dupack) {
2094 struct tcp_sock *tp = tcp_sk(sk);
2095 u32 prior_sacked = tp->sacked_out;
2096 s32 delivered;
2097
2098 tp->sacked_out += num_dupack;
2099 tcp_check_reno_reordering(sk, 0);
2100 delivered = tp->sacked_out - prior_sacked;
2101 if (delivered > 0)
2102 tcp_count_delivered(tp, delivered, ece_ack);
2103 tcp_verify_left_out(tp);
2104 }
2105 }
2106
2107 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2108
tcp_remove_reno_sacks(struct sock * sk,int acked,bool ece_ack)2109 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2110 {
2111 struct tcp_sock *tp = tcp_sk(sk);
2112
2113 if (acked > 0) {
2114 /* One ACK acked hole. The rest eat duplicate ACKs. */
2115 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2116 ece_ack);
2117 if (acked - 1 >= tp->sacked_out)
2118 tp->sacked_out = 0;
2119 else
2120 tp->sacked_out -= acked - 1;
2121 }
2122 tcp_check_reno_reordering(sk, acked);
2123 tcp_verify_left_out(tp);
2124 }
2125
tcp_reset_reno_sack(struct tcp_sock * tp)2126 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2127 {
2128 tp->sacked_out = 0;
2129 }
2130
tcp_clear_retrans(struct tcp_sock * tp)2131 void tcp_clear_retrans(struct tcp_sock *tp)
2132 {
2133 tp->retrans_out = 0;
2134 tp->lost_out = 0;
2135 tp->undo_marker = 0;
2136 tp->undo_retrans = -1;
2137 tp->sacked_out = 0;
2138 tp->rto_stamp = 0;
2139 tp->total_rto = 0;
2140 tp->total_rto_recoveries = 0;
2141 tp->total_rto_time = 0;
2142 }
2143
tcp_init_undo(struct tcp_sock * tp)2144 static inline void tcp_init_undo(struct tcp_sock *tp)
2145 {
2146 tp->undo_marker = tp->snd_una;
2147
2148 /* Retransmission still in flight may cause DSACKs later. */
2149 /* First, account for regular retransmits in flight: */
2150 tp->undo_retrans = tp->retrans_out;
2151 /* Next, account for TLP retransmits in flight: */
2152 if (tp->tlp_high_seq && tp->tlp_retrans)
2153 tp->undo_retrans++;
2154 /* Finally, avoid 0, because undo_retrans==0 means "can undo now": */
2155 if (!tp->undo_retrans)
2156 tp->undo_retrans = -1;
2157 }
2158
tcp_is_rack(const struct sock * sk)2159 static bool tcp_is_rack(const struct sock *sk)
2160 {
2161 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
2162 TCP_RACK_LOSS_DETECTION;
2163 }
2164
2165 /* If we detect SACK reneging, forget all SACK information
2166 * and reset tags completely, otherwise preserve SACKs. If receiver
2167 * dropped its ofo queue, we will know this due to reneging detection.
2168 */
tcp_timeout_mark_lost(struct sock * sk)2169 static void tcp_timeout_mark_lost(struct sock *sk)
2170 {
2171 struct tcp_sock *tp = tcp_sk(sk);
2172 struct sk_buff *skb, *head;
2173 bool is_reneg; /* is receiver reneging on SACKs? */
2174
2175 head = tcp_rtx_queue_head(sk);
2176 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2177 if (is_reneg) {
2178 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2179 tp->sacked_out = 0;
2180 /* Mark SACK reneging until we recover from this loss event. */
2181 tp->is_sack_reneg = 1;
2182 } else if (tcp_is_reno(tp)) {
2183 tcp_reset_reno_sack(tp);
2184 }
2185
2186 skb = head;
2187 skb_rbtree_walk_from(skb) {
2188 if (is_reneg)
2189 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2190 else if (tcp_is_rack(sk) && skb != head &&
2191 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2192 continue; /* Don't mark recently sent ones lost yet */
2193 tcp_mark_skb_lost(sk, skb);
2194 }
2195 tcp_verify_left_out(tp);
2196 tcp_clear_all_retrans_hints(tp);
2197 }
2198
2199 /* Enter Loss state. */
tcp_enter_loss(struct sock * sk)2200 void tcp_enter_loss(struct sock *sk)
2201 {
2202 const struct inet_connection_sock *icsk = inet_csk(sk);
2203 struct tcp_sock *tp = tcp_sk(sk);
2204 struct net *net = sock_net(sk);
2205 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2206 u8 reordering;
2207
2208 tcp_timeout_mark_lost(sk);
2209
2210 /* Reduce ssthresh if it has not yet been made inside this window. */
2211 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2212 !after(tp->high_seq, tp->snd_una) ||
2213 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2214 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2215 tp->prior_cwnd = tcp_snd_cwnd(tp);
2216 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2217 tcp_ca_event(sk, CA_EVENT_LOSS);
2218 tcp_init_undo(tp);
2219 }
2220 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1);
2221 tp->snd_cwnd_cnt = 0;
2222 tp->snd_cwnd_stamp = tcp_jiffies32;
2223
2224 /* Timeout in disordered state after receiving substantial DUPACKs
2225 * suggests that the degree of reordering is over-estimated.
2226 */
2227 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2228 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2229 tp->sacked_out >= reordering)
2230 tp->reordering = min_t(unsigned int, tp->reordering,
2231 reordering);
2232
2233 tcp_set_ca_state(sk, TCP_CA_Loss);
2234 tp->high_seq = tp->snd_nxt;
2235 tp->tlp_high_seq = 0;
2236 tcp_ecn_queue_cwr(tp);
2237
2238 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2239 * loss recovery is underway except recurring timeout(s) on
2240 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2241 */
2242 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2243 (new_recovery || icsk->icsk_retransmits) &&
2244 !inet_csk(sk)->icsk_mtup.probe_size;
2245 }
2246
2247 /* If ACK arrived pointing to a remembered SACK, it means that our
2248 * remembered SACKs do not reflect real state of receiver i.e.
2249 * receiver _host_ is heavily congested (or buggy).
2250 *
2251 * To avoid big spurious retransmission bursts due to transient SACK
2252 * scoreboard oddities that look like reneging, we give the receiver a
2253 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2254 * restore sanity to the SACK scoreboard. If the apparent reneging
2255 * persists until this RTO then we'll clear the SACK scoreboard.
2256 */
tcp_check_sack_reneging(struct sock * sk,int * ack_flag)2257 static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag)
2258 {
2259 if (*ack_flag & FLAG_SACK_RENEGING &&
2260 *ack_flag & FLAG_SND_UNA_ADVANCED) {
2261 struct tcp_sock *tp = tcp_sk(sk);
2262 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2263 msecs_to_jiffies(10));
2264
2265 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, delay, false);
2266 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2267 return true;
2268 }
2269 return false;
2270 }
2271
2272 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2273 * counter when SACK is enabled (without SACK, sacked_out is used for
2274 * that purpose).
2275 *
2276 * With reordering, holes may still be in flight, so RFC3517 recovery
2277 * uses pure sacked_out (total number of SACKed segments) even though
2278 * it violates the RFC that uses duplicate ACKs, often these are equal
2279 * but when e.g. out-of-window ACKs or packet duplication occurs,
2280 * they differ. Since neither occurs due to loss, TCP should really
2281 * ignore them.
2282 */
tcp_dupack_heuristics(const struct tcp_sock * tp)2283 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2284 {
2285 return tp->sacked_out + 1;
2286 }
2287
2288 /* Linux NewReno/SACK/ECN state machine.
2289 * --------------------------------------
2290 *
2291 * "Open" Normal state, no dubious events, fast path.
2292 * "Disorder" In all the respects it is "Open",
2293 * but requires a bit more attention. It is entered when
2294 * we see some SACKs or dupacks. It is split of "Open"
2295 * mainly to move some processing from fast path to slow one.
2296 * "CWR" CWND was reduced due to some Congestion Notification event.
2297 * It can be ECN, ICMP source quench, local device congestion.
2298 * "Recovery" CWND was reduced, we are fast-retransmitting.
2299 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2300 *
2301 * tcp_fastretrans_alert() is entered:
2302 * - each incoming ACK, if state is not "Open"
2303 * - when arrived ACK is unusual, namely:
2304 * * SACK
2305 * * Duplicate ACK.
2306 * * ECN ECE.
2307 *
2308 * Counting packets in flight is pretty simple.
2309 *
2310 * in_flight = packets_out - left_out + retrans_out
2311 *
2312 * packets_out is SND.NXT-SND.UNA counted in packets.
2313 *
2314 * retrans_out is number of retransmitted segments.
2315 *
2316 * left_out is number of segments left network, but not ACKed yet.
2317 *
2318 * left_out = sacked_out + lost_out
2319 *
2320 * sacked_out: Packets, which arrived to receiver out of order
2321 * and hence not ACKed. With SACKs this number is simply
2322 * amount of SACKed data. Even without SACKs
2323 * it is easy to give pretty reliable estimate of this number,
2324 * counting duplicate ACKs.
2325 *
2326 * lost_out: Packets lost by network. TCP has no explicit
2327 * "loss notification" feedback from network (for now).
2328 * It means that this number can be only _guessed_.
2329 * Actually, it is the heuristics to predict lossage that
2330 * distinguishes different algorithms.
2331 *
2332 * F.e. after RTO, when all the queue is considered as lost,
2333 * lost_out = packets_out and in_flight = retrans_out.
2334 *
2335 * Essentially, we have now a few algorithms detecting
2336 * lost packets.
2337 *
2338 * If the receiver supports SACK:
2339 *
2340 * RFC6675/3517: It is the conventional algorithm. A packet is
2341 * considered lost if the number of higher sequence packets
2342 * SACKed is greater than or equal the DUPACK thoreshold
2343 * (reordering). This is implemented in tcp_mark_head_lost and
2344 * tcp_update_scoreboard.
2345 *
2346 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2347 * (2017-) that checks timing instead of counting DUPACKs.
2348 * Essentially a packet is considered lost if it's not S/ACKed
2349 * after RTT + reordering_window, where both metrics are
2350 * dynamically measured and adjusted. This is implemented in
2351 * tcp_rack_mark_lost.
2352 *
2353 * If the receiver does not support SACK:
2354 *
2355 * NewReno (RFC6582): in Recovery we assume that one segment
2356 * is lost (classic Reno). While we are in Recovery and
2357 * a partial ACK arrives, we assume that one more packet
2358 * is lost (NewReno). This heuristics are the same in NewReno
2359 * and SACK.
2360 *
2361 * Really tricky (and requiring careful tuning) part of algorithm
2362 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2363 * The first determines the moment _when_ we should reduce CWND and,
2364 * hence, slow down forward transmission. In fact, it determines the moment
2365 * when we decide that hole is caused by loss, rather than by a reorder.
2366 *
2367 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2368 * holes, caused by lost packets.
2369 *
2370 * And the most logically complicated part of algorithm is undo
2371 * heuristics. We detect false retransmits due to both too early
2372 * fast retransmit (reordering) and underestimated RTO, analyzing
2373 * timestamps and D-SACKs. When we detect that some segments were
2374 * retransmitted by mistake and CWND reduction was wrong, we undo
2375 * window reduction and abort recovery phase. This logic is hidden
2376 * inside several functions named tcp_try_undo_<something>.
2377 */
2378
2379 /* This function decides, when we should leave Disordered state
2380 * and enter Recovery phase, reducing congestion window.
2381 *
2382 * Main question: may we further continue forward transmission
2383 * with the same cwnd?
2384 */
tcp_time_to_recover(struct sock * sk,int flag)2385 static bool tcp_time_to_recover(struct sock *sk, int flag)
2386 {
2387 struct tcp_sock *tp = tcp_sk(sk);
2388
2389 /* Trick#1: The loss is proven. */
2390 if (tp->lost_out)
2391 return true;
2392
2393 /* Not-A-Trick#2 : Classic rule... */
2394 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2395 return true;
2396
2397 return false;
2398 }
2399
2400 /* Detect loss in event "A" above by marking head of queue up as lost.
2401 * For RFC3517 SACK, a segment is considered lost if it
2402 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2403 * the maximum SACKed segments to pass before reaching this limit.
2404 */
tcp_mark_head_lost(struct sock * sk,int packets,int mark_head)2405 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2406 {
2407 struct tcp_sock *tp = tcp_sk(sk);
2408 struct sk_buff *skb;
2409 int cnt;
2410 /* Use SACK to deduce losses of new sequences sent during recovery */
2411 const u32 loss_high = tp->snd_nxt;
2412
2413 WARN_ON(packets > tp->packets_out);
2414 skb = tp->lost_skb_hint;
2415 if (skb) {
2416 /* Head already handled? */
2417 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2418 return;
2419 cnt = tp->lost_cnt_hint;
2420 } else {
2421 skb = tcp_rtx_queue_head(sk);
2422 cnt = 0;
2423 }
2424
2425 skb_rbtree_walk_from(skb) {
2426 /* TODO: do this better */
2427 /* this is not the most efficient way to do this... */
2428 tp->lost_skb_hint = skb;
2429 tp->lost_cnt_hint = cnt;
2430
2431 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2432 break;
2433
2434 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2435 cnt += tcp_skb_pcount(skb);
2436
2437 if (cnt > packets)
2438 break;
2439
2440 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2441 tcp_mark_skb_lost(sk, skb);
2442
2443 if (mark_head)
2444 break;
2445 }
2446 tcp_verify_left_out(tp);
2447 }
2448
2449 /* Account newly detected lost packet(s) */
2450
tcp_update_scoreboard(struct sock * sk,int fast_rexmit)2451 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2452 {
2453 struct tcp_sock *tp = tcp_sk(sk);
2454
2455 if (tcp_is_sack(tp)) {
2456 int sacked_upto = tp->sacked_out - tp->reordering;
2457 if (sacked_upto >= 0)
2458 tcp_mark_head_lost(sk, sacked_upto, 0);
2459 else if (fast_rexmit)
2460 tcp_mark_head_lost(sk, 1, 1);
2461 }
2462 }
2463
tcp_tsopt_ecr_before(const struct tcp_sock * tp,u32 when)2464 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2465 {
2466 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2467 before(tp->rx_opt.rcv_tsecr, when);
2468 }
2469
2470 /* skb is spurious retransmitted if the returned timestamp echo
2471 * reply is prior to the skb transmission time
2472 */
tcp_skb_spurious_retrans(const struct tcp_sock * tp,const struct sk_buff * skb)2473 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2474 const struct sk_buff *skb)
2475 {
2476 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2477 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb));
2478 }
2479
2480 /* Nothing was retransmitted or returned timestamp is less
2481 * than timestamp of the first retransmission.
2482 */
tcp_packet_delayed(const struct tcp_sock * tp)2483 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2484 {
2485 const struct sock *sk = (const struct sock *)tp;
2486
2487 if (tp->retrans_stamp &&
2488 tcp_tsopt_ecr_before(tp, tp->retrans_stamp))
2489 return true; /* got echoed TS before first retransmission */
2490
2491 /* Check if nothing was retransmitted (retrans_stamp==0), which may
2492 * happen in fast recovery due to TSQ. But we ignore zero retrans_stamp
2493 * in TCP_SYN_SENT, since when we set FLAG_SYN_ACKED we also clear
2494 * retrans_stamp even if we had retransmitted the SYN.
2495 */
2496 if (!tp->retrans_stamp && /* no record of a retransmit/SYN? */
2497 sk->sk_state != TCP_SYN_SENT) /* not the FLAG_SYN_ACKED case? */
2498 return true; /* nothing was retransmitted */
2499
2500 return false;
2501 }
2502
2503 /* Undo procedures. */
2504
2505 /* We can clear retrans_stamp when there are no retransmissions in the
2506 * window. It would seem that it is trivially available for us in
2507 * tp->retrans_out, however, that kind of assumptions doesn't consider
2508 * what will happen if errors occur when sending retransmission for the
2509 * second time. ...It could the that such segment has only
2510 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2511 * the head skb is enough except for some reneging corner cases that
2512 * are not worth the effort.
2513 *
2514 * Main reason for all this complexity is the fact that connection dying
2515 * time now depends on the validity of the retrans_stamp, in particular,
2516 * that successive retransmissions of a segment must not advance
2517 * retrans_stamp under any conditions.
2518 */
tcp_any_retrans_done(const struct sock * sk)2519 static bool tcp_any_retrans_done(const struct sock *sk)
2520 {
2521 const struct tcp_sock *tp = tcp_sk(sk);
2522 struct sk_buff *skb;
2523
2524 if (tp->retrans_out)
2525 return true;
2526
2527 skb = tcp_rtx_queue_head(sk);
2528 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2529 return true;
2530
2531 return false;
2532 }
2533
2534 /* If loss recovery is finished and there are no retransmits out in the
2535 * network, then we clear retrans_stamp so that upon the next loss recovery
2536 * retransmits_timed_out() and timestamp-undo are using the correct value.
2537 */
tcp_retrans_stamp_cleanup(struct sock * sk)2538 static void tcp_retrans_stamp_cleanup(struct sock *sk)
2539 {
2540 if (!tcp_any_retrans_done(sk))
2541 tcp_sk(sk)->retrans_stamp = 0;
2542 }
2543
DBGUNDO(struct sock * sk,const char * msg)2544 static void DBGUNDO(struct sock *sk, const char *msg)
2545 {
2546 #if FASTRETRANS_DEBUG > 1
2547 struct tcp_sock *tp = tcp_sk(sk);
2548 struct inet_sock *inet = inet_sk(sk);
2549
2550 if (sk->sk_family == AF_INET) {
2551 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2552 msg,
2553 &inet->inet_daddr, ntohs(inet->inet_dport),
2554 tcp_snd_cwnd(tp), tcp_left_out(tp),
2555 tp->snd_ssthresh, tp->prior_ssthresh,
2556 tp->packets_out);
2557 }
2558 #if IS_ENABLED(CONFIG_IPV6)
2559 else if (sk->sk_family == AF_INET6) {
2560 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2561 msg,
2562 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2563 tcp_snd_cwnd(tp), tcp_left_out(tp),
2564 tp->snd_ssthresh, tp->prior_ssthresh,
2565 tp->packets_out);
2566 }
2567 #endif
2568 #endif
2569 }
2570
tcp_undo_cwnd_reduction(struct sock * sk,bool unmark_loss)2571 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2572 {
2573 struct tcp_sock *tp = tcp_sk(sk);
2574
2575 if (unmark_loss) {
2576 struct sk_buff *skb;
2577
2578 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2579 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2580 }
2581 tp->lost_out = 0;
2582 tcp_clear_all_retrans_hints(tp);
2583 }
2584
2585 if (tp->prior_ssthresh) {
2586 const struct inet_connection_sock *icsk = inet_csk(sk);
2587
2588 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk));
2589
2590 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2591 tp->snd_ssthresh = tp->prior_ssthresh;
2592 tcp_ecn_withdraw_cwr(tp);
2593 }
2594 }
2595 tp->snd_cwnd_stamp = tcp_jiffies32;
2596 tp->undo_marker = 0;
2597 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2598 }
2599
tcp_may_undo(const struct tcp_sock * tp)2600 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2601 {
2602 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2603 }
2604
tcp_is_non_sack_preventing_reopen(struct sock * sk)2605 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2606 {
2607 struct tcp_sock *tp = tcp_sk(sk);
2608
2609 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2610 /* Hold old state until something *above* high_seq
2611 * is ACKed. For Reno it is MUST to prevent false
2612 * fast retransmits (RFC2582). SACK TCP is safe. */
2613 if (!tcp_any_retrans_done(sk))
2614 tp->retrans_stamp = 0;
2615 return true;
2616 }
2617 return false;
2618 }
2619
2620 /* People celebrate: "We love our President!" */
tcp_try_undo_recovery(struct sock * sk)2621 static bool tcp_try_undo_recovery(struct sock *sk)
2622 {
2623 struct tcp_sock *tp = tcp_sk(sk);
2624
2625 if (tcp_may_undo(tp)) {
2626 int mib_idx;
2627
2628 /* Happy end! We did not retransmit anything
2629 * or our original transmission succeeded.
2630 */
2631 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2632 tcp_undo_cwnd_reduction(sk, false);
2633 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2634 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2635 else
2636 mib_idx = LINUX_MIB_TCPFULLUNDO;
2637
2638 NET_INC_STATS(sock_net(sk), mib_idx);
2639 } else if (tp->rack.reo_wnd_persist) {
2640 tp->rack.reo_wnd_persist--;
2641 }
2642 if (tcp_is_non_sack_preventing_reopen(sk))
2643 return true;
2644 tcp_set_ca_state(sk, TCP_CA_Open);
2645 tp->is_sack_reneg = 0;
2646 return false;
2647 }
2648
2649 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
tcp_try_undo_dsack(struct sock * sk)2650 static bool tcp_try_undo_dsack(struct sock *sk)
2651 {
2652 struct tcp_sock *tp = tcp_sk(sk);
2653
2654 if (tp->undo_marker && !tp->undo_retrans) {
2655 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2656 tp->rack.reo_wnd_persist + 1);
2657 DBGUNDO(sk, "D-SACK");
2658 tcp_undo_cwnd_reduction(sk, false);
2659 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2660 return true;
2661 }
2662 return false;
2663 }
2664
2665 /* Undo during loss recovery after partial ACK or using F-RTO. */
tcp_try_undo_loss(struct sock * sk,bool frto_undo)2666 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2667 {
2668 struct tcp_sock *tp = tcp_sk(sk);
2669
2670 if (frto_undo || tcp_may_undo(tp)) {
2671 tcp_undo_cwnd_reduction(sk, true);
2672
2673 DBGUNDO(sk, "partial loss");
2674 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2675 if (frto_undo)
2676 NET_INC_STATS(sock_net(sk),
2677 LINUX_MIB_TCPSPURIOUSRTOS);
2678 inet_csk(sk)->icsk_retransmits = 0;
2679 if (tcp_is_non_sack_preventing_reopen(sk))
2680 return true;
2681 if (frto_undo || tcp_is_sack(tp)) {
2682 tcp_set_ca_state(sk, TCP_CA_Open);
2683 tp->is_sack_reneg = 0;
2684 }
2685 return true;
2686 }
2687 return false;
2688 }
2689
2690 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2691 * It computes the number of packets to send (sndcnt) based on packets newly
2692 * delivered:
2693 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2694 * cwnd reductions across a full RTT.
2695 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2696 * But when SND_UNA is acked without further losses,
2697 * slow starts cwnd up to ssthresh to speed up the recovery.
2698 */
tcp_init_cwnd_reduction(struct sock * sk)2699 static void tcp_init_cwnd_reduction(struct sock *sk)
2700 {
2701 struct tcp_sock *tp = tcp_sk(sk);
2702
2703 tp->high_seq = tp->snd_nxt;
2704 tp->tlp_high_seq = 0;
2705 tp->snd_cwnd_cnt = 0;
2706 tp->prior_cwnd = tcp_snd_cwnd(tp);
2707 tp->prr_delivered = 0;
2708 tp->prr_out = 0;
2709 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2710 tcp_ecn_queue_cwr(tp);
2711 }
2712
tcp_cwnd_reduction(struct sock * sk,int newly_acked_sacked,int newly_lost,int flag)2713 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag)
2714 {
2715 struct tcp_sock *tp = tcp_sk(sk);
2716 int sndcnt = 0;
2717 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2718
2719 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2720 return;
2721
2722 trace_tcp_cwnd_reduction_tp(sk, newly_acked_sacked, newly_lost, flag);
2723
2724 tp->prr_delivered += newly_acked_sacked;
2725 if (delta < 0) {
2726 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2727 tp->prior_cwnd - 1;
2728 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2729 } else {
2730 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out,
2731 newly_acked_sacked);
2732 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost)
2733 sndcnt++;
2734 sndcnt = min(delta, sndcnt);
2735 }
2736 /* Force a fast retransmit upon entering fast recovery */
2737 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2738 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt);
2739 }
2740
tcp_end_cwnd_reduction(struct sock * sk)2741 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2742 {
2743 struct tcp_sock *tp = tcp_sk(sk);
2744
2745 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2746 return;
2747
2748 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2749 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2750 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2751 tcp_snd_cwnd_set(tp, tp->snd_ssthresh);
2752 tp->snd_cwnd_stamp = tcp_jiffies32;
2753 }
2754 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2755 }
2756
2757 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
tcp_enter_cwr(struct sock * sk)2758 void tcp_enter_cwr(struct sock *sk)
2759 {
2760 struct tcp_sock *tp = tcp_sk(sk);
2761
2762 tp->prior_ssthresh = 0;
2763 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2764 tp->undo_marker = 0;
2765 tcp_init_cwnd_reduction(sk);
2766 tcp_set_ca_state(sk, TCP_CA_CWR);
2767 }
2768 }
2769 EXPORT_SYMBOL(tcp_enter_cwr);
2770
tcp_try_keep_open(struct sock * sk)2771 static void tcp_try_keep_open(struct sock *sk)
2772 {
2773 struct tcp_sock *tp = tcp_sk(sk);
2774 int state = TCP_CA_Open;
2775
2776 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2777 state = TCP_CA_Disorder;
2778
2779 if (inet_csk(sk)->icsk_ca_state != state) {
2780 tcp_set_ca_state(sk, state);
2781 tp->high_seq = tp->snd_nxt;
2782 }
2783 }
2784
tcp_try_to_open(struct sock * sk,int flag)2785 static void tcp_try_to_open(struct sock *sk, int flag)
2786 {
2787 struct tcp_sock *tp = tcp_sk(sk);
2788
2789 tcp_verify_left_out(tp);
2790
2791 if (!tcp_any_retrans_done(sk))
2792 tp->retrans_stamp = 0;
2793
2794 if (flag & FLAG_ECE)
2795 tcp_enter_cwr(sk);
2796
2797 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2798 tcp_try_keep_open(sk);
2799 }
2800 }
2801
tcp_mtup_probe_failed(struct sock * sk)2802 static void tcp_mtup_probe_failed(struct sock *sk)
2803 {
2804 struct inet_connection_sock *icsk = inet_csk(sk);
2805
2806 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2807 icsk->icsk_mtup.probe_size = 0;
2808 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2809 }
2810
tcp_mtup_probe_success(struct sock * sk)2811 static void tcp_mtup_probe_success(struct sock *sk)
2812 {
2813 struct tcp_sock *tp = tcp_sk(sk);
2814 struct inet_connection_sock *icsk = inet_csk(sk);
2815 u64 val;
2816
2817 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2818
2819 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache);
2820 do_div(val, icsk->icsk_mtup.probe_size);
2821 DEBUG_NET_WARN_ON_ONCE((u32)val != val);
2822 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val));
2823
2824 tp->snd_cwnd_cnt = 0;
2825 tp->snd_cwnd_stamp = tcp_jiffies32;
2826 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2827
2828 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2829 icsk->icsk_mtup.probe_size = 0;
2830 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2831 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2832 }
2833
2834 /* Sometimes we deduce that packets have been dropped due to reasons other than
2835 * congestion, like path MTU reductions or failed client TFO attempts. In these
2836 * cases we call this function to retransmit as many packets as cwnd allows,
2837 * without reducing cwnd. Given that retransmits will set retrans_stamp to a
2838 * non-zero value (and may do so in a later calling context due to TSQ), we
2839 * also enter CA_Loss so that we track when all retransmitted packets are ACKed
2840 * and clear retrans_stamp when that happens (to ensure later recurring RTOs
2841 * are using the correct retrans_stamp and don't declare ETIMEDOUT
2842 * prematurely).
2843 */
tcp_non_congestion_loss_retransmit(struct sock * sk)2844 static void tcp_non_congestion_loss_retransmit(struct sock *sk)
2845 {
2846 const struct inet_connection_sock *icsk = inet_csk(sk);
2847 struct tcp_sock *tp = tcp_sk(sk);
2848
2849 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2850 tp->high_seq = tp->snd_nxt;
2851 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2852 tp->prior_ssthresh = 0;
2853 tp->undo_marker = 0;
2854 tcp_set_ca_state(sk, TCP_CA_Loss);
2855 }
2856 tcp_xmit_retransmit_queue(sk);
2857 }
2858
2859 /* Do a simple retransmit without using the backoff mechanisms in
2860 * tcp_timer. This is used for path mtu discovery.
2861 * The socket is already locked here.
2862 */
tcp_simple_retransmit(struct sock * sk)2863 void tcp_simple_retransmit(struct sock *sk)
2864 {
2865 struct tcp_sock *tp = tcp_sk(sk);
2866 struct sk_buff *skb;
2867 int mss;
2868
2869 /* A fastopen SYN request is stored as two separate packets within
2870 * the retransmit queue, this is done by tcp_send_syn_data().
2871 * As a result simply checking the MSS of the frames in the queue
2872 * will not work for the SYN packet.
2873 *
2874 * Us being here is an indication of a path MTU issue so we can
2875 * assume that the fastopen SYN was lost and just mark all the
2876 * frames in the retransmit queue as lost. We will use an MSS of
2877 * -1 to mark all frames as lost, otherwise compute the current MSS.
2878 */
2879 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT)
2880 mss = -1;
2881 else
2882 mss = tcp_current_mss(sk);
2883
2884 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2885 if (tcp_skb_seglen(skb) > mss)
2886 tcp_mark_skb_lost(sk, skb);
2887 }
2888
2889 tcp_clear_retrans_hints_partial(tp);
2890
2891 if (!tp->lost_out)
2892 return;
2893
2894 if (tcp_is_reno(tp))
2895 tcp_limit_reno_sacked(tp);
2896
2897 tcp_verify_left_out(tp);
2898
2899 /* Don't muck with the congestion window here.
2900 * Reason is that we do not increase amount of _data_
2901 * in network, but units changed and effective
2902 * cwnd/ssthresh really reduced now.
2903 */
2904 tcp_non_congestion_loss_retransmit(sk);
2905 }
2906 EXPORT_IPV6_MOD(tcp_simple_retransmit);
2907
tcp_enter_recovery(struct sock * sk,bool ece_ack)2908 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2909 {
2910 struct tcp_sock *tp = tcp_sk(sk);
2911 int mib_idx;
2912
2913 /* Start the clock with our fast retransmit, for undo and ETIMEDOUT. */
2914 tcp_retrans_stamp_cleanup(sk);
2915
2916 if (tcp_is_reno(tp))
2917 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2918 else
2919 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2920
2921 NET_INC_STATS(sock_net(sk), mib_idx);
2922
2923 tp->prior_ssthresh = 0;
2924 tcp_init_undo(tp);
2925
2926 if (!tcp_in_cwnd_reduction(sk)) {
2927 if (!ece_ack)
2928 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2929 tcp_init_cwnd_reduction(sk);
2930 }
2931 tcp_set_ca_state(sk, TCP_CA_Recovery);
2932 }
2933
tcp_update_rto_time(struct tcp_sock * tp)2934 static void tcp_update_rto_time(struct tcp_sock *tp)
2935 {
2936 if (tp->rto_stamp) {
2937 tp->total_rto_time += tcp_time_stamp_ms(tp) - tp->rto_stamp;
2938 tp->rto_stamp = 0;
2939 }
2940 }
2941
2942 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2943 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2944 */
tcp_process_loss(struct sock * sk,int flag,int num_dupack,int * rexmit)2945 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2946 int *rexmit)
2947 {
2948 struct tcp_sock *tp = tcp_sk(sk);
2949 bool recovered = !before(tp->snd_una, tp->high_seq);
2950
2951 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2952 tcp_try_undo_loss(sk, false))
2953 return;
2954
2955 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2956 /* Step 3.b. A timeout is spurious if not all data are
2957 * lost, i.e., never-retransmitted data are (s)acked.
2958 */
2959 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2960 tcp_try_undo_loss(sk, true))
2961 return;
2962
2963 if (after(tp->snd_nxt, tp->high_seq)) {
2964 if (flag & FLAG_DATA_SACKED || num_dupack)
2965 tp->frto = 0; /* Step 3.a. loss was real */
2966 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2967 tp->high_seq = tp->snd_nxt;
2968 /* Step 2.b. Try send new data (but deferred until cwnd
2969 * is updated in tcp_ack()). Otherwise fall back to
2970 * the conventional recovery.
2971 */
2972 if (!tcp_write_queue_empty(sk) &&
2973 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2974 *rexmit = REXMIT_NEW;
2975 return;
2976 }
2977 tp->frto = 0;
2978 }
2979 }
2980
2981 if (recovered) {
2982 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2983 tcp_try_undo_recovery(sk);
2984 return;
2985 }
2986 if (tcp_is_reno(tp)) {
2987 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2988 * delivered. Lower inflight to clock out (re)transmissions.
2989 */
2990 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2991 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2992 else if (flag & FLAG_SND_UNA_ADVANCED)
2993 tcp_reset_reno_sack(tp);
2994 }
2995 *rexmit = REXMIT_LOST;
2996 }
2997
tcp_force_fast_retransmit(struct sock * sk)2998 static bool tcp_force_fast_retransmit(struct sock *sk)
2999 {
3000 struct tcp_sock *tp = tcp_sk(sk);
3001
3002 return after(tcp_highest_sack_seq(tp),
3003 tp->snd_una + tp->reordering * tp->mss_cache);
3004 }
3005
3006 /* Undo during fast recovery after partial ACK. */
tcp_try_undo_partial(struct sock * sk,u32 prior_snd_una,bool * do_lost)3007 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
3008 bool *do_lost)
3009 {
3010 struct tcp_sock *tp = tcp_sk(sk);
3011
3012 if (tp->undo_marker && tcp_packet_delayed(tp)) {
3013 /* Plain luck! Hole if filled with delayed
3014 * packet, rather than with a retransmit. Check reordering.
3015 */
3016 tcp_check_sack_reordering(sk, prior_snd_una, 1);
3017
3018 /* We are getting evidence that the reordering degree is higher
3019 * than we realized. If there are no retransmits out then we
3020 * can undo. Otherwise we clock out new packets but do not
3021 * mark more packets lost or retransmit more.
3022 */
3023 if (tp->retrans_out)
3024 return true;
3025
3026 if (!tcp_any_retrans_done(sk))
3027 tp->retrans_stamp = 0;
3028
3029 DBGUNDO(sk, "partial recovery");
3030 tcp_undo_cwnd_reduction(sk, true);
3031 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
3032 tcp_try_keep_open(sk);
3033 } else {
3034 /* Partial ACK arrived. Force fast retransmit. */
3035 *do_lost = tcp_force_fast_retransmit(sk);
3036 }
3037 return false;
3038 }
3039
tcp_identify_packet_loss(struct sock * sk,int * ack_flag)3040 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
3041 {
3042 struct tcp_sock *tp = tcp_sk(sk);
3043
3044 if (tcp_rtx_queue_empty(sk))
3045 return;
3046
3047 if (unlikely(tcp_is_reno(tp))) {
3048 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
3049 } else if (tcp_is_rack(sk)) {
3050 u32 prior_retrans = tp->retrans_out;
3051
3052 if (tcp_rack_mark_lost(sk))
3053 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
3054 if (prior_retrans > tp->retrans_out)
3055 *ack_flag |= FLAG_LOST_RETRANS;
3056 }
3057 }
3058
3059 /* Process an event, which can update packets-in-flight not trivially.
3060 * Main goal of this function is to calculate new estimate for left_out,
3061 * taking into account both packets sitting in receiver's buffer and
3062 * packets lost by network.
3063 *
3064 * Besides that it updates the congestion state when packet loss or ECN
3065 * is detected. But it does not reduce the cwnd, it is done by the
3066 * congestion control later.
3067 *
3068 * It does _not_ decide what to send, it is made in function
3069 * tcp_xmit_retransmit_queue().
3070 */
tcp_fastretrans_alert(struct sock * sk,const u32 prior_snd_una,int num_dupack,int * ack_flag,int * rexmit)3071 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
3072 int num_dupack, int *ack_flag, int *rexmit)
3073 {
3074 struct inet_connection_sock *icsk = inet_csk(sk);
3075 struct tcp_sock *tp = tcp_sk(sk);
3076 int fast_rexmit = 0, flag = *ack_flag;
3077 bool ece_ack = flag & FLAG_ECE;
3078 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
3079 tcp_force_fast_retransmit(sk));
3080
3081 if (!tp->packets_out && tp->sacked_out)
3082 tp->sacked_out = 0;
3083
3084 /* Now state machine starts.
3085 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3086 if (ece_ack)
3087 tp->prior_ssthresh = 0;
3088
3089 /* B. In all the states check for reneging SACKs. */
3090 if (tcp_check_sack_reneging(sk, ack_flag))
3091 return;
3092
3093 /* C. Check consistency of the current state. */
3094 tcp_verify_left_out(tp);
3095
3096 /* D. Check state exit conditions. State can be terminated
3097 * when high_seq is ACKed. */
3098 if (icsk->icsk_ca_state == TCP_CA_Open) {
3099 WARN_ON(tp->retrans_out != 0 && !tp->syn_data);
3100 tp->retrans_stamp = 0;
3101 } else if (!before(tp->snd_una, tp->high_seq)) {
3102 switch (icsk->icsk_ca_state) {
3103 case TCP_CA_CWR:
3104 /* CWR is to be held something *above* high_seq
3105 * is ACKed for CWR bit to reach receiver. */
3106 if (tp->snd_una != tp->high_seq) {
3107 tcp_end_cwnd_reduction(sk);
3108 tcp_set_ca_state(sk, TCP_CA_Open);
3109 }
3110 break;
3111
3112 case TCP_CA_Recovery:
3113 if (tcp_is_reno(tp))
3114 tcp_reset_reno_sack(tp);
3115 if (tcp_try_undo_recovery(sk))
3116 return;
3117 tcp_end_cwnd_reduction(sk);
3118 break;
3119 }
3120 }
3121
3122 /* E. Process state. */
3123 switch (icsk->icsk_ca_state) {
3124 case TCP_CA_Recovery:
3125 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3126 if (tcp_is_reno(tp))
3127 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3128 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
3129 return;
3130
3131 if (tcp_try_undo_dsack(sk))
3132 tcp_try_to_open(sk, flag);
3133
3134 tcp_identify_packet_loss(sk, ack_flag);
3135 if (icsk->icsk_ca_state != TCP_CA_Recovery) {
3136 if (!tcp_time_to_recover(sk, flag))
3137 return;
3138 /* Undo reverts the recovery state. If loss is evident,
3139 * starts a new recovery (e.g. reordering then loss);
3140 */
3141 tcp_enter_recovery(sk, ece_ack);
3142 }
3143 break;
3144 case TCP_CA_Loss:
3145 tcp_process_loss(sk, flag, num_dupack, rexmit);
3146 if (icsk->icsk_ca_state != TCP_CA_Loss)
3147 tcp_update_rto_time(tp);
3148 tcp_identify_packet_loss(sk, ack_flag);
3149 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
3150 (*ack_flag & FLAG_LOST_RETRANS)))
3151 return;
3152 /* Change state if cwnd is undone or retransmits are lost */
3153 fallthrough;
3154 default:
3155 if (tcp_is_reno(tp)) {
3156 if (flag & FLAG_SND_UNA_ADVANCED)
3157 tcp_reset_reno_sack(tp);
3158 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3159 }
3160
3161 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3162 tcp_try_undo_dsack(sk);
3163
3164 tcp_identify_packet_loss(sk, ack_flag);
3165 if (!tcp_time_to_recover(sk, flag)) {
3166 tcp_try_to_open(sk, flag);
3167 return;
3168 }
3169
3170 /* MTU probe failure: don't reduce cwnd */
3171 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3172 icsk->icsk_mtup.probe_size &&
3173 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3174 tcp_mtup_probe_failed(sk);
3175 /* Restores the reduction we did in tcp_mtup_probe() */
3176 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
3177 tcp_simple_retransmit(sk);
3178 return;
3179 }
3180
3181 /* Otherwise enter Recovery state */
3182 tcp_enter_recovery(sk, ece_ack);
3183 fast_rexmit = 1;
3184 }
3185
3186 if (!tcp_is_rack(sk) && do_lost)
3187 tcp_update_scoreboard(sk, fast_rexmit);
3188 *rexmit = REXMIT_LOST;
3189 }
3190
tcp_update_rtt_min(struct sock * sk,u32 rtt_us,const int flag)3191 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3192 {
3193 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
3194 struct tcp_sock *tp = tcp_sk(sk);
3195
3196 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3197 /* If the remote keeps returning delayed ACKs, eventually
3198 * the min filter would pick it up and overestimate the
3199 * prop. delay when it expires. Skip suspected delayed ACKs.
3200 */
3201 return;
3202 }
3203 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3204 rtt_us ? : jiffies_to_usecs(1));
3205 }
3206
tcp_ack_update_rtt(struct sock * sk,const int flag,long seq_rtt_us,long sack_rtt_us,long ca_rtt_us,struct rate_sample * rs)3207 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3208 long seq_rtt_us, long sack_rtt_us,
3209 long ca_rtt_us, struct rate_sample *rs)
3210 {
3211 const struct tcp_sock *tp = tcp_sk(sk);
3212
3213 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3214 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3215 * Karn's algorithm forbids taking RTT if some retransmitted data
3216 * is acked (RFC6298).
3217 */
3218 if (seq_rtt_us < 0)
3219 seq_rtt_us = sack_rtt_us;
3220
3221 /* RTTM Rule: A TSecr value received in a segment is used to
3222 * update the averaged RTT measurement only if the segment
3223 * acknowledges some new data, i.e., only if it advances the
3224 * left edge of the send window.
3225 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3226 */
3227 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp &&
3228 tp->rx_opt.rcv_tsecr && flag & FLAG_ACKED)
3229 seq_rtt_us = ca_rtt_us = tcp_rtt_tsopt_us(tp);
3230
3231 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3232 if (seq_rtt_us < 0)
3233 return false;
3234
3235 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3236 * always taken together with ACK, SACK, or TS-opts. Any negative
3237 * values will be skipped with the seq_rtt_us < 0 check above.
3238 */
3239 tcp_update_rtt_min(sk, ca_rtt_us, flag);
3240 tcp_rtt_estimator(sk, seq_rtt_us);
3241 tcp_set_rto(sk);
3242
3243 /* RFC6298: only reset backoff on valid RTT measurement. */
3244 inet_csk(sk)->icsk_backoff = 0;
3245 return true;
3246 }
3247
3248 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
tcp_synack_rtt_meas(struct sock * sk,struct request_sock * req)3249 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3250 {
3251 struct rate_sample rs;
3252 long rtt_us = -1L;
3253
3254 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3255 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3256
3257 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3258 }
3259
3260
tcp_cong_avoid(struct sock * sk,u32 ack,u32 acked)3261 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3262 {
3263 const struct inet_connection_sock *icsk = inet_csk(sk);
3264
3265 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3266 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3267 }
3268
3269 /* Restart timer after forward progress on connection.
3270 * RFC2988 recommends to restart timer to now+rto.
3271 */
tcp_rearm_rto(struct sock * sk)3272 void tcp_rearm_rto(struct sock *sk)
3273 {
3274 const struct inet_connection_sock *icsk = inet_csk(sk);
3275 struct tcp_sock *tp = tcp_sk(sk);
3276
3277 /* If the retrans timer is currently being used by Fast Open
3278 * for SYN-ACK retrans purpose, stay put.
3279 */
3280 if (rcu_access_pointer(tp->fastopen_rsk))
3281 return;
3282
3283 if (!tp->packets_out) {
3284 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3285 } else {
3286 u32 rto = inet_csk(sk)->icsk_rto;
3287 /* Offset the time elapsed after installing regular RTO */
3288 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3289 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3290 s64 delta_us = tcp_rto_delta_us(sk);
3291 /* delta_us may not be positive if the socket is locked
3292 * when the retrans timer fires and is rescheduled.
3293 */
3294 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3295 }
3296 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto, true);
3297 }
3298 }
3299
3300 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
tcp_set_xmit_timer(struct sock * sk)3301 static void tcp_set_xmit_timer(struct sock *sk)
3302 {
3303 if (!tcp_schedule_loss_probe(sk, true))
3304 tcp_rearm_rto(sk);
3305 }
3306
3307 /* If we get here, the whole TSO packet has not been acked. */
tcp_tso_acked(struct sock * sk,struct sk_buff * skb)3308 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3309 {
3310 struct tcp_sock *tp = tcp_sk(sk);
3311 u32 packets_acked;
3312
3313 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3314
3315 packets_acked = tcp_skb_pcount(skb);
3316 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3317 return 0;
3318 packets_acked -= tcp_skb_pcount(skb);
3319
3320 if (packets_acked) {
3321 BUG_ON(tcp_skb_pcount(skb) == 0);
3322 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3323 }
3324
3325 return packets_acked;
3326 }
3327
tcp_ack_tstamp(struct sock * sk,struct sk_buff * skb,const struct sk_buff * ack_skb,u32 prior_snd_una)3328 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3329 const struct sk_buff *ack_skb, u32 prior_snd_una)
3330 {
3331 const struct skb_shared_info *shinfo;
3332
3333 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3334 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3335 return;
3336
3337 shinfo = skb_shinfo(skb);
3338 if (!before(shinfo->tskey, prior_snd_una) &&
3339 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3340 tcp_skb_tsorted_save(skb) {
3341 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK);
3342 } tcp_skb_tsorted_restore(skb);
3343 }
3344 }
3345
3346 /* Remove acknowledged frames from the retransmission queue. If our packet
3347 * is before the ack sequence we can discard it as it's confirmed to have
3348 * arrived at the other end.
3349 */
tcp_clean_rtx_queue(struct sock * sk,const struct sk_buff * ack_skb,u32 prior_fack,u32 prior_snd_una,struct tcp_sacktag_state * sack,bool ece_ack)3350 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb,
3351 u32 prior_fack, u32 prior_snd_una,
3352 struct tcp_sacktag_state *sack, bool ece_ack)
3353 {
3354 const struct inet_connection_sock *icsk = inet_csk(sk);
3355 u64 first_ackt, last_ackt;
3356 struct tcp_sock *tp = tcp_sk(sk);
3357 u32 prior_sacked = tp->sacked_out;
3358 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3359 struct sk_buff *skb, *next;
3360 bool fully_acked = true;
3361 long sack_rtt_us = -1L;
3362 long seq_rtt_us = -1L;
3363 long ca_rtt_us = -1L;
3364 u32 pkts_acked = 0;
3365 bool rtt_update;
3366 int flag = 0;
3367
3368 first_ackt = 0;
3369
3370 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3371 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3372 const u32 start_seq = scb->seq;
3373 u8 sacked = scb->sacked;
3374 u32 acked_pcount;
3375
3376 /* Determine how many packets and what bytes were acked, tso and else */
3377 if (after(scb->end_seq, tp->snd_una)) {
3378 if (tcp_skb_pcount(skb) == 1 ||
3379 !after(tp->snd_una, scb->seq))
3380 break;
3381
3382 acked_pcount = tcp_tso_acked(sk, skb);
3383 if (!acked_pcount)
3384 break;
3385 fully_acked = false;
3386 } else {
3387 acked_pcount = tcp_skb_pcount(skb);
3388 }
3389
3390 if (unlikely(sacked & TCPCB_RETRANS)) {
3391 if (sacked & TCPCB_SACKED_RETRANS)
3392 tp->retrans_out -= acked_pcount;
3393 flag |= FLAG_RETRANS_DATA_ACKED;
3394 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3395 last_ackt = tcp_skb_timestamp_us(skb);
3396 WARN_ON_ONCE(last_ackt == 0);
3397 if (!first_ackt)
3398 first_ackt = last_ackt;
3399
3400 if (before(start_seq, reord))
3401 reord = start_seq;
3402 if (!after(scb->end_seq, tp->high_seq))
3403 flag |= FLAG_ORIG_SACK_ACKED;
3404 }
3405
3406 if (sacked & TCPCB_SACKED_ACKED) {
3407 tp->sacked_out -= acked_pcount;
3408 } else if (tcp_is_sack(tp)) {
3409 tcp_count_delivered(tp, acked_pcount, ece_ack);
3410 if (!tcp_skb_spurious_retrans(tp, skb))
3411 tcp_rack_advance(tp, sacked, scb->end_seq,
3412 tcp_skb_timestamp_us(skb));
3413 }
3414 if (sacked & TCPCB_LOST)
3415 tp->lost_out -= acked_pcount;
3416
3417 tp->packets_out -= acked_pcount;
3418 pkts_acked += acked_pcount;
3419 tcp_rate_skb_delivered(sk, skb, sack->rate);
3420
3421 /* Initial outgoing SYN's get put onto the write_queue
3422 * just like anything else we transmit. It is not
3423 * true data, and if we misinform our callers that
3424 * this ACK acks real data, we will erroneously exit
3425 * connection startup slow start one packet too
3426 * quickly. This is severely frowned upon behavior.
3427 */
3428 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3429 flag |= FLAG_DATA_ACKED;
3430 } else {
3431 flag |= FLAG_SYN_ACKED;
3432 tp->retrans_stamp = 0;
3433 }
3434
3435 if (!fully_acked)
3436 break;
3437
3438 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3439
3440 next = skb_rb_next(skb);
3441 if (unlikely(skb == tp->retransmit_skb_hint))
3442 tp->retransmit_skb_hint = NULL;
3443 if (unlikely(skb == tp->lost_skb_hint))
3444 tp->lost_skb_hint = NULL;
3445 tcp_highest_sack_replace(sk, skb, next);
3446 tcp_rtx_queue_unlink_and_free(skb, sk);
3447 }
3448
3449 if (!skb)
3450 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3451
3452 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3453 tp->snd_up = tp->snd_una;
3454
3455 if (skb) {
3456 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3457 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3458 flag |= FLAG_SACK_RENEGING;
3459 }
3460
3461 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3462 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3463 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3464
3465 if (pkts_acked == 1 && fully_acked && !prior_sacked &&
3466 (tp->snd_una - prior_snd_una) < tp->mss_cache &&
3467 sack->rate->prior_delivered + 1 == tp->delivered &&
3468 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3469 /* Conservatively mark a delayed ACK. It's typically
3470 * from a lone runt packet over the round trip to
3471 * a receiver w/o out-of-order or CE events.
3472 */
3473 flag |= FLAG_ACK_MAYBE_DELAYED;
3474 }
3475 }
3476 if (sack->first_sackt) {
3477 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3478 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3479 }
3480 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3481 ca_rtt_us, sack->rate);
3482
3483 if (flag & FLAG_ACKED) {
3484 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3485 if (unlikely(icsk->icsk_mtup.probe_size &&
3486 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3487 tcp_mtup_probe_success(sk);
3488 }
3489
3490 if (tcp_is_reno(tp)) {
3491 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3492
3493 /* If any of the cumulatively ACKed segments was
3494 * retransmitted, non-SACK case cannot confirm that
3495 * progress was due to original transmission due to
3496 * lack of TCPCB_SACKED_ACKED bits even if some of
3497 * the packets may have been never retransmitted.
3498 */
3499 if (flag & FLAG_RETRANS_DATA_ACKED)
3500 flag &= ~FLAG_ORIG_SACK_ACKED;
3501 } else {
3502 int delta;
3503
3504 /* Non-retransmitted hole got filled? That's reordering */
3505 if (before(reord, prior_fack))
3506 tcp_check_sack_reordering(sk, reord, 0);
3507
3508 delta = prior_sacked - tp->sacked_out;
3509 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3510 }
3511 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3512 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3513 tcp_skb_timestamp_us(skb))) {
3514 /* Do not re-arm RTO if the sack RTT is measured from data sent
3515 * after when the head was last (re)transmitted. Otherwise the
3516 * timeout may continue to extend in loss recovery.
3517 */
3518 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3519 }
3520
3521 if (icsk->icsk_ca_ops->pkts_acked) {
3522 struct ack_sample sample = { .pkts_acked = pkts_acked,
3523 .rtt_us = sack->rate->rtt_us };
3524
3525 sample.in_flight = tp->mss_cache *
3526 (tp->delivered - sack->rate->prior_delivered);
3527 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3528 }
3529
3530 #if FASTRETRANS_DEBUG > 0
3531 WARN_ON((int)tp->sacked_out < 0);
3532 WARN_ON((int)tp->lost_out < 0);
3533 WARN_ON((int)tp->retrans_out < 0);
3534 if (!tp->packets_out && tcp_is_sack(tp)) {
3535 icsk = inet_csk(sk);
3536 if (tp->lost_out) {
3537 pr_debug("Leak l=%u %d\n",
3538 tp->lost_out, icsk->icsk_ca_state);
3539 tp->lost_out = 0;
3540 }
3541 if (tp->sacked_out) {
3542 pr_debug("Leak s=%u %d\n",
3543 tp->sacked_out, icsk->icsk_ca_state);
3544 tp->sacked_out = 0;
3545 }
3546 if (tp->retrans_out) {
3547 pr_debug("Leak r=%u %d\n",
3548 tp->retrans_out, icsk->icsk_ca_state);
3549 tp->retrans_out = 0;
3550 }
3551 }
3552 #endif
3553 return flag;
3554 }
3555
tcp_ack_probe(struct sock * sk)3556 static void tcp_ack_probe(struct sock *sk)
3557 {
3558 struct inet_connection_sock *icsk = inet_csk(sk);
3559 struct sk_buff *head = tcp_send_head(sk);
3560 const struct tcp_sock *tp = tcp_sk(sk);
3561
3562 /* Was it a usable window open? */
3563 if (!head)
3564 return;
3565 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3566 icsk->icsk_backoff = 0;
3567 icsk->icsk_probes_tstamp = 0;
3568 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3569 /* Socket must be waked up by subsequent tcp_data_snd_check().
3570 * This function is not for random using!
3571 */
3572 } else {
3573 unsigned long when = tcp_probe0_when(sk, tcp_rto_max(sk));
3574
3575 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3576 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, true);
3577 }
3578 }
3579
tcp_ack_is_dubious(const struct sock * sk,const int flag)3580 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3581 {
3582 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3583 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3584 }
3585
3586 /* Decide wheather to run the increase function of congestion control. */
tcp_may_raise_cwnd(const struct sock * sk,const int flag)3587 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3588 {
3589 /* If reordering is high then always grow cwnd whenever data is
3590 * delivered regardless of its ordering. Otherwise stay conservative
3591 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3592 * new SACK or ECE mark may first advance cwnd here and later reduce
3593 * cwnd in tcp_fastretrans_alert() based on more states.
3594 */
3595 if (tcp_sk(sk)->reordering >
3596 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3597 return flag & FLAG_FORWARD_PROGRESS;
3598
3599 return flag & FLAG_DATA_ACKED;
3600 }
3601
3602 /* The "ultimate" congestion control function that aims to replace the rigid
3603 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3604 * It's called toward the end of processing an ACK with precise rate
3605 * information. All transmission or retransmission are delayed afterwards.
3606 */
tcp_cong_control(struct sock * sk,u32 ack,u32 acked_sacked,int flag,const struct rate_sample * rs)3607 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3608 int flag, const struct rate_sample *rs)
3609 {
3610 const struct inet_connection_sock *icsk = inet_csk(sk);
3611
3612 if (icsk->icsk_ca_ops->cong_control) {
3613 icsk->icsk_ca_ops->cong_control(sk, ack, flag, rs);
3614 return;
3615 }
3616
3617 if (tcp_in_cwnd_reduction(sk)) {
3618 /* Reduce cwnd if state mandates */
3619 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag);
3620 } else if (tcp_may_raise_cwnd(sk, flag)) {
3621 /* Advance cwnd if state allows */
3622 tcp_cong_avoid(sk, ack, acked_sacked);
3623 }
3624 tcp_update_pacing_rate(sk);
3625 }
3626
3627 /* Check that window update is acceptable.
3628 * The function assumes that snd_una<=ack<=snd_next.
3629 */
tcp_may_update_window(const struct tcp_sock * tp,const u32 ack,const u32 ack_seq,const u32 nwin)3630 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3631 const u32 ack, const u32 ack_seq,
3632 const u32 nwin)
3633 {
3634 return after(ack, tp->snd_una) ||
3635 after(ack_seq, tp->snd_wl1) ||
3636 (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin));
3637 }
3638
tcp_snd_sne_update(struct tcp_sock * tp,u32 ack)3639 static void tcp_snd_sne_update(struct tcp_sock *tp, u32 ack)
3640 {
3641 #ifdef CONFIG_TCP_AO
3642 struct tcp_ao_info *ao;
3643
3644 if (!static_branch_unlikely(&tcp_ao_needed.key))
3645 return;
3646
3647 ao = rcu_dereference_protected(tp->ao_info,
3648 lockdep_sock_is_held((struct sock *)tp));
3649 if (ao && ack < tp->snd_una) {
3650 ao->snd_sne++;
3651 trace_tcp_ao_snd_sne_update((struct sock *)tp, ao->snd_sne);
3652 }
3653 #endif
3654 }
3655
3656 /* If we update tp->snd_una, also update tp->bytes_acked */
tcp_snd_una_update(struct tcp_sock * tp,u32 ack)3657 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3658 {
3659 u32 delta = ack - tp->snd_una;
3660
3661 sock_owned_by_me((struct sock *)tp);
3662 tp->bytes_acked += delta;
3663 tcp_snd_sne_update(tp, ack);
3664 tp->snd_una = ack;
3665 }
3666
tcp_rcv_sne_update(struct tcp_sock * tp,u32 seq)3667 static void tcp_rcv_sne_update(struct tcp_sock *tp, u32 seq)
3668 {
3669 #ifdef CONFIG_TCP_AO
3670 struct tcp_ao_info *ao;
3671
3672 if (!static_branch_unlikely(&tcp_ao_needed.key))
3673 return;
3674
3675 ao = rcu_dereference_protected(tp->ao_info,
3676 lockdep_sock_is_held((struct sock *)tp));
3677 if (ao && seq < tp->rcv_nxt) {
3678 ao->rcv_sne++;
3679 trace_tcp_ao_rcv_sne_update((struct sock *)tp, ao->rcv_sne);
3680 }
3681 #endif
3682 }
3683
3684 /* If we update tp->rcv_nxt, also update tp->bytes_received */
tcp_rcv_nxt_update(struct tcp_sock * tp,u32 seq)3685 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3686 {
3687 u32 delta = seq - tp->rcv_nxt;
3688
3689 sock_owned_by_me((struct sock *)tp);
3690 tp->bytes_received += delta;
3691 tcp_rcv_sne_update(tp, seq);
3692 WRITE_ONCE(tp->rcv_nxt, seq);
3693 }
3694
3695 /* Update our send window.
3696 *
3697 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3698 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3699 */
tcp_ack_update_window(struct sock * sk,const struct sk_buff * skb,u32 ack,u32 ack_seq)3700 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3701 u32 ack_seq)
3702 {
3703 struct tcp_sock *tp = tcp_sk(sk);
3704 int flag = 0;
3705 u32 nwin = ntohs(tcp_hdr(skb)->window);
3706
3707 if (likely(!tcp_hdr(skb)->syn))
3708 nwin <<= tp->rx_opt.snd_wscale;
3709
3710 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3711 flag |= FLAG_WIN_UPDATE;
3712 tcp_update_wl(tp, ack_seq);
3713
3714 if (tp->snd_wnd != nwin) {
3715 tp->snd_wnd = nwin;
3716
3717 /* Note, it is the only place, where
3718 * fast path is recovered for sending TCP.
3719 */
3720 tp->pred_flags = 0;
3721 tcp_fast_path_check(sk);
3722
3723 if (!tcp_write_queue_empty(sk))
3724 tcp_slow_start_after_idle_check(sk);
3725
3726 if (nwin > tp->max_window) {
3727 tp->max_window = nwin;
3728 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3729 }
3730 }
3731 }
3732
3733 tcp_snd_una_update(tp, ack);
3734
3735 return flag;
3736 }
3737
__tcp_oow_rate_limited(struct net * net,int mib_idx,u32 * last_oow_ack_time)3738 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3739 u32 *last_oow_ack_time)
3740 {
3741 /* Paired with the WRITE_ONCE() in this function. */
3742 u32 val = READ_ONCE(*last_oow_ack_time);
3743
3744 if (val) {
3745 s32 elapsed = (s32)(tcp_jiffies32 - val);
3746
3747 if (0 <= elapsed &&
3748 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3749 NET_INC_STATS(net, mib_idx);
3750 return true; /* rate-limited: don't send yet! */
3751 }
3752 }
3753
3754 /* Paired with the prior READ_ONCE() and with itself,
3755 * as we might be lockless.
3756 */
3757 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32);
3758
3759 return false; /* not rate-limited: go ahead, send dupack now! */
3760 }
3761
3762 /* Return true if we're currently rate-limiting out-of-window ACKs and
3763 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3764 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3765 * attacks that send repeated SYNs or ACKs for the same connection. To
3766 * do this, we do not send a duplicate SYNACK or ACK if the remote
3767 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3768 */
tcp_oow_rate_limited(struct net * net,const struct sk_buff * skb,int mib_idx,u32 * last_oow_ack_time)3769 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3770 int mib_idx, u32 *last_oow_ack_time)
3771 {
3772 /* Data packets without SYNs are not likely part of an ACK loop. */
3773 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3774 !tcp_hdr(skb)->syn)
3775 return false;
3776
3777 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3778 }
3779
3780 /* RFC 5961 7 [ACK Throttling] */
tcp_send_challenge_ack(struct sock * sk)3781 static void tcp_send_challenge_ack(struct sock *sk)
3782 {
3783 struct tcp_sock *tp = tcp_sk(sk);
3784 struct net *net = sock_net(sk);
3785 u32 count, now, ack_limit;
3786
3787 /* First check our per-socket dupack rate limit. */
3788 if (__tcp_oow_rate_limited(net,
3789 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3790 &tp->last_oow_ack_time))
3791 return;
3792
3793 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3794 if (ack_limit == INT_MAX)
3795 goto send_ack;
3796
3797 /* Then check host-wide RFC 5961 rate limit. */
3798 now = jiffies / HZ;
3799 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) {
3800 u32 half = (ack_limit + 1) >> 1;
3801
3802 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now);
3803 WRITE_ONCE(net->ipv4.tcp_challenge_count,
3804 get_random_u32_inclusive(half, ack_limit + half - 1));
3805 }
3806 count = READ_ONCE(net->ipv4.tcp_challenge_count);
3807 if (count > 0) {
3808 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1);
3809 send_ack:
3810 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3811 tcp_send_ack(sk);
3812 }
3813 }
3814
tcp_store_ts_recent(struct tcp_sock * tp)3815 static void tcp_store_ts_recent(struct tcp_sock *tp)
3816 {
3817 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3818 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3819 }
3820
__tcp_replace_ts_recent(struct tcp_sock * tp,s32 tstamp_delta)3821 static int __tcp_replace_ts_recent(struct tcp_sock *tp, s32 tstamp_delta)
3822 {
3823 tcp_store_ts_recent(tp);
3824 return tstamp_delta > 0 ? FLAG_TS_PROGRESS : 0;
3825 }
3826
tcp_replace_ts_recent(struct tcp_sock * tp,u32 seq)3827 static int tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3828 {
3829 s32 delta;
3830
3831 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3832 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3833 * extra check below makes sure this can only happen
3834 * for pure ACK frames. -DaveM
3835 *
3836 * Not only, also it occurs for expired timestamps.
3837 */
3838
3839 if (tcp_paws_check(&tp->rx_opt, 0)) {
3840 delta = tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent;
3841 return __tcp_replace_ts_recent(tp, delta);
3842 }
3843 }
3844
3845 return 0;
3846 }
3847
3848 /* This routine deals with acks during a TLP episode and ends an episode by
3849 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3850 */
tcp_process_tlp_ack(struct sock * sk,u32 ack,int flag)3851 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3852 {
3853 struct tcp_sock *tp = tcp_sk(sk);
3854
3855 if (before(ack, tp->tlp_high_seq))
3856 return;
3857
3858 if (!tp->tlp_retrans) {
3859 /* TLP of new data has been acknowledged */
3860 tp->tlp_high_seq = 0;
3861 } else if (flag & FLAG_DSACK_TLP) {
3862 /* This DSACK means original and TLP probe arrived; no loss */
3863 tp->tlp_high_seq = 0;
3864 } else if (after(ack, tp->tlp_high_seq)) {
3865 /* ACK advances: there was a loss, so reduce cwnd. Reset
3866 * tlp_high_seq in tcp_init_cwnd_reduction()
3867 */
3868 tcp_init_cwnd_reduction(sk);
3869 tcp_set_ca_state(sk, TCP_CA_CWR);
3870 tcp_end_cwnd_reduction(sk);
3871 tcp_try_keep_open(sk);
3872 NET_INC_STATS(sock_net(sk),
3873 LINUX_MIB_TCPLOSSPROBERECOVERY);
3874 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3875 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3876 /* Pure dupack: original and TLP probe arrived; no loss */
3877 tp->tlp_high_seq = 0;
3878 }
3879 }
3880
tcp_in_ack_event(struct sock * sk,int flag)3881 static void tcp_in_ack_event(struct sock *sk, int flag)
3882 {
3883 const struct inet_connection_sock *icsk = inet_csk(sk);
3884
3885 if (icsk->icsk_ca_ops->in_ack_event) {
3886 u32 ack_ev_flags = 0;
3887
3888 if (flag & FLAG_WIN_UPDATE)
3889 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3890 if (flag & FLAG_SLOWPATH) {
3891 ack_ev_flags |= CA_ACK_SLOWPATH;
3892 if (flag & FLAG_ECE)
3893 ack_ev_flags |= CA_ACK_ECE;
3894 }
3895
3896 icsk->icsk_ca_ops->in_ack_event(sk, ack_ev_flags);
3897 }
3898 }
3899
3900 /* Congestion control has updated the cwnd already. So if we're in
3901 * loss recovery then now we do any new sends (for FRTO) or
3902 * retransmits (for CA_Loss or CA_recovery) that make sense.
3903 */
tcp_xmit_recovery(struct sock * sk,int rexmit)3904 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3905 {
3906 struct tcp_sock *tp = tcp_sk(sk);
3907
3908 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3909 return;
3910
3911 if (unlikely(rexmit == REXMIT_NEW)) {
3912 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3913 TCP_NAGLE_OFF);
3914 if (after(tp->snd_nxt, tp->high_seq))
3915 return;
3916 tp->frto = 0;
3917 }
3918 tcp_xmit_retransmit_queue(sk);
3919 }
3920
3921 /* Returns the number of packets newly acked or sacked by the current ACK */
tcp_newly_delivered(struct sock * sk,u32 prior_delivered,int flag)3922 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3923 {
3924 const struct net *net = sock_net(sk);
3925 struct tcp_sock *tp = tcp_sk(sk);
3926 u32 delivered;
3927
3928 delivered = tp->delivered - prior_delivered;
3929 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3930 if (flag & FLAG_ECE)
3931 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3932
3933 return delivered;
3934 }
3935
3936 /* This routine deals with incoming acks, but not outgoing ones. */
tcp_ack(struct sock * sk,const struct sk_buff * skb,int flag)3937 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3938 {
3939 struct inet_connection_sock *icsk = inet_csk(sk);
3940 struct tcp_sock *tp = tcp_sk(sk);
3941 struct tcp_sacktag_state sack_state;
3942 struct rate_sample rs = { .prior_delivered = 0 };
3943 u32 prior_snd_una = tp->snd_una;
3944 bool is_sack_reneg = tp->is_sack_reneg;
3945 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3946 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3947 int num_dupack = 0;
3948 int prior_packets = tp->packets_out;
3949 u32 delivered = tp->delivered;
3950 u32 lost = tp->lost;
3951 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3952 u32 prior_fack;
3953
3954 sack_state.first_sackt = 0;
3955 sack_state.rate = &rs;
3956 sack_state.sack_delivered = 0;
3957
3958 /* We very likely will need to access rtx queue. */
3959 prefetch(sk->tcp_rtx_queue.rb_node);
3960
3961 /* If the ack is older than previous acks
3962 * then we can probably ignore it.
3963 */
3964 if (before(ack, prior_snd_una)) {
3965 u32 max_window;
3966
3967 /* do not accept ACK for bytes we never sent. */
3968 max_window = min_t(u64, tp->max_window, tp->bytes_acked);
3969 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3970 if (before(ack, prior_snd_una - max_window)) {
3971 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3972 tcp_send_challenge_ack(sk);
3973 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK;
3974 }
3975 goto old_ack;
3976 }
3977
3978 /* If the ack includes data we haven't sent yet, discard
3979 * this segment (RFC793 Section 3.9).
3980 */
3981 if (after(ack, tp->snd_nxt))
3982 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA;
3983
3984 if (after(ack, prior_snd_una)) {
3985 flag |= FLAG_SND_UNA_ADVANCED;
3986 icsk->icsk_retransmits = 0;
3987
3988 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3989 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3990 if (tp->tcp_clean_acked)
3991 tp->tcp_clean_acked(sk, ack);
3992 #endif
3993 }
3994
3995 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3996 rs.prior_in_flight = tcp_packets_in_flight(tp);
3997
3998 /* ts_recent update must be made after we are sure that the packet
3999 * is in window.
4000 */
4001 if (flag & FLAG_UPDATE_TS_RECENT)
4002 flag |= tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4003
4004 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
4005 FLAG_SND_UNA_ADVANCED) {
4006 /* Window is constant, pure forward advance.
4007 * No more checks are required.
4008 * Note, we use the fact that SND.UNA>=SND.WL2.
4009 */
4010 tcp_update_wl(tp, ack_seq);
4011 tcp_snd_una_update(tp, ack);
4012 flag |= FLAG_WIN_UPDATE;
4013
4014 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
4015 } else {
4016 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
4017 flag |= FLAG_DATA;
4018 else
4019 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
4020
4021 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
4022
4023 if (TCP_SKB_CB(skb)->sacked)
4024 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
4025 &sack_state);
4026
4027 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb)))
4028 flag |= FLAG_ECE;
4029
4030 if (sack_state.sack_delivered)
4031 tcp_count_delivered(tp, sack_state.sack_delivered,
4032 flag & FLAG_ECE);
4033 }
4034
4035 /* This is a deviation from RFC3168 since it states that:
4036 * "When the TCP data sender is ready to set the CWR bit after reducing
4037 * the congestion window, it SHOULD set the CWR bit only on the first
4038 * new data packet that it transmits."
4039 * We accept CWR on pure ACKs to be more robust
4040 * with widely-deployed TCP implementations that do this.
4041 */
4042 tcp_ecn_accept_cwr(sk, skb);
4043
4044 /* We passed data and got it acked, remove any soft error
4045 * log. Something worked...
4046 */
4047 WRITE_ONCE(sk->sk_err_soft, 0);
4048 icsk->icsk_probes_out = 0;
4049 tp->rcv_tstamp = tcp_jiffies32;
4050 if (!prior_packets)
4051 goto no_queue;
4052
4053 /* See if we can take anything off of the retransmit queue. */
4054 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una,
4055 &sack_state, flag & FLAG_ECE);
4056
4057 tcp_rack_update_reo_wnd(sk, &rs);
4058
4059 tcp_in_ack_event(sk, flag);
4060
4061 if (tp->tlp_high_seq)
4062 tcp_process_tlp_ack(sk, ack, flag);
4063
4064 if (tcp_ack_is_dubious(sk, flag)) {
4065 if (!(flag & (FLAG_SND_UNA_ADVANCED |
4066 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
4067 num_dupack = 1;
4068 /* Consider if pure acks were aggregated in tcp_add_backlog() */
4069 if (!(flag & FLAG_DATA))
4070 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4071 }
4072 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4073 &rexmit);
4074 }
4075
4076 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
4077 if (flag & FLAG_SET_XMIT_TIMER)
4078 tcp_set_xmit_timer(sk);
4079
4080 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
4081 sk_dst_confirm(sk);
4082
4083 delivered = tcp_newly_delivered(sk, delivered, flag);
4084 lost = tp->lost - lost; /* freshly marked lost */
4085 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
4086 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
4087 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
4088 tcp_xmit_recovery(sk, rexmit);
4089 return 1;
4090
4091 no_queue:
4092 tcp_in_ack_event(sk, flag);
4093 /* If data was DSACKed, see if we can undo a cwnd reduction. */
4094 if (flag & FLAG_DSACKING_ACK) {
4095 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4096 &rexmit);
4097 tcp_newly_delivered(sk, delivered, flag);
4098 }
4099 /* If this ack opens up a zero window, clear backoff. It was
4100 * being used to time the probes, and is probably far higher than
4101 * it needs to be for normal retransmission.
4102 */
4103 tcp_ack_probe(sk);
4104
4105 if (tp->tlp_high_seq)
4106 tcp_process_tlp_ack(sk, ack, flag);
4107 return 1;
4108
4109 old_ack:
4110 /* If data was SACKed, tag it and see if we should send more data.
4111 * If data was DSACKed, see if we can undo a cwnd reduction.
4112 */
4113 if (TCP_SKB_CB(skb)->sacked) {
4114 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
4115 &sack_state);
4116 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4117 &rexmit);
4118 tcp_newly_delivered(sk, delivered, flag);
4119 tcp_xmit_recovery(sk, rexmit);
4120 }
4121
4122 return 0;
4123 }
4124
tcp_parse_fastopen_option(int len,const unsigned char * cookie,bool syn,struct tcp_fastopen_cookie * foc,bool exp_opt)4125 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
4126 bool syn, struct tcp_fastopen_cookie *foc,
4127 bool exp_opt)
4128 {
4129 /* Valid only in SYN or SYN-ACK with an even length. */
4130 if (!foc || !syn || len < 0 || (len & 1))
4131 return;
4132
4133 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
4134 len <= TCP_FASTOPEN_COOKIE_MAX)
4135 memcpy(foc->val, cookie, len);
4136 else if (len != 0)
4137 len = -1;
4138 foc->len = len;
4139 foc->exp = exp_opt;
4140 }
4141
smc_parse_options(const struct tcphdr * th,struct tcp_options_received * opt_rx,const unsigned char * ptr,int opsize)4142 static bool smc_parse_options(const struct tcphdr *th,
4143 struct tcp_options_received *opt_rx,
4144 const unsigned char *ptr,
4145 int opsize)
4146 {
4147 #if IS_ENABLED(CONFIG_SMC)
4148 if (static_branch_unlikely(&tcp_have_smc)) {
4149 if (th->syn && !(opsize & 1) &&
4150 opsize >= TCPOLEN_EXP_SMC_BASE &&
4151 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
4152 opt_rx->smc_ok = 1;
4153 return true;
4154 }
4155 }
4156 #endif
4157 return false;
4158 }
4159
4160 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
4161 * value on success.
4162 */
tcp_parse_mss_option(const struct tcphdr * th,u16 user_mss)4163 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
4164 {
4165 const unsigned char *ptr = (const unsigned char *)(th + 1);
4166 int length = (th->doff * 4) - sizeof(struct tcphdr);
4167 u16 mss = 0;
4168
4169 while (length > 0) {
4170 int opcode = *ptr++;
4171 int opsize;
4172
4173 switch (opcode) {
4174 case TCPOPT_EOL:
4175 return mss;
4176 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4177 length--;
4178 continue;
4179 default:
4180 if (length < 2)
4181 return mss;
4182 opsize = *ptr++;
4183 if (opsize < 2) /* "silly options" */
4184 return mss;
4185 if (opsize > length)
4186 return mss; /* fail on partial options */
4187 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
4188 u16 in_mss = get_unaligned_be16(ptr);
4189
4190 if (in_mss) {
4191 if (user_mss && user_mss < in_mss)
4192 in_mss = user_mss;
4193 mss = in_mss;
4194 }
4195 }
4196 ptr += opsize - 2;
4197 length -= opsize;
4198 }
4199 }
4200 return mss;
4201 }
4202
4203 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
4204 * But, this can also be called on packets in the established flow when
4205 * the fast version below fails.
4206 */
tcp_parse_options(const struct net * net,const struct sk_buff * skb,struct tcp_options_received * opt_rx,int estab,struct tcp_fastopen_cookie * foc)4207 void tcp_parse_options(const struct net *net,
4208 const struct sk_buff *skb,
4209 struct tcp_options_received *opt_rx, int estab,
4210 struct tcp_fastopen_cookie *foc)
4211 {
4212 const unsigned char *ptr;
4213 const struct tcphdr *th = tcp_hdr(skb);
4214 int length = (th->doff * 4) - sizeof(struct tcphdr);
4215
4216 ptr = (const unsigned char *)(th + 1);
4217 opt_rx->saw_tstamp = 0;
4218 opt_rx->saw_unknown = 0;
4219
4220 while (length > 0) {
4221 int opcode = *ptr++;
4222 int opsize;
4223
4224 switch (opcode) {
4225 case TCPOPT_EOL:
4226 return;
4227 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4228 length--;
4229 continue;
4230 default:
4231 if (length < 2)
4232 return;
4233 opsize = *ptr++;
4234 if (opsize < 2) /* "silly options" */
4235 return;
4236 if (opsize > length)
4237 return; /* don't parse partial options */
4238 switch (opcode) {
4239 case TCPOPT_MSS:
4240 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
4241 u16 in_mss = get_unaligned_be16(ptr);
4242 if (in_mss) {
4243 if (opt_rx->user_mss &&
4244 opt_rx->user_mss < in_mss)
4245 in_mss = opt_rx->user_mss;
4246 opt_rx->mss_clamp = in_mss;
4247 }
4248 }
4249 break;
4250 case TCPOPT_WINDOW:
4251 if (opsize == TCPOLEN_WINDOW && th->syn &&
4252 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) {
4253 __u8 snd_wscale = *(__u8 *)ptr;
4254 opt_rx->wscale_ok = 1;
4255 if (snd_wscale > TCP_MAX_WSCALE) {
4256 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4257 __func__,
4258 snd_wscale,
4259 TCP_MAX_WSCALE);
4260 snd_wscale = TCP_MAX_WSCALE;
4261 }
4262 opt_rx->snd_wscale = snd_wscale;
4263 }
4264 break;
4265 case TCPOPT_TIMESTAMP:
4266 if ((opsize == TCPOLEN_TIMESTAMP) &&
4267 ((estab && opt_rx->tstamp_ok) ||
4268 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) {
4269 opt_rx->saw_tstamp = 1;
4270 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4271 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4272 }
4273 break;
4274 case TCPOPT_SACK_PERM:
4275 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4276 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) {
4277 opt_rx->sack_ok = TCP_SACK_SEEN;
4278 tcp_sack_reset(opt_rx);
4279 }
4280 break;
4281
4282 case TCPOPT_SACK:
4283 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4284 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4285 opt_rx->sack_ok) {
4286 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4287 }
4288 break;
4289 #ifdef CONFIG_TCP_MD5SIG
4290 case TCPOPT_MD5SIG:
4291 /* The MD5 Hash has already been
4292 * checked (see tcp_v{4,6}_rcv()).
4293 */
4294 break;
4295 #endif
4296 #ifdef CONFIG_TCP_AO
4297 case TCPOPT_AO:
4298 /* TCP AO has already been checked
4299 * (see tcp_inbound_ao_hash()).
4300 */
4301 break;
4302 #endif
4303 case TCPOPT_FASTOPEN:
4304 tcp_parse_fastopen_option(
4305 opsize - TCPOLEN_FASTOPEN_BASE,
4306 ptr, th->syn, foc, false);
4307 break;
4308
4309 case TCPOPT_EXP:
4310 /* Fast Open option shares code 254 using a
4311 * 16 bits magic number.
4312 */
4313 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4314 get_unaligned_be16(ptr) ==
4315 TCPOPT_FASTOPEN_MAGIC) {
4316 tcp_parse_fastopen_option(opsize -
4317 TCPOLEN_EXP_FASTOPEN_BASE,
4318 ptr + 2, th->syn, foc, true);
4319 break;
4320 }
4321
4322 if (smc_parse_options(th, opt_rx, ptr, opsize))
4323 break;
4324
4325 opt_rx->saw_unknown = 1;
4326 break;
4327
4328 default:
4329 opt_rx->saw_unknown = 1;
4330 }
4331 ptr += opsize-2;
4332 length -= opsize;
4333 }
4334 }
4335 }
4336 EXPORT_SYMBOL(tcp_parse_options);
4337
tcp_parse_aligned_timestamp(struct tcp_sock * tp,const struct tcphdr * th)4338 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4339 {
4340 const __be32 *ptr = (const __be32 *)(th + 1);
4341
4342 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4343 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4344 tp->rx_opt.saw_tstamp = 1;
4345 ++ptr;
4346 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4347 ++ptr;
4348 if (*ptr)
4349 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4350 else
4351 tp->rx_opt.rcv_tsecr = 0;
4352 return true;
4353 }
4354 return false;
4355 }
4356
4357 /* Fast parse options. This hopes to only see timestamps.
4358 * If it is wrong it falls back on tcp_parse_options().
4359 */
tcp_fast_parse_options(const struct net * net,const struct sk_buff * skb,const struct tcphdr * th,struct tcp_sock * tp)4360 static bool tcp_fast_parse_options(const struct net *net,
4361 const struct sk_buff *skb,
4362 const struct tcphdr *th, struct tcp_sock *tp)
4363 {
4364 /* In the spirit of fast parsing, compare doff directly to constant
4365 * values. Because equality is used, short doff can be ignored here.
4366 */
4367 if (th->doff == (sizeof(*th) / 4)) {
4368 tp->rx_opt.saw_tstamp = 0;
4369 return false;
4370 } else if (tp->rx_opt.tstamp_ok &&
4371 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4372 if (tcp_parse_aligned_timestamp(tp, th))
4373 return true;
4374 }
4375
4376 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4377 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4378 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4379
4380 return true;
4381 }
4382
4383 #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
4384 /*
4385 * Parse Signature options
4386 */
tcp_do_parse_auth_options(const struct tcphdr * th,const u8 ** md5_hash,const u8 ** ao_hash)4387 int tcp_do_parse_auth_options(const struct tcphdr *th,
4388 const u8 **md5_hash, const u8 **ao_hash)
4389 {
4390 int length = (th->doff << 2) - sizeof(*th);
4391 const u8 *ptr = (const u8 *)(th + 1);
4392 unsigned int minlen = TCPOLEN_MD5SIG;
4393
4394 if (IS_ENABLED(CONFIG_TCP_AO))
4395 minlen = sizeof(struct tcp_ao_hdr) + 1;
4396
4397 *md5_hash = NULL;
4398 *ao_hash = NULL;
4399
4400 /* If not enough data remaining, we can short cut */
4401 while (length >= minlen) {
4402 int opcode = *ptr++;
4403 int opsize;
4404
4405 switch (opcode) {
4406 case TCPOPT_EOL:
4407 return 0;
4408 case TCPOPT_NOP:
4409 length--;
4410 continue;
4411 default:
4412 opsize = *ptr++;
4413 if (opsize < 2 || opsize > length)
4414 return -EINVAL;
4415 if (opcode == TCPOPT_MD5SIG) {
4416 if (opsize != TCPOLEN_MD5SIG)
4417 return -EINVAL;
4418 if (unlikely(*md5_hash || *ao_hash))
4419 return -EEXIST;
4420 *md5_hash = ptr;
4421 } else if (opcode == TCPOPT_AO) {
4422 if (opsize <= sizeof(struct tcp_ao_hdr))
4423 return -EINVAL;
4424 if (unlikely(*md5_hash || *ao_hash))
4425 return -EEXIST;
4426 *ao_hash = ptr;
4427 }
4428 }
4429 ptr += opsize - 2;
4430 length -= opsize;
4431 }
4432 return 0;
4433 }
4434 EXPORT_SYMBOL(tcp_do_parse_auth_options);
4435 #endif
4436
4437 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4438 *
4439 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4440 * it can pass through stack. So, the following predicate verifies that
4441 * this segment is not used for anything but congestion avoidance or
4442 * fast retransmit. Moreover, we even are able to eliminate most of such
4443 * second order effects, if we apply some small "replay" window (~RTO)
4444 * to timestamp space.
4445 *
4446 * All these measures still do not guarantee that we reject wrapped ACKs
4447 * on networks with high bandwidth, when sequence space is recycled fastly,
4448 * but it guarantees that such events will be very rare and do not affect
4449 * connection seriously. This doesn't look nice, but alas, PAWS is really
4450 * buggy extension.
4451 *
4452 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4453 * states that events when retransmit arrives after original data are rare.
4454 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4455 * the biggest problem on large power networks even with minor reordering.
4456 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4457 * up to bandwidth of 18Gigabit/sec. 8) ]
4458 */
4459
4460 /* Estimates max number of increments of remote peer TSval in
4461 * a replay window (based on our current RTO estimation).
4462 */
tcp_tsval_replay(const struct sock * sk)4463 static u32 tcp_tsval_replay(const struct sock *sk)
4464 {
4465 /* If we use usec TS resolution,
4466 * then expect the remote peer to use the same resolution.
4467 */
4468 if (tcp_sk(sk)->tcp_usec_ts)
4469 return inet_csk(sk)->icsk_rto * (USEC_PER_SEC / HZ);
4470
4471 /* RFC 7323 recommends a TSval clock between 1ms and 1sec.
4472 * We know that some OS (including old linux) can use 1200 Hz.
4473 */
4474 return inet_csk(sk)->icsk_rto * 1200 / HZ;
4475 }
4476
tcp_disordered_ack_check(const struct sock * sk,const struct sk_buff * skb)4477 static enum skb_drop_reason tcp_disordered_ack_check(const struct sock *sk,
4478 const struct sk_buff *skb)
4479 {
4480 const struct tcp_sock *tp = tcp_sk(sk);
4481 const struct tcphdr *th = tcp_hdr(skb);
4482 SKB_DR_INIT(reason, TCP_RFC7323_PAWS);
4483 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4484 u32 seq = TCP_SKB_CB(skb)->seq;
4485
4486 /* 1. Is this not a pure ACK ? */
4487 if (!th->ack || seq != TCP_SKB_CB(skb)->end_seq)
4488 return reason;
4489
4490 /* 2. Is its sequence not the expected one ? */
4491 if (seq != tp->rcv_nxt)
4492 return before(seq, tp->rcv_nxt) ?
4493 SKB_DROP_REASON_TCP_RFC7323_PAWS_ACK :
4494 reason;
4495
4496 /* 3. Is this not a duplicate ACK ? */
4497 if (ack != tp->snd_una)
4498 return reason;
4499
4500 /* 4. Is this updating the window ? */
4501 if (tcp_may_update_window(tp, ack, seq, ntohs(th->window) <<
4502 tp->rx_opt.snd_wscale))
4503 return reason;
4504
4505 /* 5. Is this not in the replay window ? */
4506 if ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) >
4507 tcp_tsval_replay(sk))
4508 return reason;
4509
4510 return 0;
4511 }
4512
4513 /* Check segment sequence number for validity.
4514 *
4515 * Segment controls are considered valid, if the segment
4516 * fits to the window after truncation to the window. Acceptability
4517 * of data (and SYN, FIN, of course) is checked separately.
4518 * See tcp_data_queue(), for example.
4519 *
4520 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4521 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4522 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4523 * (borrowed from freebsd)
4524 */
4525
tcp_sequence(const struct tcp_sock * tp,u32 seq,u32 end_seq)4526 static enum skb_drop_reason tcp_sequence(const struct tcp_sock *tp,
4527 u32 seq, u32 end_seq)
4528 {
4529 if (before(end_seq, tp->rcv_wup))
4530 return SKB_DROP_REASON_TCP_OLD_SEQUENCE;
4531
4532 if (after(seq, tp->rcv_nxt + tcp_receive_window(tp)))
4533 return SKB_DROP_REASON_TCP_INVALID_SEQUENCE;
4534
4535 return SKB_NOT_DROPPED_YET;
4536 }
4537
4538
tcp_done_with_error(struct sock * sk,int err)4539 void tcp_done_with_error(struct sock *sk, int err)
4540 {
4541 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4542 WRITE_ONCE(sk->sk_err, err);
4543 smp_wmb();
4544
4545 tcp_write_queue_purge(sk);
4546 tcp_done(sk);
4547
4548 if (!sock_flag(sk, SOCK_DEAD))
4549 sk_error_report(sk);
4550 }
4551 EXPORT_IPV6_MOD(tcp_done_with_error);
4552
4553 /* When we get a reset we do this. */
tcp_reset(struct sock * sk,struct sk_buff * skb)4554 void tcp_reset(struct sock *sk, struct sk_buff *skb)
4555 {
4556 int err;
4557
4558 trace_tcp_receive_reset(sk);
4559
4560 /* mptcp can't tell us to ignore reset pkts,
4561 * so just ignore the return value of mptcp_incoming_options().
4562 */
4563 if (sk_is_mptcp(sk))
4564 mptcp_incoming_options(sk, skb);
4565
4566 /* We want the right error as BSD sees it (and indeed as we do). */
4567 switch (sk->sk_state) {
4568 case TCP_SYN_SENT:
4569 err = ECONNREFUSED;
4570 break;
4571 case TCP_CLOSE_WAIT:
4572 err = EPIPE;
4573 break;
4574 case TCP_CLOSE:
4575 return;
4576 default:
4577 err = ECONNRESET;
4578 }
4579 tcp_done_with_error(sk, err);
4580 }
4581
4582 /*
4583 * Process the FIN bit. This now behaves as it is supposed to work
4584 * and the FIN takes effect when it is validly part of sequence
4585 * space. Not before when we get holes.
4586 *
4587 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4588 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4589 * TIME-WAIT)
4590 *
4591 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4592 * close and we go into CLOSING (and later onto TIME-WAIT)
4593 *
4594 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4595 */
tcp_fin(struct sock * sk)4596 void tcp_fin(struct sock *sk)
4597 {
4598 struct tcp_sock *tp = tcp_sk(sk);
4599
4600 inet_csk_schedule_ack(sk);
4601
4602 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
4603 sock_set_flag(sk, SOCK_DONE);
4604
4605 switch (sk->sk_state) {
4606 case TCP_SYN_RECV:
4607 case TCP_ESTABLISHED:
4608 /* Move to CLOSE_WAIT */
4609 tcp_set_state(sk, TCP_CLOSE_WAIT);
4610 inet_csk_enter_pingpong_mode(sk);
4611 break;
4612
4613 case TCP_CLOSE_WAIT:
4614 case TCP_CLOSING:
4615 /* Received a retransmission of the FIN, do
4616 * nothing.
4617 */
4618 break;
4619 case TCP_LAST_ACK:
4620 /* RFC793: Remain in the LAST-ACK state. */
4621 break;
4622
4623 case TCP_FIN_WAIT1:
4624 /* This case occurs when a simultaneous close
4625 * happens, we must ack the received FIN and
4626 * enter the CLOSING state.
4627 */
4628 tcp_send_ack(sk);
4629 tcp_set_state(sk, TCP_CLOSING);
4630 break;
4631 case TCP_FIN_WAIT2:
4632 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4633 tcp_send_ack(sk);
4634 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4635 break;
4636 default:
4637 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4638 * cases we should never reach this piece of code.
4639 */
4640 pr_err("%s: Impossible, sk->sk_state=%d\n",
4641 __func__, sk->sk_state);
4642 break;
4643 }
4644
4645 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4646 * Probably, we should reset in this case. For now drop them.
4647 */
4648 skb_rbtree_purge(&tp->out_of_order_queue);
4649 if (tcp_is_sack(tp))
4650 tcp_sack_reset(&tp->rx_opt);
4651
4652 if (!sock_flag(sk, SOCK_DEAD)) {
4653 sk->sk_state_change(sk);
4654
4655 /* Do not send POLL_HUP for half duplex close. */
4656 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4657 sk->sk_state == TCP_CLOSE)
4658 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4659 else
4660 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4661 }
4662 }
4663
tcp_sack_extend(struct tcp_sack_block * sp,u32 seq,u32 end_seq)4664 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4665 u32 end_seq)
4666 {
4667 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4668 if (before(seq, sp->start_seq))
4669 sp->start_seq = seq;
4670 if (after(end_seq, sp->end_seq))
4671 sp->end_seq = end_seq;
4672 return true;
4673 }
4674 return false;
4675 }
4676
tcp_dsack_set(struct sock * sk,u32 seq,u32 end_seq)4677 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4678 {
4679 struct tcp_sock *tp = tcp_sk(sk);
4680
4681 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4682 int mib_idx;
4683
4684 if (before(seq, tp->rcv_nxt))
4685 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4686 else
4687 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4688
4689 NET_INC_STATS(sock_net(sk), mib_idx);
4690
4691 tp->rx_opt.dsack = 1;
4692 tp->duplicate_sack[0].start_seq = seq;
4693 tp->duplicate_sack[0].end_seq = end_seq;
4694 }
4695 }
4696
tcp_dsack_extend(struct sock * sk,u32 seq,u32 end_seq)4697 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4698 {
4699 struct tcp_sock *tp = tcp_sk(sk);
4700
4701 if (!tp->rx_opt.dsack)
4702 tcp_dsack_set(sk, seq, end_seq);
4703 else
4704 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4705 }
4706
tcp_rcv_spurious_retrans(struct sock * sk,const struct sk_buff * skb)4707 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4708 {
4709 /* When the ACK path fails or drops most ACKs, the sender would
4710 * timeout and spuriously retransmit the same segment repeatedly.
4711 * If it seems our ACKs are not reaching the other side,
4712 * based on receiving a duplicate data segment with new flowlabel
4713 * (suggesting the sender suffered an RTO), and we are not already
4714 * repathing due to our own RTO, then rehash the socket to repath our
4715 * packets.
4716 */
4717 #if IS_ENABLED(CONFIG_IPV6)
4718 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss &&
4719 skb->protocol == htons(ETH_P_IPV6) &&
4720 (tcp_sk(sk)->inet_conn.icsk_ack.lrcv_flowlabel !=
4721 ntohl(ip6_flowlabel(ipv6_hdr(skb)))) &&
4722 sk_rethink_txhash(sk))
4723 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4724
4725 /* Save last flowlabel after a spurious retrans. */
4726 tcp_save_lrcv_flowlabel(sk, skb);
4727 #endif
4728 }
4729
tcp_send_dupack(struct sock * sk,const struct sk_buff * skb)4730 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4731 {
4732 struct tcp_sock *tp = tcp_sk(sk);
4733
4734 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4735 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4736 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4737 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4738
4739 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4740 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4741
4742 tcp_rcv_spurious_retrans(sk, skb);
4743 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4744 end_seq = tp->rcv_nxt;
4745 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4746 }
4747 }
4748
4749 tcp_send_ack(sk);
4750 }
4751
4752 /* These routines update the SACK block as out-of-order packets arrive or
4753 * in-order packets close up the sequence space.
4754 */
tcp_sack_maybe_coalesce(struct tcp_sock * tp)4755 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4756 {
4757 int this_sack;
4758 struct tcp_sack_block *sp = &tp->selective_acks[0];
4759 struct tcp_sack_block *swalk = sp + 1;
4760
4761 /* See if the recent change to the first SACK eats into
4762 * or hits the sequence space of other SACK blocks, if so coalesce.
4763 */
4764 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4765 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4766 int i;
4767
4768 /* Zap SWALK, by moving every further SACK up by one slot.
4769 * Decrease num_sacks.
4770 */
4771 tp->rx_opt.num_sacks--;
4772 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4773 sp[i] = sp[i + 1];
4774 continue;
4775 }
4776 this_sack++;
4777 swalk++;
4778 }
4779 }
4780
tcp_sack_compress_send_ack(struct sock * sk)4781 void tcp_sack_compress_send_ack(struct sock *sk)
4782 {
4783 struct tcp_sock *tp = tcp_sk(sk);
4784
4785 if (!tp->compressed_ack)
4786 return;
4787
4788 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4789 __sock_put(sk);
4790
4791 /* Since we have to send one ack finally,
4792 * substract one from tp->compressed_ack to keep
4793 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4794 */
4795 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4796 tp->compressed_ack - 1);
4797
4798 tp->compressed_ack = 0;
4799 tcp_send_ack(sk);
4800 }
4801
4802 /* Reasonable amount of sack blocks included in TCP SACK option
4803 * The max is 4, but this becomes 3 if TCP timestamps are there.
4804 * Given that SACK packets might be lost, be conservative and use 2.
4805 */
4806 #define TCP_SACK_BLOCKS_EXPECTED 2
4807
tcp_sack_new_ofo_skb(struct sock * sk,u32 seq,u32 end_seq)4808 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4809 {
4810 struct tcp_sock *tp = tcp_sk(sk);
4811 struct tcp_sack_block *sp = &tp->selective_acks[0];
4812 int cur_sacks = tp->rx_opt.num_sacks;
4813 int this_sack;
4814
4815 if (!cur_sacks)
4816 goto new_sack;
4817
4818 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4819 if (tcp_sack_extend(sp, seq, end_seq)) {
4820 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4821 tcp_sack_compress_send_ack(sk);
4822 /* Rotate this_sack to the first one. */
4823 for (; this_sack > 0; this_sack--, sp--)
4824 swap(*sp, *(sp - 1));
4825 if (cur_sacks > 1)
4826 tcp_sack_maybe_coalesce(tp);
4827 return;
4828 }
4829 }
4830
4831 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4832 tcp_sack_compress_send_ack(sk);
4833
4834 /* Could not find an adjacent existing SACK, build a new one,
4835 * put it at the front, and shift everyone else down. We
4836 * always know there is at least one SACK present already here.
4837 *
4838 * If the sack array is full, forget about the last one.
4839 */
4840 if (this_sack >= TCP_NUM_SACKS) {
4841 this_sack--;
4842 tp->rx_opt.num_sacks--;
4843 sp--;
4844 }
4845 for (; this_sack > 0; this_sack--, sp--)
4846 *sp = *(sp - 1);
4847
4848 new_sack:
4849 /* Build the new head SACK, and we're done. */
4850 sp->start_seq = seq;
4851 sp->end_seq = end_seq;
4852 tp->rx_opt.num_sacks++;
4853 }
4854
4855 /* RCV.NXT advances, some SACKs should be eaten. */
4856
tcp_sack_remove(struct tcp_sock * tp)4857 static void tcp_sack_remove(struct tcp_sock *tp)
4858 {
4859 struct tcp_sack_block *sp = &tp->selective_acks[0];
4860 int num_sacks = tp->rx_opt.num_sacks;
4861 int this_sack;
4862
4863 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4864 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4865 tp->rx_opt.num_sacks = 0;
4866 return;
4867 }
4868
4869 for (this_sack = 0; this_sack < num_sacks;) {
4870 /* Check if the start of the sack is covered by RCV.NXT. */
4871 if (!before(tp->rcv_nxt, sp->start_seq)) {
4872 int i;
4873
4874 /* RCV.NXT must cover all the block! */
4875 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4876
4877 /* Zap this SACK, by moving forward any other SACKS. */
4878 for (i = this_sack+1; i < num_sacks; i++)
4879 tp->selective_acks[i-1] = tp->selective_acks[i];
4880 num_sacks--;
4881 continue;
4882 }
4883 this_sack++;
4884 sp++;
4885 }
4886 tp->rx_opt.num_sacks = num_sacks;
4887 }
4888
4889 /**
4890 * tcp_try_coalesce - try to merge skb to prior one
4891 * @sk: socket
4892 * @to: prior buffer
4893 * @from: buffer to add in queue
4894 * @fragstolen: pointer to boolean
4895 *
4896 * Before queueing skb @from after @to, try to merge them
4897 * to reduce overall memory use and queue lengths, if cost is small.
4898 * Packets in ofo or receive queues can stay a long time.
4899 * Better try to coalesce them right now to avoid future collapses.
4900 * Returns true if caller should free @from instead of queueing it
4901 */
tcp_try_coalesce(struct sock * sk,struct sk_buff * to,struct sk_buff * from,bool * fragstolen)4902 static bool tcp_try_coalesce(struct sock *sk,
4903 struct sk_buff *to,
4904 struct sk_buff *from,
4905 bool *fragstolen)
4906 {
4907 int delta;
4908
4909 *fragstolen = false;
4910
4911 /* Its possible this segment overlaps with prior segment in queue */
4912 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4913 return false;
4914
4915 if (!tcp_skb_can_collapse_rx(to, from))
4916 return false;
4917
4918 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4919 return false;
4920
4921 atomic_add(delta, &sk->sk_rmem_alloc);
4922 sk_mem_charge(sk, delta);
4923 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4924 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4925 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4926 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4927
4928 if (TCP_SKB_CB(from)->has_rxtstamp) {
4929 TCP_SKB_CB(to)->has_rxtstamp = true;
4930 to->tstamp = from->tstamp;
4931 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4932 }
4933
4934 return true;
4935 }
4936
tcp_ooo_try_coalesce(struct sock * sk,struct sk_buff * to,struct sk_buff * from,bool * fragstolen)4937 static bool tcp_ooo_try_coalesce(struct sock *sk,
4938 struct sk_buff *to,
4939 struct sk_buff *from,
4940 bool *fragstolen)
4941 {
4942 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4943
4944 /* In case tcp_drop_reason() is called later, update to->gso_segs */
4945 if (res) {
4946 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4947 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4948
4949 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4950 }
4951 return res;
4952 }
4953
4954 noinline_for_tracing static void
tcp_drop_reason(struct sock * sk,struct sk_buff * skb,enum skb_drop_reason reason)4955 tcp_drop_reason(struct sock *sk, struct sk_buff *skb, enum skb_drop_reason reason)
4956 {
4957 sk_drops_add(sk, skb);
4958 sk_skb_reason_drop(sk, skb, reason);
4959 }
4960
4961 /* This one checks to see if we can put data from the
4962 * out_of_order queue into the receive_queue.
4963 */
tcp_ofo_queue(struct sock * sk)4964 static void tcp_ofo_queue(struct sock *sk)
4965 {
4966 struct tcp_sock *tp = tcp_sk(sk);
4967 __u32 dsack_high = tp->rcv_nxt;
4968 bool fin, fragstolen, eaten;
4969 struct sk_buff *skb, *tail;
4970 struct rb_node *p;
4971
4972 p = rb_first(&tp->out_of_order_queue);
4973 while (p) {
4974 skb = rb_to_skb(p);
4975 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4976 break;
4977
4978 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4979 __u32 dsack = dsack_high;
4980 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4981 dsack_high = TCP_SKB_CB(skb)->end_seq;
4982 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4983 }
4984 p = rb_next(p);
4985 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4986
4987 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4988 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP);
4989 continue;
4990 }
4991
4992 tail = skb_peek_tail(&sk->sk_receive_queue);
4993 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4994 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4995 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4996 if (!eaten)
4997 tcp_add_receive_queue(sk, skb);
4998 else
4999 kfree_skb_partial(skb, fragstolen);
5000
5001 if (unlikely(fin)) {
5002 tcp_fin(sk);
5003 /* tcp_fin() purges tp->out_of_order_queue,
5004 * so we must end this loop right now.
5005 */
5006 break;
5007 }
5008 }
5009 }
5010
5011 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb);
5012 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb);
5013
tcp_try_rmem_schedule(struct sock * sk,struct sk_buff * skb,unsigned int size)5014 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
5015 unsigned int size)
5016 {
5017 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
5018 !sk_rmem_schedule(sk, skb, size)) {
5019
5020 if (tcp_prune_queue(sk, skb) < 0)
5021 return -1;
5022
5023 while (!sk_rmem_schedule(sk, skb, size)) {
5024 if (!tcp_prune_ofo_queue(sk, skb))
5025 return -1;
5026 }
5027 }
5028 return 0;
5029 }
5030
tcp_data_queue_ofo(struct sock * sk,struct sk_buff * skb)5031 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
5032 {
5033 struct tcp_sock *tp = tcp_sk(sk);
5034 struct rb_node **p, *parent;
5035 struct sk_buff *skb1;
5036 u32 seq, end_seq;
5037 bool fragstolen;
5038
5039 tcp_save_lrcv_flowlabel(sk, skb);
5040 tcp_data_ecn_check(sk, skb);
5041
5042 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
5043 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
5044 sk->sk_data_ready(sk);
5045 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM);
5046 return;
5047 }
5048
5049 /* Disable header prediction. */
5050 tp->pred_flags = 0;
5051 inet_csk_schedule_ack(sk);
5052
5053 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
5054 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
5055 seq = TCP_SKB_CB(skb)->seq;
5056 end_seq = TCP_SKB_CB(skb)->end_seq;
5057
5058 p = &tp->out_of_order_queue.rb_node;
5059 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5060 /* Initial out of order segment, build 1 SACK. */
5061 if (tcp_is_sack(tp)) {
5062 tp->rx_opt.num_sacks = 1;
5063 tp->selective_acks[0].start_seq = seq;
5064 tp->selective_acks[0].end_seq = end_seq;
5065 }
5066 rb_link_node(&skb->rbnode, NULL, p);
5067 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
5068 tp->ooo_last_skb = skb;
5069 goto end;
5070 }
5071
5072 /* In the typical case, we are adding an skb to the end of the list.
5073 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
5074 */
5075 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
5076 skb, &fragstolen)) {
5077 coalesce_done:
5078 /* For non sack flows, do not grow window to force DUPACK
5079 * and trigger fast retransmit.
5080 */
5081 if (tcp_is_sack(tp))
5082 tcp_grow_window(sk, skb, true);
5083 kfree_skb_partial(skb, fragstolen);
5084 skb = NULL;
5085 goto add_sack;
5086 }
5087 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
5088 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
5089 parent = &tp->ooo_last_skb->rbnode;
5090 p = &parent->rb_right;
5091 goto insert;
5092 }
5093
5094 /* Find place to insert this segment. Handle overlaps on the way. */
5095 parent = NULL;
5096 while (*p) {
5097 parent = *p;
5098 skb1 = rb_to_skb(parent);
5099 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
5100 p = &parent->rb_left;
5101 continue;
5102 }
5103 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
5104 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
5105 /* All the bits are present. Drop. */
5106 NET_INC_STATS(sock_net(sk),
5107 LINUX_MIB_TCPOFOMERGE);
5108 tcp_drop_reason(sk, skb,
5109 SKB_DROP_REASON_TCP_OFOMERGE);
5110 skb = NULL;
5111 tcp_dsack_set(sk, seq, end_seq);
5112 goto add_sack;
5113 }
5114 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
5115 /* Partial overlap. */
5116 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
5117 } else {
5118 /* skb's seq == skb1's seq and skb covers skb1.
5119 * Replace skb1 with skb.
5120 */
5121 rb_replace_node(&skb1->rbnode, &skb->rbnode,
5122 &tp->out_of_order_queue);
5123 tcp_dsack_extend(sk,
5124 TCP_SKB_CB(skb1)->seq,
5125 TCP_SKB_CB(skb1)->end_seq);
5126 NET_INC_STATS(sock_net(sk),
5127 LINUX_MIB_TCPOFOMERGE);
5128 tcp_drop_reason(sk, skb1,
5129 SKB_DROP_REASON_TCP_OFOMERGE);
5130 goto merge_right;
5131 }
5132 } else if (tcp_ooo_try_coalesce(sk, skb1,
5133 skb, &fragstolen)) {
5134 goto coalesce_done;
5135 }
5136 p = &parent->rb_right;
5137 }
5138 insert:
5139 /* Insert segment into RB tree. */
5140 rb_link_node(&skb->rbnode, parent, p);
5141 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
5142
5143 merge_right:
5144 /* Remove other segments covered by skb. */
5145 while ((skb1 = skb_rb_next(skb)) != NULL) {
5146 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
5147 break;
5148 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
5149 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
5150 end_seq);
5151 break;
5152 }
5153 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
5154 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
5155 TCP_SKB_CB(skb1)->end_seq);
5156 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
5157 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE);
5158 }
5159 /* If there is no skb after us, we are the last_skb ! */
5160 if (!skb1)
5161 tp->ooo_last_skb = skb;
5162
5163 add_sack:
5164 if (tcp_is_sack(tp))
5165 tcp_sack_new_ofo_skb(sk, seq, end_seq);
5166 end:
5167 if (skb) {
5168 /* For non sack flows, do not grow window to force DUPACK
5169 * and trigger fast retransmit.
5170 */
5171 if (tcp_is_sack(tp))
5172 tcp_grow_window(sk, skb, false);
5173 skb_condense(skb);
5174 tcp_skb_set_owner_r(skb, sk);
5175 }
5176 }
5177
tcp_queue_rcv(struct sock * sk,struct sk_buff * skb,bool * fragstolen)5178 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
5179 bool *fragstolen)
5180 {
5181 int eaten;
5182 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
5183
5184 eaten = (tail &&
5185 tcp_try_coalesce(sk, tail,
5186 skb, fragstolen)) ? 1 : 0;
5187 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
5188 if (!eaten) {
5189 tcp_add_receive_queue(sk, skb);
5190 tcp_skb_set_owner_r(skb, sk);
5191 }
5192 return eaten;
5193 }
5194
tcp_send_rcvq(struct sock * sk,struct msghdr * msg,size_t size)5195 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
5196 {
5197 struct sk_buff *skb;
5198 int err = -ENOMEM;
5199 int data_len = 0;
5200 bool fragstolen;
5201
5202 if (size == 0)
5203 return 0;
5204
5205 if (size > PAGE_SIZE) {
5206 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
5207
5208 data_len = npages << PAGE_SHIFT;
5209 size = data_len + (size & ~PAGE_MASK);
5210 }
5211 skb = alloc_skb_with_frags(size - data_len, data_len,
5212 PAGE_ALLOC_COSTLY_ORDER,
5213 &err, sk->sk_allocation);
5214 if (!skb)
5215 goto err;
5216
5217 skb_put(skb, size - data_len);
5218 skb->data_len = data_len;
5219 skb->len = size;
5220
5221 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5222 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5223 goto err_free;
5224 }
5225
5226 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
5227 if (err)
5228 goto err_free;
5229
5230 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
5231 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
5232 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
5233
5234 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
5235 WARN_ON_ONCE(fragstolen); /* should not happen */
5236 __kfree_skb(skb);
5237 }
5238 return size;
5239
5240 err_free:
5241 kfree_skb(skb);
5242 err:
5243 return err;
5244
5245 }
5246
tcp_data_ready(struct sock * sk)5247 void tcp_data_ready(struct sock *sk)
5248 {
5249 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE))
5250 sk->sk_data_ready(sk);
5251 }
5252
tcp_data_queue(struct sock * sk,struct sk_buff * skb)5253 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
5254 {
5255 struct tcp_sock *tp = tcp_sk(sk);
5256 enum skb_drop_reason reason;
5257 bool fragstolen;
5258 int eaten;
5259
5260 /* If a subflow has been reset, the packet should not continue
5261 * to be processed, drop the packet.
5262 */
5263 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) {
5264 __kfree_skb(skb);
5265 return;
5266 }
5267
5268 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
5269 __kfree_skb(skb);
5270 return;
5271 }
5272 tcp_cleanup_skb(skb);
5273 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
5274
5275 reason = SKB_DROP_REASON_NOT_SPECIFIED;
5276 tp->rx_opt.dsack = 0;
5277
5278 /* Queue data for delivery to the user.
5279 * Packets in sequence go to the receive queue.
5280 * Out of sequence packets to the out_of_order_queue.
5281 */
5282 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5283 if (tcp_receive_window(tp) == 0) {
5284 /* Some stacks are known to send bare FIN packets
5285 * in a loop even if we send RWIN 0 in our ACK.
5286 * Accepting this FIN does not hurt memory pressure
5287 * because the FIN flag will simply be merged to the
5288 * receive queue tail skb in most cases.
5289 */
5290 if (!skb->len &&
5291 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
5292 goto queue_and_out;
5293
5294 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5295 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5296 goto out_of_window;
5297 }
5298
5299 /* Ok. In sequence. In window. */
5300 queue_and_out:
5301 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5302 /* TODO: maybe ratelimit these WIN 0 ACK ? */
5303 inet_csk(sk)->icsk_ack.pending |=
5304 (ICSK_ACK_NOMEM | ICSK_ACK_NOW);
5305 inet_csk_schedule_ack(sk);
5306 sk->sk_data_ready(sk);
5307
5308 if (skb_queue_len(&sk->sk_receive_queue) && skb->len) {
5309 reason = SKB_DROP_REASON_PROTO_MEM;
5310 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5311 goto drop;
5312 }
5313 sk_forced_mem_schedule(sk, skb->truesize);
5314 }
5315
5316 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5317 if (skb->len)
5318 tcp_event_data_recv(sk, skb);
5319 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
5320 tcp_fin(sk);
5321
5322 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5323 tcp_ofo_queue(sk);
5324
5325 /* RFC5681. 4.2. SHOULD send immediate ACK, when
5326 * gap in queue is filled.
5327 */
5328 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5329 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
5330 }
5331
5332 if (tp->rx_opt.num_sacks)
5333 tcp_sack_remove(tp);
5334
5335 tcp_fast_path_check(sk);
5336
5337 if (eaten > 0)
5338 kfree_skb_partial(skb, fragstolen);
5339 if (!sock_flag(sk, SOCK_DEAD))
5340 tcp_data_ready(sk);
5341 return;
5342 }
5343
5344 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
5345 tcp_rcv_spurious_retrans(sk, skb);
5346 /* A retransmit, 2nd most common case. Force an immediate ack. */
5347 reason = SKB_DROP_REASON_TCP_OLD_DATA;
5348 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5349 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5350
5351 out_of_window:
5352 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5353 inet_csk_schedule_ack(sk);
5354 drop:
5355 tcp_drop_reason(sk, skb, reason);
5356 return;
5357 }
5358
5359 /* Out of window. F.e. zero window probe. */
5360 if (!before(TCP_SKB_CB(skb)->seq,
5361 tp->rcv_nxt + tcp_receive_window(tp))) {
5362 reason = SKB_DROP_REASON_TCP_OVERWINDOW;
5363 goto out_of_window;
5364 }
5365
5366 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5367 /* Partial packet, seq < rcv_next < end_seq */
5368 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5369
5370 /* If window is closed, drop tail of packet. But after
5371 * remembering D-SACK for its head made in previous line.
5372 */
5373 if (!tcp_receive_window(tp)) {
5374 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5375 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5376 goto out_of_window;
5377 }
5378 goto queue_and_out;
5379 }
5380
5381 tcp_data_queue_ofo(sk, skb);
5382 }
5383
tcp_skb_next(struct sk_buff * skb,struct sk_buff_head * list)5384 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5385 {
5386 if (list)
5387 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5388
5389 return skb_rb_next(skb);
5390 }
5391
tcp_collapse_one(struct sock * sk,struct sk_buff * skb,struct sk_buff_head * list,struct rb_root * root)5392 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5393 struct sk_buff_head *list,
5394 struct rb_root *root)
5395 {
5396 struct sk_buff *next = tcp_skb_next(skb, list);
5397
5398 if (list)
5399 __skb_unlink(skb, list);
5400 else
5401 rb_erase(&skb->rbnode, root);
5402
5403 __kfree_skb(skb);
5404 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5405
5406 return next;
5407 }
5408
5409 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
tcp_rbtree_insert(struct rb_root * root,struct sk_buff * skb)5410 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5411 {
5412 struct rb_node **p = &root->rb_node;
5413 struct rb_node *parent = NULL;
5414 struct sk_buff *skb1;
5415
5416 while (*p) {
5417 parent = *p;
5418 skb1 = rb_to_skb(parent);
5419 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5420 p = &parent->rb_left;
5421 else
5422 p = &parent->rb_right;
5423 }
5424 rb_link_node(&skb->rbnode, parent, p);
5425 rb_insert_color(&skb->rbnode, root);
5426 }
5427
5428 /* Collapse contiguous sequence of skbs head..tail with
5429 * sequence numbers start..end.
5430 *
5431 * If tail is NULL, this means until the end of the queue.
5432 *
5433 * Segments with FIN/SYN are not collapsed (only because this
5434 * simplifies code)
5435 */
5436 static void
tcp_collapse(struct sock * sk,struct sk_buff_head * list,struct rb_root * root,struct sk_buff * head,struct sk_buff * tail,u32 start,u32 end)5437 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5438 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5439 {
5440 struct sk_buff *skb = head, *n;
5441 struct sk_buff_head tmp;
5442 bool end_of_skbs;
5443
5444 /* First, check that queue is collapsible and find
5445 * the point where collapsing can be useful.
5446 */
5447 restart:
5448 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5449 n = tcp_skb_next(skb, list);
5450
5451 if (!skb_frags_readable(skb))
5452 goto skip_this;
5453
5454 /* No new bits? It is possible on ofo queue. */
5455 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5456 skb = tcp_collapse_one(sk, skb, list, root);
5457 if (!skb)
5458 break;
5459 goto restart;
5460 }
5461
5462 /* The first skb to collapse is:
5463 * - not SYN/FIN and
5464 * - bloated or contains data before "start" or
5465 * overlaps to the next one and mptcp allow collapsing.
5466 */
5467 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5468 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5469 before(TCP_SKB_CB(skb)->seq, start))) {
5470 end_of_skbs = false;
5471 break;
5472 }
5473
5474 if (n && n != tail && skb_frags_readable(n) &&
5475 tcp_skb_can_collapse_rx(skb, n) &&
5476 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5477 end_of_skbs = false;
5478 break;
5479 }
5480
5481 skip_this:
5482 /* Decided to skip this, advance start seq. */
5483 start = TCP_SKB_CB(skb)->end_seq;
5484 }
5485 if (end_of_skbs ||
5486 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) ||
5487 !skb_frags_readable(skb))
5488 return;
5489
5490 __skb_queue_head_init(&tmp);
5491
5492 while (before(start, end)) {
5493 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5494 struct sk_buff *nskb;
5495
5496 nskb = alloc_skb(copy, GFP_ATOMIC);
5497 if (!nskb)
5498 break;
5499
5500 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5501 skb_copy_decrypted(nskb, skb);
5502 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5503 if (list)
5504 __skb_queue_before(list, skb, nskb);
5505 else
5506 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5507 tcp_skb_set_owner_r(nskb, sk);
5508 mptcp_skb_ext_move(nskb, skb);
5509
5510 /* Copy data, releasing collapsed skbs. */
5511 while (copy > 0) {
5512 int offset = start - TCP_SKB_CB(skb)->seq;
5513 int size = TCP_SKB_CB(skb)->end_seq - start;
5514
5515 BUG_ON(offset < 0);
5516 if (size > 0) {
5517 size = min(copy, size);
5518 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5519 BUG();
5520 TCP_SKB_CB(nskb)->end_seq += size;
5521 copy -= size;
5522 start += size;
5523 }
5524 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5525 skb = tcp_collapse_one(sk, skb, list, root);
5526 if (!skb ||
5527 skb == tail ||
5528 !tcp_skb_can_collapse_rx(nskb, skb) ||
5529 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) ||
5530 !skb_frags_readable(skb))
5531 goto end;
5532 }
5533 }
5534 }
5535 end:
5536 skb_queue_walk_safe(&tmp, skb, n)
5537 tcp_rbtree_insert(root, skb);
5538 }
5539
5540 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5541 * and tcp_collapse() them until all the queue is collapsed.
5542 */
tcp_collapse_ofo_queue(struct sock * sk)5543 static void tcp_collapse_ofo_queue(struct sock *sk)
5544 {
5545 struct tcp_sock *tp = tcp_sk(sk);
5546 u32 range_truesize, sum_tiny = 0;
5547 struct sk_buff *skb, *head;
5548 u32 start, end;
5549
5550 skb = skb_rb_first(&tp->out_of_order_queue);
5551 new_range:
5552 if (!skb) {
5553 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5554 return;
5555 }
5556 start = TCP_SKB_CB(skb)->seq;
5557 end = TCP_SKB_CB(skb)->end_seq;
5558 range_truesize = skb->truesize;
5559
5560 for (head = skb;;) {
5561 skb = skb_rb_next(skb);
5562
5563 /* Range is terminated when we see a gap or when
5564 * we are at the queue end.
5565 */
5566 if (!skb ||
5567 after(TCP_SKB_CB(skb)->seq, end) ||
5568 before(TCP_SKB_CB(skb)->end_seq, start)) {
5569 /* Do not attempt collapsing tiny skbs */
5570 if (range_truesize != head->truesize ||
5571 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) {
5572 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5573 head, skb, start, end);
5574 } else {
5575 sum_tiny += range_truesize;
5576 if (sum_tiny > sk->sk_rcvbuf >> 3)
5577 return;
5578 }
5579 goto new_range;
5580 }
5581
5582 range_truesize += skb->truesize;
5583 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5584 start = TCP_SKB_CB(skb)->seq;
5585 if (after(TCP_SKB_CB(skb)->end_seq, end))
5586 end = TCP_SKB_CB(skb)->end_seq;
5587 }
5588 }
5589
5590 /*
5591 * Clean the out-of-order queue to make room.
5592 * We drop high sequences packets to :
5593 * 1) Let a chance for holes to be filled.
5594 * This means we do not drop packets from ooo queue if their sequence
5595 * is before incoming packet sequence.
5596 * 2) not add too big latencies if thousands of packets sit there.
5597 * (But if application shrinks SO_RCVBUF, we could still end up
5598 * freeing whole queue here)
5599 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5600 *
5601 * Return true if queue has shrunk.
5602 */
tcp_prune_ofo_queue(struct sock * sk,const struct sk_buff * in_skb)5603 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb)
5604 {
5605 struct tcp_sock *tp = tcp_sk(sk);
5606 struct rb_node *node, *prev;
5607 bool pruned = false;
5608 int goal;
5609
5610 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5611 return false;
5612
5613 goal = sk->sk_rcvbuf >> 3;
5614 node = &tp->ooo_last_skb->rbnode;
5615
5616 do {
5617 struct sk_buff *skb = rb_to_skb(node);
5618
5619 /* If incoming skb would land last in ofo queue, stop pruning. */
5620 if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq))
5621 break;
5622 pruned = true;
5623 prev = rb_prev(node);
5624 rb_erase(node, &tp->out_of_order_queue);
5625 goal -= skb->truesize;
5626 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE);
5627 tp->ooo_last_skb = rb_to_skb(prev);
5628 if (!prev || goal <= 0) {
5629 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5630 !tcp_under_memory_pressure(sk))
5631 break;
5632 goal = sk->sk_rcvbuf >> 3;
5633 }
5634 node = prev;
5635 } while (node);
5636
5637 if (pruned) {
5638 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5639 /* Reset SACK state. A conforming SACK implementation will
5640 * do the same at a timeout based retransmit. When a connection
5641 * is in a sad state like this, we care only about integrity
5642 * of the connection not performance.
5643 */
5644 if (tp->rx_opt.sack_ok)
5645 tcp_sack_reset(&tp->rx_opt);
5646 }
5647 return pruned;
5648 }
5649
5650 /* Reduce allocated memory if we can, trying to get
5651 * the socket within its memory limits again.
5652 *
5653 * Return less than zero if we should start dropping frames
5654 * until the socket owning process reads some of the data
5655 * to stabilize the situation.
5656 */
tcp_prune_queue(struct sock * sk,const struct sk_buff * in_skb)5657 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb)
5658 {
5659 struct tcp_sock *tp = tcp_sk(sk);
5660
5661 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5662
5663 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5664 tcp_clamp_window(sk);
5665 else if (tcp_under_memory_pressure(sk))
5666 tcp_adjust_rcv_ssthresh(sk);
5667
5668 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5669 return 0;
5670
5671 tcp_collapse_ofo_queue(sk);
5672 if (!skb_queue_empty(&sk->sk_receive_queue))
5673 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5674 skb_peek(&sk->sk_receive_queue),
5675 NULL,
5676 tp->copied_seq, tp->rcv_nxt);
5677
5678 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5679 return 0;
5680
5681 /* Collapsing did not help, destructive actions follow.
5682 * This must not ever occur. */
5683
5684 tcp_prune_ofo_queue(sk, in_skb);
5685
5686 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5687 return 0;
5688
5689 /* If we are really being abused, tell the caller to silently
5690 * drop receive data on the floor. It will get retransmitted
5691 * and hopefully then we'll have sufficient space.
5692 */
5693 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5694
5695 /* Massive buffer overcommit. */
5696 tp->pred_flags = 0;
5697 return -1;
5698 }
5699
tcp_should_expand_sndbuf(struct sock * sk)5700 static bool tcp_should_expand_sndbuf(struct sock *sk)
5701 {
5702 const struct tcp_sock *tp = tcp_sk(sk);
5703
5704 /* If the user specified a specific send buffer setting, do
5705 * not modify it.
5706 */
5707 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5708 return false;
5709
5710 /* If we are under global TCP memory pressure, do not expand. */
5711 if (tcp_under_memory_pressure(sk)) {
5712 int unused_mem = sk_unused_reserved_mem(sk);
5713
5714 /* Adjust sndbuf according to reserved mem. But make sure
5715 * it never goes below SOCK_MIN_SNDBUF.
5716 * See sk_stream_moderate_sndbuf() for more details.
5717 */
5718 if (unused_mem > SOCK_MIN_SNDBUF)
5719 WRITE_ONCE(sk->sk_sndbuf, unused_mem);
5720
5721 return false;
5722 }
5723
5724 /* If we are under soft global TCP memory pressure, do not expand. */
5725 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5726 return false;
5727
5728 /* If we filled the congestion window, do not expand. */
5729 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp))
5730 return false;
5731
5732 return true;
5733 }
5734
tcp_new_space(struct sock * sk)5735 static void tcp_new_space(struct sock *sk)
5736 {
5737 struct tcp_sock *tp = tcp_sk(sk);
5738
5739 if (tcp_should_expand_sndbuf(sk)) {
5740 tcp_sndbuf_expand(sk);
5741 tp->snd_cwnd_stamp = tcp_jiffies32;
5742 }
5743
5744 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
5745 }
5746
5747 /* Caller made space either from:
5748 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5749 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5750 *
5751 * We might be able to generate EPOLLOUT to the application if:
5752 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5753 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5754 * small enough that tcp_stream_memory_free() decides it
5755 * is time to generate EPOLLOUT.
5756 */
tcp_check_space(struct sock * sk)5757 void tcp_check_space(struct sock *sk)
5758 {
5759 /* pairs with tcp_poll() */
5760 smp_mb();
5761 if (sk->sk_socket &&
5762 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5763 tcp_new_space(sk);
5764 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5765 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5766 }
5767 }
5768
tcp_data_snd_check(struct sock * sk)5769 static inline void tcp_data_snd_check(struct sock *sk)
5770 {
5771 tcp_push_pending_frames(sk);
5772 tcp_check_space(sk);
5773 }
5774
5775 /*
5776 * Check if sending an ack is needed.
5777 */
__tcp_ack_snd_check(struct sock * sk,int ofo_possible)5778 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5779 {
5780 struct tcp_sock *tp = tcp_sk(sk);
5781 unsigned long rtt, delay;
5782
5783 /* More than one full frame received... */
5784 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5785 /* ... and right edge of window advances far enough.
5786 * (tcp_recvmsg() will send ACK otherwise).
5787 * If application uses SO_RCVLOWAT, we want send ack now if
5788 * we have not received enough bytes to satisfy the condition.
5789 */
5790 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5791 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5792 /* We ACK each frame or... */
5793 tcp_in_quickack_mode(sk) ||
5794 /* Protocol state mandates a one-time immediate ACK */
5795 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5796 /* If we are running from __release_sock() in user context,
5797 * Defer the ack until tcp_release_cb().
5798 */
5799 if (sock_owned_by_user_nocheck(sk) &&
5800 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_backlog_ack_defer)) {
5801 set_bit(TCP_ACK_DEFERRED, &sk->sk_tsq_flags);
5802 return;
5803 }
5804 send_now:
5805 tcp_send_ack(sk);
5806 return;
5807 }
5808
5809 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5810 tcp_send_delayed_ack(sk);
5811 return;
5812 }
5813
5814 if (!tcp_is_sack(tp) ||
5815 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr))
5816 goto send_now;
5817
5818 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5819 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5820 tp->dup_ack_counter = 0;
5821 }
5822 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5823 tp->dup_ack_counter++;
5824 goto send_now;
5825 }
5826 tp->compressed_ack++;
5827 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5828 return;
5829
5830 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5831
5832 rtt = tp->rcv_rtt_est.rtt_us;
5833 if (tp->srtt_us && tp->srtt_us < rtt)
5834 rtt = tp->srtt_us;
5835
5836 delay = min_t(unsigned long,
5837 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5838 rtt * (NSEC_PER_USEC >> 3)/20);
5839 sock_hold(sk);
5840 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5841 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns),
5842 HRTIMER_MODE_REL_PINNED_SOFT);
5843 }
5844
tcp_ack_snd_check(struct sock * sk)5845 static inline void tcp_ack_snd_check(struct sock *sk)
5846 {
5847 if (!inet_csk_ack_scheduled(sk)) {
5848 /* We sent a data segment already. */
5849 return;
5850 }
5851 __tcp_ack_snd_check(sk, 1);
5852 }
5853
5854 /*
5855 * This routine is only called when we have urgent data
5856 * signaled. Its the 'slow' part of tcp_urg. It could be
5857 * moved inline now as tcp_urg is only called from one
5858 * place. We handle URGent data wrong. We have to - as
5859 * BSD still doesn't use the correction from RFC961.
5860 * For 1003.1g we should support a new option TCP_STDURG to permit
5861 * either form (or just set the sysctl tcp_stdurg).
5862 */
5863
tcp_check_urg(struct sock * sk,const struct tcphdr * th)5864 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5865 {
5866 struct tcp_sock *tp = tcp_sk(sk);
5867 u32 ptr = ntohs(th->urg_ptr);
5868
5869 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5870 ptr--;
5871 ptr += ntohl(th->seq);
5872
5873 /* Ignore urgent data that we've already seen and read. */
5874 if (after(tp->copied_seq, ptr))
5875 return;
5876
5877 /* Do not replay urg ptr.
5878 *
5879 * NOTE: interesting situation not covered by specs.
5880 * Misbehaving sender may send urg ptr, pointing to segment,
5881 * which we already have in ofo queue. We are not able to fetch
5882 * such data and will stay in TCP_URG_NOTYET until will be eaten
5883 * by recvmsg(). Seems, we are not obliged to handle such wicked
5884 * situations. But it is worth to think about possibility of some
5885 * DoSes using some hypothetical application level deadlock.
5886 */
5887 if (before(ptr, tp->rcv_nxt))
5888 return;
5889
5890 /* Do we already have a newer (or duplicate) urgent pointer? */
5891 if (tp->urg_data && !after(ptr, tp->urg_seq))
5892 return;
5893
5894 /* Tell the world about our new urgent pointer. */
5895 sk_send_sigurg(sk);
5896
5897 /* We may be adding urgent data when the last byte read was
5898 * urgent. To do this requires some care. We cannot just ignore
5899 * tp->copied_seq since we would read the last urgent byte again
5900 * as data, nor can we alter copied_seq until this data arrives
5901 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5902 *
5903 * NOTE. Double Dutch. Rendering to plain English: author of comment
5904 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5905 * and expect that both A and B disappear from stream. This is _wrong_.
5906 * Though this happens in BSD with high probability, this is occasional.
5907 * Any application relying on this is buggy. Note also, that fix "works"
5908 * only in this artificial test. Insert some normal data between A and B and we will
5909 * decline of BSD again. Verdict: it is better to remove to trap
5910 * buggy users.
5911 */
5912 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5913 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5914 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5915 tp->copied_seq++;
5916 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5917 __skb_unlink(skb, &sk->sk_receive_queue);
5918 __kfree_skb(skb);
5919 }
5920 }
5921
5922 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET);
5923 WRITE_ONCE(tp->urg_seq, ptr);
5924
5925 /* Disable header prediction. */
5926 tp->pred_flags = 0;
5927 }
5928
5929 /* This is the 'fast' part of urgent handling. */
tcp_urg(struct sock * sk,struct sk_buff * skb,const struct tcphdr * th)5930 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5931 {
5932 struct tcp_sock *tp = tcp_sk(sk);
5933
5934 /* Check if we get a new urgent pointer - normally not. */
5935 if (unlikely(th->urg))
5936 tcp_check_urg(sk, th);
5937
5938 /* Do we wait for any urgent data? - normally not... */
5939 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) {
5940 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5941 th->syn;
5942
5943 /* Is the urgent pointer pointing into this packet? */
5944 if (ptr < skb->len) {
5945 u8 tmp;
5946 if (skb_copy_bits(skb, ptr, &tmp, 1))
5947 BUG();
5948 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp);
5949 if (!sock_flag(sk, SOCK_DEAD))
5950 sk->sk_data_ready(sk);
5951 }
5952 }
5953 }
5954
5955 /* Accept RST for rcv_nxt - 1 after a FIN.
5956 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5957 * FIN is sent followed by a RST packet. The RST is sent with the same
5958 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5959 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5960 * ACKs on the closed socket. In addition middleboxes can drop either the
5961 * challenge ACK or a subsequent RST.
5962 */
tcp_reset_check(const struct sock * sk,const struct sk_buff * skb)5963 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5964 {
5965 const struct tcp_sock *tp = tcp_sk(sk);
5966
5967 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5968 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5969 TCPF_CLOSING));
5970 }
5971
5972 /* Does PAWS and seqno based validation of an incoming segment, flags will
5973 * play significant role here.
5974 */
tcp_validate_incoming(struct sock * sk,struct sk_buff * skb,const struct tcphdr * th,int syn_inerr)5975 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5976 const struct tcphdr *th, int syn_inerr)
5977 {
5978 struct tcp_sock *tp = tcp_sk(sk);
5979 SKB_DR(reason);
5980
5981 /* RFC1323: H1. Apply PAWS check first. */
5982 if (!tcp_fast_parse_options(sock_net(sk), skb, th, tp) ||
5983 !tp->rx_opt.saw_tstamp ||
5984 tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW))
5985 goto step1;
5986
5987 reason = tcp_disordered_ack_check(sk, skb);
5988 if (!reason)
5989 goto step1;
5990 /* Reset is accepted even if it did not pass PAWS. */
5991 if (th->rst)
5992 goto step1;
5993 if (unlikely(th->syn))
5994 goto syn_challenge;
5995
5996 /* Old ACK are common, increment PAWS_OLD_ACK
5997 * and do not send a dupack.
5998 */
5999 if (reason == SKB_DROP_REASON_TCP_RFC7323_PAWS_ACK) {
6000 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWS_OLD_ACK);
6001 goto discard;
6002 }
6003 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
6004 if (!tcp_oow_rate_limited(sock_net(sk), skb,
6005 LINUX_MIB_TCPACKSKIPPEDPAWS,
6006 &tp->last_oow_ack_time))
6007 tcp_send_dupack(sk, skb);
6008 goto discard;
6009
6010 step1:
6011 /* Step 1: check sequence number */
6012 reason = tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
6013 if (reason) {
6014 /* RFC793, page 37: "In all states except SYN-SENT, all reset
6015 * (RST) segments are validated by checking their SEQ-fields."
6016 * And page 69: "If an incoming segment is not acceptable,
6017 * an acknowledgment should be sent in reply (unless the RST
6018 * bit is set, if so drop the segment and return)".
6019 */
6020 if (!th->rst) {
6021 if (th->syn)
6022 goto syn_challenge;
6023 if (!tcp_oow_rate_limited(sock_net(sk), skb,
6024 LINUX_MIB_TCPACKSKIPPEDSEQ,
6025 &tp->last_oow_ack_time))
6026 tcp_send_dupack(sk, skb);
6027 } else if (tcp_reset_check(sk, skb)) {
6028 goto reset;
6029 }
6030 goto discard;
6031 }
6032
6033 /* Step 2: check RST bit */
6034 if (th->rst) {
6035 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
6036 * FIN and SACK too if available):
6037 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
6038 * the right-most SACK block,
6039 * then
6040 * RESET the connection
6041 * else
6042 * Send a challenge ACK
6043 */
6044 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
6045 tcp_reset_check(sk, skb))
6046 goto reset;
6047
6048 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
6049 struct tcp_sack_block *sp = &tp->selective_acks[0];
6050 int max_sack = sp[0].end_seq;
6051 int this_sack;
6052
6053 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
6054 ++this_sack) {
6055 max_sack = after(sp[this_sack].end_seq,
6056 max_sack) ?
6057 sp[this_sack].end_seq : max_sack;
6058 }
6059
6060 if (TCP_SKB_CB(skb)->seq == max_sack)
6061 goto reset;
6062 }
6063
6064 /* Disable TFO if RST is out-of-order
6065 * and no data has been received
6066 * for current active TFO socket
6067 */
6068 if (tp->syn_fastopen && !tp->data_segs_in &&
6069 sk->sk_state == TCP_ESTABLISHED)
6070 tcp_fastopen_active_disable(sk);
6071 tcp_send_challenge_ack(sk);
6072 SKB_DR_SET(reason, TCP_RESET);
6073 goto discard;
6074 }
6075
6076 /* step 3: check security and precedence [ignored] */
6077
6078 /* step 4: Check for a SYN
6079 * RFC 5961 4.2 : Send a challenge ack
6080 */
6081 if (th->syn) {
6082 if (sk->sk_state == TCP_SYN_RECV && sk->sk_socket && th->ack &&
6083 TCP_SKB_CB(skb)->seq + 1 == TCP_SKB_CB(skb)->end_seq &&
6084 TCP_SKB_CB(skb)->seq + 1 == tp->rcv_nxt &&
6085 TCP_SKB_CB(skb)->ack_seq == tp->snd_nxt)
6086 goto pass;
6087 syn_challenge:
6088 if (syn_inerr)
6089 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6090 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
6091 tcp_send_challenge_ack(sk);
6092 SKB_DR_SET(reason, TCP_INVALID_SYN);
6093 goto discard;
6094 }
6095
6096 pass:
6097 bpf_skops_parse_hdr(sk, skb);
6098
6099 return true;
6100
6101 discard:
6102 tcp_drop_reason(sk, skb, reason);
6103 return false;
6104
6105 reset:
6106 tcp_reset(sk, skb);
6107 __kfree_skb(skb);
6108 return false;
6109 }
6110
6111 /*
6112 * TCP receive function for the ESTABLISHED state.
6113 *
6114 * It is split into a fast path and a slow path. The fast path is
6115 * disabled when:
6116 * - A zero window was announced from us - zero window probing
6117 * is only handled properly in the slow path.
6118 * - Out of order segments arrived.
6119 * - Urgent data is expected.
6120 * - There is no buffer space left
6121 * - Unexpected TCP flags/window values/header lengths are received
6122 * (detected by checking the TCP header against pred_flags)
6123 * - Data is sent in both directions. Fast path only supports pure senders
6124 * or pure receivers (this means either the sequence number or the ack
6125 * value must stay constant)
6126 * - Unexpected TCP option.
6127 *
6128 * When these conditions are not satisfied it drops into a standard
6129 * receive procedure patterned after RFC793 to handle all cases.
6130 * The first three cases are guaranteed by proper pred_flags setting,
6131 * the rest is checked inline. Fast processing is turned on in
6132 * tcp_data_queue when everything is OK.
6133 */
tcp_rcv_established(struct sock * sk,struct sk_buff * skb)6134 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
6135 {
6136 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED;
6137 const struct tcphdr *th = (const struct tcphdr *)skb->data;
6138 struct tcp_sock *tp = tcp_sk(sk);
6139 unsigned int len = skb->len;
6140
6141 /* TCP congestion window tracking */
6142 trace_tcp_probe(sk, skb);
6143
6144 tcp_mstamp_refresh(tp);
6145 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
6146 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
6147 /*
6148 * Header prediction.
6149 * The code loosely follows the one in the famous
6150 * "30 instruction TCP receive" Van Jacobson mail.
6151 *
6152 * Van's trick is to deposit buffers into socket queue
6153 * on a device interrupt, to call tcp_recv function
6154 * on the receive process context and checksum and copy
6155 * the buffer to user space. smart...
6156 *
6157 * Our current scheme is not silly either but we take the
6158 * extra cost of the net_bh soft interrupt processing...
6159 * We do checksum and copy also but from device to kernel.
6160 */
6161
6162 tp->rx_opt.saw_tstamp = 0;
6163
6164 /* pred_flags is 0xS?10 << 16 + snd_wnd
6165 * if header_prediction is to be made
6166 * 'S' will always be tp->tcp_header_len >> 2
6167 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
6168 * turn it off (when there are holes in the receive
6169 * space for instance)
6170 * PSH flag is ignored.
6171 */
6172
6173 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
6174 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
6175 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6176 int tcp_header_len = tp->tcp_header_len;
6177 s32 delta = 0;
6178 int flag = 0;
6179
6180 /* Timestamp header prediction: tcp_header_len
6181 * is automatically equal to th->doff*4 due to pred_flags
6182 * match.
6183 */
6184
6185 /* Check timestamp */
6186 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
6187 /* No? Slow path! */
6188 if (!tcp_parse_aligned_timestamp(tp, th))
6189 goto slow_path;
6190
6191 delta = tp->rx_opt.rcv_tsval -
6192 tp->rx_opt.ts_recent;
6193 /* If PAWS failed, check it more carefully in slow path */
6194 if (delta < 0)
6195 goto slow_path;
6196
6197 /* DO NOT update ts_recent here, if checksum fails
6198 * and timestamp was corrupted part, it will result
6199 * in a hung connection since we will drop all
6200 * future packets due to the PAWS test.
6201 */
6202 }
6203
6204 if (len <= tcp_header_len) {
6205 /* Bulk data transfer: sender */
6206 if (len == tcp_header_len) {
6207 /* Predicted packet is in window by definition.
6208 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
6209 * Hence, check seq<=rcv_wup reduces to:
6210 */
6211 if (tcp_header_len ==
6212 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
6213 tp->rcv_nxt == tp->rcv_wup)
6214 flag |= __tcp_replace_ts_recent(tp,
6215 delta);
6216
6217 /* We know that such packets are checksummed
6218 * on entry.
6219 */
6220 tcp_ack(sk, skb, flag);
6221 __kfree_skb(skb);
6222 tcp_data_snd_check(sk);
6223 /* When receiving pure ack in fast path, update
6224 * last ts ecr directly instead of calling
6225 * tcp_rcv_rtt_measure_ts()
6226 */
6227 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
6228 return;
6229 } else { /* Header too small */
6230 reason = SKB_DROP_REASON_PKT_TOO_SMALL;
6231 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6232 goto discard;
6233 }
6234 } else {
6235 int eaten = 0;
6236 bool fragstolen = false;
6237
6238 if (tcp_checksum_complete(skb))
6239 goto csum_error;
6240
6241 if ((int)skb->truesize > sk->sk_forward_alloc)
6242 goto step5;
6243
6244 /* Predicted packet is in window by definition.
6245 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
6246 * Hence, check seq<=rcv_wup reduces to:
6247 */
6248 if (tcp_header_len ==
6249 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
6250 tp->rcv_nxt == tp->rcv_wup)
6251 flag |= __tcp_replace_ts_recent(tp,
6252 delta);
6253
6254 tcp_rcv_rtt_measure_ts(sk, skb);
6255
6256 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
6257
6258 /* Bulk data transfer: receiver */
6259 tcp_cleanup_skb(skb);
6260 __skb_pull(skb, tcp_header_len);
6261 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
6262
6263 tcp_event_data_recv(sk, skb);
6264
6265 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
6266 /* Well, only one small jumplet in fast path... */
6267 tcp_ack(sk, skb, flag | FLAG_DATA);
6268 tcp_data_snd_check(sk);
6269 if (!inet_csk_ack_scheduled(sk))
6270 goto no_ack;
6271 } else {
6272 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
6273 }
6274
6275 __tcp_ack_snd_check(sk, 0);
6276 no_ack:
6277 if (eaten)
6278 kfree_skb_partial(skb, fragstolen);
6279 tcp_data_ready(sk);
6280 return;
6281 }
6282 }
6283
6284 slow_path:
6285 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
6286 goto csum_error;
6287
6288 if (!th->ack && !th->rst && !th->syn) {
6289 reason = SKB_DROP_REASON_TCP_FLAGS;
6290 goto discard;
6291 }
6292
6293 /*
6294 * Standard slow path.
6295 */
6296
6297 if (!tcp_validate_incoming(sk, skb, th, 1))
6298 return;
6299
6300 step5:
6301 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT);
6302 if ((int)reason < 0) {
6303 reason = -reason;
6304 goto discard;
6305 }
6306 tcp_rcv_rtt_measure_ts(sk, skb);
6307
6308 /* Process urgent data. */
6309 tcp_urg(sk, skb, th);
6310
6311 /* step 7: process the segment text */
6312 tcp_data_queue(sk, skb);
6313
6314 tcp_data_snd_check(sk);
6315 tcp_ack_snd_check(sk);
6316 return;
6317
6318 csum_error:
6319 reason = SKB_DROP_REASON_TCP_CSUM;
6320 trace_tcp_bad_csum(skb);
6321 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
6322 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6323
6324 discard:
6325 tcp_drop_reason(sk, skb, reason);
6326 }
6327 EXPORT_IPV6_MOD(tcp_rcv_established);
6328
tcp_init_transfer(struct sock * sk,int bpf_op,struct sk_buff * skb)6329 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
6330 {
6331 struct inet_connection_sock *icsk = inet_csk(sk);
6332 struct tcp_sock *tp = tcp_sk(sk);
6333
6334 tcp_mtup_init(sk);
6335 icsk->icsk_af_ops->rebuild_header(sk);
6336 tcp_init_metrics(sk);
6337
6338 /* Initialize the congestion window to start the transfer.
6339 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
6340 * retransmitted. In light of RFC6298 more aggressive 1sec
6341 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
6342 * retransmission has occurred.
6343 */
6344 if (tp->total_retrans > 1 && tp->undo_marker)
6345 tcp_snd_cwnd_set(tp, 1);
6346 else
6347 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk)));
6348 tp->snd_cwnd_stamp = tcp_jiffies32;
6349
6350 bpf_skops_established(sk, bpf_op, skb);
6351 /* Initialize congestion control unless BPF initialized it already: */
6352 if (!icsk->icsk_ca_initialized)
6353 tcp_init_congestion_control(sk);
6354 tcp_init_buffer_space(sk);
6355 }
6356
tcp_finish_connect(struct sock * sk,struct sk_buff * skb)6357 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
6358 {
6359 struct tcp_sock *tp = tcp_sk(sk);
6360 struct inet_connection_sock *icsk = inet_csk(sk);
6361
6362 tcp_ao_finish_connect(sk, skb);
6363 tcp_set_state(sk, TCP_ESTABLISHED);
6364 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
6365
6366 if (skb) {
6367 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
6368 security_inet_conn_established(sk, skb);
6369 sk_mark_napi_id(sk, skb);
6370 }
6371
6372 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
6373
6374 /* Prevent spurious tcp_cwnd_restart() on first data
6375 * packet.
6376 */
6377 tp->lsndtime = tcp_jiffies32;
6378
6379 if (sock_flag(sk, SOCK_KEEPOPEN))
6380 tcp_reset_keepalive_timer(sk, keepalive_time_when(tp));
6381
6382 if (!tp->rx_opt.snd_wscale)
6383 __tcp_fast_path_on(tp, tp->snd_wnd);
6384 else
6385 tp->pred_flags = 0;
6386 }
6387
tcp_rcv_fastopen_synack(struct sock * sk,struct sk_buff * synack,struct tcp_fastopen_cookie * cookie)6388 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6389 struct tcp_fastopen_cookie *cookie)
6390 {
6391 struct tcp_sock *tp = tcp_sk(sk);
6392 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
6393 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6394 bool syn_drop = false;
6395
6396 if (mss == tp->rx_opt.user_mss) {
6397 struct tcp_options_received opt;
6398
6399 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6400 tcp_clear_options(&opt);
6401 opt.user_mss = opt.mss_clamp = 0;
6402 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
6403 mss = opt.mss_clamp;
6404 }
6405
6406 if (!tp->syn_fastopen) {
6407 /* Ignore an unsolicited cookie */
6408 cookie->len = -1;
6409 } else if (tp->total_retrans) {
6410 /* SYN timed out and the SYN-ACK neither has a cookie nor
6411 * acknowledges data. Presumably the remote received only
6412 * the retransmitted (regular) SYNs: either the original
6413 * SYN-data or the corresponding SYN-ACK was dropped.
6414 */
6415 syn_drop = (cookie->len < 0 && data);
6416 } else if (cookie->len < 0 && !tp->syn_data) {
6417 /* We requested a cookie but didn't get it. If we did not use
6418 * the (old) exp opt format then try so next time (try_exp=1).
6419 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6420 */
6421 try_exp = tp->syn_fastopen_exp ? 2 : 1;
6422 }
6423
6424 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
6425
6426 if (data) { /* Retransmit unacked data in SYN */
6427 if (tp->total_retrans)
6428 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6429 else
6430 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
6431 skb_rbtree_walk_from(data)
6432 tcp_mark_skb_lost(sk, data);
6433 tcp_non_congestion_loss_retransmit(sk);
6434 NET_INC_STATS(sock_net(sk),
6435 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6436 return true;
6437 }
6438 tp->syn_data_acked = tp->syn_data;
6439 if (tp->syn_data_acked) {
6440 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6441 /* SYN-data is counted as two separate packets in tcp_ack() */
6442 if (tp->delivered > 1)
6443 --tp->delivered;
6444 }
6445
6446 tcp_fastopen_add_skb(sk, synack);
6447
6448 return false;
6449 }
6450
smc_check_reset_syn(struct tcp_sock * tp)6451 static void smc_check_reset_syn(struct tcp_sock *tp)
6452 {
6453 #if IS_ENABLED(CONFIG_SMC)
6454 if (static_branch_unlikely(&tcp_have_smc)) {
6455 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6456 tp->syn_smc = 0;
6457 }
6458 #endif
6459 }
6460
tcp_try_undo_spurious_syn(struct sock * sk)6461 static void tcp_try_undo_spurious_syn(struct sock *sk)
6462 {
6463 struct tcp_sock *tp = tcp_sk(sk);
6464 u32 syn_stamp;
6465
6466 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6467 * spurious if the ACK's timestamp option echo value matches the
6468 * original SYN timestamp.
6469 */
6470 syn_stamp = tp->retrans_stamp;
6471 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6472 syn_stamp == tp->rx_opt.rcv_tsecr)
6473 tp->undo_marker = 0;
6474 }
6475
tcp_rcv_synsent_state_process(struct sock * sk,struct sk_buff * skb,const struct tcphdr * th)6476 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6477 const struct tcphdr *th)
6478 {
6479 struct inet_connection_sock *icsk = inet_csk(sk);
6480 struct tcp_sock *tp = tcp_sk(sk);
6481 struct tcp_fastopen_cookie foc = { .len = -1 };
6482 int saved_clamp = tp->rx_opt.mss_clamp;
6483 bool fastopen_fail;
6484 SKB_DR(reason);
6485
6486 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6487 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6488 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6489
6490 if (th->ack) {
6491 /* rfc793:
6492 * "If the state is SYN-SENT then
6493 * first check the ACK bit
6494 * If the ACK bit is set
6495 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6496 * a reset (unless the RST bit is set, if so drop
6497 * the segment and return)"
6498 */
6499 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6500 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6501 /* Previous FIN/ACK or RST/ACK might be ignored. */
6502 if (icsk->icsk_retransmits == 0)
6503 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
6504 TCP_TIMEOUT_MIN, false);
6505 SKB_DR_SET(reason, TCP_INVALID_ACK_SEQUENCE);
6506 goto reset_and_undo;
6507 }
6508
6509 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6510 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6511 tcp_time_stamp_ts(tp))) {
6512 NET_INC_STATS(sock_net(sk),
6513 LINUX_MIB_PAWSACTIVEREJECTED);
6514 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6515 goto reset_and_undo;
6516 }
6517
6518 /* Now ACK is acceptable.
6519 *
6520 * "If the RST bit is set
6521 * If the ACK was acceptable then signal the user "error:
6522 * connection reset", drop the segment, enter CLOSED state,
6523 * delete TCB, and return."
6524 */
6525
6526 if (th->rst) {
6527 tcp_reset(sk, skb);
6528 consume:
6529 __kfree_skb(skb);
6530 return 0;
6531 }
6532
6533 /* rfc793:
6534 * "fifth, if neither of the SYN or RST bits is set then
6535 * drop the segment and return."
6536 *
6537 * See note below!
6538 * --ANK(990513)
6539 */
6540 if (!th->syn) {
6541 SKB_DR_SET(reason, TCP_FLAGS);
6542 goto discard_and_undo;
6543 }
6544 /* rfc793:
6545 * "If the SYN bit is on ...
6546 * are acceptable then ...
6547 * (our SYN has been ACKed), change the connection
6548 * state to ESTABLISHED..."
6549 */
6550
6551 tcp_ecn_rcv_synack(tp, th);
6552
6553 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6554 tcp_try_undo_spurious_syn(sk);
6555 tcp_ack(sk, skb, FLAG_SLOWPATH);
6556
6557 /* Ok.. it's good. Set up sequence numbers and
6558 * move to established.
6559 */
6560 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6561 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6562
6563 /* RFC1323: The window in SYN & SYN/ACK segments is
6564 * never scaled.
6565 */
6566 tp->snd_wnd = ntohs(th->window);
6567
6568 if (!tp->rx_opt.wscale_ok) {
6569 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6570 WRITE_ONCE(tp->window_clamp,
6571 min(tp->window_clamp, 65535U));
6572 }
6573
6574 if (tp->rx_opt.saw_tstamp) {
6575 tp->rx_opt.tstamp_ok = 1;
6576 tp->tcp_header_len =
6577 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6578 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6579 tcp_store_ts_recent(tp);
6580 } else {
6581 tp->tcp_header_len = sizeof(struct tcphdr);
6582 }
6583
6584 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6585 tcp_initialize_rcv_mss(sk);
6586
6587 /* Remember, tcp_poll() does not lock socket!
6588 * Change state from SYN-SENT only after copied_seq
6589 * is initialized. */
6590 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6591
6592 smc_check_reset_syn(tp);
6593
6594 smp_mb();
6595
6596 tcp_finish_connect(sk, skb);
6597
6598 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6599 tcp_rcv_fastopen_synack(sk, skb, &foc);
6600
6601 if (!sock_flag(sk, SOCK_DEAD)) {
6602 sk->sk_state_change(sk);
6603 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6604 }
6605 if (fastopen_fail)
6606 return -1;
6607 if (sk->sk_write_pending ||
6608 READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) ||
6609 inet_csk_in_pingpong_mode(sk)) {
6610 /* Save one ACK. Data will be ready after
6611 * several ticks, if write_pending is set.
6612 *
6613 * It may be deleted, but with this feature tcpdumps
6614 * look so _wonderfully_ clever, that I was not able
6615 * to stand against the temptation 8) --ANK
6616 */
6617 inet_csk_schedule_ack(sk);
6618 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6619 tcp_reset_xmit_timer(sk, ICSK_TIME_DACK,
6620 TCP_DELACK_MAX, false);
6621 goto consume;
6622 }
6623 tcp_send_ack(sk);
6624 return -1;
6625 }
6626
6627 /* No ACK in the segment */
6628
6629 if (th->rst) {
6630 /* rfc793:
6631 * "If the RST bit is set
6632 *
6633 * Otherwise (no ACK) drop the segment and return."
6634 */
6635 SKB_DR_SET(reason, TCP_RESET);
6636 goto discard_and_undo;
6637 }
6638
6639 /* PAWS check. */
6640 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6641 tcp_paws_reject(&tp->rx_opt, 0)) {
6642 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6643 goto discard_and_undo;
6644 }
6645 if (th->syn) {
6646 /* We see SYN without ACK. It is attempt of
6647 * simultaneous connect with crossed SYNs.
6648 * Particularly, it can be connect to self.
6649 */
6650 #ifdef CONFIG_TCP_AO
6651 struct tcp_ao_info *ao;
6652
6653 ao = rcu_dereference_protected(tp->ao_info,
6654 lockdep_sock_is_held(sk));
6655 if (ao) {
6656 WRITE_ONCE(ao->risn, th->seq);
6657 ao->rcv_sne = 0;
6658 }
6659 #endif
6660 tcp_set_state(sk, TCP_SYN_RECV);
6661
6662 if (tp->rx_opt.saw_tstamp) {
6663 tp->rx_opt.tstamp_ok = 1;
6664 tcp_store_ts_recent(tp);
6665 tp->tcp_header_len =
6666 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6667 } else {
6668 tp->tcp_header_len = sizeof(struct tcphdr);
6669 }
6670
6671 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6672 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6673 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6674
6675 /* RFC1323: The window in SYN & SYN/ACK segments is
6676 * never scaled.
6677 */
6678 tp->snd_wnd = ntohs(th->window);
6679 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6680 tp->max_window = tp->snd_wnd;
6681
6682 tcp_ecn_rcv_syn(tp, th);
6683
6684 tcp_mtup_init(sk);
6685 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6686 tcp_initialize_rcv_mss(sk);
6687
6688 tcp_send_synack(sk);
6689 #if 0
6690 /* Note, we could accept data and URG from this segment.
6691 * There are no obstacles to make this (except that we must
6692 * either change tcp_recvmsg() to prevent it from returning data
6693 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6694 *
6695 * However, if we ignore data in ACKless segments sometimes,
6696 * we have no reasons to accept it sometimes.
6697 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6698 * is not flawless. So, discard packet for sanity.
6699 * Uncomment this return to process the data.
6700 */
6701 return -1;
6702 #else
6703 goto consume;
6704 #endif
6705 }
6706 /* "fifth, if neither of the SYN or RST bits is set then
6707 * drop the segment and return."
6708 */
6709
6710 discard_and_undo:
6711 tcp_clear_options(&tp->rx_opt);
6712 tp->rx_opt.mss_clamp = saved_clamp;
6713 tcp_drop_reason(sk, skb, reason);
6714 return 0;
6715
6716 reset_and_undo:
6717 tcp_clear_options(&tp->rx_opt);
6718 tp->rx_opt.mss_clamp = saved_clamp;
6719 /* we can reuse/return @reason to its caller to handle the exception */
6720 return reason;
6721 }
6722
tcp_rcv_synrecv_state_fastopen(struct sock * sk)6723 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6724 {
6725 struct tcp_sock *tp = tcp_sk(sk);
6726 struct request_sock *req;
6727
6728 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6729 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6730 */
6731 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out)
6732 tcp_try_undo_recovery(sk);
6733
6734 tcp_update_rto_time(tp);
6735 inet_csk(sk)->icsk_retransmits = 0;
6736 /* In tcp_fastopen_synack_timer() on the first SYNACK RTO we set
6737 * retrans_stamp but don't enter CA_Loss, so in case that happened we
6738 * need to zero retrans_stamp here to prevent spurious
6739 * retransmits_timed_out(). However, if the ACK of our SYNACK caused us
6740 * to enter CA_Recovery then we need to leave retrans_stamp as it was
6741 * set entering CA_Recovery, for correct retransmits_timed_out() and
6742 * undo behavior.
6743 */
6744 tcp_retrans_stamp_cleanup(sk);
6745
6746 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6747 * we no longer need req so release it.
6748 */
6749 req = rcu_dereference_protected(tp->fastopen_rsk,
6750 lockdep_sock_is_held(sk));
6751 reqsk_fastopen_remove(sk, req, false);
6752
6753 /* Re-arm the timer because data may have been sent out.
6754 * This is similar to the regular data transmission case
6755 * when new data has just been ack'ed.
6756 *
6757 * (TFO) - we could try to be more aggressive and
6758 * retransmitting any data sooner based on when they
6759 * are sent out.
6760 */
6761 tcp_rearm_rto(sk);
6762 }
6763
6764 /*
6765 * This function implements the receiving procedure of RFC 793 for
6766 * all states except ESTABLISHED and TIME_WAIT.
6767 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6768 * address independent.
6769 */
6770
6771 enum skb_drop_reason
tcp_rcv_state_process(struct sock * sk,struct sk_buff * skb)6772 tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6773 {
6774 struct tcp_sock *tp = tcp_sk(sk);
6775 struct inet_connection_sock *icsk = inet_csk(sk);
6776 const struct tcphdr *th = tcp_hdr(skb);
6777 struct request_sock *req;
6778 int queued = 0;
6779 SKB_DR(reason);
6780
6781 switch (sk->sk_state) {
6782 case TCP_CLOSE:
6783 SKB_DR_SET(reason, TCP_CLOSE);
6784 goto discard;
6785
6786 case TCP_LISTEN:
6787 if (th->ack)
6788 return SKB_DROP_REASON_TCP_FLAGS;
6789
6790 if (th->rst) {
6791 SKB_DR_SET(reason, TCP_RESET);
6792 goto discard;
6793 }
6794 if (th->syn) {
6795 if (th->fin) {
6796 SKB_DR_SET(reason, TCP_FLAGS);
6797 goto discard;
6798 }
6799 /* It is possible that we process SYN packets from backlog,
6800 * so we need to make sure to disable BH and RCU right there.
6801 */
6802 rcu_read_lock();
6803 local_bh_disable();
6804 icsk->icsk_af_ops->conn_request(sk, skb);
6805 local_bh_enable();
6806 rcu_read_unlock();
6807
6808 consume_skb(skb);
6809 return 0;
6810 }
6811 SKB_DR_SET(reason, TCP_FLAGS);
6812 goto discard;
6813
6814 case TCP_SYN_SENT:
6815 tp->rx_opt.saw_tstamp = 0;
6816 tcp_mstamp_refresh(tp);
6817 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6818 if (queued >= 0)
6819 return queued;
6820
6821 /* Do step6 onward by hand. */
6822 tcp_urg(sk, skb, th);
6823 __kfree_skb(skb);
6824 tcp_data_snd_check(sk);
6825 return 0;
6826 }
6827
6828 tcp_mstamp_refresh(tp);
6829 tp->rx_opt.saw_tstamp = 0;
6830 req = rcu_dereference_protected(tp->fastopen_rsk,
6831 lockdep_sock_is_held(sk));
6832 if (req) {
6833 bool req_stolen;
6834
6835 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6836 sk->sk_state != TCP_FIN_WAIT1);
6837
6838 SKB_DR_SET(reason, TCP_FASTOPEN);
6839 if (!tcp_check_req(sk, skb, req, true, &req_stolen, &reason))
6840 goto discard;
6841 }
6842
6843 if (!th->ack && !th->rst && !th->syn) {
6844 SKB_DR_SET(reason, TCP_FLAGS);
6845 goto discard;
6846 }
6847 if (!tcp_validate_incoming(sk, skb, th, 0))
6848 return 0;
6849
6850 /* step 5: check the ACK field */
6851 reason = tcp_ack(sk, skb, FLAG_SLOWPATH |
6852 FLAG_UPDATE_TS_RECENT |
6853 FLAG_NO_CHALLENGE_ACK);
6854
6855 if ((int)reason <= 0) {
6856 if (sk->sk_state == TCP_SYN_RECV) {
6857 /* send one RST */
6858 if (!reason)
6859 return SKB_DROP_REASON_TCP_OLD_ACK;
6860 return -reason;
6861 }
6862 /* accept old ack during closing */
6863 if ((int)reason < 0) {
6864 tcp_send_challenge_ack(sk);
6865 reason = -reason;
6866 goto discard;
6867 }
6868 }
6869 SKB_DR_SET(reason, NOT_SPECIFIED);
6870 switch (sk->sk_state) {
6871 case TCP_SYN_RECV:
6872 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6873 if (!tp->srtt_us)
6874 tcp_synack_rtt_meas(sk, req);
6875
6876 if (req) {
6877 tcp_rcv_synrecv_state_fastopen(sk);
6878 } else {
6879 tcp_try_undo_spurious_syn(sk);
6880 tp->retrans_stamp = 0;
6881 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6882 skb);
6883 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6884 }
6885 tcp_ao_established(sk);
6886 smp_mb();
6887 tcp_set_state(sk, TCP_ESTABLISHED);
6888 sk->sk_state_change(sk);
6889
6890 /* Note, that this wakeup is only for marginal crossed SYN case.
6891 * Passively open sockets are not waked up, because
6892 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6893 */
6894 if (sk->sk_socket)
6895 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6896
6897 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6898 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6899 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6900
6901 if (tp->rx_opt.tstamp_ok)
6902 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6903
6904 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6905 tcp_update_pacing_rate(sk);
6906
6907 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6908 tp->lsndtime = tcp_jiffies32;
6909
6910 tcp_initialize_rcv_mss(sk);
6911 tcp_fast_path_on(tp);
6912 if (sk->sk_shutdown & SEND_SHUTDOWN)
6913 tcp_shutdown(sk, SEND_SHUTDOWN);
6914 break;
6915
6916 case TCP_FIN_WAIT1: {
6917 int tmo;
6918
6919 if (req)
6920 tcp_rcv_synrecv_state_fastopen(sk);
6921
6922 if (tp->snd_una != tp->write_seq)
6923 break;
6924
6925 tcp_set_state(sk, TCP_FIN_WAIT2);
6926 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN);
6927
6928 sk_dst_confirm(sk);
6929
6930 if (!sock_flag(sk, SOCK_DEAD)) {
6931 /* Wake up lingering close() */
6932 sk->sk_state_change(sk);
6933 break;
6934 }
6935
6936 if (READ_ONCE(tp->linger2) < 0) {
6937 tcp_done(sk);
6938 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6939 return SKB_DROP_REASON_TCP_ABORT_ON_DATA;
6940 }
6941 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6942 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6943 /* Receive out of order FIN after close() */
6944 if (tp->syn_fastopen && th->fin)
6945 tcp_fastopen_active_disable(sk);
6946 tcp_done(sk);
6947 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6948 return SKB_DROP_REASON_TCP_ABORT_ON_DATA;
6949 }
6950
6951 tmo = tcp_fin_time(sk);
6952 if (tmo > TCP_TIMEWAIT_LEN) {
6953 tcp_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6954 } else if (th->fin || sock_owned_by_user(sk)) {
6955 /* Bad case. We could lose such FIN otherwise.
6956 * It is not a big problem, but it looks confusing
6957 * and not so rare event. We still can lose it now,
6958 * if it spins in bh_lock_sock(), but it is really
6959 * marginal case.
6960 */
6961 tcp_reset_keepalive_timer(sk, tmo);
6962 } else {
6963 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6964 goto consume;
6965 }
6966 break;
6967 }
6968
6969 case TCP_CLOSING:
6970 if (tp->snd_una == tp->write_seq) {
6971 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6972 goto consume;
6973 }
6974 break;
6975
6976 case TCP_LAST_ACK:
6977 if (tp->snd_una == tp->write_seq) {
6978 tcp_update_metrics(sk);
6979 tcp_done(sk);
6980 goto consume;
6981 }
6982 break;
6983 }
6984
6985 /* step 6: check the URG bit */
6986 tcp_urg(sk, skb, th);
6987
6988 /* step 7: process the segment text */
6989 switch (sk->sk_state) {
6990 case TCP_CLOSE_WAIT:
6991 case TCP_CLOSING:
6992 case TCP_LAST_ACK:
6993 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6994 /* If a subflow has been reset, the packet should not
6995 * continue to be processed, drop the packet.
6996 */
6997 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb))
6998 goto discard;
6999 break;
7000 }
7001 fallthrough;
7002 case TCP_FIN_WAIT1:
7003 case TCP_FIN_WAIT2:
7004 /* RFC 793 says to queue data in these states,
7005 * RFC 1122 says we MUST send a reset.
7006 * BSD 4.4 also does reset.
7007 */
7008 if (sk->sk_shutdown & RCV_SHUTDOWN) {
7009 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
7010 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
7011 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
7012 tcp_reset(sk, skb);
7013 return SKB_DROP_REASON_TCP_ABORT_ON_DATA;
7014 }
7015 }
7016 fallthrough;
7017 case TCP_ESTABLISHED:
7018 tcp_data_queue(sk, skb);
7019 queued = 1;
7020 break;
7021 }
7022
7023 /* tcp_data could move socket to TIME-WAIT */
7024 if (sk->sk_state != TCP_CLOSE) {
7025 tcp_data_snd_check(sk);
7026 tcp_ack_snd_check(sk);
7027 }
7028
7029 if (!queued) {
7030 discard:
7031 tcp_drop_reason(sk, skb, reason);
7032 }
7033 return 0;
7034
7035 consume:
7036 __kfree_skb(skb);
7037 return 0;
7038 }
7039 EXPORT_IPV6_MOD(tcp_rcv_state_process);
7040
pr_drop_req(struct request_sock * req,__u16 port,int family)7041 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
7042 {
7043 struct inet_request_sock *ireq = inet_rsk(req);
7044
7045 if (family == AF_INET)
7046 net_dbg_ratelimited("drop open request from %pI4/%u\n",
7047 &ireq->ir_rmt_addr, port);
7048 #if IS_ENABLED(CONFIG_IPV6)
7049 else if (family == AF_INET6)
7050 net_dbg_ratelimited("drop open request from %pI6/%u\n",
7051 &ireq->ir_v6_rmt_addr, port);
7052 #endif
7053 }
7054
7055 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
7056 *
7057 * If we receive a SYN packet with these bits set, it means a
7058 * network is playing bad games with TOS bits. In order to
7059 * avoid possible false congestion notifications, we disable
7060 * TCP ECN negotiation.
7061 *
7062 * Exception: tcp_ca wants ECN. This is required for DCTCP
7063 * congestion control: Linux DCTCP asserts ECT on all packets,
7064 * including SYN, which is most optimal solution; however,
7065 * others, such as FreeBSD do not.
7066 *
7067 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
7068 * set, indicating the use of a future TCP extension (such as AccECN). See
7069 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
7070 * extensions.
7071 */
tcp_ecn_create_request(struct request_sock * req,const struct sk_buff * skb,const struct sock * listen_sk,const struct dst_entry * dst)7072 static void tcp_ecn_create_request(struct request_sock *req,
7073 const struct sk_buff *skb,
7074 const struct sock *listen_sk,
7075 const struct dst_entry *dst)
7076 {
7077 const struct tcphdr *th = tcp_hdr(skb);
7078 const struct net *net = sock_net(listen_sk);
7079 bool th_ecn = th->ece && th->cwr;
7080 bool ect, ecn_ok;
7081 u32 ecn_ok_dst;
7082
7083 if (!th_ecn)
7084 return;
7085
7086 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
7087 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
7088 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst;
7089
7090 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
7091 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
7092 tcp_bpf_ca_needs_ecn((struct sock *)req))
7093 inet_rsk(req)->ecn_ok = 1;
7094 }
7095
tcp_openreq_init(struct request_sock * req,const struct tcp_options_received * rx_opt,struct sk_buff * skb,const struct sock * sk)7096 static void tcp_openreq_init(struct request_sock *req,
7097 const struct tcp_options_received *rx_opt,
7098 struct sk_buff *skb, const struct sock *sk)
7099 {
7100 struct inet_request_sock *ireq = inet_rsk(req);
7101
7102 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
7103 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
7104 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
7105 tcp_rsk(req)->snt_synack = 0;
7106 tcp_rsk(req)->snt_tsval_first = 0;
7107 tcp_rsk(req)->last_oow_ack_time = 0;
7108 req->mss = rx_opt->mss_clamp;
7109 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
7110 ireq->tstamp_ok = rx_opt->tstamp_ok;
7111 ireq->sack_ok = rx_opt->sack_ok;
7112 ireq->snd_wscale = rx_opt->snd_wscale;
7113 ireq->wscale_ok = rx_opt->wscale_ok;
7114 ireq->acked = 0;
7115 ireq->ecn_ok = 0;
7116 ireq->ir_rmt_port = tcp_hdr(skb)->source;
7117 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
7118 ireq->ir_mark = inet_request_mark(sk, skb);
7119 #if IS_ENABLED(CONFIG_SMC)
7120 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested &&
7121 tcp_sk(sk)->smc_hs_congested(sk));
7122 #endif
7123 }
7124
7125 /*
7126 * Return true if a syncookie should be sent
7127 */
tcp_syn_flood_action(struct sock * sk,const char * proto)7128 static bool tcp_syn_flood_action(struct sock *sk, const char *proto)
7129 {
7130 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
7131 const char *msg = "Dropping request";
7132 struct net *net = sock_net(sk);
7133 bool want_cookie = false;
7134 u8 syncookies;
7135
7136 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
7137
7138 #ifdef CONFIG_SYN_COOKIES
7139 if (syncookies) {
7140 msg = "Sending cookies";
7141 want_cookie = true;
7142 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
7143 } else
7144 #endif
7145 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
7146
7147 if (!READ_ONCE(queue->synflood_warned) && syncookies != 2 &&
7148 xchg(&queue->synflood_warned, 1) == 0) {
7149 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) {
7150 net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n",
7151 proto, inet6_rcv_saddr(sk),
7152 sk->sk_num, msg);
7153 } else {
7154 net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n",
7155 proto, &sk->sk_rcv_saddr,
7156 sk->sk_num, msg);
7157 }
7158 }
7159
7160 return want_cookie;
7161 }
7162
tcp_reqsk_record_syn(const struct sock * sk,struct request_sock * req,const struct sk_buff * skb)7163 static void tcp_reqsk_record_syn(const struct sock *sk,
7164 struct request_sock *req,
7165 const struct sk_buff *skb)
7166 {
7167 if (tcp_sk(sk)->save_syn) {
7168 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
7169 struct saved_syn *saved_syn;
7170 u32 mac_hdrlen;
7171 void *base;
7172
7173 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
7174 base = skb_mac_header(skb);
7175 mac_hdrlen = skb_mac_header_len(skb);
7176 len += mac_hdrlen;
7177 } else {
7178 base = skb_network_header(skb);
7179 mac_hdrlen = 0;
7180 }
7181
7182 saved_syn = kmalloc(struct_size(saved_syn, data, len),
7183 GFP_ATOMIC);
7184 if (saved_syn) {
7185 saved_syn->mac_hdrlen = mac_hdrlen;
7186 saved_syn->network_hdrlen = skb_network_header_len(skb);
7187 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
7188 memcpy(saved_syn->data, base, len);
7189 req->saved_syn = saved_syn;
7190 }
7191 }
7192 }
7193
7194 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
7195 * used for SYN cookie generation.
7196 */
tcp_get_syncookie_mss(struct request_sock_ops * rsk_ops,const struct tcp_request_sock_ops * af_ops,struct sock * sk,struct tcphdr * th)7197 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
7198 const struct tcp_request_sock_ops *af_ops,
7199 struct sock *sk, struct tcphdr *th)
7200 {
7201 struct tcp_sock *tp = tcp_sk(sk);
7202 u16 mss;
7203
7204 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 &&
7205 !inet_csk_reqsk_queue_is_full(sk))
7206 return 0;
7207
7208 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
7209 return 0;
7210
7211 if (sk_acceptq_is_full(sk)) {
7212 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
7213 return 0;
7214 }
7215
7216 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
7217 if (!mss)
7218 mss = af_ops->mss_clamp;
7219
7220 return mss;
7221 }
7222 EXPORT_IPV6_MOD_GPL(tcp_get_syncookie_mss);
7223
tcp_conn_request(struct request_sock_ops * rsk_ops,const struct tcp_request_sock_ops * af_ops,struct sock * sk,struct sk_buff * skb)7224 int tcp_conn_request(struct request_sock_ops *rsk_ops,
7225 const struct tcp_request_sock_ops *af_ops,
7226 struct sock *sk, struct sk_buff *skb)
7227 {
7228 struct tcp_fastopen_cookie foc = { .len = -1 };
7229 struct tcp_options_received tmp_opt;
7230 struct tcp_sock *tp = tcp_sk(sk);
7231 struct net *net = sock_net(sk);
7232 struct sock *fastopen_sk = NULL;
7233 struct request_sock *req;
7234 bool want_cookie = false;
7235 struct dst_entry *dst;
7236 struct flowi fl;
7237 u8 syncookies;
7238 u32 isn;
7239
7240 #ifdef CONFIG_TCP_AO
7241 const struct tcp_ao_hdr *aoh;
7242 #endif
7243
7244 isn = __this_cpu_read(tcp_tw_isn);
7245 if (isn) {
7246 /* TW buckets are converted to open requests without
7247 * limitations, they conserve resources and peer is
7248 * evidently real one.
7249 */
7250 __this_cpu_write(tcp_tw_isn, 0);
7251 } else {
7252 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
7253
7254 if (syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) {
7255 want_cookie = tcp_syn_flood_action(sk,
7256 rsk_ops->slab_name);
7257 if (!want_cookie)
7258 goto drop;
7259 }
7260 }
7261
7262 if (sk_acceptq_is_full(sk)) {
7263 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
7264 goto drop;
7265 }
7266
7267 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
7268 if (!req)
7269 goto drop;
7270
7271 req->syncookie = want_cookie;
7272 tcp_rsk(req)->af_specific = af_ops;
7273 tcp_rsk(req)->ts_off = 0;
7274 tcp_rsk(req)->req_usec_ts = false;
7275 #if IS_ENABLED(CONFIG_MPTCP)
7276 tcp_rsk(req)->is_mptcp = 0;
7277 #endif
7278
7279 tcp_clear_options(&tmp_opt);
7280 tmp_opt.mss_clamp = af_ops->mss_clamp;
7281 tmp_opt.user_mss = tp->rx_opt.user_mss;
7282 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
7283 want_cookie ? NULL : &foc);
7284
7285 if (want_cookie && !tmp_opt.saw_tstamp)
7286 tcp_clear_options(&tmp_opt);
7287
7288 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
7289 tmp_opt.smc_ok = 0;
7290
7291 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
7292 tcp_openreq_init(req, &tmp_opt, skb, sk);
7293 inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk);
7294
7295 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
7296 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
7297
7298 dst = af_ops->route_req(sk, skb, &fl, req, isn);
7299 if (!dst)
7300 goto drop_and_free;
7301
7302 if (tmp_opt.tstamp_ok) {
7303 tcp_rsk(req)->req_usec_ts = dst_tcp_usec_ts(dst);
7304 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
7305 }
7306 if (!want_cookie && !isn) {
7307 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog);
7308
7309 /* Kill the following clause, if you dislike this way. */
7310 if (!syncookies &&
7311 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
7312 (max_syn_backlog >> 2)) &&
7313 !tcp_peer_is_proven(req, dst)) {
7314 /* Without syncookies last quarter of
7315 * backlog is filled with destinations,
7316 * proven to be alive.
7317 * It means that we continue to communicate
7318 * to destinations, already remembered
7319 * to the moment of synflood.
7320 */
7321 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
7322 rsk_ops->family);
7323 goto drop_and_release;
7324 }
7325
7326 isn = af_ops->init_seq(skb);
7327 }
7328
7329 tcp_ecn_create_request(req, skb, sk, dst);
7330
7331 if (want_cookie) {
7332 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
7333 if (!tmp_opt.tstamp_ok)
7334 inet_rsk(req)->ecn_ok = 0;
7335 }
7336
7337 #ifdef CONFIG_TCP_AO
7338 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh))
7339 goto drop_and_release; /* Invalid TCP options */
7340 if (aoh) {
7341 tcp_rsk(req)->used_tcp_ao = true;
7342 tcp_rsk(req)->ao_rcv_next = aoh->keyid;
7343 tcp_rsk(req)->ao_keyid = aoh->rnext_keyid;
7344
7345 } else {
7346 tcp_rsk(req)->used_tcp_ao = false;
7347 }
7348 #endif
7349 tcp_rsk(req)->snt_isn = isn;
7350 tcp_rsk(req)->txhash = net_tx_rndhash();
7351 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
7352 tcp_openreq_init_rwin(req, sk, dst);
7353 sk_rx_queue_set(req_to_sk(req), skb);
7354 if (!want_cookie) {
7355 tcp_reqsk_record_syn(sk, req, skb);
7356 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
7357 }
7358 if (fastopen_sk) {
7359 af_ops->send_synack(fastopen_sk, dst, &fl, req,
7360 &foc, TCP_SYNACK_FASTOPEN, skb);
7361 /* Add the child socket directly into the accept queue */
7362 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
7363 reqsk_fastopen_remove(fastopen_sk, req, false);
7364 bh_unlock_sock(fastopen_sk);
7365 sock_put(fastopen_sk);
7366 goto drop_and_free;
7367 }
7368 sk->sk_data_ready(sk);
7369 bh_unlock_sock(fastopen_sk);
7370 sock_put(fastopen_sk);
7371 } else {
7372 tcp_rsk(req)->tfo_listener = false;
7373 if (!want_cookie) {
7374 req->timeout = tcp_timeout_init((struct sock *)req);
7375 if (unlikely(!inet_csk_reqsk_queue_hash_add(sk, req,
7376 req->timeout))) {
7377 reqsk_free(req);
7378 dst_release(dst);
7379 return 0;
7380 }
7381
7382 }
7383 af_ops->send_synack(sk, dst, &fl, req, &foc,
7384 !want_cookie ? TCP_SYNACK_NORMAL :
7385 TCP_SYNACK_COOKIE,
7386 skb);
7387 if (want_cookie) {
7388 reqsk_free(req);
7389 return 0;
7390 }
7391 }
7392 reqsk_put(req);
7393 return 0;
7394
7395 drop_and_release:
7396 dst_release(dst);
7397 drop_and_free:
7398 __reqsk_free(req);
7399 drop:
7400 tcp_listendrop(sk);
7401 return 0;
7402 }
7403 EXPORT_IPV6_MOD(tcp_conn_request);
7404