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 <asm/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
106 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
107 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
108 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
109 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
110
111 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
112 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
113
114 #define REXMIT_NONE 0 /* no loss recovery to do */
115 #define REXMIT_LOST 1 /* retransmit packets marked lost */
116 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
117
118 #if IS_ENABLED(CONFIG_TLS_DEVICE)
119 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ);
120
clean_acked_data_enable(struct inet_connection_sock * icsk,void (* cad)(struct sock * sk,u32 ack_seq))121 void clean_acked_data_enable(struct inet_connection_sock *icsk,
122 void (*cad)(struct sock *sk, u32 ack_seq))
123 {
124 icsk->icsk_clean_acked = cad;
125 static_branch_deferred_inc(&clean_acked_data_enabled);
126 }
127 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
128
clean_acked_data_disable(struct inet_connection_sock * icsk)129 void clean_acked_data_disable(struct inet_connection_sock *icsk)
130 {
131 static_branch_slow_dec_deferred(&clean_acked_data_enabled);
132 icsk->icsk_clean_acked = NULL;
133 }
134 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
135
clean_acked_data_flush(void)136 void clean_acked_data_flush(void)
137 {
138 static_key_deferred_flush(&clean_acked_data_enabled);
139 }
140 EXPORT_SYMBOL_GPL(clean_acked_data_flush);
141 #endif
142
143 #ifdef CONFIG_CGROUP_BPF
bpf_skops_parse_hdr(struct sock * sk,struct sk_buff * skb)144 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
145 {
146 bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown &&
147 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
148 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG);
149 bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
150 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG);
151 struct bpf_sock_ops_kern sock_ops;
152
153 if (likely(!unknown_opt && !parse_all_opt))
154 return;
155
156 /* The skb will be handled in the
157 * bpf_skops_established() or
158 * bpf_skops_write_hdr_opt().
159 */
160 switch (sk->sk_state) {
161 case TCP_SYN_RECV:
162 case TCP_SYN_SENT:
163 case TCP_LISTEN:
164 return;
165 }
166
167 sock_owned_by_me(sk);
168
169 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
170 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB;
171 sock_ops.is_fullsock = 1;
172 sock_ops.sk = sk;
173 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
174
175 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
176 }
177
bpf_skops_established(struct sock * sk,int bpf_op,struct sk_buff * skb)178 static void bpf_skops_established(struct sock *sk, int bpf_op,
179 struct sk_buff *skb)
180 {
181 struct bpf_sock_ops_kern sock_ops;
182
183 sock_owned_by_me(sk);
184
185 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
186 sock_ops.op = bpf_op;
187 sock_ops.is_fullsock = 1;
188 sock_ops.sk = sk;
189 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */
190 if (skb)
191 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
192
193 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
194 }
195 #else
bpf_skops_parse_hdr(struct sock * sk,struct sk_buff * skb)196 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
197 {
198 }
199
bpf_skops_established(struct sock * sk,int bpf_op,struct sk_buff * skb)200 static void bpf_skops_established(struct sock *sk, int bpf_op,
201 struct sk_buff *skb)
202 {
203 }
204 #endif
205
tcp_gro_dev_warn(const struct sock * sk,const struct sk_buff * skb,unsigned int len)206 static __cold void tcp_gro_dev_warn(const struct sock *sk, const struct sk_buff *skb,
207 unsigned int len)
208 {
209 struct net_device *dev;
210
211 rcu_read_lock();
212 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
213 if (!dev || len >= READ_ONCE(dev->mtu))
214 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
215 dev ? dev->name : "Unknown driver");
216 rcu_read_unlock();
217 }
218
219 /* Adapt the MSS value used to make delayed ack decision to the
220 * real world.
221 */
tcp_measure_rcv_mss(struct sock * sk,const struct sk_buff * skb)222 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
223 {
224 struct inet_connection_sock *icsk = inet_csk(sk);
225 const unsigned int lss = icsk->icsk_ack.last_seg_size;
226 unsigned int len;
227
228 icsk->icsk_ack.last_seg_size = 0;
229
230 /* skb->len may jitter because of SACKs, even if peer
231 * sends good full-sized frames.
232 */
233 len = skb_shinfo(skb)->gso_size ? : skb->len;
234 if (len >= icsk->icsk_ack.rcv_mss) {
235 /* Note: divides are still a bit expensive.
236 * For the moment, only adjust scaling_ratio
237 * when we update icsk_ack.rcv_mss.
238 */
239 if (unlikely(len != icsk->icsk_ack.rcv_mss)) {
240 u64 val = (u64)skb->len << TCP_RMEM_TO_WIN_SCALE;
241
242 do_div(val, skb->truesize);
243 tcp_sk(sk)->scaling_ratio = val ? val : 1;
244 }
245 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
246 tcp_sk(sk)->advmss);
247 /* Account for possibly-removed options */
248 DO_ONCE_LITE_IF(len > icsk->icsk_ack.rcv_mss + MAX_TCP_OPTION_SPACE,
249 tcp_gro_dev_warn, sk, skb, len);
250 /* If the skb has a len of exactly 1*MSS and has the PSH bit
251 * set then it is likely the end of an application write. So
252 * more data may not be arriving soon, and yet the data sender
253 * may be waiting for an ACK if cwnd-bound or using TX zero
254 * copy. So we set ICSK_ACK_PUSHED here so that
255 * tcp_cleanup_rbuf() will send an ACK immediately if the app
256 * reads all of the data and is not ping-pong. If len > MSS
257 * then this logic does not matter (and does not hurt) because
258 * tcp_cleanup_rbuf() will always ACK immediately if the app
259 * reads data and there is more than an MSS of unACKed data.
260 */
261 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH)
262 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
263 } else {
264 /* Otherwise, we make more careful check taking into account,
265 * that SACKs block is variable.
266 *
267 * "len" is invariant segment length, including TCP header.
268 */
269 len += skb->data - skb_transport_header(skb);
270 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
271 /* If PSH is not set, packet should be
272 * full sized, provided peer TCP is not badly broken.
273 * This observation (if it is correct 8)) allows
274 * to handle super-low mtu links fairly.
275 */
276 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
277 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
278 /* Subtract also invariant (if peer is RFC compliant),
279 * tcp header plus fixed timestamp option length.
280 * Resulting "len" is MSS free of SACK jitter.
281 */
282 len -= tcp_sk(sk)->tcp_header_len;
283 icsk->icsk_ack.last_seg_size = len;
284 if (len == lss) {
285 icsk->icsk_ack.rcv_mss = len;
286 return;
287 }
288 }
289 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
290 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
291 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
292 }
293 }
294
tcp_incr_quickack(struct sock * sk,unsigned int max_quickacks)295 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
296 {
297 struct inet_connection_sock *icsk = inet_csk(sk);
298 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
299
300 if (quickacks == 0)
301 quickacks = 2;
302 quickacks = min(quickacks, max_quickacks);
303 if (quickacks > icsk->icsk_ack.quick)
304 icsk->icsk_ack.quick = quickacks;
305 }
306
tcp_enter_quickack_mode(struct sock * sk,unsigned int max_quickacks)307 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
308 {
309 struct inet_connection_sock *icsk = inet_csk(sk);
310
311 tcp_incr_quickack(sk, max_quickacks);
312 inet_csk_exit_pingpong_mode(sk);
313 icsk->icsk_ack.ato = TCP_ATO_MIN;
314 }
315
316 /* Send ACKs quickly, if "quick" count is not exhausted
317 * and the session is not interactive.
318 */
319
tcp_in_quickack_mode(struct sock * sk)320 static bool tcp_in_quickack_mode(struct sock *sk)
321 {
322 const struct inet_connection_sock *icsk = inet_csk(sk);
323 const struct dst_entry *dst = __sk_dst_get(sk);
324
325 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
326 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
327 }
328
tcp_ecn_queue_cwr(struct tcp_sock * tp)329 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
330 {
331 if (tp->ecn_flags & TCP_ECN_OK)
332 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
333 }
334
tcp_ecn_accept_cwr(struct sock * sk,const struct sk_buff * skb)335 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
336 {
337 if (tcp_hdr(skb)->cwr) {
338 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
339
340 /* If the sender is telling us it has entered CWR, then its
341 * cwnd may be very low (even just 1 packet), so we should ACK
342 * immediately.
343 */
344 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
345 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
346 }
347 }
348
tcp_ecn_withdraw_cwr(struct tcp_sock * tp)349 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
350 {
351 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
352 }
353
__tcp_ecn_check_ce(struct sock * sk,const struct sk_buff * skb)354 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
355 {
356 struct tcp_sock *tp = tcp_sk(sk);
357
358 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
359 case INET_ECN_NOT_ECT:
360 /* Funny extension: if ECT is not set on a segment,
361 * and we already seen ECT on a previous segment,
362 * it is probably a retransmit.
363 */
364 if (tp->ecn_flags & TCP_ECN_SEEN)
365 tcp_enter_quickack_mode(sk, 2);
366 break;
367 case INET_ECN_CE:
368 if (tcp_ca_needs_ecn(sk))
369 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
370
371 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
372 /* Better not delay acks, sender can have a very low cwnd */
373 tcp_enter_quickack_mode(sk, 2);
374 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
375 }
376 tp->ecn_flags |= TCP_ECN_SEEN;
377 break;
378 default:
379 if (tcp_ca_needs_ecn(sk))
380 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
381 tp->ecn_flags |= TCP_ECN_SEEN;
382 break;
383 }
384 }
385
tcp_ecn_check_ce(struct sock * sk,const struct sk_buff * skb)386 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
387 {
388 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
389 __tcp_ecn_check_ce(sk, skb);
390 }
391
tcp_ecn_rcv_synack(struct tcp_sock * tp,const struct tcphdr * th)392 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
393 {
394 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
395 tp->ecn_flags &= ~TCP_ECN_OK;
396 }
397
tcp_ecn_rcv_syn(struct tcp_sock * tp,const struct tcphdr * th)398 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
399 {
400 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
401 tp->ecn_flags &= ~TCP_ECN_OK;
402 }
403
tcp_ecn_rcv_ecn_echo(const struct tcp_sock * tp,const struct tcphdr * th)404 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
405 {
406 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
407 return true;
408 return false;
409 }
410
411 /* Buffer size and advertised window tuning.
412 *
413 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
414 */
415
tcp_sndbuf_expand(struct sock * sk)416 static void tcp_sndbuf_expand(struct sock *sk)
417 {
418 const struct tcp_sock *tp = tcp_sk(sk);
419 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
420 int sndmem, per_mss;
421 u32 nr_segs;
422
423 /* Worst case is non GSO/TSO : each frame consumes one skb
424 * and skb->head is kmalloced using power of two area of memory
425 */
426 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
427 MAX_TCP_HEADER +
428 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
429
430 per_mss = roundup_pow_of_two(per_mss) +
431 SKB_DATA_ALIGN(sizeof(struct sk_buff));
432
433 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp));
434 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
435
436 /* Fast Recovery (RFC 5681 3.2) :
437 * Cubic needs 1.7 factor, rounded to 2 to include
438 * extra cushion (application might react slowly to EPOLLOUT)
439 */
440 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
441 sndmem *= nr_segs * per_mss;
442
443 if (sk->sk_sndbuf < sndmem)
444 WRITE_ONCE(sk->sk_sndbuf,
445 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2])));
446 }
447
448 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
449 *
450 * All tcp_full_space() is split to two parts: "network" buffer, allocated
451 * forward and advertised in receiver window (tp->rcv_wnd) and
452 * "application buffer", required to isolate scheduling/application
453 * latencies from network.
454 * window_clamp is maximal advertised window. It can be less than
455 * tcp_full_space(), in this case tcp_full_space() - window_clamp
456 * is reserved for "application" buffer. The less window_clamp is
457 * the smoother our behaviour from viewpoint of network, but the lower
458 * throughput and the higher sensitivity of the connection to losses. 8)
459 *
460 * rcv_ssthresh is more strict window_clamp used at "slow start"
461 * phase to predict further behaviour of this connection.
462 * It is used for two goals:
463 * - to enforce header prediction at sender, even when application
464 * requires some significant "application buffer". It is check #1.
465 * - to prevent pruning of receive queue because of misprediction
466 * of receiver window. Check #2.
467 *
468 * The scheme does not work when sender sends good segments opening
469 * window and then starts to feed us spaghetti. But it should work
470 * in common situations. Otherwise, we have to rely on queue collapsing.
471 */
472
473 /* Slow part of check#2. */
__tcp_grow_window(const struct sock * sk,const struct sk_buff * skb,unsigned int skbtruesize)474 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb,
475 unsigned int skbtruesize)
476 {
477 const struct tcp_sock *tp = tcp_sk(sk);
478 /* Optimize this! */
479 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1;
480 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1;
481
482 while (tp->rcv_ssthresh <= window) {
483 if (truesize <= skb->len)
484 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
485
486 truesize >>= 1;
487 window >>= 1;
488 }
489 return 0;
490 }
491
492 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing
493 * can play nice with us, as sk_buff and skb->head might be either
494 * freed or shared with up to MAX_SKB_FRAGS segments.
495 * Only give a boost to drivers using page frag(s) to hold the frame(s),
496 * and if no payload was pulled in skb->head before reaching us.
497 */
truesize_adjust(bool adjust,const struct sk_buff * skb)498 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb)
499 {
500 u32 truesize = skb->truesize;
501
502 if (adjust && !skb_headlen(skb)) {
503 truesize -= SKB_TRUESIZE(skb_end_offset(skb));
504 /* paranoid check, some drivers might be buggy */
505 if (unlikely((int)truesize < (int)skb->len))
506 truesize = skb->truesize;
507 }
508 return truesize;
509 }
510
tcp_grow_window(struct sock * sk,const struct sk_buff * skb,bool adjust)511 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb,
512 bool adjust)
513 {
514 struct tcp_sock *tp = tcp_sk(sk);
515 int room;
516
517 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
518
519 if (room <= 0)
520 return;
521
522 /* Check #1 */
523 if (!tcp_under_memory_pressure(sk)) {
524 unsigned int truesize = truesize_adjust(adjust, skb);
525 int incr;
526
527 /* Check #2. Increase window, if skb with such overhead
528 * will fit to rcvbuf in future.
529 */
530 if (tcp_win_from_space(sk, truesize) <= skb->len)
531 incr = 2 * tp->advmss;
532 else
533 incr = __tcp_grow_window(sk, skb, truesize);
534
535 if (incr) {
536 incr = max_t(int, incr, 2 * skb->len);
537 tp->rcv_ssthresh += min(room, incr);
538 inet_csk(sk)->icsk_ack.quick |= 1;
539 }
540 } else {
541 /* Under pressure:
542 * Adjust rcv_ssthresh according to reserved mem
543 */
544 tcp_adjust_rcv_ssthresh(sk);
545 }
546 }
547
548 /* 3. Try to fixup all. It is made immediately after connection enters
549 * established state.
550 */
tcp_init_buffer_space(struct sock * sk)551 static void tcp_init_buffer_space(struct sock *sk)
552 {
553 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win);
554 struct tcp_sock *tp = tcp_sk(sk);
555 int maxwin;
556
557 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
558 tcp_sndbuf_expand(sk);
559
560 tcp_mstamp_refresh(tp);
561 tp->rcvq_space.time = tp->tcp_mstamp;
562 tp->rcvq_space.seq = tp->copied_seq;
563
564 maxwin = tcp_full_space(sk);
565
566 if (tp->window_clamp >= maxwin) {
567 WRITE_ONCE(tp->window_clamp, maxwin);
568
569 if (tcp_app_win && maxwin > 4 * tp->advmss)
570 WRITE_ONCE(tp->window_clamp,
571 max(maxwin - (maxwin >> tcp_app_win),
572 4 * tp->advmss));
573 }
574
575 /* Force reservation of one segment. */
576 if (tcp_app_win &&
577 tp->window_clamp > 2 * tp->advmss &&
578 tp->window_clamp + tp->advmss > maxwin)
579 WRITE_ONCE(tp->window_clamp,
580 max(2 * tp->advmss, maxwin - tp->advmss));
581
582 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
583 tp->snd_cwnd_stamp = tcp_jiffies32;
584 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
585 (u32)TCP_INIT_CWND * tp->advmss);
586 }
587
588 /* 4. Recalculate window clamp after socket hit its memory bounds. */
tcp_clamp_window(struct sock * sk)589 static void tcp_clamp_window(struct sock *sk)
590 {
591 struct tcp_sock *tp = tcp_sk(sk);
592 struct inet_connection_sock *icsk = inet_csk(sk);
593 struct net *net = sock_net(sk);
594 int rmem2;
595
596 icsk->icsk_ack.quick = 0;
597 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
598
599 if (sk->sk_rcvbuf < rmem2 &&
600 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
601 !tcp_under_memory_pressure(sk) &&
602 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
603 WRITE_ONCE(sk->sk_rcvbuf,
604 min(atomic_read(&sk->sk_rmem_alloc), rmem2));
605 }
606 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
607 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
608 }
609
610 /* Initialize RCV_MSS value.
611 * RCV_MSS is an our guess about MSS used by the peer.
612 * We haven't any direct information about the MSS.
613 * It's better to underestimate the RCV_MSS rather than overestimate.
614 * Overestimations make us ACKing less frequently than needed.
615 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
616 */
tcp_initialize_rcv_mss(struct sock * sk)617 void tcp_initialize_rcv_mss(struct sock *sk)
618 {
619 const struct tcp_sock *tp = tcp_sk(sk);
620 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
621
622 hint = min(hint, tp->rcv_wnd / 2);
623 hint = min(hint, TCP_MSS_DEFAULT);
624 hint = max(hint, TCP_MIN_MSS);
625
626 inet_csk(sk)->icsk_ack.rcv_mss = hint;
627 }
628 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
629
630 /* Receiver "autotuning" code.
631 *
632 * The algorithm for RTT estimation w/o timestamps is based on
633 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
634 * <https://public.lanl.gov/radiant/pubs.html#DRS>
635 *
636 * More detail on this code can be found at
637 * <http://staff.psc.edu/jheffner/>,
638 * though this reference is out of date. A new paper
639 * is pending.
640 */
tcp_rcv_rtt_update(struct tcp_sock * tp,u32 sample,int win_dep)641 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
642 {
643 u32 new_sample = tp->rcv_rtt_est.rtt_us;
644 long m = sample;
645
646 if (new_sample != 0) {
647 /* If we sample in larger samples in the non-timestamp
648 * case, we could grossly overestimate the RTT especially
649 * with chatty applications or bulk transfer apps which
650 * are stalled on filesystem I/O.
651 *
652 * Also, since we are only going for a minimum in the
653 * non-timestamp case, we do not smooth things out
654 * else with timestamps disabled convergence takes too
655 * long.
656 */
657 if (!win_dep) {
658 m -= (new_sample >> 3);
659 new_sample += m;
660 } else {
661 m <<= 3;
662 if (m < new_sample)
663 new_sample = m;
664 }
665 } else {
666 /* No previous measure. */
667 new_sample = m << 3;
668 }
669
670 tp->rcv_rtt_est.rtt_us = new_sample;
671 }
672
tcp_rcv_rtt_measure(struct tcp_sock * tp)673 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
674 {
675 u32 delta_us;
676
677 if (tp->rcv_rtt_est.time == 0)
678 goto new_measure;
679 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
680 return;
681 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
682 if (!delta_us)
683 delta_us = 1;
684 tcp_rcv_rtt_update(tp, delta_us, 1);
685
686 new_measure:
687 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
688 tp->rcv_rtt_est.time = tp->tcp_mstamp;
689 }
690
tcp_rtt_tsopt_us(const struct tcp_sock * tp)691 static s32 tcp_rtt_tsopt_us(const struct tcp_sock *tp)
692 {
693 u32 delta, delta_us;
694
695 delta = tcp_time_stamp_ts(tp) - tp->rx_opt.rcv_tsecr;
696 if (tp->tcp_usec_ts)
697 return delta;
698
699 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
700 if (!delta)
701 delta = 1;
702 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
703 return delta_us;
704 }
705 return -1;
706 }
707
tcp_rcv_rtt_measure_ts(struct sock * sk,const struct sk_buff * skb)708 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
709 const struct sk_buff *skb)
710 {
711 struct tcp_sock *tp = tcp_sk(sk);
712
713 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
714 return;
715 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
716
717 if (TCP_SKB_CB(skb)->end_seq -
718 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
719 s32 delta = tcp_rtt_tsopt_us(tp);
720
721 if (delta >= 0)
722 tcp_rcv_rtt_update(tp, delta, 0);
723 }
724 }
725
726 /*
727 * This function should be called every time data is copied to user space.
728 * It calculates the appropriate TCP receive buffer space.
729 */
tcp_rcv_space_adjust(struct sock * sk)730 void tcp_rcv_space_adjust(struct sock *sk)
731 {
732 struct tcp_sock *tp = tcp_sk(sk);
733 u32 copied;
734 int time;
735
736 trace_tcp_rcv_space_adjust(sk);
737
738 tcp_mstamp_refresh(tp);
739 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
740 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
741 return;
742
743 /* Number of bytes copied to user in last RTT */
744 copied = tp->copied_seq - tp->rcvq_space.seq;
745 if (copied <= tp->rcvq_space.space)
746 goto new_measure;
747
748 /* A bit of theory :
749 * copied = bytes received in previous RTT, our base window
750 * To cope with packet losses, we need a 2x factor
751 * To cope with slow start, and sender growing its cwin by 100 %
752 * every RTT, we need a 4x factor, because the ACK we are sending
753 * now is for the next RTT, not the current one :
754 * <prev RTT . ><current RTT .. ><next RTT .... >
755 */
756
757 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
758 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
759 u64 rcvwin, grow;
760 int rcvbuf;
761
762 /* minimal window to cope with packet losses, assuming
763 * steady state. Add some cushion because of small variations.
764 */
765 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
766
767 /* Accommodate for sender rate increase (eg. slow start) */
768 grow = rcvwin * (copied - tp->rcvq_space.space);
769 do_div(grow, tp->rcvq_space.space);
770 rcvwin += (grow << 1);
771
772 rcvbuf = min_t(u64, tcp_space_from_win(sk, rcvwin),
773 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
774 if (rcvbuf > sk->sk_rcvbuf) {
775 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
776
777 /* Make the window clamp follow along. */
778 WRITE_ONCE(tp->window_clamp,
779 tcp_win_from_space(sk, rcvbuf));
780 }
781 }
782 tp->rcvq_space.space = copied;
783
784 new_measure:
785 tp->rcvq_space.seq = tp->copied_seq;
786 tp->rcvq_space.time = tp->tcp_mstamp;
787 }
788
tcp_save_lrcv_flowlabel(struct sock * sk,const struct sk_buff * skb)789 static void tcp_save_lrcv_flowlabel(struct sock *sk, const struct sk_buff *skb)
790 {
791 #if IS_ENABLED(CONFIG_IPV6)
792 struct inet_connection_sock *icsk = inet_csk(sk);
793
794 if (skb->protocol == htons(ETH_P_IPV6))
795 icsk->icsk_ack.lrcv_flowlabel = ntohl(ip6_flowlabel(ipv6_hdr(skb)));
796 #endif
797 }
798
799 /* There is something which you must keep in mind when you analyze the
800 * behavior of the tp->ato delayed ack timeout interval. When a
801 * connection starts up, we want to ack as quickly as possible. The
802 * problem is that "good" TCP's do slow start at the beginning of data
803 * transmission. The means that until we send the first few ACK's the
804 * sender will sit on his end and only queue most of his data, because
805 * he can only send snd_cwnd unacked packets at any given time. For
806 * each ACK we send, he increments snd_cwnd and transmits more of his
807 * queue. -DaveM
808 */
tcp_event_data_recv(struct sock * sk,struct sk_buff * skb)809 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
810 {
811 struct tcp_sock *tp = tcp_sk(sk);
812 struct inet_connection_sock *icsk = inet_csk(sk);
813 u32 now;
814
815 inet_csk_schedule_ack(sk);
816
817 tcp_measure_rcv_mss(sk, skb);
818
819 tcp_rcv_rtt_measure(tp);
820
821 now = tcp_jiffies32;
822
823 if (!icsk->icsk_ack.ato) {
824 /* The _first_ data packet received, initialize
825 * delayed ACK engine.
826 */
827 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
828 icsk->icsk_ack.ato = TCP_ATO_MIN;
829 } else {
830 int m = now - icsk->icsk_ack.lrcvtime;
831
832 if (m <= TCP_ATO_MIN / 2) {
833 /* The fastest case is the first. */
834 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
835 } else if (m < icsk->icsk_ack.ato) {
836 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
837 if (icsk->icsk_ack.ato > icsk->icsk_rto)
838 icsk->icsk_ack.ato = icsk->icsk_rto;
839 } else if (m > icsk->icsk_rto) {
840 /* Too long gap. Apparently sender failed to
841 * restart window, so that we send ACKs quickly.
842 */
843 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
844 }
845 }
846 icsk->icsk_ack.lrcvtime = now;
847 tcp_save_lrcv_flowlabel(sk, skb);
848
849 tcp_ecn_check_ce(sk, skb);
850
851 if (skb->len >= 128)
852 tcp_grow_window(sk, skb, true);
853 }
854
855 /* Called to compute a smoothed rtt estimate. The data fed to this
856 * routine either comes from timestamps, or from segments that were
857 * known _not_ to have been retransmitted [see Karn/Partridge
858 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
859 * piece by Van Jacobson.
860 * NOTE: the next three routines used to be one big routine.
861 * To save cycles in the RFC 1323 implementation it was better to break
862 * it up into three procedures. -- erics
863 */
tcp_rtt_estimator(struct sock * sk,long mrtt_us)864 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
865 {
866 struct tcp_sock *tp = tcp_sk(sk);
867 long m = mrtt_us; /* RTT */
868 u32 srtt = tp->srtt_us;
869
870 /* The following amusing code comes from Jacobson's
871 * article in SIGCOMM '88. Note that rtt and mdev
872 * are scaled versions of rtt and mean deviation.
873 * This is designed to be as fast as possible
874 * m stands for "measurement".
875 *
876 * On a 1990 paper the rto value is changed to:
877 * RTO = rtt + 4 * mdev
878 *
879 * Funny. This algorithm seems to be very broken.
880 * These formulae increase RTO, when it should be decreased, increase
881 * too slowly, when it should be increased quickly, decrease too quickly
882 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
883 * does not matter how to _calculate_ it. Seems, it was trap
884 * that VJ failed to avoid. 8)
885 */
886 if (srtt != 0) {
887 m -= (srtt >> 3); /* m is now error in rtt est */
888 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
889 if (m < 0) {
890 m = -m; /* m is now abs(error) */
891 m -= (tp->mdev_us >> 2); /* similar update on mdev */
892 /* This is similar to one of Eifel findings.
893 * Eifel blocks mdev updates when rtt decreases.
894 * This solution is a bit different: we use finer gain
895 * for mdev in this case (alpha*beta).
896 * Like Eifel it also prevents growth of rto,
897 * but also it limits too fast rto decreases,
898 * happening in pure Eifel.
899 */
900 if (m > 0)
901 m >>= 3;
902 } else {
903 m -= (tp->mdev_us >> 2); /* similar update on mdev */
904 }
905 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
906 if (tp->mdev_us > tp->mdev_max_us) {
907 tp->mdev_max_us = tp->mdev_us;
908 if (tp->mdev_max_us > tp->rttvar_us)
909 tp->rttvar_us = tp->mdev_max_us;
910 }
911 if (after(tp->snd_una, tp->rtt_seq)) {
912 if (tp->mdev_max_us < tp->rttvar_us)
913 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
914 tp->rtt_seq = tp->snd_nxt;
915 tp->mdev_max_us = tcp_rto_min_us(sk);
916
917 tcp_bpf_rtt(sk, mrtt_us, srtt);
918 }
919 } else {
920 /* no previous measure. */
921 srtt = m << 3; /* take the measured time to be rtt */
922 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
923 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
924 tp->mdev_max_us = tp->rttvar_us;
925 tp->rtt_seq = tp->snd_nxt;
926
927 tcp_bpf_rtt(sk, mrtt_us, srtt);
928 }
929 tp->srtt_us = max(1U, srtt);
930 }
931
tcp_update_pacing_rate(struct sock * sk)932 static void tcp_update_pacing_rate(struct sock *sk)
933 {
934 const struct tcp_sock *tp = tcp_sk(sk);
935 u64 rate;
936
937 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
938 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
939
940 /* current rate is (cwnd * mss) / srtt
941 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
942 * In Congestion Avoidance phase, set it to 120 % the current rate.
943 *
944 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
945 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
946 * end of slow start and should slow down.
947 */
948 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2)
949 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio);
950 else
951 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio);
952
953 rate *= max(tcp_snd_cwnd(tp), tp->packets_out);
954
955 if (likely(tp->srtt_us))
956 do_div(rate, tp->srtt_us);
957
958 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
959 * without any lock. We want to make sure compiler wont store
960 * intermediate values in this location.
961 */
962 WRITE_ONCE(sk->sk_pacing_rate,
963 min_t(u64, rate, READ_ONCE(sk->sk_max_pacing_rate)));
964 }
965
966 /* Calculate rto without backoff. This is the second half of Van Jacobson's
967 * routine referred to above.
968 */
tcp_set_rto(struct sock * sk)969 static void tcp_set_rto(struct sock *sk)
970 {
971 const struct tcp_sock *tp = tcp_sk(sk);
972 /* Old crap is replaced with new one. 8)
973 *
974 * More seriously:
975 * 1. If rtt variance happened to be less 50msec, it is hallucination.
976 * It cannot be less due to utterly erratic ACK generation made
977 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
978 * to do with delayed acks, because at cwnd>2 true delack timeout
979 * is invisible. Actually, Linux-2.4 also generates erratic
980 * ACKs in some circumstances.
981 */
982 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
983
984 /* 2. Fixups made earlier cannot be right.
985 * If we do not estimate RTO correctly without them,
986 * all the algo is pure shit and should be replaced
987 * with correct one. It is exactly, which we pretend to do.
988 */
989
990 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
991 * guarantees that rto is higher.
992 */
993 tcp_bound_rto(sk);
994 }
995
tcp_init_cwnd(const struct tcp_sock * tp,const struct dst_entry * dst)996 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
997 {
998 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
999
1000 if (!cwnd)
1001 cwnd = TCP_INIT_CWND;
1002 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
1003 }
1004
1005 struct tcp_sacktag_state {
1006 /* Timestamps for earliest and latest never-retransmitted segment
1007 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1008 * but congestion control should still get an accurate delay signal.
1009 */
1010 u64 first_sackt;
1011 u64 last_sackt;
1012 u32 reord;
1013 u32 sack_delivered;
1014 int flag;
1015 unsigned int mss_now;
1016 struct rate_sample *rate;
1017 };
1018
1019 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
1020 * and spurious retransmission information if this DSACK is unlikely caused by
1021 * sender's action:
1022 * - DSACKed sequence range is larger than maximum receiver's window.
1023 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
1024 */
tcp_dsack_seen(struct tcp_sock * tp,u32 start_seq,u32 end_seq,struct tcp_sacktag_state * state)1025 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
1026 u32 end_seq, struct tcp_sacktag_state *state)
1027 {
1028 u32 seq_len, dup_segs = 1;
1029
1030 if (!before(start_seq, end_seq))
1031 return 0;
1032
1033 seq_len = end_seq - start_seq;
1034 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
1035 if (seq_len > tp->max_window)
1036 return 0;
1037 if (seq_len > tp->mss_cache)
1038 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
1039 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq)
1040 state->flag |= FLAG_DSACK_TLP;
1041
1042 tp->dsack_dups += dup_segs;
1043 /* Skip the DSACK if dup segs weren't retransmitted by sender */
1044 if (tp->dsack_dups > tp->total_retrans)
1045 return 0;
1046
1047 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
1048 /* We increase the RACK ordering window in rounds where we receive
1049 * DSACKs that may have been due to reordering causing RACK to trigger
1050 * a spurious fast recovery. Thus RACK ignores DSACKs that happen
1051 * without having seen reordering, or that match TLP probes (TLP
1052 * is timer-driven, not triggered by RACK).
1053 */
1054 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP))
1055 tp->rack.dsack_seen = 1;
1056
1057 state->flag |= FLAG_DSACKING_ACK;
1058 /* A spurious retransmission is delivered */
1059 state->sack_delivered += dup_segs;
1060
1061 return dup_segs;
1062 }
1063
1064 /* It's reordering when higher sequence was delivered (i.e. sacked) before
1065 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1066 * distance is approximated in full-mss packet distance ("reordering").
1067 */
tcp_check_sack_reordering(struct sock * sk,const u32 low_seq,const int ts)1068 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1069 const int ts)
1070 {
1071 struct tcp_sock *tp = tcp_sk(sk);
1072 const u32 mss = tp->mss_cache;
1073 u32 fack, metric;
1074
1075 fack = tcp_highest_sack_seq(tp);
1076 if (!before(low_seq, fack))
1077 return;
1078
1079 metric = fack - low_seq;
1080 if ((metric > tp->reordering * mss) && mss) {
1081 #if FASTRETRANS_DEBUG > 1
1082 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1083 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1084 tp->reordering,
1085 0,
1086 tp->sacked_out,
1087 tp->undo_marker ? tp->undo_retrans : 0);
1088 #endif
1089 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1090 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1091 }
1092
1093 /* This exciting event is worth to be remembered. 8) */
1094 tp->reord_seen++;
1095 NET_INC_STATS(sock_net(sk),
1096 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1097 }
1098
1099 /* This must be called before lost_out or retrans_out are updated
1100 * on a new loss, because we want to know if all skbs previously
1101 * known to be lost have already been retransmitted, indicating
1102 * that this newly lost skb is our next skb to retransmit.
1103 */
tcp_verify_retransmit_hint(struct tcp_sock * tp,struct sk_buff * skb)1104 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1105 {
1106 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1107 (tp->retransmit_skb_hint &&
1108 before(TCP_SKB_CB(skb)->seq,
1109 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1110 tp->retransmit_skb_hint = skb;
1111 }
1112
1113 /* Sum the number of packets on the wire we have marked as lost, and
1114 * notify the congestion control module that the given skb was marked lost.
1115 */
tcp_notify_skb_loss_event(struct tcp_sock * tp,const struct sk_buff * skb)1116 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1117 {
1118 tp->lost += tcp_skb_pcount(skb);
1119 }
1120
tcp_mark_skb_lost(struct sock * sk,struct sk_buff * skb)1121 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1122 {
1123 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1124 struct tcp_sock *tp = tcp_sk(sk);
1125
1126 if (sacked & TCPCB_SACKED_ACKED)
1127 return;
1128
1129 tcp_verify_retransmit_hint(tp, skb);
1130 if (sacked & TCPCB_LOST) {
1131 if (sacked & TCPCB_SACKED_RETRANS) {
1132 /* Account for retransmits that are lost again */
1133 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1134 tp->retrans_out -= tcp_skb_pcount(skb);
1135 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1136 tcp_skb_pcount(skb));
1137 tcp_notify_skb_loss_event(tp, skb);
1138 }
1139 } else {
1140 tp->lost_out += tcp_skb_pcount(skb);
1141 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1142 tcp_notify_skb_loss_event(tp, skb);
1143 }
1144 }
1145
1146 /* Updates the delivered and delivered_ce counts */
tcp_count_delivered(struct tcp_sock * tp,u32 delivered,bool ece_ack)1147 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1148 bool ece_ack)
1149 {
1150 tp->delivered += delivered;
1151 if (ece_ack)
1152 tp->delivered_ce += delivered;
1153 }
1154
1155 /* This procedure tags the retransmission queue when SACKs arrive.
1156 *
1157 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1158 * Packets in queue with these bits set are counted in variables
1159 * sacked_out, retrans_out and lost_out, correspondingly.
1160 *
1161 * Valid combinations are:
1162 * Tag InFlight Description
1163 * 0 1 - orig segment is in flight.
1164 * S 0 - nothing flies, orig reached receiver.
1165 * L 0 - nothing flies, orig lost by net.
1166 * R 2 - both orig and retransmit are in flight.
1167 * L|R 1 - orig is lost, retransmit is in flight.
1168 * S|R 1 - orig reached receiver, retrans is still in flight.
1169 * (L|S|R is logically valid, it could occur when L|R is sacked,
1170 * but it is equivalent to plain S and code short-circuits it to S.
1171 * L|S is logically invalid, it would mean -1 packet in flight 8))
1172 *
1173 * These 6 states form finite state machine, controlled by the following events:
1174 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1175 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1176 * 3. Loss detection event of two flavors:
1177 * A. Scoreboard estimator decided the packet is lost.
1178 * A'. Reno "three dupacks" marks head of queue lost.
1179 * B. SACK arrives sacking SND.NXT at the moment, when the
1180 * segment was retransmitted.
1181 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1182 *
1183 * It is pleasant to note, that state diagram turns out to be commutative,
1184 * so that we are allowed not to be bothered by order of our actions,
1185 * when multiple events arrive simultaneously. (see the function below).
1186 *
1187 * Reordering detection.
1188 * --------------------
1189 * Reordering metric is maximal distance, which a packet can be displaced
1190 * in packet stream. With SACKs we can estimate it:
1191 *
1192 * 1. SACK fills old hole and the corresponding segment was not
1193 * ever retransmitted -> reordering. Alas, we cannot use it
1194 * when segment was retransmitted.
1195 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1196 * for retransmitted and already SACKed segment -> reordering..
1197 * Both of these heuristics are not used in Loss state, when we cannot
1198 * account for retransmits accurately.
1199 *
1200 * SACK block validation.
1201 * ----------------------
1202 *
1203 * SACK block range validation checks that the received SACK block fits to
1204 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1205 * Note that SND.UNA is not included to the range though being valid because
1206 * it means that the receiver is rather inconsistent with itself reporting
1207 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1208 * perfectly valid, however, in light of RFC2018 which explicitly states
1209 * that "SACK block MUST reflect the newest segment. Even if the newest
1210 * segment is going to be discarded ...", not that it looks very clever
1211 * in case of head skb. Due to potentional receiver driven attacks, we
1212 * choose to avoid immediate execution of a walk in write queue due to
1213 * reneging and defer head skb's loss recovery to standard loss recovery
1214 * procedure that will eventually trigger (nothing forbids us doing this).
1215 *
1216 * Implements also blockage to start_seq wrap-around. Problem lies in the
1217 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1218 * there's no guarantee that it will be before snd_nxt (n). The problem
1219 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1220 * wrap (s_w):
1221 *
1222 * <- outs wnd -> <- wrapzone ->
1223 * u e n u_w e_w s n_w
1224 * | | | | | | |
1225 * |<------------+------+----- TCP seqno space --------------+---------->|
1226 * ...-- <2^31 ->| |<--------...
1227 * ...---- >2^31 ------>| |<--------...
1228 *
1229 * Current code wouldn't be vulnerable but it's better still to discard such
1230 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1231 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1232 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1233 * equal to the ideal case (infinite seqno space without wrap caused issues).
1234 *
1235 * With D-SACK the lower bound is extended to cover sequence space below
1236 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1237 * again, D-SACK block must not to go across snd_una (for the same reason as
1238 * for the normal SACK blocks, explained above). But there all simplicity
1239 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1240 * fully below undo_marker they do not affect behavior in anyway and can
1241 * therefore be safely ignored. In rare cases (which are more or less
1242 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1243 * fragmentation and packet reordering past skb's retransmission. To consider
1244 * them correctly, the acceptable range must be extended even more though
1245 * the exact amount is rather hard to quantify. However, tp->max_window can
1246 * be used as an exaggerated estimate.
1247 */
tcp_is_sackblock_valid(struct tcp_sock * tp,bool is_dsack,u32 start_seq,u32 end_seq)1248 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1249 u32 start_seq, u32 end_seq)
1250 {
1251 /* Too far in future, or reversed (interpretation is ambiguous) */
1252 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1253 return false;
1254
1255 /* Nasty start_seq wrap-around check (see comments above) */
1256 if (!before(start_seq, tp->snd_nxt))
1257 return false;
1258
1259 /* In outstanding window? ...This is valid exit for D-SACKs too.
1260 * start_seq == snd_una is non-sensical (see comments above)
1261 */
1262 if (after(start_seq, tp->snd_una))
1263 return true;
1264
1265 if (!is_dsack || !tp->undo_marker)
1266 return false;
1267
1268 /* ...Then it's D-SACK, and must reside below snd_una completely */
1269 if (after(end_seq, tp->snd_una))
1270 return false;
1271
1272 if (!before(start_seq, tp->undo_marker))
1273 return true;
1274
1275 /* Too old */
1276 if (!after(end_seq, tp->undo_marker))
1277 return false;
1278
1279 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1280 * start_seq < undo_marker and end_seq >= undo_marker.
1281 */
1282 return !before(start_seq, end_seq - tp->max_window);
1283 }
1284
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)1285 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1286 struct tcp_sack_block_wire *sp, int num_sacks,
1287 u32 prior_snd_una, struct tcp_sacktag_state *state)
1288 {
1289 struct tcp_sock *tp = tcp_sk(sk);
1290 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1291 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1292 u32 dup_segs;
1293
1294 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1295 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1296 } else if (num_sacks > 1) {
1297 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1298 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1299
1300 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1301 return false;
1302 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1303 } else {
1304 return false;
1305 }
1306
1307 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1308 if (!dup_segs) { /* Skip dubious DSACK */
1309 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1310 return false;
1311 }
1312
1313 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1314
1315 /* D-SACK for already forgotten data... Do dumb counting. */
1316 if (tp->undo_marker && tp->undo_retrans > 0 &&
1317 !after(end_seq_0, prior_snd_una) &&
1318 after(end_seq_0, tp->undo_marker))
1319 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1320
1321 return true;
1322 }
1323
1324 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1325 * the incoming SACK may not exactly match but we can find smaller MSS
1326 * aligned portion of it that matches. Therefore we might need to fragment
1327 * which may fail and creates some hassle (caller must handle error case
1328 * returns).
1329 *
1330 * FIXME: this could be merged to shift decision code
1331 */
tcp_match_skb_to_sack(struct sock * sk,struct sk_buff * skb,u32 start_seq,u32 end_seq)1332 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1333 u32 start_seq, u32 end_seq)
1334 {
1335 int err;
1336 bool in_sack;
1337 unsigned int pkt_len;
1338 unsigned int mss;
1339
1340 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1341 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1342
1343 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1344 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1345 mss = tcp_skb_mss(skb);
1346 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1347
1348 if (!in_sack) {
1349 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1350 if (pkt_len < mss)
1351 pkt_len = mss;
1352 } else {
1353 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1354 if (pkt_len < mss)
1355 return -EINVAL;
1356 }
1357
1358 /* Round if necessary so that SACKs cover only full MSSes
1359 * and/or the remaining small portion (if present)
1360 */
1361 if (pkt_len > mss) {
1362 unsigned int new_len = (pkt_len / mss) * mss;
1363 if (!in_sack && new_len < pkt_len)
1364 new_len += mss;
1365 pkt_len = new_len;
1366 }
1367
1368 if (pkt_len >= skb->len && !in_sack)
1369 return 0;
1370
1371 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1372 pkt_len, mss, GFP_ATOMIC);
1373 if (err < 0)
1374 return err;
1375 }
1376
1377 return in_sack;
1378 }
1379
1380 /* 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)1381 static u8 tcp_sacktag_one(struct sock *sk,
1382 struct tcp_sacktag_state *state, u8 sacked,
1383 u32 start_seq, u32 end_seq,
1384 int dup_sack, int pcount,
1385 u64 xmit_time)
1386 {
1387 struct tcp_sock *tp = tcp_sk(sk);
1388
1389 /* Account D-SACK for retransmitted packet. */
1390 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1391 if (tp->undo_marker && tp->undo_retrans > 0 &&
1392 after(end_seq, tp->undo_marker))
1393 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1394 if ((sacked & TCPCB_SACKED_ACKED) &&
1395 before(start_seq, state->reord))
1396 state->reord = start_seq;
1397 }
1398
1399 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1400 if (!after(end_seq, tp->snd_una))
1401 return sacked;
1402
1403 if (!(sacked & TCPCB_SACKED_ACKED)) {
1404 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1405
1406 if (sacked & TCPCB_SACKED_RETRANS) {
1407 /* If the segment is not tagged as lost,
1408 * we do not clear RETRANS, believing
1409 * that retransmission is still in flight.
1410 */
1411 if (sacked & TCPCB_LOST) {
1412 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1413 tp->lost_out -= pcount;
1414 tp->retrans_out -= pcount;
1415 }
1416 } else {
1417 if (!(sacked & TCPCB_RETRANS)) {
1418 /* New sack for not retransmitted frame,
1419 * which was in hole. It is reordering.
1420 */
1421 if (before(start_seq,
1422 tcp_highest_sack_seq(tp)) &&
1423 before(start_seq, state->reord))
1424 state->reord = start_seq;
1425
1426 if (!after(end_seq, tp->high_seq))
1427 state->flag |= FLAG_ORIG_SACK_ACKED;
1428 if (state->first_sackt == 0)
1429 state->first_sackt = xmit_time;
1430 state->last_sackt = xmit_time;
1431 }
1432
1433 if (sacked & TCPCB_LOST) {
1434 sacked &= ~TCPCB_LOST;
1435 tp->lost_out -= pcount;
1436 }
1437 }
1438
1439 sacked |= TCPCB_SACKED_ACKED;
1440 state->flag |= FLAG_DATA_SACKED;
1441 tp->sacked_out += pcount;
1442 /* Out-of-order packets delivered */
1443 state->sack_delivered += pcount;
1444
1445 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1446 if (tp->lost_skb_hint &&
1447 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1448 tp->lost_cnt_hint += pcount;
1449 }
1450
1451 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1452 * frames and clear it. undo_retrans is decreased above, L|R frames
1453 * are accounted above as well.
1454 */
1455 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1456 sacked &= ~TCPCB_SACKED_RETRANS;
1457 tp->retrans_out -= pcount;
1458 }
1459
1460 return sacked;
1461 }
1462
1463 /* Shift newly-SACKed bytes from this skb to the immediately previous
1464 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1465 */
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)1466 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1467 struct sk_buff *skb,
1468 struct tcp_sacktag_state *state,
1469 unsigned int pcount, int shifted, int mss,
1470 bool dup_sack)
1471 {
1472 struct tcp_sock *tp = tcp_sk(sk);
1473 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1474 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1475
1476 BUG_ON(!pcount);
1477
1478 /* Adjust counters and hints for the newly sacked sequence
1479 * range but discard the return value since prev is already
1480 * marked. We must tag the range first because the seq
1481 * advancement below implicitly advances
1482 * tcp_highest_sack_seq() when skb is highest_sack.
1483 */
1484 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1485 start_seq, end_seq, dup_sack, pcount,
1486 tcp_skb_timestamp_us(skb));
1487 tcp_rate_skb_delivered(sk, skb, state->rate);
1488
1489 if (skb == tp->lost_skb_hint)
1490 tp->lost_cnt_hint += pcount;
1491
1492 TCP_SKB_CB(prev)->end_seq += shifted;
1493 TCP_SKB_CB(skb)->seq += shifted;
1494
1495 tcp_skb_pcount_add(prev, pcount);
1496 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1497 tcp_skb_pcount_add(skb, -pcount);
1498
1499 /* When we're adding to gso_segs == 1, gso_size will be zero,
1500 * in theory this shouldn't be necessary but as long as DSACK
1501 * code can come after this skb later on it's better to keep
1502 * setting gso_size to something.
1503 */
1504 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1505 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1506
1507 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1508 if (tcp_skb_pcount(skb) <= 1)
1509 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1510
1511 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1512 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1513
1514 if (skb->len > 0) {
1515 BUG_ON(!tcp_skb_pcount(skb));
1516 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1517 return false;
1518 }
1519
1520 /* Whole SKB was eaten :-) */
1521
1522 if (skb == tp->retransmit_skb_hint)
1523 tp->retransmit_skb_hint = prev;
1524 if (skb == tp->lost_skb_hint) {
1525 tp->lost_skb_hint = prev;
1526 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1527 }
1528
1529 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1530 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1531 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1532 TCP_SKB_CB(prev)->end_seq++;
1533
1534 if (skb == tcp_highest_sack(sk))
1535 tcp_advance_highest_sack(sk, skb);
1536
1537 tcp_skb_collapse_tstamp(prev, skb);
1538 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1539 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1540
1541 tcp_rtx_queue_unlink_and_free(skb, sk);
1542
1543 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1544
1545 return true;
1546 }
1547
1548 /* I wish gso_size would have a bit more sane initialization than
1549 * something-or-zero which complicates things
1550 */
tcp_skb_seglen(const struct sk_buff * skb)1551 static int tcp_skb_seglen(const struct sk_buff *skb)
1552 {
1553 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1554 }
1555
1556 /* Shifting pages past head area doesn't work */
skb_can_shift(const struct sk_buff * skb)1557 static int skb_can_shift(const struct sk_buff *skb)
1558 {
1559 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1560 }
1561
tcp_skb_shift(struct sk_buff * to,struct sk_buff * from,int pcount,int shiftlen)1562 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1563 int pcount, int shiftlen)
1564 {
1565 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1566 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1567 * to make sure not storing more than 65535 * 8 bytes per skb,
1568 * even if current MSS is bigger.
1569 */
1570 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1571 return 0;
1572 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1573 return 0;
1574 return skb_shift(to, from, shiftlen);
1575 }
1576
1577 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1578 * skb.
1579 */
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)1580 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1581 struct tcp_sacktag_state *state,
1582 u32 start_seq, u32 end_seq,
1583 bool dup_sack)
1584 {
1585 struct tcp_sock *tp = tcp_sk(sk);
1586 struct sk_buff *prev;
1587 int mss;
1588 int pcount = 0;
1589 int len;
1590 int in_sack;
1591
1592 /* Normally R but no L won't result in plain S */
1593 if (!dup_sack &&
1594 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1595 goto fallback;
1596 if (!skb_can_shift(skb))
1597 goto fallback;
1598 /* This frame is about to be dropped (was ACKed). */
1599 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1600 goto fallback;
1601
1602 /* Can only happen with delayed DSACK + discard craziness */
1603 prev = skb_rb_prev(skb);
1604 if (!prev)
1605 goto fallback;
1606
1607 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1608 goto fallback;
1609
1610 if (!tcp_skb_can_collapse(prev, skb))
1611 goto fallback;
1612
1613 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1614 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1615
1616 if (in_sack) {
1617 len = skb->len;
1618 pcount = tcp_skb_pcount(skb);
1619 mss = tcp_skb_seglen(skb);
1620
1621 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1622 * drop this restriction as unnecessary
1623 */
1624 if (mss != tcp_skb_seglen(prev))
1625 goto fallback;
1626 } else {
1627 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1628 goto noop;
1629 /* CHECKME: This is non-MSS split case only?, this will
1630 * cause skipped skbs due to advancing loop btw, original
1631 * has that feature too
1632 */
1633 if (tcp_skb_pcount(skb) <= 1)
1634 goto noop;
1635
1636 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1637 if (!in_sack) {
1638 /* TODO: head merge to next could be attempted here
1639 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1640 * though it might not be worth of the additional hassle
1641 *
1642 * ...we can probably just fallback to what was done
1643 * previously. We could try merging non-SACKed ones
1644 * as well but it probably isn't going to buy off
1645 * because later SACKs might again split them, and
1646 * it would make skb timestamp tracking considerably
1647 * harder problem.
1648 */
1649 goto fallback;
1650 }
1651
1652 len = end_seq - TCP_SKB_CB(skb)->seq;
1653 BUG_ON(len < 0);
1654 BUG_ON(len > skb->len);
1655
1656 /* MSS boundaries should be honoured or else pcount will
1657 * severely break even though it makes things bit trickier.
1658 * Optimize common case to avoid most of the divides
1659 */
1660 mss = tcp_skb_mss(skb);
1661
1662 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1663 * drop this restriction as unnecessary
1664 */
1665 if (mss != tcp_skb_seglen(prev))
1666 goto fallback;
1667
1668 if (len == mss) {
1669 pcount = 1;
1670 } else if (len < mss) {
1671 goto noop;
1672 } else {
1673 pcount = len / mss;
1674 len = pcount * mss;
1675 }
1676 }
1677
1678 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1679 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1680 goto fallback;
1681
1682 if (!tcp_skb_shift(prev, skb, pcount, len))
1683 goto fallback;
1684 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1685 goto out;
1686
1687 /* Hole filled allows collapsing with the next as well, this is very
1688 * useful when hole on every nth skb pattern happens
1689 */
1690 skb = skb_rb_next(prev);
1691 if (!skb)
1692 goto out;
1693
1694 if (!skb_can_shift(skb) ||
1695 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1696 (mss != tcp_skb_seglen(skb)))
1697 goto out;
1698
1699 if (!tcp_skb_can_collapse(prev, skb))
1700 goto out;
1701 len = skb->len;
1702 pcount = tcp_skb_pcount(skb);
1703 if (tcp_skb_shift(prev, skb, pcount, len))
1704 tcp_shifted_skb(sk, prev, skb, state, pcount,
1705 len, mss, 0);
1706
1707 out:
1708 return prev;
1709
1710 noop:
1711 return skb;
1712
1713 fallback:
1714 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1715 return NULL;
1716 }
1717
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)1718 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1719 struct tcp_sack_block *next_dup,
1720 struct tcp_sacktag_state *state,
1721 u32 start_seq, u32 end_seq,
1722 bool dup_sack_in)
1723 {
1724 struct tcp_sock *tp = tcp_sk(sk);
1725 struct sk_buff *tmp;
1726
1727 skb_rbtree_walk_from(skb) {
1728 int in_sack = 0;
1729 bool dup_sack = dup_sack_in;
1730
1731 /* queue is in-order => we can short-circuit the walk early */
1732 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1733 break;
1734
1735 if (next_dup &&
1736 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1737 in_sack = tcp_match_skb_to_sack(sk, skb,
1738 next_dup->start_seq,
1739 next_dup->end_seq);
1740 if (in_sack > 0)
1741 dup_sack = true;
1742 }
1743
1744 /* skb reference here is a bit tricky to get right, since
1745 * shifting can eat and free both this skb and the next,
1746 * so not even _safe variant of the loop is enough.
1747 */
1748 if (in_sack <= 0) {
1749 tmp = tcp_shift_skb_data(sk, skb, state,
1750 start_seq, end_seq, dup_sack);
1751 if (tmp) {
1752 if (tmp != skb) {
1753 skb = tmp;
1754 continue;
1755 }
1756
1757 in_sack = 0;
1758 } else {
1759 in_sack = tcp_match_skb_to_sack(sk, skb,
1760 start_seq,
1761 end_seq);
1762 }
1763 }
1764
1765 if (unlikely(in_sack < 0))
1766 break;
1767
1768 if (in_sack) {
1769 TCP_SKB_CB(skb)->sacked =
1770 tcp_sacktag_one(sk,
1771 state,
1772 TCP_SKB_CB(skb)->sacked,
1773 TCP_SKB_CB(skb)->seq,
1774 TCP_SKB_CB(skb)->end_seq,
1775 dup_sack,
1776 tcp_skb_pcount(skb),
1777 tcp_skb_timestamp_us(skb));
1778 tcp_rate_skb_delivered(sk, skb, state->rate);
1779 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1780 list_del_init(&skb->tcp_tsorted_anchor);
1781
1782 if (!before(TCP_SKB_CB(skb)->seq,
1783 tcp_highest_sack_seq(tp)))
1784 tcp_advance_highest_sack(sk, skb);
1785 }
1786 }
1787 return skb;
1788 }
1789
tcp_sacktag_bsearch(struct sock * sk,u32 seq)1790 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1791 {
1792 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1793 struct sk_buff *skb;
1794
1795 while (*p) {
1796 parent = *p;
1797 skb = rb_to_skb(parent);
1798 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1799 p = &parent->rb_left;
1800 continue;
1801 }
1802 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1803 p = &parent->rb_right;
1804 continue;
1805 }
1806 return skb;
1807 }
1808 return NULL;
1809 }
1810
tcp_sacktag_skip(struct sk_buff * skb,struct sock * sk,u32 skip_to_seq)1811 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1812 u32 skip_to_seq)
1813 {
1814 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1815 return skb;
1816
1817 return tcp_sacktag_bsearch(sk, skip_to_seq);
1818 }
1819
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)1820 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1821 struct sock *sk,
1822 struct tcp_sack_block *next_dup,
1823 struct tcp_sacktag_state *state,
1824 u32 skip_to_seq)
1825 {
1826 if (!next_dup)
1827 return skb;
1828
1829 if (before(next_dup->start_seq, skip_to_seq)) {
1830 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1831 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1832 next_dup->start_seq, next_dup->end_seq,
1833 1);
1834 }
1835
1836 return skb;
1837 }
1838
tcp_sack_cache_ok(const struct tcp_sock * tp,const struct tcp_sack_block * cache)1839 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1840 {
1841 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1842 }
1843
1844 static int
tcp_sacktag_write_queue(struct sock * sk,const struct sk_buff * ack_skb,u32 prior_snd_una,struct tcp_sacktag_state * state)1845 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1846 u32 prior_snd_una, struct tcp_sacktag_state *state)
1847 {
1848 struct tcp_sock *tp = tcp_sk(sk);
1849 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1850 TCP_SKB_CB(ack_skb)->sacked);
1851 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1852 struct tcp_sack_block sp[TCP_NUM_SACKS];
1853 struct tcp_sack_block *cache;
1854 struct sk_buff *skb;
1855 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1856 int used_sacks;
1857 bool found_dup_sack = false;
1858 int i, j;
1859 int first_sack_index;
1860
1861 state->flag = 0;
1862 state->reord = tp->snd_nxt;
1863
1864 if (!tp->sacked_out)
1865 tcp_highest_sack_reset(sk);
1866
1867 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1868 num_sacks, prior_snd_una, state);
1869
1870 /* Eliminate too old ACKs, but take into
1871 * account more or less fresh ones, they can
1872 * contain valid SACK info.
1873 */
1874 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1875 return 0;
1876
1877 if (!tp->packets_out)
1878 goto out;
1879
1880 used_sacks = 0;
1881 first_sack_index = 0;
1882 for (i = 0; i < num_sacks; i++) {
1883 bool dup_sack = !i && found_dup_sack;
1884
1885 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1886 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1887
1888 if (!tcp_is_sackblock_valid(tp, dup_sack,
1889 sp[used_sacks].start_seq,
1890 sp[used_sacks].end_seq)) {
1891 int mib_idx;
1892
1893 if (dup_sack) {
1894 if (!tp->undo_marker)
1895 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1896 else
1897 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1898 } else {
1899 /* Don't count olds caused by ACK reordering */
1900 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1901 !after(sp[used_sacks].end_seq, tp->snd_una))
1902 continue;
1903 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1904 }
1905
1906 NET_INC_STATS(sock_net(sk), mib_idx);
1907 if (i == 0)
1908 first_sack_index = -1;
1909 continue;
1910 }
1911
1912 /* Ignore very old stuff early */
1913 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1914 if (i == 0)
1915 first_sack_index = -1;
1916 continue;
1917 }
1918
1919 used_sacks++;
1920 }
1921
1922 /* order SACK blocks to allow in order walk of the retrans queue */
1923 for (i = used_sacks - 1; i > 0; i--) {
1924 for (j = 0; j < i; j++) {
1925 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1926 swap(sp[j], sp[j + 1]);
1927
1928 /* Track where the first SACK block goes to */
1929 if (j == first_sack_index)
1930 first_sack_index = j + 1;
1931 }
1932 }
1933 }
1934
1935 state->mss_now = tcp_current_mss(sk);
1936 skb = NULL;
1937 i = 0;
1938
1939 if (!tp->sacked_out) {
1940 /* It's already past, so skip checking against it */
1941 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1942 } else {
1943 cache = tp->recv_sack_cache;
1944 /* Skip empty blocks in at head of the cache */
1945 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1946 !cache->end_seq)
1947 cache++;
1948 }
1949
1950 while (i < used_sacks) {
1951 u32 start_seq = sp[i].start_seq;
1952 u32 end_seq = sp[i].end_seq;
1953 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1954 struct tcp_sack_block *next_dup = NULL;
1955
1956 if (found_dup_sack && ((i + 1) == first_sack_index))
1957 next_dup = &sp[i + 1];
1958
1959 /* Skip too early cached blocks */
1960 while (tcp_sack_cache_ok(tp, cache) &&
1961 !before(start_seq, cache->end_seq))
1962 cache++;
1963
1964 /* Can skip some work by looking recv_sack_cache? */
1965 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1966 after(end_seq, cache->start_seq)) {
1967
1968 /* Head todo? */
1969 if (before(start_seq, cache->start_seq)) {
1970 skb = tcp_sacktag_skip(skb, sk, start_seq);
1971 skb = tcp_sacktag_walk(skb, sk, next_dup,
1972 state,
1973 start_seq,
1974 cache->start_seq,
1975 dup_sack);
1976 }
1977
1978 /* Rest of the block already fully processed? */
1979 if (!after(end_seq, cache->end_seq))
1980 goto advance_sp;
1981
1982 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1983 state,
1984 cache->end_seq);
1985
1986 /* ...tail remains todo... */
1987 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1988 /* ...but better entrypoint exists! */
1989 skb = tcp_highest_sack(sk);
1990 if (!skb)
1991 break;
1992 cache++;
1993 goto walk;
1994 }
1995
1996 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1997 /* Check overlap against next cached too (past this one already) */
1998 cache++;
1999 continue;
2000 }
2001
2002 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
2003 skb = tcp_highest_sack(sk);
2004 if (!skb)
2005 break;
2006 }
2007 skb = tcp_sacktag_skip(skb, sk, start_seq);
2008
2009 walk:
2010 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
2011 start_seq, end_seq, dup_sack);
2012
2013 advance_sp:
2014 i++;
2015 }
2016
2017 /* Clear the head of the cache sack blocks so we can skip it next time */
2018 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
2019 tp->recv_sack_cache[i].start_seq = 0;
2020 tp->recv_sack_cache[i].end_seq = 0;
2021 }
2022 for (j = 0; j < used_sacks; j++)
2023 tp->recv_sack_cache[i++] = sp[j];
2024
2025 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
2026 tcp_check_sack_reordering(sk, state->reord, 0);
2027
2028 tcp_verify_left_out(tp);
2029 out:
2030
2031 #if FASTRETRANS_DEBUG > 0
2032 WARN_ON((int)tp->sacked_out < 0);
2033 WARN_ON((int)tp->lost_out < 0);
2034 WARN_ON((int)tp->retrans_out < 0);
2035 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
2036 #endif
2037 return state->flag;
2038 }
2039
2040 /* Limits sacked_out so that sum with lost_out isn't ever larger than
2041 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
2042 */
tcp_limit_reno_sacked(struct tcp_sock * tp)2043 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
2044 {
2045 u32 holes;
2046
2047 holes = max(tp->lost_out, 1U);
2048 holes = min(holes, tp->packets_out);
2049
2050 if ((tp->sacked_out + holes) > tp->packets_out) {
2051 tp->sacked_out = tp->packets_out - holes;
2052 return true;
2053 }
2054 return false;
2055 }
2056
2057 /* If we receive more dupacks than we expected counting segments
2058 * in assumption of absent reordering, interpret this as reordering.
2059 * The only another reason could be bug in receiver TCP.
2060 */
tcp_check_reno_reordering(struct sock * sk,const int addend)2061 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2062 {
2063 struct tcp_sock *tp = tcp_sk(sk);
2064
2065 if (!tcp_limit_reno_sacked(tp))
2066 return;
2067
2068 tp->reordering = min_t(u32, tp->packets_out + addend,
2069 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
2070 tp->reord_seen++;
2071 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
2072 }
2073
2074 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2075
tcp_add_reno_sack(struct sock * sk,int num_dupack,bool ece_ack)2076 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2077 {
2078 if (num_dupack) {
2079 struct tcp_sock *tp = tcp_sk(sk);
2080 u32 prior_sacked = tp->sacked_out;
2081 s32 delivered;
2082
2083 tp->sacked_out += num_dupack;
2084 tcp_check_reno_reordering(sk, 0);
2085 delivered = tp->sacked_out - prior_sacked;
2086 if (delivered > 0)
2087 tcp_count_delivered(tp, delivered, ece_ack);
2088 tcp_verify_left_out(tp);
2089 }
2090 }
2091
2092 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2093
tcp_remove_reno_sacks(struct sock * sk,int acked,bool ece_ack)2094 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2095 {
2096 struct tcp_sock *tp = tcp_sk(sk);
2097
2098 if (acked > 0) {
2099 /* One ACK acked hole. The rest eat duplicate ACKs. */
2100 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2101 ece_ack);
2102 if (acked - 1 >= tp->sacked_out)
2103 tp->sacked_out = 0;
2104 else
2105 tp->sacked_out -= acked - 1;
2106 }
2107 tcp_check_reno_reordering(sk, acked);
2108 tcp_verify_left_out(tp);
2109 }
2110
tcp_reset_reno_sack(struct tcp_sock * tp)2111 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2112 {
2113 tp->sacked_out = 0;
2114 }
2115
tcp_clear_retrans(struct tcp_sock * tp)2116 void tcp_clear_retrans(struct tcp_sock *tp)
2117 {
2118 tp->retrans_out = 0;
2119 tp->lost_out = 0;
2120 tp->undo_marker = 0;
2121 tp->undo_retrans = -1;
2122 tp->sacked_out = 0;
2123 tp->rto_stamp = 0;
2124 tp->total_rto = 0;
2125 tp->total_rto_recoveries = 0;
2126 tp->total_rto_time = 0;
2127 }
2128
tcp_init_undo(struct tcp_sock * tp)2129 static inline void tcp_init_undo(struct tcp_sock *tp)
2130 {
2131 tp->undo_marker = tp->snd_una;
2132 /* Retransmission still in flight may cause DSACKs later. */
2133 tp->undo_retrans = tp->retrans_out ? : -1;
2134 }
2135
tcp_is_rack(const struct sock * sk)2136 static bool tcp_is_rack(const struct sock *sk)
2137 {
2138 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
2139 TCP_RACK_LOSS_DETECTION;
2140 }
2141
2142 /* If we detect SACK reneging, forget all SACK information
2143 * and reset tags completely, otherwise preserve SACKs. If receiver
2144 * dropped its ofo queue, we will know this due to reneging detection.
2145 */
tcp_timeout_mark_lost(struct sock * sk)2146 static void tcp_timeout_mark_lost(struct sock *sk)
2147 {
2148 struct tcp_sock *tp = tcp_sk(sk);
2149 struct sk_buff *skb, *head;
2150 bool is_reneg; /* is receiver reneging on SACKs? */
2151
2152 head = tcp_rtx_queue_head(sk);
2153 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2154 if (is_reneg) {
2155 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2156 tp->sacked_out = 0;
2157 /* Mark SACK reneging until we recover from this loss event. */
2158 tp->is_sack_reneg = 1;
2159 } else if (tcp_is_reno(tp)) {
2160 tcp_reset_reno_sack(tp);
2161 }
2162
2163 skb = head;
2164 skb_rbtree_walk_from(skb) {
2165 if (is_reneg)
2166 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2167 else if (tcp_is_rack(sk) && skb != head &&
2168 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2169 continue; /* Don't mark recently sent ones lost yet */
2170 tcp_mark_skb_lost(sk, skb);
2171 }
2172 tcp_verify_left_out(tp);
2173 tcp_clear_all_retrans_hints(tp);
2174 }
2175
2176 /* Enter Loss state. */
tcp_enter_loss(struct sock * sk)2177 void tcp_enter_loss(struct sock *sk)
2178 {
2179 const struct inet_connection_sock *icsk = inet_csk(sk);
2180 struct tcp_sock *tp = tcp_sk(sk);
2181 struct net *net = sock_net(sk);
2182 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2183 u8 reordering;
2184
2185 tcp_timeout_mark_lost(sk);
2186
2187 /* Reduce ssthresh if it has not yet been made inside this window. */
2188 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2189 !after(tp->high_seq, tp->snd_una) ||
2190 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2191 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2192 tp->prior_cwnd = tcp_snd_cwnd(tp);
2193 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2194 tcp_ca_event(sk, CA_EVENT_LOSS);
2195 tcp_init_undo(tp);
2196 }
2197 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1);
2198 tp->snd_cwnd_cnt = 0;
2199 tp->snd_cwnd_stamp = tcp_jiffies32;
2200
2201 /* Timeout in disordered state after receiving substantial DUPACKs
2202 * suggests that the degree of reordering is over-estimated.
2203 */
2204 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2205 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2206 tp->sacked_out >= reordering)
2207 tp->reordering = min_t(unsigned int, tp->reordering,
2208 reordering);
2209
2210 tcp_set_ca_state(sk, TCP_CA_Loss);
2211 tp->high_seq = tp->snd_nxt;
2212 tcp_ecn_queue_cwr(tp);
2213
2214 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2215 * loss recovery is underway except recurring timeout(s) on
2216 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2217 */
2218 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2219 (new_recovery || icsk->icsk_retransmits) &&
2220 !inet_csk(sk)->icsk_mtup.probe_size;
2221 }
2222
2223 /* If ACK arrived pointing to a remembered SACK, it means that our
2224 * remembered SACKs do not reflect real state of receiver i.e.
2225 * receiver _host_ is heavily congested (or buggy).
2226 *
2227 * To avoid big spurious retransmission bursts due to transient SACK
2228 * scoreboard oddities that look like reneging, we give the receiver a
2229 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2230 * restore sanity to the SACK scoreboard. If the apparent reneging
2231 * persists until this RTO then we'll clear the SACK scoreboard.
2232 */
tcp_check_sack_reneging(struct sock * sk,int * ack_flag)2233 static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag)
2234 {
2235 if (*ack_flag & FLAG_SACK_RENEGING &&
2236 *ack_flag & FLAG_SND_UNA_ADVANCED) {
2237 struct tcp_sock *tp = tcp_sk(sk);
2238 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2239 msecs_to_jiffies(10));
2240
2241 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2242 delay, TCP_RTO_MAX);
2243 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2244 return true;
2245 }
2246 return false;
2247 }
2248
2249 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2250 * counter when SACK is enabled (without SACK, sacked_out is used for
2251 * that purpose).
2252 *
2253 * With reordering, holes may still be in flight, so RFC3517 recovery
2254 * uses pure sacked_out (total number of SACKed segments) even though
2255 * it violates the RFC that uses duplicate ACKs, often these are equal
2256 * but when e.g. out-of-window ACKs or packet duplication occurs,
2257 * they differ. Since neither occurs due to loss, TCP should really
2258 * ignore them.
2259 */
tcp_dupack_heuristics(const struct tcp_sock * tp)2260 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2261 {
2262 return tp->sacked_out + 1;
2263 }
2264
2265 /* Linux NewReno/SACK/ECN state machine.
2266 * --------------------------------------
2267 *
2268 * "Open" Normal state, no dubious events, fast path.
2269 * "Disorder" In all the respects it is "Open",
2270 * but requires a bit more attention. It is entered when
2271 * we see some SACKs or dupacks. It is split of "Open"
2272 * mainly to move some processing from fast path to slow one.
2273 * "CWR" CWND was reduced due to some Congestion Notification event.
2274 * It can be ECN, ICMP source quench, local device congestion.
2275 * "Recovery" CWND was reduced, we are fast-retransmitting.
2276 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2277 *
2278 * tcp_fastretrans_alert() is entered:
2279 * - each incoming ACK, if state is not "Open"
2280 * - when arrived ACK is unusual, namely:
2281 * * SACK
2282 * * Duplicate ACK.
2283 * * ECN ECE.
2284 *
2285 * Counting packets in flight is pretty simple.
2286 *
2287 * in_flight = packets_out - left_out + retrans_out
2288 *
2289 * packets_out is SND.NXT-SND.UNA counted in packets.
2290 *
2291 * retrans_out is number of retransmitted segments.
2292 *
2293 * left_out is number of segments left network, but not ACKed yet.
2294 *
2295 * left_out = sacked_out + lost_out
2296 *
2297 * sacked_out: Packets, which arrived to receiver out of order
2298 * and hence not ACKed. With SACKs this number is simply
2299 * amount of SACKed data. Even without SACKs
2300 * it is easy to give pretty reliable estimate of this number,
2301 * counting duplicate ACKs.
2302 *
2303 * lost_out: Packets lost by network. TCP has no explicit
2304 * "loss notification" feedback from network (for now).
2305 * It means that this number can be only _guessed_.
2306 * Actually, it is the heuristics to predict lossage that
2307 * distinguishes different algorithms.
2308 *
2309 * F.e. after RTO, when all the queue is considered as lost,
2310 * lost_out = packets_out and in_flight = retrans_out.
2311 *
2312 * Essentially, we have now a few algorithms detecting
2313 * lost packets.
2314 *
2315 * If the receiver supports SACK:
2316 *
2317 * RFC6675/3517: It is the conventional algorithm. A packet is
2318 * considered lost if the number of higher sequence packets
2319 * SACKed is greater than or equal the DUPACK thoreshold
2320 * (reordering). This is implemented in tcp_mark_head_lost and
2321 * tcp_update_scoreboard.
2322 *
2323 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2324 * (2017-) that checks timing instead of counting DUPACKs.
2325 * Essentially a packet is considered lost if it's not S/ACKed
2326 * after RTT + reordering_window, where both metrics are
2327 * dynamically measured and adjusted. This is implemented in
2328 * tcp_rack_mark_lost.
2329 *
2330 * If the receiver does not support SACK:
2331 *
2332 * NewReno (RFC6582): in Recovery we assume that one segment
2333 * is lost (classic Reno). While we are in Recovery and
2334 * a partial ACK arrives, we assume that one more packet
2335 * is lost (NewReno). This heuristics are the same in NewReno
2336 * and SACK.
2337 *
2338 * Really tricky (and requiring careful tuning) part of algorithm
2339 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2340 * The first determines the moment _when_ we should reduce CWND and,
2341 * hence, slow down forward transmission. In fact, it determines the moment
2342 * when we decide that hole is caused by loss, rather than by a reorder.
2343 *
2344 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2345 * holes, caused by lost packets.
2346 *
2347 * And the most logically complicated part of algorithm is undo
2348 * heuristics. We detect false retransmits due to both too early
2349 * fast retransmit (reordering) and underestimated RTO, analyzing
2350 * timestamps and D-SACKs. When we detect that some segments were
2351 * retransmitted by mistake and CWND reduction was wrong, we undo
2352 * window reduction and abort recovery phase. This logic is hidden
2353 * inside several functions named tcp_try_undo_<something>.
2354 */
2355
2356 /* This function decides, when we should leave Disordered state
2357 * and enter Recovery phase, reducing congestion window.
2358 *
2359 * Main question: may we further continue forward transmission
2360 * with the same cwnd?
2361 */
tcp_time_to_recover(struct sock * sk,int flag)2362 static bool tcp_time_to_recover(struct sock *sk, int flag)
2363 {
2364 struct tcp_sock *tp = tcp_sk(sk);
2365
2366 /* Trick#1: The loss is proven. */
2367 if (tp->lost_out)
2368 return true;
2369
2370 /* Not-A-Trick#2 : Classic rule... */
2371 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2372 return true;
2373
2374 return false;
2375 }
2376
2377 /* Detect loss in event "A" above by marking head of queue up as lost.
2378 * For RFC3517 SACK, a segment is considered lost if it
2379 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2380 * the maximum SACKed segments to pass before reaching this limit.
2381 */
tcp_mark_head_lost(struct sock * sk,int packets,int mark_head)2382 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2383 {
2384 struct tcp_sock *tp = tcp_sk(sk);
2385 struct sk_buff *skb;
2386 int cnt;
2387 /* Use SACK to deduce losses of new sequences sent during recovery */
2388 const u32 loss_high = tp->snd_nxt;
2389
2390 WARN_ON(packets > tp->packets_out);
2391 skb = tp->lost_skb_hint;
2392 if (skb) {
2393 /* Head already handled? */
2394 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2395 return;
2396 cnt = tp->lost_cnt_hint;
2397 } else {
2398 skb = tcp_rtx_queue_head(sk);
2399 cnt = 0;
2400 }
2401
2402 skb_rbtree_walk_from(skb) {
2403 /* TODO: do this better */
2404 /* this is not the most efficient way to do this... */
2405 tp->lost_skb_hint = skb;
2406 tp->lost_cnt_hint = cnt;
2407
2408 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2409 break;
2410
2411 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2412 cnt += tcp_skb_pcount(skb);
2413
2414 if (cnt > packets)
2415 break;
2416
2417 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2418 tcp_mark_skb_lost(sk, skb);
2419
2420 if (mark_head)
2421 break;
2422 }
2423 tcp_verify_left_out(tp);
2424 }
2425
2426 /* Account newly detected lost packet(s) */
2427
tcp_update_scoreboard(struct sock * sk,int fast_rexmit)2428 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2429 {
2430 struct tcp_sock *tp = tcp_sk(sk);
2431
2432 if (tcp_is_sack(tp)) {
2433 int sacked_upto = tp->sacked_out - tp->reordering;
2434 if (sacked_upto >= 0)
2435 tcp_mark_head_lost(sk, sacked_upto, 0);
2436 else if (fast_rexmit)
2437 tcp_mark_head_lost(sk, 1, 1);
2438 }
2439 }
2440
tcp_tsopt_ecr_before(const struct tcp_sock * tp,u32 when)2441 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2442 {
2443 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2444 before(tp->rx_opt.rcv_tsecr, when);
2445 }
2446
2447 /* skb is spurious retransmitted if the returned timestamp echo
2448 * reply is prior to the skb transmission time
2449 */
tcp_skb_spurious_retrans(const struct tcp_sock * tp,const struct sk_buff * skb)2450 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2451 const struct sk_buff *skb)
2452 {
2453 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2454 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb));
2455 }
2456
2457 /* Nothing was retransmitted or returned timestamp is less
2458 * than timestamp of the first retransmission.
2459 */
tcp_packet_delayed(const struct tcp_sock * tp)2460 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2461 {
2462 return tp->retrans_stamp &&
2463 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2464 }
2465
2466 /* Undo procedures. */
2467
2468 /* We can clear retrans_stamp when there are no retransmissions in the
2469 * window. It would seem that it is trivially available for us in
2470 * tp->retrans_out, however, that kind of assumptions doesn't consider
2471 * what will happen if errors occur when sending retransmission for the
2472 * second time. ...It could the that such segment has only
2473 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2474 * the head skb is enough except for some reneging corner cases that
2475 * are not worth the effort.
2476 *
2477 * Main reason for all this complexity is the fact that connection dying
2478 * time now depends on the validity of the retrans_stamp, in particular,
2479 * that successive retransmissions of a segment must not advance
2480 * retrans_stamp under any conditions.
2481 */
tcp_any_retrans_done(const struct sock * sk)2482 static bool tcp_any_retrans_done(const struct sock *sk)
2483 {
2484 const struct tcp_sock *tp = tcp_sk(sk);
2485 struct sk_buff *skb;
2486
2487 if (tp->retrans_out)
2488 return true;
2489
2490 skb = tcp_rtx_queue_head(sk);
2491 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2492 return true;
2493
2494 return false;
2495 }
2496
DBGUNDO(struct sock * sk,const char * msg)2497 static void DBGUNDO(struct sock *sk, const char *msg)
2498 {
2499 #if FASTRETRANS_DEBUG > 1
2500 struct tcp_sock *tp = tcp_sk(sk);
2501 struct inet_sock *inet = inet_sk(sk);
2502
2503 if (sk->sk_family == AF_INET) {
2504 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2505 msg,
2506 &inet->inet_daddr, ntohs(inet->inet_dport),
2507 tcp_snd_cwnd(tp), tcp_left_out(tp),
2508 tp->snd_ssthresh, tp->prior_ssthresh,
2509 tp->packets_out);
2510 }
2511 #if IS_ENABLED(CONFIG_IPV6)
2512 else if (sk->sk_family == AF_INET6) {
2513 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2514 msg,
2515 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2516 tcp_snd_cwnd(tp), tcp_left_out(tp),
2517 tp->snd_ssthresh, tp->prior_ssthresh,
2518 tp->packets_out);
2519 }
2520 #endif
2521 #endif
2522 }
2523
tcp_undo_cwnd_reduction(struct sock * sk,bool unmark_loss)2524 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2525 {
2526 struct tcp_sock *tp = tcp_sk(sk);
2527
2528 if (unmark_loss) {
2529 struct sk_buff *skb;
2530
2531 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2532 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2533 }
2534 tp->lost_out = 0;
2535 tcp_clear_all_retrans_hints(tp);
2536 }
2537
2538 if (tp->prior_ssthresh) {
2539 const struct inet_connection_sock *icsk = inet_csk(sk);
2540
2541 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk));
2542
2543 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2544 tp->snd_ssthresh = tp->prior_ssthresh;
2545 tcp_ecn_withdraw_cwr(tp);
2546 }
2547 }
2548 tp->snd_cwnd_stamp = tcp_jiffies32;
2549 tp->undo_marker = 0;
2550 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2551 }
2552
tcp_may_undo(const struct tcp_sock * tp)2553 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2554 {
2555 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2556 }
2557
tcp_is_non_sack_preventing_reopen(struct sock * sk)2558 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2559 {
2560 struct tcp_sock *tp = tcp_sk(sk);
2561
2562 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2563 /* Hold old state until something *above* high_seq
2564 * is ACKed. For Reno it is MUST to prevent false
2565 * fast retransmits (RFC2582). SACK TCP is safe. */
2566 if (!tcp_any_retrans_done(sk))
2567 tp->retrans_stamp = 0;
2568 return true;
2569 }
2570 return false;
2571 }
2572
2573 /* People celebrate: "We love our President!" */
tcp_try_undo_recovery(struct sock * sk)2574 static bool tcp_try_undo_recovery(struct sock *sk)
2575 {
2576 struct tcp_sock *tp = tcp_sk(sk);
2577
2578 if (tcp_may_undo(tp)) {
2579 int mib_idx;
2580
2581 /* Happy end! We did not retransmit anything
2582 * or our original transmission succeeded.
2583 */
2584 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2585 tcp_undo_cwnd_reduction(sk, false);
2586 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2587 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2588 else
2589 mib_idx = LINUX_MIB_TCPFULLUNDO;
2590
2591 NET_INC_STATS(sock_net(sk), mib_idx);
2592 } else if (tp->rack.reo_wnd_persist) {
2593 tp->rack.reo_wnd_persist--;
2594 }
2595 if (tcp_is_non_sack_preventing_reopen(sk))
2596 return true;
2597 tcp_set_ca_state(sk, TCP_CA_Open);
2598 tp->is_sack_reneg = 0;
2599 return false;
2600 }
2601
2602 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
tcp_try_undo_dsack(struct sock * sk)2603 static bool tcp_try_undo_dsack(struct sock *sk)
2604 {
2605 struct tcp_sock *tp = tcp_sk(sk);
2606
2607 if (tp->undo_marker && !tp->undo_retrans) {
2608 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2609 tp->rack.reo_wnd_persist + 1);
2610 DBGUNDO(sk, "D-SACK");
2611 tcp_undo_cwnd_reduction(sk, false);
2612 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2613 return true;
2614 }
2615 return false;
2616 }
2617
2618 /* Undo during loss recovery after partial ACK or using F-RTO. */
tcp_try_undo_loss(struct sock * sk,bool frto_undo)2619 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2620 {
2621 struct tcp_sock *tp = tcp_sk(sk);
2622
2623 if (frto_undo || tcp_may_undo(tp)) {
2624 tcp_undo_cwnd_reduction(sk, true);
2625
2626 DBGUNDO(sk, "partial loss");
2627 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2628 if (frto_undo)
2629 NET_INC_STATS(sock_net(sk),
2630 LINUX_MIB_TCPSPURIOUSRTOS);
2631 inet_csk(sk)->icsk_retransmits = 0;
2632 if (tcp_is_non_sack_preventing_reopen(sk))
2633 return true;
2634 if (frto_undo || tcp_is_sack(tp)) {
2635 tcp_set_ca_state(sk, TCP_CA_Open);
2636 tp->is_sack_reneg = 0;
2637 }
2638 return true;
2639 }
2640 return false;
2641 }
2642
2643 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2644 * It computes the number of packets to send (sndcnt) based on packets newly
2645 * delivered:
2646 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2647 * cwnd reductions across a full RTT.
2648 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2649 * But when SND_UNA is acked without further losses,
2650 * slow starts cwnd up to ssthresh to speed up the recovery.
2651 */
tcp_init_cwnd_reduction(struct sock * sk)2652 static void tcp_init_cwnd_reduction(struct sock *sk)
2653 {
2654 struct tcp_sock *tp = tcp_sk(sk);
2655
2656 tp->high_seq = tp->snd_nxt;
2657 tp->tlp_high_seq = 0;
2658 tp->snd_cwnd_cnt = 0;
2659 tp->prior_cwnd = tcp_snd_cwnd(tp);
2660 tp->prr_delivered = 0;
2661 tp->prr_out = 0;
2662 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2663 tcp_ecn_queue_cwr(tp);
2664 }
2665
tcp_cwnd_reduction(struct sock * sk,int newly_acked_sacked,int newly_lost,int flag)2666 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag)
2667 {
2668 struct tcp_sock *tp = tcp_sk(sk);
2669 int sndcnt = 0;
2670 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2671
2672 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2673 return;
2674
2675 tp->prr_delivered += newly_acked_sacked;
2676 if (delta < 0) {
2677 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2678 tp->prior_cwnd - 1;
2679 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2680 } else {
2681 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out,
2682 newly_acked_sacked);
2683 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost)
2684 sndcnt++;
2685 sndcnt = min(delta, sndcnt);
2686 }
2687 /* Force a fast retransmit upon entering fast recovery */
2688 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2689 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt);
2690 }
2691
tcp_end_cwnd_reduction(struct sock * sk)2692 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2693 {
2694 struct tcp_sock *tp = tcp_sk(sk);
2695
2696 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2697 return;
2698
2699 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2700 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2701 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2702 tcp_snd_cwnd_set(tp, tp->snd_ssthresh);
2703 tp->snd_cwnd_stamp = tcp_jiffies32;
2704 }
2705 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2706 }
2707
2708 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
tcp_enter_cwr(struct sock * sk)2709 void tcp_enter_cwr(struct sock *sk)
2710 {
2711 struct tcp_sock *tp = tcp_sk(sk);
2712
2713 tp->prior_ssthresh = 0;
2714 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2715 tp->undo_marker = 0;
2716 tcp_init_cwnd_reduction(sk);
2717 tcp_set_ca_state(sk, TCP_CA_CWR);
2718 }
2719 }
2720 EXPORT_SYMBOL(tcp_enter_cwr);
2721
tcp_try_keep_open(struct sock * sk)2722 static void tcp_try_keep_open(struct sock *sk)
2723 {
2724 struct tcp_sock *tp = tcp_sk(sk);
2725 int state = TCP_CA_Open;
2726
2727 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2728 state = TCP_CA_Disorder;
2729
2730 if (inet_csk(sk)->icsk_ca_state != state) {
2731 tcp_set_ca_state(sk, state);
2732 tp->high_seq = tp->snd_nxt;
2733 }
2734 }
2735
tcp_try_to_open(struct sock * sk,int flag)2736 static void tcp_try_to_open(struct sock *sk, int flag)
2737 {
2738 struct tcp_sock *tp = tcp_sk(sk);
2739
2740 tcp_verify_left_out(tp);
2741
2742 if (!tcp_any_retrans_done(sk))
2743 tp->retrans_stamp = 0;
2744
2745 if (flag & FLAG_ECE)
2746 tcp_enter_cwr(sk);
2747
2748 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2749 tcp_try_keep_open(sk);
2750 }
2751 }
2752
tcp_mtup_probe_failed(struct sock * sk)2753 static void tcp_mtup_probe_failed(struct sock *sk)
2754 {
2755 struct inet_connection_sock *icsk = inet_csk(sk);
2756
2757 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2758 icsk->icsk_mtup.probe_size = 0;
2759 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2760 }
2761
tcp_mtup_probe_success(struct sock * sk)2762 static void tcp_mtup_probe_success(struct sock *sk)
2763 {
2764 struct tcp_sock *tp = tcp_sk(sk);
2765 struct inet_connection_sock *icsk = inet_csk(sk);
2766 u64 val;
2767
2768 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2769
2770 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache);
2771 do_div(val, icsk->icsk_mtup.probe_size);
2772 DEBUG_NET_WARN_ON_ONCE((u32)val != val);
2773 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val));
2774
2775 tp->snd_cwnd_cnt = 0;
2776 tp->snd_cwnd_stamp = tcp_jiffies32;
2777 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2778
2779 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2780 icsk->icsk_mtup.probe_size = 0;
2781 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2782 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2783 }
2784
2785 /* Do a simple retransmit without using the backoff mechanisms in
2786 * tcp_timer. This is used for path mtu discovery.
2787 * The socket is already locked here.
2788 */
tcp_simple_retransmit(struct sock * sk)2789 void tcp_simple_retransmit(struct sock *sk)
2790 {
2791 const struct inet_connection_sock *icsk = inet_csk(sk);
2792 struct tcp_sock *tp = tcp_sk(sk);
2793 struct sk_buff *skb;
2794 int mss;
2795
2796 /* A fastopen SYN request is stored as two separate packets within
2797 * the retransmit queue, this is done by tcp_send_syn_data().
2798 * As a result simply checking the MSS of the frames in the queue
2799 * will not work for the SYN packet.
2800 *
2801 * Us being here is an indication of a path MTU issue so we can
2802 * assume that the fastopen SYN was lost and just mark all the
2803 * frames in the retransmit queue as lost. We will use an MSS of
2804 * -1 to mark all frames as lost, otherwise compute the current MSS.
2805 */
2806 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT)
2807 mss = -1;
2808 else
2809 mss = tcp_current_mss(sk);
2810
2811 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2812 if (tcp_skb_seglen(skb) > mss)
2813 tcp_mark_skb_lost(sk, skb);
2814 }
2815
2816 tcp_clear_retrans_hints_partial(tp);
2817
2818 if (!tp->lost_out)
2819 return;
2820
2821 if (tcp_is_reno(tp))
2822 tcp_limit_reno_sacked(tp);
2823
2824 tcp_verify_left_out(tp);
2825
2826 /* Don't muck with the congestion window here.
2827 * Reason is that we do not increase amount of _data_
2828 * in network, but units changed and effective
2829 * cwnd/ssthresh really reduced now.
2830 */
2831 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2832 tp->high_seq = tp->snd_nxt;
2833 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2834 tp->prior_ssthresh = 0;
2835 tp->undo_marker = 0;
2836 tcp_set_ca_state(sk, TCP_CA_Loss);
2837 }
2838 tcp_xmit_retransmit_queue(sk);
2839 }
2840 EXPORT_SYMBOL(tcp_simple_retransmit);
2841
tcp_enter_recovery(struct sock * sk,bool ece_ack)2842 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2843 {
2844 struct tcp_sock *tp = tcp_sk(sk);
2845 int mib_idx;
2846
2847 if (tcp_is_reno(tp))
2848 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2849 else
2850 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2851
2852 NET_INC_STATS(sock_net(sk), mib_idx);
2853
2854 tp->prior_ssthresh = 0;
2855 tcp_init_undo(tp);
2856
2857 if (!tcp_in_cwnd_reduction(sk)) {
2858 if (!ece_ack)
2859 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2860 tcp_init_cwnd_reduction(sk);
2861 }
2862 tcp_set_ca_state(sk, TCP_CA_Recovery);
2863 }
2864
tcp_update_rto_time(struct tcp_sock * tp)2865 static void tcp_update_rto_time(struct tcp_sock *tp)
2866 {
2867 if (tp->rto_stamp) {
2868 tp->total_rto_time += tcp_time_stamp_ms(tp) - tp->rto_stamp;
2869 tp->rto_stamp = 0;
2870 }
2871 }
2872
2873 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2874 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2875 */
tcp_process_loss(struct sock * sk,int flag,int num_dupack,int * rexmit)2876 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2877 int *rexmit)
2878 {
2879 struct tcp_sock *tp = tcp_sk(sk);
2880 bool recovered = !before(tp->snd_una, tp->high_seq);
2881
2882 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2883 tcp_try_undo_loss(sk, false))
2884 return;
2885
2886 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2887 /* Step 3.b. A timeout is spurious if not all data are
2888 * lost, i.e., never-retransmitted data are (s)acked.
2889 */
2890 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2891 tcp_try_undo_loss(sk, true))
2892 return;
2893
2894 if (after(tp->snd_nxt, tp->high_seq)) {
2895 if (flag & FLAG_DATA_SACKED || num_dupack)
2896 tp->frto = 0; /* Step 3.a. loss was real */
2897 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2898 tp->high_seq = tp->snd_nxt;
2899 /* Step 2.b. Try send new data (but deferred until cwnd
2900 * is updated in tcp_ack()). Otherwise fall back to
2901 * the conventional recovery.
2902 */
2903 if (!tcp_write_queue_empty(sk) &&
2904 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2905 *rexmit = REXMIT_NEW;
2906 return;
2907 }
2908 tp->frto = 0;
2909 }
2910 }
2911
2912 if (recovered) {
2913 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2914 tcp_try_undo_recovery(sk);
2915 return;
2916 }
2917 if (tcp_is_reno(tp)) {
2918 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2919 * delivered. Lower inflight to clock out (re)transmissions.
2920 */
2921 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2922 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2923 else if (flag & FLAG_SND_UNA_ADVANCED)
2924 tcp_reset_reno_sack(tp);
2925 }
2926 *rexmit = REXMIT_LOST;
2927 }
2928
tcp_force_fast_retransmit(struct sock * sk)2929 static bool tcp_force_fast_retransmit(struct sock *sk)
2930 {
2931 struct tcp_sock *tp = tcp_sk(sk);
2932
2933 return after(tcp_highest_sack_seq(tp),
2934 tp->snd_una + tp->reordering * tp->mss_cache);
2935 }
2936
2937 /* Undo during fast recovery after partial ACK. */
tcp_try_undo_partial(struct sock * sk,u32 prior_snd_una,bool * do_lost)2938 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
2939 bool *do_lost)
2940 {
2941 struct tcp_sock *tp = tcp_sk(sk);
2942
2943 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2944 /* Plain luck! Hole if filled with delayed
2945 * packet, rather than with a retransmit. Check reordering.
2946 */
2947 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2948
2949 /* We are getting evidence that the reordering degree is higher
2950 * than we realized. If there are no retransmits out then we
2951 * can undo. Otherwise we clock out new packets but do not
2952 * mark more packets lost or retransmit more.
2953 */
2954 if (tp->retrans_out)
2955 return true;
2956
2957 if (!tcp_any_retrans_done(sk))
2958 tp->retrans_stamp = 0;
2959
2960 DBGUNDO(sk, "partial recovery");
2961 tcp_undo_cwnd_reduction(sk, true);
2962 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2963 tcp_try_keep_open(sk);
2964 } else {
2965 /* Partial ACK arrived. Force fast retransmit. */
2966 *do_lost = tcp_force_fast_retransmit(sk);
2967 }
2968 return false;
2969 }
2970
tcp_identify_packet_loss(struct sock * sk,int * ack_flag)2971 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2972 {
2973 struct tcp_sock *tp = tcp_sk(sk);
2974
2975 if (tcp_rtx_queue_empty(sk))
2976 return;
2977
2978 if (unlikely(tcp_is_reno(tp))) {
2979 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2980 } else if (tcp_is_rack(sk)) {
2981 u32 prior_retrans = tp->retrans_out;
2982
2983 if (tcp_rack_mark_lost(sk))
2984 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2985 if (prior_retrans > tp->retrans_out)
2986 *ack_flag |= FLAG_LOST_RETRANS;
2987 }
2988 }
2989
2990 /* Process an event, which can update packets-in-flight not trivially.
2991 * Main goal of this function is to calculate new estimate for left_out,
2992 * taking into account both packets sitting in receiver's buffer and
2993 * packets lost by network.
2994 *
2995 * Besides that it updates the congestion state when packet loss or ECN
2996 * is detected. But it does not reduce the cwnd, it is done by the
2997 * congestion control later.
2998 *
2999 * It does _not_ decide what to send, it is made in function
3000 * tcp_xmit_retransmit_queue().
3001 */
tcp_fastretrans_alert(struct sock * sk,const u32 prior_snd_una,int num_dupack,int * ack_flag,int * rexmit)3002 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
3003 int num_dupack, int *ack_flag, int *rexmit)
3004 {
3005 struct inet_connection_sock *icsk = inet_csk(sk);
3006 struct tcp_sock *tp = tcp_sk(sk);
3007 int fast_rexmit = 0, flag = *ack_flag;
3008 bool ece_ack = flag & FLAG_ECE;
3009 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
3010 tcp_force_fast_retransmit(sk));
3011
3012 if (!tp->packets_out && tp->sacked_out)
3013 tp->sacked_out = 0;
3014
3015 /* Now state machine starts.
3016 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3017 if (ece_ack)
3018 tp->prior_ssthresh = 0;
3019
3020 /* B. In all the states check for reneging SACKs. */
3021 if (tcp_check_sack_reneging(sk, ack_flag))
3022 return;
3023
3024 /* C. Check consistency of the current state. */
3025 tcp_verify_left_out(tp);
3026
3027 /* D. Check state exit conditions. State can be terminated
3028 * when high_seq is ACKed. */
3029 if (icsk->icsk_ca_state == TCP_CA_Open) {
3030 WARN_ON(tp->retrans_out != 0 && !tp->syn_data);
3031 tp->retrans_stamp = 0;
3032 } else if (!before(tp->snd_una, tp->high_seq)) {
3033 switch (icsk->icsk_ca_state) {
3034 case TCP_CA_CWR:
3035 /* CWR is to be held something *above* high_seq
3036 * is ACKed for CWR bit to reach receiver. */
3037 if (tp->snd_una != tp->high_seq) {
3038 tcp_end_cwnd_reduction(sk);
3039 tcp_set_ca_state(sk, TCP_CA_Open);
3040 }
3041 break;
3042
3043 case TCP_CA_Recovery:
3044 if (tcp_is_reno(tp))
3045 tcp_reset_reno_sack(tp);
3046 if (tcp_try_undo_recovery(sk))
3047 return;
3048 tcp_end_cwnd_reduction(sk);
3049 break;
3050 }
3051 }
3052
3053 /* E. Process state. */
3054 switch (icsk->icsk_ca_state) {
3055 case TCP_CA_Recovery:
3056 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3057 if (tcp_is_reno(tp))
3058 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3059 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
3060 return;
3061
3062 if (tcp_try_undo_dsack(sk))
3063 tcp_try_keep_open(sk);
3064
3065 tcp_identify_packet_loss(sk, ack_flag);
3066 if (icsk->icsk_ca_state != TCP_CA_Recovery) {
3067 if (!tcp_time_to_recover(sk, flag))
3068 return;
3069 /* Undo reverts the recovery state. If loss is evident,
3070 * starts a new recovery (e.g. reordering then loss);
3071 */
3072 tcp_enter_recovery(sk, ece_ack);
3073 }
3074 break;
3075 case TCP_CA_Loss:
3076 tcp_process_loss(sk, flag, num_dupack, rexmit);
3077 if (icsk->icsk_ca_state != TCP_CA_Loss)
3078 tcp_update_rto_time(tp);
3079 tcp_identify_packet_loss(sk, ack_flag);
3080 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
3081 (*ack_flag & FLAG_LOST_RETRANS)))
3082 return;
3083 /* Change state if cwnd is undone or retransmits are lost */
3084 fallthrough;
3085 default:
3086 if (tcp_is_reno(tp)) {
3087 if (flag & FLAG_SND_UNA_ADVANCED)
3088 tcp_reset_reno_sack(tp);
3089 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3090 }
3091
3092 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3093 tcp_try_undo_dsack(sk);
3094
3095 tcp_identify_packet_loss(sk, ack_flag);
3096 if (!tcp_time_to_recover(sk, flag)) {
3097 tcp_try_to_open(sk, flag);
3098 return;
3099 }
3100
3101 /* MTU probe failure: don't reduce cwnd */
3102 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3103 icsk->icsk_mtup.probe_size &&
3104 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3105 tcp_mtup_probe_failed(sk);
3106 /* Restores the reduction we did in tcp_mtup_probe() */
3107 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
3108 tcp_simple_retransmit(sk);
3109 return;
3110 }
3111
3112 /* Otherwise enter Recovery state */
3113 tcp_enter_recovery(sk, ece_ack);
3114 fast_rexmit = 1;
3115 }
3116
3117 if (!tcp_is_rack(sk) && do_lost)
3118 tcp_update_scoreboard(sk, fast_rexmit);
3119 *rexmit = REXMIT_LOST;
3120 }
3121
tcp_update_rtt_min(struct sock * sk,u32 rtt_us,const int flag)3122 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3123 {
3124 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
3125 struct tcp_sock *tp = tcp_sk(sk);
3126
3127 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3128 /* If the remote keeps returning delayed ACKs, eventually
3129 * the min filter would pick it up and overestimate the
3130 * prop. delay when it expires. Skip suspected delayed ACKs.
3131 */
3132 return;
3133 }
3134 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3135 rtt_us ? : jiffies_to_usecs(1));
3136 }
3137
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)3138 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3139 long seq_rtt_us, long sack_rtt_us,
3140 long ca_rtt_us, struct rate_sample *rs)
3141 {
3142 const struct tcp_sock *tp = tcp_sk(sk);
3143
3144 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3145 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3146 * Karn's algorithm forbids taking RTT if some retransmitted data
3147 * is acked (RFC6298).
3148 */
3149 if (seq_rtt_us < 0)
3150 seq_rtt_us = sack_rtt_us;
3151
3152 /* RTTM Rule: A TSecr value received in a segment is used to
3153 * update the averaged RTT measurement only if the segment
3154 * acknowledges some new data, i.e., only if it advances the
3155 * left edge of the send window.
3156 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3157 */
3158 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp &&
3159 tp->rx_opt.rcv_tsecr && flag & FLAG_ACKED)
3160 seq_rtt_us = ca_rtt_us = tcp_rtt_tsopt_us(tp);
3161
3162 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3163 if (seq_rtt_us < 0)
3164 return false;
3165
3166 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3167 * always taken together with ACK, SACK, or TS-opts. Any negative
3168 * values will be skipped with the seq_rtt_us < 0 check above.
3169 */
3170 tcp_update_rtt_min(sk, ca_rtt_us, flag);
3171 tcp_rtt_estimator(sk, seq_rtt_us);
3172 tcp_set_rto(sk);
3173
3174 /* RFC6298: only reset backoff on valid RTT measurement. */
3175 inet_csk(sk)->icsk_backoff = 0;
3176 return true;
3177 }
3178
3179 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
tcp_synack_rtt_meas(struct sock * sk,struct request_sock * req)3180 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3181 {
3182 struct rate_sample rs;
3183 long rtt_us = -1L;
3184
3185 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3186 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3187
3188 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3189 }
3190
3191
tcp_cong_avoid(struct sock * sk,u32 ack,u32 acked)3192 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3193 {
3194 const struct inet_connection_sock *icsk = inet_csk(sk);
3195
3196 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3197 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3198 }
3199
3200 /* Restart timer after forward progress on connection.
3201 * RFC2988 recommends to restart timer to now+rto.
3202 */
tcp_rearm_rto(struct sock * sk)3203 void tcp_rearm_rto(struct sock *sk)
3204 {
3205 const struct inet_connection_sock *icsk = inet_csk(sk);
3206 struct tcp_sock *tp = tcp_sk(sk);
3207
3208 /* If the retrans timer is currently being used by Fast Open
3209 * for SYN-ACK retrans purpose, stay put.
3210 */
3211 if (rcu_access_pointer(tp->fastopen_rsk))
3212 return;
3213
3214 if (!tp->packets_out) {
3215 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3216 } else {
3217 u32 rto = inet_csk(sk)->icsk_rto;
3218 /* Offset the time elapsed after installing regular RTO */
3219 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3220 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3221 s64 delta_us = tcp_rto_delta_us(sk);
3222 /* delta_us may not be positive if the socket is locked
3223 * when the retrans timer fires and is rescheduled.
3224 */
3225 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3226 }
3227 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3228 TCP_RTO_MAX);
3229 }
3230 }
3231
3232 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
tcp_set_xmit_timer(struct sock * sk)3233 static void tcp_set_xmit_timer(struct sock *sk)
3234 {
3235 if (!tcp_schedule_loss_probe(sk, true))
3236 tcp_rearm_rto(sk);
3237 }
3238
3239 /* If we get here, the whole TSO packet has not been acked. */
tcp_tso_acked(struct sock * sk,struct sk_buff * skb)3240 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3241 {
3242 struct tcp_sock *tp = tcp_sk(sk);
3243 u32 packets_acked;
3244
3245 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3246
3247 packets_acked = tcp_skb_pcount(skb);
3248 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3249 return 0;
3250 packets_acked -= tcp_skb_pcount(skb);
3251
3252 if (packets_acked) {
3253 BUG_ON(tcp_skb_pcount(skb) == 0);
3254 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3255 }
3256
3257 return packets_acked;
3258 }
3259
tcp_ack_tstamp(struct sock * sk,struct sk_buff * skb,const struct sk_buff * ack_skb,u32 prior_snd_una)3260 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3261 const struct sk_buff *ack_skb, u32 prior_snd_una)
3262 {
3263 const struct skb_shared_info *shinfo;
3264
3265 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3266 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3267 return;
3268
3269 shinfo = skb_shinfo(skb);
3270 if (!before(shinfo->tskey, prior_snd_una) &&
3271 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3272 tcp_skb_tsorted_save(skb) {
3273 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK);
3274 } tcp_skb_tsorted_restore(skb);
3275 }
3276 }
3277
3278 /* Remove acknowledged frames from the retransmission queue. If our packet
3279 * is before the ack sequence we can discard it as it's confirmed to have
3280 * arrived at the other end.
3281 */
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)3282 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb,
3283 u32 prior_fack, u32 prior_snd_una,
3284 struct tcp_sacktag_state *sack, bool ece_ack)
3285 {
3286 const struct inet_connection_sock *icsk = inet_csk(sk);
3287 u64 first_ackt, last_ackt;
3288 struct tcp_sock *tp = tcp_sk(sk);
3289 u32 prior_sacked = tp->sacked_out;
3290 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3291 struct sk_buff *skb, *next;
3292 bool fully_acked = true;
3293 long sack_rtt_us = -1L;
3294 long seq_rtt_us = -1L;
3295 long ca_rtt_us = -1L;
3296 u32 pkts_acked = 0;
3297 bool rtt_update;
3298 int flag = 0;
3299
3300 first_ackt = 0;
3301
3302 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3303 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3304 const u32 start_seq = scb->seq;
3305 u8 sacked = scb->sacked;
3306 u32 acked_pcount;
3307
3308 /* Determine how many packets and what bytes were acked, tso and else */
3309 if (after(scb->end_seq, tp->snd_una)) {
3310 if (tcp_skb_pcount(skb) == 1 ||
3311 !after(tp->snd_una, scb->seq))
3312 break;
3313
3314 acked_pcount = tcp_tso_acked(sk, skb);
3315 if (!acked_pcount)
3316 break;
3317 fully_acked = false;
3318 } else {
3319 acked_pcount = tcp_skb_pcount(skb);
3320 }
3321
3322 if (unlikely(sacked & TCPCB_RETRANS)) {
3323 if (sacked & TCPCB_SACKED_RETRANS)
3324 tp->retrans_out -= acked_pcount;
3325 flag |= FLAG_RETRANS_DATA_ACKED;
3326 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3327 last_ackt = tcp_skb_timestamp_us(skb);
3328 WARN_ON_ONCE(last_ackt == 0);
3329 if (!first_ackt)
3330 first_ackt = last_ackt;
3331
3332 if (before(start_seq, reord))
3333 reord = start_seq;
3334 if (!after(scb->end_seq, tp->high_seq))
3335 flag |= FLAG_ORIG_SACK_ACKED;
3336 }
3337
3338 if (sacked & TCPCB_SACKED_ACKED) {
3339 tp->sacked_out -= acked_pcount;
3340 } else if (tcp_is_sack(tp)) {
3341 tcp_count_delivered(tp, acked_pcount, ece_ack);
3342 if (!tcp_skb_spurious_retrans(tp, skb))
3343 tcp_rack_advance(tp, sacked, scb->end_seq,
3344 tcp_skb_timestamp_us(skb));
3345 }
3346 if (sacked & TCPCB_LOST)
3347 tp->lost_out -= acked_pcount;
3348
3349 tp->packets_out -= acked_pcount;
3350 pkts_acked += acked_pcount;
3351 tcp_rate_skb_delivered(sk, skb, sack->rate);
3352
3353 /* Initial outgoing SYN's get put onto the write_queue
3354 * just like anything else we transmit. It is not
3355 * true data, and if we misinform our callers that
3356 * this ACK acks real data, we will erroneously exit
3357 * connection startup slow start one packet too
3358 * quickly. This is severely frowned upon behavior.
3359 */
3360 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3361 flag |= FLAG_DATA_ACKED;
3362 } else {
3363 flag |= FLAG_SYN_ACKED;
3364 tp->retrans_stamp = 0;
3365 }
3366
3367 if (!fully_acked)
3368 break;
3369
3370 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3371
3372 next = skb_rb_next(skb);
3373 if (unlikely(skb == tp->retransmit_skb_hint))
3374 tp->retransmit_skb_hint = NULL;
3375 if (unlikely(skb == tp->lost_skb_hint))
3376 tp->lost_skb_hint = NULL;
3377 tcp_highest_sack_replace(sk, skb, next);
3378 tcp_rtx_queue_unlink_and_free(skb, sk);
3379 }
3380
3381 if (!skb)
3382 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3383
3384 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3385 tp->snd_up = tp->snd_una;
3386
3387 if (skb) {
3388 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3389 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3390 flag |= FLAG_SACK_RENEGING;
3391 }
3392
3393 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3394 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3395 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3396
3397 if (pkts_acked == 1 && fully_acked && !prior_sacked &&
3398 (tp->snd_una - prior_snd_una) < tp->mss_cache &&
3399 sack->rate->prior_delivered + 1 == tp->delivered &&
3400 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3401 /* Conservatively mark a delayed ACK. It's typically
3402 * from a lone runt packet over the round trip to
3403 * a receiver w/o out-of-order or CE events.
3404 */
3405 flag |= FLAG_ACK_MAYBE_DELAYED;
3406 }
3407 }
3408 if (sack->first_sackt) {
3409 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3410 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3411 }
3412 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3413 ca_rtt_us, sack->rate);
3414
3415 if (flag & FLAG_ACKED) {
3416 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3417 if (unlikely(icsk->icsk_mtup.probe_size &&
3418 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3419 tcp_mtup_probe_success(sk);
3420 }
3421
3422 if (tcp_is_reno(tp)) {
3423 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3424
3425 /* If any of the cumulatively ACKed segments was
3426 * retransmitted, non-SACK case cannot confirm that
3427 * progress was due to original transmission due to
3428 * lack of TCPCB_SACKED_ACKED bits even if some of
3429 * the packets may have been never retransmitted.
3430 */
3431 if (flag & FLAG_RETRANS_DATA_ACKED)
3432 flag &= ~FLAG_ORIG_SACK_ACKED;
3433 } else {
3434 int delta;
3435
3436 /* Non-retransmitted hole got filled? That's reordering */
3437 if (before(reord, prior_fack))
3438 tcp_check_sack_reordering(sk, reord, 0);
3439
3440 delta = prior_sacked - tp->sacked_out;
3441 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3442 }
3443 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3444 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3445 tcp_skb_timestamp_us(skb))) {
3446 /* Do not re-arm RTO if the sack RTT is measured from data sent
3447 * after when the head was last (re)transmitted. Otherwise the
3448 * timeout may continue to extend in loss recovery.
3449 */
3450 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3451 }
3452
3453 if (icsk->icsk_ca_ops->pkts_acked) {
3454 struct ack_sample sample = { .pkts_acked = pkts_acked,
3455 .rtt_us = sack->rate->rtt_us };
3456
3457 sample.in_flight = tp->mss_cache *
3458 (tp->delivered - sack->rate->prior_delivered);
3459 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3460 }
3461
3462 #if FASTRETRANS_DEBUG > 0
3463 WARN_ON((int)tp->sacked_out < 0);
3464 WARN_ON((int)tp->lost_out < 0);
3465 WARN_ON((int)tp->retrans_out < 0);
3466 if (!tp->packets_out && tcp_is_sack(tp)) {
3467 icsk = inet_csk(sk);
3468 if (tp->lost_out) {
3469 pr_debug("Leak l=%u %d\n",
3470 tp->lost_out, icsk->icsk_ca_state);
3471 tp->lost_out = 0;
3472 }
3473 if (tp->sacked_out) {
3474 pr_debug("Leak s=%u %d\n",
3475 tp->sacked_out, icsk->icsk_ca_state);
3476 tp->sacked_out = 0;
3477 }
3478 if (tp->retrans_out) {
3479 pr_debug("Leak r=%u %d\n",
3480 tp->retrans_out, icsk->icsk_ca_state);
3481 tp->retrans_out = 0;
3482 }
3483 }
3484 #endif
3485 return flag;
3486 }
3487
tcp_ack_probe(struct sock * sk)3488 static void tcp_ack_probe(struct sock *sk)
3489 {
3490 struct inet_connection_sock *icsk = inet_csk(sk);
3491 struct sk_buff *head = tcp_send_head(sk);
3492 const struct tcp_sock *tp = tcp_sk(sk);
3493
3494 /* Was it a usable window open? */
3495 if (!head)
3496 return;
3497 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3498 icsk->icsk_backoff = 0;
3499 icsk->icsk_probes_tstamp = 0;
3500 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3501 /* Socket must be waked up by subsequent tcp_data_snd_check().
3502 * This function is not for random using!
3503 */
3504 } else {
3505 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3506
3507 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3508 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
3509 }
3510 }
3511
tcp_ack_is_dubious(const struct sock * sk,const int flag)3512 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3513 {
3514 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3515 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3516 }
3517
3518 /* Decide wheather to run the increase function of congestion control. */
tcp_may_raise_cwnd(const struct sock * sk,const int flag)3519 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3520 {
3521 /* If reordering is high then always grow cwnd whenever data is
3522 * delivered regardless of its ordering. Otherwise stay conservative
3523 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3524 * new SACK or ECE mark may first advance cwnd here and later reduce
3525 * cwnd in tcp_fastretrans_alert() based on more states.
3526 */
3527 if (tcp_sk(sk)->reordering >
3528 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3529 return flag & FLAG_FORWARD_PROGRESS;
3530
3531 return flag & FLAG_DATA_ACKED;
3532 }
3533
3534 /* The "ultimate" congestion control function that aims to replace the rigid
3535 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3536 * It's called toward the end of processing an ACK with precise rate
3537 * information. All transmission or retransmission are delayed afterwards.
3538 */
tcp_cong_control(struct sock * sk,u32 ack,u32 acked_sacked,int flag,const struct rate_sample * rs)3539 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3540 int flag, const struct rate_sample *rs)
3541 {
3542 const struct inet_connection_sock *icsk = inet_csk(sk);
3543
3544 if (icsk->icsk_ca_ops->cong_control) {
3545 icsk->icsk_ca_ops->cong_control(sk, ack, flag, rs);
3546 return;
3547 }
3548
3549 if (tcp_in_cwnd_reduction(sk)) {
3550 /* Reduce cwnd if state mandates */
3551 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag);
3552 } else if (tcp_may_raise_cwnd(sk, flag)) {
3553 /* Advance cwnd if state allows */
3554 tcp_cong_avoid(sk, ack, acked_sacked);
3555 }
3556 tcp_update_pacing_rate(sk);
3557 }
3558
3559 /* Check that window update is acceptable.
3560 * The function assumes that snd_una<=ack<=snd_next.
3561 */
tcp_may_update_window(const struct tcp_sock * tp,const u32 ack,const u32 ack_seq,const u32 nwin)3562 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3563 const u32 ack, const u32 ack_seq,
3564 const u32 nwin)
3565 {
3566 return after(ack, tp->snd_una) ||
3567 after(ack_seq, tp->snd_wl1) ||
3568 (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin));
3569 }
3570
tcp_snd_sne_update(struct tcp_sock * tp,u32 ack)3571 static void tcp_snd_sne_update(struct tcp_sock *tp, u32 ack)
3572 {
3573 #ifdef CONFIG_TCP_AO
3574 struct tcp_ao_info *ao;
3575
3576 if (!static_branch_unlikely(&tcp_ao_needed.key))
3577 return;
3578
3579 ao = rcu_dereference_protected(tp->ao_info,
3580 lockdep_sock_is_held((struct sock *)tp));
3581 if (ao && ack < tp->snd_una)
3582 ao->snd_sne++;
3583 #endif
3584 }
3585
3586 /* If we update tp->snd_una, also update tp->bytes_acked */
tcp_snd_una_update(struct tcp_sock * tp,u32 ack)3587 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3588 {
3589 u32 delta = ack - tp->snd_una;
3590
3591 sock_owned_by_me((struct sock *)tp);
3592 tp->bytes_acked += delta;
3593 tcp_snd_sne_update(tp, ack);
3594 tp->snd_una = ack;
3595 }
3596
tcp_rcv_sne_update(struct tcp_sock * tp,u32 seq)3597 static void tcp_rcv_sne_update(struct tcp_sock *tp, u32 seq)
3598 {
3599 #ifdef CONFIG_TCP_AO
3600 struct tcp_ao_info *ao;
3601
3602 if (!static_branch_unlikely(&tcp_ao_needed.key))
3603 return;
3604
3605 ao = rcu_dereference_protected(tp->ao_info,
3606 lockdep_sock_is_held((struct sock *)tp));
3607 if (ao && seq < tp->rcv_nxt)
3608 ao->rcv_sne++;
3609 #endif
3610 }
3611
3612 /* If we update tp->rcv_nxt, also update tp->bytes_received */
tcp_rcv_nxt_update(struct tcp_sock * tp,u32 seq)3613 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3614 {
3615 u32 delta = seq - tp->rcv_nxt;
3616
3617 sock_owned_by_me((struct sock *)tp);
3618 tp->bytes_received += delta;
3619 tcp_rcv_sne_update(tp, seq);
3620 WRITE_ONCE(tp->rcv_nxt, seq);
3621 }
3622
3623 /* Update our send window.
3624 *
3625 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3626 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3627 */
tcp_ack_update_window(struct sock * sk,const struct sk_buff * skb,u32 ack,u32 ack_seq)3628 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3629 u32 ack_seq)
3630 {
3631 struct tcp_sock *tp = tcp_sk(sk);
3632 int flag = 0;
3633 u32 nwin = ntohs(tcp_hdr(skb)->window);
3634
3635 if (likely(!tcp_hdr(skb)->syn))
3636 nwin <<= tp->rx_opt.snd_wscale;
3637
3638 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3639 flag |= FLAG_WIN_UPDATE;
3640 tcp_update_wl(tp, ack_seq);
3641
3642 if (tp->snd_wnd != nwin) {
3643 tp->snd_wnd = nwin;
3644
3645 /* Note, it is the only place, where
3646 * fast path is recovered for sending TCP.
3647 */
3648 tp->pred_flags = 0;
3649 tcp_fast_path_check(sk);
3650
3651 if (!tcp_write_queue_empty(sk))
3652 tcp_slow_start_after_idle_check(sk);
3653
3654 if (nwin > tp->max_window) {
3655 tp->max_window = nwin;
3656 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3657 }
3658 }
3659 }
3660
3661 tcp_snd_una_update(tp, ack);
3662
3663 return flag;
3664 }
3665
__tcp_oow_rate_limited(struct net * net,int mib_idx,u32 * last_oow_ack_time)3666 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3667 u32 *last_oow_ack_time)
3668 {
3669 /* Paired with the WRITE_ONCE() in this function. */
3670 u32 val = READ_ONCE(*last_oow_ack_time);
3671
3672 if (val) {
3673 s32 elapsed = (s32)(tcp_jiffies32 - val);
3674
3675 if (0 <= elapsed &&
3676 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3677 NET_INC_STATS(net, mib_idx);
3678 return true; /* rate-limited: don't send yet! */
3679 }
3680 }
3681
3682 /* Paired with the prior READ_ONCE() and with itself,
3683 * as we might be lockless.
3684 */
3685 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32);
3686
3687 return false; /* not rate-limited: go ahead, send dupack now! */
3688 }
3689
3690 /* Return true if we're currently rate-limiting out-of-window ACKs and
3691 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3692 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3693 * attacks that send repeated SYNs or ACKs for the same connection. To
3694 * do this, we do not send a duplicate SYNACK or ACK if the remote
3695 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3696 */
tcp_oow_rate_limited(struct net * net,const struct sk_buff * skb,int mib_idx,u32 * last_oow_ack_time)3697 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3698 int mib_idx, u32 *last_oow_ack_time)
3699 {
3700 /* Data packets without SYNs are not likely part of an ACK loop. */
3701 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3702 !tcp_hdr(skb)->syn)
3703 return false;
3704
3705 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3706 }
3707
3708 /* RFC 5961 7 [ACK Throttling] */
tcp_send_challenge_ack(struct sock * sk)3709 static void tcp_send_challenge_ack(struct sock *sk)
3710 {
3711 struct tcp_sock *tp = tcp_sk(sk);
3712 struct net *net = sock_net(sk);
3713 u32 count, now, ack_limit;
3714
3715 /* First check our per-socket dupack rate limit. */
3716 if (__tcp_oow_rate_limited(net,
3717 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3718 &tp->last_oow_ack_time))
3719 return;
3720
3721 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3722 if (ack_limit == INT_MAX)
3723 goto send_ack;
3724
3725 /* Then check host-wide RFC 5961 rate limit. */
3726 now = jiffies / HZ;
3727 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) {
3728 u32 half = (ack_limit + 1) >> 1;
3729
3730 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now);
3731 WRITE_ONCE(net->ipv4.tcp_challenge_count,
3732 get_random_u32_inclusive(half, ack_limit + half - 1));
3733 }
3734 count = READ_ONCE(net->ipv4.tcp_challenge_count);
3735 if (count > 0) {
3736 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1);
3737 send_ack:
3738 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3739 tcp_send_ack(sk);
3740 }
3741 }
3742
tcp_store_ts_recent(struct tcp_sock * tp)3743 static void tcp_store_ts_recent(struct tcp_sock *tp)
3744 {
3745 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3746 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3747 }
3748
tcp_replace_ts_recent(struct tcp_sock * tp,u32 seq)3749 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3750 {
3751 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3752 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3753 * extra check below makes sure this can only happen
3754 * for pure ACK frames. -DaveM
3755 *
3756 * Not only, also it occurs for expired timestamps.
3757 */
3758
3759 if (tcp_paws_check(&tp->rx_opt, 0))
3760 tcp_store_ts_recent(tp);
3761 }
3762 }
3763
3764 /* This routine deals with acks during a TLP episode and ends an episode by
3765 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3766 */
tcp_process_tlp_ack(struct sock * sk,u32 ack,int flag)3767 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3768 {
3769 struct tcp_sock *tp = tcp_sk(sk);
3770
3771 if (before(ack, tp->tlp_high_seq))
3772 return;
3773
3774 if (!tp->tlp_retrans) {
3775 /* TLP of new data has been acknowledged */
3776 tp->tlp_high_seq = 0;
3777 } else if (flag & FLAG_DSACK_TLP) {
3778 /* This DSACK means original and TLP probe arrived; no loss */
3779 tp->tlp_high_seq = 0;
3780 } else if (after(ack, tp->tlp_high_seq)) {
3781 /* ACK advances: there was a loss, so reduce cwnd. Reset
3782 * tlp_high_seq in tcp_init_cwnd_reduction()
3783 */
3784 tcp_init_cwnd_reduction(sk);
3785 tcp_set_ca_state(sk, TCP_CA_CWR);
3786 tcp_end_cwnd_reduction(sk);
3787 tcp_try_keep_open(sk);
3788 NET_INC_STATS(sock_net(sk),
3789 LINUX_MIB_TCPLOSSPROBERECOVERY);
3790 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3791 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3792 /* Pure dupack: original and TLP probe arrived; no loss */
3793 tp->tlp_high_seq = 0;
3794 }
3795 }
3796
tcp_in_ack_event(struct sock * sk,u32 flags)3797 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3798 {
3799 const struct inet_connection_sock *icsk = inet_csk(sk);
3800
3801 if (icsk->icsk_ca_ops->in_ack_event)
3802 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3803 }
3804
3805 /* Congestion control has updated the cwnd already. So if we're in
3806 * loss recovery then now we do any new sends (for FRTO) or
3807 * retransmits (for CA_Loss or CA_recovery) that make sense.
3808 */
tcp_xmit_recovery(struct sock * sk,int rexmit)3809 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3810 {
3811 struct tcp_sock *tp = tcp_sk(sk);
3812
3813 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3814 return;
3815
3816 if (unlikely(rexmit == REXMIT_NEW)) {
3817 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3818 TCP_NAGLE_OFF);
3819 if (after(tp->snd_nxt, tp->high_seq))
3820 return;
3821 tp->frto = 0;
3822 }
3823 tcp_xmit_retransmit_queue(sk);
3824 }
3825
3826 /* Returns the number of packets newly acked or sacked by the current ACK */
tcp_newly_delivered(struct sock * sk,u32 prior_delivered,int flag)3827 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3828 {
3829 const struct net *net = sock_net(sk);
3830 struct tcp_sock *tp = tcp_sk(sk);
3831 u32 delivered;
3832
3833 delivered = tp->delivered - prior_delivered;
3834 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3835 if (flag & FLAG_ECE)
3836 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3837
3838 return delivered;
3839 }
3840
3841 /* This routine deals with incoming acks, but not outgoing ones. */
tcp_ack(struct sock * sk,const struct sk_buff * skb,int flag)3842 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3843 {
3844 struct inet_connection_sock *icsk = inet_csk(sk);
3845 struct tcp_sock *tp = tcp_sk(sk);
3846 struct tcp_sacktag_state sack_state;
3847 struct rate_sample rs = { .prior_delivered = 0 };
3848 u32 prior_snd_una = tp->snd_una;
3849 bool is_sack_reneg = tp->is_sack_reneg;
3850 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3851 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3852 int num_dupack = 0;
3853 int prior_packets = tp->packets_out;
3854 u32 delivered = tp->delivered;
3855 u32 lost = tp->lost;
3856 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3857 u32 prior_fack;
3858
3859 sack_state.first_sackt = 0;
3860 sack_state.rate = &rs;
3861 sack_state.sack_delivered = 0;
3862
3863 /* We very likely will need to access rtx queue. */
3864 prefetch(sk->tcp_rtx_queue.rb_node);
3865
3866 /* If the ack is older than previous acks
3867 * then we can probably ignore it.
3868 */
3869 if (before(ack, prior_snd_una)) {
3870 u32 max_window;
3871
3872 /* do not accept ACK for bytes we never sent. */
3873 max_window = min_t(u64, tp->max_window, tp->bytes_acked);
3874 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3875 if (before(ack, prior_snd_una - max_window)) {
3876 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3877 tcp_send_challenge_ack(sk);
3878 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK;
3879 }
3880 goto old_ack;
3881 }
3882
3883 /* If the ack includes data we haven't sent yet, discard
3884 * this segment (RFC793 Section 3.9).
3885 */
3886 if (after(ack, tp->snd_nxt))
3887 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA;
3888
3889 if (after(ack, prior_snd_una)) {
3890 flag |= FLAG_SND_UNA_ADVANCED;
3891 icsk->icsk_retransmits = 0;
3892
3893 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3894 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3895 if (icsk->icsk_clean_acked)
3896 icsk->icsk_clean_acked(sk, ack);
3897 #endif
3898 }
3899
3900 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3901 rs.prior_in_flight = tcp_packets_in_flight(tp);
3902
3903 /* ts_recent update must be made after we are sure that the packet
3904 * is in window.
3905 */
3906 if (flag & FLAG_UPDATE_TS_RECENT)
3907 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3908
3909 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3910 FLAG_SND_UNA_ADVANCED) {
3911 /* Window is constant, pure forward advance.
3912 * No more checks are required.
3913 * Note, we use the fact that SND.UNA>=SND.WL2.
3914 */
3915 tcp_update_wl(tp, ack_seq);
3916 tcp_snd_una_update(tp, ack);
3917 flag |= FLAG_WIN_UPDATE;
3918
3919 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3920
3921 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3922 } else {
3923 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3924
3925 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3926 flag |= FLAG_DATA;
3927 else
3928 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3929
3930 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3931
3932 if (TCP_SKB_CB(skb)->sacked)
3933 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3934 &sack_state);
3935
3936 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3937 flag |= FLAG_ECE;
3938 ack_ev_flags |= CA_ACK_ECE;
3939 }
3940
3941 if (sack_state.sack_delivered)
3942 tcp_count_delivered(tp, sack_state.sack_delivered,
3943 flag & FLAG_ECE);
3944
3945 if (flag & FLAG_WIN_UPDATE)
3946 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3947
3948 tcp_in_ack_event(sk, ack_ev_flags);
3949 }
3950
3951 /* This is a deviation from RFC3168 since it states that:
3952 * "When the TCP data sender is ready to set the CWR bit after reducing
3953 * the congestion window, it SHOULD set the CWR bit only on the first
3954 * new data packet that it transmits."
3955 * We accept CWR on pure ACKs to be more robust
3956 * with widely-deployed TCP implementations that do this.
3957 */
3958 tcp_ecn_accept_cwr(sk, skb);
3959
3960 /* We passed data and got it acked, remove any soft error
3961 * log. Something worked...
3962 */
3963 WRITE_ONCE(sk->sk_err_soft, 0);
3964 icsk->icsk_probes_out = 0;
3965 tp->rcv_tstamp = tcp_jiffies32;
3966 if (!prior_packets)
3967 goto no_queue;
3968
3969 /* See if we can take anything off of the retransmit queue. */
3970 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una,
3971 &sack_state, flag & FLAG_ECE);
3972
3973 tcp_rack_update_reo_wnd(sk, &rs);
3974
3975 if (tp->tlp_high_seq)
3976 tcp_process_tlp_ack(sk, ack, flag);
3977
3978 if (tcp_ack_is_dubious(sk, flag)) {
3979 if (!(flag & (FLAG_SND_UNA_ADVANCED |
3980 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
3981 num_dupack = 1;
3982 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3983 if (!(flag & FLAG_DATA))
3984 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3985 }
3986 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3987 &rexmit);
3988 }
3989
3990 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3991 if (flag & FLAG_SET_XMIT_TIMER)
3992 tcp_set_xmit_timer(sk);
3993
3994 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3995 sk_dst_confirm(sk);
3996
3997 delivered = tcp_newly_delivered(sk, delivered, flag);
3998 lost = tp->lost - lost; /* freshly marked lost */
3999 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
4000 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
4001 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
4002 tcp_xmit_recovery(sk, rexmit);
4003 return 1;
4004
4005 no_queue:
4006 /* If data was DSACKed, see if we can undo a cwnd reduction. */
4007 if (flag & FLAG_DSACKING_ACK) {
4008 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4009 &rexmit);
4010 tcp_newly_delivered(sk, delivered, flag);
4011 }
4012 /* If this ack opens up a zero window, clear backoff. It was
4013 * being used to time the probes, and is probably far higher than
4014 * it needs to be for normal retransmission.
4015 */
4016 tcp_ack_probe(sk);
4017
4018 if (tp->tlp_high_seq)
4019 tcp_process_tlp_ack(sk, ack, flag);
4020 return 1;
4021
4022 old_ack:
4023 /* If data was SACKed, tag it and see if we should send more data.
4024 * If data was DSACKed, see if we can undo a cwnd reduction.
4025 */
4026 if (TCP_SKB_CB(skb)->sacked) {
4027 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
4028 &sack_state);
4029 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4030 &rexmit);
4031 tcp_newly_delivered(sk, delivered, flag);
4032 tcp_xmit_recovery(sk, rexmit);
4033 }
4034
4035 return 0;
4036 }
4037
tcp_parse_fastopen_option(int len,const unsigned char * cookie,bool syn,struct tcp_fastopen_cookie * foc,bool exp_opt)4038 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
4039 bool syn, struct tcp_fastopen_cookie *foc,
4040 bool exp_opt)
4041 {
4042 /* Valid only in SYN or SYN-ACK with an even length. */
4043 if (!foc || !syn || len < 0 || (len & 1))
4044 return;
4045
4046 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
4047 len <= TCP_FASTOPEN_COOKIE_MAX)
4048 memcpy(foc->val, cookie, len);
4049 else if (len != 0)
4050 len = -1;
4051 foc->len = len;
4052 foc->exp = exp_opt;
4053 }
4054
smc_parse_options(const struct tcphdr * th,struct tcp_options_received * opt_rx,const unsigned char * ptr,int opsize)4055 static bool smc_parse_options(const struct tcphdr *th,
4056 struct tcp_options_received *opt_rx,
4057 const unsigned char *ptr,
4058 int opsize)
4059 {
4060 #if IS_ENABLED(CONFIG_SMC)
4061 if (static_branch_unlikely(&tcp_have_smc)) {
4062 if (th->syn && !(opsize & 1) &&
4063 opsize >= TCPOLEN_EXP_SMC_BASE &&
4064 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
4065 opt_rx->smc_ok = 1;
4066 return true;
4067 }
4068 }
4069 #endif
4070 return false;
4071 }
4072
4073 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
4074 * value on success.
4075 */
tcp_parse_mss_option(const struct tcphdr * th,u16 user_mss)4076 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
4077 {
4078 const unsigned char *ptr = (const unsigned char *)(th + 1);
4079 int length = (th->doff * 4) - sizeof(struct tcphdr);
4080 u16 mss = 0;
4081
4082 while (length > 0) {
4083 int opcode = *ptr++;
4084 int opsize;
4085
4086 switch (opcode) {
4087 case TCPOPT_EOL:
4088 return mss;
4089 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4090 length--;
4091 continue;
4092 default:
4093 if (length < 2)
4094 return mss;
4095 opsize = *ptr++;
4096 if (opsize < 2) /* "silly options" */
4097 return mss;
4098 if (opsize > length)
4099 return mss; /* fail on partial options */
4100 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
4101 u16 in_mss = get_unaligned_be16(ptr);
4102
4103 if (in_mss) {
4104 if (user_mss && user_mss < in_mss)
4105 in_mss = user_mss;
4106 mss = in_mss;
4107 }
4108 }
4109 ptr += opsize - 2;
4110 length -= opsize;
4111 }
4112 }
4113 return mss;
4114 }
4115 EXPORT_SYMBOL_GPL(tcp_parse_mss_option);
4116
4117 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
4118 * But, this can also be called on packets in the established flow when
4119 * the fast version below fails.
4120 */
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)4121 void tcp_parse_options(const struct net *net,
4122 const struct sk_buff *skb,
4123 struct tcp_options_received *opt_rx, int estab,
4124 struct tcp_fastopen_cookie *foc)
4125 {
4126 const unsigned char *ptr;
4127 const struct tcphdr *th = tcp_hdr(skb);
4128 int length = (th->doff * 4) - sizeof(struct tcphdr);
4129
4130 ptr = (const unsigned char *)(th + 1);
4131 opt_rx->saw_tstamp = 0;
4132 opt_rx->saw_unknown = 0;
4133
4134 while (length > 0) {
4135 int opcode = *ptr++;
4136 int opsize;
4137
4138 switch (opcode) {
4139 case TCPOPT_EOL:
4140 return;
4141 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4142 length--;
4143 continue;
4144 default:
4145 if (length < 2)
4146 return;
4147 opsize = *ptr++;
4148 if (opsize < 2) /* "silly options" */
4149 return;
4150 if (opsize > length)
4151 return; /* don't parse partial options */
4152 switch (opcode) {
4153 case TCPOPT_MSS:
4154 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
4155 u16 in_mss = get_unaligned_be16(ptr);
4156 if (in_mss) {
4157 if (opt_rx->user_mss &&
4158 opt_rx->user_mss < in_mss)
4159 in_mss = opt_rx->user_mss;
4160 opt_rx->mss_clamp = in_mss;
4161 }
4162 }
4163 break;
4164 case TCPOPT_WINDOW:
4165 if (opsize == TCPOLEN_WINDOW && th->syn &&
4166 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) {
4167 __u8 snd_wscale = *(__u8 *)ptr;
4168 opt_rx->wscale_ok = 1;
4169 if (snd_wscale > TCP_MAX_WSCALE) {
4170 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4171 __func__,
4172 snd_wscale,
4173 TCP_MAX_WSCALE);
4174 snd_wscale = TCP_MAX_WSCALE;
4175 }
4176 opt_rx->snd_wscale = snd_wscale;
4177 }
4178 break;
4179 case TCPOPT_TIMESTAMP:
4180 if ((opsize == TCPOLEN_TIMESTAMP) &&
4181 ((estab && opt_rx->tstamp_ok) ||
4182 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) {
4183 opt_rx->saw_tstamp = 1;
4184 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4185 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4186 }
4187 break;
4188 case TCPOPT_SACK_PERM:
4189 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4190 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) {
4191 opt_rx->sack_ok = TCP_SACK_SEEN;
4192 tcp_sack_reset(opt_rx);
4193 }
4194 break;
4195
4196 case TCPOPT_SACK:
4197 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4198 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4199 opt_rx->sack_ok) {
4200 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4201 }
4202 break;
4203 #ifdef CONFIG_TCP_MD5SIG
4204 case TCPOPT_MD5SIG:
4205 /* The MD5 Hash has already been
4206 * checked (see tcp_v{4,6}_rcv()).
4207 */
4208 break;
4209 #endif
4210 case TCPOPT_FASTOPEN:
4211 tcp_parse_fastopen_option(
4212 opsize - TCPOLEN_FASTOPEN_BASE,
4213 ptr, th->syn, foc, false);
4214 break;
4215
4216 case TCPOPT_EXP:
4217 /* Fast Open option shares code 254 using a
4218 * 16 bits magic number.
4219 */
4220 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4221 get_unaligned_be16(ptr) ==
4222 TCPOPT_FASTOPEN_MAGIC) {
4223 tcp_parse_fastopen_option(opsize -
4224 TCPOLEN_EXP_FASTOPEN_BASE,
4225 ptr + 2, th->syn, foc, true);
4226 break;
4227 }
4228
4229 if (smc_parse_options(th, opt_rx, ptr, opsize))
4230 break;
4231
4232 opt_rx->saw_unknown = 1;
4233 break;
4234
4235 default:
4236 opt_rx->saw_unknown = 1;
4237 }
4238 ptr += opsize-2;
4239 length -= opsize;
4240 }
4241 }
4242 }
4243 EXPORT_SYMBOL(tcp_parse_options);
4244
tcp_parse_aligned_timestamp(struct tcp_sock * tp,const struct tcphdr * th)4245 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4246 {
4247 const __be32 *ptr = (const __be32 *)(th + 1);
4248
4249 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4250 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4251 tp->rx_opt.saw_tstamp = 1;
4252 ++ptr;
4253 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4254 ++ptr;
4255 if (*ptr)
4256 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4257 else
4258 tp->rx_opt.rcv_tsecr = 0;
4259 return true;
4260 }
4261 return false;
4262 }
4263
4264 /* Fast parse options. This hopes to only see timestamps.
4265 * If it is wrong it falls back on tcp_parse_options().
4266 */
tcp_fast_parse_options(const struct net * net,const struct sk_buff * skb,const struct tcphdr * th,struct tcp_sock * tp)4267 static bool tcp_fast_parse_options(const struct net *net,
4268 const struct sk_buff *skb,
4269 const struct tcphdr *th, struct tcp_sock *tp)
4270 {
4271 /* In the spirit of fast parsing, compare doff directly to constant
4272 * values. Because equality is used, short doff can be ignored here.
4273 */
4274 if (th->doff == (sizeof(*th) / 4)) {
4275 tp->rx_opt.saw_tstamp = 0;
4276 return false;
4277 } else if (tp->rx_opt.tstamp_ok &&
4278 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4279 if (tcp_parse_aligned_timestamp(tp, th))
4280 return true;
4281 }
4282
4283 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4284 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4285 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4286
4287 return true;
4288 }
4289
4290 #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
4291 /*
4292 * Parse Signature options
4293 */
tcp_do_parse_auth_options(const struct tcphdr * th,const u8 ** md5_hash,const u8 ** ao_hash)4294 int tcp_do_parse_auth_options(const struct tcphdr *th,
4295 const u8 **md5_hash, const u8 **ao_hash)
4296 {
4297 int length = (th->doff << 2) - sizeof(*th);
4298 const u8 *ptr = (const u8 *)(th + 1);
4299 unsigned int minlen = TCPOLEN_MD5SIG;
4300
4301 if (IS_ENABLED(CONFIG_TCP_AO))
4302 minlen = sizeof(struct tcp_ao_hdr) + 1;
4303
4304 *md5_hash = NULL;
4305 *ao_hash = NULL;
4306
4307 /* If not enough data remaining, we can short cut */
4308 while (length >= minlen) {
4309 int opcode = *ptr++;
4310 int opsize;
4311
4312 switch (opcode) {
4313 case TCPOPT_EOL:
4314 return 0;
4315 case TCPOPT_NOP:
4316 length--;
4317 continue;
4318 default:
4319 opsize = *ptr++;
4320 if (opsize < 2 || opsize > length)
4321 return -EINVAL;
4322 if (opcode == TCPOPT_MD5SIG) {
4323 if (opsize != TCPOLEN_MD5SIG)
4324 return -EINVAL;
4325 if (unlikely(*md5_hash || *ao_hash))
4326 return -EEXIST;
4327 *md5_hash = ptr;
4328 } else if (opcode == TCPOPT_AO) {
4329 if (opsize <= sizeof(struct tcp_ao_hdr))
4330 return -EINVAL;
4331 if (unlikely(*md5_hash || *ao_hash))
4332 return -EEXIST;
4333 *ao_hash = ptr;
4334 }
4335 }
4336 ptr += opsize - 2;
4337 length -= opsize;
4338 }
4339 return 0;
4340 }
4341 EXPORT_SYMBOL(tcp_do_parse_auth_options);
4342 #endif
4343
4344 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4345 *
4346 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4347 * it can pass through stack. So, the following predicate verifies that
4348 * this segment is not used for anything but congestion avoidance or
4349 * fast retransmit. Moreover, we even are able to eliminate most of such
4350 * second order effects, if we apply some small "replay" window (~RTO)
4351 * to timestamp space.
4352 *
4353 * All these measures still do not guarantee that we reject wrapped ACKs
4354 * on networks with high bandwidth, when sequence space is recycled fastly,
4355 * but it guarantees that such events will be very rare and do not affect
4356 * connection seriously. This doesn't look nice, but alas, PAWS is really
4357 * buggy extension.
4358 *
4359 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4360 * states that events when retransmit arrives after original data are rare.
4361 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4362 * the biggest problem on large power networks even with minor reordering.
4363 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4364 * up to bandwidth of 18Gigabit/sec. 8) ]
4365 */
4366
4367 /* Estimates max number of increments of remote peer TSval in
4368 * a replay window (based on our current RTO estimation).
4369 */
tcp_tsval_replay(const struct sock * sk)4370 static u32 tcp_tsval_replay(const struct sock *sk)
4371 {
4372 /* If we use usec TS resolution,
4373 * then expect the remote peer to use the same resolution.
4374 */
4375 if (tcp_sk(sk)->tcp_usec_ts)
4376 return inet_csk(sk)->icsk_rto * (USEC_PER_SEC / HZ);
4377
4378 /* RFC 7323 recommends a TSval clock between 1ms and 1sec.
4379 * We know that some OS (including old linux) can use 1200 Hz.
4380 */
4381 return inet_csk(sk)->icsk_rto * 1200 / HZ;
4382 }
4383
tcp_disordered_ack(const struct sock * sk,const struct sk_buff * skb)4384 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4385 {
4386 const struct tcp_sock *tp = tcp_sk(sk);
4387 const struct tcphdr *th = tcp_hdr(skb);
4388 u32 seq = TCP_SKB_CB(skb)->seq;
4389 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4390
4391 return /* 1. Pure ACK with correct sequence number. */
4392 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4393
4394 /* 2. ... and duplicate ACK. */
4395 ack == tp->snd_una &&
4396
4397 /* 3. ... and does not update window. */
4398 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4399
4400 /* 4. ... and sits in replay window. */
4401 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <=
4402 tcp_tsval_replay(sk);
4403 }
4404
tcp_paws_discard(const struct sock * sk,const struct sk_buff * skb)4405 static inline bool tcp_paws_discard(const struct sock *sk,
4406 const struct sk_buff *skb)
4407 {
4408 const struct tcp_sock *tp = tcp_sk(sk);
4409
4410 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4411 !tcp_disordered_ack(sk, skb);
4412 }
4413
4414 /* Check segment sequence number for validity.
4415 *
4416 * Segment controls are considered valid, if the segment
4417 * fits to the window after truncation to the window. Acceptability
4418 * of data (and SYN, FIN, of course) is checked separately.
4419 * See tcp_data_queue(), for example.
4420 *
4421 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4422 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4423 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4424 * (borrowed from freebsd)
4425 */
4426
tcp_sequence(const struct tcp_sock * tp,u32 seq,u32 end_seq)4427 static enum skb_drop_reason tcp_sequence(const struct tcp_sock *tp,
4428 u32 seq, u32 end_seq)
4429 {
4430 if (before(end_seq, tp->rcv_wup))
4431 return SKB_DROP_REASON_TCP_OLD_SEQUENCE;
4432
4433 if (after(seq, tp->rcv_nxt + tcp_receive_window(tp)))
4434 return SKB_DROP_REASON_TCP_INVALID_SEQUENCE;
4435
4436 return SKB_NOT_DROPPED_YET;
4437 }
4438
4439 /* When we get a reset we do this. */
tcp_reset(struct sock * sk,struct sk_buff * skb)4440 void tcp_reset(struct sock *sk, struct sk_buff *skb)
4441 {
4442 trace_tcp_receive_reset(sk);
4443
4444 /* mptcp can't tell us to ignore reset pkts,
4445 * so just ignore the return value of mptcp_incoming_options().
4446 */
4447 if (sk_is_mptcp(sk))
4448 mptcp_incoming_options(sk, skb);
4449
4450 /* We want the right error as BSD sees it (and indeed as we do). */
4451 switch (sk->sk_state) {
4452 case TCP_SYN_SENT:
4453 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
4454 break;
4455 case TCP_CLOSE_WAIT:
4456 WRITE_ONCE(sk->sk_err, EPIPE);
4457 break;
4458 case TCP_CLOSE:
4459 return;
4460 default:
4461 WRITE_ONCE(sk->sk_err, ECONNRESET);
4462 }
4463 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4464 smp_wmb();
4465
4466 tcp_write_queue_purge(sk);
4467 tcp_done(sk);
4468
4469 if (!sock_flag(sk, SOCK_DEAD))
4470 sk_error_report(sk);
4471 }
4472
4473 /*
4474 * Process the FIN bit. This now behaves as it is supposed to work
4475 * and the FIN takes effect when it is validly part of sequence
4476 * space. Not before when we get holes.
4477 *
4478 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4479 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4480 * TIME-WAIT)
4481 *
4482 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4483 * close and we go into CLOSING (and later onto TIME-WAIT)
4484 *
4485 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4486 */
tcp_fin(struct sock * sk)4487 void tcp_fin(struct sock *sk)
4488 {
4489 struct tcp_sock *tp = tcp_sk(sk);
4490
4491 inet_csk_schedule_ack(sk);
4492
4493 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
4494 sock_set_flag(sk, SOCK_DONE);
4495
4496 switch (sk->sk_state) {
4497 case TCP_SYN_RECV:
4498 case TCP_ESTABLISHED:
4499 /* Move to CLOSE_WAIT */
4500 tcp_set_state(sk, TCP_CLOSE_WAIT);
4501 inet_csk_enter_pingpong_mode(sk);
4502 break;
4503
4504 case TCP_CLOSE_WAIT:
4505 case TCP_CLOSING:
4506 /* Received a retransmission of the FIN, do
4507 * nothing.
4508 */
4509 break;
4510 case TCP_LAST_ACK:
4511 /* RFC793: Remain in the LAST-ACK state. */
4512 break;
4513
4514 case TCP_FIN_WAIT1:
4515 /* This case occurs when a simultaneous close
4516 * happens, we must ack the received FIN and
4517 * enter the CLOSING state.
4518 */
4519 tcp_send_ack(sk);
4520 tcp_set_state(sk, TCP_CLOSING);
4521 break;
4522 case TCP_FIN_WAIT2:
4523 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4524 tcp_send_ack(sk);
4525 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4526 break;
4527 default:
4528 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4529 * cases we should never reach this piece of code.
4530 */
4531 pr_err("%s: Impossible, sk->sk_state=%d\n",
4532 __func__, sk->sk_state);
4533 break;
4534 }
4535
4536 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4537 * Probably, we should reset in this case. For now drop them.
4538 */
4539 skb_rbtree_purge(&tp->out_of_order_queue);
4540 if (tcp_is_sack(tp))
4541 tcp_sack_reset(&tp->rx_opt);
4542
4543 if (!sock_flag(sk, SOCK_DEAD)) {
4544 sk->sk_state_change(sk);
4545
4546 /* Do not send POLL_HUP for half duplex close. */
4547 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4548 sk->sk_state == TCP_CLOSE)
4549 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4550 else
4551 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4552 }
4553 }
4554
tcp_sack_extend(struct tcp_sack_block * sp,u32 seq,u32 end_seq)4555 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4556 u32 end_seq)
4557 {
4558 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4559 if (before(seq, sp->start_seq))
4560 sp->start_seq = seq;
4561 if (after(end_seq, sp->end_seq))
4562 sp->end_seq = end_seq;
4563 return true;
4564 }
4565 return false;
4566 }
4567
tcp_dsack_set(struct sock * sk,u32 seq,u32 end_seq)4568 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4569 {
4570 struct tcp_sock *tp = tcp_sk(sk);
4571
4572 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4573 int mib_idx;
4574
4575 if (before(seq, tp->rcv_nxt))
4576 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4577 else
4578 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4579
4580 NET_INC_STATS(sock_net(sk), mib_idx);
4581
4582 tp->rx_opt.dsack = 1;
4583 tp->duplicate_sack[0].start_seq = seq;
4584 tp->duplicate_sack[0].end_seq = end_seq;
4585 }
4586 }
4587
tcp_dsack_extend(struct sock * sk,u32 seq,u32 end_seq)4588 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4589 {
4590 struct tcp_sock *tp = tcp_sk(sk);
4591
4592 if (!tp->rx_opt.dsack)
4593 tcp_dsack_set(sk, seq, end_seq);
4594 else
4595 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4596 }
4597
tcp_rcv_spurious_retrans(struct sock * sk,const struct sk_buff * skb)4598 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4599 {
4600 /* When the ACK path fails or drops most ACKs, the sender would
4601 * timeout and spuriously retransmit the same segment repeatedly.
4602 * If it seems our ACKs are not reaching the other side,
4603 * based on receiving a duplicate data segment with new flowlabel
4604 * (suggesting the sender suffered an RTO), and we are not already
4605 * repathing due to our own RTO, then rehash the socket to repath our
4606 * packets.
4607 */
4608 #if IS_ENABLED(CONFIG_IPV6)
4609 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss &&
4610 skb->protocol == htons(ETH_P_IPV6) &&
4611 (tcp_sk(sk)->inet_conn.icsk_ack.lrcv_flowlabel !=
4612 ntohl(ip6_flowlabel(ipv6_hdr(skb)))) &&
4613 sk_rethink_txhash(sk))
4614 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4615
4616 /* Save last flowlabel after a spurious retrans. */
4617 tcp_save_lrcv_flowlabel(sk, skb);
4618 #endif
4619 }
4620
tcp_send_dupack(struct sock * sk,const struct sk_buff * skb)4621 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4622 {
4623 struct tcp_sock *tp = tcp_sk(sk);
4624
4625 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4626 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4627 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4628 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4629
4630 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4631 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4632
4633 tcp_rcv_spurious_retrans(sk, skb);
4634 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4635 end_seq = tp->rcv_nxt;
4636 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4637 }
4638 }
4639
4640 tcp_send_ack(sk);
4641 }
4642
4643 /* These routines update the SACK block as out-of-order packets arrive or
4644 * in-order packets close up the sequence space.
4645 */
tcp_sack_maybe_coalesce(struct tcp_sock * tp)4646 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4647 {
4648 int this_sack;
4649 struct tcp_sack_block *sp = &tp->selective_acks[0];
4650 struct tcp_sack_block *swalk = sp + 1;
4651
4652 /* See if the recent change to the first SACK eats into
4653 * or hits the sequence space of other SACK blocks, if so coalesce.
4654 */
4655 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4656 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4657 int i;
4658
4659 /* Zap SWALK, by moving every further SACK up by one slot.
4660 * Decrease num_sacks.
4661 */
4662 tp->rx_opt.num_sacks--;
4663 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4664 sp[i] = sp[i + 1];
4665 continue;
4666 }
4667 this_sack++;
4668 swalk++;
4669 }
4670 }
4671
tcp_sack_compress_send_ack(struct sock * sk)4672 void tcp_sack_compress_send_ack(struct sock *sk)
4673 {
4674 struct tcp_sock *tp = tcp_sk(sk);
4675
4676 if (!tp->compressed_ack)
4677 return;
4678
4679 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4680 __sock_put(sk);
4681
4682 /* Since we have to send one ack finally,
4683 * substract one from tp->compressed_ack to keep
4684 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4685 */
4686 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4687 tp->compressed_ack - 1);
4688
4689 tp->compressed_ack = 0;
4690 tcp_send_ack(sk);
4691 }
4692
4693 /* Reasonable amount of sack blocks included in TCP SACK option
4694 * The max is 4, but this becomes 3 if TCP timestamps are there.
4695 * Given that SACK packets might be lost, be conservative and use 2.
4696 */
4697 #define TCP_SACK_BLOCKS_EXPECTED 2
4698
tcp_sack_new_ofo_skb(struct sock * sk,u32 seq,u32 end_seq)4699 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4700 {
4701 struct tcp_sock *tp = tcp_sk(sk);
4702 struct tcp_sack_block *sp = &tp->selective_acks[0];
4703 int cur_sacks = tp->rx_opt.num_sacks;
4704 int this_sack;
4705
4706 if (!cur_sacks)
4707 goto new_sack;
4708
4709 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4710 if (tcp_sack_extend(sp, seq, end_seq)) {
4711 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4712 tcp_sack_compress_send_ack(sk);
4713 /* Rotate this_sack to the first one. */
4714 for (; this_sack > 0; this_sack--, sp--)
4715 swap(*sp, *(sp - 1));
4716 if (cur_sacks > 1)
4717 tcp_sack_maybe_coalesce(tp);
4718 return;
4719 }
4720 }
4721
4722 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4723 tcp_sack_compress_send_ack(sk);
4724
4725 /* Could not find an adjacent existing SACK, build a new one,
4726 * put it at the front, and shift everyone else down. We
4727 * always know there is at least one SACK present already here.
4728 *
4729 * If the sack array is full, forget about the last one.
4730 */
4731 if (this_sack >= TCP_NUM_SACKS) {
4732 this_sack--;
4733 tp->rx_opt.num_sacks--;
4734 sp--;
4735 }
4736 for (; this_sack > 0; this_sack--, sp--)
4737 *sp = *(sp - 1);
4738
4739 new_sack:
4740 /* Build the new head SACK, and we're done. */
4741 sp->start_seq = seq;
4742 sp->end_seq = end_seq;
4743 tp->rx_opt.num_sacks++;
4744 }
4745
4746 /* RCV.NXT advances, some SACKs should be eaten. */
4747
tcp_sack_remove(struct tcp_sock * tp)4748 static void tcp_sack_remove(struct tcp_sock *tp)
4749 {
4750 struct tcp_sack_block *sp = &tp->selective_acks[0];
4751 int num_sacks = tp->rx_opt.num_sacks;
4752 int this_sack;
4753
4754 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4755 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4756 tp->rx_opt.num_sacks = 0;
4757 return;
4758 }
4759
4760 for (this_sack = 0; this_sack < num_sacks;) {
4761 /* Check if the start of the sack is covered by RCV.NXT. */
4762 if (!before(tp->rcv_nxt, sp->start_seq)) {
4763 int i;
4764
4765 /* RCV.NXT must cover all the block! */
4766 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4767
4768 /* Zap this SACK, by moving forward any other SACKS. */
4769 for (i = this_sack+1; i < num_sacks; i++)
4770 tp->selective_acks[i-1] = tp->selective_acks[i];
4771 num_sacks--;
4772 continue;
4773 }
4774 this_sack++;
4775 sp++;
4776 }
4777 tp->rx_opt.num_sacks = num_sacks;
4778 }
4779
4780 /**
4781 * tcp_try_coalesce - try to merge skb to prior one
4782 * @sk: socket
4783 * @to: prior buffer
4784 * @from: buffer to add in queue
4785 * @fragstolen: pointer to boolean
4786 *
4787 * Before queueing skb @from after @to, try to merge them
4788 * to reduce overall memory use and queue lengths, if cost is small.
4789 * Packets in ofo or receive queues can stay a long time.
4790 * Better try to coalesce them right now to avoid future collapses.
4791 * Returns true if caller should free @from instead of queueing it
4792 */
tcp_try_coalesce(struct sock * sk,struct sk_buff * to,struct sk_buff * from,bool * fragstolen)4793 static bool tcp_try_coalesce(struct sock *sk,
4794 struct sk_buff *to,
4795 struct sk_buff *from,
4796 bool *fragstolen)
4797 {
4798 int delta;
4799
4800 *fragstolen = false;
4801
4802 /* Its possible this segment overlaps with prior segment in queue */
4803 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4804 return false;
4805
4806 if (!mptcp_skb_can_collapse(to, from))
4807 return false;
4808
4809 if (skb_cmp_decrypted(from, to))
4810 return false;
4811
4812 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4813 return false;
4814
4815 atomic_add(delta, &sk->sk_rmem_alloc);
4816 sk_mem_charge(sk, delta);
4817 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4818 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4819 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4820 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4821
4822 if (TCP_SKB_CB(from)->has_rxtstamp) {
4823 TCP_SKB_CB(to)->has_rxtstamp = true;
4824 to->tstamp = from->tstamp;
4825 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4826 }
4827
4828 return true;
4829 }
4830
tcp_ooo_try_coalesce(struct sock * sk,struct sk_buff * to,struct sk_buff * from,bool * fragstolen)4831 static bool tcp_ooo_try_coalesce(struct sock *sk,
4832 struct sk_buff *to,
4833 struct sk_buff *from,
4834 bool *fragstolen)
4835 {
4836 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4837
4838 /* In case tcp_drop_reason() is called later, update to->gso_segs */
4839 if (res) {
4840 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4841 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4842
4843 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4844 }
4845 return res;
4846 }
4847
tcp_drop_reason(struct sock * sk,struct sk_buff * skb,enum skb_drop_reason reason)4848 static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb,
4849 enum skb_drop_reason reason)
4850 {
4851 sk_drops_add(sk, skb);
4852 kfree_skb_reason(skb, reason);
4853 }
4854
4855 /* This one checks to see if we can put data from the
4856 * out_of_order queue into the receive_queue.
4857 */
tcp_ofo_queue(struct sock * sk)4858 static void tcp_ofo_queue(struct sock *sk)
4859 {
4860 struct tcp_sock *tp = tcp_sk(sk);
4861 __u32 dsack_high = tp->rcv_nxt;
4862 bool fin, fragstolen, eaten;
4863 struct sk_buff *skb, *tail;
4864 struct rb_node *p;
4865
4866 p = rb_first(&tp->out_of_order_queue);
4867 while (p) {
4868 skb = rb_to_skb(p);
4869 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4870 break;
4871
4872 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4873 __u32 dsack = dsack_high;
4874 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4875 dsack_high = TCP_SKB_CB(skb)->end_seq;
4876 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4877 }
4878 p = rb_next(p);
4879 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4880
4881 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4882 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP);
4883 continue;
4884 }
4885
4886 tail = skb_peek_tail(&sk->sk_receive_queue);
4887 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4888 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4889 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4890 if (!eaten)
4891 __skb_queue_tail(&sk->sk_receive_queue, skb);
4892 else
4893 kfree_skb_partial(skb, fragstolen);
4894
4895 if (unlikely(fin)) {
4896 tcp_fin(sk);
4897 /* tcp_fin() purges tp->out_of_order_queue,
4898 * so we must end this loop right now.
4899 */
4900 break;
4901 }
4902 }
4903 }
4904
4905 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb);
4906 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb);
4907
tcp_try_rmem_schedule(struct sock * sk,struct sk_buff * skb,unsigned int size)4908 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4909 unsigned int size)
4910 {
4911 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4912 !sk_rmem_schedule(sk, skb, size)) {
4913
4914 if (tcp_prune_queue(sk, skb) < 0)
4915 return -1;
4916
4917 while (!sk_rmem_schedule(sk, skb, size)) {
4918 if (!tcp_prune_ofo_queue(sk, skb))
4919 return -1;
4920 }
4921 }
4922 return 0;
4923 }
4924
tcp_data_queue_ofo(struct sock * sk,struct sk_buff * skb)4925 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4926 {
4927 struct tcp_sock *tp = tcp_sk(sk);
4928 struct rb_node **p, *parent;
4929 struct sk_buff *skb1;
4930 u32 seq, end_seq;
4931 bool fragstolen;
4932
4933 tcp_save_lrcv_flowlabel(sk, skb);
4934 tcp_ecn_check_ce(sk, skb);
4935
4936 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4937 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4938 sk->sk_data_ready(sk);
4939 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM);
4940 return;
4941 }
4942
4943 /* Disable header prediction. */
4944 tp->pred_flags = 0;
4945 inet_csk_schedule_ack(sk);
4946
4947 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4948 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4949 seq = TCP_SKB_CB(skb)->seq;
4950 end_seq = TCP_SKB_CB(skb)->end_seq;
4951
4952 p = &tp->out_of_order_queue.rb_node;
4953 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4954 /* Initial out of order segment, build 1 SACK. */
4955 if (tcp_is_sack(tp)) {
4956 tp->rx_opt.num_sacks = 1;
4957 tp->selective_acks[0].start_seq = seq;
4958 tp->selective_acks[0].end_seq = end_seq;
4959 }
4960 rb_link_node(&skb->rbnode, NULL, p);
4961 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4962 tp->ooo_last_skb = skb;
4963 goto end;
4964 }
4965
4966 /* In the typical case, we are adding an skb to the end of the list.
4967 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4968 */
4969 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4970 skb, &fragstolen)) {
4971 coalesce_done:
4972 /* For non sack flows, do not grow window to force DUPACK
4973 * and trigger fast retransmit.
4974 */
4975 if (tcp_is_sack(tp))
4976 tcp_grow_window(sk, skb, true);
4977 kfree_skb_partial(skb, fragstolen);
4978 skb = NULL;
4979 goto add_sack;
4980 }
4981 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4982 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4983 parent = &tp->ooo_last_skb->rbnode;
4984 p = &parent->rb_right;
4985 goto insert;
4986 }
4987
4988 /* Find place to insert this segment. Handle overlaps on the way. */
4989 parent = NULL;
4990 while (*p) {
4991 parent = *p;
4992 skb1 = rb_to_skb(parent);
4993 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4994 p = &parent->rb_left;
4995 continue;
4996 }
4997 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4998 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4999 /* All the bits are present. Drop. */
5000 NET_INC_STATS(sock_net(sk),
5001 LINUX_MIB_TCPOFOMERGE);
5002 tcp_drop_reason(sk, skb,
5003 SKB_DROP_REASON_TCP_OFOMERGE);
5004 skb = NULL;
5005 tcp_dsack_set(sk, seq, end_seq);
5006 goto add_sack;
5007 }
5008 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
5009 /* Partial overlap. */
5010 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
5011 } else {
5012 /* skb's seq == skb1's seq and skb covers skb1.
5013 * Replace skb1 with skb.
5014 */
5015 rb_replace_node(&skb1->rbnode, &skb->rbnode,
5016 &tp->out_of_order_queue);
5017 tcp_dsack_extend(sk,
5018 TCP_SKB_CB(skb1)->seq,
5019 TCP_SKB_CB(skb1)->end_seq);
5020 NET_INC_STATS(sock_net(sk),
5021 LINUX_MIB_TCPOFOMERGE);
5022 tcp_drop_reason(sk, skb1,
5023 SKB_DROP_REASON_TCP_OFOMERGE);
5024 goto merge_right;
5025 }
5026 } else if (tcp_ooo_try_coalesce(sk, skb1,
5027 skb, &fragstolen)) {
5028 goto coalesce_done;
5029 }
5030 p = &parent->rb_right;
5031 }
5032 insert:
5033 /* Insert segment into RB tree. */
5034 rb_link_node(&skb->rbnode, parent, p);
5035 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
5036
5037 merge_right:
5038 /* Remove other segments covered by skb. */
5039 while ((skb1 = skb_rb_next(skb)) != NULL) {
5040 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
5041 break;
5042 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
5043 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
5044 end_seq);
5045 break;
5046 }
5047 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
5048 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
5049 TCP_SKB_CB(skb1)->end_seq);
5050 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
5051 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE);
5052 }
5053 /* If there is no skb after us, we are the last_skb ! */
5054 if (!skb1)
5055 tp->ooo_last_skb = skb;
5056
5057 add_sack:
5058 if (tcp_is_sack(tp))
5059 tcp_sack_new_ofo_skb(sk, seq, end_seq);
5060 end:
5061 if (skb) {
5062 /* For non sack flows, do not grow window to force DUPACK
5063 * and trigger fast retransmit.
5064 */
5065 if (tcp_is_sack(tp))
5066 tcp_grow_window(sk, skb, false);
5067 skb_condense(skb);
5068 skb_set_owner_r(skb, sk);
5069 }
5070 }
5071
tcp_queue_rcv(struct sock * sk,struct sk_buff * skb,bool * fragstolen)5072 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
5073 bool *fragstolen)
5074 {
5075 int eaten;
5076 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
5077
5078 eaten = (tail &&
5079 tcp_try_coalesce(sk, tail,
5080 skb, fragstolen)) ? 1 : 0;
5081 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
5082 if (!eaten) {
5083 __skb_queue_tail(&sk->sk_receive_queue, skb);
5084 skb_set_owner_r(skb, sk);
5085 }
5086 return eaten;
5087 }
5088
tcp_send_rcvq(struct sock * sk,struct msghdr * msg,size_t size)5089 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
5090 {
5091 struct sk_buff *skb;
5092 int err = -ENOMEM;
5093 int data_len = 0;
5094 bool fragstolen;
5095
5096 if (size == 0)
5097 return 0;
5098
5099 if (size > PAGE_SIZE) {
5100 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
5101
5102 data_len = npages << PAGE_SHIFT;
5103 size = data_len + (size & ~PAGE_MASK);
5104 }
5105 skb = alloc_skb_with_frags(size - data_len, data_len,
5106 PAGE_ALLOC_COSTLY_ORDER,
5107 &err, sk->sk_allocation);
5108 if (!skb)
5109 goto err;
5110
5111 skb_put(skb, size - data_len);
5112 skb->data_len = data_len;
5113 skb->len = size;
5114
5115 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5116 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5117 goto err_free;
5118 }
5119
5120 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
5121 if (err)
5122 goto err_free;
5123
5124 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
5125 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
5126 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
5127
5128 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
5129 WARN_ON_ONCE(fragstolen); /* should not happen */
5130 __kfree_skb(skb);
5131 }
5132 return size;
5133
5134 err_free:
5135 kfree_skb(skb);
5136 err:
5137 return err;
5138
5139 }
5140
tcp_data_ready(struct sock * sk)5141 void tcp_data_ready(struct sock *sk)
5142 {
5143 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE))
5144 sk->sk_data_ready(sk);
5145 }
5146
tcp_data_queue(struct sock * sk,struct sk_buff * skb)5147 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
5148 {
5149 struct tcp_sock *tp = tcp_sk(sk);
5150 enum skb_drop_reason reason;
5151 bool fragstolen;
5152 int eaten;
5153
5154 /* If a subflow has been reset, the packet should not continue
5155 * to be processed, drop the packet.
5156 */
5157 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) {
5158 __kfree_skb(skb);
5159 return;
5160 }
5161
5162 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
5163 __kfree_skb(skb);
5164 return;
5165 }
5166 skb_dst_drop(skb);
5167 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
5168
5169 reason = SKB_DROP_REASON_NOT_SPECIFIED;
5170 tp->rx_opt.dsack = 0;
5171
5172 /* Queue data for delivery to the user.
5173 * Packets in sequence go to the receive queue.
5174 * Out of sequence packets to the out_of_order_queue.
5175 */
5176 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5177 if (tcp_receive_window(tp) == 0) {
5178 /* Some stacks are known to send bare FIN packets
5179 * in a loop even if we send RWIN 0 in our ACK.
5180 * Accepting this FIN does not hurt memory pressure
5181 * because the FIN flag will simply be merged to the
5182 * receive queue tail skb in most cases.
5183 */
5184 if (!skb->len &&
5185 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
5186 goto queue_and_out;
5187
5188 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5189 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5190 goto out_of_window;
5191 }
5192
5193 /* Ok. In sequence. In window. */
5194 queue_and_out:
5195 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5196 /* TODO: maybe ratelimit these WIN 0 ACK ? */
5197 inet_csk(sk)->icsk_ack.pending |=
5198 (ICSK_ACK_NOMEM | ICSK_ACK_NOW);
5199 inet_csk_schedule_ack(sk);
5200 sk->sk_data_ready(sk);
5201
5202 if (skb_queue_len(&sk->sk_receive_queue) && skb->len) {
5203 reason = SKB_DROP_REASON_PROTO_MEM;
5204 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5205 goto drop;
5206 }
5207 sk_forced_mem_schedule(sk, skb->truesize);
5208 }
5209
5210 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5211 if (skb->len)
5212 tcp_event_data_recv(sk, skb);
5213 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
5214 tcp_fin(sk);
5215
5216 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5217 tcp_ofo_queue(sk);
5218
5219 /* RFC5681. 4.2. SHOULD send immediate ACK, when
5220 * gap in queue is filled.
5221 */
5222 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5223 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
5224 }
5225
5226 if (tp->rx_opt.num_sacks)
5227 tcp_sack_remove(tp);
5228
5229 tcp_fast_path_check(sk);
5230
5231 if (eaten > 0)
5232 kfree_skb_partial(skb, fragstolen);
5233 if (!sock_flag(sk, SOCK_DEAD))
5234 tcp_data_ready(sk);
5235 return;
5236 }
5237
5238 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
5239 tcp_rcv_spurious_retrans(sk, skb);
5240 /* A retransmit, 2nd most common case. Force an immediate ack. */
5241 reason = SKB_DROP_REASON_TCP_OLD_DATA;
5242 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5243 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5244
5245 out_of_window:
5246 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5247 inet_csk_schedule_ack(sk);
5248 drop:
5249 tcp_drop_reason(sk, skb, reason);
5250 return;
5251 }
5252
5253 /* Out of window. F.e. zero window probe. */
5254 if (!before(TCP_SKB_CB(skb)->seq,
5255 tp->rcv_nxt + tcp_receive_window(tp))) {
5256 reason = SKB_DROP_REASON_TCP_OVERWINDOW;
5257 goto out_of_window;
5258 }
5259
5260 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5261 /* Partial packet, seq < rcv_next < end_seq */
5262 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5263
5264 /* If window is closed, drop tail of packet. But after
5265 * remembering D-SACK for its head made in previous line.
5266 */
5267 if (!tcp_receive_window(tp)) {
5268 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5269 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5270 goto out_of_window;
5271 }
5272 goto queue_and_out;
5273 }
5274
5275 tcp_data_queue_ofo(sk, skb);
5276 }
5277
tcp_skb_next(struct sk_buff * skb,struct sk_buff_head * list)5278 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5279 {
5280 if (list)
5281 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5282
5283 return skb_rb_next(skb);
5284 }
5285
tcp_collapse_one(struct sock * sk,struct sk_buff * skb,struct sk_buff_head * list,struct rb_root * root)5286 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5287 struct sk_buff_head *list,
5288 struct rb_root *root)
5289 {
5290 struct sk_buff *next = tcp_skb_next(skb, list);
5291
5292 if (list)
5293 __skb_unlink(skb, list);
5294 else
5295 rb_erase(&skb->rbnode, root);
5296
5297 __kfree_skb(skb);
5298 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5299
5300 return next;
5301 }
5302
5303 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
tcp_rbtree_insert(struct rb_root * root,struct sk_buff * skb)5304 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5305 {
5306 struct rb_node **p = &root->rb_node;
5307 struct rb_node *parent = NULL;
5308 struct sk_buff *skb1;
5309
5310 while (*p) {
5311 parent = *p;
5312 skb1 = rb_to_skb(parent);
5313 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5314 p = &parent->rb_left;
5315 else
5316 p = &parent->rb_right;
5317 }
5318 rb_link_node(&skb->rbnode, parent, p);
5319 rb_insert_color(&skb->rbnode, root);
5320 }
5321
5322 /* Collapse contiguous sequence of skbs head..tail with
5323 * sequence numbers start..end.
5324 *
5325 * If tail is NULL, this means until the end of the queue.
5326 *
5327 * Segments with FIN/SYN are not collapsed (only because this
5328 * simplifies code)
5329 */
5330 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)5331 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5332 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5333 {
5334 struct sk_buff *skb = head, *n;
5335 struct sk_buff_head tmp;
5336 bool end_of_skbs;
5337
5338 /* First, check that queue is collapsible and find
5339 * the point where collapsing can be useful.
5340 */
5341 restart:
5342 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5343 n = tcp_skb_next(skb, list);
5344
5345 /* No new bits? It is possible on ofo queue. */
5346 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5347 skb = tcp_collapse_one(sk, skb, list, root);
5348 if (!skb)
5349 break;
5350 goto restart;
5351 }
5352
5353 /* The first skb to collapse is:
5354 * - not SYN/FIN and
5355 * - bloated or contains data before "start" or
5356 * overlaps to the next one and mptcp allow collapsing.
5357 */
5358 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5359 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5360 before(TCP_SKB_CB(skb)->seq, start))) {
5361 end_of_skbs = false;
5362 break;
5363 }
5364
5365 if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
5366 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5367 end_of_skbs = false;
5368 break;
5369 }
5370
5371 /* Decided to skip this, advance start seq. */
5372 start = TCP_SKB_CB(skb)->end_seq;
5373 }
5374 if (end_of_skbs ||
5375 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5376 return;
5377
5378 __skb_queue_head_init(&tmp);
5379
5380 while (before(start, end)) {
5381 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5382 struct sk_buff *nskb;
5383
5384 nskb = alloc_skb(copy, GFP_ATOMIC);
5385 if (!nskb)
5386 break;
5387
5388 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5389 skb_copy_decrypted(nskb, skb);
5390 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5391 if (list)
5392 __skb_queue_before(list, skb, nskb);
5393 else
5394 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5395 skb_set_owner_r(nskb, sk);
5396 mptcp_skb_ext_move(nskb, skb);
5397
5398 /* Copy data, releasing collapsed skbs. */
5399 while (copy > 0) {
5400 int offset = start - TCP_SKB_CB(skb)->seq;
5401 int size = TCP_SKB_CB(skb)->end_seq - start;
5402
5403 BUG_ON(offset < 0);
5404 if (size > 0) {
5405 size = min(copy, size);
5406 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5407 BUG();
5408 TCP_SKB_CB(nskb)->end_seq += size;
5409 copy -= size;
5410 start += size;
5411 }
5412 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5413 skb = tcp_collapse_one(sk, skb, list, root);
5414 if (!skb ||
5415 skb == tail ||
5416 !mptcp_skb_can_collapse(nskb, skb) ||
5417 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5418 goto end;
5419 if (skb_cmp_decrypted(skb, nskb))
5420 goto end;
5421 }
5422 }
5423 }
5424 end:
5425 skb_queue_walk_safe(&tmp, skb, n)
5426 tcp_rbtree_insert(root, skb);
5427 }
5428
5429 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5430 * and tcp_collapse() them until all the queue is collapsed.
5431 */
tcp_collapse_ofo_queue(struct sock * sk)5432 static void tcp_collapse_ofo_queue(struct sock *sk)
5433 {
5434 struct tcp_sock *tp = tcp_sk(sk);
5435 u32 range_truesize, sum_tiny = 0;
5436 struct sk_buff *skb, *head;
5437 u32 start, end;
5438
5439 skb = skb_rb_first(&tp->out_of_order_queue);
5440 new_range:
5441 if (!skb) {
5442 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5443 return;
5444 }
5445 start = TCP_SKB_CB(skb)->seq;
5446 end = TCP_SKB_CB(skb)->end_seq;
5447 range_truesize = skb->truesize;
5448
5449 for (head = skb;;) {
5450 skb = skb_rb_next(skb);
5451
5452 /* Range is terminated when we see a gap or when
5453 * we are at the queue end.
5454 */
5455 if (!skb ||
5456 after(TCP_SKB_CB(skb)->seq, end) ||
5457 before(TCP_SKB_CB(skb)->end_seq, start)) {
5458 /* Do not attempt collapsing tiny skbs */
5459 if (range_truesize != head->truesize ||
5460 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) {
5461 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5462 head, skb, start, end);
5463 } else {
5464 sum_tiny += range_truesize;
5465 if (sum_tiny > sk->sk_rcvbuf >> 3)
5466 return;
5467 }
5468 goto new_range;
5469 }
5470
5471 range_truesize += skb->truesize;
5472 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5473 start = TCP_SKB_CB(skb)->seq;
5474 if (after(TCP_SKB_CB(skb)->end_seq, end))
5475 end = TCP_SKB_CB(skb)->end_seq;
5476 }
5477 }
5478
5479 /*
5480 * Clean the out-of-order queue to make room.
5481 * We drop high sequences packets to :
5482 * 1) Let a chance for holes to be filled.
5483 * This means we do not drop packets from ooo queue if their sequence
5484 * is before incoming packet sequence.
5485 * 2) not add too big latencies if thousands of packets sit there.
5486 * (But if application shrinks SO_RCVBUF, we could still end up
5487 * freeing whole queue here)
5488 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5489 *
5490 * Return true if queue has shrunk.
5491 */
tcp_prune_ofo_queue(struct sock * sk,const struct sk_buff * in_skb)5492 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb)
5493 {
5494 struct tcp_sock *tp = tcp_sk(sk);
5495 struct rb_node *node, *prev;
5496 bool pruned = false;
5497 int goal;
5498
5499 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5500 return false;
5501
5502 goal = sk->sk_rcvbuf >> 3;
5503 node = &tp->ooo_last_skb->rbnode;
5504
5505 do {
5506 struct sk_buff *skb = rb_to_skb(node);
5507
5508 /* If incoming skb would land last in ofo queue, stop pruning. */
5509 if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq))
5510 break;
5511 pruned = true;
5512 prev = rb_prev(node);
5513 rb_erase(node, &tp->out_of_order_queue);
5514 goal -= skb->truesize;
5515 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE);
5516 tp->ooo_last_skb = rb_to_skb(prev);
5517 if (!prev || goal <= 0) {
5518 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5519 !tcp_under_memory_pressure(sk))
5520 break;
5521 goal = sk->sk_rcvbuf >> 3;
5522 }
5523 node = prev;
5524 } while (node);
5525
5526 if (pruned) {
5527 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5528 /* Reset SACK state. A conforming SACK implementation will
5529 * do the same at a timeout based retransmit. When a connection
5530 * is in a sad state like this, we care only about integrity
5531 * of the connection not performance.
5532 */
5533 if (tp->rx_opt.sack_ok)
5534 tcp_sack_reset(&tp->rx_opt);
5535 }
5536 return pruned;
5537 }
5538
5539 /* Reduce allocated memory if we can, trying to get
5540 * the socket within its memory limits again.
5541 *
5542 * Return less than zero if we should start dropping frames
5543 * until the socket owning process reads some of the data
5544 * to stabilize the situation.
5545 */
tcp_prune_queue(struct sock * sk,const struct sk_buff * in_skb)5546 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb)
5547 {
5548 struct tcp_sock *tp = tcp_sk(sk);
5549
5550 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5551
5552 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5553 tcp_clamp_window(sk);
5554 else if (tcp_under_memory_pressure(sk))
5555 tcp_adjust_rcv_ssthresh(sk);
5556
5557 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5558 return 0;
5559
5560 tcp_collapse_ofo_queue(sk);
5561 if (!skb_queue_empty(&sk->sk_receive_queue))
5562 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5563 skb_peek(&sk->sk_receive_queue),
5564 NULL,
5565 tp->copied_seq, tp->rcv_nxt);
5566
5567 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5568 return 0;
5569
5570 /* Collapsing did not help, destructive actions follow.
5571 * This must not ever occur. */
5572
5573 tcp_prune_ofo_queue(sk, in_skb);
5574
5575 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5576 return 0;
5577
5578 /* If we are really being abused, tell the caller to silently
5579 * drop receive data on the floor. It will get retransmitted
5580 * and hopefully then we'll have sufficient space.
5581 */
5582 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5583
5584 /* Massive buffer overcommit. */
5585 tp->pred_flags = 0;
5586 return -1;
5587 }
5588
tcp_should_expand_sndbuf(struct sock * sk)5589 static bool tcp_should_expand_sndbuf(struct sock *sk)
5590 {
5591 const struct tcp_sock *tp = tcp_sk(sk);
5592
5593 /* If the user specified a specific send buffer setting, do
5594 * not modify it.
5595 */
5596 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5597 return false;
5598
5599 /* If we are under global TCP memory pressure, do not expand. */
5600 if (tcp_under_memory_pressure(sk)) {
5601 int unused_mem = sk_unused_reserved_mem(sk);
5602
5603 /* Adjust sndbuf according to reserved mem. But make sure
5604 * it never goes below SOCK_MIN_SNDBUF.
5605 * See sk_stream_moderate_sndbuf() for more details.
5606 */
5607 if (unused_mem > SOCK_MIN_SNDBUF)
5608 WRITE_ONCE(sk->sk_sndbuf, unused_mem);
5609
5610 return false;
5611 }
5612
5613 /* If we are under soft global TCP memory pressure, do not expand. */
5614 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5615 return false;
5616
5617 /* If we filled the congestion window, do not expand. */
5618 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp))
5619 return false;
5620
5621 return true;
5622 }
5623
tcp_new_space(struct sock * sk)5624 static void tcp_new_space(struct sock *sk)
5625 {
5626 struct tcp_sock *tp = tcp_sk(sk);
5627
5628 if (tcp_should_expand_sndbuf(sk)) {
5629 tcp_sndbuf_expand(sk);
5630 tp->snd_cwnd_stamp = tcp_jiffies32;
5631 }
5632
5633 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
5634 }
5635
5636 /* Caller made space either from:
5637 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5638 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5639 *
5640 * We might be able to generate EPOLLOUT to the application if:
5641 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5642 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5643 * small enough that tcp_stream_memory_free() decides it
5644 * is time to generate EPOLLOUT.
5645 */
tcp_check_space(struct sock * sk)5646 void tcp_check_space(struct sock *sk)
5647 {
5648 /* pairs with tcp_poll() */
5649 smp_mb();
5650 if (sk->sk_socket &&
5651 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5652 tcp_new_space(sk);
5653 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5654 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5655 }
5656 }
5657
tcp_data_snd_check(struct sock * sk)5658 static inline void tcp_data_snd_check(struct sock *sk)
5659 {
5660 tcp_push_pending_frames(sk);
5661 tcp_check_space(sk);
5662 }
5663
5664 /*
5665 * Check if sending an ack is needed.
5666 */
__tcp_ack_snd_check(struct sock * sk,int ofo_possible)5667 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5668 {
5669 struct tcp_sock *tp = tcp_sk(sk);
5670 unsigned long rtt, delay;
5671
5672 /* More than one full frame received... */
5673 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5674 /* ... and right edge of window advances far enough.
5675 * (tcp_recvmsg() will send ACK otherwise).
5676 * If application uses SO_RCVLOWAT, we want send ack now if
5677 * we have not received enough bytes to satisfy the condition.
5678 */
5679 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5680 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5681 /* We ACK each frame or... */
5682 tcp_in_quickack_mode(sk) ||
5683 /* Protocol state mandates a one-time immediate ACK */
5684 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5685 /* If we are running from __release_sock() in user context,
5686 * Defer the ack until tcp_release_cb().
5687 */
5688 if (sock_owned_by_user_nocheck(sk) &&
5689 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_backlog_ack_defer)) {
5690 set_bit(TCP_ACK_DEFERRED, &sk->sk_tsq_flags);
5691 return;
5692 }
5693 send_now:
5694 tcp_send_ack(sk);
5695 return;
5696 }
5697
5698 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5699 tcp_send_delayed_ack(sk);
5700 return;
5701 }
5702
5703 if (!tcp_is_sack(tp) ||
5704 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr))
5705 goto send_now;
5706
5707 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5708 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5709 tp->dup_ack_counter = 0;
5710 }
5711 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5712 tp->dup_ack_counter++;
5713 goto send_now;
5714 }
5715 tp->compressed_ack++;
5716 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5717 return;
5718
5719 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5720
5721 rtt = tp->rcv_rtt_est.rtt_us;
5722 if (tp->srtt_us && tp->srtt_us < rtt)
5723 rtt = tp->srtt_us;
5724
5725 delay = min_t(unsigned long,
5726 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5727 rtt * (NSEC_PER_USEC >> 3)/20);
5728 sock_hold(sk);
5729 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5730 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns),
5731 HRTIMER_MODE_REL_PINNED_SOFT);
5732 }
5733
tcp_ack_snd_check(struct sock * sk)5734 static inline void tcp_ack_snd_check(struct sock *sk)
5735 {
5736 if (!inet_csk_ack_scheduled(sk)) {
5737 /* We sent a data segment already. */
5738 return;
5739 }
5740 __tcp_ack_snd_check(sk, 1);
5741 }
5742
5743 /*
5744 * This routine is only called when we have urgent data
5745 * signaled. Its the 'slow' part of tcp_urg. It could be
5746 * moved inline now as tcp_urg is only called from one
5747 * place. We handle URGent data wrong. We have to - as
5748 * BSD still doesn't use the correction from RFC961.
5749 * For 1003.1g we should support a new option TCP_STDURG to permit
5750 * either form (or just set the sysctl tcp_stdurg).
5751 */
5752
tcp_check_urg(struct sock * sk,const struct tcphdr * th)5753 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5754 {
5755 struct tcp_sock *tp = tcp_sk(sk);
5756 u32 ptr = ntohs(th->urg_ptr);
5757
5758 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5759 ptr--;
5760 ptr += ntohl(th->seq);
5761
5762 /* Ignore urgent data that we've already seen and read. */
5763 if (after(tp->copied_seq, ptr))
5764 return;
5765
5766 /* Do not replay urg ptr.
5767 *
5768 * NOTE: interesting situation not covered by specs.
5769 * Misbehaving sender may send urg ptr, pointing to segment,
5770 * which we already have in ofo queue. We are not able to fetch
5771 * such data and will stay in TCP_URG_NOTYET until will be eaten
5772 * by recvmsg(). Seems, we are not obliged to handle such wicked
5773 * situations. But it is worth to think about possibility of some
5774 * DoSes using some hypothetical application level deadlock.
5775 */
5776 if (before(ptr, tp->rcv_nxt))
5777 return;
5778
5779 /* Do we already have a newer (or duplicate) urgent pointer? */
5780 if (tp->urg_data && !after(ptr, tp->urg_seq))
5781 return;
5782
5783 /* Tell the world about our new urgent pointer. */
5784 sk_send_sigurg(sk);
5785
5786 /* We may be adding urgent data when the last byte read was
5787 * urgent. To do this requires some care. We cannot just ignore
5788 * tp->copied_seq since we would read the last urgent byte again
5789 * as data, nor can we alter copied_seq until this data arrives
5790 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5791 *
5792 * NOTE. Double Dutch. Rendering to plain English: author of comment
5793 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5794 * and expect that both A and B disappear from stream. This is _wrong_.
5795 * Though this happens in BSD with high probability, this is occasional.
5796 * Any application relying on this is buggy. Note also, that fix "works"
5797 * only in this artificial test. Insert some normal data between A and B and we will
5798 * decline of BSD again. Verdict: it is better to remove to trap
5799 * buggy users.
5800 */
5801 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5802 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5803 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5804 tp->copied_seq++;
5805 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5806 __skb_unlink(skb, &sk->sk_receive_queue);
5807 __kfree_skb(skb);
5808 }
5809 }
5810
5811 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET);
5812 WRITE_ONCE(tp->urg_seq, ptr);
5813
5814 /* Disable header prediction. */
5815 tp->pred_flags = 0;
5816 }
5817
5818 /* This is the 'fast' part of urgent handling. */
tcp_urg(struct sock * sk,struct sk_buff * skb,const struct tcphdr * th)5819 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5820 {
5821 struct tcp_sock *tp = tcp_sk(sk);
5822
5823 /* Check if we get a new urgent pointer - normally not. */
5824 if (unlikely(th->urg))
5825 tcp_check_urg(sk, th);
5826
5827 /* Do we wait for any urgent data? - normally not... */
5828 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) {
5829 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5830 th->syn;
5831
5832 /* Is the urgent pointer pointing into this packet? */
5833 if (ptr < skb->len) {
5834 u8 tmp;
5835 if (skb_copy_bits(skb, ptr, &tmp, 1))
5836 BUG();
5837 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp);
5838 if (!sock_flag(sk, SOCK_DEAD))
5839 sk->sk_data_ready(sk);
5840 }
5841 }
5842 }
5843
5844 /* Accept RST for rcv_nxt - 1 after a FIN.
5845 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5846 * FIN is sent followed by a RST packet. The RST is sent with the same
5847 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5848 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5849 * ACKs on the closed socket. In addition middleboxes can drop either the
5850 * challenge ACK or a subsequent RST.
5851 */
tcp_reset_check(const struct sock * sk,const struct sk_buff * skb)5852 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5853 {
5854 const struct tcp_sock *tp = tcp_sk(sk);
5855
5856 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5857 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5858 TCPF_CLOSING));
5859 }
5860
5861 /* Does PAWS and seqno based validation of an incoming segment, flags will
5862 * play significant role here.
5863 */
tcp_validate_incoming(struct sock * sk,struct sk_buff * skb,const struct tcphdr * th,int syn_inerr)5864 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5865 const struct tcphdr *th, int syn_inerr)
5866 {
5867 struct tcp_sock *tp = tcp_sk(sk);
5868 SKB_DR(reason);
5869
5870 /* RFC1323: H1. Apply PAWS check first. */
5871 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5872 tp->rx_opt.saw_tstamp &&
5873 tcp_paws_discard(sk, skb)) {
5874 if (!th->rst) {
5875 if (unlikely(th->syn))
5876 goto syn_challenge;
5877 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5878 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5879 LINUX_MIB_TCPACKSKIPPEDPAWS,
5880 &tp->last_oow_ack_time))
5881 tcp_send_dupack(sk, skb);
5882 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
5883 goto discard;
5884 }
5885 /* Reset is accepted even if it did not pass PAWS. */
5886 }
5887
5888 /* Step 1: check sequence number */
5889 reason = tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5890 if (reason) {
5891 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5892 * (RST) segments are validated by checking their SEQ-fields."
5893 * And page 69: "If an incoming segment is not acceptable,
5894 * an acknowledgment should be sent in reply (unless the RST
5895 * bit is set, if so drop the segment and return)".
5896 */
5897 if (!th->rst) {
5898 if (th->syn)
5899 goto syn_challenge;
5900 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5901 LINUX_MIB_TCPACKSKIPPEDSEQ,
5902 &tp->last_oow_ack_time))
5903 tcp_send_dupack(sk, skb);
5904 } else if (tcp_reset_check(sk, skb)) {
5905 goto reset;
5906 }
5907 goto discard;
5908 }
5909
5910 /* Step 2: check RST bit */
5911 if (th->rst) {
5912 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5913 * FIN and SACK too if available):
5914 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5915 * the right-most SACK block,
5916 * then
5917 * RESET the connection
5918 * else
5919 * Send a challenge ACK
5920 */
5921 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5922 tcp_reset_check(sk, skb))
5923 goto reset;
5924
5925 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5926 struct tcp_sack_block *sp = &tp->selective_acks[0];
5927 int max_sack = sp[0].end_seq;
5928 int this_sack;
5929
5930 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5931 ++this_sack) {
5932 max_sack = after(sp[this_sack].end_seq,
5933 max_sack) ?
5934 sp[this_sack].end_seq : max_sack;
5935 }
5936
5937 if (TCP_SKB_CB(skb)->seq == max_sack)
5938 goto reset;
5939 }
5940
5941 /* Disable TFO if RST is out-of-order
5942 * and no data has been received
5943 * for current active TFO socket
5944 */
5945 if (tp->syn_fastopen && !tp->data_segs_in &&
5946 sk->sk_state == TCP_ESTABLISHED)
5947 tcp_fastopen_active_disable(sk);
5948 tcp_send_challenge_ack(sk);
5949 SKB_DR_SET(reason, TCP_RESET);
5950 goto discard;
5951 }
5952
5953 /* step 3: check security and precedence [ignored] */
5954
5955 /* step 4: Check for a SYN
5956 * RFC 5961 4.2 : Send a challenge ack
5957 */
5958 if (th->syn) {
5959 syn_challenge:
5960 if (syn_inerr)
5961 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5962 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5963 tcp_send_challenge_ack(sk);
5964 SKB_DR_SET(reason, TCP_INVALID_SYN);
5965 goto discard;
5966 }
5967
5968 bpf_skops_parse_hdr(sk, skb);
5969
5970 return true;
5971
5972 discard:
5973 tcp_drop_reason(sk, skb, reason);
5974 return false;
5975
5976 reset:
5977 tcp_reset(sk, skb);
5978 __kfree_skb(skb);
5979 return false;
5980 }
5981
5982 /*
5983 * TCP receive function for the ESTABLISHED state.
5984 *
5985 * It is split into a fast path and a slow path. The fast path is
5986 * disabled when:
5987 * - A zero window was announced from us - zero window probing
5988 * is only handled properly in the slow path.
5989 * - Out of order segments arrived.
5990 * - Urgent data is expected.
5991 * - There is no buffer space left
5992 * - Unexpected TCP flags/window values/header lengths are received
5993 * (detected by checking the TCP header against pred_flags)
5994 * - Data is sent in both directions. Fast path only supports pure senders
5995 * or pure receivers (this means either the sequence number or the ack
5996 * value must stay constant)
5997 * - Unexpected TCP option.
5998 *
5999 * When these conditions are not satisfied it drops into a standard
6000 * receive procedure patterned after RFC793 to handle all cases.
6001 * The first three cases are guaranteed by proper pred_flags setting,
6002 * the rest is checked inline. Fast processing is turned on in
6003 * tcp_data_queue when everything is OK.
6004 */
tcp_rcv_established(struct sock * sk,struct sk_buff * skb)6005 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
6006 {
6007 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED;
6008 const struct tcphdr *th = (const struct tcphdr *)skb->data;
6009 struct tcp_sock *tp = tcp_sk(sk);
6010 unsigned int len = skb->len;
6011
6012 /* TCP congestion window tracking */
6013 trace_tcp_probe(sk, skb);
6014
6015 tcp_mstamp_refresh(tp);
6016 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
6017 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
6018 /*
6019 * Header prediction.
6020 * The code loosely follows the one in the famous
6021 * "30 instruction TCP receive" Van Jacobson mail.
6022 *
6023 * Van's trick is to deposit buffers into socket queue
6024 * on a device interrupt, to call tcp_recv function
6025 * on the receive process context and checksum and copy
6026 * the buffer to user space. smart...
6027 *
6028 * Our current scheme is not silly either but we take the
6029 * extra cost of the net_bh soft interrupt processing...
6030 * We do checksum and copy also but from device to kernel.
6031 */
6032
6033 tp->rx_opt.saw_tstamp = 0;
6034
6035 /* pred_flags is 0xS?10 << 16 + snd_wnd
6036 * if header_prediction is to be made
6037 * 'S' will always be tp->tcp_header_len >> 2
6038 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
6039 * turn it off (when there are holes in the receive
6040 * space for instance)
6041 * PSH flag is ignored.
6042 */
6043
6044 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
6045 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
6046 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6047 int tcp_header_len = tp->tcp_header_len;
6048
6049 /* Timestamp header prediction: tcp_header_len
6050 * is automatically equal to th->doff*4 due to pred_flags
6051 * match.
6052 */
6053
6054 /* Check timestamp */
6055 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
6056 /* No? Slow path! */
6057 if (!tcp_parse_aligned_timestamp(tp, th))
6058 goto slow_path;
6059
6060 /* If PAWS failed, check it more carefully in slow path */
6061 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
6062 goto slow_path;
6063
6064 /* DO NOT update ts_recent here, if checksum fails
6065 * and timestamp was corrupted part, it will result
6066 * in a hung connection since we will drop all
6067 * future packets due to the PAWS test.
6068 */
6069 }
6070
6071 if (len <= tcp_header_len) {
6072 /* Bulk data transfer: sender */
6073 if (len == tcp_header_len) {
6074 /* Predicted packet is in window by definition.
6075 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
6076 * Hence, check seq<=rcv_wup reduces to:
6077 */
6078 if (tcp_header_len ==
6079 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
6080 tp->rcv_nxt == tp->rcv_wup)
6081 tcp_store_ts_recent(tp);
6082
6083 /* We know that such packets are checksummed
6084 * on entry.
6085 */
6086 tcp_ack(sk, skb, 0);
6087 __kfree_skb(skb);
6088 tcp_data_snd_check(sk);
6089 /* When receiving pure ack in fast path, update
6090 * last ts ecr directly instead of calling
6091 * tcp_rcv_rtt_measure_ts()
6092 */
6093 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
6094 return;
6095 } else { /* Header too small */
6096 reason = SKB_DROP_REASON_PKT_TOO_SMALL;
6097 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6098 goto discard;
6099 }
6100 } else {
6101 int eaten = 0;
6102 bool fragstolen = false;
6103
6104 if (tcp_checksum_complete(skb))
6105 goto csum_error;
6106
6107 if ((int)skb->truesize > sk->sk_forward_alloc)
6108 goto step5;
6109
6110 /* Predicted packet is in window by definition.
6111 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
6112 * Hence, check seq<=rcv_wup reduces to:
6113 */
6114 if (tcp_header_len ==
6115 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
6116 tp->rcv_nxt == tp->rcv_wup)
6117 tcp_store_ts_recent(tp);
6118
6119 tcp_rcv_rtt_measure_ts(sk, skb);
6120
6121 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
6122
6123 /* Bulk data transfer: receiver */
6124 skb_dst_drop(skb);
6125 __skb_pull(skb, tcp_header_len);
6126 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
6127
6128 tcp_event_data_recv(sk, skb);
6129
6130 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
6131 /* Well, only one small jumplet in fast path... */
6132 tcp_ack(sk, skb, FLAG_DATA);
6133 tcp_data_snd_check(sk);
6134 if (!inet_csk_ack_scheduled(sk))
6135 goto no_ack;
6136 } else {
6137 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
6138 }
6139
6140 __tcp_ack_snd_check(sk, 0);
6141 no_ack:
6142 if (eaten)
6143 kfree_skb_partial(skb, fragstolen);
6144 tcp_data_ready(sk);
6145 return;
6146 }
6147 }
6148
6149 slow_path:
6150 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
6151 goto csum_error;
6152
6153 if (!th->ack && !th->rst && !th->syn) {
6154 reason = SKB_DROP_REASON_TCP_FLAGS;
6155 goto discard;
6156 }
6157
6158 /*
6159 * Standard slow path.
6160 */
6161
6162 if (!tcp_validate_incoming(sk, skb, th, 1))
6163 return;
6164
6165 step5:
6166 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT);
6167 if ((int)reason < 0) {
6168 reason = -reason;
6169 goto discard;
6170 }
6171 tcp_rcv_rtt_measure_ts(sk, skb);
6172
6173 /* Process urgent data. */
6174 tcp_urg(sk, skb, th);
6175
6176 /* step 7: process the segment text */
6177 tcp_data_queue(sk, skb);
6178
6179 tcp_data_snd_check(sk);
6180 tcp_ack_snd_check(sk);
6181 return;
6182
6183 csum_error:
6184 reason = SKB_DROP_REASON_TCP_CSUM;
6185 trace_tcp_bad_csum(skb);
6186 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
6187 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6188
6189 discard:
6190 tcp_drop_reason(sk, skb, reason);
6191 }
6192 EXPORT_SYMBOL(tcp_rcv_established);
6193
tcp_init_transfer(struct sock * sk,int bpf_op,struct sk_buff * skb)6194 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
6195 {
6196 struct inet_connection_sock *icsk = inet_csk(sk);
6197 struct tcp_sock *tp = tcp_sk(sk);
6198
6199 tcp_mtup_init(sk);
6200 icsk->icsk_af_ops->rebuild_header(sk);
6201 tcp_init_metrics(sk);
6202
6203 /* Initialize the congestion window to start the transfer.
6204 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
6205 * retransmitted. In light of RFC6298 more aggressive 1sec
6206 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
6207 * retransmission has occurred.
6208 */
6209 if (tp->total_retrans > 1 && tp->undo_marker)
6210 tcp_snd_cwnd_set(tp, 1);
6211 else
6212 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk)));
6213 tp->snd_cwnd_stamp = tcp_jiffies32;
6214
6215 bpf_skops_established(sk, bpf_op, skb);
6216 /* Initialize congestion control unless BPF initialized it already: */
6217 if (!icsk->icsk_ca_initialized)
6218 tcp_init_congestion_control(sk);
6219 tcp_init_buffer_space(sk);
6220 }
6221
tcp_finish_connect(struct sock * sk,struct sk_buff * skb)6222 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
6223 {
6224 struct tcp_sock *tp = tcp_sk(sk);
6225 struct inet_connection_sock *icsk = inet_csk(sk);
6226
6227 tcp_ao_finish_connect(sk, skb);
6228 tcp_set_state(sk, TCP_ESTABLISHED);
6229 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
6230
6231 if (skb) {
6232 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
6233 security_inet_conn_established(sk, skb);
6234 sk_mark_napi_id(sk, skb);
6235 }
6236
6237 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
6238
6239 /* Prevent spurious tcp_cwnd_restart() on first data
6240 * packet.
6241 */
6242 tp->lsndtime = tcp_jiffies32;
6243
6244 if (sock_flag(sk, SOCK_KEEPOPEN))
6245 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
6246
6247 if (!tp->rx_opt.snd_wscale)
6248 __tcp_fast_path_on(tp, tp->snd_wnd);
6249 else
6250 tp->pred_flags = 0;
6251 }
6252
tcp_rcv_fastopen_synack(struct sock * sk,struct sk_buff * synack,struct tcp_fastopen_cookie * cookie)6253 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6254 struct tcp_fastopen_cookie *cookie)
6255 {
6256 struct tcp_sock *tp = tcp_sk(sk);
6257 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
6258 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6259 bool syn_drop = false;
6260
6261 if (mss == tp->rx_opt.user_mss) {
6262 struct tcp_options_received opt;
6263
6264 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6265 tcp_clear_options(&opt);
6266 opt.user_mss = opt.mss_clamp = 0;
6267 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
6268 mss = opt.mss_clamp;
6269 }
6270
6271 if (!tp->syn_fastopen) {
6272 /* Ignore an unsolicited cookie */
6273 cookie->len = -1;
6274 } else if (tp->total_retrans) {
6275 /* SYN timed out and the SYN-ACK neither has a cookie nor
6276 * acknowledges data. Presumably the remote received only
6277 * the retransmitted (regular) SYNs: either the original
6278 * SYN-data or the corresponding SYN-ACK was dropped.
6279 */
6280 syn_drop = (cookie->len < 0 && data);
6281 } else if (cookie->len < 0 && !tp->syn_data) {
6282 /* We requested a cookie but didn't get it. If we did not use
6283 * the (old) exp opt format then try so next time (try_exp=1).
6284 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6285 */
6286 try_exp = tp->syn_fastopen_exp ? 2 : 1;
6287 }
6288
6289 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
6290
6291 if (data) { /* Retransmit unacked data in SYN */
6292 if (tp->total_retrans)
6293 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6294 else
6295 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
6296 skb_rbtree_walk_from(data)
6297 tcp_mark_skb_lost(sk, data);
6298 tcp_xmit_retransmit_queue(sk);
6299 NET_INC_STATS(sock_net(sk),
6300 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6301 return true;
6302 }
6303 tp->syn_data_acked = tp->syn_data;
6304 if (tp->syn_data_acked) {
6305 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6306 /* SYN-data is counted as two separate packets in tcp_ack() */
6307 if (tp->delivered > 1)
6308 --tp->delivered;
6309 }
6310
6311 tcp_fastopen_add_skb(sk, synack);
6312
6313 return false;
6314 }
6315
smc_check_reset_syn(struct tcp_sock * tp)6316 static void smc_check_reset_syn(struct tcp_sock *tp)
6317 {
6318 #if IS_ENABLED(CONFIG_SMC)
6319 if (static_branch_unlikely(&tcp_have_smc)) {
6320 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6321 tp->syn_smc = 0;
6322 }
6323 #endif
6324 }
6325
tcp_try_undo_spurious_syn(struct sock * sk)6326 static void tcp_try_undo_spurious_syn(struct sock *sk)
6327 {
6328 struct tcp_sock *tp = tcp_sk(sk);
6329 u32 syn_stamp;
6330
6331 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6332 * spurious if the ACK's timestamp option echo value matches the
6333 * original SYN timestamp.
6334 */
6335 syn_stamp = tp->retrans_stamp;
6336 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6337 syn_stamp == tp->rx_opt.rcv_tsecr)
6338 tp->undo_marker = 0;
6339 }
6340
tcp_rcv_synsent_state_process(struct sock * sk,struct sk_buff * skb,const struct tcphdr * th)6341 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6342 const struct tcphdr *th)
6343 {
6344 struct inet_connection_sock *icsk = inet_csk(sk);
6345 struct tcp_sock *tp = tcp_sk(sk);
6346 struct tcp_fastopen_cookie foc = { .len = -1 };
6347 int saved_clamp = tp->rx_opt.mss_clamp;
6348 bool fastopen_fail;
6349 SKB_DR(reason);
6350
6351 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6352 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6353 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6354
6355 if (th->ack) {
6356 /* rfc793:
6357 * "If the state is SYN-SENT then
6358 * first check the ACK bit
6359 * If the ACK bit is set
6360 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6361 * a reset (unless the RST bit is set, if so drop
6362 * the segment and return)"
6363 */
6364 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6365 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6366 /* Previous FIN/ACK or RST/ACK might be ignored. */
6367 if (icsk->icsk_retransmits == 0)
6368 inet_csk_reset_xmit_timer(sk,
6369 ICSK_TIME_RETRANS,
6370 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
6371 SKB_DR_SET(reason, TCP_INVALID_ACK_SEQUENCE);
6372 goto reset_and_undo;
6373 }
6374
6375 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6376 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6377 tcp_time_stamp_ts(tp))) {
6378 NET_INC_STATS(sock_net(sk),
6379 LINUX_MIB_PAWSACTIVEREJECTED);
6380 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6381 goto reset_and_undo;
6382 }
6383
6384 /* Now ACK is acceptable.
6385 *
6386 * "If the RST bit is set
6387 * If the ACK was acceptable then signal the user "error:
6388 * connection reset", drop the segment, enter CLOSED state,
6389 * delete TCB, and return."
6390 */
6391
6392 if (th->rst) {
6393 tcp_reset(sk, skb);
6394 consume:
6395 __kfree_skb(skb);
6396 return 0;
6397 }
6398
6399 /* rfc793:
6400 * "fifth, if neither of the SYN or RST bits is set then
6401 * drop the segment and return."
6402 *
6403 * See note below!
6404 * --ANK(990513)
6405 */
6406 if (!th->syn) {
6407 SKB_DR_SET(reason, TCP_FLAGS);
6408 goto discard_and_undo;
6409 }
6410 /* rfc793:
6411 * "If the SYN bit is on ...
6412 * are acceptable then ...
6413 * (our SYN has been ACKed), change the connection
6414 * state to ESTABLISHED..."
6415 */
6416
6417 tcp_ecn_rcv_synack(tp, th);
6418
6419 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6420 tcp_try_undo_spurious_syn(sk);
6421 tcp_ack(sk, skb, FLAG_SLOWPATH);
6422
6423 /* Ok.. it's good. Set up sequence numbers and
6424 * move to established.
6425 */
6426 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6427 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6428
6429 /* RFC1323: The window in SYN & SYN/ACK segments is
6430 * never scaled.
6431 */
6432 tp->snd_wnd = ntohs(th->window);
6433
6434 if (!tp->rx_opt.wscale_ok) {
6435 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6436 WRITE_ONCE(tp->window_clamp,
6437 min(tp->window_clamp, 65535U));
6438 }
6439
6440 if (tp->rx_opt.saw_tstamp) {
6441 tp->rx_opt.tstamp_ok = 1;
6442 tp->tcp_header_len =
6443 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6444 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6445 tcp_store_ts_recent(tp);
6446 } else {
6447 tp->tcp_header_len = sizeof(struct tcphdr);
6448 }
6449
6450 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6451 tcp_initialize_rcv_mss(sk);
6452
6453 /* Remember, tcp_poll() does not lock socket!
6454 * Change state from SYN-SENT only after copied_seq
6455 * is initialized. */
6456 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6457
6458 smc_check_reset_syn(tp);
6459
6460 smp_mb();
6461
6462 tcp_finish_connect(sk, skb);
6463
6464 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6465 tcp_rcv_fastopen_synack(sk, skb, &foc);
6466
6467 if (!sock_flag(sk, SOCK_DEAD)) {
6468 sk->sk_state_change(sk);
6469 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6470 }
6471 if (fastopen_fail)
6472 return -1;
6473 if (sk->sk_write_pending ||
6474 READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) ||
6475 inet_csk_in_pingpong_mode(sk)) {
6476 /* Save one ACK. Data will be ready after
6477 * several ticks, if write_pending is set.
6478 *
6479 * It may be deleted, but with this feature tcpdumps
6480 * look so _wonderfully_ clever, that I was not able
6481 * to stand against the temptation 8) --ANK
6482 */
6483 inet_csk_schedule_ack(sk);
6484 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6485 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6486 TCP_DELACK_MAX, TCP_RTO_MAX);
6487 goto consume;
6488 }
6489 tcp_send_ack(sk);
6490 return -1;
6491 }
6492
6493 /* No ACK in the segment */
6494
6495 if (th->rst) {
6496 /* rfc793:
6497 * "If the RST bit is set
6498 *
6499 * Otherwise (no ACK) drop the segment and return."
6500 */
6501 SKB_DR_SET(reason, TCP_RESET);
6502 goto discard_and_undo;
6503 }
6504
6505 /* PAWS check. */
6506 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6507 tcp_paws_reject(&tp->rx_opt, 0)) {
6508 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6509 goto discard_and_undo;
6510 }
6511 if (th->syn) {
6512 /* We see SYN without ACK. It is attempt of
6513 * simultaneous connect with crossed SYNs.
6514 * Particularly, it can be connect to self.
6515 */
6516 #ifdef CONFIG_TCP_AO
6517 struct tcp_ao_info *ao;
6518
6519 ao = rcu_dereference_protected(tp->ao_info,
6520 lockdep_sock_is_held(sk));
6521 if (ao) {
6522 WRITE_ONCE(ao->risn, th->seq);
6523 ao->rcv_sne = 0;
6524 }
6525 #endif
6526 tcp_set_state(sk, TCP_SYN_RECV);
6527
6528 if (tp->rx_opt.saw_tstamp) {
6529 tp->rx_opt.tstamp_ok = 1;
6530 tcp_store_ts_recent(tp);
6531 tp->tcp_header_len =
6532 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6533 } else {
6534 tp->tcp_header_len = sizeof(struct tcphdr);
6535 }
6536
6537 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6538 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6539 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6540
6541 /* RFC1323: The window in SYN & SYN/ACK segments is
6542 * never scaled.
6543 */
6544 tp->snd_wnd = ntohs(th->window);
6545 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6546 tp->max_window = tp->snd_wnd;
6547
6548 tcp_ecn_rcv_syn(tp, th);
6549
6550 tcp_mtup_init(sk);
6551 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6552 tcp_initialize_rcv_mss(sk);
6553
6554 tcp_send_synack(sk);
6555 #if 0
6556 /* Note, we could accept data and URG from this segment.
6557 * There are no obstacles to make this (except that we must
6558 * either change tcp_recvmsg() to prevent it from returning data
6559 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6560 *
6561 * However, if we ignore data in ACKless segments sometimes,
6562 * we have no reasons to accept it sometimes.
6563 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6564 * is not flawless. So, discard packet for sanity.
6565 * Uncomment this return to process the data.
6566 */
6567 return -1;
6568 #else
6569 goto consume;
6570 #endif
6571 }
6572 /* "fifth, if neither of the SYN or RST bits is set then
6573 * drop the segment and return."
6574 */
6575
6576 discard_and_undo:
6577 tcp_clear_options(&tp->rx_opt);
6578 tp->rx_opt.mss_clamp = saved_clamp;
6579 tcp_drop_reason(sk, skb, reason);
6580 return 0;
6581
6582 reset_and_undo:
6583 tcp_clear_options(&tp->rx_opt);
6584 tp->rx_opt.mss_clamp = saved_clamp;
6585 /* we can reuse/return @reason to its caller to handle the exception */
6586 return reason;
6587 }
6588
tcp_rcv_synrecv_state_fastopen(struct sock * sk)6589 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6590 {
6591 struct tcp_sock *tp = tcp_sk(sk);
6592 struct request_sock *req;
6593
6594 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6595 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6596 */
6597 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out)
6598 tcp_try_undo_recovery(sk);
6599
6600 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6601 tcp_update_rto_time(tp);
6602 tp->retrans_stamp = 0;
6603 inet_csk(sk)->icsk_retransmits = 0;
6604
6605 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6606 * we no longer need req so release it.
6607 */
6608 req = rcu_dereference_protected(tp->fastopen_rsk,
6609 lockdep_sock_is_held(sk));
6610 reqsk_fastopen_remove(sk, req, false);
6611
6612 /* Re-arm the timer because data may have been sent out.
6613 * This is similar to the regular data transmission case
6614 * when new data has just been ack'ed.
6615 *
6616 * (TFO) - we could try to be more aggressive and
6617 * retransmitting any data sooner based on when they
6618 * are sent out.
6619 */
6620 tcp_rearm_rto(sk);
6621 }
6622
6623 /*
6624 * This function implements the receiving procedure of RFC 793 for
6625 * all states except ESTABLISHED and TIME_WAIT.
6626 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6627 * address independent.
6628 */
6629
6630 enum skb_drop_reason
tcp_rcv_state_process(struct sock * sk,struct sk_buff * skb)6631 tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6632 {
6633 struct tcp_sock *tp = tcp_sk(sk);
6634 struct inet_connection_sock *icsk = inet_csk(sk);
6635 const struct tcphdr *th = tcp_hdr(skb);
6636 struct request_sock *req;
6637 int queued = 0;
6638 SKB_DR(reason);
6639
6640 switch (sk->sk_state) {
6641 case TCP_CLOSE:
6642 SKB_DR_SET(reason, TCP_CLOSE);
6643 goto discard;
6644
6645 case TCP_LISTEN:
6646 if (th->ack)
6647 return SKB_DROP_REASON_TCP_FLAGS;
6648
6649 if (th->rst) {
6650 SKB_DR_SET(reason, TCP_RESET);
6651 goto discard;
6652 }
6653 if (th->syn) {
6654 if (th->fin) {
6655 SKB_DR_SET(reason, TCP_FLAGS);
6656 goto discard;
6657 }
6658 /* It is possible that we process SYN packets from backlog,
6659 * so we need to make sure to disable BH and RCU right there.
6660 */
6661 rcu_read_lock();
6662 local_bh_disable();
6663 icsk->icsk_af_ops->conn_request(sk, skb);
6664 local_bh_enable();
6665 rcu_read_unlock();
6666
6667 consume_skb(skb);
6668 return 0;
6669 }
6670 SKB_DR_SET(reason, TCP_FLAGS);
6671 goto discard;
6672
6673 case TCP_SYN_SENT:
6674 tp->rx_opt.saw_tstamp = 0;
6675 tcp_mstamp_refresh(tp);
6676 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6677 if (queued >= 0)
6678 return queued;
6679
6680 /* Do step6 onward by hand. */
6681 tcp_urg(sk, skb, th);
6682 __kfree_skb(skb);
6683 tcp_data_snd_check(sk);
6684 return 0;
6685 }
6686
6687 tcp_mstamp_refresh(tp);
6688 tp->rx_opt.saw_tstamp = 0;
6689 req = rcu_dereference_protected(tp->fastopen_rsk,
6690 lockdep_sock_is_held(sk));
6691 if (req) {
6692 bool req_stolen;
6693
6694 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6695 sk->sk_state != TCP_FIN_WAIT1);
6696
6697 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) {
6698 SKB_DR_SET(reason, TCP_FASTOPEN);
6699 goto discard;
6700 }
6701 }
6702
6703 if (!th->ack && !th->rst && !th->syn) {
6704 SKB_DR_SET(reason, TCP_FLAGS);
6705 goto discard;
6706 }
6707 if (!tcp_validate_incoming(sk, skb, th, 0))
6708 return 0;
6709
6710 /* step 5: check the ACK field */
6711 reason = tcp_ack(sk, skb, FLAG_SLOWPATH |
6712 FLAG_UPDATE_TS_RECENT |
6713 FLAG_NO_CHALLENGE_ACK);
6714
6715 if ((int)reason <= 0) {
6716 if (sk->sk_state == TCP_SYN_RECV) {
6717 /* send one RST */
6718 if (!reason)
6719 return SKB_DROP_REASON_TCP_OLD_ACK;
6720 return -reason;
6721 }
6722 /* accept old ack during closing */
6723 if ((int)reason < 0) {
6724 tcp_send_challenge_ack(sk);
6725 reason = -reason;
6726 goto discard;
6727 }
6728 }
6729 SKB_DR_SET(reason, NOT_SPECIFIED);
6730 switch (sk->sk_state) {
6731 case TCP_SYN_RECV:
6732 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6733 if (!tp->srtt_us)
6734 tcp_synack_rtt_meas(sk, req);
6735
6736 if (req) {
6737 tcp_rcv_synrecv_state_fastopen(sk);
6738 } else {
6739 tcp_try_undo_spurious_syn(sk);
6740 tp->retrans_stamp = 0;
6741 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6742 skb);
6743 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6744 }
6745 tcp_ao_established(sk);
6746 smp_mb();
6747 tcp_set_state(sk, TCP_ESTABLISHED);
6748 sk->sk_state_change(sk);
6749
6750 /* Note, that this wakeup is only for marginal crossed SYN case.
6751 * Passively open sockets are not waked up, because
6752 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6753 */
6754 if (sk->sk_socket)
6755 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6756
6757 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6758 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6759 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6760
6761 if (tp->rx_opt.tstamp_ok)
6762 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6763
6764 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6765 tcp_update_pacing_rate(sk);
6766
6767 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6768 tp->lsndtime = tcp_jiffies32;
6769
6770 tcp_initialize_rcv_mss(sk);
6771 tcp_fast_path_on(tp);
6772 if (sk->sk_shutdown & SEND_SHUTDOWN)
6773 tcp_shutdown(sk, SEND_SHUTDOWN);
6774 break;
6775
6776 case TCP_FIN_WAIT1: {
6777 int tmo;
6778
6779 if (req)
6780 tcp_rcv_synrecv_state_fastopen(sk);
6781
6782 if (tp->snd_una != tp->write_seq)
6783 break;
6784
6785 tcp_set_state(sk, TCP_FIN_WAIT2);
6786 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN);
6787
6788 sk_dst_confirm(sk);
6789
6790 if (!sock_flag(sk, SOCK_DEAD)) {
6791 /* Wake up lingering close() */
6792 sk->sk_state_change(sk);
6793 break;
6794 }
6795
6796 if (READ_ONCE(tp->linger2) < 0) {
6797 tcp_done(sk);
6798 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6799 return SKB_DROP_REASON_TCP_ABORT_ON_DATA;
6800 }
6801 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6802 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6803 /* Receive out of order FIN after close() */
6804 if (tp->syn_fastopen && th->fin)
6805 tcp_fastopen_active_disable(sk);
6806 tcp_done(sk);
6807 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6808 return SKB_DROP_REASON_TCP_ABORT_ON_DATA;
6809 }
6810
6811 tmo = tcp_fin_time(sk);
6812 if (tmo > TCP_TIMEWAIT_LEN) {
6813 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6814 } else if (th->fin || sock_owned_by_user(sk)) {
6815 /* Bad case. We could lose such FIN otherwise.
6816 * It is not a big problem, but it looks confusing
6817 * and not so rare event. We still can lose it now,
6818 * if it spins in bh_lock_sock(), but it is really
6819 * marginal case.
6820 */
6821 inet_csk_reset_keepalive_timer(sk, tmo);
6822 } else {
6823 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6824 goto consume;
6825 }
6826 break;
6827 }
6828
6829 case TCP_CLOSING:
6830 if (tp->snd_una == tp->write_seq) {
6831 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6832 goto consume;
6833 }
6834 break;
6835
6836 case TCP_LAST_ACK:
6837 if (tp->snd_una == tp->write_seq) {
6838 tcp_update_metrics(sk);
6839 tcp_done(sk);
6840 goto consume;
6841 }
6842 break;
6843 }
6844
6845 /* step 6: check the URG bit */
6846 tcp_urg(sk, skb, th);
6847
6848 /* step 7: process the segment text */
6849 switch (sk->sk_state) {
6850 case TCP_CLOSE_WAIT:
6851 case TCP_CLOSING:
6852 case TCP_LAST_ACK:
6853 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6854 /* If a subflow has been reset, the packet should not
6855 * continue to be processed, drop the packet.
6856 */
6857 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb))
6858 goto discard;
6859 break;
6860 }
6861 fallthrough;
6862 case TCP_FIN_WAIT1:
6863 case TCP_FIN_WAIT2:
6864 /* RFC 793 says to queue data in these states,
6865 * RFC 1122 says we MUST send a reset.
6866 * BSD 4.4 also does reset.
6867 */
6868 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6869 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6870 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6871 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6872 tcp_reset(sk, skb);
6873 return SKB_DROP_REASON_TCP_ABORT_ON_DATA;
6874 }
6875 }
6876 fallthrough;
6877 case TCP_ESTABLISHED:
6878 tcp_data_queue(sk, skb);
6879 queued = 1;
6880 break;
6881 }
6882
6883 /* tcp_data could move socket to TIME-WAIT */
6884 if (sk->sk_state != TCP_CLOSE) {
6885 tcp_data_snd_check(sk);
6886 tcp_ack_snd_check(sk);
6887 }
6888
6889 if (!queued) {
6890 discard:
6891 tcp_drop_reason(sk, skb, reason);
6892 }
6893 return 0;
6894
6895 consume:
6896 __kfree_skb(skb);
6897 return 0;
6898 }
6899 EXPORT_SYMBOL(tcp_rcv_state_process);
6900
pr_drop_req(struct request_sock * req,__u16 port,int family)6901 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6902 {
6903 struct inet_request_sock *ireq = inet_rsk(req);
6904
6905 if (family == AF_INET)
6906 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6907 &ireq->ir_rmt_addr, port);
6908 #if IS_ENABLED(CONFIG_IPV6)
6909 else if (family == AF_INET6)
6910 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6911 &ireq->ir_v6_rmt_addr, port);
6912 #endif
6913 }
6914
6915 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6916 *
6917 * If we receive a SYN packet with these bits set, it means a
6918 * network is playing bad games with TOS bits. In order to
6919 * avoid possible false congestion notifications, we disable
6920 * TCP ECN negotiation.
6921 *
6922 * Exception: tcp_ca wants ECN. This is required for DCTCP
6923 * congestion control: Linux DCTCP asserts ECT on all packets,
6924 * including SYN, which is most optimal solution; however,
6925 * others, such as FreeBSD do not.
6926 *
6927 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6928 * set, indicating the use of a future TCP extension (such as AccECN). See
6929 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6930 * extensions.
6931 */
tcp_ecn_create_request(struct request_sock * req,const struct sk_buff * skb,const struct sock * listen_sk,const struct dst_entry * dst)6932 static void tcp_ecn_create_request(struct request_sock *req,
6933 const struct sk_buff *skb,
6934 const struct sock *listen_sk,
6935 const struct dst_entry *dst)
6936 {
6937 const struct tcphdr *th = tcp_hdr(skb);
6938 const struct net *net = sock_net(listen_sk);
6939 bool th_ecn = th->ece && th->cwr;
6940 bool ect, ecn_ok;
6941 u32 ecn_ok_dst;
6942
6943 if (!th_ecn)
6944 return;
6945
6946 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6947 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6948 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst;
6949
6950 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6951 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6952 tcp_bpf_ca_needs_ecn((struct sock *)req))
6953 inet_rsk(req)->ecn_ok = 1;
6954 }
6955
tcp_openreq_init(struct request_sock * req,const struct tcp_options_received * rx_opt,struct sk_buff * skb,const struct sock * sk)6956 static void tcp_openreq_init(struct request_sock *req,
6957 const struct tcp_options_received *rx_opt,
6958 struct sk_buff *skb, const struct sock *sk)
6959 {
6960 struct inet_request_sock *ireq = inet_rsk(req);
6961
6962 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6963 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6964 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6965 tcp_rsk(req)->snt_synack = 0;
6966 tcp_rsk(req)->last_oow_ack_time = 0;
6967 req->mss = rx_opt->mss_clamp;
6968 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6969 ireq->tstamp_ok = rx_opt->tstamp_ok;
6970 ireq->sack_ok = rx_opt->sack_ok;
6971 ireq->snd_wscale = rx_opt->snd_wscale;
6972 ireq->wscale_ok = rx_opt->wscale_ok;
6973 ireq->acked = 0;
6974 ireq->ecn_ok = 0;
6975 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6976 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6977 ireq->ir_mark = inet_request_mark(sk, skb);
6978 #if IS_ENABLED(CONFIG_SMC)
6979 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested &&
6980 tcp_sk(sk)->smc_hs_congested(sk));
6981 #endif
6982 }
6983
inet_reqsk_alloc(const struct request_sock_ops * ops,struct sock * sk_listener,bool attach_listener)6984 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6985 struct sock *sk_listener,
6986 bool attach_listener)
6987 {
6988 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6989 attach_listener);
6990
6991 if (req) {
6992 struct inet_request_sock *ireq = inet_rsk(req);
6993
6994 ireq->ireq_opt = NULL;
6995 #if IS_ENABLED(CONFIG_IPV6)
6996 ireq->pktopts = NULL;
6997 #endif
6998 atomic64_set(&ireq->ir_cookie, 0);
6999 ireq->ireq_state = TCP_NEW_SYN_RECV;
7000 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
7001 ireq->ireq_family = sk_listener->sk_family;
7002 req->timeout = TCP_TIMEOUT_INIT;
7003 }
7004
7005 return req;
7006 }
7007 EXPORT_SYMBOL(inet_reqsk_alloc);
7008
7009 /*
7010 * Return true if a syncookie should be sent
7011 */
tcp_syn_flood_action(struct sock * sk,const char * proto)7012 static bool tcp_syn_flood_action(struct sock *sk, const char *proto)
7013 {
7014 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
7015 const char *msg = "Dropping request";
7016 struct net *net = sock_net(sk);
7017 bool want_cookie = false;
7018 u8 syncookies;
7019
7020 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
7021
7022 #ifdef CONFIG_SYN_COOKIES
7023 if (syncookies) {
7024 msg = "Sending cookies";
7025 want_cookie = true;
7026 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
7027 } else
7028 #endif
7029 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
7030
7031 if (!READ_ONCE(queue->synflood_warned) && syncookies != 2 &&
7032 xchg(&queue->synflood_warned, 1) == 0) {
7033 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) {
7034 net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n",
7035 proto, inet6_rcv_saddr(sk),
7036 sk->sk_num, msg);
7037 } else {
7038 net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n",
7039 proto, &sk->sk_rcv_saddr,
7040 sk->sk_num, msg);
7041 }
7042 }
7043
7044 return want_cookie;
7045 }
7046
tcp_reqsk_record_syn(const struct sock * sk,struct request_sock * req,const struct sk_buff * skb)7047 static void tcp_reqsk_record_syn(const struct sock *sk,
7048 struct request_sock *req,
7049 const struct sk_buff *skb)
7050 {
7051 if (tcp_sk(sk)->save_syn) {
7052 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
7053 struct saved_syn *saved_syn;
7054 u32 mac_hdrlen;
7055 void *base;
7056
7057 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
7058 base = skb_mac_header(skb);
7059 mac_hdrlen = skb_mac_header_len(skb);
7060 len += mac_hdrlen;
7061 } else {
7062 base = skb_network_header(skb);
7063 mac_hdrlen = 0;
7064 }
7065
7066 saved_syn = kmalloc(struct_size(saved_syn, data, len),
7067 GFP_ATOMIC);
7068 if (saved_syn) {
7069 saved_syn->mac_hdrlen = mac_hdrlen;
7070 saved_syn->network_hdrlen = skb_network_header_len(skb);
7071 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
7072 memcpy(saved_syn->data, base, len);
7073 req->saved_syn = saved_syn;
7074 }
7075 }
7076 }
7077
7078 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
7079 * used for SYN cookie generation.
7080 */
tcp_get_syncookie_mss(struct request_sock_ops * rsk_ops,const struct tcp_request_sock_ops * af_ops,struct sock * sk,struct tcphdr * th)7081 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
7082 const struct tcp_request_sock_ops *af_ops,
7083 struct sock *sk, struct tcphdr *th)
7084 {
7085 struct tcp_sock *tp = tcp_sk(sk);
7086 u16 mss;
7087
7088 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 &&
7089 !inet_csk_reqsk_queue_is_full(sk))
7090 return 0;
7091
7092 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
7093 return 0;
7094
7095 if (sk_acceptq_is_full(sk)) {
7096 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
7097 return 0;
7098 }
7099
7100 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
7101 if (!mss)
7102 mss = af_ops->mss_clamp;
7103
7104 return mss;
7105 }
7106 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
7107
tcp_conn_request(struct request_sock_ops * rsk_ops,const struct tcp_request_sock_ops * af_ops,struct sock * sk,struct sk_buff * skb)7108 int tcp_conn_request(struct request_sock_ops *rsk_ops,
7109 const struct tcp_request_sock_ops *af_ops,
7110 struct sock *sk, struct sk_buff *skb)
7111 {
7112 struct tcp_fastopen_cookie foc = { .len = -1 };
7113 struct tcp_options_received tmp_opt;
7114 struct tcp_sock *tp = tcp_sk(sk);
7115 struct net *net = sock_net(sk);
7116 struct sock *fastopen_sk = NULL;
7117 struct request_sock *req;
7118 bool want_cookie = false;
7119 struct dst_entry *dst;
7120 struct flowi fl;
7121 u8 syncookies;
7122 u32 isn;
7123
7124 #ifdef CONFIG_TCP_AO
7125 const struct tcp_ao_hdr *aoh;
7126 #endif
7127
7128 isn = __this_cpu_read(tcp_tw_isn);
7129 if (isn) {
7130 /* TW buckets are converted to open requests without
7131 * limitations, they conserve resources and peer is
7132 * evidently real one.
7133 */
7134 __this_cpu_write(tcp_tw_isn, 0);
7135 } else {
7136 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
7137
7138 if (syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) {
7139 want_cookie = tcp_syn_flood_action(sk,
7140 rsk_ops->slab_name);
7141 if (!want_cookie)
7142 goto drop;
7143 }
7144 }
7145
7146 if (sk_acceptq_is_full(sk)) {
7147 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
7148 goto drop;
7149 }
7150
7151 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
7152 if (!req)
7153 goto drop;
7154
7155 req->syncookie = want_cookie;
7156 tcp_rsk(req)->af_specific = af_ops;
7157 tcp_rsk(req)->ts_off = 0;
7158 tcp_rsk(req)->req_usec_ts = false;
7159 #if IS_ENABLED(CONFIG_MPTCP)
7160 tcp_rsk(req)->is_mptcp = 0;
7161 #endif
7162
7163 tcp_clear_options(&tmp_opt);
7164 tmp_opt.mss_clamp = af_ops->mss_clamp;
7165 tmp_opt.user_mss = tp->rx_opt.user_mss;
7166 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
7167 want_cookie ? NULL : &foc);
7168
7169 if (want_cookie && !tmp_opt.saw_tstamp)
7170 tcp_clear_options(&tmp_opt);
7171
7172 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
7173 tmp_opt.smc_ok = 0;
7174
7175 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
7176 tcp_openreq_init(req, &tmp_opt, skb, sk);
7177 inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk);
7178
7179 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
7180 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
7181
7182 dst = af_ops->route_req(sk, skb, &fl, req, isn);
7183 if (!dst)
7184 goto drop_and_free;
7185
7186 if (tmp_opt.tstamp_ok) {
7187 tcp_rsk(req)->req_usec_ts = dst_tcp_usec_ts(dst);
7188 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
7189 }
7190 if (!want_cookie && !isn) {
7191 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog);
7192
7193 /* Kill the following clause, if you dislike this way. */
7194 if (!syncookies &&
7195 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
7196 (max_syn_backlog >> 2)) &&
7197 !tcp_peer_is_proven(req, dst)) {
7198 /* Without syncookies last quarter of
7199 * backlog is filled with destinations,
7200 * proven to be alive.
7201 * It means that we continue to communicate
7202 * to destinations, already remembered
7203 * to the moment of synflood.
7204 */
7205 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
7206 rsk_ops->family);
7207 goto drop_and_release;
7208 }
7209
7210 isn = af_ops->init_seq(skb);
7211 }
7212
7213 tcp_ecn_create_request(req, skb, sk, dst);
7214
7215 if (want_cookie) {
7216 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
7217 if (!tmp_opt.tstamp_ok)
7218 inet_rsk(req)->ecn_ok = 0;
7219 }
7220
7221 #ifdef CONFIG_TCP_AO
7222 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh))
7223 goto drop_and_release; /* Invalid TCP options */
7224 if (aoh) {
7225 tcp_rsk(req)->used_tcp_ao = true;
7226 tcp_rsk(req)->ao_rcv_next = aoh->keyid;
7227 tcp_rsk(req)->ao_keyid = aoh->rnext_keyid;
7228
7229 } else {
7230 tcp_rsk(req)->used_tcp_ao = false;
7231 }
7232 #endif
7233 tcp_rsk(req)->snt_isn = isn;
7234 tcp_rsk(req)->txhash = net_tx_rndhash();
7235 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
7236 tcp_openreq_init_rwin(req, sk, dst);
7237 sk_rx_queue_set(req_to_sk(req), skb);
7238 if (!want_cookie) {
7239 tcp_reqsk_record_syn(sk, req, skb);
7240 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
7241 }
7242 if (fastopen_sk) {
7243 af_ops->send_synack(fastopen_sk, dst, &fl, req,
7244 &foc, TCP_SYNACK_FASTOPEN, skb);
7245 /* Add the child socket directly into the accept queue */
7246 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
7247 reqsk_fastopen_remove(fastopen_sk, req, false);
7248 bh_unlock_sock(fastopen_sk);
7249 sock_put(fastopen_sk);
7250 goto drop_and_free;
7251 }
7252 sk->sk_data_ready(sk);
7253 bh_unlock_sock(fastopen_sk);
7254 sock_put(fastopen_sk);
7255 } else {
7256 tcp_rsk(req)->tfo_listener = false;
7257 if (!want_cookie) {
7258 req->timeout = tcp_timeout_init((struct sock *)req);
7259 inet_csk_reqsk_queue_hash_add(sk, req, req->timeout);
7260 }
7261 af_ops->send_synack(sk, dst, &fl, req, &foc,
7262 !want_cookie ? TCP_SYNACK_NORMAL :
7263 TCP_SYNACK_COOKIE,
7264 skb);
7265 if (want_cookie) {
7266 reqsk_free(req);
7267 return 0;
7268 }
7269 }
7270 reqsk_put(req);
7271 return 0;
7272
7273 drop_and_release:
7274 dst_release(dst);
7275 drop_and_free:
7276 __reqsk_free(req);
7277 drop:
7278 tcp_listendrop(sk);
7279 return 0;
7280 }
7281 EXPORT_SYMBOL(tcp_conn_request);
7282