1 | // SPDX-License-Identifier: GPL-2.0 |
---|---|
2 | /* |
3 | * INET An implementation of the TCP/IP protocol suite for the LINUX |
4 | * operating system. INET is implemented using the BSD Socket |
5 | * interface as the means of communication with the user level. |
6 | * |
7 | * Implementation of the Transmission Control Protocol(TCP). |
8 | * |
9 | * Authors: Ross Biro |
10 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
11 | * Mark Evans, <evansmp@uhura.aston.ac.uk> |
12 | * Corey Minyard <wf-rch!minyard@relay.EU.net> |
13 | * Florian La Roche, <flla@stud.uni-sb.de> |
14 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> |
15 | * Linus Torvalds, <torvalds@cs.helsinki.fi> |
16 | * Alan Cox, <gw4pts@gw4pts.ampr.org> |
17 | * Matthew Dillon, <dillon@apollo.west.oic.com> |
18 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> |
19 | * Jorge Cwik, <jorge@laser.satlink.net> |
20 | */ |
21 | |
22 | /* |
23 | * Changes: |
24 | * Pedro Roque : Fast Retransmit/Recovery. |
25 | * Two receive queues. |
26 | * Retransmit queue handled by TCP. |
27 | * Better retransmit timer handling. |
28 | * New congestion avoidance. |
29 | * Header prediction. |
30 | * Variable renaming. |
31 | * |
32 | * Eric : Fast Retransmit. |
33 | * Randy Scott : MSS option defines. |
34 | * Eric Schenk : Fixes to slow start algorithm. |
35 | * Eric Schenk : Yet another double ACK bug. |
36 | * Eric Schenk : Delayed ACK bug fixes. |
37 | * Eric Schenk : Floyd style fast retrans war avoidance. |
38 | * David S. Miller : Don't allow zero congestion window. |
39 | * Eric Schenk : Fix retransmitter so that it sends |
40 | * next packet on ack of previous packet. |
41 | * Andi Kleen : Moved open_request checking here |
42 | * and process RSTs for open_requests. |
43 | * Andi Kleen : Better prune_queue, and other fixes. |
44 | * Andrey Savochkin: Fix RTT measurements in the presence of |
45 | * timestamps. |
46 | * Andrey Savochkin: Check sequence numbers correctly when |
47 | * removing SACKs due to in sequence incoming |
48 | * data segments. |
49 | * Andi Kleen: Make sure we never ack data there is not |
50 | * enough room for. Also make this condition |
51 | * a fatal error if it might still happen. |
52 | * Andi Kleen: Add tcp_measure_rcv_mss to make |
53 | * connections with MSS<min(MTU,ann. MSS) |
54 | * work without delayed acks. |
55 | * Andi Kleen: Process packets with PSH set in the |
56 | * fast path. |
57 | * J Hadi Salim: ECN support |
58 | * Andrei Gurtov, |
59 | * Pasi Sarolahti, |
60 | * Panu Kuhlberg: Experimental audit of TCP (re)transmission |
61 | * engine. Lots of bugs are found. |
62 | * Pasi Sarolahti: F-RTO for dealing with spurious RTOs |
63 | */ |
64 | |
65 | #define pr_fmt(fmt) "TCP: " fmt |
66 | |
67 | #include <linux/mm.h> |
68 | #include <linux/slab.h> |
69 | #include <linux/module.h> |
70 | #include <linux/sysctl.h> |
71 | #include <linux/kernel.h> |
72 | #include <linux/prefetch.h> |
73 | #include <net/dst.h> |
74 | #include <net/tcp.h> |
75 | #include <net/proto_memory.h> |
76 | #include <net/inet_common.h> |
77 | #include <linux/ipsec.h> |
78 | #include <linux/unaligned.h> |
79 | #include <linux/errqueue.h> |
80 | #include <trace/events/tcp.h> |
81 | #include <linux/jump_label_ratelimit.h> |
82 | #include <net/busy_poll.h> |
83 | #include <net/mptcp.h> |
84 | |
85 | int sysctl_tcp_max_orphans __read_mostly = NR_FILE; |
86 | |
87 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ |
88 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ |
89 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ |
90 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ |
91 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ |
92 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ |
93 | #define FLAG_ECE 0x40 /* ECE in this ACK */ |
94 | #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */ |
95 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ |
96 | #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */ |
97 | #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ |
98 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ |
99 | #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */ |
100 | #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ |
101 | #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */ |
102 | #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */ |
103 | #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */ |
104 | #define FLAG_DSACK_TLP 0x20000 /* DSACK for tail loss probe */ |
105 | #define FLAG_TS_PROGRESS 0x40000 /* Positive timestamp delta */ |
106 | |
107 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) |
108 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) |
109 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK) |
110 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) |
111 | |
112 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) |
113 | #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) |
114 | |
115 | #define REXMIT_NONE 0 /* no loss recovery to do */ |
116 | #define REXMIT_LOST 1 /* retransmit packets marked lost */ |
117 | #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */ |
118 | |
119 | #if IS_ENABLED(CONFIG_TLS_DEVICE) |
120 | static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ); |
121 | |
122 | void clean_acked_data_enable(struct tcp_sock *tp, |
123 | void (*cad)(struct sock *sk, u32 ack_seq)) |
124 | { |
125 | tp->tcp_clean_acked = cad; |
126 | static_branch_deferred_inc(&clean_acked_data_enabled); |
127 | } |
128 | EXPORT_SYMBOL_GPL(clean_acked_data_enable); |
129 | |
130 | void clean_acked_data_disable(struct tcp_sock *tp) |
131 | { |
132 | static_branch_slow_dec_deferred(&clean_acked_data_enabled); |
133 | tp->tcp_clean_acked = NULL; |
134 | } |
135 | EXPORT_SYMBOL_GPL(clean_acked_data_disable); |
136 | |
137 | void clean_acked_data_flush(void) |
138 | { |
139 | static_key_deferred_flush(&clean_acked_data_enabled); |
140 | } |
141 | EXPORT_SYMBOL_GPL(clean_acked_data_flush); |
142 | #endif |
143 | |
144 | #ifdef CONFIG_CGROUP_BPF |
145 | static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) |
146 | { |
147 | bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown && |
148 | BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), |
149 | BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG); |
150 | bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), |
151 | BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG); |
152 | struct bpf_sock_ops_kern sock_ops; |
153 | |
154 | if (likely(!unknown_opt && !parse_all_opt)) |
155 | return; |
156 | |
157 | /* The skb will be handled in the |
158 | * bpf_skops_established() or |
159 | * bpf_skops_write_hdr_opt(). |
160 | */ |
161 | switch (sk->sk_state) { |
162 | case TCP_SYN_RECV: |
163 | case TCP_SYN_SENT: |
164 | case TCP_LISTEN: |
165 | return; |
166 | } |
167 | |
168 | sock_owned_by_me(sk); |
169 | |
170 | memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); |
171 | sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB; |
172 | sock_ops.is_fullsock = 1; |
173 | sock_ops.is_locked_tcp_sock = 1; |
174 | sock_ops.sk = sk; |
175 | bpf_skops_init_skb(skops: &sock_ops, skb, end_offset: tcp_hdrlen(skb)); |
176 | |
177 | BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops); |
178 | } |
179 | |
180 | static void bpf_skops_established(struct sock *sk, int bpf_op, |
181 | struct sk_buff *skb) |
182 | { |
183 | struct bpf_sock_ops_kern sock_ops; |
184 | |
185 | sock_owned_by_me(sk); |
186 | |
187 | memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); |
188 | sock_ops.op = bpf_op; |
189 | sock_ops.is_fullsock = 1; |
190 | sock_ops.is_locked_tcp_sock = 1; |
191 | sock_ops.sk = sk; |
192 | /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */ |
193 | if (skb) |
194 | bpf_skops_init_skb(skops: &sock_ops, skb, end_offset: tcp_hdrlen(skb)); |
195 | |
196 | BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops); |
197 | } |
198 | #else |
199 | static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) |
200 | { |
201 | } |
202 | |
203 | static void bpf_skops_established(struct sock *sk, int bpf_op, |
204 | struct sk_buff *skb) |
205 | { |
206 | } |
207 | #endif |
208 | |
209 | static __cold void tcp_gro_dev_warn(const struct sock *sk, const struct sk_buff *skb, |
210 | unsigned int len) |
211 | { |
212 | struct net_device *dev; |
213 | |
214 | rcu_read_lock(); |
215 | dev = dev_get_by_index_rcu(net: sock_net(sk), ifindex: skb->skb_iif); |
216 | if (!dev || len >= READ_ONCE(dev->mtu)) |
217 | pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n", |
218 | dev ? dev->name : "Unknown driver"); |
219 | rcu_read_unlock(); |
220 | } |
221 | |
222 | /* Adapt the MSS value used to make delayed ack decision to the |
223 | * real world. |
224 | */ |
225 | static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) |
226 | { |
227 | struct inet_connection_sock *icsk = inet_csk(sk); |
228 | const unsigned int lss = icsk->icsk_ack.last_seg_size; |
229 | unsigned int len; |
230 | |
231 | icsk->icsk_ack.last_seg_size = 0; |
232 | |
233 | /* skb->len may jitter because of SACKs, even if peer |
234 | * sends good full-sized frames. |
235 | */ |
236 | len = skb_shinfo(skb)->gso_size ? : skb->len; |
237 | if (len >= icsk->icsk_ack.rcv_mss) { |
238 | /* Note: divides are still a bit expensive. |
239 | * For the moment, only adjust scaling_ratio |
240 | * when we update icsk_ack.rcv_mss. |
241 | */ |
242 | if (unlikely(len != icsk->icsk_ack.rcv_mss)) { |
243 | u64 val = (u64)skb->len << TCP_RMEM_TO_WIN_SCALE; |
244 | u8 old_ratio = tcp_sk(sk)->scaling_ratio; |
245 | |
246 | do_div(val, skb->truesize); |
247 | tcp_sk(sk)->scaling_ratio = val ? val : 1; |
248 | |
249 | if (old_ratio != tcp_sk(sk)->scaling_ratio) { |
250 | struct tcp_sock *tp = tcp_sk(sk); |
251 | |
252 | val = tcp_win_from_space(sk, space: sk->sk_rcvbuf); |
253 | tcp_set_window_clamp(sk, val); |
254 | |
255 | if (tp->window_clamp < tp->rcvq_space.space) |
256 | tp->rcvq_space.space = tp->window_clamp; |
257 | } |
258 | } |
259 | icsk->icsk_ack.rcv_mss = min_t(unsigned int, len, |
260 | tcp_sk(sk)->advmss); |
261 | /* Account for possibly-removed options */ |
262 | DO_ONCE_LITE_IF(len > icsk->icsk_ack.rcv_mss + MAX_TCP_OPTION_SPACE, |
263 | tcp_gro_dev_warn, sk, skb, len); |
264 | /* If the skb has a len of exactly 1*MSS and has the PSH bit |
265 | * set then it is likely the end of an application write. So |
266 | * more data may not be arriving soon, and yet the data sender |
267 | * may be waiting for an ACK if cwnd-bound or using TX zero |
268 | * copy. So we set ICSK_ACK_PUSHED here so that |
269 | * tcp_cleanup_rbuf() will send an ACK immediately if the app |
270 | * reads all of the data and is not ping-pong. If len > MSS |
271 | * then this logic does not matter (and does not hurt) because |
272 | * tcp_cleanup_rbuf() will always ACK immediately if the app |
273 | * reads data and there is more than an MSS of unACKed data. |
274 | */ |
275 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH) |
276 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
277 | } else { |
278 | /* Otherwise, we make more careful check taking into account, |
279 | * that SACKs block is variable. |
280 | * |
281 | * "len" is invariant segment length, including TCP header. |
282 | */ |
283 | len += skb->data - skb_transport_header(skb); |
284 | if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || |
285 | /* If PSH is not set, packet should be |
286 | * full sized, provided peer TCP is not badly broken. |
287 | * This observation (if it is correct 8)) allows |
288 | * to handle super-low mtu links fairly. |
289 | */ |
290 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && |
291 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { |
292 | /* Subtract also invariant (if peer is RFC compliant), |
293 | * tcp header plus fixed timestamp option length. |
294 | * Resulting "len" is MSS free of SACK jitter. |
295 | */ |
296 | len -= tcp_sk(sk)->tcp_header_len; |
297 | icsk->icsk_ack.last_seg_size = len; |
298 | if (len == lss) { |
299 | icsk->icsk_ack.rcv_mss = len; |
300 | return; |
301 | } |
302 | } |
303 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) |
304 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; |
305 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
306 | } |
307 | } |
308 | |
309 | static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks) |
310 | { |
311 | struct inet_connection_sock *icsk = inet_csk(sk); |
312 | unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); |
313 | |
314 | if (quickacks == 0) |
315 | quickacks = 2; |
316 | quickacks = min(quickacks, max_quickacks); |
317 | if (quickacks > icsk->icsk_ack.quick) |
318 | icsk->icsk_ack.quick = quickacks; |
319 | } |
320 | |
321 | static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks) |
322 | { |
323 | struct inet_connection_sock *icsk = inet_csk(sk); |
324 | |
325 | tcp_incr_quickack(sk, max_quickacks); |
326 | inet_csk_exit_pingpong_mode(sk); |
327 | icsk->icsk_ack.ato = TCP_ATO_MIN; |
328 | } |
329 | |
330 | /* Send ACKs quickly, if "quick" count is not exhausted |
331 | * and the session is not interactive. |
332 | */ |
333 | |
334 | static bool tcp_in_quickack_mode(struct sock *sk) |
335 | { |
336 | const struct inet_connection_sock *icsk = inet_csk(sk); |
337 | |
338 | return icsk->icsk_ack.dst_quick_ack || |
339 | (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk)); |
340 | } |
341 | |
342 | static void tcp_ecn_queue_cwr(struct tcp_sock *tp) |
343 | { |
344 | if (tcp_ecn_mode_rfc3168(tp)) |
345 | tp->ecn_flags |= TCP_ECN_QUEUE_CWR; |
346 | } |
347 | |
348 | static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb) |
349 | { |
350 | if (tcp_hdr(skb)->cwr) { |
351 | tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR; |
352 | |
353 | /* If the sender is telling us it has entered CWR, then its |
354 | * cwnd may be very low (even just 1 packet), so we should ACK |
355 | * immediately. |
356 | */ |
357 | if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) |
358 | inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; |
359 | } |
360 | } |
361 | |
362 | static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp) |
363 | { |
364 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
365 | } |
366 | |
367 | static void tcp_data_ecn_check(struct sock *sk, const struct sk_buff *skb) |
368 | { |
369 | struct tcp_sock *tp = tcp_sk(sk); |
370 | |
371 | if (tcp_ecn_disabled(tp)) |
372 | return; |
373 | |
374 | switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { |
375 | case INET_ECN_NOT_ECT: |
376 | /* Funny extension: if ECT is not set on a segment, |
377 | * and we already seen ECT on a previous segment, |
378 | * it is probably a retransmit. |
379 | */ |
380 | if (tp->ecn_flags & TCP_ECN_SEEN) |
381 | tcp_enter_quickack_mode(sk, max_quickacks: 2); |
382 | break; |
383 | case INET_ECN_CE: |
384 | if (tcp_ca_needs_ecn(sk)) |
385 | tcp_ca_event(sk, event: CA_EVENT_ECN_IS_CE); |
386 | |
387 | if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { |
388 | /* Better not delay acks, sender can have a very low cwnd */ |
389 | tcp_enter_quickack_mode(sk, max_quickacks: 2); |
390 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; |
391 | } |
392 | tp->ecn_flags |= TCP_ECN_SEEN; |
393 | break; |
394 | default: |
395 | if (tcp_ca_needs_ecn(sk)) |
396 | tcp_ca_event(sk, event: CA_EVENT_ECN_NO_CE); |
397 | tp->ecn_flags |= TCP_ECN_SEEN; |
398 | break; |
399 | } |
400 | } |
401 | |
402 | static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) |
403 | { |
404 | if (tcp_ecn_mode_rfc3168(tp) && (!th->ece || th->cwr)) |
405 | tcp_ecn_mode_set(tp, TCP_ECN_DISABLED); |
406 | } |
407 | |
408 | static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) |
409 | { |
410 | if (tcp_ecn_mode_rfc3168(tp) && (!th->ece || !th->cwr)) |
411 | tcp_ecn_mode_set(tp, TCP_ECN_DISABLED); |
412 | } |
413 | |
414 | static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) |
415 | { |
416 | if (th->ece && !th->syn && tcp_ecn_mode_rfc3168(tp)) |
417 | return true; |
418 | return false; |
419 | } |
420 | |
421 | static void tcp_count_delivered_ce(struct tcp_sock *tp, u32 ecn_count) |
422 | { |
423 | tp->delivered_ce += ecn_count; |
424 | } |
425 | |
426 | /* Updates the delivered and delivered_ce counts */ |
427 | static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered, |
428 | bool ece_ack) |
429 | { |
430 | tp->delivered += delivered; |
431 | if (ece_ack) |
432 | tcp_count_delivered_ce(tp, ecn_count: delivered); |
433 | } |
434 | |
435 | /* Buffer size and advertised window tuning. |
436 | * |
437 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. |
438 | */ |
439 | |
440 | static void tcp_sndbuf_expand(struct sock *sk) |
441 | { |
442 | const struct tcp_sock *tp = tcp_sk(sk); |
443 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; |
444 | int sndmem, per_mss; |
445 | u32 nr_segs; |
446 | |
447 | /* Worst case is non GSO/TSO : each frame consumes one skb |
448 | * and skb->head is kmalloced using power of two area of memory |
449 | */ |
450 | per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + |
451 | MAX_TCP_HEADER + |
452 | SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); |
453 | |
454 | per_mss = roundup_pow_of_two(per_mss) + |
455 | SKB_DATA_ALIGN(sizeof(struct sk_buff)); |
456 | |
457 | nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp)); |
458 | nr_segs = max_t(u32, nr_segs, tp->reordering + 1); |
459 | |
460 | /* Fast Recovery (RFC 5681 3.2) : |
461 | * Cubic needs 1.7 factor, rounded to 2 to include |
462 | * extra cushion (application might react slowly to EPOLLOUT) |
463 | */ |
464 | sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2; |
465 | sndmem *= nr_segs * per_mss; |
466 | |
467 | if (sk->sk_sndbuf < sndmem) |
468 | WRITE_ONCE(sk->sk_sndbuf, |
469 | min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2]))); |
470 | } |
471 | |
472 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) |
473 | * |
474 | * All tcp_full_space() is split to two parts: "network" buffer, allocated |
475 | * forward and advertised in receiver window (tp->rcv_wnd) and |
476 | * "application buffer", required to isolate scheduling/application |
477 | * latencies from network. |
478 | * window_clamp is maximal advertised window. It can be less than |
479 | * tcp_full_space(), in this case tcp_full_space() - window_clamp |
480 | * is reserved for "application" buffer. The less window_clamp is |
481 | * the smoother our behaviour from viewpoint of network, but the lower |
482 | * throughput and the higher sensitivity of the connection to losses. 8) |
483 | * |
484 | * rcv_ssthresh is more strict window_clamp used at "slow start" |
485 | * phase to predict further behaviour of this connection. |
486 | * It is used for two goals: |
487 | * - to enforce header prediction at sender, even when application |
488 | * requires some significant "application buffer". It is check #1. |
489 | * - to prevent pruning of receive queue because of misprediction |
490 | * of receiver window. Check #2. |
491 | * |
492 | * The scheme does not work when sender sends good segments opening |
493 | * window and then starts to feed us spaghetti. But it should work |
494 | * in common situations. Otherwise, we have to rely on queue collapsing. |
495 | */ |
496 | |
497 | /* Slow part of check#2. */ |
498 | static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb, |
499 | unsigned int skbtruesize) |
500 | { |
501 | const struct tcp_sock *tp = tcp_sk(sk); |
502 | /* Optimize this! */ |
503 | int truesize = tcp_win_from_space(sk, space: skbtruesize) >> 1; |
504 | int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1; |
505 | |
506 | while (tp->rcv_ssthresh <= window) { |
507 | if (truesize <= skb->len) |
508 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; |
509 | |
510 | truesize >>= 1; |
511 | window >>= 1; |
512 | } |
513 | return 0; |
514 | } |
515 | |
516 | /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing |
517 | * can play nice with us, as sk_buff and skb->head might be either |
518 | * freed or shared with up to MAX_SKB_FRAGS segments. |
519 | * Only give a boost to drivers using page frag(s) to hold the frame(s), |
520 | * and if no payload was pulled in skb->head before reaching us. |
521 | */ |
522 | static u32 truesize_adjust(bool adjust, const struct sk_buff *skb) |
523 | { |
524 | u32 truesize = skb->truesize; |
525 | |
526 | if (adjust && !skb_headlen(skb)) { |
527 | truesize -= SKB_TRUESIZE(skb_end_offset(skb)); |
528 | /* paranoid check, some drivers might be buggy */ |
529 | if (unlikely((int)truesize < (int)skb->len)) |
530 | truesize = skb->truesize; |
531 | } |
532 | return truesize; |
533 | } |
534 | |
535 | static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb, |
536 | bool adjust) |
537 | { |
538 | struct tcp_sock *tp = tcp_sk(sk); |
539 | int room; |
540 | |
541 | room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh; |
542 | |
543 | if (room <= 0) |
544 | return; |
545 | |
546 | /* Check #1 */ |
547 | if (!tcp_under_memory_pressure(sk)) { |
548 | unsigned int truesize = truesize_adjust(adjust, skb); |
549 | int incr; |
550 | |
551 | /* Check #2. Increase window, if skb with such overhead |
552 | * will fit to rcvbuf in future. |
553 | */ |
554 | if (tcp_win_from_space(sk, space: truesize) <= skb->len) |
555 | incr = 2 * tp->advmss; |
556 | else |
557 | incr = __tcp_grow_window(sk, skb, skbtruesize: truesize); |
558 | |
559 | if (incr) { |
560 | incr = max_t(int, incr, 2 * skb->len); |
561 | tp->rcv_ssthresh += min(room, incr); |
562 | inet_csk(sk)->icsk_ack.quick |= 1; |
563 | } |
564 | } else { |
565 | /* Under pressure: |
566 | * Adjust rcv_ssthresh according to reserved mem |
567 | */ |
568 | tcp_adjust_rcv_ssthresh(sk); |
569 | } |
570 | } |
571 | |
572 | /* 3. Try to fixup all. It is made immediately after connection enters |
573 | * established state. |
574 | */ |
575 | static void tcp_init_buffer_space(struct sock *sk) |
576 | { |
577 | int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win); |
578 | struct tcp_sock *tp = tcp_sk(sk); |
579 | int maxwin; |
580 | |
581 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) |
582 | tcp_sndbuf_expand(sk); |
583 | |
584 | tcp_mstamp_refresh(tp); |
585 | tp->rcvq_space.time = tp->tcp_mstamp; |
586 | tp->rcvq_space.seq = tp->copied_seq; |
587 | |
588 | maxwin = tcp_full_space(sk); |
589 | |
590 | if (tp->window_clamp >= maxwin) { |
591 | WRITE_ONCE(tp->window_clamp, maxwin); |
592 | |
593 | if (tcp_app_win && maxwin > 4 * tp->advmss) |
594 | WRITE_ONCE(tp->window_clamp, |
595 | max(maxwin - (maxwin >> tcp_app_win), |
596 | 4 * tp->advmss)); |
597 | } |
598 | |
599 | /* Force reservation of one segment. */ |
600 | if (tcp_app_win && |
601 | tp->window_clamp > 2 * tp->advmss && |
602 | tp->window_clamp + tp->advmss > maxwin) |
603 | WRITE_ONCE(tp->window_clamp, |
604 | max(2 * tp->advmss, maxwin - tp->advmss)); |
605 | |
606 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); |
607 | tp->snd_cwnd_stamp = tcp_jiffies32; |
608 | tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd, |
609 | (u32)TCP_INIT_CWND * tp->advmss); |
610 | } |
611 | |
612 | /* 4. Recalculate window clamp after socket hit its memory bounds. */ |
613 | static void tcp_clamp_window(struct sock *sk) |
614 | { |
615 | struct tcp_sock *tp = tcp_sk(sk); |
616 | struct inet_connection_sock *icsk = inet_csk(sk); |
617 | struct net *net = sock_net(sk); |
618 | int rmem2; |
619 | |
620 | icsk->icsk_ack.quick = 0; |
621 | rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]); |
622 | |
623 | if (sk->sk_rcvbuf < rmem2 && |
624 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && |
625 | !tcp_under_memory_pressure(sk) && |
626 | sk_memory_allocated(sk) < sk_prot_mem_limits(sk, index: 0)) { |
627 | WRITE_ONCE(sk->sk_rcvbuf, |
628 | min(atomic_read(&sk->sk_rmem_alloc), rmem2)); |
629 | } |
630 | if (atomic_read(v: &sk->sk_rmem_alloc) > sk->sk_rcvbuf) |
631 | tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); |
632 | } |
633 | |
634 | /* Initialize RCV_MSS value. |
635 | * RCV_MSS is an our guess about MSS used by the peer. |
636 | * We haven't any direct information about the MSS. |
637 | * It's better to underestimate the RCV_MSS rather than overestimate. |
638 | * Overestimations make us ACKing less frequently than needed. |
639 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). |
640 | */ |
641 | void tcp_initialize_rcv_mss(struct sock *sk) |
642 | { |
643 | const struct tcp_sock *tp = tcp_sk(sk); |
644 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); |
645 | |
646 | hint = min(hint, tp->rcv_wnd / 2); |
647 | hint = min(hint, TCP_MSS_DEFAULT); |
648 | hint = max(hint, TCP_MIN_MSS); |
649 | |
650 | inet_csk(sk)->icsk_ack.rcv_mss = hint; |
651 | } |
652 | EXPORT_IPV6_MOD(tcp_initialize_rcv_mss); |
653 | |
654 | /* Receiver "autotuning" code. |
655 | * |
656 | * The algorithm for RTT estimation w/o timestamps is based on |
657 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. |
658 | * <https://public.lanl.gov/radiant/pubs.html#DRS> |
659 | * |
660 | * More detail on this code can be found at |
661 | * <http://staff.psc.edu/jheffner/>, |
662 | * though this reference is out of date. A new paper |
663 | * is pending. |
664 | */ |
665 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) |
666 | { |
667 | u32 new_sample, old_sample = tp->rcv_rtt_est.rtt_us; |
668 | long m = sample << 3; |
669 | |
670 | if (old_sample == 0 || m < old_sample) { |
671 | new_sample = m; |
672 | } else { |
673 | /* If we sample in larger samples in the non-timestamp |
674 | * case, we could grossly overestimate the RTT especially |
675 | * with chatty applications or bulk transfer apps which |
676 | * are stalled on filesystem I/O. |
677 | * |
678 | * Also, since we are only going for a minimum in the |
679 | * non-timestamp case, we do not smooth things out |
680 | * else with timestamps disabled convergence takes too |
681 | * long. |
682 | */ |
683 | if (win_dep) |
684 | return; |
685 | /* Do not use this sample if receive queue is not empty. */ |
686 | if (tp->rcv_nxt != tp->copied_seq) |
687 | return; |
688 | new_sample = old_sample - (old_sample >> 3) + sample; |
689 | } |
690 | |
691 | tp->rcv_rtt_est.rtt_us = new_sample; |
692 | } |
693 | |
694 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) |
695 | { |
696 | u32 delta_us; |
697 | |
698 | if (tp->rcv_rtt_est.time == 0) |
699 | goto new_measure; |
700 | if (before(seq1: tp->rcv_nxt, seq2: tp->rcv_rtt_est.seq)) |
701 | return; |
702 | delta_us = tcp_stamp_us_delta(t1: tp->tcp_mstamp, t0: tp->rcv_rtt_est.time); |
703 | if (!delta_us) |
704 | delta_us = 1; |
705 | tcp_rcv_rtt_update(tp, sample: delta_us, win_dep: 1); |
706 | |
707 | new_measure: |
708 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; |
709 | tp->rcv_rtt_est.time = tp->tcp_mstamp; |
710 | } |
711 | |
712 | static s32 tcp_rtt_tsopt_us(const struct tcp_sock *tp, u32 min_delta) |
713 | { |
714 | u32 delta, delta_us; |
715 | |
716 | delta = tcp_time_stamp_ts(tp) - tp->rx_opt.rcv_tsecr; |
717 | if (tp->tcp_usec_ts) |
718 | return delta; |
719 | |
720 | if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) { |
721 | if (!delta) |
722 | delta = min_delta; |
723 | delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ); |
724 | return delta_us; |
725 | } |
726 | return -1; |
727 | } |
728 | |
729 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, |
730 | const struct sk_buff *skb) |
731 | { |
732 | struct tcp_sock *tp = tcp_sk(sk); |
733 | |
734 | if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr) |
735 | return; |
736 | tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; |
737 | |
738 | if (TCP_SKB_CB(skb)->end_seq - |
739 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) { |
740 | s32 delta = tcp_rtt_tsopt_us(tp, min_delta: 0); |
741 | |
742 | if (delta > 0) |
743 | tcp_rcv_rtt_update(tp, sample: delta, win_dep: 0); |
744 | } |
745 | } |
746 | |
747 | static void tcp_rcvbuf_grow(struct sock *sk) |
748 | { |
749 | const struct net *net = sock_net(sk); |
750 | struct tcp_sock *tp = tcp_sk(sk); |
751 | int rcvwin, rcvbuf, cap; |
752 | |
753 | if (!READ_ONCE(net->ipv4.sysctl_tcp_moderate_rcvbuf) || |
754 | (sk->sk_userlocks & SOCK_RCVBUF_LOCK)) |
755 | return; |
756 | |
757 | /* slow start: allow the sender to double its rate. */ |
758 | rcvwin = tp->rcvq_space.space << 1; |
759 | |
760 | if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) |
761 | rcvwin += TCP_SKB_CB(tp->ooo_last_skb)->end_seq - tp->rcv_nxt; |
762 | |
763 | cap = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]); |
764 | |
765 | rcvbuf = min_t(u32, tcp_space_from_win(sk, rcvwin), cap); |
766 | if (rcvbuf > sk->sk_rcvbuf) { |
767 | WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); |
768 | /* Make the window clamp follow along. */ |
769 | WRITE_ONCE(tp->window_clamp, |
770 | tcp_win_from_space(sk, rcvbuf)); |
771 | } |
772 | } |
773 | /* |
774 | * This function should be called every time data is copied to user space. |
775 | * It calculates the appropriate TCP receive buffer space. |
776 | */ |
777 | void tcp_rcv_space_adjust(struct sock *sk) |
778 | { |
779 | struct tcp_sock *tp = tcp_sk(sk); |
780 | int time, inq, copied; |
781 | |
782 | trace_tcp_rcv_space_adjust(sk); |
783 | |
784 | tcp_mstamp_refresh(tp); |
785 | time = tcp_stamp_us_delta(t1: tp->tcp_mstamp, t0: tp->rcvq_space.time); |
786 | if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0) |
787 | return; |
788 | |
789 | /* Number of bytes copied to user in last RTT */ |
790 | copied = tp->copied_seq - tp->rcvq_space.seq; |
791 | /* Number of bytes in receive queue. */ |
792 | inq = tp->rcv_nxt - tp->copied_seq; |
793 | copied -= inq; |
794 | if (copied <= tp->rcvq_space.space) |
795 | goto new_measure; |
796 | |
797 | trace_tcp_rcvbuf_grow(sk, time); |
798 | |
799 | tp->rcvq_space.space = copied; |
800 | |
801 | tcp_rcvbuf_grow(sk); |
802 | |
803 | new_measure: |
804 | tp->rcvq_space.seq = tp->copied_seq; |
805 | tp->rcvq_space.time = tp->tcp_mstamp; |
806 | } |
807 | |
808 | static void tcp_save_lrcv_flowlabel(struct sock *sk, const struct sk_buff *skb) |
809 | { |
810 | #if IS_ENABLED(CONFIG_IPV6) |
811 | struct inet_connection_sock *icsk = inet_csk(sk); |
812 | |
813 | if (skb->protocol == htons(ETH_P_IPV6)) |
814 | icsk->icsk_ack.lrcv_flowlabel = ntohl(ip6_flowlabel(ipv6_hdr(skb))); |
815 | #endif |
816 | } |
817 | |
818 | /* There is something which you must keep in mind when you analyze the |
819 | * behavior of the tp->ato delayed ack timeout interval. When a |
820 | * connection starts up, we want to ack as quickly as possible. The |
821 | * problem is that "good" TCP's do slow start at the beginning of data |
822 | * transmission. The means that until we send the first few ACK's the |
823 | * sender will sit on his end and only queue most of his data, because |
824 | * he can only send snd_cwnd unacked packets at any given time. For |
825 | * each ACK we send, he increments snd_cwnd and transmits more of his |
826 | * queue. -DaveM |
827 | */ |
828 | static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) |
829 | { |
830 | struct tcp_sock *tp = tcp_sk(sk); |
831 | struct inet_connection_sock *icsk = inet_csk(sk); |
832 | u32 now; |
833 | |
834 | inet_csk_schedule_ack(sk); |
835 | |
836 | tcp_measure_rcv_mss(sk, skb); |
837 | |
838 | tcp_rcv_rtt_measure(tp); |
839 | |
840 | now = tcp_jiffies32; |
841 | |
842 | if (!icsk->icsk_ack.ato) { |
843 | /* The _first_ data packet received, initialize |
844 | * delayed ACK engine. |
845 | */ |
846 | tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); |
847 | icsk->icsk_ack.ato = TCP_ATO_MIN; |
848 | } else { |
849 | int m = now - icsk->icsk_ack.lrcvtime; |
850 | |
851 | if (m <= TCP_ATO_MIN / 2) { |
852 | /* The fastest case is the first. */ |
853 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; |
854 | } else if (m < icsk->icsk_ack.ato) { |
855 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; |
856 | if (icsk->icsk_ack.ato > icsk->icsk_rto) |
857 | icsk->icsk_ack.ato = icsk->icsk_rto; |
858 | } else if (m > icsk->icsk_rto) { |
859 | /* Too long gap. Apparently sender failed to |
860 | * restart window, so that we send ACKs quickly. |
861 | */ |
862 | tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); |
863 | } |
864 | } |
865 | icsk->icsk_ack.lrcvtime = now; |
866 | tcp_save_lrcv_flowlabel(sk, skb); |
867 | |
868 | tcp_data_ecn_check(sk, skb); |
869 | |
870 | if (skb->len >= 128) |
871 | tcp_grow_window(sk, skb, adjust: true); |
872 | } |
873 | |
874 | /* Called to compute a smoothed rtt estimate. The data fed to this |
875 | * routine either comes from timestamps, or from segments that were |
876 | * known _not_ to have been retransmitted [see Karn/Partridge |
877 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 |
878 | * piece by Van Jacobson. |
879 | * NOTE: the next three routines used to be one big routine. |
880 | * To save cycles in the RFC 1323 implementation it was better to break |
881 | * it up into three procedures. -- erics |
882 | */ |
883 | static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) |
884 | { |
885 | struct tcp_sock *tp = tcp_sk(sk); |
886 | long m = mrtt_us; /* RTT */ |
887 | u32 srtt = tp->srtt_us; |
888 | |
889 | /* The following amusing code comes from Jacobson's |
890 | * article in SIGCOMM '88. Note that rtt and mdev |
891 | * are scaled versions of rtt and mean deviation. |
892 | * This is designed to be as fast as possible |
893 | * m stands for "measurement". |
894 | * |
895 | * On a 1990 paper the rto value is changed to: |
896 | * RTO = rtt + 4 * mdev |
897 | * |
898 | * Funny. This algorithm seems to be very broken. |
899 | * These formulae increase RTO, when it should be decreased, increase |
900 | * too slowly, when it should be increased quickly, decrease too quickly |
901 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely |
902 | * does not matter how to _calculate_ it. Seems, it was trap |
903 | * that VJ failed to avoid. 8) |
904 | */ |
905 | if (srtt != 0) { |
906 | m -= (srtt >> 3); /* m is now error in rtt est */ |
907 | srtt += m; /* rtt = 7/8 rtt + 1/8 new */ |
908 | if (m < 0) { |
909 | m = -m; /* m is now abs(error) */ |
910 | m -= (tp->mdev_us >> 2); /* similar update on mdev */ |
911 | /* This is similar to one of Eifel findings. |
912 | * Eifel blocks mdev updates when rtt decreases. |
913 | * This solution is a bit different: we use finer gain |
914 | * for mdev in this case (alpha*beta). |
915 | * Like Eifel it also prevents growth of rto, |
916 | * but also it limits too fast rto decreases, |
917 | * happening in pure Eifel. |
918 | */ |
919 | if (m > 0) |
920 | m >>= 3; |
921 | } else { |
922 | m -= (tp->mdev_us >> 2); /* similar update on mdev */ |
923 | } |
924 | tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ |
925 | if (tp->mdev_us > tp->mdev_max_us) { |
926 | tp->mdev_max_us = tp->mdev_us; |
927 | if (tp->mdev_max_us > tp->rttvar_us) |
928 | tp->rttvar_us = tp->mdev_max_us; |
929 | } |
930 | if (after(tp->snd_una, tp->rtt_seq)) { |
931 | if (tp->mdev_max_us < tp->rttvar_us) |
932 | tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; |
933 | tp->rtt_seq = tp->snd_nxt; |
934 | tp->mdev_max_us = tcp_rto_min_us(sk); |
935 | |
936 | tcp_bpf_rtt(sk, mrtt: mrtt_us, srtt); |
937 | } |
938 | } else { |
939 | /* no previous measure. */ |
940 | srtt = m << 3; /* take the measured time to be rtt */ |
941 | tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ |
942 | tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); |
943 | tp->mdev_max_us = tp->rttvar_us; |
944 | tp->rtt_seq = tp->snd_nxt; |
945 | |
946 | tcp_bpf_rtt(sk, mrtt: mrtt_us, srtt); |
947 | } |
948 | tp->srtt_us = max(1U, srtt); |
949 | } |
950 | |
951 | static void tcp_update_pacing_rate(struct sock *sk) |
952 | { |
953 | const struct tcp_sock *tp = tcp_sk(sk); |
954 | u64 rate; |
955 | |
956 | /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ |
957 | rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); |
958 | |
959 | /* current rate is (cwnd * mss) / srtt |
960 | * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. |
961 | * In Congestion Avoidance phase, set it to 120 % the current rate. |
962 | * |
963 | * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) |
964 | * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching |
965 | * end of slow start and should slow down. |
966 | */ |
967 | if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2) |
968 | rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio); |
969 | else |
970 | rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio); |
971 | |
972 | rate *= max(tcp_snd_cwnd(tp), tp->packets_out); |
973 | |
974 | if (likely(tp->srtt_us)) |
975 | do_div(rate, tp->srtt_us); |
976 | |
977 | /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate |
978 | * without any lock. We want to make sure compiler wont store |
979 | * intermediate values in this location. |
980 | */ |
981 | WRITE_ONCE(sk->sk_pacing_rate, |
982 | min_t(u64, rate, READ_ONCE(sk->sk_max_pacing_rate))); |
983 | } |
984 | |
985 | /* Calculate rto without backoff. This is the second half of Van Jacobson's |
986 | * routine referred to above. |
987 | */ |
988 | static void tcp_set_rto(struct sock *sk) |
989 | { |
990 | const struct tcp_sock *tp = tcp_sk(sk); |
991 | /* Old crap is replaced with new one. 8) |
992 | * |
993 | * More seriously: |
994 | * 1. If rtt variance happened to be less 50msec, it is hallucination. |
995 | * It cannot be less due to utterly erratic ACK generation made |
996 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ |
997 | * to do with delayed acks, because at cwnd>2 true delack timeout |
998 | * is invisible. Actually, Linux-2.4 also generates erratic |
999 | * ACKs in some circumstances. |
1000 | */ |
1001 | inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); |
1002 | |
1003 | /* 2. Fixups made earlier cannot be right. |
1004 | * If we do not estimate RTO correctly without them, |
1005 | * all the algo is pure shit and should be replaced |
1006 | * with correct one. It is exactly, which we pretend to do. |
1007 | */ |
1008 | |
1009 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo |
1010 | * guarantees that rto is higher. |
1011 | */ |
1012 | tcp_bound_rto(sk); |
1013 | } |
1014 | |
1015 | __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) |
1016 | { |
1017 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); |
1018 | |
1019 | if (!cwnd) |
1020 | cwnd = TCP_INIT_CWND; |
1021 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); |
1022 | } |
1023 | |
1024 | struct tcp_sacktag_state { |
1025 | /* Timestamps for earliest and latest never-retransmitted segment |
1026 | * that was SACKed. RTO needs the earliest RTT to stay conservative, |
1027 | * but congestion control should still get an accurate delay signal. |
1028 | */ |
1029 | u64 first_sackt; |
1030 | u64 last_sackt; |
1031 | u32 reord; |
1032 | u32 sack_delivered; |
1033 | int flag; |
1034 | unsigned int mss_now; |
1035 | struct rate_sample *rate; |
1036 | }; |
1037 | |
1038 | /* Take a notice that peer is sending D-SACKs. Skip update of data delivery |
1039 | * and spurious retransmission information if this DSACK is unlikely caused by |
1040 | * sender's action: |
1041 | * - DSACKed sequence range is larger than maximum receiver's window. |
1042 | * - Total no. of DSACKed segments exceed the total no. of retransmitted segs. |
1043 | */ |
1044 | static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq, |
1045 | u32 end_seq, struct tcp_sacktag_state *state) |
1046 | { |
1047 | u32 seq_len, dup_segs = 1; |
1048 | |
1049 | if (!before(seq1: start_seq, seq2: end_seq)) |
1050 | return 0; |
1051 | |
1052 | seq_len = end_seq - start_seq; |
1053 | /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */ |
1054 | if (seq_len > tp->max_window) |
1055 | return 0; |
1056 | if (seq_len > tp->mss_cache) |
1057 | dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache); |
1058 | else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq) |
1059 | state->flag |= FLAG_DSACK_TLP; |
1060 | |
1061 | tp->dsack_dups += dup_segs; |
1062 | /* Skip the DSACK if dup segs weren't retransmitted by sender */ |
1063 | if (tp->dsack_dups > tp->total_retrans) |
1064 | return 0; |
1065 | |
1066 | tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; |
1067 | /* We increase the RACK ordering window in rounds where we receive |
1068 | * DSACKs that may have been due to reordering causing RACK to trigger |
1069 | * a spurious fast recovery. Thus RACK ignores DSACKs that happen |
1070 | * without having seen reordering, or that match TLP probes (TLP |
1071 | * is timer-driven, not triggered by RACK). |
1072 | */ |
1073 | if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP)) |
1074 | tp->rack.dsack_seen = 1; |
1075 | |
1076 | state->flag |= FLAG_DSACKING_ACK; |
1077 | /* A spurious retransmission is delivered */ |
1078 | state->sack_delivered += dup_segs; |
1079 | |
1080 | return dup_segs; |
1081 | } |
1082 | |
1083 | /* It's reordering when higher sequence was delivered (i.e. sacked) before |
1084 | * some lower never-retransmitted sequence ("low_seq"). The maximum reordering |
1085 | * distance is approximated in full-mss packet distance ("reordering"). |
1086 | */ |
1087 | static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq, |
1088 | const int ts) |
1089 | { |
1090 | struct tcp_sock *tp = tcp_sk(sk); |
1091 | const u32 mss = tp->mss_cache; |
1092 | u32 fack, metric; |
1093 | |
1094 | fack = tcp_highest_sack_seq(tp); |
1095 | if (!before(seq1: low_seq, seq2: fack)) |
1096 | return; |
1097 | |
1098 | metric = fack - low_seq; |
1099 | if ((metric > tp->reordering * mss) && mss) { |
1100 | #if FASTRETRANS_DEBUG > 1 |
1101 | pr_debug("Disorder%d %d %u f%u s%u rr%d\n", |
1102 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, |
1103 | tp->reordering, |
1104 | 0, |
1105 | tp->sacked_out, |
1106 | tp->undo_marker ? tp->undo_retrans : 0); |
1107 | #endif |
1108 | tp->reordering = min_t(u32, (metric + mss - 1) / mss, |
1109 | READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)); |
1110 | } |
1111 | |
1112 | /* This exciting event is worth to be remembered. 8) */ |
1113 | tp->reord_seen++; |
1114 | NET_INC_STATS(sock_net(sk), |
1115 | ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER); |
1116 | } |
1117 | |
1118 | /* This must be called before lost_out or retrans_out are updated |
1119 | * on a new loss, because we want to know if all skbs previously |
1120 | * known to be lost have already been retransmitted, indicating |
1121 | * that this newly lost skb is our next skb to retransmit. |
1122 | */ |
1123 | static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) |
1124 | { |
1125 | if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) || |
1126 | (tp->retransmit_skb_hint && |
1127 | before(TCP_SKB_CB(skb)->seq, |
1128 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq))) |
1129 | tp->retransmit_skb_hint = skb; |
1130 | } |
1131 | |
1132 | /* Sum the number of packets on the wire we have marked as lost, and |
1133 | * notify the congestion control module that the given skb was marked lost. |
1134 | */ |
1135 | static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb) |
1136 | { |
1137 | tp->lost += tcp_skb_pcount(skb); |
1138 | } |
1139 | |
1140 | void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb) |
1141 | { |
1142 | __u8 sacked = TCP_SKB_CB(skb)->sacked; |
1143 | struct tcp_sock *tp = tcp_sk(sk); |
1144 | |
1145 | if (sacked & TCPCB_SACKED_ACKED) |
1146 | return; |
1147 | |
1148 | tcp_verify_retransmit_hint(tp, skb); |
1149 | if (sacked & TCPCB_LOST) { |
1150 | if (sacked & TCPCB_SACKED_RETRANS) { |
1151 | /* Account for retransmits that are lost again */ |
1152 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
1153 | tp->retrans_out -= tcp_skb_pcount(skb); |
1154 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT, |
1155 | tcp_skb_pcount(skb)); |
1156 | tcp_notify_skb_loss_event(tp, skb); |
1157 | } |
1158 | } else { |
1159 | tp->lost_out += tcp_skb_pcount(skb); |
1160 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
1161 | tcp_notify_skb_loss_event(tp, skb); |
1162 | } |
1163 | } |
1164 | |
1165 | /* This procedure tags the retransmission queue when SACKs arrive. |
1166 | * |
1167 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). |
1168 | * Packets in queue with these bits set are counted in variables |
1169 | * sacked_out, retrans_out and lost_out, correspondingly. |
1170 | * |
1171 | * Valid combinations are: |
1172 | * Tag InFlight Description |
1173 | * 0 1 - orig segment is in flight. |
1174 | * S 0 - nothing flies, orig reached receiver. |
1175 | * L 0 - nothing flies, orig lost by net. |
1176 | * R 2 - both orig and retransmit are in flight. |
1177 | * L|R 1 - orig is lost, retransmit is in flight. |
1178 | * S|R 1 - orig reached receiver, retrans is still in flight. |
1179 | * (L|S|R is logically valid, it could occur when L|R is sacked, |
1180 | * but it is equivalent to plain S and code short-circuits it to S. |
1181 | * L|S is logically invalid, it would mean -1 packet in flight 8)) |
1182 | * |
1183 | * These 6 states form finite state machine, controlled by the following events: |
1184 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) |
1185 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) |
1186 | * 3. Loss detection event of two flavors: |
1187 | * A. Scoreboard estimator decided the packet is lost. |
1188 | * A'. Reno "three dupacks" marks head of queue lost. |
1189 | * B. SACK arrives sacking SND.NXT at the moment, when the |
1190 | * segment was retransmitted. |
1191 | * 4. D-SACK added new rule: D-SACK changes any tag to S. |
1192 | * |
1193 | * It is pleasant to note, that state diagram turns out to be commutative, |
1194 | * so that we are allowed not to be bothered by order of our actions, |
1195 | * when multiple events arrive simultaneously. (see the function below). |
1196 | * |
1197 | * Reordering detection. |
1198 | * -------------------- |
1199 | * Reordering metric is maximal distance, which a packet can be displaced |
1200 | * in packet stream. With SACKs we can estimate it: |
1201 | * |
1202 | * 1. SACK fills old hole and the corresponding segment was not |
1203 | * ever retransmitted -> reordering. Alas, we cannot use it |
1204 | * when segment was retransmitted. |
1205 | * 2. The last flaw is solved with D-SACK. D-SACK arrives |
1206 | * for retransmitted and already SACKed segment -> reordering.. |
1207 | * Both of these heuristics are not used in Loss state, when we cannot |
1208 | * account for retransmits accurately. |
1209 | * |
1210 | * SACK block validation. |
1211 | * ---------------------- |
1212 | * |
1213 | * SACK block range validation checks that the received SACK block fits to |
1214 | * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. |
1215 | * Note that SND.UNA is not included to the range though being valid because |
1216 | * it means that the receiver is rather inconsistent with itself reporting |
1217 | * SACK reneging when it should advance SND.UNA. Such SACK block this is |
1218 | * perfectly valid, however, in light of RFC2018 which explicitly states |
1219 | * that "SACK block MUST reflect the newest segment. Even if the newest |
1220 | * segment is going to be discarded ...", not that it looks very clever |
1221 | * in case of head skb. Due to potentional receiver driven attacks, we |
1222 | * choose to avoid immediate execution of a walk in write queue due to |
1223 | * reneging and defer head skb's loss recovery to standard loss recovery |
1224 | * procedure that will eventually trigger (nothing forbids us doing this). |
1225 | * |
1226 | * Implements also blockage to start_seq wrap-around. Problem lies in the |
1227 | * fact that though start_seq (s) is before end_seq (i.e., not reversed), |
1228 | * there's no guarantee that it will be before snd_nxt (n). The problem |
1229 | * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt |
1230 | * wrap (s_w): |
1231 | * |
1232 | * <- outs wnd -> <- wrapzone -> |
1233 | * u e n u_w e_w s n_w |
1234 | * | | | | | | | |
1235 | * |<------------+------+----- TCP seqno space --------------+---------->| |
1236 | * ...-- <2^31 ->| |<--------... |
1237 | * ...---- >2^31 ------>| |<--------... |
1238 | * |
1239 | * Current code wouldn't be vulnerable but it's better still to discard such |
1240 | * crazy SACK blocks. Doing this check for start_seq alone closes somewhat |
1241 | * similar case (end_seq after snd_nxt wrap) as earlier reversed check in |
1242 | * snd_nxt wrap -> snd_una region will then become "well defined", i.e., |
1243 | * equal to the ideal case (infinite seqno space without wrap caused issues). |
1244 | * |
1245 | * With D-SACK the lower bound is extended to cover sequence space below |
1246 | * SND.UNA down to undo_marker, which is the last point of interest. Yet |
1247 | * again, D-SACK block must not to go across snd_una (for the same reason as |
1248 | * for the normal SACK blocks, explained above). But there all simplicity |
1249 | * ends, TCP might receive valid D-SACKs below that. As long as they reside |
1250 | * fully below undo_marker they do not affect behavior in anyway and can |
1251 | * therefore be safely ignored. In rare cases (which are more or less |
1252 | * theoretical ones), the D-SACK will nicely cross that boundary due to skb |
1253 | * fragmentation and packet reordering past skb's retransmission. To consider |
1254 | * them correctly, the acceptable range must be extended even more though |
1255 | * the exact amount is rather hard to quantify. However, tp->max_window can |
1256 | * be used as an exaggerated estimate. |
1257 | */ |
1258 | static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, |
1259 | u32 start_seq, u32 end_seq) |
1260 | { |
1261 | /* Too far in future, or reversed (interpretation is ambiguous) */ |
1262 | if (after(end_seq, tp->snd_nxt) || !before(seq1: start_seq, seq2: end_seq)) |
1263 | return false; |
1264 | |
1265 | /* Nasty start_seq wrap-around check (see comments above) */ |
1266 | if (!before(seq1: start_seq, seq2: tp->snd_nxt)) |
1267 | return false; |
1268 | |
1269 | /* In outstanding window? ...This is valid exit for D-SACKs too. |
1270 | * start_seq == snd_una is non-sensical (see comments above) |
1271 | */ |
1272 | if (after(start_seq, tp->snd_una)) |
1273 | return true; |
1274 | |
1275 | if (!is_dsack || !tp->undo_marker) |
1276 | return false; |
1277 | |
1278 | /* ...Then it's D-SACK, and must reside below snd_una completely */ |
1279 | if (after(end_seq, tp->snd_una)) |
1280 | return false; |
1281 | |
1282 | if (!before(seq1: start_seq, seq2: tp->undo_marker)) |
1283 | return true; |
1284 | |
1285 | /* Too old */ |
1286 | if (!after(end_seq, tp->undo_marker)) |
1287 | return false; |
1288 | |
1289 | /* Undo_marker boundary crossing (overestimates a lot). Known already: |
1290 | * start_seq < undo_marker and end_seq >= undo_marker. |
1291 | */ |
1292 | return !before(seq1: start_seq, seq2: end_seq - tp->max_window); |
1293 | } |
1294 | |
1295 | static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, |
1296 | struct tcp_sack_block_wire *sp, int num_sacks, |
1297 | u32 prior_snd_una, struct tcp_sacktag_state *state) |
1298 | { |
1299 | struct tcp_sock *tp = tcp_sk(sk); |
1300 | u32 start_seq_0 = get_unaligned_be32(p: &sp[0].start_seq); |
1301 | u32 end_seq_0 = get_unaligned_be32(p: &sp[0].end_seq); |
1302 | u32 dup_segs; |
1303 | |
1304 | if (before(seq1: start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { |
1305 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV); |
1306 | } else if (num_sacks > 1) { |
1307 | u32 end_seq_1 = get_unaligned_be32(p: &sp[1].end_seq); |
1308 | u32 start_seq_1 = get_unaligned_be32(p: &sp[1].start_seq); |
1309 | |
1310 | if (after(end_seq_0, end_seq_1) || before(seq1: start_seq_0, seq2: start_seq_1)) |
1311 | return false; |
1312 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV); |
1313 | } else { |
1314 | return false; |
1315 | } |
1316 | |
1317 | dup_segs = tcp_dsack_seen(tp, start_seq: start_seq_0, end_seq: end_seq_0, state); |
1318 | if (!dup_segs) { /* Skip dubious DSACK */ |
1319 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS); |
1320 | return false; |
1321 | } |
1322 | |
1323 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs); |
1324 | |
1325 | /* D-SACK for already forgotten data... Do dumb counting. */ |
1326 | if (tp->undo_marker && tp->undo_retrans > 0 && |
1327 | !after(end_seq_0, prior_snd_una) && |
1328 | after(end_seq_0, tp->undo_marker)) |
1329 | tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs); |
1330 | |
1331 | return true; |
1332 | } |
1333 | |
1334 | /* Check if skb is fully within the SACK block. In presence of GSO skbs, |
1335 | * the incoming SACK may not exactly match but we can find smaller MSS |
1336 | * aligned portion of it that matches. Therefore we might need to fragment |
1337 | * which may fail and creates some hassle (caller must handle error case |
1338 | * returns). |
1339 | * |
1340 | * FIXME: this could be merged to shift decision code |
1341 | */ |
1342 | static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, |
1343 | u32 start_seq, u32 end_seq) |
1344 | { |
1345 | int err; |
1346 | bool in_sack; |
1347 | unsigned int pkt_len; |
1348 | unsigned int mss; |
1349 | |
1350 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && |
1351 | !before(seq1: end_seq, TCP_SKB_CB(skb)->end_seq); |
1352 | |
1353 | if (tcp_skb_pcount(skb) > 1 && !in_sack && |
1354 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { |
1355 | mss = tcp_skb_mss(skb); |
1356 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); |
1357 | |
1358 | if (!in_sack) { |
1359 | pkt_len = start_seq - TCP_SKB_CB(skb)->seq; |
1360 | if (pkt_len < mss) |
1361 | pkt_len = mss; |
1362 | } else { |
1363 | pkt_len = end_seq - TCP_SKB_CB(skb)->seq; |
1364 | if (pkt_len < mss) |
1365 | return -EINVAL; |
1366 | } |
1367 | |
1368 | /* Round if necessary so that SACKs cover only full MSSes |
1369 | * and/or the remaining small portion (if present) |
1370 | */ |
1371 | if (pkt_len > mss) { |
1372 | unsigned int new_len = (pkt_len / mss) * mss; |
1373 | if (!in_sack && new_len < pkt_len) |
1374 | new_len += mss; |
1375 | pkt_len = new_len; |
1376 | } |
1377 | |
1378 | if (pkt_len >= skb->len && !in_sack) |
1379 | return 0; |
1380 | |
1381 | err = tcp_fragment(sk, tcp_queue: TCP_FRAG_IN_RTX_QUEUE, skb, |
1382 | len: pkt_len, mss_now: mss, GFP_ATOMIC); |
1383 | if (err < 0) |
1384 | return err; |
1385 | } |
1386 | |
1387 | return in_sack; |
1388 | } |
1389 | |
1390 | /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ |
1391 | static u8 tcp_sacktag_one(struct sock *sk, |
1392 | struct tcp_sacktag_state *state, u8 sacked, |
1393 | u32 start_seq, u32 end_seq, |
1394 | int dup_sack, int pcount, |
1395 | u64 xmit_time) |
1396 | { |
1397 | struct tcp_sock *tp = tcp_sk(sk); |
1398 | |
1399 | /* Account D-SACK for retransmitted packet. */ |
1400 | if (dup_sack && (sacked & TCPCB_RETRANS)) { |
1401 | if (tp->undo_marker && tp->undo_retrans > 0 && |
1402 | after(end_seq, tp->undo_marker)) |
1403 | tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount); |
1404 | if ((sacked & TCPCB_SACKED_ACKED) && |
1405 | before(seq1: start_seq, seq2: state->reord)) |
1406 | state->reord = start_seq; |
1407 | } |
1408 | |
1409 | /* Nothing to do; acked frame is about to be dropped (was ACKed). */ |
1410 | if (!after(end_seq, tp->snd_una)) |
1411 | return sacked; |
1412 | |
1413 | if (!(sacked & TCPCB_SACKED_ACKED)) { |
1414 | tcp_rack_advance(tp, sacked, end_seq, xmit_time); |
1415 | |
1416 | if (sacked & TCPCB_SACKED_RETRANS) { |
1417 | /* If the segment is not tagged as lost, |
1418 | * we do not clear RETRANS, believing |
1419 | * that retransmission is still in flight. |
1420 | */ |
1421 | if (sacked & TCPCB_LOST) { |
1422 | sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); |
1423 | tp->lost_out -= pcount; |
1424 | tp->retrans_out -= pcount; |
1425 | } |
1426 | } else { |
1427 | if (!(sacked & TCPCB_RETRANS)) { |
1428 | /* New sack for not retransmitted frame, |
1429 | * which was in hole. It is reordering. |
1430 | */ |
1431 | if (before(seq1: start_seq, |
1432 | seq2: tcp_highest_sack_seq(tp)) && |
1433 | before(seq1: start_seq, seq2: state->reord)) |
1434 | state->reord = start_seq; |
1435 | |
1436 | if (!after(end_seq, tp->high_seq)) |
1437 | state->flag |= FLAG_ORIG_SACK_ACKED; |
1438 | if (state->first_sackt == 0) |
1439 | state->first_sackt = xmit_time; |
1440 | state->last_sackt = xmit_time; |
1441 | } |
1442 | |
1443 | if (sacked & TCPCB_LOST) { |
1444 | sacked &= ~TCPCB_LOST; |
1445 | tp->lost_out -= pcount; |
1446 | } |
1447 | } |
1448 | |
1449 | sacked |= TCPCB_SACKED_ACKED; |
1450 | state->flag |= FLAG_DATA_SACKED; |
1451 | tp->sacked_out += pcount; |
1452 | /* Out-of-order packets delivered */ |
1453 | state->sack_delivered += pcount; |
1454 | |
1455 | /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ |
1456 | if (tp->lost_skb_hint && |
1457 | before(seq1: start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) |
1458 | tp->lost_cnt_hint += pcount; |
1459 | } |
1460 | |
1461 | /* D-SACK. We can detect redundant retransmission in S|R and plain R |
1462 | * frames and clear it. undo_retrans is decreased above, L|R frames |
1463 | * are accounted above as well. |
1464 | */ |
1465 | if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { |
1466 | sacked &= ~TCPCB_SACKED_RETRANS; |
1467 | tp->retrans_out -= pcount; |
1468 | } |
1469 | |
1470 | return sacked; |
1471 | } |
1472 | |
1473 | /* Shift newly-SACKed bytes from this skb to the immediately previous |
1474 | * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. |
1475 | */ |
1476 | static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev, |
1477 | struct sk_buff *skb, |
1478 | struct tcp_sacktag_state *state, |
1479 | unsigned int pcount, int shifted, int mss, |
1480 | bool dup_sack) |
1481 | { |
1482 | struct tcp_sock *tp = tcp_sk(sk); |
1483 | u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ |
1484 | u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ |
1485 | |
1486 | BUG_ON(!pcount); |
1487 | |
1488 | /* Adjust counters and hints for the newly sacked sequence |
1489 | * range but discard the return value since prev is already |
1490 | * marked. We must tag the range first because the seq |
1491 | * advancement below implicitly advances |
1492 | * tcp_highest_sack_seq() when skb is highest_sack. |
1493 | */ |
1494 | tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, |
1495 | start_seq, end_seq, dup_sack, pcount, |
1496 | xmit_time: tcp_skb_timestamp_us(skb)); |
1497 | tcp_rate_skb_delivered(sk, skb, rs: state->rate); |
1498 | |
1499 | if (skb == tp->lost_skb_hint) |
1500 | tp->lost_cnt_hint += pcount; |
1501 | |
1502 | TCP_SKB_CB(prev)->end_seq += shifted; |
1503 | TCP_SKB_CB(skb)->seq += shifted; |
1504 | |
1505 | tcp_skb_pcount_add(skb: prev, segs: pcount); |
1506 | WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount); |
1507 | tcp_skb_pcount_add(skb, segs: -pcount); |
1508 | |
1509 | /* When we're adding to gso_segs == 1, gso_size will be zero, |
1510 | * in theory this shouldn't be necessary but as long as DSACK |
1511 | * code can come after this skb later on it's better to keep |
1512 | * setting gso_size to something. |
1513 | */ |
1514 | if (!TCP_SKB_CB(prev)->tcp_gso_size) |
1515 | TCP_SKB_CB(prev)->tcp_gso_size = mss; |
1516 | |
1517 | /* CHECKME: To clear or not to clear? Mimics normal skb currently */ |
1518 | if (tcp_skb_pcount(skb) <= 1) |
1519 | TCP_SKB_CB(skb)->tcp_gso_size = 0; |
1520 | |
1521 | /* Difference in this won't matter, both ACKed by the same cumul. ACK */ |
1522 | TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); |
1523 | |
1524 | if (skb->len > 0) { |
1525 | BUG_ON(!tcp_skb_pcount(skb)); |
1526 | NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED); |
1527 | return false; |
1528 | } |
1529 | |
1530 | /* Whole SKB was eaten :-) */ |
1531 | |
1532 | if (skb == tp->retransmit_skb_hint) |
1533 | tp->retransmit_skb_hint = prev; |
1534 | if (skb == tp->lost_skb_hint) { |
1535 | tp->lost_skb_hint = prev; |
1536 | tp->lost_cnt_hint -= tcp_skb_pcount(skb: prev); |
1537 | } |
1538 | |
1539 | TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
1540 | TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor; |
1541 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) |
1542 | TCP_SKB_CB(prev)->end_seq++; |
1543 | |
1544 | if (skb == tcp_highest_sack(sk)) |
1545 | tcp_advance_highest_sack(sk, skb); |
1546 | |
1547 | tcp_skb_collapse_tstamp(skb: prev, next_skb: skb); |
1548 | if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp)) |
1549 | TCP_SKB_CB(prev)->tx.delivered_mstamp = 0; |
1550 | |
1551 | tcp_rtx_queue_unlink_and_free(skb, sk); |
1552 | |
1553 | NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED); |
1554 | |
1555 | return true; |
1556 | } |
1557 | |
1558 | /* I wish gso_size would have a bit more sane initialization than |
1559 | * something-or-zero which complicates things |
1560 | */ |
1561 | static int tcp_skb_seglen(const struct sk_buff *skb) |
1562 | { |
1563 | return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); |
1564 | } |
1565 | |
1566 | /* Shifting pages past head area doesn't work */ |
1567 | static int skb_can_shift(const struct sk_buff *skb) |
1568 | { |
1569 | return !skb_headlen(skb) && skb_is_nonlinear(skb); |
1570 | } |
1571 | |
1572 | int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from, |
1573 | int pcount, int shiftlen) |
1574 | { |
1575 | /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE) |
1576 | * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need |
1577 | * to make sure not storing more than 65535 * 8 bytes per skb, |
1578 | * even if current MSS is bigger. |
1579 | */ |
1580 | if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE)) |
1581 | return 0; |
1582 | if (unlikely(tcp_skb_pcount(to) + pcount > 65535)) |
1583 | return 0; |
1584 | return skb_shift(tgt: to, skb: from, shiftlen); |
1585 | } |
1586 | |
1587 | /* Try collapsing SACK blocks spanning across multiple skbs to a single |
1588 | * skb. |
1589 | */ |
1590 | static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, |
1591 | struct tcp_sacktag_state *state, |
1592 | u32 start_seq, u32 end_seq, |
1593 | bool dup_sack) |
1594 | { |
1595 | struct tcp_sock *tp = tcp_sk(sk); |
1596 | struct sk_buff *prev; |
1597 | int mss; |
1598 | int pcount = 0; |
1599 | int len; |
1600 | int in_sack; |
1601 | |
1602 | /* Normally R but no L won't result in plain S */ |
1603 | if (!dup_sack && |
1604 | (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) |
1605 | goto fallback; |
1606 | if (!skb_can_shift(skb)) |
1607 | goto fallback; |
1608 | /* This frame is about to be dropped (was ACKed). */ |
1609 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) |
1610 | goto fallback; |
1611 | |
1612 | /* Can only happen with delayed DSACK + discard craziness */ |
1613 | prev = skb_rb_prev(skb); |
1614 | if (!prev) |
1615 | goto fallback; |
1616 | |
1617 | if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) |
1618 | goto fallback; |
1619 | |
1620 | if (!tcp_skb_can_collapse(to: prev, from: skb)) |
1621 | goto fallback; |
1622 | |
1623 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && |
1624 | !before(seq1: end_seq, TCP_SKB_CB(skb)->end_seq); |
1625 | |
1626 | if (in_sack) { |
1627 | len = skb->len; |
1628 | pcount = tcp_skb_pcount(skb); |
1629 | mss = tcp_skb_seglen(skb); |
1630 | |
1631 | /* TODO: Fix DSACKs to not fragment already SACKed and we can |
1632 | * drop this restriction as unnecessary |
1633 | */ |
1634 | if (mss != tcp_skb_seglen(skb: prev)) |
1635 | goto fallback; |
1636 | } else { |
1637 | if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) |
1638 | goto noop; |
1639 | /* CHECKME: This is non-MSS split case only?, this will |
1640 | * cause skipped skbs due to advancing loop btw, original |
1641 | * has that feature too |
1642 | */ |
1643 | if (tcp_skb_pcount(skb) <= 1) |
1644 | goto noop; |
1645 | |
1646 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); |
1647 | if (!in_sack) { |
1648 | /* TODO: head merge to next could be attempted here |
1649 | * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), |
1650 | * though it might not be worth of the additional hassle |
1651 | * |
1652 | * ...we can probably just fallback to what was done |
1653 | * previously. We could try merging non-SACKed ones |
1654 | * as well but it probably isn't going to buy off |
1655 | * because later SACKs might again split them, and |
1656 | * it would make skb timestamp tracking considerably |
1657 | * harder problem. |
1658 | */ |
1659 | goto fallback; |
1660 | } |
1661 | |
1662 | len = end_seq - TCP_SKB_CB(skb)->seq; |
1663 | BUG_ON(len < 0); |
1664 | BUG_ON(len > skb->len); |
1665 | |
1666 | /* MSS boundaries should be honoured or else pcount will |
1667 | * severely break even though it makes things bit trickier. |
1668 | * Optimize common case to avoid most of the divides |
1669 | */ |
1670 | mss = tcp_skb_mss(skb); |
1671 | |
1672 | /* TODO: Fix DSACKs to not fragment already SACKed and we can |
1673 | * drop this restriction as unnecessary |
1674 | */ |
1675 | if (mss != tcp_skb_seglen(skb: prev)) |
1676 | goto fallback; |
1677 | |
1678 | if (len == mss) { |
1679 | pcount = 1; |
1680 | } else if (len < mss) { |
1681 | goto noop; |
1682 | } else { |
1683 | pcount = len / mss; |
1684 | len = pcount * mss; |
1685 | } |
1686 | } |
1687 | |
1688 | /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ |
1689 | if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) |
1690 | goto fallback; |
1691 | |
1692 | if (!tcp_skb_shift(to: prev, from: skb, pcount, shiftlen: len)) |
1693 | goto fallback; |
1694 | if (!tcp_shifted_skb(sk, prev, skb, state, pcount, shifted: len, mss, dup_sack)) |
1695 | goto out; |
1696 | |
1697 | /* Hole filled allows collapsing with the next as well, this is very |
1698 | * useful when hole on every nth skb pattern happens |
1699 | */ |
1700 | skb = skb_rb_next(prev); |
1701 | if (!skb) |
1702 | goto out; |
1703 | |
1704 | if (!skb_can_shift(skb) || |
1705 | ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || |
1706 | (mss != tcp_skb_seglen(skb))) |
1707 | goto out; |
1708 | |
1709 | if (!tcp_skb_can_collapse(to: prev, from: skb)) |
1710 | goto out; |
1711 | len = skb->len; |
1712 | pcount = tcp_skb_pcount(skb); |
1713 | if (tcp_skb_shift(to: prev, from: skb, pcount, shiftlen: len)) |
1714 | tcp_shifted_skb(sk, prev, skb, state, pcount, |
1715 | shifted: len, mss, dup_sack: 0); |
1716 | |
1717 | out: |
1718 | return prev; |
1719 | |
1720 | noop: |
1721 | return skb; |
1722 | |
1723 | fallback: |
1724 | NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); |
1725 | return NULL; |
1726 | } |
1727 | |
1728 | static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, |
1729 | struct tcp_sack_block *next_dup, |
1730 | struct tcp_sacktag_state *state, |
1731 | u32 start_seq, u32 end_seq, |
1732 | bool dup_sack_in) |
1733 | { |
1734 | struct tcp_sock *tp = tcp_sk(sk); |
1735 | struct sk_buff *tmp; |
1736 | |
1737 | skb_rbtree_walk_from(skb) { |
1738 | int in_sack = 0; |
1739 | bool dup_sack = dup_sack_in; |
1740 | |
1741 | /* queue is in-order => we can short-circuit the walk early */ |
1742 | if (!before(TCP_SKB_CB(skb)->seq, seq2: end_seq)) |
1743 | break; |
1744 | |
1745 | if (next_dup && |
1746 | before(TCP_SKB_CB(skb)->seq, seq2: next_dup->end_seq)) { |
1747 | in_sack = tcp_match_skb_to_sack(sk, skb, |
1748 | start_seq: next_dup->start_seq, |
1749 | end_seq: next_dup->end_seq); |
1750 | if (in_sack > 0) |
1751 | dup_sack = true; |
1752 | } |
1753 | |
1754 | /* skb reference here is a bit tricky to get right, since |
1755 | * shifting can eat and free both this skb and the next, |
1756 | * so not even _safe variant of the loop is enough. |
1757 | */ |
1758 | if (in_sack <= 0) { |
1759 | tmp = tcp_shift_skb_data(sk, skb, state, |
1760 | start_seq, end_seq, dup_sack); |
1761 | if (tmp) { |
1762 | if (tmp != skb) { |
1763 | skb = tmp; |
1764 | continue; |
1765 | } |
1766 | |
1767 | in_sack = 0; |
1768 | } else { |
1769 | in_sack = tcp_match_skb_to_sack(sk, skb, |
1770 | start_seq, |
1771 | end_seq); |
1772 | } |
1773 | } |
1774 | |
1775 | if (unlikely(in_sack < 0)) |
1776 | break; |
1777 | |
1778 | if (in_sack) { |
1779 | TCP_SKB_CB(skb)->sacked = |
1780 | tcp_sacktag_one(sk, |
1781 | state, |
1782 | TCP_SKB_CB(skb)->sacked, |
1783 | TCP_SKB_CB(skb)->seq, |
1784 | TCP_SKB_CB(skb)->end_seq, |
1785 | dup_sack, |
1786 | pcount: tcp_skb_pcount(skb), |
1787 | xmit_time: tcp_skb_timestamp_us(skb)); |
1788 | tcp_rate_skb_delivered(sk, skb, rs: state->rate); |
1789 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
1790 | list_del_init(entry: &skb->tcp_tsorted_anchor); |
1791 | |
1792 | if (!before(TCP_SKB_CB(skb)->seq, |
1793 | seq2: tcp_highest_sack_seq(tp))) |
1794 | tcp_advance_highest_sack(sk, skb); |
1795 | } |
1796 | } |
1797 | return skb; |
1798 | } |
1799 | |
1800 | static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq) |
1801 | { |
1802 | struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node; |
1803 | struct sk_buff *skb; |
1804 | |
1805 | while (*p) { |
1806 | parent = *p; |
1807 | skb = rb_to_skb(parent); |
1808 | if (before(seq1: seq, TCP_SKB_CB(skb)->seq)) { |
1809 | p = &parent->rb_left; |
1810 | continue; |
1811 | } |
1812 | if (!before(seq1: seq, TCP_SKB_CB(skb)->end_seq)) { |
1813 | p = &parent->rb_right; |
1814 | continue; |
1815 | } |
1816 | return skb; |
1817 | } |
1818 | return NULL; |
1819 | } |
1820 | |
1821 | static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, |
1822 | u32 skip_to_seq) |
1823 | { |
1824 | if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq)) |
1825 | return skb; |
1826 | |
1827 | return tcp_sacktag_bsearch(sk, seq: skip_to_seq); |
1828 | } |
1829 | |
1830 | static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, |
1831 | struct sock *sk, |
1832 | struct tcp_sack_block *next_dup, |
1833 | struct tcp_sacktag_state *state, |
1834 | u32 skip_to_seq) |
1835 | { |
1836 | if (!next_dup) |
1837 | return skb; |
1838 | |
1839 | if (before(seq1: next_dup->start_seq, seq2: skip_to_seq)) { |
1840 | skb = tcp_sacktag_skip(skb, sk, skip_to_seq: next_dup->start_seq); |
1841 | skb = tcp_sacktag_walk(skb, sk, NULL, state, |
1842 | start_seq: next_dup->start_seq, end_seq: next_dup->end_seq, |
1843 | dup_sack_in: 1); |
1844 | } |
1845 | |
1846 | return skb; |
1847 | } |
1848 | |
1849 | static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) |
1850 | { |
1851 | return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); |
1852 | } |
1853 | |
1854 | static int |
1855 | tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, |
1856 | u32 prior_snd_una, struct tcp_sacktag_state *state) |
1857 | { |
1858 | struct tcp_sock *tp = tcp_sk(sk); |
1859 | const unsigned char *ptr = (skb_transport_header(skb: ack_skb) + |
1860 | TCP_SKB_CB(ack_skb)->sacked); |
1861 | struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); |
1862 | struct tcp_sack_block sp[TCP_NUM_SACKS]; |
1863 | struct tcp_sack_block *cache; |
1864 | struct sk_buff *skb; |
1865 | int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); |
1866 | int used_sacks; |
1867 | bool found_dup_sack = false; |
1868 | int i, j; |
1869 | int first_sack_index; |
1870 | |
1871 | state->flag = 0; |
1872 | state->reord = tp->snd_nxt; |
1873 | |
1874 | if (!tp->sacked_out) |
1875 | tcp_highest_sack_reset(sk); |
1876 | |
1877 | found_dup_sack = tcp_check_dsack(sk, ack_skb, sp: sp_wire, |
1878 | num_sacks, prior_snd_una, state); |
1879 | |
1880 | /* Eliminate too old ACKs, but take into |
1881 | * account more or less fresh ones, they can |
1882 | * contain valid SACK info. |
1883 | */ |
1884 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, seq2: prior_snd_una - tp->max_window)) |
1885 | return 0; |
1886 | |
1887 | if (!tp->packets_out) |
1888 | goto out; |
1889 | |
1890 | used_sacks = 0; |
1891 | first_sack_index = 0; |
1892 | for (i = 0; i < num_sacks; i++) { |
1893 | bool dup_sack = !i && found_dup_sack; |
1894 | |
1895 | sp[used_sacks].start_seq = get_unaligned_be32(p: &sp_wire[i].start_seq); |
1896 | sp[used_sacks].end_seq = get_unaligned_be32(p: &sp_wire[i].end_seq); |
1897 | |
1898 | if (!tcp_is_sackblock_valid(tp, is_dsack: dup_sack, |
1899 | start_seq: sp[used_sacks].start_seq, |
1900 | end_seq: sp[used_sacks].end_seq)) { |
1901 | int mib_idx; |
1902 | |
1903 | if (dup_sack) { |
1904 | if (!tp->undo_marker) |
1905 | mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; |
1906 | else |
1907 | mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; |
1908 | } else { |
1909 | /* Don't count olds caused by ACK reordering */ |
1910 | if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && |
1911 | !after(sp[used_sacks].end_seq, tp->snd_una)) |
1912 | continue; |
1913 | mib_idx = LINUX_MIB_TCPSACKDISCARD; |
1914 | } |
1915 | |
1916 | NET_INC_STATS(sock_net(sk), mib_idx); |
1917 | if (i == 0) |
1918 | first_sack_index = -1; |
1919 | continue; |
1920 | } |
1921 | |
1922 | /* Ignore very old stuff early */ |
1923 | if (!after(sp[used_sacks].end_seq, prior_snd_una)) { |
1924 | if (i == 0) |
1925 | first_sack_index = -1; |
1926 | continue; |
1927 | } |
1928 | |
1929 | used_sacks++; |
1930 | } |
1931 | |
1932 | /* order SACK blocks to allow in order walk of the retrans queue */ |
1933 | for (i = used_sacks - 1; i > 0; i--) { |
1934 | for (j = 0; j < i; j++) { |
1935 | if (after(sp[j].start_seq, sp[j + 1].start_seq)) { |
1936 | swap(sp[j], sp[j + 1]); |
1937 | |
1938 | /* Track where the first SACK block goes to */ |
1939 | if (j == first_sack_index) |
1940 | first_sack_index = j + 1; |
1941 | } |
1942 | } |
1943 | } |
1944 | |
1945 | state->mss_now = tcp_current_mss(sk); |
1946 | skb = NULL; |
1947 | i = 0; |
1948 | |
1949 | if (!tp->sacked_out) { |
1950 | /* It's already past, so skip checking against it */ |
1951 | cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); |
1952 | } else { |
1953 | cache = tp->recv_sack_cache; |
1954 | /* Skip empty blocks in at head of the cache */ |
1955 | while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && |
1956 | !cache->end_seq) |
1957 | cache++; |
1958 | } |
1959 | |
1960 | while (i < used_sacks) { |
1961 | u32 start_seq = sp[i].start_seq; |
1962 | u32 end_seq = sp[i].end_seq; |
1963 | bool dup_sack = (found_dup_sack && (i == first_sack_index)); |
1964 | struct tcp_sack_block *next_dup = NULL; |
1965 | |
1966 | if (found_dup_sack && ((i + 1) == first_sack_index)) |
1967 | next_dup = &sp[i + 1]; |
1968 | |
1969 | /* Skip too early cached blocks */ |
1970 | while (tcp_sack_cache_ok(tp, cache) && |
1971 | !before(seq1: start_seq, seq2: cache->end_seq)) |
1972 | cache++; |
1973 | |
1974 | /* Can skip some work by looking recv_sack_cache? */ |
1975 | if (tcp_sack_cache_ok(tp, cache) && !dup_sack && |
1976 | after(end_seq, cache->start_seq)) { |
1977 | |
1978 | /* Head todo? */ |
1979 | if (before(seq1: start_seq, seq2: cache->start_seq)) { |
1980 | skb = tcp_sacktag_skip(skb, sk, skip_to_seq: start_seq); |
1981 | skb = tcp_sacktag_walk(skb, sk, next_dup, |
1982 | state, |
1983 | start_seq, |
1984 | end_seq: cache->start_seq, |
1985 | dup_sack_in: dup_sack); |
1986 | } |
1987 | |
1988 | /* Rest of the block already fully processed? */ |
1989 | if (!after(end_seq, cache->end_seq)) |
1990 | goto advance_sp; |
1991 | |
1992 | skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, |
1993 | state, |
1994 | skip_to_seq: cache->end_seq); |
1995 | |
1996 | /* ...tail remains todo... */ |
1997 | if (tcp_highest_sack_seq(tp) == cache->end_seq) { |
1998 | /* ...but better entrypoint exists! */ |
1999 | skb = tcp_highest_sack(sk); |
2000 | if (!skb) |
2001 | break; |
2002 | cache++; |
2003 | goto walk; |
2004 | } |
2005 | |
2006 | skb = tcp_sacktag_skip(skb, sk, skip_to_seq: cache->end_seq); |
2007 | /* Check overlap against next cached too (past this one already) */ |
2008 | cache++; |
2009 | continue; |
2010 | } |
2011 | |
2012 | if (!before(seq1: start_seq, seq2: tcp_highest_sack_seq(tp))) { |
2013 | skb = tcp_highest_sack(sk); |
2014 | if (!skb) |
2015 | break; |
2016 | } |
2017 | skb = tcp_sacktag_skip(skb, sk, skip_to_seq: start_seq); |
2018 | |
2019 | walk: |
2020 | skb = tcp_sacktag_walk(skb, sk, next_dup, state, |
2021 | start_seq, end_seq, dup_sack_in: dup_sack); |
2022 | |
2023 | advance_sp: |
2024 | i++; |
2025 | } |
2026 | |
2027 | /* Clear the head of the cache sack blocks so we can skip it next time */ |
2028 | for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { |
2029 | tp->recv_sack_cache[i].start_seq = 0; |
2030 | tp->recv_sack_cache[i].end_seq = 0; |
2031 | } |
2032 | for (j = 0; j < used_sacks; j++) |
2033 | tp->recv_sack_cache[i++] = sp[j]; |
2034 | |
2035 | if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker) |
2036 | tcp_check_sack_reordering(sk, low_seq: state->reord, ts: 0); |
2037 | |
2038 | tcp_verify_left_out(tp); |
2039 | out: |
2040 | |
2041 | #if FASTRETRANS_DEBUG > 0 |
2042 | WARN_ON((int)tp->sacked_out < 0); |
2043 | WARN_ON((int)tp->lost_out < 0); |
2044 | WARN_ON((int)tp->retrans_out < 0); |
2045 | WARN_ON((int)tcp_packets_in_flight(tp) < 0); |
2046 | #endif |
2047 | return state->flag; |
2048 | } |
2049 | |
2050 | /* Limits sacked_out so that sum with lost_out isn't ever larger than |
2051 | * packets_out. Returns false if sacked_out adjustement wasn't necessary. |
2052 | */ |
2053 | static bool tcp_limit_reno_sacked(struct tcp_sock *tp) |
2054 | { |
2055 | u32 holes; |
2056 | |
2057 | holes = max(tp->lost_out, 1U); |
2058 | holes = min(holes, tp->packets_out); |
2059 | |
2060 | if ((tp->sacked_out + holes) > tp->packets_out) { |
2061 | tp->sacked_out = tp->packets_out - holes; |
2062 | return true; |
2063 | } |
2064 | return false; |
2065 | } |
2066 | |
2067 | /* If we receive more dupacks than we expected counting segments |
2068 | * in assumption of absent reordering, interpret this as reordering. |
2069 | * The only another reason could be bug in receiver TCP. |
2070 | */ |
2071 | static void tcp_check_reno_reordering(struct sock *sk, const int addend) |
2072 | { |
2073 | struct tcp_sock *tp = tcp_sk(sk); |
2074 | |
2075 | if (!tcp_limit_reno_sacked(tp)) |
2076 | return; |
2077 | |
2078 | tp->reordering = min_t(u32, tp->packets_out + addend, |
2079 | READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)); |
2080 | tp->reord_seen++; |
2081 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER); |
2082 | } |
2083 | |
2084 | /* Emulate SACKs for SACKless connection: account for a new dupack. */ |
2085 | |
2086 | static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack) |
2087 | { |
2088 | if (num_dupack) { |
2089 | struct tcp_sock *tp = tcp_sk(sk); |
2090 | u32 prior_sacked = tp->sacked_out; |
2091 | s32 delivered; |
2092 | |
2093 | tp->sacked_out += num_dupack; |
2094 | tcp_check_reno_reordering(sk, addend: 0); |
2095 | delivered = tp->sacked_out - prior_sacked; |
2096 | if (delivered > 0) |
2097 | tcp_count_delivered(tp, delivered, ece_ack); |
2098 | tcp_verify_left_out(tp); |
2099 | } |
2100 | } |
2101 | |
2102 | /* Account for ACK, ACKing some data in Reno Recovery phase. */ |
2103 | |
2104 | static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack) |
2105 | { |
2106 | struct tcp_sock *tp = tcp_sk(sk); |
2107 | |
2108 | if (acked > 0) { |
2109 | /* One ACK acked hole. The rest eat duplicate ACKs. */ |
2110 | tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1), |
2111 | ece_ack); |
2112 | if (acked - 1 >= tp->sacked_out) |
2113 | tp->sacked_out = 0; |
2114 | else |
2115 | tp->sacked_out -= acked - 1; |
2116 | } |
2117 | tcp_check_reno_reordering(sk, addend: acked); |
2118 | tcp_verify_left_out(tp); |
2119 | } |
2120 | |
2121 | static inline void tcp_reset_reno_sack(struct tcp_sock *tp) |
2122 | { |
2123 | tp->sacked_out = 0; |
2124 | } |
2125 | |
2126 | void tcp_clear_retrans(struct tcp_sock *tp) |
2127 | { |
2128 | tp->retrans_out = 0; |
2129 | tp->lost_out = 0; |
2130 | tp->undo_marker = 0; |
2131 | tp->undo_retrans = -1; |
2132 | tp->sacked_out = 0; |
2133 | tp->rto_stamp = 0; |
2134 | tp->total_rto = 0; |
2135 | tp->total_rto_recoveries = 0; |
2136 | tp->total_rto_time = 0; |
2137 | } |
2138 | |
2139 | static inline void tcp_init_undo(struct tcp_sock *tp) |
2140 | { |
2141 | tp->undo_marker = tp->snd_una; |
2142 | |
2143 | /* Retransmission still in flight may cause DSACKs later. */ |
2144 | /* First, account for regular retransmits in flight: */ |
2145 | tp->undo_retrans = tp->retrans_out; |
2146 | /* Next, account for TLP retransmits in flight: */ |
2147 | if (tp->tlp_high_seq && tp->tlp_retrans) |
2148 | tp->undo_retrans++; |
2149 | /* Finally, avoid 0, because undo_retrans==0 means "can undo now": */ |
2150 | if (!tp->undo_retrans) |
2151 | tp->undo_retrans = -1; |
2152 | } |
2153 | |
2154 | static bool tcp_is_rack(const struct sock *sk) |
2155 | { |
2156 | return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) & |
2157 | TCP_RACK_LOSS_DETECTION; |
2158 | } |
2159 | |
2160 | /* If we detect SACK reneging, forget all SACK information |
2161 | * and reset tags completely, otherwise preserve SACKs. If receiver |
2162 | * dropped its ofo queue, we will know this due to reneging detection. |
2163 | */ |
2164 | static void tcp_timeout_mark_lost(struct sock *sk) |
2165 | { |
2166 | struct tcp_sock *tp = tcp_sk(sk); |
2167 | struct sk_buff *skb, *head; |
2168 | bool is_reneg; /* is receiver reneging on SACKs? */ |
2169 | |
2170 | head = tcp_rtx_queue_head(sk); |
2171 | is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED); |
2172 | if (is_reneg) { |
2173 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); |
2174 | tp->sacked_out = 0; |
2175 | /* Mark SACK reneging until we recover from this loss event. */ |
2176 | tp->is_sack_reneg = 1; |
2177 | } else if (tcp_is_reno(tp)) { |
2178 | tcp_reset_reno_sack(tp); |
2179 | } |
2180 | |
2181 | skb = head; |
2182 | skb_rbtree_walk_from(skb) { |
2183 | if (is_reneg) |
2184 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; |
2185 | else if (tcp_is_rack(sk) && skb != head && |
2186 | tcp_rack_skb_timeout(tp, skb, reo_wnd: 0) > 0) |
2187 | continue; /* Don't mark recently sent ones lost yet */ |
2188 | tcp_mark_skb_lost(sk, skb); |
2189 | } |
2190 | tcp_verify_left_out(tp); |
2191 | tcp_clear_all_retrans_hints(tp); |
2192 | } |
2193 | |
2194 | /* Enter Loss state. */ |
2195 | void tcp_enter_loss(struct sock *sk) |
2196 | { |
2197 | const struct inet_connection_sock *icsk = inet_csk(sk); |
2198 | struct tcp_sock *tp = tcp_sk(sk); |
2199 | struct net *net = sock_net(sk); |
2200 | bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery; |
2201 | u8 reordering; |
2202 | |
2203 | tcp_timeout_mark_lost(sk); |
2204 | |
2205 | /* Reduce ssthresh if it has not yet been made inside this window. */ |
2206 | if (icsk->icsk_ca_state <= TCP_CA_Disorder || |
2207 | !after(tp->high_seq, tp->snd_una) || |
2208 | (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { |
2209 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
2210 | tp->prior_cwnd = tcp_snd_cwnd(tp); |
2211 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); |
2212 | tcp_ca_event(sk, event: CA_EVENT_LOSS); |
2213 | tcp_init_undo(tp); |
2214 | } |
2215 | tcp_snd_cwnd_set(tp, val: tcp_packets_in_flight(tp) + 1); |
2216 | tp->snd_cwnd_cnt = 0; |
2217 | tp->snd_cwnd_stamp = tcp_jiffies32; |
2218 | |
2219 | /* Timeout in disordered state after receiving substantial DUPACKs |
2220 | * suggests that the degree of reordering is over-estimated. |
2221 | */ |
2222 | reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering); |
2223 | if (icsk->icsk_ca_state <= TCP_CA_Disorder && |
2224 | tp->sacked_out >= reordering) |
2225 | tp->reordering = min_t(unsigned int, tp->reordering, |
2226 | reordering); |
2227 | |
2228 | tcp_set_ca_state(sk, ca_state: TCP_CA_Loss); |
2229 | tp->high_seq = tp->snd_nxt; |
2230 | tp->tlp_high_seq = 0; |
2231 | tcp_ecn_queue_cwr(tp); |
2232 | |
2233 | /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous |
2234 | * loss recovery is underway except recurring timeout(s) on |
2235 | * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing |
2236 | */ |
2237 | tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) && |
2238 | (new_recovery || icsk->icsk_retransmits) && |
2239 | !inet_csk(sk)->icsk_mtup.probe_size; |
2240 | } |
2241 | |
2242 | /* If ACK arrived pointing to a remembered SACK, it means that our |
2243 | * remembered SACKs do not reflect real state of receiver i.e. |
2244 | * receiver _host_ is heavily congested (or buggy). |
2245 | * |
2246 | * To avoid big spurious retransmission bursts due to transient SACK |
2247 | * scoreboard oddities that look like reneging, we give the receiver a |
2248 | * little time (max(RTT/2, 10ms)) to send us some more ACKs that will |
2249 | * restore sanity to the SACK scoreboard. If the apparent reneging |
2250 | * persists until this RTO then we'll clear the SACK scoreboard. |
2251 | */ |
2252 | static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag) |
2253 | { |
2254 | if (*ack_flag & FLAG_SACK_RENEGING && |
2255 | *ack_flag & FLAG_SND_UNA_ADVANCED) { |
2256 | struct tcp_sock *tp = tcp_sk(sk); |
2257 | unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4), |
2258 | msecs_to_jiffies(10)); |
2259 | |
2260 | tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, when: delay, pace_delay: false); |
2261 | *ack_flag &= ~FLAG_SET_XMIT_TIMER; |
2262 | return true; |
2263 | } |
2264 | return false; |
2265 | } |
2266 | |
2267 | /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs |
2268 | * counter when SACK is enabled (without SACK, sacked_out is used for |
2269 | * that purpose). |
2270 | * |
2271 | * With reordering, holes may still be in flight, so RFC3517 recovery |
2272 | * uses pure sacked_out (total number of SACKed segments) even though |
2273 | * it violates the RFC that uses duplicate ACKs, often these are equal |
2274 | * but when e.g. out-of-window ACKs or packet duplication occurs, |
2275 | * they differ. Since neither occurs due to loss, TCP should really |
2276 | * ignore them. |
2277 | */ |
2278 | static inline int tcp_dupack_heuristics(const struct tcp_sock *tp) |
2279 | { |
2280 | return tp->sacked_out + 1; |
2281 | } |
2282 | |
2283 | /* Linux NewReno/SACK/ECN state machine. |
2284 | * -------------------------------------- |
2285 | * |
2286 | * "Open" Normal state, no dubious events, fast path. |
2287 | * "Disorder" In all the respects it is "Open", |
2288 | * but requires a bit more attention. It is entered when |
2289 | * we see some SACKs or dupacks. It is split of "Open" |
2290 | * mainly to move some processing from fast path to slow one. |
2291 | * "CWR" CWND was reduced due to some Congestion Notification event. |
2292 | * It can be ECN, ICMP source quench, local device congestion. |
2293 | * "Recovery" CWND was reduced, we are fast-retransmitting. |
2294 | * "Loss" CWND was reduced due to RTO timeout or SACK reneging. |
2295 | * |
2296 | * tcp_fastretrans_alert() is entered: |
2297 | * - each incoming ACK, if state is not "Open" |
2298 | * - when arrived ACK is unusual, namely: |
2299 | * * SACK |
2300 | * * Duplicate ACK. |
2301 | * * ECN ECE. |
2302 | * |
2303 | * Counting packets in flight is pretty simple. |
2304 | * |
2305 | * in_flight = packets_out - left_out + retrans_out |
2306 | * |
2307 | * packets_out is SND.NXT-SND.UNA counted in packets. |
2308 | * |
2309 | * retrans_out is number of retransmitted segments. |
2310 | * |
2311 | * left_out is number of segments left network, but not ACKed yet. |
2312 | * |
2313 | * left_out = sacked_out + lost_out |
2314 | * |
2315 | * sacked_out: Packets, which arrived to receiver out of order |
2316 | * and hence not ACKed. With SACKs this number is simply |
2317 | * amount of SACKed data. Even without SACKs |
2318 | * it is easy to give pretty reliable estimate of this number, |
2319 | * counting duplicate ACKs. |
2320 | * |
2321 | * lost_out: Packets lost by network. TCP has no explicit |
2322 | * "loss notification" feedback from network (for now). |
2323 | * It means that this number can be only _guessed_. |
2324 | * Actually, it is the heuristics to predict lossage that |
2325 | * distinguishes different algorithms. |
2326 | * |
2327 | * F.e. after RTO, when all the queue is considered as lost, |
2328 | * lost_out = packets_out and in_flight = retrans_out. |
2329 | * |
2330 | * Essentially, we have now a few algorithms detecting |
2331 | * lost packets. |
2332 | * |
2333 | * If the receiver supports SACK: |
2334 | * |
2335 | * RFC6675/3517: It is the conventional algorithm. A packet is |
2336 | * considered lost if the number of higher sequence packets |
2337 | * SACKed is greater than or equal the DUPACK thoreshold |
2338 | * (reordering). This is implemented in tcp_mark_head_lost and |
2339 | * tcp_update_scoreboard. |
2340 | * |
2341 | * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm |
2342 | * (2017-) that checks timing instead of counting DUPACKs. |
2343 | * Essentially a packet is considered lost if it's not S/ACKed |
2344 | * after RTT + reordering_window, where both metrics are |
2345 | * dynamically measured and adjusted. This is implemented in |
2346 | * tcp_rack_mark_lost. |
2347 | * |
2348 | * If the receiver does not support SACK: |
2349 | * |
2350 | * NewReno (RFC6582): in Recovery we assume that one segment |
2351 | * is lost (classic Reno). While we are in Recovery and |
2352 | * a partial ACK arrives, we assume that one more packet |
2353 | * is lost (NewReno). This heuristics are the same in NewReno |
2354 | * and SACK. |
2355 | * |
2356 | * Really tricky (and requiring careful tuning) part of algorithm |
2357 | * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). |
2358 | * The first determines the moment _when_ we should reduce CWND and, |
2359 | * hence, slow down forward transmission. In fact, it determines the moment |
2360 | * when we decide that hole is caused by loss, rather than by a reorder. |
2361 | * |
2362 | * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill |
2363 | * holes, caused by lost packets. |
2364 | * |
2365 | * And the most logically complicated part of algorithm is undo |
2366 | * heuristics. We detect false retransmits due to both too early |
2367 | * fast retransmit (reordering) and underestimated RTO, analyzing |
2368 | * timestamps and D-SACKs. When we detect that some segments were |
2369 | * retransmitted by mistake and CWND reduction was wrong, we undo |
2370 | * window reduction and abort recovery phase. This logic is hidden |
2371 | * inside several functions named tcp_try_undo_<something>. |
2372 | */ |
2373 | |
2374 | /* This function decides, when we should leave Disordered state |
2375 | * and enter Recovery phase, reducing congestion window. |
2376 | * |
2377 | * Main question: may we further continue forward transmission |
2378 | * with the same cwnd? |
2379 | */ |
2380 | static bool tcp_time_to_recover(struct sock *sk, int flag) |
2381 | { |
2382 | struct tcp_sock *tp = tcp_sk(sk); |
2383 | |
2384 | /* Trick#1: The loss is proven. */ |
2385 | if (tp->lost_out) |
2386 | return true; |
2387 | |
2388 | /* Not-A-Trick#2 : Classic rule... */ |
2389 | if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering) |
2390 | return true; |
2391 | |
2392 | return false; |
2393 | } |
2394 | |
2395 | /* Detect loss in event "A" above by marking head of queue up as lost. |
2396 | * For RFC3517 SACK, a segment is considered lost if it |
2397 | * has at least tp->reordering SACKed seqments above it; "packets" refers to |
2398 | * the maximum SACKed segments to pass before reaching this limit. |
2399 | */ |
2400 | static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head) |
2401 | { |
2402 | struct tcp_sock *tp = tcp_sk(sk); |
2403 | struct sk_buff *skb; |
2404 | int cnt; |
2405 | /* Use SACK to deduce losses of new sequences sent during recovery */ |
2406 | const u32 loss_high = tp->snd_nxt; |
2407 | |
2408 | WARN_ON(packets > tp->packets_out); |
2409 | skb = tp->lost_skb_hint; |
2410 | if (skb) { |
2411 | /* Head already handled? */ |
2412 | if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una)) |
2413 | return; |
2414 | cnt = tp->lost_cnt_hint; |
2415 | } else { |
2416 | skb = tcp_rtx_queue_head(sk); |
2417 | cnt = 0; |
2418 | } |
2419 | |
2420 | skb_rbtree_walk_from(skb) { |
2421 | /* TODO: do this better */ |
2422 | /* this is not the most efficient way to do this... */ |
2423 | tp->lost_skb_hint = skb; |
2424 | tp->lost_cnt_hint = cnt; |
2425 | |
2426 | if (after(TCP_SKB_CB(skb)->end_seq, loss_high)) |
2427 | break; |
2428 | |
2429 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
2430 | cnt += tcp_skb_pcount(skb); |
2431 | |
2432 | if (cnt > packets) |
2433 | break; |
2434 | |
2435 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST)) |
2436 | tcp_mark_skb_lost(sk, skb); |
2437 | |
2438 | if (mark_head) |
2439 | break; |
2440 | } |
2441 | tcp_verify_left_out(tp); |
2442 | } |
2443 | |
2444 | /* Account newly detected lost packet(s) */ |
2445 | |
2446 | static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) |
2447 | { |
2448 | struct tcp_sock *tp = tcp_sk(sk); |
2449 | |
2450 | if (tcp_is_sack(tp)) { |
2451 | int sacked_upto = tp->sacked_out - tp->reordering; |
2452 | if (sacked_upto >= 0) |
2453 | tcp_mark_head_lost(sk, packets: sacked_upto, mark_head: 0); |
2454 | else if (fast_rexmit) |
2455 | tcp_mark_head_lost(sk, packets: 1, mark_head: 1); |
2456 | } |
2457 | } |
2458 | |
2459 | static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when) |
2460 | { |
2461 | return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && |
2462 | before(seq1: tp->rx_opt.rcv_tsecr, seq2: when); |
2463 | } |
2464 | |
2465 | /* skb is spurious retransmitted if the returned timestamp echo |
2466 | * reply is prior to the skb transmission time |
2467 | */ |
2468 | static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp, |
2469 | const struct sk_buff *skb) |
2470 | { |
2471 | return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) && |
2472 | tcp_tsopt_ecr_before(tp, when: tcp_skb_timestamp_ts(usec_ts: tp->tcp_usec_ts, skb)); |
2473 | } |
2474 | |
2475 | /* Nothing was retransmitted or returned timestamp is less |
2476 | * than timestamp of the first retransmission. |
2477 | */ |
2478 | static inline bool tcp_packet_delayed(const struct tcp_sock *tp) |
2479 | { |
2480 | const struct sock *sk = (const struct sock *)tp; |
2481 | |
2482 | if (tp->retrans_stamp && |
2483 | tcp_tsopt_ecr_before(tp, when: tp->retrans_stamp)) |
2484 | return true; /* got echoed TS before first retransmission */ |
2485 | |
2486 | /* Check if nothing was retransmitted (retrans_stamp==0), which may |
2487 | * happen in fast recovery due to TSQ. But we ignore zero retrans_stamp |
2488 | * in TCP_SYN_SENT, since when we set FLAG_SYN_ACKED we also clear |
2489 | * retrans_stamp even if we had retransmitted the SYN. |
2490 | */ |
2491 | if (!tp->retrans_stamp && /* no record of a retransmit/SYN? */ |
2492 | sk->sk_state != TCP_SYN_SENT) /* not the FLAG_SYN_ACKED case? */ |
2493 | return true; /* nothing was retransmitted */ |
2494 | |
2495 | return false; |
2496 | } |
2497 | |
2498 | /* Undo procedures. */ |
2499 | |
2500 | /* We can clear retrans_stamp when there are no retransmissions in the |
2501 | * window. It would seem that it is trivially available for us in |
2502 | * tp->retrans_out, however, that kind of assumptions doesn't consider |
2503 | * what will happen if errors occur when sending retransmission for the |
2504 | * second time. ...It could the that such segment has only |
2505 | * TCPCB_EVER_RETRANS set at the present time. It seems that checking |
2506 | * the head skb is enough except for some reneging corner cases that |
2507 | * are not worth the effort. |
2508 | * |
2509 | * Main reason for all this complexity is the fact that connection dying |
2510 | * time now depends on the validity of the retrans_stamp, in particular, |
2511 | * that successive retransmissions of a segment must not advance |
2512 | * retrans_stamp under any conditions. |
2513 | */ |
2514 | static bool tcp_any_retrans_done(const struct sock *sk) |
2515 | { |
2516 | const struct tcp_sock *tp = tcp_sk(sk); |
2517 | struct sk_buff *skb; |
2518 | |
2519 | if (tp->retrans_out) |
2520 | return true; |
2521 | |
2522 | skb = tcp_rtx_queue_head(sk); |
2523 | if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) |
2524 | return true; |
2525 | |
2526 | return false; |
2527 | } |
2528 | |
2529 | /* If loss recovery is finished and there are no retransmits out in the |
2530 | * network, then we clear retrans_stamp so that upon the next loss recovery |
2531 | * retransmits_timed_out() and timestamp-undo are using the correct value. |
2532 | */ |
2533 | static void tcp_retrans_stamp_cleanup(struct sock *sk) |
2534 | { |
2535 | if (!tcp_any_retrans_done(sk)) |
2536 | tcp_sk(sk)->retrans_stamp = 0; |
2537 | } |
2538 | |
2539 | static void DBGUNDO(struct sock *sk, const char *msg) |
2540 | { |
2541 | #if FASTRETRANS_DEBUG > 1 |
2542 | struct tcp_sock *tp = tcp_sk(sk); |
2543 | struct inet_sock *inet = inet_sk(sk); |
2544 | |
2545 | if (sk->sk_family == AF_INET) { |
2546 | pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", |
2547 | msg, |
2548 | &inet->inet_daddr, ntohs(inet->inet_dport), |
2549 | tcp_snd_cwnd(tp), tcp_left_out(tp), |
2550 | tp->snd_ssthresh, tp->prior_ssthresh, |
2551 | tp->packets_out); |
2552 | } |
2553 | #if IS_ENABLED(CONFIG_IPV6) |
2554 | else if (sk->sk_family == AF_INET6) { |
2555 | pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", |
2556 | msg, |
2557 | &sk->sk_v6_daddr, ntohs(inet->inet_dport), |
2558 | tcp_snd_cwnd(tp), tcp_left_out(tp), |
2559 | tp->snd_ssthresh, tp->prior_ssthresh, |
2560 | tp->packets_out); |
2561 | } |
2562 | #endif |
2563 | #endif |
2564 | } |
2565 | |
2566 | static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss) |
2567 | { |
2568 | struct tcp_sock *tp = tcp_sk(sk); |
2569 | |
2570 | if (unmark_loss) { |
2571 | struct sk_buff *skb; |
2572 | |
2573 | skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { |
2574 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; |
2575 | } |
2576 | tp->lost_out = 0; |
2577 | tcp_clear_all_retrans_hints(tp); |
2578 | } |
2579 | |
2580 | if (tp->prior_ssthresh) { |
2581 | const struct inet_connection_sock *icsk = inet_csk(sk); |
2582 | |
2583 | tcp_snd_cwnd_set(tp, val: icsk->icsk_ca_ops->undo_cwnd(sk)); |
2584 | |
2585 | if (tp->prior_ssthresh > tp->snd_ssthresh) { |
2586 | tp->snd_ssthresh = tp->prior_ssthresh; |
2587 | tcp_ecn_withdraw_cwr(tp); |
2588 | } |
2589 | } |
2590 | tp->snd_cwnd_stamp = tcp_jiffies32; |
2591 | tp->undo_marker = 0; |
2592 | tp->rack.advanced = 1; /* Force RACK to re-exam losses */ |
2593 | } |
2594 | |
2595 | static inline bool tcp_may_undo(const struct tcp_sock *tp) |
2596 | { |
2597 | return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); |
2598 | } |
2599 | |
2600 | static bool tcp_is_non_sack_preventing_reopen(struct sock *sk) |
2601 | { |
2602 | struct tcp_sock *tp = tcp_sk(sk); |
2603 | |
2604 | if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { |
2605 | /* Hold old state until something *above* high_seq |
2606 | * is ACKed. For Reno it is MUST to prevent false |
2607 | * fast retransmits (RFC2582). SACK TCP is safe. */ |
2608 | if (!tcp_any_retrans_done(sk)) |
2609 | tp->retrans_stamp = 0; |
2610 | return true; |
2611 | } |
2612 | return false; |
2613 | } |
2614 | |
2615 | /* People celebrate: "We love our President!" */ |
2616 | static bool tcp_try_undo_recovery(struct sock *sk) |
2617 | { |
2618 | struct tcp_sock *tp = tcp_sk(sk); |
2619 | |
2620 | if (tcp_may_undo(tp)) { |
2621 | int mib_idx; |
2622 | |
2623 | /* Happy end! We did not retransmit anything |
2624 | * or our original transmission succeeded. |
2625 | */ |
2626 | DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss": "retrans"); |
2627 | tcp_undo_cwnd_reduction(sk, unmark_loss: false); |
2628 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) |
2629 | mib_idx = LINUX_MIB_TCPLOSSUNDO; |
2630 | else |
2631 | mib_idx = LINUX_MIB_TCPFULLUNDO; |
2632 | |
2633 | NET_INC_STATS(sock_net(sk), mib_idx); |
2634 | } else if (tp->rack.reo_wnd_persist) { |
2635 | tp->rack.reo_wnd_persist--; |
2636 | } |
2637 | if (tcp_is_non_sack_preventing_reopen(sk)) |
2638 | return true; |
2639 | tcp_set_ca_state(sk, ca_state: TCP_CA_Open); |
2640 | tp->is_sack_reneg = 0; |
2641 | return false; |
2642 | } |
2643 | |
2644 | /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ |
2645 | static bool tcp_try_undo_dsack(struct sock *sk) |
2646 | { |
2647 | struct tcp_sock *tp = tcp_sk(sk); |
2648 | |
2649 | if (tp->undo_marker && !tp->undo_retrans) { |
2650 | tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH, |
2651 | tp->rack.reo_wnd_persist + 1); |
2652 | DBGUNDO(sk, msg: "D-SACK"); |
2653 | tcp_undo_cwnd_reduction(sk, unmark_loss: false); |
2654 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); |
2655 | return true; |
2656 | } |
2657 | return false; |
2658 | } |
2659 | |
2660 | /* Undo during loss recovery after partial ACK or using F-RTO. */ |
2661 | static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo) |
2662 | { |
2663 | struct tcp_sock *tp = tcp_sk(sk); |
2664 | |
2665 | if (frto_undo || tcp_may_undo(tp)) { |
2666 | tcp_undo_cwnd_reduction(sk, unmark_loss: true); |
2667 | |
2668 | DBGUNDO(sk, msg: "partial loss"); |
2669 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); |
2670 | if (frto_undo) |
2671 | NET_INC_STATS(sock_net(sk), |
2672 | LINUX_MIB_TCPSPURIOUSRTOS); |
2673 | inet_csk(sk)->icsk_retransmits = 0; |
2674 | if (tcp_is_non_sack_preventing_reopen(sk)) |
2675 | return true; |
2676 | if (frto_undo || tcp_is_sack(tp)) { |
2677 | tcp_set_ca_state(sk, ca_state: TCP_CA_Open); |
2678 | tp->is_sack_reneg = 0; |
2679 | } |
2680 | return true; |
2681 | } |
2682 | return false; |
2683 | } |
2684 | |
2685 | /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937. |
2686 | * It computes the number of packets to send (sndcnt) based on packets newly |
2687 | * delivered: |
2688 | * 1) If the packets in flight is larger than ssthresh, PRR spreads the |
2689 | * cwnd reductions across a full RTT. |
2690 | * 2) Otherwise PRR uses packet conservation to send as much as delivered. |
2691 | * But when SND_UNA is acked without further losses, |
2692 | * slow starts cwnd up to ssthresh to speed up the recovery. |
2693 | */ |
2694 | static void tcp_init_cwnd_reduction(struct sock *sk) |
2695 | { |
2696 | struct tcp_sock *tp = tcp_sk(sk); |
2697 | |
2698 | tp->high_seq = tp->snd_nxt; |
2699 | tp->tlp_high_seq = 0; |
2700 | tp->snd_cwnd_cnt = 0; |
2701 | tp->prior_cwnd = tcp_snd_cwnd(tp); |
2702 | tp->prr_delivered = 0; |
2703 | tp->prr_out = 0; |
2704 | tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); |
2705 | tcp_ecn_queue_cwr(tp); |
2706 | } |
2707 | |
2708 | void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag) |
2709 | { |
2710 | struct tcp_sock *tp = tcp_sk(sk); |
2711 | int sndcnt = 0; |
2712 | int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); |
2713 | |
2714 | if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd)) |
2715 | return; |
2716 | |
2717 | trace_tcp_cwnd_reduction_tp(sk, newly_acked_sacked, newly_lost, flag); |
2718 | |
2719 | tp->prr_delivered += newly_acked_sacked; |
2720 | if (delta < 0) { |
2721 | u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + |
2722 | tp->prior_cwnd - 1; |
2723 | sndcnt = div_u64(dividend, divisor: tp->prior_cwnd) - tp->prr_out; |
2724 | } else { |
2725 | sndcnt = max_t(int, tp->prr_delivered - tp->prr_out, |
2726 | newly_acked_sacked); |
2727 | if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost) |
2728 | sndcnt++; |
2729 | sndcnt = min(delta, sndcnt); |
2730 | } |
2731 | /* Force a fast retransmit upon entering fast recovery */ |
2732 | sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1)); |
2733 | tcp_snd_cwnd_set(tp, val: tcp_packets_in_flight(tp) + sndcnt); |
2734 | } |
2735 | |
2736 | static inline void tcp_end_cwnd_reduction(struct sock *sk) |
2737 | { |
2738 | struct tcp_sock *tp = tcp_sk(sk); |
2739 | |
2740 | if (inet_csk(sk)->icsk_ca_ops->cong_control) |
2741 | return; |
2742 | |
2743 | /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */ |
2744 | if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH && |
2745 | (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) { |
2746 | tcp_snd_cwnd_set(tp, val: tp->snd_ssthresh); |
2747 | tp->snd_cwnd_stamp = tcp_jiffies32; |
2748 | } |
2749 | tcp_ca_event(sk, event: CA_EVENT_COMPLETE_CWR); |
2750 | } |
2751 | |
2752 | /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */ |
2753 | void tcp_enter_cwr(struct sock *sk) |
2754 | { |
2755 | struct tcp_sock *tp = tcp_sk(sk); |
2756 | |
2757 | tp->prior_ssthresh = 0; |
2758 | if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { |
2759 | tp->undo_marker = 0; |
2760 | tcp_init_cwnd_reduction(sk); |
2761 | tcp_set_ca_state(sk, ca_state: TCP_CA_CWR); |
2762 | } |
2763 | } |
2764 | EXPORT_SYMBOL(tcp_enter_cwr); |
2765 | |
2766 | static void tcp_try_keep_open(struct sock *sk) |
2767 | { |
2768 | struct tcp_sock *tp = tcp_sk(sk); |
2769 | int state = TCP_CA_Open; |
2770 | |
2771 | if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) |
2772 | state = TCP_CA_Disorder; |
2773 | |
2774 | if (inet_csk(sk)->icsk_ca_state != state) { |
2775 | tcp_set_ca_state(sk, ca_state: state); |
2776 | tp->high_seq = tp->snd_nxt; |
2777 | } |
2778 | } |
2779 | |
2780 | static void tcp_try_to_open(struct sock *sk, int flag) |
2781 | { |
2782 | struct tcp_sock *tp = tcp_sk(sk); |
2783 | |
2784 | tcp_verify_left_out(tp); |
2785 | |
2786 | if (!tcp_any_retrans_done(sk)) |
2787 | tp->retrans_stamp = 0; |
2788 | |
2789 | if (flag & FLAG_ECE) |
2790 | tcp_enter_cwr(sk); |
2791 | |
2792 | if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { |
2793 | tcp_try_keep_open(sk); |
2794 | } |
2795 | } |
2796 | |
2797 | static void tcp_mtup_probe_failed(struct sock *sk) |
2798 | { |
2799 | struct inet_connection_sock *icsk = inet_csk(sk); |
2800 | |
2801 | icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; |
2802 | icsk->icsk_mtup.probe_size = 0; |
2803 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL); |
2804 | } |
2805 | |
2806 | static void tcp_mtup_probe_success(struct sock *sk) |
2807 | { |
2808 | struct tcp_sock *tp = tcp_sk(sk); |
2809 | struct inet_connection_sock *icsk = inet_csk(sk); |
2810 | u64 val; |
2811 | |
2812 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
2813 | |
2814 | val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, mss: tp->mss_cache); |
2815 | do_div(val, icsk->icsk_mtup.probe_size); |
2816 | DEBUG_NET_WARN_ON_ONCE((u32)val != val); |
2817 | tcp_snd_cwnd_set(tp, max_t(u32, 1U, val)); |
2818 | |
2819 | tp->snd_cwnd_cnt = 0; |
2820 | tp->snd_cwnd_stamp = tcp_jiffies32; |
2821 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
2822 | |
2823 | icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; |
2824 | icsk->icsk_mtup.probe_size = 0; |
2825 | tcp_sync_mss(sk, pmtu: icsk->icsk_pmtu_cookie); |
2826 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS); |
2827 | } |
2828 | |
2829 | /* Sometimes we deduce that packets have been dropped due to reasons other than |
2830 | * congestion, like path MTU reductions or failed client TFO attempts. In these |
2831 | * cases we call this function to retransmit as many packets as cwnd allows, |
2832 | * without reducing cwnd. Given that retransmits will set retrans_stamp to a |
2833 | * non-zero value (and may do so in a later calling context due to TSQ), we |
2834 | * also enter CA_Loss so that we track when all retransmitted packets are ACKed |
2835 | * and clear retrans_stamp when that happens (to ensure later recurring RTOs |
2836 | * are using the correct retrans_stamp and don't declare ETIMEDOUT |
2837 | * prematurely). |
2838 | */ |
2839 | static void tcp_non_congestion_loss_retransmit(struct sock *sk) |
2840 | { |
2841 | const struct inet_connection_sock *icsk = inet_csk(sk); |
2842 | struct tcp_sock *tp = tcp_sk(sk); |
2843 | |
2844 | if (icsk->icsk_ca_state != TCP_CA_Loss) { |
2845 | tp->high_seq = tp->snd_nxt; |
2846 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
2847 | tp->prior_ssthresh = 0; |
2848 | tp->undo_marker = 0; |
2849 | tcp_set_ca_state(sk, ca_state: TCP_CA_Loss); |
2850 | } |
2851 | tcp_xmit_retransmit_queue(sk); |
2852 | } |
2853 | |
2854 | /* Do a simple retransmit without using the backoff mechanisms in |
2855 | * tcp_timer. This is used for path mtu discovery. |
2856 | * The socket is already locked here. |
2857 | */ |
2858 | void tcp_simple_retransmit(struct sock *sk) |
2859 | { |
2860 | struct tcp_sock *tp = tcp_sk(sk); |
2861 | struct sk_buff *skb; |
2862 | int mss; |
2863 | |
2864 | /* A fastopen SYN request is stored as two separate packets within |
2865 | * the retransmit queue, this is done by tcp_send_syn_data(). |
2866 | * As a result simply checking the MSS of the frames in the queue |
2867 | * will not work for the SYN packet. |
2868 | * |
2869 | * Us being here is an indication of a path MTU issue so we can |
2870 | * assume that the fastopen SYN was lost and just mark all the |
2871 | * frames in the retransmit queue as lost. We will use an MSS of |
2872 | * -1 to mark all frames as lost, otherwise compute the current MSS. |
2873 | */ |
2874 | if (tp->syn_data && sk->sk_state == TCP_SYN_SENT) |
2875 | mss = -1; |
2876 | else |
2877 | mss = tcp_current_mss(sk); |
2878 | |
2879 | skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { |
2880 | if (tcp_skb_seglen(skb) > mss) |
2881 | tcp_mark_skb_lost(sk, skb); |
2882 | } |
2883 | |
2884 | tcp_clear_retrans_hints_partial(tp); |
2885 | |
2886 | if (!tp->lost_out) |
2887 | return; |
2888 | |
2889 | if (tcp_is_reno(tp)) |
2890 | tcp_limit_reno_sacked(tp); |
2891 | |
2892 | tcp_verify_left_out(tp); |
2893 | |
2894 | /* Don't muck with the congestion window here. |
2895 | * Reason is that we do not increase amount of _data_ |
2896 | * in network, but units changed and effective |
2897 | * cwnd/ssthresh really reduced now. |
2898 | */ |
2899 | tcp_non_congestion_loss_retransmit(sk); |
2900 | } |
2901 | EXPORT_IPV6_MOD(tcp_simple_retransmit); |
2902 | |
2903 | void tcp_enter_recovery(struct sock *sk, bool ece_ack) |
2904 | { |
2905 | struct tcp_sock *tp = tcp_sk(sk); |
2906 | int mib_idx; |
2907 | |
2908 | /* Start the clock with our fast retransmit, for undo and ETIMEDOUT. */ |
2909 | tcp_retrans_stamp_cleanup(sk); |
2910 | |
2911 | if (tcp_is_reno(tp)) |
2912 | mib_idx = LINUX_MIB_TCPRENORECOVERY; |
2913 | else |
2914 | mib_idx = LINUX_MIB_TCPSACKRECOVERY; |
2915 | |
2916 | NET_INC_STATS(sock_net(sk), mib_idx); |
2917 | |
2918 | tp->prior_ssthresh = 0; |
2919 | tcp_init_undo(tp); |
2920 | |
2921 | if (!tcp_in_cwnd_reduction(sk)) { |
2922 | if (!ece_ack) |
2923 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
2924 | tcp_init_cwnd_reduction(sk); |
2925 | } |
2926 | tcp_set_ca_state(sk, ca_state: TCP_CA_Recovery); |
2927 | } |
2928 | |
2929 | static void tcp_update_rto_time(struct tcp_sock *tp) |
2930 | { |
2931 | if (tp->rto_stamp) { |
2932 | tp->total_rto_time += tcp_time_stamp_ms(tp) - tp->rto_stamp; |
2933 | tp->rto_stamp = 0; |
2934 | } |
2935 | } |
2936 | |
2937 | /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are |
2938 | * recovered or spurious. Otherwise retransmits more on partial ACKs. |
2939 | */ |
2940 | static void tcp_process_loss(struct sock *sk, int flag, int num_dupack, |
2941 | int *rexmit) |
2942 | { |
2943 | struct tcp_sock *tp = tcp_sk(sk); |
2944 | bool recovered = !before(seq1: tp->snd_una, seq2: tp->high_seq); |
2945 | |
2946 | if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) && |
2947 | tcp_try_undo_loss(sk, frto_undo: false)) |
2948 | return; |
2949 | |
2950 | if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */ |
2951 | /* Step 3.b. A timeout is spurious if not all data are |
2952 | * lost, i.e., never-retransmitted data are (s)acked. |
2953 | */ |
2954 | if ((flag & FLAG_ORIG_SACK_ACKED) && |
2955 | tcp_try_undo_loss(sk, frto_undo: true)) |
2956 | return; |
2957 | |
2958 | if (after(tp->snd_nxt, tp->high_seq)) { |
2959 | if (flag & FLAG_DATA_SACKED || num_dupack) |
2960 | tp->frto = 0; /* Step 3.a. loss was real */ |
2961 | } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) { |
2962 | tp->high_seq = tp->snd_nxt; |
2963 | /* Step 2.b. Try send new data (but deferred until cwnd |
2964 | * is updated in tcp_ack()). Otherwise fall back to |
2965 | * the conventional recovery. |
2966 | */ |
2967 | if (!tcp_write_queue_empty(sk) && |
2968 | after(tcp_wnd_end(tp), tp->snd_nxt)) { |
2969 | *rexmit = REXMIT_NEW; |
2970 | return; |
2971 | } |
2972 | tp->frto = 0; |
2973 | } |
2974 | } |
2975 | |
2976 | if (recovered) { |
2977 | /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */ |
2978 | tcp_try_undo_recovery(sk); |
2979 | return; |
2980 | } |
2981 | if (tcp_is_reno(tp)) { |
2982 | /* A Reno DUPACK means new data in F-RTO step 2.b above are |
2983 | * delivered. Lower inflight to clock out (re)transmissions. |
2984 | */ |
2985 | if (after(tp->snd_nxt, tp->high_seq) && num_dupack) |
2986 | tcp_add_reno_sack(sk, num_dupack, ece_ack: flag & FLAG_ECE); |
2987 | else if (flag & FLAG_SND_UNA_ADVANCED) |
2988 | tcp_reset_reno_sack(tp); |
2989 | } |
2990 | *rexmit = REXMIT_LOST; |
2991 | } |
2992 | |
2993 | static bool tcp_force_fast_retransmit(struct sock *sk) |
2994 | { |
2995 | struct tcp_sock *tp = tcp_sk(sk); |
2996 | |
2997 | return after(tcp_highest_sack_seq(tp), |
2998 | tp->snd_una + tp->reordering * tp->mss_cache); |
2999 | } |
3000 | |
3001 | /* Undo during fast recovery after partial ACK. */ |
3002 | static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una, |
3003 | bool *do_lost) |
3004 | { |
3005 | struct tcp_sock *tp = tcp_sk(sk); |
3006 | |
3007 | if (tp->undo_marker && tcp_packet_delayed(tp)) { |
3008 | /* Plain luck! Hole if filled with delayed |
3009 | * packet, rather than with a retransmit. Check reordering. |
3010 | */ |
3011 | tcp_check_sack_reordering(sk, low_seq: prior_snd_una, ts: 1); |
3012 | |
3013 | /* We are getting evidence that the reordering degree is higher |
3014 | * than we realized. If there are no retransmits out then we |
3015 | * can undo. Otherwise we clock out new packets but do not |
3016 | * mark more packets lost or retransmit more. |
3017 | */ |
3018 | if (tp->retrans_out) |
3019 | return true; |
3020 | |
3021 | if (!tcp_any_retrans_done(sk)) |
3022 | tp->retrans_stamp = 0; |
3023 | |
3024 | DBGUNDO(sk, msg: "partial recovery"); |
3025 | tcp_undo_cwnd_reduction(sk, unmark_loss: true); |
3026 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); |
3027 | tcp_try_keep_open(sk); |
3028 | } else { |
3029 | /* Partial ACK arrived. Force fast retransmit. */ |
3030 | *do_lost = tcp_force_fast_retransmit(sk); |
3031 | } |
3032 | return false; |
3033 | } |
3034 | |
3035 | static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag) |
3036 | { |
3037 | struct tcp_sock *tp = tcp_sk(sk); |
3038 | |
3039 | if (tcp_rtx_queue_empty(sk)) |
3040 | return; |
3041 | |
3042 | if (unlikely(tcp_is_reno(tp))) { |
3043 | tcp_newreno_mark_lost(sk, snd_una_advanced: *ack_flag & FLAG_SND_UNA_ADVANCED); |
3044 | } else if (tcp_is_rack(sk)) { |
3045 | u32 prior_retrans = tp->retrans_out; |
3046 | |
3047 | if (tcp_rack_mark_lost(sk)) |
3048 | *ack_flag &= ~FLAG_SET_XMIT_TIMER; |
3049 | if (prior_retrans > tp->retrans_out) |
3050 | *ack_flag |= FLAG_LOST_RETRANS; |
3051 | } |
3052 | } |
3053 | |
3054 | /* Process an event, which can update packets-in-flight not trivially. |
3055 | * Main goal of this function is to calculate new estimate for left_out, |
3056 | * taking into account both packets sitting in receiver's buffer and |
3057 | * packets lost by network. |
3058 | * |
3059 | * Besides that it updates the congestion state when packet loss or ECN |
3060 | * is detected. But it does not reduce the cwnd, it is done by the |
3061 | * congestion control later. |
3062 | * |
3063 | * It does _not_ decide what to send, it is made in function |
3064 | * tcp_xmit_retransmit_queue(). |
3065 | */ |
3066 | static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una, |
3067 | int num_dupack, int *ack_flag, int *rexmit) |
3068 | { |
3069 | struct inet_connection_sock *icsk = inet_csk(sk); |
3070 | struct tcp_sock *tp = tcp_sk(sk); |
3071 | int fast_rexmit = 0, flag = *ack_flag; |
3072 | bool ece_ack = flag & FLAG_ECE; |
3073 | bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) && |
3074 | tcp_force_fast_retransmit(sk)); |
3075 | |
3076 | if (!tp->packets_out && tp->sacked_out) |
3077 | tp->sacked_out = 0; |
3078 | |
3079 | /* Now state machine starts. |
3080 | * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ |
3081 | if (ece_ack) |
3082 | tp->prior_ssthresh = 0; |
3083 | |
3084 | /* B. In all the states check for reneging SACKs. */ |
3085 | if (tcp_check_sack_reneging(sk, ack_flag)) |
3086 | return; |
3087 | |
3088 | /* C. Check consistency of the current state. */ |
3089 | tcp_verify_left_out(tp); |
3090 | |
3091 | /* D. Check state exit conditions. State can be terminated |
3092 | * when high_seq is ACKed. */ |
3093 | if (icsk->icsk_ca_state == TCP_CA_Open) { |
3094 | WARN_ON(tp->retrans_out != 0 && !tp->syn_data); |
3095 | tp->retrans_stamp = 0; |
3096 | } else if (!before(seq1: tp->snd_una, seq2: tp->high_seq)) { |
3097 | switch (icsk->icsk_ca_state) { |
3098 | case TCP_CA_CWR: |
3099 | /* CWR is to be held something *above* high_seq |
3100 | * is ACKed for CWR bit to reach receiver. */ |
3101 | if (tp->snd_una != tp->high_seq) { |
3102 | tcp_end_cwnd_reduction(sk); |
3103 | tcp_set_ca_state(sk, ca_state: TCP_CA_Open); |
3104 | } |
3105 | break; |
3106 | |
3107 | case TCP_CA_Recovery: |
3108 | if (tcp_is_reno(tp)) |
3109 | tcp_reset_reno_sack(tp); |
3110 | if (tcp_try_undo_recovery(sk)) |
3111 | return; |
3112 | tcp_end_cwnd_reduction(sk); |
3113 | break; |
3114 | } |
3115 | } |
3116 | |
3117 | /* E. Process state. */ |
3118 | switch (icsk->icsk_ca_state) { |
3119 | case TCP_CA_Recovery: |
3120 | if (!(flag & FLAG_SND_UNA_ADVANCED)) { |
3121 | if (tcp_is_reno(tp)) |
3122 | tcp_add_reno_sack(sk, num_dupack, ece_ack); |
3123 | } else if (tcp_try_undo_partial(sk, prior_snd_una, do_lost: &do_lost)) |
3124 | return; |
3125 | |
3126 | if (tcp_try_undo_dsack(sk)) |
3127 | tcp_try_to_open(sk, flag); |
3128 | |
3129 | tcp_identify_packet_loss(sk, ack_flag); |
3130 | if (icsk->icsk_ca_state != TCP_CA_Recovery) { |
3131 | if (!tcp_time_to_recover(sk, flag)) |
3132 | return; |
3133 | /* Undo reverts the recovery state. If loss is evident, |
3134 | * starts a new recovery (e.g. reordering then loss); |
3135 | */ |
3136 | tcp_enter_recovery(sk, ece_ack); |
3137 | } |
3138 | break; |
3139 | case TCP_CA_Loss: |
3140 | tcp_process_loss(sk, flag, num_dupack, rexmit); |
3141 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
3142 | tcp_update_rto_time(tp); |
3143 | tcp_identify_packet_loss(sk, ack_flag); |
3144 | if (!(icsk->icsk_ca_state == TCP_CA_Open || |
3145 | (*ack_flag & FLAG_LOST_RETRANS))) |
3146 | return; |
3147 | /* Change state if cwnd is undone or retransmits are lost */ |
3148 | fallthrough; |
3149 | default: |
3150 | if (tcp_is_reno(tp)) { |
3151 | if (flag & FLAG_SND_UNA_ADVANCED) |
3152 | tcp_reset_reno_sack(tp); |
3153 | tcp_add_reno_sack(sk, num_dupack, ece_ack); |
3154 | } |
3155 | |
3156 | if (icsk->icsk_ca_state <= TCP_CA_Disorder) |
3157 | tcp_try_undo_dsack(sk); |
3158 | |
3159 | tcp_identify_packet_loss(sk, ack_flag); |
3160 | if (!tcp_time_to_recover(sk, flag)) { |
3161 | tcp_try_to_open(sk, flag); |
3162 | return; |
3163 | } |
3164 | |
3165 | /* MTU probe failure: don't reduce cwnd */ |
3166 | if (icsk->icsk_ca_state < TCP_CA_CWR && |
3167 | icsk->icsk_mtup.probe_size && |
3168 | tp->snd_una == tp->mtu_probe.probe_seq_start) { |
3169 | tcp_mtup_probe_failed(sk); |
3170 | /* Restores the reduction we did in tcp_mtup_probe() */ |
3171 | tcp_snd_cwnd_set(tp, val: tcp_snd_cwnd(tp) + 1); |
3172 | tcp_simple_retransmit(sk); |
3173 | return; |
3174 | } |
3175 | |
3176 | /* Otherwise enter Recovery state */ |
3177 | tcp_enter_recovery(sk, ece_ack); |
3178 | fast_rexmit = 1; |
3179 | } |
3180 | |
3181 | if (!tcp_is_rack(sk) && do_lost) |
3182 | tcp_update_scoreboard(sk, fast_rexmit); |
3183 | *rexmit = REXMIT_LOST; |
3184 | } |
3185 | |
3186 | static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag) |
3187 | { |
3188 | u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ; |
3189 | struct tcp_sock *tp = tcp_sk(sk); |
3190 | |
3191 | if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) { |
3192 | /* If the remote keeps returning delayed ACKs, eventually |
3193 | * the min filter would pick it up and overestimate the |
3194 | * prop. delay when it expires. Skip suspected delayed ACKs. |
3195 | */ |
3196 | return; |
3197 | } |
3198 | minmax_running_min(m: &tp->rtt_min, win: wlen, tcp_jiffies32, |
3199 | meas: rtt_us ? : jiffies_to_usecs(j: 1)); |
3200 | } |
3201 | |
3202 | static bool tcp_ack_update_rtt(struct sock *sk, const int flag, |
3203 | long seq_rtt_us, long sack_rtt_us, |
3204 | long ca_rtt_us, struct rate_sample *rs) |
3205 | { |
3206 | const struct tcp_sock *tp = tcp_sk(sk); |
3207 | |
3208 | /* Prefer RTT measured from ACK's timing to TS-ECR. This is because |
3209 | * broken middle-boxes or peers may corrupt TS-ECR fields. But |
3210 | * Karn's algorithm forbids taking RTT if some retransmitted data |
3211 | * is acked (RFC6298). |
3212 | */ |
3213 | if (seq_rtt_us < 0) |
3214 | seq_rtt_us = sack_rtt_us; |
3215 | |
3216 | /* RTTM Rule: A TSecr value received in a segment is used to |
3217 | * update the averaged RTT measurement only if the segment |
3218 | * acknowledges some new data, i.e., only if it advances the |
3219 | * left edge of the send window. |
3220 | * See draft-ietf-tcplw-high-performance-00, section 3.3. |
3221 | */ |
3222 | if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && |
3223 | tp->rx_opt.rcv_tsecr && flag & FLAG_ACKED) |
3224 | seq_rtt_us = ca_rtt_us = tcp_rtt_tsopt_us(tp, min_delta: 1); |
3225 | |
3226 | rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */ |
3227 | if (seq_rtt_us < 0) |
3228 | return false; |
3229 | |
3230 | /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is |
3231 | * always taken together with ACK, SACK, or TS-opts. Any negative |
3232 | * values will be skipped with the seq_rtt_us < 0 check above. |
3233 | */ |
3234 | tcp_update_rtt_min(sk, rtt_us: ca_rtt_us, flag); |
3235 | tcp_rtt_estimator(sk, mrtt_us: seq_rtt_us); |
3236 | tcp_set_rto(sk); |
3237 | |
3238 | /* RFC6298: only reset backoff on valid RTT measurement. */ |
3239 | inet_csk(sk)->icsk_backoff = 0; |
3240 | return true; |
3241 | } |
3242 | |
3243 | /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ |
3244 | void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req) |
3245 | { |
3246 | struct rate_sample rs; |
3247 | long rtt_us = -1L; |
3248 | |
3249 | if (req && !req->num_retrans && tcp_rsk(req)->snt_synack) |
3250 | rtt_us = tcp_stamp_us_delta(t1: tcp_clock_us(), t0: tcp_rsk(req)->snt_synack); |
3251 | |
3252 | tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt_us: rtt_us, sack_rtt_us: -1L, ca_rtt_us: rtt_us, rs: &rs); |
3253 | } |
3254 | |
3255 | |
3256 | static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) |
3257 | { |
3258 | const struct inet_connection_sock *icsk = inet_csk(sk); |
3259 | |
3260 | icsk->icsk_ca_ops->cong_avoid(sk, ack, acked); |
3261 | tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32; |
3262 | } |
3263 | |
3264 | /* Restart timer after forward progress on connection. |
3265 | * RFC2988 recommends to restart timer to now+rto. |
3266 | */ |
3267 | void tcp_rearm_rto(struct sock *sk) |
3268 | { |
3269 | const struct inet_connection_sock *icsk = inet_csk(sk); |
3270 | struct tcp_sock *tp = tcp_sk(sk); |
3271 | |
3272 | /* If the retrans timer is currently being used by Fast Open |
3273 | * for SYN-ACK retrans purpose, stay put. |
3274 | */ |
3275 | if (rcu_access_pointer(tp->fastopen_rsk)) |
3276 | return; |
3277 | |
3278 | if (!tp->packets_out) { |
3279 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); |
3280 | } else { |
3281 | u32 rto = inet_csk(sk)->icsk_rto; |
3282 | /* Offset the time elapsed after installing regular RTO */ |
3283 | if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT || |
3284 | icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { |
3285 | s64 delta_us = tcp_rto_delta_us(sk); |
3286 | /* delta_us may not be positive if the socket is locked |
3287 | * when the retrans timer fires and is rescheduled. |
3288 | */ |
3289 | rto = usecs_to_jiffies(max_t(int, delta_us, 1)); |
3290 | } |
3291 | tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, when: rto, pace_delay: true); |
3292 | } |
3293 | } |
3294 | |
3295 | /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */ |
3296 | static void tcp_set_xmit_timer(struct sock *sk) |
3297 | { |
3298 | if (!tcp_schedule_loss_probe(sk, advancing_rto: true)) |
3299 | tcp_rearm_rto(sk); |
3300 | } |
3301 | |
3302 | /* If we get here, the whole TSO packet has not been acked. */ |
3303 | static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) |
3304 | { |
3305 | struct tcp_sock *tp = tcp_sk(sk); |
3306 | u32 packets_acked; |
3307 | |
3308 | BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); |
3309 | |
3310 | packets_acked = tcp_skb_pcount(skb); |
3311 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
3312 | return 0; |
3313 | packets_acked -= tcp_skb_pcount(skb); |
3314 | |
3315 | if (packets_acked) { |
3316 | BUG_ON(tcp_skb_pcount(skb) == 0); |
3317 | BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); |
3318 | } |
3319 | |
3320 | return packets_acked; |
3321 | } |
3322 | |
3323 | static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb, |
3324 | const struct sk_buff *ack_skb, u32 prior_snd_una) |
3325 | { |
3326 | const struct skb_shared_info *shinfo; |
3327 | |
3328 | /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */ |
3329 | if (likely(!TCP_SKB_CB(skb)->txstamp_ack)) |
3330 | return; |
3331 | |
3332 | shinfo = skb_shinfo(skb); |
3333 | if (!before(seq1: shinfo->tskey, seq2: prior_snd_una) && |
3334 | before(seq1: shinfo->tskey, tcp_sk(sk)->snd_una)) { |
3335 | tcp_skb_tsorted_save(skb) { |
3336 | __skb_tstamp_tx(orig_skb: skb, ack_skb, NULL, sk, tstype: SCM_TSTAMP_ACK); |
3337 | } tcp_skb_tsorted_restore(skb); |
3338 | } |
3339 | } |
3340 | |
3341 | /* Remove acknowledged frames from the retransmission queue. If our packet |
3342 | * is before the ack sequence we can discard it as it's confirmed to have |
3343 | * arrived at the other end. |
3344 | */ |
3345 | static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb, |
3346 | u32 prior_fack, u32 prior_snd_una, |
3347 | struct tcp_sacktag_state *sack, bool ece_ack) |
3348 | { |
3349 | const struct inet_connection_sock *icsk = inet_csk(sk); |
3350 | u64 first_ackt, last_ackt; |
3351 | struct tcp_sock *tp = tcp_sk(sk); |
3352 | u32 prior_sacked = tp->sacked_out; |
3353 | u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */ |
3354 | struct sk_buff *skb, *next; |
3355 | bool fully_acked = true; |
3356 | long sack_rtt_us = -1L; |
3357 | long seq_rtt_us = -1L; |
3358 | long ca_rtt_us = -1L; |
3359 | u32 pkts_acked = 0; |
3360 | bool rtt_update; |
3361 | int flag = 0; |
3362 | |
3363 | first_ackt = 0; |
3364 | |
3365 | for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) { |
3366 | struct tcp_skb_cb *scb = TCP_SKB_CB(skb); |
3367 | const u32 start_seq = scb->seq; |
3368 | u8 sacked = scb->sacked; |
3369 | u32 acked_pcount; |
3370 | |
3371 | /* Determine how many packets and what bytes were acked, tso and else */ |
3372 | if (after(scb->end_seq, tp->snd_una)) { |
3373 | if (tcp_skb_pcount(skb) == 1 || |
3374 | !after(tp->snd_una, scb->seq)) |
3375 | break; |
3376 | |
3377 | acked_pcount = tcp_tso_acked(sk, skb); |
3378 | if (!acked_pcount) |
3379 | break; |
3380 | fully_acked = false; |
3381 | } else { |
3382 | acked_pcount = tcp_skb_pcount(skb); |
3383 | } |
3384 | |
3385 | if (unlikely(sacked & TCPCB_RETRANS)) { |
3386 | if (sacked & TCPCB_SACKED_RETRANS) |
3387 | tp->retrans_out -= acked_pcount; |
3388 | flag |= FLAG_RETRANS_DATA_ACKED; |
3389 | } else if (!(sacked & TCPCB_SACKED_ACKED)) { |
3390 | last_ackt = tcp_skb_timestamp_us(skb); |
3391 | WARN_ON_ONCE(last_ackt == 0); |
3392 | if (!first_ackt) |
3393 | first_ackt = last_ackt; |
3394 | |
3395 | if (before(seq1: start_seq, seq2: reord)) |
3396 | reord = start_seq; |
3397 | if (!after(scb->end_seq, tp->high_seq)) |
3398 | flag |= FLAG_ORIG_SACK_ACKED; |
3399 | } |
3400 | |
3401 | if (sacked & TCPCB_SACKED_ACKED) { |
3402 | tp->sacked_out -= acked_pcount; |
3403 | } else if (tcp_is_sack(tp)) { |
3404 | tcp_count_delivered(tp, delivered: acked_pcount, ece_ack); |
3405 | if (!tcp_skb_spurious_retrans(tp, skb)) |
3406 | tcp_rack_advance(tp, sacked, end_seq: scb->end_seq, |
3407 | xmit_time: tcp_skb_timestamp_us(skb)); |
3408 | } |
3409 | if (sacked & TCPCB_LOST) |
3410 | tp->lost_out -= acked_pcount; |
3411 | |
3412 | tp->packets_out -= acked_pcount; |
3413 | pkts_acked += acked_pcount; |
3414 | tcp_rate_skb_delivered(sk, skb, rs: sack->rate); |
3415 | |
3416 | /* Initial outgoing SYN's get put onto the write_queue |
3417 | * just like anything else we transmit. It is not |
3418 | * true data, and if we misinform our callers that |
3419 | * this ACK acks real data, we will erroneously exit |
3420 | * connection startup slow start one packet too |
3421 | * quickly. This is severely frowned upon behavior. |
3422 | */ |
3423 | if (likely(!(scb->tcp_flags & TCPHDR_SYN))) { |
3424 | flag |= FLAG_DATA_ACKED; |
3425 | } else { |
3426 | flag |= FLAG_SYN_ACKED; |
3427 | tp->retrans_stamp = 0; |
3428 | } |
3429 | |
3430 | if (!fully_acked) |
3431 | break; |
3432 | |
3433 | tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una); |
3434 | |
3435 | next = skb_rb_next(skb); |
3436 | if (unlikely(skb == tp->retransmit_skb_hint)) |
3437 | tp->retransmit_skb_hint = NULL; |
3438 | if (unlikely(skb == tp->lost_skb_hint)) |
3439 | tp->lost_skb_hint = NULL; |
3440 | tcp_highest_sack_replace(sk, old: skb, new: next); |
3441 | tcp_rtx_queue_unlink_and_free(skb, sk); |
3442 | } |
3443 | |
3444 | if (!skb) |
3445 | tcp_chrono_stop(sk, type: TCP_CHRONO_BUSY); |
3446 | |
3447 | if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) |
3448 | tp->snd_up = tp->snd_una; |
3449 | |
3450 | if (skb) { |
3451 | tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una); |
3452 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
3453 | flag |= FLAG_SACK_RENEGING; |
3454 | } |
3455 | |
3456 | if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) { |
3457 | seq_rtt_us = tcp_stamp_us_delta(t1: tp->tcp_mstamp, t0: first_ackt); |
3458 | ca_rtt_us = tcp_stamp_us_delta(t1: tp->tcp_mstamp, t0: last_ackt); |
3459 | |
3460 | if (pkts_acked == 1 && fully_acked && !prior_sacked && |
3461 | (tp->snd_una - prior_snd_una) < tp->mss_cache && |
3462 | sack->rate->prior_delivered + 1 == tp->delivered && |
3463 | !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) { |
3464 | /* Conservatively mark a delayed ACK. It's typically |
3465 | * from a lone runt packet over the round trip to |
3466 | * a receiver w/o out-of-order or CE events. |
3467 | */ |
3468 | flag |= FLAG_ACK_MAYBE_DELAYED; |
3469 | } |
3470 | } |
3471 | if (sack->first_sackt) { |
3472 | sack_rtt_us = tcp_stamp_us_delta(t1: tp->tcp_mstamp, t0: sack->first_sackt); |
3473 | ca_rtt_us = tcp_stamp_us_delta(t1: tp->tcp_mstamp, t0: sack->last_sackt); |
3474 | } |
3475 | rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us, |
3476 | ca_rtt_us, rs: sack->rate); |
3477 | |
3478 | if (flag & FLAG_ACKED) { |
3479 | flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ |
3480 | if (unlikely(icsk->icsk_mtup.probe_size && |
3481 | !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { |
3482 | tcp_mtup_probe_success(sk); |
3483 | } |
3484 | |
3485 | if (tcp_is_reno(tp)) { |
3486 | tcp_remove_reno_sacks(sk, acked: pkts_acked, ece_ack); |
3487 | |
3488 | /* If any of the cumulatively ACKed segments was |
3489 | * retransmitted, non-SACK case cannot confirm that |
3490 | * progress was due to original transmission due to |
3491 | * lack of TCPCB_SACKED_ACKED bits even if some of |
3492 | * the packets may have been never retransmitted. |
3493 | */ |
3494 | if (flag & FLAG_RETRANS_DATA_ACKED) |
3495 | flag &= ~FLAG_ORIG_SACK_ACKED; |
3496 | } else { |
3497 | int delta; |
3498 | |
3499 | /* Non-retransmitted hole got filled? That's reordering */ |
3500 | if (before(seq1: reord, seq2: prior_fack)) |
3501 | tcp_check_sack_reordering(sk, low_seq: reord, ts: 0); |
3502 | |
3503 | delta = prior_sacked - tp->sacked_out; |
3504 | tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); |
3505 | } |
3506 | } else if (skb && rtt_update && sack_rtt_us >= 0 && |
3507 | sack_rtt_us > tcp_stamp_us_delta(t1: tp->tcp_mstamp, |
3508 | t0: tcp_skb_timestamp_us(skb))) { |
3509 | /* Do not re-arm RTO if the sack RTT is measured from data sent |
3510 | * after when the head was last (re)transmitted. Otherwise the |
3511 | * timeout may continue to extend in loss recovery. |
3512 | */ |
3513 | flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ |
3514 | } |
3515 | |
3516 | if (icsk->icsk_ca_ops->pkts_acked) { |
3517 | struct ack_sample sample = { .pkts_acked = pkts_acked, |
3518 | .rtt_us = sack->rate->rtt_us }; |
3519 | |
3520 | sample.in_flight = tp->mss_cache * |
3521 | (tp->delivered - sack->rate->prior_delivered); |
3522 | icsk->icsk_ca_ops->pkts_acked(sk, &sample); |
3523 | } |
3524 | |
3525 | #if FASTRETRANS_DEBUG > 0 |
3526 | WARN_ON((int)tp->sacked_out < 0); |
3527 | WARN_ON((int)tp->lost_out < 0); |
3528 | WARN_ON((int)tp->retrans_out < 0); |
3529 | if (!tp->packets_out && tcp_is_sack(tp)) { |
3530 | icsk = inet_csk(sk); |
3531 | if (tp->lost_out) { |
3532 | pr_debug("Leak l=%u %d\n", |
3533 | tp->lost_out, icsk->icsk_ca_state); |
3534 | tp->lost_out = 0; |
3535 | } |
3536 | if (tp->sacked_out) { |
3537 | pr_debug("Leak s=%u %d\n", |
3538 | tp->sacked_out, icsk->icsk_ca_state); |
3539 | tp->sacked_out = 0; |
3540 | } |
3541 | if (tp->retrans_out) { |
3542 | pr_debug("Leak r=%u %d\n", |
3543 | tp->retrans_out, icsk->icsk_ca_state); |
3544 | tp->retrans_out = 0; |
3545 | } |
3546 | } |
3547 | #endif |
3548 | return flag; |
3549 | } |
3550 | |
3551 | static void tcp_ack_probe(struct sock *sk) |
3552 | { |
3553 | struct inet_connection_sock *icsk = inet_csk(sk); |
3554 | struct sk_buff *head = tcp_send_head(sk); |
3555 | const struct tcp_sock *tp = tcp_sk(sk); |
3556 | |
3557 | /* Was it a usable window open? */ |
3558 | if (!head) |
3559 | return; |
3560 | if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) { |
3561 | icsk->icsk_backoff = 0; |
3562 | icsk->icsk_probes_tstamp = 0; |
3563 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); |
3564 | /* Socket must be waked up by subsequent tcp_data_snd_check(). |
3565 | * This function is not for random using! |
3566 | */ |
3567 | } else { |
3568 | unsigned long when = tcp_probe0_when(sk, max_when: tcp_rto_max(sk)); |
3569 | |
3570 | when = tcp_clamp_probe0_to_user_timeout(sk, when); |
3571 | tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, pace_delay: true); |
3572 | } |
3573 | } |
3574 | |
3575 | static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag) |
3576 | { |
3577 | return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || |
3578 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open; |
3579 | } |
3580 | |
3581 | /* Decide wheather to run the increase function of congestion control. */ |
3582 | static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) |
3583 | { |
3584 | /* If reordering is high then always grow cwnd whenever data is |
3585 | * delivered regardless of its ordering. Otherwise stay conservative |
3586 | * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/ |
3587 | * new SACK or ECE mark may first advance cwnd here and later reduce |
3588 | * cwnd in tcp_fastretrans_alert() based on more states. |
3589 | */ |
3590 | if (tcp_sk(sk)->reordering > |
3591 | READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering)) |
3592 | return flag & FLAG_FORWARD_PROGRESS; |
3593 | |
3594 | return flag & FLAG_DATA_ACKED; |
3595 | } |
3596 | |
3597 | /* The "ultimate" congestion control function that aims to replace the rigid |
3598 | * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction). |
3599 | * It's called toward the end of processing an ACK with precise rate |
3600 | * information. All transmission or retransmission are delayed afterwards. |
3601 | */ |
3602 | static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked, |
3603 | int flag, const struct rate_sample *rs) |
3604 | { |
3605 | const struct inet_connection_sock *icsk = inet_csk(sk); |
3606 | |
3607 | if (icsk->icsk_ca_ops->cong_control) { |
3608 | icsk->icsk_ca_ops->cong_control(sk, ack, flag, rs); |
3609 | return; |
3610 | } |
3611 | |
3612 | if (tcp_in_cwnd_reduction(sk)) { |
3613 | /* Reduce cwnd if state mandates */ |
3614 | tcp_cwnd_reduction(sk, newly_acked_sacked: acked_sacked, newly_lost: rs->losses, flag); |
3615 | } else if (tcp_may_raise_cwnd(sk, flag)) { |
3616 | /* Advance cwnd if state allows */ |
3617 | tcp_cong_avoid(sk, ack, acked: acked_sacked); |
3618 | } |
3619 | tcp_update_pacing_rate(sk); |
3620 | } |
3621 | |
3622 | /* Check that window update is acceptable. |
3623 | * The function assumes that snd_una<=ack<=snd_next. |
3624 | */ |
3625 | static inline bool tcp_may_update_window(const struct tcp_sock *tp, |
3626 | const u32 ack, const u32 ack_seq, |
3627 | const u32 nwin) |
3628 | { |
3629 | return after(ack, tp->snd_una) || |
3630 | after(ack_seq, tp->snd_wl1) || |
3631 | (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin)); |
3632 | } |
3633 | |
3634 | static void tcp_snd_sne_update(struct tcp_sock *tp, u32 ack) |
3635 | { |
3636 | #ifdef CONFIG_TCP_AO |
3637 | struct tcp_ao_info *ao; |
3638 | |
3639 | if (!static_branch_unlikely(&tcp_ao_needed.key)) |
3640 | return; |
3641 | |
3642 | ao = rcu_dereference_protected(tp->ao_info, |
3643 | lockdep_sock_is_held((struct sock *)tp)); |
3644 | if (ao && ack < tp->snd_una) { |
3645 | ao->snd_sne++; |
3646 | trace_tcp_ao_snd_sne_update(sk: (struct sock *)tp, new_sne: ao->snd_sne); |
3647 | } |
3648 | #endif |
3649 | } |
3650 | |
3651 | /* If we update tp->snd_una, also update tp->bytes_acked */ |
3652 | static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack) |
3653 | { |
3654 | u32 delta = ack - tp->snd_una; |
3655 | |
3656 | sock_owned_by_me(sk: (struct sock *)tp); |
3657 | tp->bytes_acked += delta; |
3658 | tcp_snd_sne_update(tp, ack); |
3659 | tp->snd_una = ack; |
3660 | } |
3661 | |
3662 | static void tcp_rcv_sne_update(struct tcp_sock *tp, u32 seq) |
3663 | { |
3664 | #ifdef CONFIG_TCP_AO |
3665 | struct tcp_ao_info *ao; |
3666 | |
3667 | if (!static_branch_unlikely(&tcp_ao_needed.key)) |
3668 | return; |
3669 | |
3670 | ao = rcu_dereference_protected(tp->ao_info, |
3671 | lockdep_sock_is_held((struct sock *)tp)); |
3672 | if (ao && seq < tp->rcv_nxt) { |
3673 | ao->rcv_sne++; |
3674 | trace_tcp_ao_rcv_sne_update(sk: (struct sock *)tp, new_sne: ao->rcv_sne); |
3675 | } |
3676 | #endif |
3677 | } |
3678 | |
3679 | /* If we update tp->rcv_nxt, also update tp->bytes_received */ |
3680 | static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq) |
3681 | { |
3682 | u32 delta = seq - tp->rcv_nxt; |
3683 | |
3684 | sock_owned_by_me(sk: (struct sock *)tp); |
3685 | tp->bytes_received += delta; |
3686 | tcp_rcv_sne_update(tp, seq); |
3687 | WRITE_ONCE(tp->rcv_nxt, seq); |
3688 | } |
3689 | |
3690 | /* Update our send window. |
3691 | * |
3692 | * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 |
3693 | * and in FreeBSD. NetBSD's one is even worse.) is wrong. |
3694 | */ |
3695 | static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, |
3696 | u32 ack_seq) |
3697 | { |
3698 | struct tcp_sock *tp = tcp_sk(sk); |
3699 | int flag = 0; |
3700 | u32 nwin = ntohs(tcp_hdr(skb)->window); |
3701 | |
3702 | if (likely(!tcp_hdr(skb)->syn)) |
3703 | nwin <<= tp->rx_opt.snd_wscale; |
3704 | |
3705 | if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { |
3706 | flag |= FLAG_WIN_UPDATE; |
3707 | tcp_update_wl(tp, seq: ack_seq); |
3708 | |
3709 | if (tp->snd_wnd != nwin) { |
3710 | tp->snd_wnd = nwin; |
3711 | |
3712 | /* Note, it is the only place, where |
3713 | * fast path is recovered for sending TCP. |
3714 | */ |
3715 | tp->pred_flags = 0; |
3716 | tcp_fast_path_check(sk); |
3717 | |
3718 | if (!tcp_write_queue_empty(sk)) |
3719 | tcp_slow_start_after_idle_check(sk); |
3720 | |
3721 | if (nwin > tp->max_window) { |
3722 | tp->max_window = nwin; |
3723 | tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); |
3724 | } |
3725 | } |
3726 | } |
3727 | |
3728 | tcp_snd_una_update(tp, ack); |
3729 | |
3730 | return flag; |
3731 | } |
3732 | |
3733 | static bool __tcp_oow_rate_limited(struct net *net, int mib_idx, |
3734 | u32 *last_oow_ack_time) |
3735 | { |
3736 | /* Paired with the WRITE_ONCE() in this function. */ |
3737 | u32 val = READ_ONCE(*last_oow_ack_time); |
3738 | |
3739 | if (val) { |
3740 | s32 elapsed = (s32)(tcp_jiffies32 - val); |
3741 | |
3742 | if (0 <= elapsed && |
3743 | elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) { |
3744 | NET_INC_STATS(net, mib_idx); |
3745 | return true; /* rate-limited: don't send yet! */ |
3746 | } |
3747 | } |
3748 | |
3749 | /* Paired with the prior READ_ONCE() and with itself, |
3750 | * as we might be lockless. |
3751 | */ |
3752 | WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32); |
3753 | |
3754 | return false; /* not rate-limited: go ahead, send dupack now! */ |
3755 | } |
3756 | |
3757 | /* Return true if we're currently rate-limiting out-of-window ACKs and |
3758 | * thus shouldn't send a dupack right now. We rate-limit dupacks in |
3759 | * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS |
3760 | * attacks that send repeated SYNs or ACKs for the same connection. To |
3761 | * do this, we do not send a duplicate SYNACK or ACK if the remote |
3762 | * endpoint is sending out-of-window SYNs or pure ACKs at a high rate. |
3763 | */ |
3764 | bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb, |
3765 | int mib_idx, u32 *last_oow_ack_time) |
3766 | { |
3767 | /* Data packets without SYNs are not likely part of an ACK loop. */ |
3768 | if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) && |
3769 | !tcp_hdr(skb)->syn) |
3770 | return false; |
3771 | |
3772 | return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time); |
3773 | } |
3774 | |
3775 | /* RFC 5961 7 [ACK Throttling] */ |
3776 | static void tcp_send_challenge_ack(struct sock *sk) |
3777 | { |
3778 | struct tcp_sock *tp = tcp_sk(sk); |
3779 | struct net *net = sock_net(sk); |
3780 | u32 count, now, ack_limit; |
3781 | |
3782 | /* First check our per-socket dupack rate limit. */ |
3783 | if (__tcp_oow_rate_limited(net, |
3784 | mib_idx: LINUX_MIB_TCPACKSKIPPEDCHALLENGE, |
3785 | last_oow_ack_time: &tp->last_oow_ack_time)) |
3786 | return; |
3787 | |
3788 | ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit); |
3789 | if (ack_limit == INT_MAX) |
3790 | goto send_ack; |
3791 | |
3792 | /* Then check host-wide RFC 5961 rate limit. */ |
3793 | now = jiffies / HZ; |
3794 | if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) { |
3795 | u32 half = (ack_limit + 1) >> 1; |
3796 | |
3797 | WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now); |
3798 | WRITE_ONCE(net->ipv4.tcp_challenge_count, |
3799 | get_random_u32_inclusive(half, ack_limit + half - 1)); |
3800 | } |
3801 | count = READ_ONCE(net->ipv4.tcp_challenge_count); |
3802 | if (count > 0) { |
3803 | WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1); |
3804 | send_ack: |
3805 | NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK); |
3806 | tcp_send_ack(sk); |
3807 | } |
3808 | } |
3809 | |
3810 | static void tcp_store_ts_recent(struct tcp_sock *tp) |
3811 | { |
3812 | tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; |
3813 | tp->rx_opt.ts_recent_stamp = ktime_get_seconds(); |
3814 | } |
3815 | |
3816 | static int __tcp_replace_ts_recent(struct tcp_sock *tp, s32 tstamp_delta) |
3817 | { |
3818 | tcp_store_ts_recent(tp); |
3819 | return tstamp_delta > 0 ? FLAG_TS_PROGRESS : 0; |
3820 | } |
3821 | |
3822 | static int tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) |
3823 | { |
3824 | s32 delta; |
3825 | |
3826 | if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { |
3827 | /* PAWS bug workaround wrt. ACK frames, the PAWS discard |
3828 | * extra check below makes sure this can only happen |
3829 | * for pure ACK frames. -DaveM |
3830 | * |
3831 | * Not only, also it occurs for expired timestamps. |
3832 | */ |
3833 | |
3834 | if (tcp_paws_check(rx_opt: &tp->rx_opt, paws_win: 0)) { |
3835 | delta = tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent; |
3836 | return __tcp_replace_ts_recent(tp, tstamp_delta: delta); |
3837 | } |
3838 | } |
3839 | |
3840 | return 0; |
3841 | } |
3842 | |
3843 | /* This routine deals with acks during a TLP episode and ends an episode by |
3844 | * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack |
3845 | */ |
3846 | static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) |
3847 | { |
3848 | struct tcp_sock *tp = tcp_sk(sk); |
3849 | |
3850 | if (before(seq1: ack, seq2: tp->tlp_high_seq)) |
3851 | return; |
3852 | |
3853 | if (!tp->tlp_retrans) { |
3854 | /* TLP of new data has been acknowledged */ |
3855 | tp->tlp_high_seq = 0; |
3856 | } else if (flag & FLAG_DSACK_TLP) { |
3857 | /* This DSACK means original and TLP probe arrived; no loss */ |
3858 | tp->tlp_high_seq = 0; |
3859 | } else if (after(ack, tp->tlp_high_seq)) { |
3860 | /* ACK advances: there was a loss, so reduce cwnd. Reset |
3861 | * tlp_high_seq in tcp_init_cwnd_reduction() |
3862 | */ |
3863 | tcp_init_cwnd_reduction(sk); |
3864 | tcp_set_ca_state(sk, ca_state: TCP_CA_CWR); |
3865 | tcp_end_cwnd_reduction(sk); |
3866 | tcp_try_keep_open(sk); |
3867 | NET_INC_STATS(sock_net(sk), |
3868 | LINUX_MIB_TCPLOSSPROBERECOVERY); |
3869 | } else if (!(flag & (FLAG_SND_UNA_ADVANCED | |
3870 | FLAG_NOT_DUP | FLAG_DATA_SACKED))) { |
3871 | /* Pure dupack: original and TLP probe arrived; no loss */ |
3872 | tp->tlp_high_seq = 0; |
3873 | } |
3874 | } |
3875 | |
3876 | static void tcp_in_ack_event(struct sock *sk, int flag) |
3877 | { |
3878 | const struct inet_connection_sock *icsk = inet_csk(sk); |
3879 | |
3880 | if (icsk->icsk_ca_ops->in_ack_event) { |
3881 | u32 ack_ev_flags = 0; |
3882 | |
3883 | if (flag & FLAG_WIN_UPDATE) |
3884 | ack_ev_flags |= CA_ACK_WIN_UPDATE; |
3885 | if (flag & FLAG_SLOWPATH) { |
3886 | ack_ev_flags |= CA_ACK_SLOWPATH; |
3887 | if (flag & FLAG_ECE) |
3888 | ack_ev_flags |= CA_ACK_ECE; |
3889 | } |
3890 | |
3891 | icsk->icsk_ca_ops->in_ack_event(sk, ack_ev_flags); |
3892 | } |
3893 | } |
3894 | |
3895 | /* Congestion control has updated the cwnd already. So if we're in |
3896 | * loss recovery then now we do any new sends (for FRTO) or |
3897 | * retransmits (for CA_Loss or CA_recovery) that make sense. |
3898 | */ |
3899 | static void tcp_xmit_recovery(struct sock *sk, int rexmit) |
3900 | { |
3901 | struct tcp_sock *tp = tcp_sk(sk); |
3902 | |
3903 | if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT) |
3904 | return; |
3905 | |
3906 | if (unlikely(rexmit == REXMIT_NEW)) { |
3907 | __tcp_push_pending_frames(sk, cur_mss: tcp_current_mss(sk), |
3908 | TCP_NAGLE_OFF); |
3909 | if (after(tp->snd_nxt, tp->high_seq)) |
3910 | return; |
3911 | tp->frto = 0; |
3912 | } |
3913 | tcp_xmit_retransmit_queue(sk); |
3914 | } |
3915 | |
3916 | /* Returns the number of packets newly acked or sacked by the current ACK */ |
3917 | static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag) |
3918 | { |
3919 | const struct net *net = sock_net(sk); |
3920 | struct tcp_sock *tp = tcp_sk(sk); |
3921 | u32 delivered; |
3922 | |
3923 | delivered = tp->delivered - prior_delivered; |
3924 | NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered); |
3925 | if (flag & FLAG_ECE) |
3926 | NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered); |
3927 | |
3928 | return delivered; |
3929 | } |
3930 | |
3931 | /* This routine deals with incoming acks, but not outgoing ones. */ |
3932 | static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) |
3933 | { |
3934 | struct inet_connection_sock *icsk = inet_csk(sk); |
3935 | struct tcp_sock *tp = tcp_sk(sk); |
3936 | struct tcp_sacktag_state sack_state; |
3937 | struct rate_sample rs = { .prior_delivered = 0 }; |
3938 | u32 prior_snd_una = tp->snd_una; |
3939 | bool is_sack_reneg = tp->is_sack_reneg; |
3940 | u32 ack_seq = TCP_SKB_CB(skb)->seq; |
3941 | u32 ack = TCP_SKB_CB(skb)->ack_seq; |
3942 | int num_dupack = 0; |
3943 | int prior_packets = tp->packets_out; |
3944 | u32 delivered = tp->delivered; |
3945 | u32 lost = tp->lost; |
3946 | int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */ |
3947 | u32 prior_fack; |
3948 | |
3949 | sack_state.first_sackt = 0; |
3950 | sack_state.rate = &rs; |
3951 | sack_state.sack_delivered = 0; |
3952 | |
3953 | /* We very likely will need to access rtx queue. */ |
3954 | prefetch(sk->tcp_rtx_queue.rb_node); |
3955 | |
3956 | /* If the ack is older than previous acks |
3957 | * then we can probably ignore it. |
3958 | */ |
3959 | if (before(seq1: ack, seq2: prior_snd_una)) { |
3960 | u32 max_window; |
3961 | |
3962 | /* do not accept ACK for bytes we never sent. */ |
3963 | max_window = min_t(u64, tp->max_window, tp->bytes_acked); |
3964 | /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */ |
3965 | if (before(seq1: ack, seq2: prior_snd_una - max_window)) { |
3966 | if (!(flag & FLAG_NO_CHALLENGE_ACK)) |
3967 | tcp_send_challenge_ack(sk); |
3968 | return -SKB_DROP_REASON_TCP_TOO_OLD_ACK; |
3969 | } |
3970 | goto old_ack; |
3971 | } |
3972 | |
3973 | /* If the ack includes data we haven't sent yet, discard |
3974 | * this segment (RFC793 Section 3.9). |
3975 | */ |
3976 | if (after(ack, tp->snd_nxt)) |
3977 | return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA; |
3978 | |
3979 | if (after(ack, prior_snd_una)) { |
3980 | flag |= FLAG_SND_UNA_ADVANCED; |
3981 | icsk->icsk_retransmits = 0; |
3982 | |
3983 | #if IS_ENABLED(CONFIG_TLS_DEVICE) |
3984 | if (static_branch_unlikely(&clean_acked_data_enabled.key)) |
3985 | if (tp->tcp_clean_acked) |
3986 | tp->tcp_clean_acked(sk, ack); |
3987 | #endif |
3988 | } |
3989 | |
3990 | prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una; |
3991 | rs.prior_in_flight = tcp_packets_in_flight(tp); |
3992 | |
3993 | /* ts_recent update must be made after we are sure that the packet |
3994 | * is in window. |
3995 | */ |
3996 | if (flag & FLAG_UPDATE_TS_RECENT) |
3997 | flag |= tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); |
3998 | |
3999 | if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) == |
4000 | FLAG_SND_UNA_ADVANCED) { |
4001 | /* Window is constant, pure forward advance. |
4002 | * No more checks are required. |
4003 | * Note, we use the fact that SND.UNA>=SND.WL2. |
4004 | */ |
4005 | tcp_update_wl(tp, seq: ack_seq); |
4006 | tcp_snd_una_update(tp, ack); |
4007 | flag |= FLAG_WIN_UPDATE; |
4008 | |
4009 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS); |
4010 | } else { |
4011 | if (ack_seq != TCP_SKB_CB(skb)->end_seq) |
4012 | flag |= FLAG_DATA; |
4013 | else |
4014 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS); |
4015 | |
4016 | flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); |
4017 | |
4018 | if (TCP_SKB_CB(skb)->sacked) |
4019 | flag |= tcp_sacktag_write_queue(sk, ack_skb: skb, prior_snd_una, |
4020 | state: &sack_state); |
4021 | |
4022 | if (tcp_ecn_rcv_ecn_echo(tp, th: tcp_hdr(skb))) |
4023 | flag |= FLAG_ECE; |
4024 | |
4025 | if (sack_state.sack_delivered) |
4026 | tcp_count_delivered(tp, delivered: sack_state.sack_delivered, |
4027 | ece_ack: flag & FLAG_ECE); |
4028 | } |
4029 | |
4030 | /* This is a deviation from RFC3168 since it states that: |
4031 | * "When the TCP data sender is ready to set the CWR bit after reducing |
4032 | * the congestion window, it SHOULD set the CWR bit only on the first |
4033 | * new data packet that it transmits." |
4034 | * We accept CWR on pure ACKs to be more robust |
4035 | * with widely-deployed TCP implementations that do this. |
4036 | */ |
4037 | tcp_ecn_accept_cwr(sk, skb); |
4038 | |
4039 | /* We passed data and got it acked, remove any soft error |
4040 | * log. Something worked... |
4041 | */ |
4042 | WRITE_ONCE(sk->sk_err_soft, 0); |
4043 | icsk->icsk_probes_out = 0; |
4044 | tp->rcv_tstamp = tcp_jiffies32; |
4045 | if (!prior_packets) |
4046 | goto no_queue; |
4047 | |
4048 | /* See if we can take anything off of the retransmit queue. */ |
4049 | flag |= tcp_clean_rtx_queue(sk, ack_skb: skb, prior_fack, prior_snd_una, |
4050 | sack: &sack_state, ece_ack: flag & FLAG_ECE); |
4051 | |
4052 | tcp_rack_update_reo_wnd(sk, rs: &rs); |
4053 | |
4054 | tcp_in_ack_event(sk, flag); |
4055 | |
4056 | if (tp->tlp_high_seq) |
4057 | tcp_process_tlp_ack(sk, ack, flag); |
4058 | |
4059 | if (tcp_ack_is_dubious(sk, flag)) { |
4060 | if (!(flag & (FLAG_SND_UNA_ADVANCED | |
4061 | FLAG_NOT_DUP | FLAG_DSACKING_ACK))) { |
4062 | num_dupack = 1; |
4063 | /* Consider if pure acks were aggregated in tcp_add_backlog() */ |
4064 | if (!(flag & FLAG_DATA)) |
4065 | num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs); |
4066 | } |
4067 | tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, ack_flag: &flag, |
4068 | rexmit: &rexmit); |
4069 | } |
4070 | |
4071 | /* If needed, reset TLP/RTO timer when RACK doesn't set. */ |
4072 | if (flag & FLAG_SET_XMIT_TIMER) |
4073 | tcp_set_xmit_timer(sk); |
4074 | |
4075 | if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) |
4076 | sk_dst_confirm(sk); |
4077 | |
4078 | delivered = tcp_newly_delivered(sk, prior_delivered: delivered, flag); |
4079 | lost = tp->lost - lost; /* freshly marked lost */ |
4080 | rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED); |
4081 | tcp_rate_gen(sk, delivered, lost, is_sack_reneg, rs: sack_state.rate); |
4082 | tcp_cong_control(sk, ack, acked_sacked: delivered, flag, rs: sack_state.rate); |
4083 | tcp_xmit_recovery(sk, rexmit); |
4084 | return 1; |
4085 | |
4086 | no_queue: |
4087 | tcp_in_ack_event(sk, flag); |
4088 | /* If data was DSACKed, see if we can undo a cwnd reduction. */ |
4089 | if (flag & FLAG_DSACKING_ACK) { |
4090 | tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, ack_flag: &flag, |
4091 | rexmit: &rexmit); |
4092 | tcp_newly_delivered(sk, prior_delivered: delivered, flag); |
4093 | } |
4094 | /* If this ack opens up a zero window, clear backoff. It was |
4095 | * being used to time the probes, and is probably far higher than |
4096 | * it needs to be for normal retransmission. |
4097 | */ |
4098 | tcp_ack_probe(sk); |
4099 | |
4100 | if (tp->tlp_high_seq) |
4101 | tcp_process_tlp_ack(sk, ack, flag); |
4102 | return 1; |
4103 | |
4104 | old_ack: |
4105 | /* If data was SACKed, tag it and see if we should send more data. |
4106 | * If data was DSACKed, see if we can undo a cwnd reduction. |
4107 | */ |
4108 | if (TCP_SKB_CB(skb)->sacked) { |
4109 | flag |= tcp_sacktag_write_queue(sk, ack_skb: skb, prior_snd_una, |
4110 | state: &sack_state); |
4111 | tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, ack_flag: &flag, |
4112 | rexmit: &rexmit); |
4113 | tcp_newly_delivered(sk, prior_delivered: delivered, flag); |
4114 | tcp_xmit_recovery(sk, rexmit); |
4115 | } |
4116 | |
4117 | return 0; |
4118 | } |
4119 | |
4120 | static void tcp_parse_fastopen_option(int len, const unsigned char *cookie, |
4121 | bool syn, struct tcp_fastopen_cookie *foc, |
4122 | bool exp_opt) |
4123 | { |
4124 | /* Valid only in SYN or SYN-ACK with an even length. */ |
4125 | if (!foc || !syn || len < 0 || (len & 1)) |
4126 | return; |
4127 | |
4128 | if (len >= TCP_FASTOPEN_COOKIE_MIN && |
4129 | len <= TCP_FASTOPEN_COOKIE_MAX) |
4130 | memcpy(foc->val, cookie, len); |
4131 | else if (len != 0) |
4132 | len = -1; |
4133 | foc->len = len; |
4134 | foc->exp = exp_opt; |
4135 | } |
4136 | |
4137 | static bool smc_parse_options(const struct tcphdr *th, |
4138 | struct tcp_options_received *opt_rx, |
4139 | const unsigned char *ptr, |
4140 | int opsize) |
4141 | { |
4142 | #if IS_ENABLED(CONFIG_SMC) |
4143 | if (static_branch_unlikely(&tcp_have_smc)) { |
4144 | if (th->syn && !(opsize & 1) && |
4145 | opsize >= TCPOLEN_EXP_SMC_BASE && |
4146 | get_unaligned_be32(p: ptr) == TCPOPT_SMC_MAGIC) { |
4147 | opt_rx->smc_ok = 1; |
4148 | return true; |
4149 | } |
4150 | } |
4151 | #endif |
4152 | return false; |
4153 | } |
4154 | |
4155 | /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped |
4156 | * value on success. |
4157 | */ |
4158 | u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss) |
4159 | { |
4160 | const unsigned char *ptr = (const unsigned char *)(th + 1); |
4161 | int length = (th->doff * 4) - sizeof(struct tcphdr); |
4162 | u16 mss = 0; |
4163 | |
4164 | while (length > 0) { |
4165 | int opcode = *ptr++; |
4166 | int opsize; |
4167 | |
4168 | switch (opcode) { |
4169 | case TCPOPT_EOL: |
4170 | return mss; |
4171 | case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ |
4172 | length--; |
4173 | continue; |
4174 | default: |
4175 | if (length < 2) |
4176 | return mss; |
4177 | opsize = *ptr++; |
4178 | if (opsize < 2) /* "silly options" */ |
4179 | return mss; |
4180 | if (opsize > length) |
4181 | return mss; /* fail on partial options */ |
4182 | if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) { |
4183 | u16 in_mss = get_unaligned_be16(p: ptr); |
4184 | |
4185 | if (in_mss) { |
4186 | if (user_mss && user_mss < in_mss) |
4187 | in_mss = user_mss; |
4188 | mss = in_mss; |
4189 | } |
4190 | } |
4191 | ptr += opsize - 2; |
4192 | length -= opsize; |
4193 | } |
4194 | } |
4195 | return mss; |
4196 | } |
4197 | |
4198 | /* Look for tcp options. Normally only called on SYN and SYNACK packets. |
4199 | * But, this can also be called on packets in the established flow when |
4200 | * the fast version below fails. |
4201 | */ |
4202 | void tcp_parse_options(const struct net *net, |
4203 | const struct sk_buff *skb, |
4204 | struct tcp_options_received *opt_rx, int estab, |
4205 | struct tcp_fastopen_cookie *foc) |
4206 | { |
4207 | const unsigned char *ptr; |
4208 | const struct tcphdr *th = tcp_hdr(skb); |
4209 | int length = (th->doff * 4) - sizeof(struct tcphdr); |
4210 | |
4211 | ptr = (const unsigned char *)(th + 1); |
4212 | opt_rx->saw_tstamp = 0; |
4213 | opt_rx->saw_unknown = 0; |
4214 | |
4215 | while (length > 0) { |
4216 | int opcode = *ptr++; |
4217 | int opsize; |
4218 | |
4219 | switch (opcode) { |
4220 | case TCPOPT_EOL: |
4221 | return; |
4222 | case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ |
4223 | length--; |
4224 | continue; |
4225 | default: |
4226 | if (length < 2) |
4227 | return; |
4228 | opsize = *ptr++; |
4229 | if (opsize < 2) /* "silly options" */ |
4230 | return; |
4231 | if (opsize > length) |
4232 | return; /* don't parse partial options */ |
4233 | switch (opcode) { |
4234 | case TCPOPT_MSS: |
4235 | if (opsize == TCPOLEN_MSS && th->syn && !estab) { |
4236 | u16 in_mss = get_unaligned_be16(p: ptr); |
4237 | if (in_mss) { |
4238 | if (opt_rx->user_mss && |
4239 | opt_rx->user_mss < in_mss) |
4240 | in_mss = opt_rx->user_mss; |
4241 | opt_rx->mss_clamp = in_mss; |
4242 | } |
4243 | } |
4244 | break; |
4245 | case TCPOPT_WINDOW: |
4246 | if (opsize == TCPOLEN_WINDOW && th->syn && |
4247 | !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) { |
4248 | __u8 snd_wscale = *(__u8 *)ptr; |
4249 | opt_rx->wscale_ok = 1; |
4250 | if (snd_wscale > TCP_MAX_WSCALE) { |
4251 | net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n", |
4252 | __func__, |
4253 | snd_wscale, |
4254 | TCP_MAX_WSCALE); |
4255 | snd_wscale = TCP_MAX_WSCALE; |
4256 | } |
4257 | opt_rx->snd_wscale = snd_wscale; |
4258 | } |
4259 | break; |
4260 | case TCPOPT_TIMESTAMP: |
4261 | if ((opsize == TCPOLEN_TIMESTAMP) && |
4262 | ((estab && opt_rx->tstamp_ok) || |
4263 | (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) { |
4264 | opt_rx->saw_tstamp = 1; |
4265 | opt_rx->rcv_tsval = get_unaligned_be32(p: ptr); |
4266 | opt_rx->rcv_tsecr = get_unaligned_be32(p: ptr + 4); |
4267 | } |
4268 | break; |
4269 | case TCPOPT_SACK_PERM: |
4270 | if (opsize == TCPOLEN_SACK_PERM && th->syn && |
4271 | !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) { |
4272 | opt_rx->sack_ok = TCP_SACK_SEEN; |
4273 | tcp_sack_reset(rx_opt: opt_rx); |
4274 | } |
4275 | break; |
4276 | |
4277 | case TCPOPT_SACK: |
4278 | if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && |
4279 | !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && |
4280 | opt_rx->sack_ok) { |
4281 | TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; |
4282 | } |
4283 | break; |
4284 | #ifdef CONFIG_TCP_MD5SIG |
4285 | case TCPOPT_MD5SIG: |
4286 | /* The MD5 Hash has already been |
4287 | * checked (see tcp_v{4,6}_rcv()). |
4288 | */ |
4289 | break; |
4290 | #endif |
4291 | #ifdef CONFIG_TCP_AO |
4292 | case TCPOPT_AO: |
4293 | /* TCP AO has already been checked |
4294 | * (see tcp_inbound_ao_hash()). |
4295 | */ |
4296 | break; |
4297 | #endif |
4298 | case TCPOPT_FASTOPEN: |
4299 | tcp_parse_fastopen_option( |
4300 | len: opsize - TCPOLEN_FASTOPEN_BASE, |
4301 | cookie: ptr, syn: th->syn, foc, exp_opt: false); |
4302 | break; |
4303 | |
4304 | case TCPOPT_EXP: |
4305 | /* Fast Open option shares code 254 using a |
4306 | * 16 bits magic number. |
4307 | */ |
4308 | if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE && |
4309 | get_unaligned_be16(p: ptr) == |
4310 | TCPOPT_FASTOPEN_MAGIC) { |
4311 | tcp_parse_fastopen_option(len: opsize - |
4312 | TCPOLEN_EXP_FASTOPEN_BASE, |
4313 | cookie: ptr + 2, syn: th->syn, foc, exp_opt: true); |
4314 | break; |
4315 | } |
4316 | |
4317 | if (smc_parse_options(th, opt_rx, ptr, opsize)) |
4318 | break; |
4319 | |
4320 | opt_rx->saw_unknown = 1; |
4321 | break; |
4322 | |
4323 | default: |
4324 | opt_rx->saw_unknown = 1; |
4325 | } |
4326 | ptr += opsize-2; |
4327 | length -= opsize; |
4328 | } |
4329 | } |
4330 | } |
4331 | EXPORT_SYMBOL(tcp_parse_options); |
4332 | |
4333 | static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) |
4334 | { |
4335 | const __be32 *ptr = (const __be32 *)(th + 1); |
4336 | |
4337 | if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
4338 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { |
4339 | tp->rx_opt.saw_tstamp = 1; |
4340 | ++ptr; |
4341 | tp->rx_opt.rcv_tsval = ntohl(*ptr); |
4342 | ++ptr; |
4343 | if (*ptr) |
4344 | tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset; |
4345 | else |
4346 | tp->rx_opt.rcv_tsecr = 0; |
4347 | return true; |
4348 | } |
4349 | return false; |
4350 | } |
4351 | |
4352 | /* Fast parse options. This hopes to only see timestamps. |
4353 | * If it is wrong it falls back on tcp_parse_options(). |
4354 | */ |
4355 | static bool tcp_fast_parse_options(const struct net *net, |
4356 | const struct sk_buff *skb, |
4357 | const struct tcphdr *th, struct tcp_sock *tp) |
4358 | { |
4359 | /* In the spirit of fast parsing, compare doff directly to constant |
4360 | * values. Because equality is used, short doff can be ignored here. |
4361 | */ |
4362 | if (th->doff == (sizeof(*th) / 4)) { |
4363 | tp->rx_opt.saw_tstamp = 0; |
4364 | return false; |
4365 | } else if (tp->rx_opt.tstamp_ok && |
4366 | th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { |
4367 | if (tcp_parse_aligned_timestamp(tp, th)) |
4368 | return true; |
4369 | } |
4370 | |
4371 | tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL); |
4372 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) |
4373 | tp->rx_opt.rcv_tsecr -= tp->tsoffset; |
4374 | |
4375 | return true; |
4376 | } |
4377 | |
4378 | #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO) |
4379 | /* |
4380 | * Parse Signature options |
4381 | */ |
4382 | int tcp_do_parse_auth_options(const struct tcphdr *th, |
4383 | const u8 **md5_hash, const u8 **ao_hash) |
4384 | { |
4385 | int length = (th->doff << 2) - sizeof(*th); |
4386 | const u8 *ptr = (const u8 *)(th + 1); |
4387 | unsigned int minlen = TCPOLEN_MD5SIG; |
4388 | |
4389 | if (IS_ENABLED(CONFIG_TCP_AO)) |
4390 | minlen = sizeof(struct tcp_ao_hdr) + 1; |
4391 | |
4392 | *md5_hash = NULL; |
4393 | *ao_hash = NULL; |
4394 | |
4395 | /* If not enough data remaining, we can short cut */ |
4396 | while (length >= minlen) { |
4397 | int opcode = *ptr++; |
4398 | int opsize; |
4399 | |
4400 | switch (opcode) { |
4401 | case TCPOPT_EOL: |
4402 | return 0; |
4403 | case TCPOPT_NOP: |
4404 | length--; |
4405 | continue; |
4406 | default: |
4407 | opsize = *ptr++; |
4408 | if (opsize < 2 || opsize > length) |
4409 | return -EINVAL; |
4410 | if (opcode == TCPOPT_MD5SIG) { |
4411 | if (opsize != TCPOLEN_MD5SIG) |
4412 | return -EINVAL; |
4413 | if (unlikely(*md5_hash || *ao_hash)) |
4414 | return -EEXIST; |
4415 | *md5_hash = ptr; |
4416 | } else if (opcode == TCPOPT_AO) { |
4417 | if (opsize <= sizeof(struct tcp_ao_hdr)) |
4418 | return -EINVAL; |
4419 | if (unlikely(*md5_hash || *ao_hash)) |
4420 | return -EEXIST; |
4421 | *ao_hash = ptr; |
4422 | } |
4423 | } |
4424 | ptr += opsize - 2; |
4425 | length -= opsize; |
4426 | } |
4427 | return 0; |
4428 | } |
4429 | EXPORT_SYMBOL(tcp_do_parse_auth_options); |
4430 | #endif |
4431 | |
4432 | /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM |
4433 | * |
4434 | * It is not fatal. If this ACK does _not_ change critical state (seqs, window) |
4435 | * it can pass through stack. So, the following predicate verifies that |
4436 | * this segment is not used for anything but congestion avoidance or |
4437 | * fast retransmit. Moreover, we even are able to eliminate most of such |
4438 | * second order effects, if we apply some small "replay" window (~RTO) |
4439 | * to timestamp space. |
4440 | * |
4441 | * All these measures still do not guarantee that we reject wrapped ACKs |
4442 | * on networks with high bandwidth, when sequence space is recycled fastly, |
4443 | * but it guarantees that such events will be very rare and do not affect |
4444 | * connection seriously. This doesn't look nice, but alas, PAWS is really |
4445 | * buggy extension. |
4446 | * |
4447 | * [ Later note. Even worse! It is buggy for segments _with_ data. RFC |
4448 | * states that events when retransmit arrives after original data are rare. |
4449 | * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is |
4450 | * the biggest problem on large power networks even with minor reordering. |
4451 | * OK, let's give it small replay window. If peer clock is even 1hz, it is safe |
4452 | * up to bandwidth of 18Gigabit/sec. 8) ] |
4453 | */ |
4454 | |
4455 | /* Estimates max number of increments of remote peer TSval in |
4456 | * a replay window (based on our current RTO estimation). |
4457 | */ |
4458 | static u32 tcp_tsval_replay(const struct sock *sk) |
4459 | { |
4460 | /* If we use usec TS resolution, |
4461 | * then expect the remote peer to use the same resolution. |
4462 | */ |
4463 | if (tcp_sk(sk)->tcp_usec_ts) |
4464 | return inet_csk(sk)->icsk_rto * (USEC_PER_SEC / HZ); |
4465 | |
4466 | /* RFC 7323 recommends a TSval clock between 1ms and 1sec. |
4467 | * We know that some OS (including old linux) can use 1200 Hz. |
4468 | */ |
4469 | return inet_csk(sk)->icsk_rto * 1200 / HZ; |
4470 | } |
4471 | |
4472 | static enum skb_drop_reason tcp_disordered_ack_check(const struct sock *sk, |
4473 | const struct sk_buff *skb) |
4474 | { |
4475 | const struct tcp_sock *tp = tcp_sk(sk); |
4476 | const struct tcphdr *th = tcp_hdr(skb); |
4477 | SKB_DR_INIT(reason, TCP_RFC7323_PAWS); |
4478 | u32 ack = TCP_SKB_CB(skb)->ack_seq; |
4479 | u32 seq = TCP_SKB_CB(skb)->seq; |
4480 | |
4481 | /* 1. Is this not a pure ACK ? */ |
4482 | if (!th->ack || seq != TCP_SKB_CB(skb)->end_seq) |
4483 | return reason; |
4484 | |
4485 | /* 2. Is its sequence not the expected one ? */ |
4486 | if (seq != tp->rcv_nxt) |
4487 | return before(seq1: seq, seq2: tp->rcv_nxt) ? |
4488 | SKB_DROP_REASON_TCP_RFC7323_PAWS_ACK : |
4489 | reason; |
4490 | |
4491 | /* 3. Is this not a duplicate ACK ? */ |
4492 | if (ack != tp->snd_una) |
4493 | return reason; |
4494 | |
4495 | /* 4. Is this updating the window ? */ |
4496 | if (tcp_may_update_window(tp, ack, ack_seq: seq, ntohs(th->window) << |
4497 | tp->rx_opt.snd_wscale)) |
4498 | return reason; |
4499 | |
4500 | /* 5. Is this not in the replay window ? */ |
4501 | if ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > |
4502 | tcp_tsval_replay(sk)) |
4503 | return reason; |
4504 | |
4505 | return 0; |
4506 | } |
4507 | |
4508 | /* Check segment sequence number for validity. |
4509 | * |
4510 | * Segment controls are considered valid, if the segment |
4511 | * fits to the window after truncation to the window. Acceptability |
4512 | * of data (and SYN, FIN, of course) is checked separately. |
4513 | * See tcp_data_queue(), for example. |
4514 | * |
4515 | * Also, controls (RST is main one) are accepted using RCV.WUP instead |
4516 | * of RCV.NXT. Peer still did not advance his SND.UNA when we |
4517 | * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. |
4518 | * (borrowed from freebsd) |
4519 | */ |
4520 | |
4521 | static enum skb_drop_reason tcp_sequence(const struct tcp_sock *tp, |
4522 | u32 seq, u32 end_seq) |
4523 | { |
4524 | if (before(seq1: end_seq, seq2: tp->rcv_wup)) |
4525 | return SKB_DROP_REASON_TCP_OLD_SEQUENCE; |
4526 | |
4527 | if (after(seq, tp->rcv_nxt + tcp_receive_window(tp))) |
4528 | return SKB_DROP_REASON_TCP_INVALID_SEQUENCE; |
4529 | |
4530 | return SKB_NOT_DROPPED_YET; |
4531 | } |
4532 | |
4533 | |
4534 | void tcp_done_with_error(struct sock *sk, int err) |
4535 | { |
4536 | /* This barrier is coupled with smp_rmb() in tcp_poll() */ |
4537 | WRITE_ONCE(sk->sk_err, err); |
4538 | smp_wmb(); |
4539 | |
4540 | tcp_write_queue_purge(sk); |
4541 | tcp_done(sk); |
4542 | |
4543 | if (!sock_flag(sk, flag: SOCK_DEAD)) |
4544 | sk_error_report(sk); |
4545 | } |
4546 | EXPORT_IPV6_MOD(tcp_done_with_error); |
4547 | |
4548 | /* When we get a reset we do this. */ |
4549 | void tcp_reset(struct sock *sk, struct sk_buff *skb) |
4550 | { |
4551 | int err; |
4552 | |
4553 | trace_tcp_receive_reset(sk); |
4554 | |
4555 | /* mptcp can't tell us to ignore reset pkts, |
4556 | * so just ignore the return value of mptcp_incoming_options(). |
4557 | */ |
4558 | if (sk_is_mptcp(sk)) |
4559 | mptcp_incoming_options(sk, skb); |
4560 | |
4561 | /* We want the right error as BSD sees it (and indeed as we do). */ |
4562 | switch (sk->sk_state) { |
4563 | case TCP_SYN_SENT: |
4564 | err = ECONNREFUSED; |
4565 | break; |
4566 | case TCP_CLOSE_WAIT: |
4567 | err = EPIPE; |
4568 | break; |
4569 | case TCP_CLOSE: |
4570 | return; |
4571 | default: |
4572 | err = ECONNRESET; |
4573 | } |
4574 | tcp_done_with_error(sk, err); |
4575 | } |
4576 | |
4577 | /* |
4578 | * Process the FIN bit. This now behaves as it is supposed to work |
4579 | * and the FIN takes effect when it is validly part of sequence |
4580 | * space. Not before when we get holes. |
4581 | * |
4582 | * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT |
4583 | * (and thence onto LAST-ACK and finally, CLOSE, we never enter |
4584 | * TIME-WAIT) |
4585 | * |
4586 | * If we are in FINWAIT-1, a received FIN indicates simultaneous |
4587 | * close and we go into CLOSING (and later onto TIME-WAIT) |
4588 | * |
4589 | * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. |
4590 | */ |
4591 | void tcp_fin(struct sock *sk) |
4592 | { |
4593 | struct tcp_sock *tp = tcp_sk(sk); |
4594 | |
4595 | inet_csk_schedule_ack(sk); |
4596 | |
4597 | WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN); |
4598 | sock_set_flag(sk, flag: SOCK_DONE); |
4599 | |
4600 | switch (sk->sk_state) { |
4601 | case TCP_SYN_RECV: |
4602 | case TCP_ESTABLISHED: |
4603 | /* Move to CLOSE_WAIT */ |
4604 | tcp_set_state(sk, state: TCP_CLOSE_WAIT); |
4605 | inet_csk_enter_pingpong_mode(sk); |
4606 | break; |
4607 | |
4608 | case TCP_CLOSE_WAIT: |
4609 | case TCP_CLOSING: |
4610 | /* Received a retransmission of the FIN, do |
4611 | * nothing. |
4612 | */ |
4613 | break; |
4614 | case TCP_LAST_ACK: |
4615 | /* RFC793: Remain in the LAST-ACK state. */ |
4616 | break; |
4617 | |
4618 | case TCP_FIN_WAIT1: |
4619 | /* This case occurs when a simultaneous close |
4620 | * happens, we must ack the received FIN and |
4621 | * enter the CLOSING state. |
4622 | */ |
4623 | tcp_send_ack(sk); |
4624 | tcp_set_state(sk, state: TCP_CLOSING); |
4625 | break; |
4626 | case TCP_FIN_WAIT2: |
4627 | /* Received a FIN -- send ACK and enter TIME_WAIT. */ |
4628 | tcp_send_ack(sk); |
4629 | tcp_time_wait(sk, state: TCP_TIME_WAIT, timeo: 0); |
4630 | break; |
4631 | default: |
4632 | /* Only TCP_LISTEN and TCP_CLOSE are left, in these |
4633 | * cases we should never reach this piece of code. |
4634 | */ |
4635 | pr_err("%s: Impossible, sk->sk_state=%d\n", |
4636 | __func__, sk->sk_state); |
4637 | break; |
4638 | } |
4639 | |
4640 | /* It _is_ possible, that we have something out-of-order _after_ FIN. |
4641 | * Probably, we should reset in this case. For now drop them. |
4642 | */ |
4643 | skb_rbtree_purge(root: &tp->out_of_order_queue); |
4644 | if (tcp_is_sack(tp)) |
4645 | tcp_sack_reset(rx_opt: &tp->rx_opt); |
4646 | |
4647 | if (!sock_flag(sk, flag: SOCK_DEAD)) { |
4648 | sk->sk_state_change(sk); |
4649 | |
4650 | /* Do not send POLL_HUP for half duplex close. */ |
4651 | if (sk->sk_shutdown == SHUTDOWN_MASK || |
4652 | sk->sk_state == TCP_CLOSE) |
4653 | sk_wake_async(sk, how: SOCK_WAKE_WAITD, POLL_HUP); |
4654 | else |
4655 | sk_wake_async(sk, how: SOCK_WAKE_WAITD, POLL_IN); |
4656 | } |
4657 | } |
4658 | |
4659 | static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, |
4660 | u32 end_seq) |
4661 | { |
4662 | if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { |
4663 | if (before(seq1: seq, seq2: sp->start_seq)) |
4664 | sp->start_seq = seq; |
4665 | if (after(end_seq, sp->end_seq)) |
4666 | sp->end_seq = end_seq; |
4667 | return true; |
4668 | } |
4669 | return false; |
4670 | } |
4671 | |
4672 | static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) |
4673 | { |
4674 | struct tcp_sock *tp = tcp_sk(sk); |
4675 | |
4676 | if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) { |
4677 | int mib_idx; |
4678 | |
4679 | if (before(seq1: seq, seq2: tp->rcv_nxt)) |
4680 | mib_idx = LINUX_MIB_TCPDSACKOLDSENT; |
4681 | else |
4682 | mib_idx = LINUX_MIB_TCPDSACKOFOSENT; |
4683 | |
4684 | NET_INC_STATS(sock_net(sk), mib_idx); |
4685 | |
4686 | tp->rx_opt.dsack = 1; |
4687 | tp->duplicate_sack[0].start_seq = seq; |
4688 | tp->duplicate_sack[0].end_seq = end_seq; |
4689 | } |
4690 | } |
4691 | |
4692 | static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) |
4693 | { |
4694 | struct tcp_sock *tp = tcp_sk(sk); |
4695 | |
4696 | if (!tp->rx_opt.dsack) |
4697 | tcp_dsack_set(sk, seq, end_seq); |
4698 | else |
4699 | tcp_sack_extend(sp: tp->duplicate_sack, seq, end_seq); |
4700 | } |
4701 | |
4702 | static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb) |
4703 | { |
4704 | /* When the ACK path fails or drops most ACKs, the sender would |
4705 | * timeout and spuriously retransmit the same segment repeatedly. |
4706 | * If it seems our ACKs are not reaching the other side, |
4707 | * based on receiving a duplicate data segment with new flowlabel |
4708 | * (suggesting the sender suffered an RTO), and we are not already |
4709 | * repathing due to our own RTO, then rehash the socket to repath our |
4710 | * packets. |
4711 | */ |
4712 | #if IS_ENABLED(CONFIG_IPV6) |
4713 | if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss && |
4714 | skb->protocol == htons(ETH_P_IPV6) && |
4715 | (tcp_sk(sk)->inet_conn.icsk_ack.lrcv_flowlabel != |
4716 | ntohl(ip6_flowlabel(ipv6_hdr(skb)))) && |
4717 | sk_rethink_txhash(sk)) |
4718 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH); |
4719 | |
4720 | /* Save last flowlabel after a spurious retrans. */ |
4721 | tcp_save_lrcv_flowlabel(sk, skb); |
4722 | #endif |
4723 | } |
4724 | |
4725 | static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) |
4726 | { |
4727 | struct tcp_sock *tp = tcp_sk(sk); |
4728 | |
4729 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && |
4730 | before(TCP_SKB_CB(skb)->seq, seq2: tp->rcv_nxt)) { |
4731 | NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); |
4732 | tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); |
4733 | |
4734 | if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) { |
4735 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; |
4736 | |
4737 | tcp_rcv_spurious_retrans(sk, skb); |
4738 | if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) |
4739 | end_seq = tp->rcv_nxt; |
4740 | tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); |
4741 | } |
4742 | } |
4743 | |
4744 | tcp_send_ack(sk); |
4745 | } |
4746 | |
4747 | /* These routines update the SACK block as out-of-order packets arrive or |
4748 | * in-order packets close up the sequence space. |
4749 | */ |
4750 | static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) |
4751 | { |
4752 | int this_sack; |
4753 | struct tcp_sack_block *sp = &tp->selective_acks[0]; |
4754 | struct tcp_sack_block *swalk = sp + 1; |
4755 | |
4756 | /* See if the recent change to the first SACK eats into |
4757 | * or hits the sequence space of other SACK blocks, if so coalesce. |
4758 | */ |
4759 | for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { |
4760 | if (tcp_sack_extend(sp, seq: swalk->start_seq, end_seq: swalk->end_seq)) { |
4761 | int i; |
4762 | |
4763 | /* Zap SWALK, by moving every further SACK up by one slot. |
4764 | * Decrease num_sacks. |
4765 | */ |
4766 | tp->rx_opt.num_sacks--; |
4767 | for (i = this_sack; i < tp->rx_opt.num_sacks; i++) |
4768 | sp[i] = sp[i + 1]; |
4769 | continue; |
4770 | } |
4771 | this_sack++; |
4772 | swalk++; |
4773 | } |
4774 | } |
4775 | |
4776 | void tcp_sack_compress_send_ack(struct sock *sk) |
4777 | { |
4778 | struct tcp_sock *tp = tcp_sk(sk); |
4779 | |
4780 | if (!tp->compressed_ack) |
4781 | return; |
4782 | |
4783 | if (hrtimer_try_to_cancel(timer: &tp->compressed_ack_timer) == 1) |
4784 | __sock_put(sk); |
4785 | |
4786 | /* Since we have to send one ack finally, |
4787 | * substract one from tp->compressed_ack to keep |
4788 | * LINUX_MIB_TCPACKCOMPRESSED accurate. |
4789 | */ |
4790 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, |
4791 | tp->compressed_ack - 1); |
4792 | |
4793 | tp->compressed_ack = 0; |
4794 | tcp_send_ack(sk); |
4795 | } |
4796 | |
4797 | /* Reasonable amount of sack blocks included in TCP SACK option |
4798 | * The max is 4, but this becomes 3 if TCP timestamps are there. |
4799 | * Given that SACK packets might be lost, be conservative and use 2. |
4800 | */ |
4801 | #define TCP_SACK_BLOCKS_EXPECTED 2 |
4802 | |
4803 | static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) |
4804 | { |
4805 | struct tcp_sock *tp = tcp_sk(sk); |
4806 | struct tcp_sack_block *sp = &tp->selective_acks[0]; |
4807 | int cur_sacks = tp->rx_opt.num_sacks; |
4808 | int this_sack; |
4809 | |
4810 | if (!cur_sacks) |
4811 | goto new_sack; |
4812 | |
4813 | for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { |
4814 | if (tcp_sack_extend(sp, seq, end_seq)) { |
4815 | if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) |
4816 | tcp_sack_compress_send_ack(sk); |
4817 | /* Rotate this_sack to the first one. */ |
4818 | for (; this_sack > 0; this_sack--, sp--) |
4819 | swap(*sp, *(sp - 1)); |
4820 | if (cur_sacks > 1) |
4821 | tcp_sack_maybe_coalesce(tp); |
4822 | return; |
4823 | } |
4824 | } |
4825 | |
4826 | if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) |
4827 | tcp_sack_compress_send_ack(sk); |
4828 | |
4829 | /* Could not find an adjacent existing SACK, build a new one, |
4830 | * put it at the front, and shift everyone else down. We |
4831 | * always know there is at least one SACK present already here. |
4832 | * |
4833 | * If the sack array is full, forget about the last one. |
4834 | */ |
4835 | if (this_sack >= TCP_NUM_SACKS) { |
4836 | this_sack--; |
4837 | tp->rx_opt.num_sacks--; |
4838 | sp--; |
4839 | } |
4840 | for (; this_sack > 0; this_sack--, sp--) |
4841 | *sp = *(sp - 1); |
4842 | |
4843 | new_sack: |
4844 | /* Build the new head SACK, and we're done. */ |
4845 | sp->start_seq = seq; |
4846 | sp->end_seq = end_seq; |
4847 | tp->rx_opt.num_sacks++; |
4848 | } |
4849 | |
4850 | /* RCV.NXT advances, some SACKs should be eaten. */ |
4851 | |
4852 | static void tcp_sack_remove(struct tcp_sock *tp) |
4853 | { |
4854 | struct tcp_sack_block *sp = &tp->selective_acks[0]; |
4855 | int num_sacks = tp->rx_opt.num_sacks; |
4856 | int this_sack; |
4857 | |
4858 | /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ |
4859 | if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { |
4860 | tp->rx_opt.num_sacks = 0; |
4861 | return; |
4862 | } |
4863 | |
4864 | for (this_sack = 0; this_sack < num_sacks;) { |
4865 | /* Check if the start of the sack is covered by RCV.NXT. */ |
4866 | if (!before(seq1: tp->rcv_nxt, seq2: sp->start_seq)) { |
4867 | int i; |
4868 | |
4869 | /* RCV.NXT must cover all the block! */ |
4870 | WARN_ON(before(tp->rcv_nxt, sp->end_seq)); |
4871 | |
4872 | /* Zap this SACK, by moving forward any other SACKS. */ |
4873 | for (i = this_sack+1; i < num_sacks; i++) |
4874 | tp->selective_acks[i-1] = tp->selective_acks[i]; |
4875 | num_sacks--; |
4876 | continue; |
4877 | } |
4878 | this_sack++; |
4879 | sp++; |
4880 | } |
4881 | tp->rx_opt.num_sacks = num_sacks; |
4882 | } |
4883 | |
4884 | /** |
4885 | * tcp_try_coalesce - try to merge skb to prior one |
4886 | * @sk: socket |
4887 | * @to: prior buffer |
4888 | * @from: buffer to add in queue |
4889 | * @fragstolen: pointer to boolean |
4890 | * |
4891 | * Before queueing skb @from after @to, try to merge them |
4892 | * to reduce overall memory use and queue lengths, if cost is small. |
4893 | * Packets in ofo or receive queues can stay a long time. |
4894 | * Better try to coalesce them right now to avoid future collapses. |
4895 | * Returns true if caller should free @from instead of queueing it |
4896 | */ |
4897 | static bool tcp_try_coalesce(struct sock *sk, |
4898 | struct sk_buff *to, |
4899 | struct sk_buff *from, |
4900 | bool *fragstolen) |
4901 | { |
4902 | int delta; |
4903 | |
4904 | *fragstolen = false; |
4905 | |
4906 | /* Its possible this segment overlaps with prior segment in queue */ |
4907 | if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq) |
4908 | return false; |
4909 | |
4910 | if (!tcp_skb_can_collapse_rx(to, from)) |
4911 | return false; |
4912 | |
4913 | if (!skb_try_coalesce(to, from, fragstolen, delta_truesize: &delta)) |
4914 | return false; |
4915 | |
4916 | atomic_add(i: delta, v: &sk->sk_rmem_alloc); |
4917 | sk_mem_charge(sk, size: delta); |
4918 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE); |
4919 | TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq; |
4920 | TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq; |
4921 | TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags; |
4922 | |
4923 | if (TCP_SKB_CB(from)->has_rxtstamp) { |
4924 | TCP_SKB_CB(to)->has_rxtstamp = true; |
4925 | to->tstamp = from->tstamp; |
4926 | skb_hwtstamps(skb: to)->hwtstamp = skb_hwtstamps(skb: from)->hwtstamp; |
4927 | } |
4928 | |
4929 | return true; |
4930 | } |
4931 | |
4932 | static bool tcp_ooo_try_coalesce(struct sock *sk, |
4933 | struct sk_buff *to, |
4934 | struct sk_buff *from, |
4935 | bool *fragstolen) |
4936 | { |
4937 | bool res = tcp_try_coalesce(sk, to, from, fragstolen); |
4938 | |
4939 | /* In case tcp_drop_reason() is called later, update to->gso_segs */ |
4940 | if (res) { |
4941 | u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) + |
4942 | max_t(u16, 1, skb_shinfo(from)->gso_segs); |
4943 | |
4944 | skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF); |
4945 | } |
4946 | return res; |
4947 | } |
4948 | |
4949 | noinline_for_tracing static void |
4950 | tcp_drop_reason(struct sock *sk, struct sk_buff *skb, enum skb_drop_reason reason) |
4951 | { |
4952 | sk_drops_add(sk, skb); |
4953 | sk_skb_reason_drop(sk, skb, reason); |
4954 | } |
4955 | |
4956 | /* This one checks to see if we can put data from the |
4957 | * out_of_order queue into the receive_queue. |
4958 | */ |
4959 | static void tcp_ofo_queue(struct sock *sk) |
4960 | { |
4961 | struct tcp_sock *tp = tcp_sk(sk); |
4962 | __u32 dsack_high = tp->rcv_nxt; |
4963 | bool fin, fragstolen, eaten; |
4964 | struct sk_buff *skb, *tail; |
4965 | struct rb_node *p; |
4966 | |
4967 | p = rb_first(&tp->out_of_order_queue); |
4968 | while (p) { |
4969 | skb = rb_to_skb(p); |
4970 | if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) |
4971 | break; |
4972 | |
4973 | if (before(TCP_SKB_CB(skb)->seq, seq2: dsack_high)) { |
4974 | __u32 dsack = dsack_high; |
4975 | if (before(TCP_SKB_CB(skb)->end_seq, seq2: dsack_high)) |
4976 | dsack_high = TCP_SKB_CB(skb)->end_seq; |
4977 | tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, end_seq: dsack); |
4978 | } |
4979 | p = rb_next(p); |
4980 | rb_erase(&skb->rbnode, &tp->out_of_order_queue); |
4981 | |
4982 | if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) { |
4983 | tcp_drop_reason(sk, skb, reason: SKB_DROP_REASON_TCP_OFO_DROP); |
4984 | continue; |
4985 | } |
4986 | |
4987 | tail = skb_peek_tail(list_: &sk->sk_receive_queue); |
4988 | eaten = tail && tcp_try_coalesce(sk, to: tail, from: skb, fragstolen: &fragstolen); |
4989 | tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); |
4990 | fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; |
4991 | if (!eaten) |
4992 | tcp_add_receive_queue(sk, skb); |
4993 | else |
4994 | kfree_skb_partial(skb, head_stolen: fragstolen); |
4995 | |
4996 | if (unlikely(fin)) { |
4997 | tcp_fin(sk); |
4998 | /* tcp_fin() purges tp->out_of_order_queue, |
4999 | * so we must end this loop right now. |
5000 | */ |
5001 | break; |
5002 | } |
5003 | } |
5004 | } |
5005 | |
5006 | static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb); |
5007 | static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb); |
5008 | |
5009 | static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb, |
5010 | unsigned int size) |
5011 | { |
5012 | if (atomic_read(v: &sk->sk_rmem_alloc) > sk->sk_rcvbuf || |
5013 | !sk_rmem_schedule(sk, skb, size)) { |
5014 | |
5015 | if (tcp_prune_queue(sk, in_skb: skb) < 0) |
5016 | return -1; |
5017 | |
5018 | while (!sk_rmem_schedule(sk, skb, size)) { |
5019 | if (!tcp_prune_ofo_queue(sk, in_skb: skb)) |
5020 | return -1; |
5021 | } |
5022 | } |
5023 | return 0; |
5024 | } |
5025 | |
5026 | static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) |
5027 | { |
5028 | struct tcp_sock *tp = tcp_sk(sk); |
5029 | struct rb_node **p, *parent; |
5030 | struct sk_buff *skb1; |
5031 | u32 seq, end_seq; |
5032 | bool fragstolen; |
5033 | |
5034 | tcp_save_lrcv_flowlabel(sk, skb); |
5035 | tcp_data_ecn_check(sk, skb); |
5036 | |
5037 | if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) { |
5038 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP); |
5039 | sk->sk_data_ready(sk); |
5040 | tcp_drop_reason(sk, skb, reason: SKB_DROP_REASON_PROTO_MEM); |
5041 | return; |
5042 | } |
5043 | |
5044 | /* Disable header prediction. */ |
5045 | tp->pred_flags = 0; |
5046 | inet_csk_schedule_ack(sk); |
5047 | |
5048 | tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs); |
5049 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE); |
5050 | seq = TCP_SKB_CB(skb)->seq; |
5051 | end_seq = TCP_SKB_CB(skb)->end_seq; |
5052 | |
5053 | p = &tp->out_of_order_queue.rb_node; |
5054 | if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { |
5055 | /* Initial out of order segment, build 1 SACK. */ |
5056 | if (tcp_is_sack(tp)) { |
5057 | tp->rx_opt.num_sacks = 1; |
5058 | tp->selective_acks[0].start_seq = seq; |
5059 | tp->selective_acks[0].end_seq = end_seq; |
5060 | } |
5061 | rb_link_node(node: &skb->rbnode, NULL, rb_link: p); |
5062 | rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); |
5063 | tp->ooo_last_skb = skb; |
5064 | goto end; |
5065 | } |
5066 | |
5067 | /* In the typical case, we are adding an skb to the end of the list. |
5068 | * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. |
5069 | */ |
5070 | if (tcp_ooo_try_coalesce(sk, to: tp->ooo_last_skb, |
5071 | from: skb, fragstolen: &fragstolen)) { |
5072 | coalesce_done: |
5073 | /* For non sack flows, do not grow window to force DUPACK |
5074 | * and trigger fast retransmit. |
5075 | */ |
5076 | if (tcp_is_sack(tp)) |
5077 | tcp_grow_window(sk, skb, adjust: true); |
5078 | kfree_skb_partial(skb, head_stolen: fragstolen); |
5079 | skb = NULL; |
5080 | goto add_sack; |
5081 | } |
5082 | /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ |
5083 | if (!before(seq1: seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) { |
5084 | parent = &tp->ooo_last_skb->rbnode; |
5085 | p = &parent->rb_right; |
5086 | goto insert; |
5087 | } |
5088 | |
5089 | /* Find place to insert this segment. Handle overlaps on the way. */ |
5090 | parent = NULL; |
5091 | while (*p) { |
5092 | parent = *p; |
5093 | skb1 = rb_to_skb(parent); |
5094 | if (before(seq1: seq, TCP_SKB_CB(skb1)->seq)) { |
5095 | p = &parent->rb_left; |
5096 | continue; |
5097 | } |
5098 | if (before(seq1: seq, TCP_SKB_CB(skb1)->end_seq)) { |
5099 | if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { |
5100 | /* All the bits are present. Drop. */ |
5101 | NET_INC_STATS(sock_net(sk), |
5102 | LINUX_MIB_TCPOFOMERGE); |
5103 | tcp_drop_reason(sk, skb, |
5104 | reason: SKB_DROP_REASON_TCP_OFOMERGE); |
5105 | skb = NULL; |
5106 | tcp_dsack_set(sk, seq, end_seq); |
5107 | goto add_sack; |
5108 | } |
5109 | if (after(seq, TCP_SKB_CB(skb1)->seq)) { |
5110 | /* Partial overlap. */ |
5111 | tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq); |
5112 | } else { |
5113 | /* skb's seq == skb1's seq and skb covers skb1. |
5114 | * Replace skb1 with skb. |
5115 | */ |
5116 | rb_replace_node(victim: &skb1->rbnode, new: &skb->rbnode, |
5117 | root: &tp->out_of_order_queue); |
5118 | tcp_dsack_extend(sk, |
5119 | TCP_SKB_CB(skb1)->seq, |
5120 | TCP_SKB_CB(skb1)->end_seq); |
5121 | NET_INC_STATS(sock_net(sk), |
5122 | LINUX_MIB_TCPOFOMERGE); |
5123 | tcp_drop_reason(sk, skb: skb1, |
5124 | reason: SKB_DROP_REASON_TCP_OFOMERGE); |
5125 | goto merge_right; |
5126 | } |
5127 | } else if (tcp_ooo_try_coalesce(sk, to: skb1, |
5128 | from: skb, fragstolen: &fragstolen)) { |
5129 | goto coalesce_done; |
5130 | } |
5131 | p = &parent->rb_right; |
5132 | } |
5133 | insert: |
5134 | /* Insert segment into RB tree. */ |
5135 | rb_link_node(node: &skb->rbnode, parent, rb_link: p); |
5136 | rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); |
5137 | |
5138 | merge_right: |
5139 | /* Remove other segments covered by skb. */ |
5140 | while ((skb1 = skb_rb_next(skb)) != NULL) { |
5141 | if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) |
5142 | break; |
5143 | if (before(seq1: end_seq, TCP_SKB_CB(skb1)->end_seq)) { |
5144 | tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, |
5145 | end_seq); |
5146 | break; |
5147 | } |
5148 | rb_erase(&skb1->rbnode, &tp->out_of_order_queue); |
5149 | tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, |
5150 | TCP_SKB_CB(skb1)->end_seq); |
5151 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE); |
5152 | tcp_drop_reason(sk, skb: skb1, reason: SKB_DROP_REASON_TCP_OFOMERGE); |
5153 | } |
5154 | /* If there is no skb after us, we are the last_skb ! */ |
5155 | if (!skb1) |
5156 | tp->ooo_last_skb = skb; |
5157 | |
5158 | add_sack: |
5159 | if (tcp_is_sack(tp)) |
5160 | tcp_sack_new_ofo_skb(sk, seq, end_seq); |
5161 | end: |
5162 | if (skb) { |
5163 | /* For non sack flows, do not grow window to force DUPACK |
5164 | * and trigger fast retransmit. |
5165 | */ |
5166 | if (tcp_is_sack(tp)) |
5167 | tcp_grow_window(sk, skb, adjust: false); |
5168 | skb_condense(skb); |
5169 | skb_set_owner_r(skb, sk); |
5170 | } |
5171 | tcp_rcvbuf_grow(sk); |
5172 | } |
5173 | |
5174 | static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, |
5175 | bool *fragstolen) |
5176 | { |
5177 | int eaten; |
5178 | struct sk_buff *tail = skb_peek_tail(list_: &sk->sk_receive_queue); |
5179 | |
5180 | eaten = (tail && |
5181 | tcp_try_coalesce(sk, to: tail, |
5182 | from: skb, fragstolen)) ? 1 : 0; |
5183 | tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq); |
5184 | if (!eaten) { |
5185 | tcp_add_receive_queue(sk, skb); |
5186 | skb_set_owner_r(skb, sk); |
5187 | } |
5188 | return eaten; |
5189 | } |
5190 | |
5191 | int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size) |
5192 | { |
5193 | struct sk_buff *skb; |
5194 | int err = -ENOMEM; |
5195 | int data_len = 0; |
5196 | bool fragstolen; |
5197 | |
5198 | if (size == 0) |
5199 | return 0; |
5200 | |
5201 | if (size > PAGE_SIZE) { |
5202 | int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS); |
5203 | |
5204 | data_len = npages << PAGE_SHIFT; |
5205 | size = data_len + (size & ~PAGE_MASK); |
5206 | } |
5207 | skb = alloc_skb_with_frags(header_len: size - data_len, data_len, |
5208 | PAGE_ALLOC_COSTLY_ORDER, |
5209 | errcode: &err, gfp_mask: sk->sk_allocation); |
5210 | if (!skb) |
5211 | goto err; |
5212 | |
5213 | skb_put(skb, len: size - data_len); |
5214 | skb->data_len = data_len; |
5215 | skb->len = size; |
5216 | |
5217 | if (tcp_try_rmem_schedule(sk, skb, size: skb->truesize)) { |
5218 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); |
5219 | goto err_free; |
5220 | } |
5221 | |
5222 | err = skb_copy_datagram_from_iter(skb, offset: 0, from: &msg->msg_iter, len: size); |
5223 | if (err) |
5224 | goto err_free; |
5225 | |
5226 | TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt; |
5227 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size; |
5228 | TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1; |
5229 | |
5230 | if (tcp_queue_rcv(sk, skb, fragstolen: &fragstolen)) { |
5231 | WARN_ON_ONCE(fragstolen); /* should not happen */ |
5232 | __kfree_skb(skb); |
5233 | } |
5234 | return size; |
5235 | |
5236 | err_free: |
5237 | kfree_skb(skb); |
5238 | err: |
5239 | return err; |
5240 | |
5241 | } |
5242 | |
5243 | void tcp_data_ready(struct sock *sk) |
5244 | { |
5245 | if (tcp_epollin_ready(sk, target: sk->sk_rcvlowat) || sock_flag(sk, flag: SOCK_DONE)) |
5246 | sk->sk_data_ready(sk); |
5247 | } |
5248 | |
5249 | static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) |
5250 | { |
5251 | struct tcp_sock *tp = tcp_sk(sk); |
5252 | enum skb_drop_reason reason; |
5253 | bool fragstolen; |
5254 | int eaten; |
5255 | |
5256 | /* If a subflow has been reset, the packet should not continue |
5257 | * to be processed, drop the packet. |
5258 | */ |
5259 | if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) { |
5260 | __kfree_skb(skb); |
5261 | return; |
5262 | } |
5263 | |
5264 | if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) { |
5265 | __kfree_skb(skb); |
5266 | return; |
5267 | } |
5268 | tcp_cleanup_skb(skb); |
5269 | __skb_pull(skb, len: tcp_hdr(skb)->doff * 4); |
5270 | |
5271 | reason = SKB_DROP_REASON_NOT_SPECIFIED; |
5272 | tp->rx_opt.dsack = 0; |
5273 | |
5274 | /* Queue data for delivery to the user. |
5275 | * Packets in sequence go to the receive queue. |
5276 | * Out of sequence packets to the out_of_order_queue. |
5277 | */ |
5278 | if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { |
5279 | if (tcp_receive_window(tp) == 0) { |
5280 | /* Some stacks are known to send bare FIN packets |
5281 | * in a loop even if we send RWIN 0 in our ACK. |
5282 | * Accepting this FIN does not hurt memory pressure |
5283 | * because the FIN flag will simply be merged to the |
5284 | * receive queue tail skb in most cases. |
5285 | */ |
5286 | if (!skb->len && |
5287 | (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) |
5288 | goto queue_and_out; |
5289 | |
5290 | reason = SKB_DROP_REASON_TCP_ZEROWINDOW; |
5291 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); |
5292 | goto out_of_window; |
5293 | } |
5294 | |
5295 | /* Ok. In sequence. In window. */ |
5296 | queue_and_out: |
5297 | if (tcp_try_rmem_schedule(sk, skb, size: skb->truesize)) { |
5298 | /* TODO: maybe ratelimit these WIN 0 ACK ? */ |
5299 | inet_csk(sk)->icsk_ack.pending |= |
5300 | (ICSK_ACK_NOMEM | ICSK_ACK_NOW); |
5301 | inet_csk_schedule_ack(sk); |
5302 | sk->sk_data_ready(sk); |
5303 | |
5304 | if (skb_queue_len(list_: &sk->sk_receive_queue) && skb->len) { |
5305 | reason = SKB_DROP_REASON_PROTO_MEM; |
5306 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); |
5307 | goto drop; |
5308 | } |
5309 | sk_forced_mem_schedule(sk, size: skb->truesize); |
5310 | } |
5311 | |
5312 | eaten = tcp_queue_rcv(sk, skb, fragstolen: &fragstolen); |
5313 | if (skb->len) |
5314 | tcp_event_data_recv(sk, skb); |
5315 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) |
5316 | tcp_fin(sk); |
5317 | |
5318 | if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) { |
5319 | tcp_ofo_queue(sk); |
5320 | |
5321 | /* RFC5681. 4.2. SHOULD send immediate ACK, when |
5322 | * gap in queue is filled. |
5323 | */ |
5324 | if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) |
5325 | inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; |
5326 | } |
5327 | |
5328 | if (tp->rx_opt.num_sacks) |
5329 | tcp_sack_remove(tp); |
5330 | |
5331 | tcp_fast_path_check(sk); |
5332 | |
5333 | if (eaten > 0) |
5334 | kfree_skb_partial(skb, head_stolen: fragstolen); |
5335 | if (!sock_flag(sk, flag: SOCK_DEAD)) |
5336 | tcp_data_ready(sk); |
5337 | return; |
5338 | } |
5339 | |
5340 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { |
5341 | tcp_rcv_spurious_retrans(sk, skb); |
5342 | /* A retransmit, 2nd most common case. Force an immediate ack. */ |
5343 | reason = SKB_DROP_REASON_TCP_OLD_DATA; |
5344 | NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); |
5345 | tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); |
5346 | |
5347 | out_of_window: |
5348 | tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); |
5349 | inet_csk_schedule_ack(sk); |
5350 | drop: |
5351 | tcp_drop_reason(sk, skb, reason); |
5352 | return; |
5353 | } |
5354 | |
5355 | /* Out of window. F.e. zero window probe. */ |
5356 | if (!before(TCP_SKB_CB(skb)->seq, |
5357 | seq2: tp->rcv_nxt + tcp_receive_window(tp))) { |
5358 | reason = SKB_DROP_REASON_TCP_OVERWINDOW; |
5359 | goto out_of_window; |
5360 | } |
5361 | |
5362 | if (before(TCP_SKB_CB(skb)->seq, seq2: tp->rcv_nxt)) { |
5363 | /* Partial packet, seq < rcv_next < end_seq */ |
5364 | tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq: tp->rcv_nxt); |
5365 | |
5366 | /* If window is closed, drop tail of packet. But after |
5367 | * remembering D-SACK for its head made in previous line. |
5368 | */ |
5369 | if (!tcp_receive_window(tp)) { |
5370 | reason = SKB_DROP_REASON_TCP_ZEROWINDOW; |
5371 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); |
5372 | goto out_of_window; |
5373 | } |
5374 | goto queue_and_out; |
5375 | } |
5376 | |
5377 | tcp_data_queue_ofo(sk, skb); |
5378 | } |
5379 | |
5380 | static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list) |
5381 | { |
5382 | if (list) |
5383 | return !skb_queue_is_last(list, skb) ? skb->next : NULL; |
5384 | |
5385 | return skb_rb_next(skb); |
5386 | } |
5387 | |
5388 | static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, |
5389 | struct sk_buff_head *list, |
5390 | struct rb_root *root) |
5391 | { |
5392 | struct sk_buff *next = tcp_skb_next(skb, list); |
5393 | |
5394 | if (list) |
5395 | __skb_unlink(skb, list); |
5396 | else |
5397 | rb_erase(&skb->rbnode, root); |
5398 | |
5399 | __kfree_skb(skb); |
5400 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); |
5401 | |
5402 | return next; |
5403 | } |
5404 | |
5405 | /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */ |
5406 | void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb) |
5407 | { |
5408 | struct rb_node **p = &root->rb_node; |
5409 | struct rb_node *parent = NULL; |
5410 | struct sk_buff *skb1; |
5411 | |
5412 | while (*p) { |
5413 | parent = *p; |
5414 | skb1 = rb_to_skb(parent); |
5415 | if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq)) |
5416 | p = &parent->rb_left; |
5417 | else |
5418 | p = &parent->rb_right; |
5419 | } |
5420 | rb_link_node(node: &skb->rbnode, parent, rb_link: p); |
5421 | rb_insert_color(&skb->rbnode, root); |
5422 | } |
5423 | |
5424 | /* Collapse contiguous sequence of skbs head..tail with |
5425 | * sequence numbers start..end. |
5426 | * |
5427 | * If tail is NULL, this means until the end of the queue. |
5428 | * |
5429 | * Segments with FIN/SYN are not collapsed (only because this |
5430 | * simplifies code) |
5431 | */ |
5432 | static void |
5433 | tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root, |
5434 | struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end) |
5435 | { |
5436 | struct sk_buff *skb = head, *n; |
5437 | struct sk_buff_head tmp; |
5438 | bool end_of_skbs; |
5439 | |
5440 | /* First, check that queue is collapsible and find |
5441 | * the point where collapsing can be useful. |
5442 | */ |
5443 | restart: |
5444 | for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) { |
5445 | n = tcp_skb_next(skb, list); |
5446 | |
5447 | if (!skb_frags_readable(skb)) |
5448 | goto skip_this; |
5449 | |
5450 | /* No new bits? It is possible on ofo queue. */ |
5451 | if (!before(seq1: start, TCP_SKB_CB(skb)->end_seq)) { |
5452 | skb = tcp_collapse_one(sk, skb, list, root); |
5453 | if (!skb) |
5454 | break; |
5455 | goto restart; |
5456 | } |
5457 | |
5458 | /* The first skb to collapse is: |
5459 | * - not SYN/FIN and |
5460 | * - bloated or contains data before "start" or |
5461 | * overlaps to the next one and mptcp allow collapsing. |
5462 | */ |
5463 | if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) && |
5464 | (tcp_win_from_space(sk, space: skb->truesize) > skb->len || |
5465 | before(TCP_SKB_CB(skb)->seq, seq2: start))) { |
5466 | end_of_skbs = false; |
5467 | break; |
5468 | } |
5469 | |
5470 | if (n && n != tail && skb_frags_readable(skb: n) && |
5471 | tcp_skb_can_collapse_rx(to: skb, from: n) && |
5472 | TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) { |
5473 | end_of_skbs = false; |
5474 | break; |
5475 | } |
5476 | |
5477 | skip_this: |
5478 | /* Decided to skip this, advance start seq. */ |
5479 | start = TCP_SKB_CB(skb)->end_seq; |
5480 | } |
5481 | if (end_of_skbs || |
5482 | (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) || |
5483 | !skb_frags_readable(skb)) |
5484 | return; |
5485 | |
5486 | __skb_queue_head_init(list: &tmp); |
5487 | |
5488 | while (before(seq1: start, seq2: end)) { |
5489 | int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start); |
5490 | struct sk_buff *nskb; |
5491 | |
5492 | nskb = alloc_skb(size: copy, GFP_ATOMIC); |
5493 | if (!nskb) |
5494 | break; |
5495 | |
5496 | memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); |
5497 | skb_copy_decrypted(to: nskb, from: skb); |
5498 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; |
5499 | if (list) |
5500 | __skb_queue_before(list, next: skb, newsk: nskb); |
5501 | else |
5502 | __skb_queue_tail(list: &tmp, newsk: nskb); /* defer rbtree insertion */ |
5503 | skb_set_owner_r(skb: nskb, sk); |
5504 | mptcp_skb_ext_move(to: nskb, from: skb); |
5505 | |
5506 | /* Copy data, releasing collapsed skbs. */ |
5507 | while (copy > 0) { |
5508 | int offset = start - TCP_SKB_CB(skb)->seq; |
5509 | int size = TCP_SKB_CB(skb)->end_seq - start; |
5510 | |
5511 | BUG_ON(offset < 0); |
5512 | if (size > 0) { |
5513 | size = min(copy, size); |
5514 | if (skb_copy_bits(skb, offset, to: skb_put(skb: nskb, len: size), len: size)) |
5515 | BUG(); |
5516 | TCP_SKB_CB(nskb)->end_seq += size; |
5517 | copy -= size; |
5518 | start += size; |
5519 | } |
5520 | if (!before(seq1: start, TCP_SKB_CB(skb)->end_seq)) { |
5521 | skb = tcp_collapse_one(sk, skb, list, root); |
5522 | if (!skb || |
5523 | skb == tail || |
5524 | !tcp_skb_can_collapse_rx(to: nskb, from: skb) || |
5525 | (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) || |
5526 | !skb_frags_readable(skb)) |
5527 | goto end; |
5528 | } |
5529 | } |
5530 | } |
5531 | end: |
5532 | skb_queue_walk_safe(&tmp, skb, n) |
5533 | tcp_rbtree_insert(root, skb); |
5534 | } |
5535 | |
5536 | /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs |
5537 | * and tcp_collapse() them until all the queue is collapsed. |
5538 | */ |
5539 | static void tcp_collapse_ofo_queue(struct sock *sk) |
5540 | { |
5541 | struct tcp_sock *tp = tcp_sk(sk); |
5542 | u32 range_truesize, sum_tiny = 0; |
5543 | struct sk_buff *skb, *head; |
5544 | u32 start, end; |
5545 | |
5546 | skb = skb_rb_first(&tp->out_of_order_queue); |
5547 | new_range: |
5548 | if (!skb) { |
5549 | tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue); |
5550 | return; |
5551 | } |
5552 | start = TCP_SKB_CB(skb)->seq; |
5553 | end = TCP_SKB_CB(skb)->end_seq; |
5554 | range_truesize = skb->truesize; |
5555 | |
5556 | for (head = skb;;) { |
5557 | skb = skb_rb_next(skb); |
5558 | |
5559 | /* Range is terminated when we see a gap or when |
5560 | * we are at the queue end. |
5561 | */ |
5562 | if (!skb || |
5563 | after(TCP_SKB_CB(skb)->seq, end) || |
5564 | before(TCP_SKB_CB(skb)->end_seq, seq2: start)) { |
5565 | /* Do not attempt collapsing tiny skbs */ |
5566 | if (range_truesize != head->truesize || |
5567 | end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) { |
5568 | tcp_collapse(sk, NULL, root: &tp->out_of_order_queue, |
5569 | head, tail: skb, start, end); |
5570 | } else { |
5571 | sum_tiny += range_truesize; |
5572 | if (sum_tiny > sk->sk_rcvbuf >> 3) |
5573 | return; |
5574 | } |
5575 | goto new_range; |
5576 | } |
5577 | |
5578 | range_truesize += skb->truesize; |
5579 | if (unlikely(before(TCP_SKB_CB(skb)->seq, start))) |
5580 | start = TCP_SKB_CB(skb)->seq; |
5581 | if (after(TCP_SKB_CB(skb)->end_seq, end)) |
5582 | end = TCP_SKB_CB(skb)->end_seq; |
5583 | } |
5584 | } |
5585 | |
5586 | /* |
5587 | * Clean the out-of-order queue to make room. |
5588 | * We drop high sequences packets to : |
5589 | * 1) Let a chance for holes to be filled. |
5590 | * This means we do not drop packets from ooo queue if their sequence |
5591 | * is before incoming packet sequence. |
5592 | * 2) not add too big latencies if thousands of packets sit there. |
5593 | * (But if application shrinks SO_RCVBUF, we could still end up |
5594 | * freeing whole queue here) |
5595 | * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks. |
5596 | * |
5597 | * Return true if queue has shrunk. |
5598 | */ |
5599 | static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb) |
5600 | { |
5601 | struct tcp_sock *tp = tcp_sk(sk); |
5602 | struct rb_node *node, *prev; |
5603 | bool pruned = false; |
5604 | int goal; |
5605 | |
5606 | if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) |
5607 | return false; |
5608 | |
5609 | goal = sk->sk_rcvbuf >> 3; |
5610 | node = &tp->ooo_last_skb->rbnode; |
5611 | |
5612 | do { |
5613 | struct sk_buff *skb = rb_to_skb(node); |
5614 | |
5615 | /* If incoming skb would land last in ofo queue, stop pruning. */ |
5616 | if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq)) |
5617 | break; |
5618 | pruned = true; |
5619 | prev = rb_prev(node); |
5620 | rb_erase(node, &tp->out_of_order_queue); |
5621 | goal -= skb->truesize; |
5622 | tcp_drop_reason(sk, skb, reason: SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE); |
5623 | tp->ooo_last_skb = rb_to_skb(prev); |
5624 | if (!prev || goal <= 0) { |
5625 | if (atomic_read(v: &sk->sk_rmem_alloc) <= sk->sk_rcvbuf && |
5626 | !tcp_under_memory_pressure(sk)) |
5627 | break; |
5628 | goal = sk->sk_rcvbuf >> 3; |
5629 | } |
5630 | node = prev; |
5631 | } while (node); |
5632 | |
5633 | if (pruned) { |
5634 | NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED); |
5635 | /* Reset SACK state. A conforming SACK implementation will |
5636 | * do the same at a timeout based retransmit. When a connection |
5637 | * is in a sad state like this, we care only about integrity |
5638 | * of the connection not performance. |
5639 | */ |
5640 | if (tp->rx_opt.sack_ok) |
5641 | tcp_sack_reset(rx_opt: &tp->rx_opt); |
5642 | } |
5643 | return pruned; |
5644 | } |
5645 | |
5646 | /* Reduce allocated memory if we can, trying to get |
5647 | * the socket within its memory limits again. |
5648 | * |
5649 | * Return less than zero if we should start dropping frames |
5650 | * until the socket owning process reads some of the data |
5651 | * to stabilize the situation. |
5652 | */ |
5653 | static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb) |
5654 | { |
5655 | struct tcp_sock *tp = tcp_sk(sk); |
5656 | |
5657 | NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED); |
5658 | |
5659 | if (atomic_read(v: &sk->sk_rmem_alloc) >= sk->sk_rcvbuf) |
5660 | tcp_clamp_window(sk); |
5661 | else if (tcp_under_memory_pressure(sk)) |
5662 | tcp_adjust_rcv_ssthresh(sk); |
5663 | |
5664 | if (atomic_read(v: &sk->sk_rmem_alloc) <= sk->sk_rcvbuf) |
5665 | return 0; |
5666 | |
5667 | tcp_collapse_ofo_queue(sk); |
5668 | if (!skb_queue_empty(list: &sk->sk_receive_queue)) |
5669 | tcp_collapse(sk, list: &sk->sk_receive_queue, NULL, |
5670 | head: skb_peek(list_: &sk->sk_receive_queue), |
5671 | NULL, |
5672 | start: tp->copied_seq, end: tp->rcv_nxt); |
5673 | |
5674 | if (atomic_read(v: &sk->sk_rmem_alloc) <= sk->sk_rcvbuf) |
5675 | return 0; |
5676 | |
5677 | /* Collapsing did not help, destructive actions follow. |
5678 | * This must not ever occur. */ |
5679 | |
5680 | tcp_prune_ofo_queue(sk, in_skb); |
5681 | |
5682 | if (atomic_read(v: &sk->sk_rmem_alloc) <= sk->sk_rcvbuf) |
5683 | return 0; |
5684 | |
5685 | /* If we are really being abused, tell the caller to silently |
5686 | * drop receive data on the floor. It will get retransmitted |
5687 | * and hopefully then we'll have sufficient space. |
5688 | */ |
5689 | NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED); |
5690 | |
5691 | /* Massive buffer overcommit. */ |
5692 | tp->pred_flags = 0; |
5693 | return -1; |
5694 | } |
5695 | |
5696 | static bool tcp_should_expand_sndbuf(struct sock *sk) |
5697 | { |
5698 | const struct tcp_sock *tp = tcp_sk(sk); |
5699 | |
5700 | /* If the user specified a specific send buffer setting, do |
5701 | * not modify it. |
5702 | */ |
5703 | if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) |
5704 | return false; |
5705 | |
5706 | /* If we are under global TCP memory pressure, do not expand. */ |
5707 | if (tcp_under_memory_pressure(sk)) { |
5708 | int unused_mem = sk_unused_reserved_mem(sk); |
5709 | |
5710 | /* Adjust sndbuf according to reserved mem. But make sure |
5711 | * it never goes below SOCK_MIN_SNDBUF. |
5712 | * See sk_stream_moderate_sndbuf() for more details. |
5713 | */ |
5714 | if (unused_mem > SOCK_MIN_SNDBUF) |
5715 | WRITE_ONCE(sk->sk_sndbuf, unused_mem); |
5716 | |
5717 | return false; |
5718 | } |
5719 | |
5720 | /* If we are under soft global TCP memory pressure, do not expand. */ |
5721 | if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, index: 0)) |
5722 | return false; |
5723 | |
5724 | /* If we filled the congestion window, do not expand. */ |
5725 | if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)) |
5726 | return false; |
5727 | |
5728 | return true; |
5729 | } |
5730 | |
5731 | static void tcp_new_space(struct sock *sk) |
5732 | { |
5733 | struct tcp_sock *tp = tcp_sk(sk); |
5734 | |
5735 | if (tcp_should_expand_sndbuf(sk)) { |
5736 | tcp_sndbuf_expand(sk); |
5737 | tp->snd_cwnd_stamp = tcp_jiffies32; |
5738 | } |
5739 | |
5740 | INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk); |
5741 | } |
5742 | |
5743 | /* Caller made space either from: |
5744 | * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced) |
5745 | * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt) |
5746 | * |
5747 | * We might be able to generate EPOLLOUT to the application if: |
5748 | * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2 |
5749 | * 2) notsent amount (tp->write_seq - tp->snd_nxt) became |
5750 | * small enough that tcp_stream_memory_free() decides it |
5751 | * is time to generate EPOLLOUT. |
5752 | */ |
5753 | void tcp_check_space(struct sock *sk) |
5754 | { |
5755 | /* pairs with tcp_poll() */ |
5756 | smp_mb(); |
5757 | if (sk->sk_socket && |
5758 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { |
5759 | tcp_new_space(sk); |
5760 | if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) |
5761 | tcp_chrono_stop(sk, type: TCP_CHRONO_SNDBUF_LIMITED); |
5762 | } |
5763 | } |
5764 | |
5765 | static inline void tcp_data_snd_check(struct sock *sk) |
5766 | { |
5767 | tcp_push_pending_frames(sk); |
5768 | tcp_check_space(sk); |
5769 | } |
5770 | |
5771 | /* |
5772 | * Check if sending an ack is needed. |
5773 | */ |
5774 | static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) |
5775 | { |
5776 | struct tcp_sock *tp = tcp_sk(sk); |
5777 | unsigned long rtt, delay; |
5778 | |
5779 | /* More than one full frame received... */ |
5780 | if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && |
5781 | /* ... and right edge of window advances far enough. |
5782 | * (tcp_recvmsg() will send ACK otherwise). |
5783 | * If application uses SO_RCVLOWAT, we want send ack now if |
5784 | * we have not received enough bytes to satisfy the condition. |
5785 | */ |
5786 | (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat || |
5787 | __tcp_select_window(sk) >= tp->rcv_wnd)) || |
5788 | /* We ACK each frame or... */ |
5789 | tcp_in_quickack_mode(sk) || |
5790 | /* Protocol state mandates a one-time immediate ACK */ |
5791 | inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) { |
5792 | /* If we are running from __release_sock() in user context, |
5793 | * Defer the ack until tcp_release_cb(). |
5794 | */ |
5795 | if (sock_owned_by_user_nocheck(sk) && |
5796 | READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_backlog_ack_defer)) { |
5797 | set_bit(nr: TCP_ACK_DEFERRED, addr: &sk->sk_tsq_flags); |
5798 | return; |
5799 | } |
5800 | send_now: |
5801 | tcp_send_ack(sk); |
5802 | return; |
5803 | } |
5804 | |
5805 | if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) { |
5806 | tcp_send_delayed_ack(sk); |
5807 | return; |
5808 | } |
5809 | |
5810 | if (!tcp_is_sack(tp) || |
5811 | tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)) |
5812 | goto send_now; |
5813 | |
5814 | if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) { |
5815 | tp->compressed_ack_rcv_nxt = tp->rcv_nxt; |
5816 | tp->dup_ack_counter = 0; |
5817 | } |
5818 | if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) { |
5819 | tp->dup_ack_counter++; |
5820 | goto send_now; |
5821 | } |
5822 | tp->compressed_ack++; |
5823 | if (hrtimer_is_queued(timer: &tp->compressed_ack_timer)) |
5824 | return; |
5825 | |
5826 | /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */ |
5827 | |
5828 | rtt = tp->rcv_rtt_est.rtt_us; |
5829 | if (tp->srtt_us && tp->srtt_us < rtt) |
5830 | rtt = tp->srtt_us; |
5831 | |
5832 | delay = min_t(unsigned long, |
5833 | READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns), |
5834 | rtt * (NSEC_PER_USEC >> 3)/20); |
5835 | sock_hold(sk); |
5836 | hrtimer_start_range_ns(timer: &tp->compressed_ack_timer, tim: ns_to_ktime(ns: delay), |
5837 | READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns), |
5838 | mode: HRTIMER_MODE_REL_PINNED_SOFT); |
5839 | } |
5840 | |
5841 | static inline void tcp_ack_snd_check(struct sock *sk) |
5842 | { |
5843 | if (!inet_csk_ack_scheduled(sk)) { |
5844 | /* We sent a data segment already. */ |
5845 | return; |
5846 | } |
5847 | __tcp_ack_snd_check(sk, ofo_possible: 1); |
5848 | } |
5849 | |
5850 | /* |
5851 | * This routine is only called when we have urgent data |
5852 | * signaled. Its the 'slow' part of tcp_urg. It could be |
5853 | * moved inline now as tcp_urg is only called from one |
5854 | * place. We handle URGent data wrong. We have to - as |
5855 | * BSD still doesn't use the correction from RFC961. |
5856 | * For 1003.1g we should support a new option TCP_STDURG to permit |
5857 | * either form (or just set the sysctl tcp_stdurg). |
5858 | */ |
5859 | |
5860 | static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) |
5861 | { |
5862 | struct tcp_sock *tp = tcp_sk(sk); |
5863 | u32 ptr = ntohs(th->urg_ptr); |
5864 | |
5865 | if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg)) |
5866 | ptr--; |
5867 | ptr += ntohl(th->seq); |
5868 | |
5869 | /* Ignore urgent data that we've already seen and read. */ |
5870 | if (after(tp->copied_seq, ptr)) |
5871 | return; |
5872 | |
5873 | /* Do not replay urg ptr. |
5874 | * |
5875 | * NOTE: interesting situation not covered by specs. |
5876 | * Misbehaving sender may send urg ptr, pointing to segment, |
5877 | * which we already have in ofo queue. We are not able to fetch |
5878 | * such data and will stay in TCP_URG_NOTYET until will be eaten |
5879 | * by recvmsg(). Seems, we are not obliged to handle such wicked |
5880 | * situations. But it is worth to think about possibility of some |
5881 | * DoSes using some hypothetical application level deadlock. |
5882 | */ |
5883 | if (before(seq1: ptr, seq2: tp->rcv_nxt)) |
5884 | return; |
5885 | |
5886 | /* Do we already have a newer (or duplicate) urgent pointer? */ |
5887 | if (tp->urg_data && !after(ptr, tp->urg_seq)) |
5888 | return; |
5889 | |
5890 | /* Tell the world about our new urgent pointer. */ |
5891 | sk_send_sigurg(sk); |
5892 | |
5893 | /* We may be adding urgent data when the last byte read was |
5894 | * urgent. To do this requires some care. We cannot just ignore |
5895 | * tp->copied_seq since we would read the last urgent byte again |
5896 | * as data, nor can we alter copied_seq until this data arrives |
5897 | * or we break the semantics of SIOCATMARK (and thus sockatmark()) |
5898 | * |
5899 | * NOTE. Double Dutch. Rendering to plain English: author of comment |
5900 | * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); |
5901 | * and expect that both A and B disappear from stream. This is _wrong_. |
5902 | * Though this happens in BSD with high probability, this is occasional. |
5903 | * Any application relying on this is buggy. Note also, that fix "works" |
5904 | * only in this artificial test. Insert some normal data between A and B and we will |
5905 | * decline of BSD again. Verdict: it is better to remove to trap |
5906 | * buggy users. |
5907 | */ |
5908 | if (tp->urg_seq == tp->copied_seq && tp->urg_data && |
5909 | !sock_flag(sk, flag: SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { |
5910 | struct sk_buff *skb = skb_peek(list_: &sk->sk_receive_queue); |
5911 | tp->copied_seq++; |
5912 | if (skb && !before(seq1: tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { |
5913 | __skb_unlink(skb, list: &sk->sk_receive_queue); |
5914 | __kfree_skb(skb); |
5915 | } |
5916 | } |
5917 | |
5918 | WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET); |
5919 | WRITE_ONCE(tp->urg_seq, ptr); |
5920 | |
5921 | /* Disable header prediction. */ |
5922 | tp->pred_flags = 0; |
5923 | } |
5924 | |
5925 | /* This is the 'fast' part of urgent handling. */ |
5926 | static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) |
5927 | { |
5928 | struct tcp_sock *tp = tcp_sk(sk); |
5929 | |
5930 | /* Check if we get a new urgent pointer - normally not. */ |
5931 | if (unlikely(th->urg)) |
5932 | tcp_check_urg(sk, th); |
5933 | |
5934 | /* Do we wait for any urgent data? - normally not... */ |
5935 | if (unlikely(tp->urg_data == TCP_URG_NOTYET)) { |
5936 | u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - |
5937 | th->syn; |
5938 | |
5939 | /* Is the urgent pointer pointing into this packet? */ |
5940 | if (ptr < skb->len) { |
5941 | u8 tmp; |
5942 | if (skb_copy_bits(skb, offset: ptr, to: &tmp, len: 1)) |
5943 | BUG(); |
5944 | WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp); |
5945 | if (!sock_flag(sk, flag: SOCK_DEAD)) |
5946 | sk->sk_data_ready(sk); |
5947 | } |
5948 | } |
5949 | } |
5950 | |
5951 | /* Accept RST for rcv_nxt - 1 after a FIN. |
5952 | * When tcp connections are abruptly terminated from Mac OSX (via ^C), a |
5953 | * FIN is sent followed by a RST packet. The RST is sent with the same |
5954 | * sequence number as the FIN, and thus according to RFC 5961 a challenge |
5955 | * ACK should be sent. However, Mac OSX rate limits replies to challenge |
5956 | * ACKs on the closed socket. In addition middleboxes can drop either the |
5957 | * challenge ACK or a subsequent RST. |
5958 | */ |
5959 | static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb) |
5960 | { |
5961 | const struct tcp_sock *tp = tcp_sk(sk); |
5962 | |
5963 | return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) && |
5964 | (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK | |
5965 | TCPF_CLOSING)); |
5966 | } |
5967 | |
5968 | /* Does PAWS and seqno based validation of an incoming segment, flags will |
5969 | * play significant role here. |
5970 | */ |
5971 | static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, |
5972 | const struct tcphdr *th, int syn_inerr) |
5973 | { |
5974 | struct tcp_sock *tp = tcp_sk(sk); |
5975 | SKB_DR(reason); |
5976 | |
5977 | /* RFC1323: H1. Apply PAWS check first. */ |
5978 | if (!tcp_fast_parse_options(net: sock_net(sk), skb, th, tp) || |
5979 | !tp->rx_opt.saw_tstamp || |
5980 | tcp_paws_check(rx_opt: &tp->rx_opt, TCP_PAWS_WINDOW)) |
5981 | goto step1; |
5982 | |
5983 | reason = tcp_disordered_ack_check(sk, skb); |
5984 | if (!reason) |
5985 | goto step1; |
5986 | /* Reset is accepted even if it did not pass PAWS. */ |
5987 | if (th->rst) |
5988 | goto step1; |
5989 | if (unlikely(th->syn)) |
5990 | goto syn_challenge; |
5991 | |
5992 | /* Old ACK are common, increment PAWS_OLD_ACK |
5993 | * and do not send a dupack. |
5994 | */ |
5995 | if (reason == SKB_DROP_REASON_TCP_RFC7323_PAWS_ACK) { |
5996 | NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWS_OLD_ACK); |
5997 | goto discard; |
5998 | } |
5999 | NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); |
6000 | if (!tcp_oow_rate_limited(net: sock_net(sk), skb, |
6001 | mib_idx: LINUX_MIB_TCPACKSKIPPEDPAWS, |
6002 | last_oow_ack_time: &tp->last_oow_ack_time)) |
6003 | tcp_send_dupack(sk, skb); |
6004 | goto discard; |
6005 | |
6006 | step1: |
6007 | /* Step 1: check sequence number */ |
6008 | reason = tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); |
6009 | if (reason) { |
6010 | /* RFC793, page 37: "In all states except SYN-SENT, all reset |
6011 | * (RST) segments are validated by checking their SEQ-fields." |
6012 | * And page 69: "If an incoming segment is not acceptable, |
6013 | * an acknowledgment should be sent in reply (unless the RST |
6014 | * bit is set, if so drop the segment and return)". |
6015 | */ |
6016 | if (!th->rst) { |
6017 | if (th->syn) |
6018 | goto syn_challenge; |
6019 | if (!tcp_oow_rate_limited(net: sock_net(sk), skb, |
6020 | mib_idx: LINUX_MIB_TCPACKSKIPPEDSEQ, |
6021 | last_oow_ack_time: &tp->last_oow_ack_time)) |
6022 | tcp_send_dupack(sk, skb); |
6023 | } else if (tcp_reset_check(sk, skb)) { |
6024 | goto reset; |
6025 | } |
6026 | goto discard; |
6027 | } |
6028 | |
6029 | /* Step 2: check RST bit */ |
6030 | if (th->rst) { |
6031 | /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a |
6032 | * FIN and SACK too if available): |
6033 | * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or |
6034 | * the right-most SACK block, |
6035 | * then |
6036 | * RESET the connection |
6037 | * else |
6038 | * Send a challenge ACK |
6039 | */ |
6040 | if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt || |
6041 | tcp_reset_check(sk, skb)) |
6042 | goto reset; |
6043 | |
6044 | if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) { |
6045 | struct tcp_sack_block *sp = &tp->selective_acks[0]; |
6046 | int max_sack = sp[0].end_seq; |
6047 | int this_sack; |
6048 | |
6049 | for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; |
6050 | ++this_sack) { |
6051 | max_sack = after(sp[this_sack].end_seq, |
6052 | max_sack) ? |
6053 | sp[this_sack].end_seq : max_sack; |
6054 | } |
6055 | |
6056 | if (TCP_SKB_CB(skb)->seq == max_sack) |
6057 | goto reset; |
6058 | } |
6059 | |
6060 | /* Disable TFO if RST is out-of-order |
6061 | * and no data has been received |
6062 | * for current active TFO socket |
6063 | */ |
6064 | if (tp->syn_fastopen && !tp->data_segs_in && |
6065 | sk->sk_state == TCP_ESTABLISHED) |
6066 | tcp_fastopen_active_disable(sk); |
6067 | tcp_send_challenge_ack(sk); |
6068 | SKB_DR_SET(reason, TCP_RESET); |
6069 | goto discard; |
6070 | } |
6071 | |
6072 | /* step 3: check security and precedence [ignored] */ |
6073 | |
6074 | /* step 4: Check for a SYN |
6075 | * RFC 5961 4.2 : Send a challenge ack |
6076 | */ |
6077 | if (th->syn) { |
6078 | if (sk->sk_state == TCP_SYN_RECV && sk->sk_socket && th->ack && |
6079 | TCP_SKB_CB(skb)->seq + 1 == TCP_SKB_CB(skb)->end_seq && |
6080 | TCP_SKB_CB(skb)->seq + 1 == tp->rcv_nxt && |
6081 | TCP_SKB_CB(skb)->ack_seq == tp->snd_nxt) |
6082 | goto pass; |
6083 | syn_challenge: |
6084 | if (syn_inerr) |
6085 | TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); |
6086 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE); |
6087 | tcp_send_challenge_ack(sk); |
6088 | SKB_DR_SET(reason, TCP_INVALID_SYN); |
6089 | goto discard; |
6090 | } |
6091 | |
6092 | pass: |
6093 | bpf_skops_parse_hdr(sk, skb); |
6094 | |
6095 | return true; |
6096 | |
6097 | discard: |
6098 | tcp_drop_reason(sk, skb, reason); |
6099 | return false; |
6100 | |
6101 | reset: |
6102 | tcp_reset(sk, skb); |
6103 | __kfree_skb(skb); |
6104 | return false; |
6105 | } |
6106 | |
6107 | /* |
6108 | * TCP receive function for the ESTABLISHED state. |
6109 | * |
6110 | * It is split into a fast path and a slow path. The fast path is |
6111 | * disabled when: |
6112 | * - A zero window was announced from us - zero window probing |
6113 | * is only handled properly in the slow path. |
6114 | * - Out of order segments arrived. |
6115 | * - Urgent data is expected. |
6116 | * - There is no buffer space left |
6117 | * - Unexpected TCP flags/window values/header lengths are received |
6118 | * (detected by checking the TCP header against pred_flags) |
6119 | * - Data is sent in both directions. Fast path only supports pure senders |
6120 | * or pure receivers (this means either the sequence number or the ack |
6121 | * value must stay constant) |
6122 | * - Unexpected TCP option. |
6123 | * |
6124 | * When these conditions are not satisfied it drops into a standard |
6125 | * receive procedure patterned after RFC793 to handle all cases. |
6126 | * The first three cases are guaranteed by proper pred_flags setting, |
6127 | * the rest is checked inline. Fast processing is turned on in |
6128 | * tcp_data_queue when everything is OK. |
6129 | */ |
6130 | void tcp_rcv_established(struct sock *sk, struct sk_buff *skb) |
6131 | { |
6132 | enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED; |
6133 | const struct tcphdr *th = (const struct tcphdr *)skb->data; |
6134 | struct tcp_sock *tp = tcp_sk(sk); |
6135 | unsigned int len = skb->len; |
6136 | |
6137 | /* TCP congestion window tracking */ |
6138 | trace_tcp_probe(sk, skb); |
6139 | |
6140 | tcp_mstamp_refresh(tp); |
6141 | if (unlikely(!rcu_access_pointer(sk->sk_rx_dst))) |
6142 | inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb); |
6143 | /* |
6144 | * Header prediction. |
6145 | * The code loosely follows the one in the famous |
6146 | * "30 instruction TCP receive" Van Jacobson mail. |
6147 | * |
6148 | * Van's trick is to deposit buffers into socket queue |
6149 | * on a device interrupt, to call tcp_recv function |
6150 | * on the receive process context and checksum and copy |
6151 | * the buffer to user space. smart... |
6152 | * |
6153 | * Our current scheme is not silly either but we take the |
6154 | * extra cost of the net_bh soft interrupt processing... |
6155 | * We do checksum and copy also but from device to kernel. |
6156 | */ |
6157 | |
6158 | tp->rx_opt.saw_tstamp = 0; |
6159 | |
6160 | /* pred_flags is 0xS?10 << 16 + snd_wnd |
6161 | * if header_prediction is to be made |
6162 | * 'S' will always be tp->tcp_header_len >> 2 |
6163 | * '?' will be 0 for the fast path, otherwise pred_flags is 0 to |
6164 | * turn it off (when there are holes in the receive |
6165 | * space for instance) |
6166 | * PSH flag is ignored. |
6167 | */ |
6168 | |
6169 | if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && |
6170 | TCP_SKB_CB(skb)->seq == tp->rcv_nxt && |
6171 | !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { |
6172 | int tcp_header_len = tp->tcp_header_len; |
6173 | s32 delta = 0; |
6174 | int flag = 0; |
6175 | |
6176 | /* Timestamp header prediction: tcp_header_len |
6177 | * is automatically equal to th->doff*4 due to pred_flags |
6178 | * match. |
6179 | */ |
6180 | |
6181 | /* Check timestamp */ |
6182 | if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { |
6183 | /* No? Slow path! */ |
6184 | if (!tcp_parse_aligned_timestamp(tp, th)) |
6185 | goto slow_path; |
6186 | |
6187 | delta = tp->rx_opt.rcv_tsval - |
6188 | tp->rx_opt.ts_recent; |
6189 | /* If PAWS failed, check it more carefully in slow path */ |
6190 | if (delta < 0) |
6191 | goto slow_path; |
6192 | |
6193 | /* DO NOT update ts_recent here, if checksum fails |
6194 | * and timestamp was corrupted part, it will result |
6195 | * in a hung connection since we will drop all |
6196 | * future packets due to the PAWS test. |
6197 | */ |
6198 | } |
6199 | |
6200 | if (len <= tcp_header_len) { |
6201 | /* Bulk data transfer: sender */ |
6202 | if (len == tcp_header_len) { |
6203 | /* Predicted packet is in window by definition. |
6204 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. |
6205 | * Hence, check seq<=rcv_wup reduces to: |
6206 | */ |
6207 | if (tcp_header_len == |
6208 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && |
6209 | tp->rcv_nxt == tp->rcv_wup) |
6210 | flag |= __tcp_replace_ts_recent(tp, |
6211 | tstamp_delta: delta); |
6212 | |
6213 | /* We know that such packets are checksummed |
6214 | * on entry. |
6215 | */ |
6216 | tcp_ack(sk, skb, flag); |
6217 | __kfree_skb(skb); |
6218 | tcp_data_snd_check(sk); |
6219 | /* When receiving pure ack in fast path, update |
6220 | * last ts ecr directly instead of calling |
6221 | * tcp_rcv_rtt_measure_ts() |
6222 | */ |
6223 | tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; |
6224 | return; |
6225 | } else { /* Header too small */ |
6226 | reason = SKB_DROP_REASON_PKT_TOO_SMALL; |
6227 | TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); |
6228 | goto discard; |
6229 | } |
6230 | } else { |
6231 | int eaten = 0; |
6232 | bool fragstolen = false; |
6233 | |
6234 | if (tcp_checksum_complete(skb)) |
6235 | goto csum_error; |
6236 | |
6237 | if ((int)skb->truesize > sk->sk_forward_alloc) |
6238 | goto step5; |
6239 | |
6240 | /* Predicted packet is in window by definition. |
6241 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. |
6242 | * Hence, check seq<=rcv_wup reduces to: |
6243 | */ |
6244 | if (tcp_header_len == |
6245 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && |
6246 | tp->rcv_nxt == tp->rcv_wup) |
6247 | flag |= __tcp_replace_ts_recent(tp, |
6248 | tstamp_delta: delta); |
6249 | |
6250 | tcp_rcv_rtt_measure_ts(sk, skb); |
6251 | |
6252 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS); |
6253 | |
6254 | /* Bulk data transfer: receiver */ |
6255 | tcp_cleanup_skb(skb); |
6256 | __skb_pull(skb, len: tcp_header_len); |
6257 | eaten = tcp_queue_rcv(sk, skb, fragstolen: &fragstolen); |
6258 | |
6259 | tcp_event_data_recv(sk, skb); |
6260 | |
6261 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { |
6262 | /* Well, only one small jumplet in fast path... */ |
6263 | tcp_ack(sk, skb, flag: flag | FLAG_DATA); |
6264 | tcp_data_snd_check(sk); |
6265 | if (!inet_csk_ack_scheduled(sk)) |
6266 | goto no_ack; |
6267 | } else { |
6268 | tcp_update_wl(tp, TCP_SKB_CB(skb)->seq); |
6269 | } |
6270 | |
6271 | __tcp_ack_snd_check(sk, ofo_possible: 0); |
6272 | no_ack: |
6273 | if (eaten) |
6274 | kfree_skb_partial(skb, head_stolen: fragstolen); |
6275 | tcp_data_ready(sk); |
6276 | return; |
6277 | } |
6278 | } |
6279 | |
6280 | slow_path: |
6281 | if (len < (th->doff << 2) || tcp_checksum_complete(skb)) |
6282 | goto csum_error; |
6283 | |
6284 | if (!th->ack && !th->rst && !th->syn) { |
6285 | reason = SKB_DROP_REASON_TCP_FLAGS; |
6286 | goto discard; |
6287 | } |
6288 | |
6289 | /* |
6290 | * Standard slow path. |
6291 | */ |
6292 | |
6293 | if (!tcp_validate_incoming(sk, skb, th, syn_inerr: 1)) |
6294 | return; |
6295 | |
6296 | step5: |
6297 | reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT); |
6298 | if ((int)reason < 0) { |
6299 | reason = -reason; |
6300 | goto discard; |
6301 | } |
6302 | tcp_rcv_rtt_measure_ts(sk, skb); |
6303 | |
6304 | /* Process urgent data. */ |
6305 | tcp_urg(sk, skb, th); |
6306 | |
6307 | /* step 7: process the segment text */ |
6308 | tcp_data_queue(sk, skb); |
6309 | |
6310 | tcp_data_snd_check(sk); |
6311 | tcp_ack_snd_check(sk); |
6312 | return; |
6313 | |
6314 | csum_error: |
6315 | reason = SKB_DROP_REASON_TCP_CSUM; |
6316 | trace_tcp_bad_csum(skb); |
6317 | TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS); |
6318 | TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); |
6319 | |
6320 | discard: |
6321 | tcp_drop_reason(sk, skb, reason); |
6322 | } |
6323 | EXPORT_IPV6_MOD(tcp_rcv_established); |
6324 | |
6325 | void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb) |
6326 | { |
6327 | struct inet_connection_sock *icsk = inet_csk(sk); |
6328 | struct tcp_sock *tp = tcp_sk(sk); |
6329 | |
6330 | tcp_mtup_init(sk); |
6331 | icsk->icsk_af_ops->rebuild_header(sk); |
6332 | tcp_init_metrics(sk); |
6333 | |
6334 | /* Initialize the congestion window to start the transfer. |
6335 | * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been |
6336 | * retransmitted. In light of RFC6298 more aggressive 1sec |
6337 | * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK |
6338 | * retransmission has occurred. |
6339 | */ |
6340 | if (tp->total_retrans > 1 && tp->undo_marker) |
6341 | tcp_snd_cwnd_set(tp, val: 1); |
6342 | else |
6343 | tcp_snd_cwnd_set(tp, val: tcp_init_cwnd(tp, dst: __sk_dst_get(sk))); |
6344 | tp->snd_cwnd_stamp = tcp_jiffies32; |
6345 | |
6346 | bpf_skops_established(sk, bpf_op, skb); |
6347 | /* Initialize congestion control unless BPF initialized it already: */ |
6348 | if (!icsk->icsk_ca_initialized) |
6349 | tcp_init_congestion_control(sk); |
6350 | tcp_init_buffer_space(sk); |
6351 | } |
6352 | |
6353 | void tcp_finish_connect(struct sock *sk, struct sk_buff *skb) |
6354 | { |
6355 | struct tcp_sock *tp = tcp_sk(sk); |
6356 | struct inet_connection_sock *icsk = inet_csk(sk); |
6357 | |
6358 | tcp_ao_finish_connect(sk, skb); |
6359 | tcp_set_state(sk, state: TCP_ESTABLISHED); |
6360 | icsk->icsk_ack.lrcvtime = tcp_jiffies32; |
6361 | |
6362 | if (skb) { |
6363 | icsk->icsk_af_ops->sk_rx_dst_set(sk, skb); |
6364 | security_inet_conn_established(sk, skb); |
6365 | sk_mark_napi_id(sk, skb); |
6366 | } |
6367 | |
6368 | tcp_init_transfer(sk, bpf_op: BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb); |
6369 | |
6370 | /* Prevent spurious tcp_cwnd_restart() on first data |
6371 | * packet. |
6372 | */ |
6373 | tp->lsndtime = tcp_jiffies32; |
6374 | |
6375 | if (sock_flag(sk, flag: SOCK_KEEPOPEN)) |
6376 | tcp_reset_keepalive_timer(sk, timeout: keepalive_time_when(tp)); |
6377 | |
6378 | if (!tp->rx_opt.snd_wscale) |
6379 | __tcp_fast_path_on(tp, snd_wnd: tp->snd_wnd); |
6380 | else |
6381 | tp->pred_flags = 0; |
6382 | } |
6383 | |
6384 | static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack, |
6385 | struct tcp_fastopen_cookie *cookie) |
6386 | { |
6387 | struct tcp_sock *tp = tcp_sk(sk); |
6388 | struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL; |
6389 | u16 mss = tp->rx_opt.mss_clamp, try_exp = 0; |
6390 | bool syn_drop = false; |
6391 | |
6392 | if (mss == tp->rx_opt.user_mss) { |
6393 | struct tcp_options_received opt; |
6394 | |
6395 | /* Get original SYNACK MSS value if user MSS sets mss_clamp */ |
6396 | tcp_clear_options(rx_opt: &opt); |
6397 | opt.user_mss = opt.mss_clamp = 0; |
6398 | tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL); |
6399 | mss = opt.mss_clamp; |
6400 | } |
6401 | |
6402 | if (!tp->syn_fastopen) { |
6403 | /* Ignore an unsolicited cookie */ |
6404 | cookie->len = -1; |
6405 | } else if (tp->total_retrans) { |
6406 | /* SYN timed out and the SYN-ACK neither has a cookie nor |
6407 | * acknowledges data. Presumably the remote received only |
6408 | * the retransmitted (regular) SYNs: either the original |
6409 | * SYN-data or the corresponding SYN-ACK was dropped. |
6410 | */ |
6411 | syn_drop = (cookie->len < 0 && data); |
6412 | } else if (cookie->len < 0 && !tp->syn_data) { |
6413 | /* We requested a cookie but didn't get it. If we did not use |
6414 | * the (old) exp opt format then try so next time (try_exp=1). |
6415 | * Otherwise we go back to use the RFC7413 opt (try_exp=2). |
6416 | */ |
6417 | try_exp = tp->syn_fastopen_exp ? 2 : 1; |
6418 | } |
6419 | |
6420 | tcp_fastopen_cache_set(sk, mss, cookie, syn_lost: syn_drop, try_exp); |
6421 | |
6422 | if (data) { /* Retransmit unacked data in SYN */ |
6423 | if (tp->total_retrans) |
6424 | tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED; |
6425 | else |
6426 | tp->fastopen_client_fail = TFO_DATA_NOT_ACKED; |
6427 | skb_rbtree_walk_from(data) |
6428 | tcp_mark_skb_lost(sk, skb: data); |
6429 | tcp_non_congestion_loss_retransmit(sk); |
6430 | NET_INC_STATS(sock_net(sk), |
6431 | LINUX_MIB_TCPFASTOPENACTIVEFAIL); |
6432 | return true; |
6433 | } |
6434 | tp->syn_data_acked = tp->syn_data; |
6435 | if (tp->syn_data_acked) { |
6436 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); |
6437 | /* SYN-data is counted as two separate packets in tcp_ack() */ |
6438 | if (tp->delivered > 1) |
6439 | --tp->delivered; |
6440 | } |
6441 | |
6442 | tcp_fastopen_add_skb(sk, skb: synack); |
6443 | |
6444 | return false; |
6445 | } |
6446 | |
6447 | static void smc_check_reset_syn(struct tcp_sock *tp) |
6448 | { |
6449 | #if IS_ENABLED(CONFIG_SMC) |
6450 | if (static_branch_unlikely(&tcp_have_smc)) { |
6451 | if (tp->syn_smc && !tp->rx_opt.smc_ok) |
6452 | tp->syn_smc = 0; |
6453 | } |
6454 | #endif |
6455 | } |
6456 | |
6457 | static void tcp_try_undo_spurious_syn(struct sock *sk) |
6458 | { |
6459 | struct tcp_sock *tp = tcp_sk(sk); |
6460 | u32 syn_stamp; |
6461 | |
6462 | /* undo_marker is set when SYN or SYNACK times out. The timeout is |
6463 | * spurious if the ACK's timestamp option echo value matches the |
6464 | * original SYN timestamp. |
6465 | */ |
6466 | syn_stamp = tp->retrans_stamp; |
6467 | if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp && |
6468 | syn_stamp == tp->rx_opt.rcv_tsecr) |
6469 | tp->undo_marker = 0; |
6470 | } |
6471 | |
6472 | static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, |
6473 | const struct tcphdr *th) |
6474 | { |
6475 | struct inet_connection_sock *icsk = inet_csk(sk); |
6476 | struct tcp_sock *tp = tcp_sk(sk); |
6477 | struct tcp_fastopen_cookie foc = { .len = -1 }; |
6478 | int saved_clamp = tp->rx_opt.mss_clamp; |
6479 | bool fastopen_fail; |
6480 | SKB_DR(reason); |
6481 | |
6482 | tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc); |
6483 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) |
6484 | tp->rx_opt.rcv_tsecr -= tp->tsoffset; |
6485 | |
6486 | if (th->ack) { |
6487 | /* rfc793: |
6488 | * "If the state is SYN-SENT then |
6489 | * first check the ACK bit |
6490 | * If the ACK bit is set |
6491 | * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send |
6492 | * a reset (unless the RST bit is set, if so drop |
6493 | * the segment and return)" |
6494 | */ |
6495 | if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) || |
6496 | after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { |
6497 | /* Previous FIN/ACK or RST/ACK might be ignored. */ |
6498 | if (icsk->icsk_retransmits == 0) |
6499 | tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
6500 | TCP_TIMEOUT_MIN, pace_delay: false); |
6501 | SKB_DR_SET(reason, TCP_INVALID_ACK_SEQUENCE); |
6502 | goto reset_and_undo; |
6503 | } |
6504 | |
6505 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && |
6506 | !between(seq1: tp->rx_opt.rcv_tsecr, seq2: tp->retrans_stamp, |
6507 | seq3: tcp_time_stamp_ts(tp))) { |
6508 | NET_INC_STATS(sock_net(sk), |
6509 | LINUX_MIB_PAWSACTIVEREJECTED); |
6510 | SKB_DR_SET(reason, TCP_RFC7323_PAWS); |
6511 | goto reset_and_undo; |
6512 | } |
6513 | |
6514 | /* Now ACK is acceptable. |
6515 | * |
6516 | * "If the RST bit is set |
6517 | * If the ACK was acceptable then signal the user "error: |
6518 | * connection reset", drop the segment, enter CLOSED state, |
6519 | * delete TCB, and return." |
6520 | */ |
6521 | |
6522 | if (th->rst) { |
6523 | tcp_reset(sk, skb); |
6524 | consume: |
6525 | __kfree_skb(skb); |
6526 | return 0; |
6527 | } |
6528 | |
6529 | /* rfc793: |
6530 | * "fifth, if neither of the SYN or RST bits is set then |
6531 | * drop the segment and return." |
6532 | * |
6533 | * See note below! |
6534 | * --ANK(990513) |
6535 | */ |
6536 | if (!th->syn) { |
6537 | SKB_DR_SET(reason, TCP_FLAGS); |
6538 | goto discard_and_undo; |
6539 | } |
6540 | /* rfc793: |
6541 | * "If the SYN bit is on ... |
6542 | * are acceptable then ... |
6543 | * (our SYN has been ACKed), change the connection |
6544 | * state to ESTABLISHED..." |
6545 | */ |
6546 | |
6547 | tcp_ecn_rcv_synack(tp, th); |
6548 | |
6549 | tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); |
6550 | tcp_try_undo_spurious_syn(sk); |
6551 | tcp_ack(sk, skb, FLAG_SLOWPATH); |
6552 | |
6553 | /* Ok.. it's good. Set up sequence numbers and |
6554 | * move to established. |
6555 | */ |
6556 | WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); |
6557 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; |
6558 | |
6559 | /* RFC1323: The window in SYN & SYN/ACK segments is |
6560 | * never scaled. |
6561 | */ |
6562 | tp->snd_wnd = ntohs(th->window); |
6563 | |
6564 | if (!tp->rx_opt.wscale_ok) { |
6565 | tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; |
6566 | WRITE_ONCE(tp->window_clamp, |
6567 | min(tp->window_clamp, 65535U)); |
6568 | } |
6569 | |
6570 | if (tp->rx_opt.saw_tstamp) { |
6571 | tp->rx_opt.tstamp_ok = 1; |
6572 | tp->tcp_header_len = |
6573 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; |
6574 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; |
6575 | tcp_store_ts_recent(tp); |
6576 | } else { |
6577 | tp->tcp_header_len = sizeof(struct tcphdr); |
6578 | } |
6579 | |
6580 | tcp_sync_mss(sk, pmtu: icsk->icsk_pmtu_cookie); |
6581 | tcp_initialize_rcv_mss(sk); |
6582 | |
6583 | /* Remember, tcp_poll() does not lock socket! |
6584 | * Change state from SYN-SENT only after copied_seq |
6585 | * is initialized. */ |
6586 | WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); |
6587 | |
6588 | smc_check_reset_syn(tp); |
6589 | |
6590 | smp_mb(); |
6591 | |
6592 | tcp_finish_connect(sk, skb); |
6593 | |
6594 | fastopen_fail = (tp->syn_fastopen || tp->syn_data) && |
6595 | tcp_rcv_fastopen_synack(sk, synack: skb, cookie: &foc); |
6596 | |
6597 | if (!sock_flag(sk, flag: SOCK_DEAD)) { |
6598 | sk->sk_state_change(sk); |
6599 | sk_wake_async(sk, how: SOCK_WAKE_IO, POLL_OUT); |
6600 | } |
6601 | if (fastopen_fail) |
6602 | return -1; |
6603 | if (sk->sk_write_pending || |
6604 | READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) || |
6605 | inet_csk_in_pingpong_mode(sk)) { |
6606 | /* Save one ACK. Data will be ready after |
6607 | * several ticks, if write_pending is set. |
6608 | * |
6609 | * It may be deleted, but with this feature tcpdumps |
6610 | * look so _wonderfully_ clever, that I was not able |
6611 | * to stand against the temptation 8) --ANK |
6612 | */ |
6613 | inet_csk_schedule_ack(sk); |
6614 | tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); |
6615 | tcp_reset_xmit_timer(sk, ICSK_TIME_DACK, |
6616 | TCP_DELACK_MAX, pace_delay: false); |
6617 | goto consume; |
6618 | } |
6619 | tcp_send_ack(sk); |
6620 | return -1; |
6621 | } |
6622 | |
6623 | /* No ACK in the segment */ |
6624 | |
6625 | if (th->rst) { |
6626 | /* rfc793: |
6627 | * "If the RST bit is set |
6628 | * |
6629 | * Otherwise (no ACK) drop the segment and return." |
6630 | */ |
6631 | SKB_DR_SET(reason, TCP_RESET); |
6632 | goto discard_and_undo; |
6633 | } |
6634 | |
6635 | /* PAWS check. */ |
6636 | if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && |
6637 | tcp_paws_reject(rx_opt: &tp->rx_opt, rst: 0)) { |
6638 | SKB_DR_SET(reason, TCP_RFC7323_PAWS); |
6639 | goto discard_and_undo; |
6640 | } |
6641 | if (th->syn) { |
6642 | /* We see SYN without ACK. It is attempt of |
6643 | * simultaneous connect with crossed SYNs. |
6644 | * Particularly, it can be connect to self. |
6645 | */ |
6646 | #ifdef CONFIG_TCP_AO |
6647 | struct tcp_ao_info *ao; |
6648 | |
6649 | ao = rcu_dereference_protected(tp->ao_info, |
6650 | lockdep_sock_is_held(sk)); |
6651 | if (ao) { |
6652 | WRITE_ONCE(ao->risn, th->seq); |
6653 | ao->rcv_sne = 0; |
6654 | } |
6655 | #endif |
6656 | tcp_set_state(sk, state: TCP_SYN_RECV); |
6657 | |
6658 | if (tp->rx_opt.saw_tstamp) { |
6659 | tp->rx_opt.tstamp_ok = 1; |
6660 | tcp_store_ts_recent(tp); |
6661 | tp->tcp_header_len = |
6662 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; |
6663 | } else { |
6664 | tp->tcp_header_len = sizeof(struct tcphdr); |
6665 | } |
6666 | |
6667 | WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); |
6668 | WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); |
6669 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; |
6670 | |
6671 | /* RFC1323: The window in SYN & SYN/ACK segments is |
6672 | * never scaled. |
6673 | */ |
6674 | tp->snd_wnd = ntohs(th->window); |
6675 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; |
6676 | tp->max_window = tp->snd_wnd; |
6677 | |
6678 | tcp_ecn_rcv_syn(tp, th); |
6679 | |
6680 | tcp_mtup_init(sk); |
6681 | tcp_sync_mss(sk, pmtu: icsk->icsk_pmtu_cookie); |
6682 | tcp_initialize_rcv_mss(sk); |
6683 | |
6684 | tcp_send_synack(sk); |
6685 | #if 0 |
6686 | /* Note, we could accept data and URG from this segment. |
6687 | * There are no obstacles to make this (except that we must |
6688 | * either change tcp_recvmsg() to prevent it from returning data |
6689 | * before 3WHS completes per RFC793, or employ TCP Fast Open). |
6690 | * |
6691 | * However, if we ignore data in ACKless segments sometimes, |
6692 | * we have no reasons to accept it sometimes. |
6693 | * Also, seems the code doing it in step6 of tcp_rcv_state_process |
6694 | * is not flawless. So, discard packet for sanity. |
6695 | * Uncomment this return to process the data. |
6696 | */ |
6697 | return -1; |
6698 | #else |
6699 | goto consume; |
6700 | #endif |
6701 | } |
6702 | /* "fifth, if neither of the SYN or RST bits is set then |
6703 | * drop the segment and return." |
6704 | */ |
6705 | |
6706 | discard_and_undo: |
6707 | tcp_clear_options(rx_opt: &tp->rx_opt); |
6708 | tp->rx_opt.mss_clamp = saved_clamp; |
6709 | tcp_drop_reason(sk, skb, reason); |
6710 | return 0; |
6711 | |
6712 | reset_and_undo: |
6713 | tcp_clear_options(rx_opt: &tp->rx_opt); |
6714 | tp->rx_opt.mss_clamp = saved_clamp; |
6715 | /* we can reuse/return @reason to its caller to handle the exception */ |
6716 | return reason; |
6717 | } |
6718 | |
6719 | static void tcp_rcv_synrecv_state_fastopen(struct sock *sk) |
6720 | { |
6721 | struct tcp_sock *tp = tcp_sk(sk); |
6722 | struct request_sock *req; |
6723 | |
6724 | /* If we are still handling the SYNACK RTO, see if timestamp ECR allows |
6725 | * undo. If peer SACKs triggered fast recovery, we can't undo here. |
6726 | */ |
6727 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out) |
6728 | tcp_try_undo_recovery(sk); |
6729 | |
6730 | tcp_update_rto_time(tp); |
6731 | inet_csk(sk)->icsk_retransmits = 0; |
6732 | /* In tcp_fastopen_synack_timer() on the first SYNACK RTO we set |
6733 | * retrans_stamp but don't enter CA_Loss, so in case that happened we |
6734 | * need to zero retrans_stamp here to prevent spurious |
6735 | * retransmits_timed_out(). However, if the ACK of our SYNACK caused us |
6736 | * to enter CA_Recovery then we need to leave retrans_stamp as it was |
6737 | * set entering CA_Recovery, for correct retransmits_timed_out() and |
6738 | * undo behavior. |
6739 | */ |
6740 | tcp_retrans_stamp_cleanup(sk); |
6741 | |
6742 | /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1, |
6743 | * we no longer need req so release it. |
6744 | */ |
6745 | req = rcu_dereference_protected(tp->fastopen_rsk, |
6746 | lockdep_sock_is_held(sk)); |
6747 | reqsk_fastopen_remove(sk, req, reset: false); |
6748 | |
6749 | /* Re-arm the timer because data may have been sent out. |
6750 | * This is similar to the regular data transmission case |
6751 | * when new data has just been ack'ed. |
6752 | * |
6753 | * (TFO) - we could try to be more aggressive and |
6754 | * retransmitting any data sooner based on when they |
6755 | * are sent out. |
6756 | */ |
6757 | tcp_rearm_rto(sk); |
6758 | } |
6759 | |
6760 | /* |
6761 | * This function implements the receiving procedure of RFC 793 for |
6762 | * all states except ESTABLISHED and TIME_WAIT. |
6763 | * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be |
6764 | * address independent. |
6765 | */ |
6766 | |
6767 | enum skb_drop_reason |
6768 | tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb) |
6769 | { |
6770 | struct tcp_sock *tp = tcp_sk(sk); |
6771 | struct inet_connection_sock *icsk = inet_csk(sk); |
6772 | const struct tcphdr *th = tcp_hdr(skb); |
6773 | struct request_sock *req; |
6774 | int queued = 0; |
6775 | SKB_DR(reason); |
6776 | |
6777 | switch (sk->sk_state) { |
6778 | case TCP_CLOSE: |
6779 | SKB_DR_SET(reason, TCP_CLOSE); |
6780 | goto discard; |
6781 | |
6782 | case TCP_LISTEN: |
6783 | if (th->ack) |
6784 | return SKB_DROP_REASON_TCP_FLAGS; |
6785 | |
6786 | if (th->rst) { |
6787 | SKB_DR_SET(reason, TCP_RESET); |
6788 | goto discard; |
6789 | } |
6790 | if (th->syn) { |
6791 | if (th->fin) { |
6792 | SKB_DR_SET(reason, TCP_FLAGS); |
6793 | goto discard; |
6794 | } |
6795 | /* It is possible that we process SYN packets from backlog, |
6796 | * so we need to make sure to disable BH and RCU right there. |
6797 | */ |
6798 | rcu_read_lock(); |
6799 | local_bh_disable(); |
6800 | icsk->icsk_af_ops->conn_request(sk, skb); |
6801 | local_bh_enable(); |
6802 | rcu_read_unlock(); |
6803 | |
6804 | consume_skb(skb); |
6805 | return 0; |
6806 | } |
6807 | SKB_DR_SET(reason, TCP_FLAGS); |
6808 | goto discard; |
6809 | |
6810 | case TCP_SYN_SENT: |
6811 | tp->rx_opt.saw_tstamp = 0; |
6812 | tcp_mstamp_refresh(tp); |
6813 | queued = tcp_rcv_synsent_state_process(sk, skb, th); |
6814 | if (queued >= 0) |
6815 | return queued; |
6816 | |
6817 | /* Do step6 onward by hand. */ |
6818 | tcp_urg(sk, skb, th); |
6819 | __kfree_skb(skb); |
6820 | tcp_data_snd_check(sk); |
6821 | return 0; |
6822 | } |
6823 | |
6824 | tcp_mstamp_refresh(tp); |
6825 | tp->rx_opt.saw_tstamp = 0; |
6826 | req = rcu_dereference_protected(tp->fastopen_rsk, |
6827 | lockdep_sock_is_held(sk)); |
6828 | if (req) { |
6829 | bool req_stolen; |
6830 | |
6831 | WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && |
6832 | sk->sk_state != TCP_FIN_WAIT1); |
6833 | |
6834 | SKB_DR_SET(reason, TCP_FASTOPEN); |
6835 | if (!tcp_check_req(sk, skb, req, fastopen: true, lost_race: &req_stolen, drop_reason: &reason)) |
6836 | goto discard; |
6837 | } |
6838 | |
6839 | if (!th->ack && !th->rst && !th->syn) { |
6840 | SKB_DR_SET(reason, TCP_FLAGS); |
6841 | goto discard; |
6842 | } |
6843 | if (!tcp_validate_incoming(sk, skb, th, syn_inerr: 0)) |
6844 | return 0; |
6845 | |
6846 | /* step 5: check the ACK field */ |
6847 | reason = tcp_ack(sk, skb, FLAG_SLOWPATH | |
6848 | FLAG_UPDATE_TS_RECENT | |
6849 | FLAG_NO_CHALLENGE_ACK); |
6850 | |
6851 | if ((int)reason <= 0) { |
6852 | if (sk->sk_state == TCP_SYN_RECV) { |
6853 | /* send one RST */ |
6854 | if (!reason) |
6855 | return SKB_DROP_REASON_TCP_OLD_ACK; |
6856 | return -reason; |
6857 | } |
6858 | /* accept old ack during closing */ |
6859 | if ((int)reason < 0) { |
6860 | tcp_send_challenge_ack(sk); |
6861 | reason = -reason; |
6862 | goto discard; |
6863 | } |
6864 | } |
6865 | SKB_DR_SET(reason, NOT_SPECIFIED); |
6866 | switch (sk->sk_state) { |
6867 | case TCP_SYN_RECV: |
6868 | tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */ |
6869 | if (!tp->srtt_us) |
6870 | tcp_synack_rtt_meas(sk, req); |
6871 | |
6872 | if (tp->rx_opt.tstamp_ok) |
6873 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; |
6874 | |
6875 | if (req) { |
6876 | tcp_rcv_synrecv_state_fastopen(sk); |
6877 | } else { |
6878 | tcp_try_undo_spurious_syn(sk); |
6879 | tp->retrans_stamp = 0; |
6880 | tcp_init_transfer(sk, bpf_op: BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, |
6881 | skb); |
6882 | WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); |
6883 | } |
6884 | tcp_ao_established(sk); |
6885 | smp_mb(); |
6886 | tcp_set_state(sk, state: TCP_ESTABLISHED); |
6887 | sk->sk_state_change(sk); |
6888 | |
6889 | /* Note, that this wakeup is only for marginal crossed SYN case. |
6890 | * Passively open sockets are not waked up, because |
6891 | * sk->sk_sleep == NULL and sk->sk_socket == NULL. |
6892 | */ |
6893 | if (sk->sk_socket) |
6894 | sk_wake_async(sk, how: SOCK_WAKE_IO, POLL_OUT); |
6895 | |
6896 | tp->snd_una = TCP_SKB_CB(skb)->ack_seq; |
6897 | tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale; |
6898 | tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); |
6899 | |
6900 | if (!inet_csk(sk)->icsk_ca_ops->cong_control) |
6901 | tcp_update_pacing_rate(sk); |
6902 | |
6903 | /* Prevent spurious tcp_cwnd_restart() on first data packet */ |
6904 | tp->lsndtime = tcp_jiffies32; |
6905 | |
6906 | tcp_initialize_rcv_mss(sk); |
6907 | tcp_fast_path_on(tp); |
6908 | if (sk->sk_shutdown & SEND_SHUTDOWN) |
6909 | tcp_shutdown(sk, SEND_SHUTDOWN); |
6910 | break; |
6911 | |
6912 | case TCP_FIN_WAIT1: { |
6913 | int tmo; |
6914 | |
6915 | if (req) |
6916 | tcp_rcv_synrecv_state_fastopen(sk); |
6917 | |
6918 | if (tp->snd_una != tp->write_seq) |
6919 | break; |
6920 | |
6921 | tcp_set_state(sk, state: TCP_FIN_WAIT2); |
6922 | WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN); |
6923 | |
6924 | sk_dst_confirm(sk); |
6925 | |
6926 | if (!sock_flag(sk, flag: SOCK_DEAD)) { |
6927 | /* Wake up lingering close() */ |
6928 | sk->sk_state_change(sk); |
6929 | break; |
6930 | } |
6931 | |
6932 | if (READ_ONCE(tp->linger2) < 0) { |
6933 | tcp_done(sk); |
6934 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); |
6935 | return SKB_DROP_REASON_TCP_ABORT_ON_DATA; |
6936 | } |
6937 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && |
6938 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { |
6939 | /* Receive out of order FIN after close() */ |
6940 | if (tp->syn_fastopen && th->fin) |
6941 | tcp_fastopen_active_disable(sk); |
6942 | tcp_done(sk); |
6943 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); |
6944 | return SKB_DROP_REASON_TCP_ABORT_ON_DATA; |
6945 | } |
6946 | |
6947 | tmo = tcp_fin_time(sk); |
6948 | if (tmo > TCP_TIMEWAIT_LEN) { |
6949 | tcp_reset_keepalive_timer(sk, timeout: tmo - TCP_TIMEWAIT_LEN); |
6950 | } else if (th->fin || sock_owned_by_user(sk)) { |
6951 | /* Bad case. We could lose such FIN otherwise. |
6952 | * It is not a big problem, but it looks confusing |
6953 | * and not so rare event. We still can lose it now, |
6954 | * if it spins in bh_lock_sock(), but it is really |
6955 | * marginal case. |
6956 | */ |
6957 | tcp_reset_keepalive_timer(sk, timeout: tmo); |
6958 | } else { |
6959 | tcp_time_wait(sk, state: TCP_FIN_WAIT2, timeo: tmo); |
6960 | goto consume; |
6961 | } |
6962 | break; |
6963 | } |
6964 | |
6965 | case TCP_CLOSING: |
6966 | if (tp->snd_una == tp->write_seq) { |
6967 | tcp_time_wait(sk, state: TCP_TIME_WAIT, timeo: 0); |
6968 | goto consume; |
6969 | } |
6970 | break; |
6971 | |
6972 | case TCP_LAST_ACK: |
6973 | if (tp->snd_una == tp->write_seq) { |
6974 | tcp_update_metrics(sk); |
6975 | tcp_done(sk); |
6976 | goto consume; |
6977 | } |
6978 | break; |
6979 | } |
6980 | |
6981 | /* step 6: check the URG bit */ |
6982 | tcp_urg(sk, skb, th); |
6983 | |
6984 | /* step 7: process the segment text */ |
6985 | switch (sk->sk_state) { |
6986 | case TCP_CLOSE_WAIT: |
6987 | case TCP_CLOSING: |
6988 | case TCP_LAST_ACK: |
6989 | if (!before(TCP_SKB_CB(skb)->seq, seq2: tp->rcv_nxt)) { |
6990 | /* If a subflow has been reset, the packet should not |
6991 | * continue to be processed, drop the packet. |
6992 | */ |
6993 | if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) |
6994 | goto discard; |
6995 | break; |
6996 | } |
6997 | fallthrough; |
6998 | case TCP_FIN_WAIT1: |
6999 | case TCP_FIN_WAIT2: |
7000 | /* RFC 793 says to queue data in these states, |
7001 | * RFC 1122 says we MUST send a reset. |
7002 | * BSD 4.4 also does reset. |
7003 | */ |
7004 | if (sk->sk_shutdown & RCV_SHUTDOWN) { |
7005 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && |
7006 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { |
7007 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); |
7008 | tcp_reset(sk, skb); |
7009 | return SKB_DROP_REASON_TCP_ABORT_ON_DATA; |
7010 | } |
7011 | } |
7012 | fallthrough; |
7013 | case TCP_ESTABLISHED: |
7014 | tcp_data_queue(sk, skb); |
7015 | queued = 1; |
7016 | break; |
7017 | } |
7018 | |
7019 | /* tcp_data could move socket to TIME-WAIT */ |
7020 | if (sk->sk_state != TCP_CLOSE) { |
7021 | tcp_data_snd_check(sk); |
7022 | tcp_ack_snd_check(sk); |
7023 | } |
7024 | |
7025 | if (!queued) { |
7026 | discard: |
7027 | tcp_drop_reason(sk, skb, reason); |
7028 | } |
7029 | return 0; |
7030 | |
7031 | consume: |
7032 | __kfree_skb(skb); |
7033 | return 0; |
7034 | } |
7035 | EXPORT_IPV6_MOD(tcp_rcv_state_process); |
7036 | |
7037 | static inline void pr_drop_req(struct request_sock *req, __u16 port, int family) |
7038 | { |
7039 | struct inet_request_sock *ireq = inet_rsk(sk: req); |
7040 | |
7041 | if (family == AF_INET) |
7042 | net_dbg_ratelimited("drop open request from %pI4/%u\n", |
7043 | &ireq->ir_rmt_addr, port); |
7044 | #if IS_ENABLED(CONFIG_IPV6) |
7045 | else if (family == AF_INET6) |
7046 | net_dbg_ratelimited("drop open request from %pI6/%u\n", |
7047 | &ireq->ir_v6_rmt_addr, port); |
7048 | #endif |
7049 | } |
7050 | |
7051 | /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set |
7052 | * |
7053 | * If we receive a SYN packet with these bits set, it means a |
7054 | * network is playing bad games with TOS bits. In order to |
7055 | * avoid possible false congestion notifications, we disable |
7056 | * TCP ECN negotiation. |
7057 | * |
7058 | * Exception: tcp_ca wants ECN. This is required for DCTCP |
7059 | * congestion control: Linux DCTCP asserts ECT on all packets, |
7060 | * including SYN, which is most optimal solution; however, |
7061 | * others, such as FreeBSD do not. |
7062 | * |
7063 | * Exception: At least one of the reserved bits of the TCP header (th->res1) is |
7064 | * set, indicating the use of a future TCP extension (such as AccECN). See |
7065 | * RFC8311 §4.3 which updates RFC3168 to allow the development of such |
7066 | * extensions. |
7067 | */ |
7068 | static void tcp_ecn_create_request(struct request_sock *req, |
7069 | const struct sk_buff *skb, |
7070 | const struct sock *listen_sk, |
7071 | const struct dst_entry *dst) |
7072 | { |
7073 | const struct tcphdr *th = tcp_hdr(skb); |
7074 | const struct net *net = sock_net(sk: listen_sk); |
7075 | bool th_ecn = th->ece && th->cwr; |
7076 | bool ect, ecn_ok; |
7077 | u32 ecn_ok_dst; |
7078 | |
7079 | if (!th_ecn) |
7080 | return; |
7081 | |
7082 | ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield); |
7083 | ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK); |
7084 | ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst; |
7085 | |
7086 | if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(sk: listen_sk) || |
7087 | (ecn_ok_dst & DST_FEATURE_ECN_CA) || |
7088 | tcp_bpf_ca_needs_ecn(sk: (struct sock *)req)) |
7089 | inet_rsk(sk: req)->ecn_ok = 1; |
7090 | } |
7091 | |
7092 | static void tcp_openreq_init(struct request_sock *req, |
7093 | const struct tcp_options_received *rx_opt, |
7094 | struct sk_buff *skb, const struct sock *sk) |
7095 | { |
7096 | struct inet_request_sock *ireq = inet_rsk(sk: req); |
7097 | |
7098 | req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */ |
7099 | tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq; |
7100 | tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; |
7101 | tcp_rsk(req)->snt_synack = 0; |
7102 | tcp_rsk(req)->snt_tsval_first = 0; |
7103 | tcp_rsk(req)->last_oow_ack_time = 0; |
7104 | req->mss = rx_opt->mss_clamp; |
7105 | req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0; |
7106 | ireq->tstamp_ok = rx_opt->tstamp_ok; |
7107 | ireq->sack_ok = rx_opt->sack_ok; |
7108 | ireq->snd_wscale = rx_opt->snd_wscale; |
7109 | ireq->wscale_ok = rx_opt->wscale_ok; |
7110 | ireq->acked = 0; |
7111 | ireq->ecn_ok = 0; |
7112 | ireq->ir_rmt_port = tcp_hdr(skb)->source; |
7113 | ireq->ir_num = ntohs(tcp_hdr(skb)->dest); |
7114 | ireq->ir_mark = inet_request_mark(sk, skb); |
7115 | #if IS_ENABLED(CONFIG_SMC) |
7116 | ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested && |
7117 | tcp_sk(sk)->smc_hs_congested(sk)); |
7118 | #endif |
7119 | } |
7120 | |
7121 | /* |
7122 | * Return true if a syncookie should be sent |
7123 | */ |
7124 | static bool tcp_syn_flood_action(struct sock *sk, const char *proto) |
7125 | { |
7126 | struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; |
7127 | const char *msg = "Dropping request"; |
7128 | struct net *net = sock_net(sk); |
7129 | bool want_cookie = false; |
7130 | u8 syncookies; |
7131 | |
7132 | syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies); |
7133 | |
7134 | #ifdef CONFIG_SYN_COOKIES |
7135 | if (syncookies) { |
7136 | msg = "Sending cookies"; |
7137 | want_cookie = true; |
7138 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES); |
7139 | } else |
7140 | #endif |
7141 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP); |
7142 | |
7143 | if (!READ_ONCE(queue->synflood_warned) && syncookies != 2 && |
7144 | xchg(&queue->synflood_warned, 1) == 0) { |
7145 | if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) { |
7146 | net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n", |
7147 | proto, inet6_rcv_saddr(sk), |
7148 | sk->sk_num, msg); |
7149 | } else { |
7150 | net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n", |
7151 | proto, &sk->sk_rcv_saddr, |
7152 | sk->sk_num, msg); |
7153 | } |
7154 | } |
7155 | |
7156 | return want_cookie; |
7157 | } |
7158 | |
7159 | static void tcp_reqsk_record_syn(const struct sock *sk, |
7160 | struct request_sock *req, |
7161 | const struct sk_buff *skb) |
7162 | { |
7163 | if (tcp_sk(sk)->save_syn) { |
7164 | u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb); |
7165 | struct saved_syn *saved_syn; |
7166 | u32 mac_hdrlen; |
7167 | void *base; |
7168 | |
7169 | if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */ |
7170 | base = skb_mac_header(skb); |
7171 | mac_hdrlen = skb_mac_header_len(skb); |
7172 | len += mac_hdrlen; |
7173 | } else { |
7174 | base = skb_network_header(skb); |
7175 | mac_hdrlen = 0; |
7176 | } |
7177 | |
7178 | saved_syn = kmalloc(struct_size(saved_syn, data, len), |
7179 | GFP_ATOMIC); |
7180 | if (saved_syn) { |
7181 | saved_syn->mac_hdrlen = mac_hdrlen; |
7182 | saved_syn->network_hdrlen = skb_network_header_len(skb); |
7183 | saved_syn->tcp_hdrlen = tcp_hdrlen(skb); |
7184 | memcpy(saved_syn->data, base, len); |
7185 | req->saved_syn = saved_syn; |
7186 | } |
7187 | } |
7188 | } |
7189 | |
7190 | /* If a SYN cookie is required and supported, returns a clamped MSS value to be |
7191 | * used for SYN cookie generation. |
7192 | */ |
7193 | u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops, |
7194 | const struct tcp_request_sock_ops *af_ops, |
7195 | struct sock *sk, struct tcphdr *th) |
7196 | { |
7197 | struct tcp_sock *tp = tcp_sk(sk); |
7198 | u16 mss; |
7199 | |
7200 | if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 && |
7201 | !inet_csk_reqsk_queue_is_full(sk)) |
7202 | return 0; |
7203 | |
7204 | if (!tcp_syn_flood_action(sk, proto: rsk_ops->slab_name)) |
7205 | return 0; |
7206 | |
7207 | if (sk_acceptq_is_full(sk)) { |
7208 | NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); |
7209 | return 0; |
7210 | } |
7211 | |
7212 | mss = tcp_parse_mss_option(th, user_mss: tp->rx_opt.user_mss); |
7213 | if (!mss) |
7214 | mss = af_ops->mss_clamp; |
7215 | |
7216 | return mss; |
7217 | } |
7218 | EXPORT_IPV6_MOD_GPL(tcp_get_syncookie_mss); |
7219 | |
7220 | int tcp_conn_request(struct request_sock_ops *rsk_ops, |
7221 | const struct tcp_request_sock_ops *af_ops, |
7222 | struct sock *sk, struct sk_buff *skb) |
7223 | { |
7224 | struct tcp_fastopen_cookie foc = { .len = -1 }; |
7225 | struct tcp_options_received tmp_opt; |
7226 | struct tcp_sock *tp = tcp_sk(sk); |
7227 | struct net *net = sock_net(sk); |
7228 | struct sock *fastopen_sk = NULL; |
7229 | struct request_sock *req; |
7230 | bool want_cookie = false; |
7231 | struct dst_entry *dst; |
7232 | struct flowi fl; |
7233 | u8 syncookies; |
7234 | u32 isn; |
7235 | |
7236 | #ifdef CONFIG_TCP_AO |
7237 | const struct tcp_ao_hdr *aoh; |
7238 | #endif |
7239 | |
7240 | isn = __this_cpu_read(tcp_tw_isn); |
7241 | if (isn) { |
7242 | /* TW buckets are converted to open requests without |
7243 | * limitations, they conserve resources and peer is |
7244 | * evidently real one. |
7245 | */ |
7246 | __this_cpu_write(tcp_tw_isn, 0); |
7247 | } else { |
7248 | syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies); |
7249 | |
7250 | if (syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) { |
7251 | want_cookie = tcp_syn_flood_action(sk, |
7252 | proto: rsk_ops->slab_name); |
7253 | if (!want_cookie) |
7254 | goto drop; |
7255 | } |
7256 | } |
7257 | |
7258 | if (sk_acceptq_is_full(sk)) { |
7259 | NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); |
7260 | goto drop; |
7261 | } |
7262 | |
7263 | req = inet_reqsk_alloc(ops: rsk_ops, sk_listener: sk, attach_listener: !want_cookie); |
7264 | if (!req) |
7265 | goto drop; |
7266 | |
7267 | req->syncookie = want_cookie; |
7268 | tcp_rsk(req)->af_specific = af_ops; |
7269 | tcp_rsk(req)->ts_off = 0; |
7270 | tcp_rsk(req)->req_usec_ts = false; |
7271 | #if IS_ENABLED(CONFIG_MPTCP) |
7272 | tcp_rsk(req)->is_mptcp = 0; |
7273 | #endif |
7274 | |
7275 | tcp_clear_options(rx_opt: &tmp_opt); |
7276 | tmp_opt.mss_clamp = af_ops->mss_clamp; |
7277 | tmp_opt.user_mss = tp->rx_opt.user_mss; |
7278 | tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, |
7279 | want_cookie ? NULL : &foc); |
7280 | |
7281 | if (want_cookie && !tmp_opt.saw_tstamp) |
7282 | tcp_clear_options(rx_opt: &tmp_opt); |
7283 | |
7284 | if (IS_ENABLED(CONFIG_SMC) && want_cookie) |
7285 | tmp_opt.smc_ok = 0; |
7286 | |
7287 | tmp_opt.tstamp_ok = tmp_opt.saw_tstamp; |
7288 | tcp_openreq_init(req, rx_opt: &tmp_opt, skb, sk); |
7289 | inet_rsk(sk: req)->no_srccheck = inet_test_bit(TRANSPARENT, sk); |
7290 | |
7291 | /* Note: tcp_v6_init_req() might override ir_iif for link locals */ |
7292 | inet_rsk(sk: req)->ir_iif = inet_request_bound_dev_if(sk, skb); |
7293 | |
7294 | dst = af_ops->route_req(sk, skb, &fl, req, isn); |
7295 | if (!dst) |
7296 | goto drop_and_free; |
7297 | |
7298 | if (tmp_opt.tstamp_ok) { |
7299 | tcp_rsk(req)->req_usec_ts = dst_tcp_usec_ts(dst); |
7300 | tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb); |
7301 | } |
7302 | if (!want_cookie && !isn) { |
7303 | int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog); |
7304 | |
7305 | /* Kill the following clause, if you dislike this way. */ |
7306 | if (!syncookies && |
7307 | (max_syn_backlog - inet_csk_reqsk_queue_len(sk) < |
7308 | (max_syn_backlog >> 2)) && |
7309 | !tcp_peer_is_proven(req, dst)) { |
7310 | /* Without syncookies last quarter of |
7311 | * backlog is filled with destinations, |
7312 | * proven to be alive. |
7313 | * It means that we continue to communicate |
7314 | * to destinations, already remembered |
7315 | * to the moment of synflood. |
7316 | */ |
7317 | pr_drop_req(req, ntohs(tcp_hdr(skb)->source), |
7318 | family: rsk_ops->family); |
7319 | goto drop_and_release; |
7320 | } |
7321 | |
7322 | isn = af_ops->init_seq(skb); |
7323 | } |
7324 | |
7325 | tcp_ecn_create_request(req, skb, listen_sk: sk, dst); |
7326 | |
7327 | if (want_cookie) { |
7328 | isn = cookie_init_sequence(ops: af_ops, sk, skb, mss: &req->mss); |
7329 | if (!tmp_opt.tstamp_ok) |
7330 | inet_rsk(sk: req)->ecn_ok = 0; |
7331 | } |
7332 | |
7333 | #ifdef CONFIG_TCP_AO |
7334 | if (tcp_parse_auth_options(th: tcp_hdr(skb), NULL, aoh: &aoh)) |
7335 | goto drop_and_release; /* Invalid TCP options */ |
7336 | if (aoh) { |
7337 | tcp_rsk(req)->used_tcp_ao = true; |
7338 | tcp_rsk(req)->ao_rcv_next = aoh->keyid; |
7339 | tcp_rsk(req)->ao_keyid = aoh->rnext_keyid; |
7340 | |
7341 | } else { |
7342 | tcp_rsk(req)->used_tcp_ao = false; |
7343 | } |
7344 | #endif |
7345 | tcp_rsk(req)->snt_isn = isn; |
7346 | tcp_rsk(req)->txhash = net_tx_rndhash(); |
7347 | tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield; |
7348 | tcp_openreq_init_rwin(req, sk_listener: sk, dst); |
7349 | sk_rx_queue_set(sk: req_to_sk(req), skb); |
7350 | if (!want_cookie) { |
7351 | tcp_reqsk_record_syn(sk, req, skb); |
7352 | fastopen_sk = tcp_try_fastopen(sk, skb, req, foc: &foc, dst); |
7353 | } |
7354 | if (fastopen_sk) { |
7355 | af_ops->send_synack(fastopen_sk, dst, &fl, req, |
7356 | &foc, TCP_SYNACK_FASTOPEN, skb); |
7357 | /* Add the child socket directly into the accept queue */ |
7358 | if (!inet_csk_reqsk_queue_add(sk, req, child: fastopen_sk)) { |
7359 | reqsk_fastopen_remove(sk: fastopen_sk, req, reset: false); |
7360 | bh_unlock_sock(fastopen_sk); |
7361 | sock_put(sk: fastopen_sk); |
7362 | goto drop_and_free; |
7363 | } |
7364 | sk->sk_data_ready(sk); |
7365 | bh_unlock_sock(fastopen_sk); |
7366 | sock_put(sk: fastopen_sk); |
7367 | } else { |
7368 | tcp_rsk(req)->tfo_listener = false; |
7369 | if (!want_cookie) { |
7370 | req->timeout = tcp_timeout_init(sk: (struct sock *)req); |
7371 | if (unlikely(!inet_csk_reqsk_queue_hash_add(sk, req, |
7372 | req->timeout))) { |
7373 | reqsk_free(req); |
7374 | dst_release(dst); |
7375 | return 0; |
7376 | } |
7377 | |
7378 | } |
7379 | af_ops->send_synack(sk, dst, &fl, req, &foc, |
7380 | !want_cookie ? TCP_SYNACK_NORMAL : |
7381 | TCP_SYNACK_COOKIE, |
7382 | skb); |
7383 | if (want_cookie) { |
7384 | reqsk_free(req); |
7385 | return 0; |
7386 | } |
7387 | } |
7388 | reqsk_put(req); |
7389 | return 0; |
7390 | |
7391 | drop_and_release: |
7392 | dst_release(dst); |
7393 | drop_and_free: |
7394 | __reqsk_free(req); |
7395 | drop: |
7396 | tcp_listendrop(sk); |
7397 | return 0; |
7398 | } |
7399 | EXPORT_IPV6_MOD(tcp_conn_request); |
7400 |
Definitions
- sysctl_tcp_max_orphans
- clean_acked_data_enabled
- clean_acked_data_enable
- clean_acked_data_disable
- clean_acked_data_flush
- bpf_skops_parse_hdr
- bpf_skops_established
- tcp_gro_dev_warn
- tcp_measure_rcv_mss
- tcp_incr_quickack
- tcp_enter_quickack_mode
- tcp_in_quickack_mode
- tcp_ecn_queue_cwr
- tcp_ecn_accept_cwr
- tcp_ecn_withdraw_cwr
- tcp_data_ecn_check
- tcp_ecn_rcv_synack
- tcp_ecn_rcv_syn
- tcp_ecn_rcv_ecn_echo
- tcp_count_delivered_ce
- tcp_count_delivered
- tcp_sndbuf_expand
- __tcp_grow_window
- truesize_adjust
- tcp_grow_window
- tcp_init_buffer_space
- tcp_clamp_window
- tcp_initialize_rcv_mss
- tcp_rcv_rtt_update
- tcp_rcv_rtt_measure
- tcp_rtt_tsopt_us
- tcp_rcv_rtt_measure_ts
- tcp_rcvbuf_grow
- tcp_rcv_space_adjust
- tcp_save_lrcv_flowlabel
- tcp_event_data_recv
- tcp_rtt_estimator
- tcp_update_pacing_rate
- tcp_set_rto
- tcp_init_cwnd
- tcp_sacktag_state
- tcp_dsack_seen
- tcp_check_sack_reordering
- tcp_verify_retransmit_hint
- tcp_notify_skb_loss_event
- tcp_mark_skb_lost
- tcp_is_sackblock_valid
- tcp_check_dsack
- tcp_match_skb_to_sack
- tcp_sacktag_one
- tcp_shifted_skb
- tcp_skb_seglen
- skb_can_shift
- tcp_skb_shift
- tcp_shift_skb_data
- tcp_sacktag_walk
- tcp_sacktag_bsearch
- tcp_sacktag_skip
- tcp_maybe_skipping_dsack
- tcp_sack_cache_ok
- tcp_sacktag_write_queue
- tcp_limit_reno_sacked
- tcp_check_reno_reordering
- tcp_add_reno_sack
- tcp_remove_reno_sacks
- tcp_reset_reno_sack
- tcp_clear_retrans
- tcp_init_undo
- tcp_is_rack
- tcp_timeout_mark_lost
- tcp_enter_loss
- tcp_check_sack_reneging
- tcp_dupack_heuristics
- tcp_time_to_recover
- tcp_mark_head_lost
- tcp_update_scoreboard
- tcp_tsopt_ecr_before
- tcp_skb_spurious_retrans
- tcp_packet_delayed
- tcp_any_retrans_done
- tcp_retrans_stamp_cleanup
- DBGUNDO
- tcp_undo_cwnd_reduction
- tcp_may_undo
- tcp_is_non_sack_preventing_reopen
- tcp_try_undo_recovery
- tcp_try_undo_dsack
- tcp_try_undo_loss
- tcp_init_cwnd_reduction
- tcp_cwnd_reduction
- tcp_end_cwnd_reduction
- tcp_enter_cwr
- tcp_try_keep_open
- tcp_try_to_open
- tcp_mtup_probe_failed
- tcp_mtup_probe_success
- tcp_non_congestion_loss_retransmit
- tcp_simple_retransmit
- tcp_enter_recovery
- tcp_update_rto_time
- tcp_process_loss
- tcp_force_fast_retransmit
- tcp_try_undo_partial
- tcp_identify_packet_loss
- tcp_fastretrans_alert
- tcp_update_rtt_min
- tcp_ack_update_rtt
- tcp_synack_rtt_meas
- tcp_cong_avoid
- tcp_rearm_rto
- tcp_set_xmit_timer
- tcp_tso_acked
- tcp_ack_tstamp
- tcp_clean_rtx_queue
- tcp_ack_probe
- tcp_ack_is_dubious
- tcp_may_raise_cwnd
- tcp_cong_control
- tcp_may_update_window
- tcp_snd_sne_update
- tcp_snd_una_update
- tcp_rcv_sne_update
- tcp_rcv_nxt_update
- tcp_ack_update_window
- __tcp_oow_rate_limited
- tcp_oow_rate_limited
- tcp_send_challenge_ack
- tcp_store_ts_recent
- __tcp_replace_ts_recent
- tcp_replace_ts_recent
- tcp_process_tlp_ack
- tcp_in_ack_event
- tcp_xmit_recovery
- tcp_newly_delivered
- tcp_ack
- tcp_parse_fastopen_option
- smc_parse_options
- tcp_parse_mss_option
- tcp_parse_options
- tcp_parse_aligned_timestamp
- tcp_fast_parse_options
- tcp_do_parse_auth_options
- tcp_tsval_replay
- tcp_disordered_ack_check
- tcp_sequence
- tcp_done_with_error
- tcp_reset
- tcp_fin
- tcp_sack_extend
- tcp_dsack_set
- tcp_dsack_extend
- tcp_rcv_spurious_retrans
- tcp_send_dupack
- tcp_sack_maybe_coalesce
- tcp_sack_compress_send_ack
- tcp_sack_new_ofo_skb
- tcp_sack_remove
- tcp_try_coalesce
- tcp_ooo_try_coalesce
- tcp_drop_reason
- tcp_ofo_queue
- tcp_try_rmem_schedule
- tcp_data_queue_ofo
- tcp_queue_rcv
- tcp_send_rcvq
- tcp_data_ready
- tcp_data_queue
- tcp_skb_next
- tcp_collapse_one
- tcp_rbtree_insert
- tcp_collapse
- tcp_collapse_ofo_queue
- tcp_prune_ofo_queue
- tcp_prune_queue
- tcp_should_expand_sndbuf
- tcp_new_space
- tcp_check_space
- tcp_data_snd_check
- __tcp_ack_snd_check
- tcp_ack_snd_check
- tcp_check_urg
- tcp_urg
- tcp_reset_check
- tcp_validate_incoming
- tcp_rcv_established
- tcp_init_transfer
- tcp_finish_connect
- tcp_rcv_fastopen_synack
- smc_check_reset_syn
- tcp_try_undo_spurious_syn
- tcp_rcv_synsent_state_process
- tcp_rcv_synrecv_state_fastopen
- tcp_rcv_state_process
- pr_drop_req
- tcp_ecn_create_request
- tcp_openreq_init
- tcp_syn_flood_action
- tcp_reqsk_record_syn
- tcp_get_syncookie_mss
Improve your Profiling and Debugging skills
Find out more