]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
1da177e4 LT |
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 | * | |
02c30a84 | 9 | * Authors: Ross Biro |
1da177e4 LT |
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. | |
caa20d9a | 44 | * Andrey Savochkin: Fix RTT measurements in the presence of |
1da177e4 LT |
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. | |
e905a9ed | 52 | * Andi Kleen: Add tcp_measure_rcv_mss to make |
1da177e4 | 53 | * connections with MSS<min(MTU,ann. MSS) |
e905a9ed | 54 | * work without delayed acks. |
1da177e4 LT |
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 | |
1da177e4 LT |
63 | */ |
64 | ||
afd46503 JP |
65 | #define pr_fmt(fmt) "TCP: " fmt |
66 | ||
1da177e4 | 67 | #include <linux/mm.h> |
5a0e3ad6 | 68 | #include <linux/slab.h> |
1da177e4 LT |
69 | #include <linux/module.h> |
70 | #include <linux/sysctl.h> | |
a0bffffc | 71 | #include <linux/kernel.h> |
ad971f61 | 72 | #include <linux/prefetch.h> |
5ffc02a1 | 73 | #include <net/dst.h> |
1da177e4 LT |
74 | #include <net/tcp.h> |
75 | #include <net/inet_common.h> | |
76 | #include <linux/ipsec.h> | |
77 | #include <asm/unaligned.h> | |
e1c8a607 | 78 | #include <linux/errqueue.h> |
5941521c | 79 | #include <trace/events/tcp.h> |
60e2a778 | 80 | #include <linux/static_key.h> |
1da177e4 | 81 | |
ab32ea5d | 82 | int sysctl_tcp_max_orphans __read_mostly = NR_FILE; |
1da177e4 | 83 | |
1da177e4 LT |
84 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ |
85 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ | |
86 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ | |
87 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ | |
88 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ | |
89 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ | |
90 | #define FLAG_ECE 0x40 /* ECE in this ACK */ | |
291a00d1 | 91 | #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */ |
31770e34 | 92 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ |
e33099f9 | 93 | #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */ |
2e605294 | 94 | #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ |
564262c1 | 95 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ |
df92c839 | 96 | #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */ |
cadbd031 | 97 | #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ |
12fb3dd9 | 98 | #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */ |
d0e1a1b5 | 99 | #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */ |
1da177e4 LT |
100 | |
101 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) | |
102 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) | |
d09b9e60 | 103 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK) |
1da177e4 LT |
104 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) |
105 | ||
1da177e4 | 106 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) |
bdf1ee5d | 107 | #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) |
1da177e4 | 108 | |
e662ca40 YC |
109 | #define REXMIT_NONE 0 /* no loss recovery to do */ |
110 | #define REXMIT_LOST 1 /* retransmit packets marked lost */ | |
111 | #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */ | |
112 | ||
0b9aefea MRL |
113 | static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb, |
114 | unsigned int len) | |
dcb17d22 MRL |
115 | { |
116 | static bool __once __read_mostly; | |
117 | ||
118 | if (!__once) { | |
119 | struct net_device *dev; | |
120 | ||
121 | __once = true; | |
122 | ||
123 | rcu_read_lock(); | |
124 | dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif); | |
0b9aefea MRL |
125 | if (!dev || len >= dev->mtu) |
126 | pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n", | |
127 | dev ? dev->name : "Unknown driver"); | |
dcb17d22 MRL |
128 | rcu_read_unlock(); |
129 | } | |
130 | } | |
131 | ||
e905a9ed | 132 | /* Adapt the MSS value used to make delayed ack decision to the |
1da177e4 | 133 | * real world. |
e905a9ed | 134 | */ |
056834d9 | 135 | static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 136 | { |
463c84b9 | 137 | struct inet_connection_sock *icsk = inet_csk(sk); |
e905a9ed | 138 | const unsigned int lss = icsk->icsk_ack.last_seg_size; |
463c84b9 | 139 | unsigned int len; |
1da177e4 | 140 | |
e905a9ed | 141 | icsk->icsk_ack.last_seg_size = 0; |
1da177e4 LT |
142 | |
143 | /* skb->len may jitter because of SACKs, even if peer | |
144 | * sends good full-sized frames. | |
145 | */ | |
056834d9 | 146 | len = skb_shinfo(skb)->gso_size ? : skb->len; |
463c84b9 | 147 | if (len >= icsk->icsk_ack.rcv_mss) { |
dcb17d22 MRL |
148 | icsk->icsk_ack.rcv_mss = min_t(unsigned int, len, |
149 | tcp_sk(sk)->advmss); | |
0b9aefea MRL |
150 | /* Account for possibly-removed options */ |
151 | if (unlikely(len > icsk->icsk_ack.rcv_mss + | |
152 | MAX_TCP_OPTION_SPACE)) | |
153 | tcp_gro_dev_warn(sk, skb, len); | |
1da177e4 LT |
154 | } else { |
155 | /* Otherwise, we make more careful check taking into account, | |
156 | * that SACKs block is variable. | |
157 | * | |
158 | * "len" is invariant segment length, including TCP header. | |
159 | */ | |
9c70220b | 160 | len += skb->data - skb_transport_header(skb); |
bee7ca9e | 161 | if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || |
1da177e4 LT |
162 | /* If PSH is not set, packet should be |
163 | * full sized, provided peer TCP is not badly broken. | |
164 | * This observation (if it is correct 8)) allows | |
165 | * to handle super-low mtu links fairly. | |
166 | */ | |
167 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && | |
aa8223c7 | 168 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { |
1da177e4 LT |
169 | /* Subtract also invariant (if peer is RFC compliant), |
170 | * tcp header plus fixed timestamp option length. | |
171 | * Resulting "len" is MSS free of SACK jitter. | |
172 | */ | |
463c84b9 ACM |
173 | len -= tcp_sk(sk)->tcp_header_len; |
174 | icsk->icsk_ack.last_seg_size = len; | |
1da177e4 | 175 | if (len == lss) { |
463c84b9 | 176 | icsk->icsk_ack.rcv_mss = len; |
1da177e4 LT |
177 | return; |
178 | } | |
179 | } | |
1ef9696c AK |
180 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) |
181 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; | |
463c84b9 | 182 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
1da177e4 LT |
183 | } |
184 | } | |
185 | ||
463c84b9 | 186 | static void tcp_incr_quickack(struct sock *sk) |
1da177e4 | 187 | { |
463c84b9 | 188 | struct inet_connection_sock *icsk = inet_csk(sk); |
95c96174 | 189 | unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); |
1da177e4 | 190 | |
056834d9 IJ |
191 | if (quickacks == 0) |
192 | quickacks = 2; | |
463c84b9 ACM |
193 | if (quickacks > icsk->icsk_ack.quick) |
194 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); | |
1da177e4 LT |
195 | } |
196 | ||
8a2ff6a5 | 197 | void tcp_enter_quickack_mode(struct sock *sk) |
1da177e4 | 198 | { |
463c84b9 ACM |
199 | struct inet_connection_sock *icsk = inet_csk(sk); |
200 | tcp_incr_quickack(sk); | |
201 | icsk->icsk_ack.pingpong = 0; | |
202 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 | 203 | } |
8a2ff6a5 | 204 | EXPORT_SYMBOL(tcp_enter_quickack_mode); |
1da177e4 LT |
205 | |
206 | /* Send ACKs quickly, if "quick" count is not exhausted | |
207 | * and the session is not interactive. | |
208 | */ | |
209 | ||
2251ae46 | 210 | static bool tcp_in_quickack_mode(struct sock *sk) |
1da177e4 | 211 | { |
463c84b9 | 212 | const struct inet_connection_sock *icsk = inet_csk(sk); |
2251ae46 | 213 | const struct dst_entry *dst = __sk_dst_get(sk); |
a2a385d6 | 214 | |
2251ae46 JM |
215 | return (dst && dst_metric(dst, RTAX_QUICKACK)) || |
216 | (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong); | |
1da177e4 LT |
217 | } |
218 | ||
735d3831 | 219 | static void tcp_ecn_queue_cwr(struct tcp_sock *tp) |
bdf1ee5d | 220 | { |
056834d9 | 221 | if (tp->ecn_flags & TCP_ECN_OK) |
bdf1ee5d IJ |
222 | tp->ecn_flags |= TCP_ECN_QUEUE_CWR; |
223 | } | |
224 | ||
735d3831 | 225 | static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb) |
bdf1ee5d IJ |
226 | { |
227 | if (tcp_hdr(skb)->cwr) | |
228 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
229 | } | |
230 | ||
735d3831 | 231 | static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp) |
bdf1ee5d IJ |
232 | { |
233 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
234 | } | |
235 | ||
735d3831 | 236 | static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) |
bdf1ee5d | 237 | { |
b82d1bb4 | 238 | switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { |
7a269ffa | 239 | case INET_ECN_NOT_ECT: |
bdf1ee5d | 240 | /* Funny extension: if ECT is not set on a segment, |
7a269ffa ED |
241 | * and we already seen ECT on a previous segment, |
242 | * it is probably a retransmit. | |
243 | */ | |
244 | if (tp->ecn_flags & TCP_ECN_SEEN) | |
bdf1ee5d | 245 | tcp_enter_quickack_mode((struct sock *)tp); |
7a269ffa ED |
246 | break; |
247 | case INET_ECN_CE: | |
9890092e FW |
248 | if (tcp_ca_needs_ecn((struct sock *)tp)) |
249 | tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE); | |
250 | ||
aae06bf5 ED |
251 | if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { |
252 | /* Better not delay acks, sender can have a very low cwnd */ | |
253 | tcp_enter_quickack_mode((struct sock *)tp); | |
254 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; | |
255 | } | |
9890092e FW |
256 | tp->ecn_flags |= TCP_ECN_SEEN; |
257 | break; | |
7a269ffa | 258 | default: |
9890092e FW |
259 | if (tcp_ca_needs_ecn((struct sock *)tp)) |
260 | tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE); | |
7a269ffa | 261 | tp->ecn_flags |= TCP_ECN_SEEN; |
9890092e | 262 | break; |
bdf1ee5d IJ |
263 | } |
264 | } | |
265 | ||
735d3831 FW |
266 | static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) |
267 | { | |
268 | if (tp->ecn_flags & TCP_ECN_OK) | |
269 | __tcp_ecn_check_ce(tp, skb); | |
270 | } | |
271 | ||
272 | static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) | |
bdf1ee5d | 273 | { |
056834d9 | 274 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) |
bdf1ee5d IJ |
275 | tp->ecn_flags &= ~TCP_ECN_OK; |
276 | } | |
277 | ||
735d3831 | 278 | static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) |
bdf1ee5d | 279 | { |
056834d9 | 280 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) |
bdf1ee5d IJ |
281 | tp->ecn_flags &= ~TCP_ECN_OK; |
282 | } | |
283 | ||
735d3831 | 284 | static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) |
bdf1ee5d | 285 | { |
056834d9 | 286 | if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) |
a2a385d6 ED |
287 | return true; |
288 | return false; | |
bdf1ee5d IJ |
289 | } |
290 | ||
1da177e4 LT |
291 | /* Buffer size and advertised window tuning. |
292 | * | |
293 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. | |
294 | */ | |
295 | ||
6ae70532 | 296 | static void tcp_sndbuf_expand(struct sock *sk) |
1da177e4 | 297 | { |
6ae70532 | 298 | const struct tcp_sock *tp = tcp_sk(sk); |
77bfc174 | 299 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; |
6ae70532 ED |
300 | int sndmem, per_mss; |
301 | u32 nr_segs; | |
302 | ||
303 | /* Worst case is non GSO/TSO : each frame consumes one skb | |
304 | * and skb->head is kmalloced using power of two area of memory | |
305 | */ | |
306 | per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + | |
307 | MAX_TCP_HEADER + | |
308 | SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); | |
309 | ||
310 | per_mss = roundup_pow_of_two(per_mss) + | |
311 | SKB_DATA_ALIGN(sizeof(struct sk_buff)); | |
312 | ||
313 | nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd); | |
314 | nr_segs = max_t(u32, nr_segs, tp->reordering + 1); | |
315 | ||
316 | /* Fast Recovery (RFC 5681 3.2) : | |
317 | * Cubic needs 1.7 factor, rounded to 2 to include | |
318 | * extra cushion (application might react slowly to POLLOUT) | |
319 | */ | |
77bfc174 YC |
320 | sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2; |
321 | sndmem *= nr_segs * per_mss; | |
1da177e4 | 322 | |
06a59ecb | 323 | if (sk->sk_sndbuf < sndmem) |
356d1833 | 324 | sk->sk_sndbuf = min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]); |
1da177e4 LT |
325 | } |
326 | ||
327 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) | |
328 | * | |
329 | * All tcp_full_space() is split to two parts: "network" buffer, allocated | |
330 | * forward and advertised in receiver window (tp->rcv_wnd) and | |
331 | * "application buffer", required to isolate scheduling/application | |
332 | * latencies from network. | |
333 | * window_clamp is maximal advertised window. It can be less than | |
334 | * tcp_full_space(), in this case tcp_full_space() - window_clamp | |
335 | * is reserved for "application" buffer. The less window_clamp is | |
336 | * the smoother our behaviour from viewpoint of network, but the lower | |
337 | * throughput and the higher sensitivity of the connection to losses. 8) | |
338 | * | |
339 | * rcv_ssthresh is more strict window_clamp used at "slow start" | |
340 | * phase to predict further behaviour of this connection. | |
341 | * It is used for two goals: | |
342 | * - to enforce header prediction at sender, even when application | |
343 | * requires some significant "application buffer". It is check #1. | |
344 | * - to prevent pruning of receive queue because of misprediction | |
345 | * of receiver window. Check #2. | |
346 | * | |
347 | * The scheme does not work when sender sends good segments opening | |
caa20d9a | 348 | * window and then starts to feed us spaghetti. But it should work |
1da177e4 LT |
349 | * in common situations. Otherwise, we have to rely on queue collapsing. |
350 | */ | |
351 | ||
352 | /* Slow part of check#2. */ | |
9e412ba7 | 353 | static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 354 | { |
9e412ba7 | 355 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 356 | /* Optimize this! */ |
94f0893e | 357 | int truesize = tcp_win_from_space(sk, skb->truesize) >> 1; |
356d1833 | 358 | int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1; |
1da177e4 LT |
359 | |
360 | while (tp->rcv_ssthresh <= window) { | |
361 | if (truesize <= skb->len) | |
463c84b9 | 362 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; |
1da177e4 LT |
363 | |
364 | truesize >>= 1; | |
365 | window >>= 1; | |
366 | } | |
367 | return 0; | |
368 | } | |
369 | ||
cf533ea5 | 370 | static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 371 | { |
9e412ba7 IJ |
372 | struct tcp_sock *tp = tcp_sk(sk); |
373 | ||
1da177e4 LT |
374 | /* Check #1 */ |
375 | if (tp->rcv_ssthresh < tp->window_clamp && | |
376 | (int)tp->rcv_ssthresh < tcp_space(sk) && | |
b8da51eb | 377 | !tcp_under_memory_pressure(sk)) { |
1da177e4 LT |
378 | int incr; |
379 | ||
380 | /* Check #2. Increase window, if skb with such overhead | |
381 | * will fit to rcvbuf in future. | |
382 | */ | |
94f0893e | 383 | if (tcp_win_from_space(sk, skb->truesize) <= skb->len) |
056834d9 | 384 | incr = 2 * tp->advmss; |
1da177e4 | 385 | else |
9e412ba7 | 386 | incr = __tcp_grow_window(sk, skb); |
1da177e4 LT |
387 | |
388 | if (incr) { | |
4d846f02 | 389 | incr = max_t(int, incr, 2 * skb->len); |
056834d9 IJ |
390 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, |
391 | tp->window_clamp); | |
463c84b9 | 392 | inet_csk(sk)->icsk_ack.quick |= 1; |
1da177e4 LT |
393 | } |
394 | } | |
395 | } | |
396 | ||
397 | /* 3. Tuning rcvbuf, when connection enters established state. */ | |
1da177e4 LT |
398 | static void tcp_fixup_rcvbuf(struct sock *sk) |
399 | { | |
e9266a02 | 400 | u32 mss = tcp_sk(sk)->advmss; |
e9266a02 | 401 | int rcvmem; |
1da177e4 | 402 | |
85f16525 YC |
403 | rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) * |
404 | tcp_default_init_rwnd(mss); | |
e9266a02 | 405 | |
b0983d3c ED |
406 | /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency |
407 | * Allow enough cushion so that sender is not limited by our window | |
408 | */ | |
4540c0cf | 409 | if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) |
b0983d3c ED |
410 | rcvmem <<= 2; |
411 | ||
e9266a02 | 412 | if (sk->sk_rcvbuf < rcvmem) |
356d1833 | 413 | sk->sk_rcvbuf = min(rcvmem, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); |
1da177e4 LT |
414 | } |
415 | ||
caa20d9a | 416 | /* 4. Try to fixup all. It is made immediately after connection enters |
1da177e4 LT |
417 | * established state. |
418 | */ | |
10467163 | 419 | void tcp_init_buffer_space(struct sock *sk) |
1da177e4 | 420 | { |
0c12654a | 421 | int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win; |
1da177e4 LT |
422 | struct tcp_sock *tp = tcp_sk(sk); |
423 | int maxwin; | |
424 | ||
425 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) | |
426 | tcp_fixup_rcvbuf(sk); | |
427 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) | |
6ae70532 | 428 | tcp_sndbuf_expand(sk); |
1da177e4 LT |
429 | |
430 | tp->rcvq_space.space = tp->rcv_wnd; | |
9a568de4 | 431 | tcp_mstamp_refresh(tp); |
645f4c6f | 432 | tp->rcvq_space.time = tp->tcp_mstamp; |
b0983d3c | 433 | tp->rcvq_space.seq = tp->copied_seq; |
1da177e4 LT |
434 | |
435 | maxwin = tcp_full_space(sk); | |
436 | ||
437 | if (tp->window_clamp >= maxwin) { | |
438 | tp->window_clamp = maxwin; | |
439 | ||
0c12654a | 440 | if (tcp_app_win && maxwin > 4 * tp->advmss) |
1da177e4 | 441 | tp->window_clamp = max(maxwin - |
0c12654a | 442 | (maxwin >> tcp_app_win), |
1da177e4 LT |
443 | 4 * tp->advmss); |
444 | } | |
445 | ||
446 | /* Force reservation of one segment. */ | |
0c12654a | 447 | if (tcp_app_win && |
1da177e4 LT |
448 | tp->window_clamp > 2 * tp->advmss && |
449 | tp->window_clamp + tp->advmss > maxwin) | |
450 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); | |
451 | ||
452 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); | |
c2203cf7 | 453 | tp->snd_cwnd_stamp = tcp_jiffies32; |
1da177e4 LT |
454 | } |
455 | ||
1da177e4 | 456 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ |
9e412ba7 | 457 | static void tcp_clamp_window(struct sock *sk) |
1da177e4 | 458 | { |
9e412ba7 | 459 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 460 | struct inet_connection_sock *icsk = inet_csk(sk); |
356d1833 | 461 | struct net *net = sock_net(sk); |
1da177e4 | 462 | |
6687e988 | 463 | icsk->icsk_ack.quick = 0; |
1da177e4 | 464 | |
356d1833 | 465 | if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] && |
326f36e9 | 466 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && |
b8da51eb | 467 | !tcp_under_memory_pressure(sk) && |
180d8cd9 | 468 | sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { |
326f36e9 | 469 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), |
356d1833 | 470 | net->ipv4.sysctl_tcp_rmem[2]); |
1da177e4 | 471 | } |
326f36e9 | 472 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) |
056834d9 | 473 | tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); |
1da177e4 LT |
474 | } |
475 | ||
40efc6fa SH |
476 | /* Initialize RCV_MSS value. |
477 | * RCV_MSS is an our guess about MSS used by the peer. | |
478 | * We haven't any direct information about the MSS. | |
479 | * It's better to underestimate the RCV_MSS rather than overestimate. | |
480 | * Overestimations make us ACKing less frequently than needed. | |
481 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). | |
482 | */ | |
483 | void tcp_initialize_rcv_mss(struct sock *sk) | |
484 | { | |
cf533ea5 | 485 | const struct tcp_sock *tp = tcp_sk(sk); |
40efc6fa SH |
486 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); |
487 | ||
056834d9 | 488 | hint = min(hint, tp->rcv_wnd / 2); |
bee7ca9e | 489 | hint = min(hint, TCP_MSS_DEFAULT); |
40efc6fa SH |
490 | hint = max(hint, TCP_MIN_MSS); |
491 | ||
492 | inet_csk(sk)->icsk_ack.rcv_mss = hint; | |
493 | } | |
4bc2f18b | 494 | EXPORT_SYMBOL(tcp_initialize_rcv_mss); |
40efc6fa | 495 | |
1da177e4 LT |
496 | /* Receiver "autotuning" code. |
497 | * | |
498 | * The algorithm for RTT estimation w/o timestamps is based on | |
499 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. | |
631dd1a8 | 500 | * <http://public.lanl.gov/radiant/pubs.html#DRS> |
1da177e4 LT |
501 | * |
502 | * More detail on this code can be found at | |
631dd1a8 | 503 | * <http://staff.psc.edu/jheffner/>, |
1da177e4 LT |
504 | * though this reference is out of date. A new paper |
505 | * is pending. | |
506 | */ | |
507 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) | |
508 | { | |
645f4c6f | 509 | u32 new_sample = tp->rcv_rtt_est.rtt_us; |
1da177e4 LT |
510 | long m = sample; |
511 | ||
1da177e4 LT |
512 | if (new_sample != 0) { |
513 | /* If we sample in larger samples in the non-timestamp | |
514 | * case, we could grossly overestimate the RTT especially | |
515 | * with chatty applications or bulk transfer apps which | |
516 | * are stalled on filesystem I/O. | |
517 | * | |
518 | * Also, since we are only going for a minimum in the | |
31f34269 | 519 | * non-timestamp case, we do not smooth things out |
caa20d9a | 520 | * else with timestamps disabled convergence takes too |
1da177e4 LT |
521 | * long. |
522 | */ | |
523 | if (!win_dep) { | |
524 | m -= (new_sample >> 3); | |
525 | new_sample += m; | |
18a223e0 NC |
526 | } else { |
527 | m <<= 3; | |
528 | if (m < new_sample) | |
529 | new_sample = m; | |
530 | } | |
1da177e4 | 531 | } else { |
caa20d9a | 532 | /* No previous measure. */ |
1da177e4 LT |
533 | new_sample = m << 3; |
534 | } | |
535 | ||
645f4c6f | 536 | tp->rcv_rtt_est.rtt_us = new_sample; |
1da177e4 LT |
537 | } |
538 | ||
539 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) | |
540 | { | |
645f4c6f ED |
541 | u32 delta_us; |
542 | ||
9a568de4 | 543 | if (tp->rcv_rtt_est.time == 0) |
1da177e4 LT |
544 | goto new_measure; |
545 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) | |
546 | return; | |
9a568de4 | 547 | delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time); |
9ee11bd0 WW |
548 | if (!delta_us) |
549 | delta_us = 1; | |
645f4c6f | 550 | tcp_rcv_rtt_update(tp, delta_us, 1); |
1da177e4 LT |
551 | |
552 | new_measure: | |
553 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; | |
645f4c6f | 554 | tp->rcv_rtt_est.time = tp->tcp_mstamp; |
1da177e4 LT |
555 | } |
556 | ||
056834d9 IJ |
557 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, |
558 | const struct sk_buff *skb) | |
1da177e4 | 559 | { |
463c84b9 | 560 | struct tcp_sock *tp = tcp_sk(sk); |
9a568de4 | 561 | |
1da177e4 LT |
562 | if (tp->rx_opt.rcv_tsecr && |
563 | (TCP_SKB_CB(skb)->end_seq - | |
9a568de4 ED |
564 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) { |
565 | u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr; | |
9ee11bd0 | 566 | u32 delta_us; |
9a568de4 | 567 | |
9ee11bd0 WW |
568 | if (!delta) |
569 | delta = 1; | |
570 | delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ); | |
9a568de4 ED |
571 | tcp_rcv_rtt_update(tp, delta_us, 0); |
572 | } | |
1da177e4 LT |
573 | } |
574 | ||
575 | /* | |
576 | * This function should be called every time data is copied to user space. | |
577 | * It calculates the appropriate TCP receive buffer space. | |
578 | */ | |
579 | void tcp_rcv_space_adjust(struct sock *sk) | |
580 | { | |
581 | struct tcp_sock *tp = tcp_sk(sk); | |
ee07c64f | 582 | u32 copied; |
1da177e4 | 583 | int time; |
e905a9ed | 584 | |
86323850 | 585 | tcp_mstamp_refresh(tp); |
9a568de4 | 586 | time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time); |
645f4c6f | 587 | if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0) |
1da177e4 | 588 | return; |
e905a9ed | 589 | |
b0983d3c ED |
590 | /* Number of bytes copied to user in last RTT */ |
591 | copied = tp->copied_seq - tp->rcvq_space.seq; | |
592 | if (copied <= tp->rcvq_space.space) | |
593 | goto new_measure; | |
594 | ||
595 | /* A bit of theory : | |
596 | * copied = bytes received in previous RTT, our base window | |
597 | * To cope with packet losses, we need a 2x factor | |
598 | * To cope with slow start, and sender growing its cwin by 100 % | |
599 | * every RTT, we need a 4x factor, because the ACK we are sending | |
600 | * now is for the next RTT, not the current one : | |
601 | * <prev RTT . ><current RTT .. ><next RTT .... > | |
602 | */ | |
603 | ||
4540c0cf | 604 | if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf && |
b0983d3c | 605 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { |
ee07c64f ED |
606 | int rcvmem, rcvbuf; |
607 | u64 rcvwin; | |
1da177e4 | 608 | |
b0983d3c ED |
609 | /* minimal window to cope with packet losses, assuming |
610 | * steady state. Add some cushion because of small variations. | |
611 | */ | |
ee07c64f | 612 | rcvwin = ((u64)copied << 1) + 16 * tp->advmss; |
1da177e4 | 613 | |
b0983d3c ED |
614 | /* If rate increased by 25%, |
615 | * assume slow start, rcvwin = 3 * copied | |
616 | * If rate increased by 50%, | |
617 | * assume sender can use 2x growth, rcvwin = 4 * copied | |
618 | */ | |
619 | if (copied >= | |
620 | tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) { | |
621 | if (copied >= | |
622 | tp->rcvq_space.space + (tp->rcvq_space.space >> 1)) | |
623 | rcvwin <<= 1; | |
624 | else | |
625 | rcvwin += (rcvwin >> 1); | |
626 | } | |
1da177e4 | 627 | |
b0983d3c | 628 | rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER); |
94f0893e | 629 | while (tcp_win_from_space(sk, rcvmem) < tp->advmss) |
b0983d3c | 630 | rcvmem += 128; |
1da177e4 | 631 | |
ee07c64f ED |
632 | do_div(rcvwin, tp->advmss); |
633 | rcvbuf = min_t(u64, rcvwin * rcvmem, | |
356d1833 | 634 | sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); |
b0983d3c ED |
635 | if (rcvbuf > sk->sk_rcvbuf) { |
636 | sk->sk_rcvbuf = rcvbuf; | |
1da177e4 | 637 | |
b0983d3c | 638 | /* Make the window clamp follow along. */ |
82d42ab7 | 639 | tp->window_clamp = tcp_win_from_space(sk, rcvbuf); |
1da177e4 LT |
640 | } |
641 | } | |
b0983d3c | 642 | tp->rcvq_space.space = copied; |
e905a9ed | 643 | |
1da177e4 LT |
644 | new_measure: |
645 | tp->rcvq_space.seq = tp->copied_seq; | |
645f4c6f | 646 | tp->rcvq_space.time = tp->tcp_mstamp; |
1da177e4 LT |
647 | } |
648 | ||
649 | /* There is something which you must keep in mind when you analyze the | |
650 | * behavior of the tp->ato delayed ack timeout interval. When a | |
651 | * connection starts up, we want to ack as quickly as possible. The | |
652 | * problem is that "good" TCP's do slow start at the beginning of data | |
653 | * transmission. The means that until we send the first few ACK's the | |
654 | * sender will sit on his end and only queue most of his data, because | |
655 | * he can only send snd_cwnd unacked packets at any given time. For | |
656 | * each ACK we send, he increments snd_cwnd and transmits more of his | |
657 | * queue. -DaveM | |
658 | */ | |
9e412ba7 | 659 | static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 660 | { |
9e412ba7 | 661 | struct tcp_sock *tp = tcp_sk(sk); |
463c84b9 | 662 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
663 | u32 now; |
664 | ||
463c84b9 | 665 | inet_csk_schedule_ack(sk); |
1da177e4 | 666 | |
463c84b9 | 667 | tcp_measure_rcv_mss(sk, skb); |
1da177e4 LT |
668 | |
669 | tcp_rcv_rtt_measure(tp); | |
e905a9ed | 670 | |
70eabf0e | 671 | now = tcp_jiffies32; |
1da177e4 | 672 | |
463c84b9 | 673 | if (!icsk->icsk_ack.ato) { |
1da177e4 LT |
674 | /* The _first_ data packet received, initialize |
675 | * delayed ACK engine. | |
676 | */ | |
463c84b9 ACM |
677 | tcp_incr_quickack(sk); |
678 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 | 679 | } else { |
463c84b9 | 680 | int m = now - icsk->icsk_ack.lrcvtime; |
1da177e4 | 681 | |
056834d9 | 682 | if (m <= TCP_ATO_MIN / 2) { |
1da177e4 | 683 | /* The fastest case is the first. */ |
463c84b9 ACM |
684 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; |
685 | } else if (m < icsk->icsk_ack.ato) { | |
686 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; | |
687 | if (icsk->icsk_ack.ato > icsk->icsk_rto) | |
688 | icsk->icsk_ack.ato = icsk->icsk_rto; | |
689 | } else if (m > icsk->icsk_rto) { | |
caa20d9a | 690 | /* Too long gap. Apparently sender failed to |
1da177e4 LT |
691 | * restart window, so that we send ACKs quickly. |
692 | */ | |
463c84b9 | 693 | tcp_incr_quickack(sk); |
3ab224be | 694 | sk_mem_reclaim(sk); |
1da177e4 LT |
695 | } |
696 | } | |
463c84b9 | 697 | icsk->icsk_ack.lrcvtime = now; |
1da177e4 | 698 | |
735d3831 | 699 | tcp_ecn_check_ce(tp, skb); |
1da177e4 LT |
700 | |
701 | if (skb->len >= 128) | |
9e412ba7 | 702 | tcp_grow_window(sk, skb); |
1da177e4 LT |
703 | } |
704 | ||
1da177e4 LT |
705 | /* Called to compute a smoothed rtt estimate. The data fed to this |
706 | * routine either comes from timestamps, or from segments that were | |
707 | * known _not_ to have been retransmitted [see Karn/Partridge | |
708 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 | |
709 | * piece by Van Jacobson. | |
710 | * NOTE: the next three routines used to be one big routine. | |
711 | * To save cycles in the RFC 1323 implementation it was better to break | |
712 | * it up into three procedures. -- erics | |
713 | */ | |
740b0f18 | 714 | static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) |
1da177e4 | 715 | { |
6687e988 | 716 | struct tcp_sock *tp = tcp_sk(sk); |
740b0f18 ED |
717 | long m = mrtt_us; /* RTT */ |
718 | u32 srtt = tp->srtt_us; | |
1da177e4 | 719 | |
1da177e4 LT |
720 | /* The following amusing code comes from Jacobson's |
721 | * article in SIGCOMM '88. Note that rtt and mdev | |
722 | * are scaled versions of rtt and mean deviation. | |
e905a9ed | 723 | * This is designed to be as fast as possible |
1da177e4 LT |
724 | * m stands for "measurement". |
725 | * | |
726 | * On a 1990 paper the rto value is changed to: | |
727 | * RTO = rtt + 4 * mdev | |
728 | * | |
729 | * Funny. This algorithm seems to be very broken. | |
730 | * These formulae increase RTO, when it should be decreased, increase | |
31f34269 | 731 | * too slowly, when it should be increased quickly, decrease too quickly |
1da177e4 LT |
732 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely |
733 | * does not matter how to _calculate_ it. Seems, it was trap | |
734 | * that VJ failed to avoid. 8) | |
735 | */ | |
4a5ab4e2 ED |
736 | if (srtt != 0) { |
737 | m -= (srtt >> 3); /* m is now error in rtt est */ | |
738 | srtt += m; /* rtt = 7/8 rtt + 1/8 new */ | |
1da177e4 LT |
739 | if (m < 0) { |
740 | m = -m; /* m is now abs(error) */ | |
740b0f18 | 741 | m -= (tp->mdev_us >> 2); /* similar update on mdev */ |
1da177e4 LT |
742 | /* This is similar to one of Eifel findings. |
743 | * Eifel blocks mdev updates when rtt decreases. | |
744 | * This solution is a bit different: we use finer gain | |
745 | * for mdev in this case (alpha*beta). | |
746 | * Like Eifel it also prevents growth of rto, | |
747 | * but also it limits too fast rto decreases, | |
748 | * happening in pure Eifel. | |
749 | */ | |
750 | if (m > 0) | |
751 | m >>= 3; | |
752 | } else { | |
740b0f18 | 753 | m -= (tp->mdev_us >> 2); /* similar update on mdev */ |
1da177e4 | 754 | } |
740b0f18 ED |
755 | tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ |
756 | if (tp->mdev_us > tp->mdev_max_us) { | |
757 | tp->mdev_max_us = tp->mdev_us; | |
758 | if (tp->mdev_max_us > tp->rttvar_us) | |
759 | tp->rttvar_us = tp->mdev_max_us; | |
1da177e4 LT |
760 | } |
761 | if (after(tp->snd_una, tp->rtt_seq)) { | |
740b0f18 ED |
762 | if (tp->mdev_max_us < tp->rttvar_us) |
763 | tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; | |
1da177e4 | 764 | tp->rtt_seq = tp->snd_nxt; |
740b0f18 | 765 | tp->mdev_max_us = tcp_rto_min_us(sk); |
1da177e4 LT |
766 | } |
767 | } else { | |
768 | /* no previous measure. */ | |
4a5ab4e2 | 769 | srtt = m << 3; /* take the measured time to be rtt */ |
740b0f18 ED |
770 | tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ |
771 | tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); | |
772 | tp->mdev_max_us = tp->rttvar_us; | |
1da177e4 LT |
773 | tp->rtt_seq = tp->snd_nxt; |
774 | } | |
740b0f18 | 775 | tp->srtt_us = max(1U, srtt); |
1da177e4 LT |
776 | } |
777 | ||
95bd09eb ED |
778 | static void tcp_update_pacing_rate(struct sock *sk) |
779 | { | |
780 | const struct tcp_sock *tp = tcp_sk(sk); | |
781 | u64 rate; | |
782 | ||
783 | /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ | |
43e122b0 ED |
784 | rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); |
785 | ||
786 | /* current rate is (cwnd * mss) / srtt | |
787 | * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. | |
788 | * In Congestion Avoidance phase, set it to 120 % the current rate. | |
789 | * | |
790 | * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) | |
791 | * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching | |
792 | * end of slow start and should slow down. | |
793 | */ | |
794 | if (tp->snd_cwnd < tp->snd_ssthresh / 2) | |
23a7102a | 795 | rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio; |
43e122b0 | 796 | else |
c26e91f8 | 797 | rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio; |
95bd09eb ED |
798 | |
799 | rate *= max(tp->snd_cwnd, tp->packets_out); | |
800 | ||
740b0f18 ED |
801 | if (likely(tp->srtt_us)) |
802 | do_div(rate, tp->srtt_us); | |
95bd09eb | 803 | |
a9da6f29 | 804 | /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate |
ba537427 ED |
805 | * without any lock. We want to make sure compiler wont store |
806 | * intermediate values in this location. | |
807 | */ | |
a9da6f29 MR |
808 | WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate, |
809 | sk->sk_max_pacing_rate)); | |
95bd09eb ED |
810 | } |
811 | ||
1da177e4 LT |
812 | /* Calculate rto without backoff. This is the second half of Van Jacobson's |
813 | * routine referred to above. | |
814 | */ | |
f7e56a76 | 815 | static void tcp_set_rto(struct sock *sk) |
1da177e4 | 816 | { |
463c84b9 | 817 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
818 | /* Old crap is replaced with new one. 8) |
819 | * | |
820 | * More seriously: | |
821 | * 1. If rtt variance happened to be less 50msec, it is hallucination. | |
822 | * It cannot be less due to utterly erratic ACK generation made | |
823 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ | |
824 | * to do with delayed acks, because at cwnd>2 true delack timeout | |
825 | * is invisible. Actually, Linux-2.4 also generates erratic | |
caa20d9a | 826 | * ACKs in some circumstances. |
1da177e4 | 827 | */ |
f1ecd5d9 | 828 | inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); |
1da177e4 LT |
829 | |
830 | /* 2. Fixups made earlier cannot be right. | |
831 | * If we do not estimate RTO correctly without them, | |
832 | * all the algo is pure shit and should be replaced | |
caa20d9a | 833 | * with correct one. It is exactly, which we pretend to do. |
1da177e4 | 834 | */ |
1da177e4 | 835 | |
ee6aac59 IJ |
836 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo |
837 | * guarantees that rto is higher. | |
838 | */ | |
f1ecd5d9 | 839 | tcp_bound_rto(sk); |
1da177e4 LT |
840 | } |
841 | ||
cf533ea5 | 842 | __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) |
1da177e4 LT |
843 | { |
844 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); | |
845 | ||
22b71c8f | 846 | if (!cwnd) |
442b9635 | 847 | cwnd = TCP_INIT_CWND; |
1da177e4 LT |
848 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); |
849 | } | |
850 | ||
564262c1 | 851 | /* Take a notice that peer is sending D-SACKs */ |
e60402d0 IJ |
852 | static void tcp_dsack_seen(struct tcp_sock *tp) |
853 | { | |
ab56222a | 854 | tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; |
1f255691 | 855 | tp->rack.dsack_seen = 1; |
e60402d0 IJ |
856 | } |
857 | ||
737ff314 YC |
858 | /* It's reordering when higher sequence was delivered (i.e. sacked) before |
859 | * some lower never-retransmitted sequence ("low_seq"). The maximum reordering | |
860 | * distance is approximated in full-mss packet distance ("reordering"). | |
861 | */ | |
862 | static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq, | |
863 | const int ts) | |
1da177e4 | 864 | { |
6687e988 | 865 | struct tcp_sock *tp = tcp_sk(sk); |
737ff314 YC |
866 | const u32 mss = tp->mss_cache; |
867 | u32 fack, metric; | |
40b215e5 | 868 | |
737ff314 YC |
869 | fack = tcp_highest_sack_seq(tp); |
870 | if (!before(low_seq, fack)) | |
6f5b24ee SHY |
871 | return; |
872 | ||
737ff314 YC |
873 | metric = fack - low_seq; |
874 | if ((metric > tp->reordering * mss) && mss) { | |
1da177e4 | 875 | #if FASTRETRANS_DEBUG > 1 |
91df42be JP |
876 | pr_debug("Disorder%d %d %u f%u s%u rr%d\n", |
877 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, | |
878 | tp->reordering, | |
737ff314 | 879 | 0, |
91df42be JP |
880 | tp->sacked_out, |
881 | tp->undo_marker ? tp->undo_retrans : 0); | |
1da177e4 | 882 | #endif |
737ff314 YC |
883 | tp->reordering = min_t(u32, (metric + mss - 1) / mss, |
884 | sock_net(sk)->ipv4.sysctl_tcp_max_reordering); | |
1da177e4 | 885 | } |
eed530b6 | 886 | |
4f41b1c5 | 887 | tp->rack.reord = 1; |
2d2517ee | 888 | /* This exciting event is worth to be remembered. 8) */ |
737ff314 YC |
889 | NET_INC_STATS(sock_net(sk), |
890 | ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER); | |
1da177e4 LT |
891 | } |
892 | ||
006f582c | 893 | /* This must be called before lost_out is incremented */ |
c8c213f2 IJ |
894 | static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) |
895 | { | |
51456b29 | 896 | if (!tp->retransmit_skb_hint || |
c8c213f2 IJ |
897 | before(TCP_SKB_CB(skb)->seq, |
898 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) | |
006f582c | 899 | tp->retransmit_skb_hint = skb; |
c8c213f2 IJ |
900 | } |
901 | ||
0682e690 NC |
902 | /* Sum the number of packets on the wire we have marked as lost. |
903 | * There are two cases we care about here: | |
904 | * a) Packet hasn't been marked lost (nor retransmitted), | |
905 | * and this is the first loss. | |
906 | * b) Packet has been marked both lost and retransmitted, | |
907 | * and this means we think it was lost again. | |
908 | */ | |
909 | static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb) | |
910 | { | |
911 | __u8 sacked = TCP_SKB_CB(skb)->sacked; | |
912 | ||
913 | if (!(sacked & TCPCB_LOST) || | |
914 | ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS))) | |
915 | tp->lost += tcp_skb_pcount(skb); | |
916 | } | |
917 | ||
41ea36e3 IJ |
918 | static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) |
919 | { | |
920 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
921 | tcp_verify_retransmit_hint(tp, skb); | |
922 | ||
923 | tp->lost_out += tcp_skb_pcount(skb); | |
0682e690 | 924 | tcp_sum_lost(tp, skb); |
41ea36e3 IJ |
925 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
926 | } | |
927 | } | |
928 | ||
4f41b1c5 | 929 | void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb) |
006f582c IJ |
930 | { |
931 | tcp_verify_retransmit_hint(tp, skb); | |
932 | ||
0682e690 | 933 | tcp_sum_lost(tp, skb); |
006f582c IJ |
934 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { |
935 | tp->lost_out += tcp_skb_pcount(skb); | |
936 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
937 | } | |
938 | } | |
939 | ||
1da177e4 LT |
940 | /* This procedure tags the retransmission queue when SACKs arrive. |
941 | * | |
942 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). | |
943 | * Packets in queue with these bits set are counted in variables | |
944 | * sacked_out, retrans_out and lost_out, correspondingly. | |
945 | * | |
946 | * Valid combinations are: | |
947 | * Tag InFlight Description | |
948 | * 0 1 - orig segment is in flight. | |
949 | * S 0 - nothing flies, orig reached receiver. | |
950 | * L 0 - nothing flies, orig lost by net. | |
951 | * R 2 - both orig and retransmit are in flight. | |
952 | * L|R 1 - orig is lost, retransmit is in flight. | |
953 | * S|R 1 - orig reached receiver, retrans is still in flight. | |
954 | * (L|S|R is logically valid, it could occur when L|R is sacked, | |
955 | * but it is equivalent to plain S and code short-curcuits it to S. | |
956 | * L|S is logically invalid, it would mean -1 packet in flight 8)) | |
957 | * | |
958 | * These 6 states form finite state machine, controlled by the following events: | |
959 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) | |
960 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) | |
974c1236 | 961 | * 3. Loss detection event of two flavors: |
1da177e4 LT |
962 | * A. Scoreboard estimator decided the packet is lost. |
963 | * A'. Reno "three dupacks" marks head of queue lost. | |
974c1236 | 964 | * B. SACK arrives sacking SND.NXT at the moment, when the |
1da177e4 LT |
965 | * segment was retransmitted. |
966 | * 4. D-SACK added new rule: D-SACK changes any tag to S. | |
967 | * | |
968 | * It is pleasant to note, that state diagram turns out to be commutative, | |
969 | * so that we are allowed not to be bothered by order of our actions, | |
970 | * when multiple events arrive simultaneously. (see the function below). | |
971 | * | |
972 | * Reordering detection. | |
973 | * -------------------- | |
974 | * Reordering metric is maximal distance, which a packet can be displaced | |
975 | * in packet stream. With SACKs we can estimate it: | |
976 | * | |
977 | * 1. SACK fills old hole and the corresponding segment was not | |
978 | * ever retransmitted -> reordering. Alas, we cannot use it | |
979 | * when segment was retransmitted. | |
980 | * 2. The last flaw is solved with D-SACK. D-SACK arrives | |
981 | * for retransmitted and already SACKed segment -> reordering.. | |
982 | * Both of these heuristics are not used in Loss state, when we cannot | |
983 | * account for retransmits accurately. | |
5b3c9882 IJ |
984 | * |
985 | * SACK block validation. | |
986 | * ---------------------- | |
987 | * | |
988 | * SACK block range validation checks that the received SACK block fits to | |
989 | * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. | |
990 | * Note that SND.UNA is not included to the range though being valid because | |
0e835331 IJ |
991 | * it means that the receiver is rather inconsistent with itself reporting |
992 | * SACK reneging when it should advance SND.UNA. Such SACK block this is | |
993 | * perfectly valid, however, in light of RFC2018 which explicitly states | |
994 | * that "SACK block MUST reflect the newest segment. Even if the newest | |
995 | * segment is going to be discarded ...", not that it looks very clever | |
996 | * in case of head skb. Due to potentional receiver driven attacks, we | |
997 | * choose to avoid immediate execution of a walk in write queue due to | |
998 | * reneging and defer head skb's loss recovery to standard loss recovery | |
999 | * procedure that will eventually trigger (nothing forbids us doing this). | |
5b3c9882 IJ |
1000 | * |
1001 | * Implements also blockage to start_seq wrap-around. Problem lies in the | |
1002 | * fact that though start_seq (s) is before end_seq (i.e., not reversed), | |
1003 | * there's no guarantee that it will be before snd_nxt (n). The problem | |
1004 | * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt | |
1005 | * wrap (s_w): | |
1006 | * | |
1007 | * <- outs wnd -> <- wrapzone -> | |
1008 | * u e n u_w e_w s n_w | |
1009 | * | | | | | | | | |
1010 | * |<------------+------+----- TCP seqno space --------------+---------->| | |
1011 | * ...-- <2^31 ->| |<--------... | |
1012 | * ...---- >2^31 ------>| |<--------... | |
1013 | * | |
1014 | * Current code wouldn't be vulnerable but it's better still to discard such | |
1015 | * crazy SACK blocks. Doing this check for start_seq alone closes somewhat | |
1016 | * similar case (end_seq after snd_nxt wrap) as earlier reversed check in | |
1017 | * snd_nxt wrap -> snd_una region will then become "well defined", i.e., | |
1018 | * equal to the ideal case (infinite seqno space without wrap caused issues). | |
1019 | * | |
1020 | * With D-SACK the lower bound is extended to cover sequence space below | |
1021 | * SND.UNA down to undo_marker, which is the last point of interest. Yet | |
564262c1 | 1022 | * again, D-SACK block must not to go across snd_una (for the same reason as |
5b3c9882 IJ |
1023 | * for the normal SACK blocks, explained above). But there all simplicity |
1024 | * ends, TCP might receive valid D-SACKs below that. As long as they reside | |
1025 | * fully below undo_marker they do not affect behavior in anyway and can | |
1026 | * therefore be safely ignored. In rare cases (which are more or less | |
1027 | * theoretical ones), the D-SACK will nicely cross that boundary due to skb | |
1028 | * fragmentation and packet reordering past skb's retransmission. To consider | |
1029 | * them correctly, the acceptable range must be extended even more though | |
1030 | * the exact amount is rather hard to quantify. However, tp->max_window can | |
1031 | * be used as an exaggerated estimate. | |
1da177e4 | 1032 | */ |
a2a385d6 ED |
1033 | static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, |
1034 | u32 start_seq, u32 end_seq) | |
5b3c9882 IJ |
1035 | { |
1036 | /* Too far in future, or reversed (interpretation is ambiguous) */ | |
1037 | if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) | |
a2a385d6 | 1038 | return false; |
5b3c9882 IJ |
1039 | |
1040 | /* Nasty start_seq wrap-around check (see comments above) */ | |
1041 | if (!before(start_seq, tp->snd_nxt)) | |
a2a385d6 | 1042 | return false; |
5b3c9882 | 1043 | |
564262c1 | 1044 | /* In outstanding window? ...This is valid exit for D-SACKs too. |
5b3c9882 IJ |
1045 | * start_seq == snd_una is non-sensical (see comments above) |
1046 | */ | |
1047 | if (after(start_seq, tp->snd_una)) | |
a2a385d6 | 1048 | return true; |
5b3c9882 IJ |
1049 | |
1050 | if (!is_dsack || !tp->undo_marker) | |
a2a385d6 | 1051 | return false; |
5b3c9882 IJ |
1052 | |
1053 | /* ...Then it's D-SACK, and must reside below snd_una completely */ | |
f779b2d6 | 1054 | if (after(end_seq, tp->snd_una)) |
a2a385d6 | 1055 | return false; |
5b3c9882 IJ |
1056 | |
1057 | if (!before(start_seq, tp->undo_marker)) | |
a2a385d6 | 1058 | return true; |
5b3c9882 IJ |
1059 | |
1060 | /* Too old */ | |
1061 | if (!after(end_seq, tp->undo_marker)) | |
a2a385d6 | 1062 | return false; |
5b3c9882 IJ |
1063 | |
1064 | /* Undo_marker boundary crossing (overestimates a lot). Known already: | |
1065 | * start_seq < undo_marker and end_seq >= undo_marker. | |
1066 | */ | |
1067 | return !before(start_seq, end_seq - tp->max_window); | |
1068 | } | |
1069 | ||
a2a385d6 ED |
1070 | static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, |
1071 | struct tcp_sack_block_wire *sp, int num_sacks, | |
1072 | u32 prior_snd_una) | |
d06e021d | 1073 | { |
1ed83465 | 1074 | struct tcp_sock *tp = tcp_sk(sk); |
d3e2ce3b HH |
1075 | u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); |
1076 | u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); | |
a2a385d6 | 1077 | bool dup_sack = false; |
d06e021d DM |
1078 | |
1079 | if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { | |
a2a385d6 | 1080 | dup_sack = true; |
e60402d0 | 1081 | tcp_dsack_seen(tp); |
c10d9310 | 1082 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV); |
d06e021d | 1083 | } else if (num_sacks > 1) { |
d3e2ce3b HH |
1084 | u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); |
1085 | u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); | |
d06e021d DM |
1086 | |
1087 | if (!after(end_seq_0, end_seq_1) && | |
1088 | !before(start_seq_0, start_seq_1)) { | |
a2a385d6 | 1089 | dup_sack = true; |
e60402d0 | 1090 | tcp_dsack_seen(tp); |
c10d9310 | 1091 | NET_INC_STATS(sock_net(sk), |
de0744af | 1092 | LINUX_MIB_TCPDSACKOFORECV); |
d06e021d DM |
1093 | } |
1094 | } | |
1095 | ||
1096 | /* D-SACK for already forgotten data... Do dumb counting. */ | |
6e08d5e3 | 1097 | if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 && |
d06e021d DM |
1098 | !after(end_seq_0, prior_snd_una) && |
1099 | after(end_seq_0, tp->undo_marker)) | |
1100 | tp->undo_retrans--; | |
1101 | ||
1102 | return dup_sack; | |
1103 | } | |
1104 | ||
a1197f5a | 1105 | struct tcp_sacktag_state { |
737ff314 | 1106 | u32 reord; |
31231a8a KKJ |
1107 | /* Timestamps for earliest and latest never-retransmitted segment |
1108 | * that was SACKed. RTO needs the earliest RTT to stay conservative, | |
1109 | * but congestion control should still get an accurate delay signal. | |
1110 | */ | |
9a568de4 ED |
1111 | u64 first_sackt; |
1112 | u64 last_sackt; | |
b9f64820 | 1113 | struct rate_sample *rate; |
740b0f18 | 1114 | int flag; |
75c119af | 1115 | unsigned int mss_now; |
a1197f5a IJ |
1116 | }; |
1117 | ||
d1935942 IJ |
1118 | /* Check if skb is fully within the SACK block. In presence of GSO skbs, |
1119 | * the incoming SACK may not exactly match but we can find smaller MSS | |
1120 | * aligned portion of it that matches. Therefore we might need to fragment | |
1121 | * which may fail and creates some hassle (caller must handle error case | |
1122 | * returns). | |
832d11c5 IJ |
1123 | * |
1124 | * FIXME: this could be merged to shift decision code | |
d1935942 | 1125 | */ |
0f79efdc | 1126 | static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, |
a2a385d6 | 1127 | u32 start_seq, u32 end_seq) |
d1935942 | 1128 | { |
a2a385d6 ED |
1129 | int err; |
1130 | bool in_sack; | |
d1935942 | 1131 | unsigned int pkt_len; |
adb92db8 | 1132 | unsigned int mss; |
d1935942 IJ |
1133 | |
1134 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1135 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1136 | ||
1137 | if (tcp_skb_pcount(skb) > 1 && !in_sack && | |
1138 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { | |
adb92db8 | 1139 | mss = tcp_skb_mss(skb); |
d1935942 IJ |
1140 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); |
1141 | ||
adb92db8 | 1142 | if (!in_sack) { |
d1935942 | 1143 | pkt_len = start_seq - TCP_SKB_CB(skb)->seq; |
adb92db8 IJ |
1144 | if (pkt_len < mss) |
1145 | pkt_len = mss; | |
1146 | } else { | |
d1935942 | 1147 | pkt_len = end_seq - TCP_SKB_CB(skb)->seq; |
adb92db8 IJ |
1148 | if (pkt_len < mss) |
1149 | return -EINVAL; | |
1150 | } | |
1151 | ||
1152 | /* Round if necessary so that SACKs cover only full MSSes | |
1153 | * and/or the remaining small portion (if present) | |
1154 | */ | |
1155 | if (pkt_len > mss) { | |
1156 | unsigned int new_len = (pkt_len / mss) * mss; | |
b451e5d2 | 1157 | if (!in_sack && new_len < pkt_len) |
adb92db8 | 1158 | new_len += mss; |
adb92db8 IJ |
1159 | pkt_len = new_len; |
1160 | } | |
b451e5d2 YC |
1161 | |
1162 | if (pkt_len >= skb->len && !in_sack) | |
1163 | return 0; | |
1164 | ||
75c119af ED |
1165 | err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, |
1166 | pkt_len, mss, GFP_ATOMIC); | |
d1935942 IJ |
1167 | if (err < 0) |
1168 | return err; | |
1169 | } | |
1170 | ||
1171 | return in_sack; | |
1172 | } | |
1173 | ||
cc9a672e NC |
1174 | /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ |
1175 | static u8 tcp_sacktag_one(struct sock *sk, | |
1176 | struct tcp_sacktag_state *state, u8 sacked, | |
1177 | u32 start_seq, u32 end_seq, | |
740b0f18 | 1178 | int dup_sack, int pcount, |
9a568de4 | 1179 | u64 xmit_time) |
9e10c47c | 1180 | { |
6859d494 | 1181 | struct tcp_sock *tp = tcp_sk(sk); |
9e10c47c IJ |
1182 | |
1183 | /* Account D-SACK for retransmitted packet. */ | |
1184 | if (dup_sack && (sacked & TCPCB_RETRANS)) { | |
6e08d5e3 | 1185 | if (tp->undo_marker && tp->undo_retrans > 0 && |
cc9a672e | 1186 | after(end_seq, tp->undo_marker)) |
9e10c47c | 1187 | tp->undo_retrans--; |
737ff314 YC |
1188 | if ((sacked & TCPCB_SACKED_ACKED) && |
1189 | before(start_seq, state->reord)) | |
1190 | state->reord = start_seq; | |
9e10c47c IJ |
1191 | } |
1192 | ||
1193 | /* Nothing to do; acked frame is about to be dropped (was ACKed). */ | |
cc9a672e | 1194 | if (!after(end_seq, tp->snd_una)) |
a1197f5a | 1195 | return sacked; |
9e10c47c IJ |
1196 | |
1197 | if (!(sacked & TCPCB_SACKED_ACKED)) { | |
d2329f10 | 1198 | tcp_rack_advance(tp, sacked, end_seq, xmit_time); |
659a8ad5 | 1199 | |
9e10c47c IJ |
1200 | if (sacked & TCPCB_SACKED_RETRANS) { |
1201 | /* If the segment is not tagged as lost, | |
1202 | * we do not clear RETRANS, believing | |
1203 | * that retransmission is still in flight. | |
1204 | */ | |
1205 | if (sacked & TCPCB_LOST) { | |
a1197f5a | 1206 | sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); |
f58b22fd IJ |
1207 | tp->lost_out -= pcount; |
1208 | tp->retrans_out -= pcount; | |
9e10c47c IJ |
1209 | } |
1210 | } else { | |
1211 | if (!(sacked & TCPCB_RETRANS)) { | |
1212 | /* New sack for not retransmitted frame, | |
1213 | * which was in hole. It is reordering. | |
1214 | */ | |
cc9a672e | 1215 | if (before(start_seq, |
737ff314 YC |
1216 | tcp_highest_sack_seq(tp)) && |
1217 | before(start_seq, state->reord)) | |
1218 | state->reord = start_seq; | |
1219 | ||
e33099f9 YC |
1220 | if (!after(end_seq, tp->high_seq)) |
1221 | state->flag |= FLAG_ORIG_SACK_ACKED; | |
9a568de4 ED |
1222 | if (state->first_sackt == 0) |
1223 | state->first_sackt = xmit_time; | |
1224 | state->last_sackt = xmit_time; | |
9e10c47c IJ |
1225 | } |
1226 | ||
1227 | if (sacked & TCPCB_LOST) { | |
a1197f5a | 1228 | sacked &= ~TCPCB_LOST; |
f58b22fd | 1229 | tp->lost_out -= pcount; |
9e10c47c IJ |
1230 | } |
1231 | } | |
1232 | ||
a1197f5a IJ |
1233 | sacked |= TCPCB_SACKED_ACKED; |
1234 | state->flag |= FLAG_DATA_SACKED; | |
f58b22fd | 1235 | tp->sacked_out += pcount; |
ddf1af6f | 1236 | tp->delivered += pcount; /* Out-of-order packets delivered */ |
9e10c47c | 1237 | |
9e10c47c | 1238 | /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ |
713bafea | 1239 | if (tp->lost_skb_hint && |
cc9a672e | 1240 | before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) |
f58b22fd | 1241 | tp->lost_cnt_hint += pcount; |
9e10c47c IJ |
1242 | } |
1243 | ||
1244 | /* D-SACK. We can detect redundant retransmission in S|R and plain R | |
1245 | * frames and clear it. undo_retrans is decreased above, L|R frames | |
1246 | * are accounted above as well. | |
1247 | */ | |
a1197f5a IJ |
1248 | if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { |
1249 | sacked &= ~TCPCB_SACKED_RETRANS; | |
f58b22fd | 1250 | tp->retrans_out -= pcount; |
9e10c47c IJ |
1251 | } |
1252 | ||
a1197f5a | 1253 | return sacked; |
9e10c47c IJ |
1254 | } |
1255 | ||
daef52ba NC |
1256 | /* Shift newly-SACKed bytes from this skb to the immediately previous |
1257 | * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. | |
1258 | */ | |
f3319816 ED |
1259 | static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev, |
1260 | struct sk_buff *skb, | |
a2a385d6 ED |
1261 | struct tcp_sacktag_state *state, |
1262 | unsigned int pcount, int shifted, int mss, | |
1263 | bool dup_sack) | |
832d11c5 IJ |
1264 | { |
1265 | struct tcp_sock *tp = tcp_sk(sk); | |
daef52ba NC |
1266 | u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ |
1267 | u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ | |
832d11c5 IJ |
1268 | |
1269 | BUG_ON(!pcount); | |
1270 | ||
4c90d3b3 NC |
1271 | /* Adjust counters and hints for the newly sacked sequence |
1272 | * range but discard the return value since prev is already | |
1273 | * marked. We must tag the range first because the seq | |
1274 | * advancement below implicitly advances | |
1275 | * tcp_highest_sack_seq() when skb is highest_sack. | |
1276 | */ | |
1277 | tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, | |
59c9af42 | 1278 | start_seq, end_seq, dup_sack, pcount, |
9a568de4 | 1279 | skb->skb_mstamp); |
b9f64820 | 1280 | tcp_rate_skb_delivered(sk, skb, state->rate); |
4c90d3b3 NC |
1281 | |
1282 | if (skb == tp->lost_skb_hint) | |
0af2a0d0 NC |
1283 | tp->lost_cnt_hint += pcount; |
1284 | ||
832d11c5 IJ |
1285 | TCP_SKB_CB(prev)->end_seq += shifted; |
1286 | TCP_SKB_CB(skb)->seq += shifted; | |
1287 | ||
cd7d8498 ED |
1288 | tcp_skb_pcount_add(prev, pcount); |
1289 | BUG_ON(tcp_skb_pcount(skb) < pcount); | |
1290 | tcp_skb_pcount_add(skb, -pcount); | |
832d11c5 IJ |
1291 | |
1292 | /* When we're adding to gso_segs == 1, gso_size will be zero, | |
1293 | * in theory this shouldn't be necessary but as long as DSACK | |
1294 | * code can come after this skb later on it's better to keep | |
1295 | * setting gso_size to something. | |
1296 | */ | |
f69ad292 ED |
1297 | if (!TCP_SKB_CB(prev)->tcp_gso_size) |
1298 | TCP_SKB_CB(prev)->tcp_gso_size = mss; | |
832d11c5 IJ |
1299 | |
1300 | /* CHECKME: To clear or not to clear? Mimics normal skb currently */ | |
51466a75 | 1301 | if (tcp_skb_pcount(skb) <= 1) |
f69ad292 | 1302 | TCP_SKB_CB(skb)->tcp_gso_size = 0; |
832d11c5 | 1303 | |
832d11c5 IJ |
1304 | /* Difference in this won't matter, both ACKed by the same cumul. ACK */ |
1305 | TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); | |
1306 | ||
832d11c5 IJ |
1307 | if (skb->len > 0) { |
1308 | BUG_ON(!tcp_skb_pcount(skb)); | |
c10d9310 | 1309 | NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED); |
a2a385d6 | 1310 | return false; |
832d11c5 IJ |
1311 | } |
1312 | ||
1313 | /* Whole SKB was eaten :-) */ | |
1314 | ||
92ee76b6 IJ |
1315 | if (skb == tp->retransmit_skb_hint) |
1316 | tp->retransmit_skb_hint = prev; | |
92ee76b6 IJ |
1317 | if (skb == tp->lost_skb_hint) { |
1318 | tp->lost_skb_hint = prev; | |
1319 | tp->lost_cnt_hint -= tcp_skb_pcount(prev); | |
1320 | } | |
1321 | ||
5e8a402f | 1322 | TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
a643b5d4 | 1323 | TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor; |
5e8a402f ED |
1324 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) |
1325 | TCP_SKB_CB(prev)->end_seq++; | |
1326 | ||
832d11c5 IJ |
1327 | if (skb == tcp_highest_sack(sk)) |
1328 | tcp_advance_highest_sack(sk, skb); | |
1329 | ||
cfea5a68 | 1330 | tcp_skb_collapse_tstamp(prev, skb); |
9a568de4 ED |
1331 | if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp)) |
1332 | TCP_SKB_CB(prev)->tx.delivered_mstamp = 0; | |
b9f64820 | 1333 | |
75c119af | 1334 | tcp_rtx_queue_unlink_and_free(skb, sk); |
832d11c5 | 1335 | |
c10d9310 | 1336 | NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED); |
111cc8b9 | 1337 | |
a2a385d6 | 1338 | return true; |
832d11c5 IJ |
1339 | } |
1340 | ||
1341 | /* I wish gso_size would have a bit more sane initialization than | |
1342 | * something-or-zero which complicates things | |
1343 | */ | |
cf533ea5 | 1344 | static int tcp_skb_seglen(const struct sk_buff *skb) |
832d11c5 | 1345 | { |
775ffabf | 1346 | return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); |
832d11c5 IJ |
1347 | } |
1348 | ||
1349 | /* Shifting pages past head area doesn't work */ | |
cf533ea5 | 1350 | static int skb_can_shift(const struct sk_buff *skb) |
832d11c5 IJ |
1351 | { |
1352 | return !skb_headlen(skb) && skb_is_nonlinear(skb); | |
1353 | } | |
1354 | ||
1355 | /* Try collapsing SACK blocks spanning across multiple skbs to a single | |
1356 | * skb. | |
1357 | */ | |
1358 | static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, | |
a1197f5a | 1359 | struct tcp_sacktag_state *state, |
832d11c5 | 1360 | u32 start_seq, u32 end_seq, |
a2a385d6 | 1361 | bool dup_sack) |
832d11c5 IJ |
1362 | { |
1363 | struct tcp_sock *tp = tcp_sk(sk); | |
1364 | struct sk_buff *prev; | |
1365 | int mss; | |
1366 | int pcount = 0; | |
1367 | int len; | |
1368 | int in_sack; | |
1369 | ||
1370 | if (!sk_can_gso(sk)) | |
1371 | goto fallback; | |
1372 | ||
1373 | /* Normally R but no L won't result in plain S */ | |
1374 | if (!dup_sack && | |
9969ca5f | 1375 | (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) |
832d11c5 IJ |
1376 | goto fallback; |
1377 | if (!skb_can_shift(skb)) | |
1378 | goto fallback; | |
1379 | /* This frame is about to be dropped (was ACKed). */ | |
1380 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1381 | goto fallback; | |
1382 | ||
1383 | /* Can only happen with delayed DSACK + discard craziness */ | |
75c119af ED |
1384 | prev = skb_rb_prev(skb); |
1385 | if (!prev) | |
832d11c5 | 1386 | goto fallback; |
832d11c5 IJ |
1387 | |
1388 | if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) | |
1389 | goto fallback; | |
1390 | ||
a643b5d4 MKL |
1391 | if (!tcp_skb_can_collapse_to(prev)) |
1392 | goto fallback; | |
1393 | ||
832d11c5 IJ |
1394 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && |
1395 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1396 | ||
1397 | if (in_sack) { | |
1398 | len = skb->len; | |
1399 | pcount = tcp_skb_pcount(skb); | |
775ffabf | 1400 | mss = tcp_skb_seglen(skb); |
832d11c5 IJ |
1401 | |
1402 | /* TODO: Fix DSACKs to not fragment already SACKed and we can | |
1403 | * drop this restriction as unnecessary | |
1404 | */ | |
775ffabf | 1405 | if (mss != tcp_skb_seglen(prev)) |
832d11c5 IJ |
1406 | goto fallback; |
1407 | } else { | |
1408 | if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) | |
1409 | goto noop; | |
1410 | /* CHECKME: This is non-MSS split case only?, this will | |
1411 | * cause skipped skbs due to advancing loop btw, original | |
1412 | * has that feature too | |
1413 | */ | |
1414 | if (tcp_skb_pcount(skb) <= 1) | |
1415 | goto noop; | |
1416 | ||
1417 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); | |
1418 | if (!in_sack) { | |
1419 | /* TODO: head merge to next could be attempted here | |
1420 | * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), | |
1421 | * though it might not be worth of the additional hassle | |
1422 | * | |
1423 | * ...we can probably just fallback to what was done | |
1424 | * previously. We could try merging non-SACKed ones | |
1425 | * as well but it probably isn't going to buy off | |
1426 | * because later SACKs might again split them, and | |
1427 | * it would make skb timestamp tracking considerably | |
1428 | * harder problem. | |
1429 | */ | |
1430 | goto fallback; | |
1431 | } | |
1432 | ||
1433 | len = end_seq - TCP_SKB_CB(skb)->seq; | |
1434 | BUG_ON(len < 0); | |
1435 | BUG_ON(len > skb->len); | |
1436 | ||
1437 | /* MSS boundaries should be honoured or else pcount will | |
1438 | * severely break even though it makes things bit trickier. | |
1439 | * Optimize common case to avoid most of the divides | |
1440 | */ | |
1441 | mss = tcp_skb_mss(skb); | |
1442 | ||
1443 | /* TODO: Fix DSACKs to not fragment already SACKed and we can | |
1444 | * drop this restriction as unnecessary | |
1445 | */ | |
775ffabf | 1446 | if (mss != tcp_skb_seglen(prev)) |
832d11c5 IJ |
1447 | goto fallback; |
1448 | ||
1449 | if (len == mss) { | |
1450 | pcount = 1; | |
1451 | } else if (len < mss) { | |
1452 | goto noop; | |
1453 | } else { | |
1454 | pcount = len / mss; | |
1455 | len = pcount * mss; | |
1456 | } | |
1457 | } | |
1458 | ||
4648dc97 NC |
1459 | /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ |
1460 | if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) | |
1461 | goto fallback; | |
1462 | ||
832d11c5 IJ |
1463 | if (!skb_shift(prev, skb, len)) |
1464 | goto fallback; | |
f3319816 | 1465 | if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack)) |
832d11c5 IJ |
1466 | goto out; |
1467 | ||
1468 | /* Hole filled allows collapsing with the next as well, this is very | |
1469 | * useful when hole on every nth skb pattern happens | |
1470 | */ | |
75c119af ED |
1471 | skb = skb_rb_next(prev); |
1472 | if (!skb) | |
832d11c5 | 1473 | goto out; |
832d11c5 | 1474 | |
f0bc52f3 | 1475 | if (!skb_can_shift(skb) || |
f0bc52f3 | 1476 | ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || |
775ffabf | 1477 | (mss != tcp_skb_seglen(skb))) |
832d11c5 IJ |
1478 | goto out; |
1479 | ||
1480 | len = skb->len; | |
1481 | if (skb_shift(prev, skb, len)) { | |
1482 | pcount += tcp_skb_pcount(skb); | |
f3319816 ED |
1483 | tcp_shifted_skb(sk, prev, skb, state, tcp_skb_pcount(skb), |
1484 | len, mss, 0); | |
832d11c5 IJ |
1485 | } |
1486 | ||
1487 | out: | |
832d11c5 IJ |
1488 | return prev; |
1489 | ||
1490 | noop: | |
1491 | return skb; | |
1492 | ||
1493 | fallback: | |
c10d9310 | 1494 | NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); |
832d11c5 IJ |
1495 | return NULL; |
1496 | } | |
1497 | ||
68f8353b IJ |
1498 | static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, |
1499 | struct tcp_sack_block *next_dup, | |
a1197f5a | 1500 | struct tcp_sacktag_state *state, |
68f8353b | 1501 | u32 start_seq, u32 end_seq, |
a2a385d6 | 1502 | bool dup_sack_in) |
68f8353b | 1503 | { |
832d11c5 IJ |
1504 | struct tcp_sock *tp = tcp_sk(sk); |
1505 | struct sk_buff *tmp; | |
1506 | ||
75c119af | 1507 | skb_rbtree_walk_from(skb) { |
68f8353b | 1508 | int in_sack = 0; |
a2a385d6 | 1509 | bool dup_sack = dup_sack_in; |
68f8353b | 1510 | |
68f8353b IJ |
1511 | /* queue is in-order => we can short-circuit the walk early */ |
1512 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) | |
1513 | break; | |
1514 | ||
00db4124 | 1515 | if (next_dup && |
68f8353b IJ |
1516 | before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { |
1517 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1518 | next_dup->start_seq, | |
1519 | next_dup->end_seq); | |
1520 | if (in_sack > 0) | |
a2a385d6 | 1521 | dup_sack = true; |
68f8353b IJ |
1522 | } |
1523 | ||
832d11c5 IJ |
1524 | /* skb reference here is a bit tricky to get right, since |
1525 | * shifting can eat and free both this skb and the next, | |
1526 | * so not even _safe variant of the loop is enough. | |
1527 | */ | |
1528 | if (in_sack <= 0) { | |
a1197f5a IJ |
1529 | tmp = tcp_shift_skb_data(sk, skb, state, |
1530 | start_seq, end_seq, dup_sack); | |
00db4124 | 1531 | if (tmp) { |
832d11c5 IJ |
1532 | if (tmp != skb) { |
1533 | skb = tmp; | |
1534 | continue; | |
1535 | } | |
1536 | ||
1537 | in_sack = 0; | |
1538 | } else { | |
1539 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1540 | start_seq, | |
1541 | end_seq); | |
1542 | } | |
1543 | } | |
1544 | ||
68f8353b IJ |
1545 | if (unlikely(in_sack < 0)) |
1546 | break; | |
1547 | ||
832d11c5 | 1548 | if (in_sack) { |
cc9a672e NC |
1549 | TCP_SKB_CB(skb)->sacked = |
1550 | tcp_sacktag_one(sk, | |
1551 | state, | |
1552 | TCP_SKB_CB(skb)->sacked, | |
1553 | TCP_SKB_CB(skb)->seq, | |
1554 | TCP_SKB_CB(skb)->end_seq, | |
1555 | dup_sack, | |
59c9af42 | 1556 | tcp_skb_pcount(skb), |
9a568de4 | 1557 | skb->skb_mstamp); |
b9f64820 | 1558 | tcp_rate_skb_delivered(sk, skb, state->rate); |
e2080072 ED |
1559 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
1560 | list_del_init(&skb->tcp_tsorted_anchor); | |
68f8353b | 1561 | |
832d11c5 IJ |
1562 | if (!before(TCP_SKB_CB(skb)->seq, |
1563 | tcp_highest_sack_seq(tp))) | |
1564 | tcp_advance_highest_sack(sk, skb); | |
1565 | } | |
68f8353b IJ |
1566 | } |
1567 | return skb; | |
1568 | } | |
1569 | ||
75c119af ED |
1570 | static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, |
1571 | struct tcp_sacktag_state *state, | |
1572 | u32 seq) | |
1573 | { | |
1574 | struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node; | |
1575 | struct sk_buff *skb; | |
75c119af ED |
1576 | |
1577 | while (*p) { | |
1578 | parent = *p; | |
1579 | skb = rb_to_skb(parent); | |
1580 | if (before(seq, TCP_SKB_CB(skb)->seq)) { | |
1581 | p = &parent->rb_left; | |
1582 | continue; | |
1583 | } | |
1584 | if (!before(seq, TCP_SKB_CB(skb)->end_seq)) { | |
1585 | p = &parent->rb_right; | |
1586 | continue; | |
1587 | } | |
75c119af ED |
1588 | return skb; |
1589 | } | |
1590 | return NULL; | |
1591 | } | |
1592 | ||
68f8353b | 1593 | static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, |
a1197f5a IJ |
1594 | struct tcp_sacktag_state *state, |
1595 | u32 skip_to_seq) | |
68f8353b | 1596 | { |
75c119af ED |
1597 | if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq)) |
1598 | return skb; | |
d152a7d8 | 1599 | |
75c119af | 1600 | return tcp_sacktag_bsearch(sk, state, skip_to_seq); |
68f8353b IJ |
1601 | } |
1602 | ||
1603 | static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, | |
1604 | struct sock *sk, | |
1605 | struct tcp_sack_block *next_dup, | |
a1197f5a IJ |
1606 | struct tcp_sacktag_state *state, |
1607 | u32 skip_to_seq) | |
68f8353b | 1608 | { |
51456b29 | 1609 | if (!next_dup) |
68f8353b IJ |
1610 | return skb; |
1611 | ||
1612 | if (before(next_dup->start_seq, skip_to_seq)) { | |
a1197f5a IJ |
1613 | skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); |
1614 | skb = tcp_sacktag_walk(skb, sk, NULL, state, | |
1615 | next_dup->start_seq, next_dup->end_seq, | |
1616 | 1); | |
68f8353b IJ |
1617 | } |
1618 | ||
1619 | return skb; | |
1620 | } | |
1621 | ||
cf533ea5 | 1622 | static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) |
68f8353b IJ |
1623 | { |
1624 | return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); | |
1625 | } | |
1626 | ||
1da177e4 | 1627 | static int |
cf533ea5 | 1628 | tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, |
196da974 | 1629 | u32 prior_snd_una, struct tcp_sacktag_state *state) |
1da177e4 LT |
1630 | { |
1631 | struct tcp_sock *tp = tcp_sk(sk); | |
cf533ea5 ED |
1632 | const unsigned char *ptr = (skb_transport_header(ack_skb) + |
1633 | TCP_SKB_CB(ack_skb)->sacked); | |
fd6dad61 | 1634 | struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); |
4389dded | 1635 | struct tcp_sack_block sp[TCP_NUM_SACKS]; |
68f8353b IJ |
1636 | struct tcp_sack_block *cache; |
1637 | struct sk_buff *skb; | |
4389dded | 1638 | int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); |
fd6dad61 | 1639 | int used_sacks; |
a2a385d6 | 1640 | bool found_dup_sack = false; |
68f8353b | 1641 | int i, j; |
fda03fbb | 1642 | int first_sack_index; |
1da177e4 | 1643 | |
196da974 | 1644 | state->flag = 0; |
737ff314 | 1645 | state->reord = tp->snd_nxt; |
a1197f5a | 1646 | |
737ff314 | 1647 | if (!tp->sacked_out) |
6859d494 | 1648 | tcp_highest_sack_reset(sk); |
1da177e4 | 1649 | |
1ed83465 | 1650 | found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, |
d06e021d | 1651 | num_sacks, prior_snd_una); |
b9f64820 | 1652 | if (found_dup_sack) { |
196da974 | 1653 | state->flag |= FLAG_DSACKING_ACK; |
b9f64820 YC |
1654 | tp->delivered++; /* A spurious retransmission is delivered */ |
1655 | } | |
6f74651a BE |
1656 | |
1657 | /* Eliminate too old ACKs, but take into | |
1658 | * account more or less fresh ones, they can | |
1659 | * contain valid SACK info. | |
1660 | */ | |
1661 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) | |
1662 | return 0; | |
1663 | ||