Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * INET An implementation of the TCP/IP protocol suite for the LINUX | |
3 | * operating system. INET is implemented using the BSD Socket | |
4 | * interface as the means of communication with the user level. | |
5 | * | |
6 | * Implementation of the Transmission Control Protocol(TCP). | |
7 | * | |
02c30a84 | 8 | * Authors: Ross Biro |
1da177e4 LT |
9 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
10 | * Mark Evans, <evansmp@uhura.aston.ac.uk> | |
11 | * Corey Minyard <wf-rch!minyard@relay.EU.net> | |
12 | * Florian La Roche, <flla@stud.uni-sb.de> | |
13 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> | |
14 | * Linus Torvalds, <torvalds@cs.helsinki.fi> | |
15 | * Alan Cox, <gw4pts@gw4pts.ampr.org> | |
16 | * Matthew Dillon, <dillon@apollo.west.oic.com> | |
17 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> | |
18 | * Jorge Cwik, <jorge@laser.satlink.net> | |
19 | */ | |
20 | ||
21 | /* | |
22 | * Changes: | |
23 | * Pedro Roque : Fast Retransmit/Recovery. | |
24 | * Two receive queues. | |
25 | * Retransmit queue handled by TCP. | |
26 | * Better retransmit timer handling. | |
27 | * New congestion avoidance. | |
28 | * Header prediction. | |
29 | * Variable renaming. | |
30 | * | |
31 | * Eric : Fast Retransmit. | |
32 | * Randy Scott : MSS option defines. | |
33 | * Eric Schenk : Fixes to slow start algorithm. | |
34 | * Eric Schenk : Yet another double ACK bug. | |
35 | * Eric Schenk : Delayed ACK bug fixes. | |
36 | * Eric Schenk : Floyd style fast retrans war avoidance. | |
37 | * David S. Miller : Don't allow zero congestion window. | |
38 | * Eric Schenk : Fix retransmitter so that it sends | |
39 | * next packet on ack of previous packet. | |
40 | * Andi Kleen : Moved open_request checking here | |
41 | * and process RSTs for open_requests. | |
42 | * Andi Kleen : Better prune_queue, and other fixes. | |
caa20d9a | 43 | * Andrey Savochkin: Fix RTT measurements in the presence of |
1da177e4 LT |
44 | * timestamps. |
45 | * Andrey Savochkin: Check sequence numbers correctly when | |
46 | * removing SACKs due to in sequence incoming | |
47 | * data segments. | |
48 | * Andi Kleen: Make sure we never ack data there is not | |
49 | * enough room for. Also make this condition | |
50 | * a fatal error if it might still happen. | |
e905a9ed | 51 | * Andi Kleen: Add tcp_measure_rcv_mss to make |
1da177e4 | 52 | * connections with MSS<min(MTU,ann. MSS) |
e905a9ed | 53 | * work without delayed acks. |
1da177e4 LT |
54 | * Andi Kleen: Process packets with PSH set in the |
55 | * fast path. | |
56 | * J Hadi Salim: ECN support | |
57 | * Andrei Gurtov, | |
58 | * Pasi Sarolahti, | |
59 | * Panu Kuhlberg: Experimental audit of TCP (re)transmission | |
60 | * engine. Lots of bugs are found. | |
61 | * Pasi Sarolahti: F-RTO for dealing with spurious RTOs | |
1da177e4 LT |
62 | */ |
63 | ||
afd46503 JP |
64 | #define pr_fmt(fmt) "TCP: " fmt |
65 | ||
1da177e4 | 66 | #include <linux/mm.h> |
5a0e3ad6 | 67 | #include <linux/slab.h> |
1da177e4 LT |
68 | #include <linux/module.h> |
69 | #include <linux/sysctl.h> | |
a0bffffc | 70 | #include <linux/kernel.h> |
ad971f61 | 71 | #include <linux/prefetch.h> |
5ffc02a1 | 72 | #include <net/dst.h> |
1da177e4 LT |
73 | #include <net/tcp.h> |
74 | #include <net/inet_common.h> | |
75 | #include <linux/ipsec.h> | |
76 | #include <asm/unaligned.h> | |
e1c8a607 | 77 | #include <linux/errqueue.h> |
1da177e4 | 78 | |
ab32ea5d BH |
79 | int sysctl_tcp_timestamps __read_mostly = 1; |
80 | int sysctl_tcp_window_scaling __read_mostly = 1; | |
81 | int sysctl_tcp_sack __read_mostly = 1; | |
82 | int sysctl_tcp_fack __read_mostly = 1; | |
83 | int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; | |
dca145ff | 84 | int sysctl_tcp_max_reordering __read_mostly = 300; |
4bc2f18b | 85 | EXPORT_SYMBOL(sysctl_tcp_reordering); |
ab32ea5d BH |
86 | int sysctl_tcp_dsack __read_mostly = 1; |
87 | int sysctl_tcp_app_win __read_mostly = 31; | |
b49960a0 | 88 | int sysctl_tcp_adv_win_scale __read_mostly = 1; |
4bc2f18b | 89 | EXPORT_SYMBOL(sysctl_tcp_adv_win_scale); |
1da177e4 | 90 | |
282f23c6 ED |
91 | /* rfc5961 challenge ack rate limiting */ |
92 | int sysctl_tcp_challenge_ack_limit = 100; | |
93 | ||
ab32ea5d BH |
94 | int sysctl_tcp_stdurg __read_mostly; |
95 | int sysctl_tcp_rfc1337 __read_mostly; | |
96 | int sysctl_tcp_max_orphans __read_mostly = NR_FILE; | |
c96fd3d4 | 97 | int sysctl_tcp_frto __read_mostly = 2; |
1da177e4 | 98 | |
7e380175 AP |
99 | int sysctl_tcp_thin_dupack __read_mostly; |
100 | ||
ab32ea5d | 101 | int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; |
6ba8a3b1 | 102 | int sysctl_tcp_early_retrans __read_mostly = 3; |
032ee423 | 103 | int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2; |
1da177e4 | 104 | |
1da177e4 LT |
105 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ |
106 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ | |
107 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ | |
108 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ | |
109 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ | |
110 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ | |
111 | #define FLAG_ECE 0x40 /* ECE in this ACK */ | |
1da177e4 | 112 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ |
e33099f9 | 113 | #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */ |
2e605294 | 114 | #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ |
564262c1 | 115 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ |
cadbd031 | 116 | #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ |
12fb3dd9 | 117 | #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */ |
1da177e4 LT |
118 | |
119 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) | |
120 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) | |
121 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) | |
122 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) | |
123 | ||
1da177e4 | 124 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) |
bdf1ee5d | 125 | #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) |
1da177e4 | 126 | |
e905a9ed | 127 | /* Adapt the MSS value used to make delayed ack decision to the |
1da177e4 | 128 | * real world. |
e905a9ed | 129 | */ |
056834d9 | 130 | static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 131 | { |
463c84b9 | 132 | struct inet_connection_sock *icsk = inet_csk(sk); |
e905a9ed | 133 | const unsigned int lss = icsk->icsk_ack.last_seg_size; |
463c84b9 | 134 | unsigned int len; |
1da177e4 | 135 | |
e905a9ed | 136 | icsk->icsk_ack.last_seg_size = 0; |
1da177e4 LT |
137 | |
138 | /* skb->len may jitter because of SACKs, even if peer | |
139 | * sends good full-sized frames. | |
140 | */ | |
056834d9 | 141 | len = skb_shinfo(skb)->gso_size ? : skb->len; |
463c84b9 ACM |
142 | if (len >= icsk->icsk_ack.rcv_mss) { |
143 | icsk->icsk_ack.rcv_mss = len; | |
1da177e4 LT |
144 | } else { |
145 | /* Otherwise, we make more careful check taking into account, | |
146 | * that SACKs block is variable. | |
147 | * | |
148 | * "len" is invariant segment length, including TCP header. | |
149 | */ | |
9c70220b | 150 | len += skb->data - skb_transport_header(skb); |
bee7ca9e | 151 | if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || |
1da177e4 LT |
152 | /* If PSH is not set, packet should be |
153 | * full sized, provided peer TCP is not badly broken. | |
154 | * This observation (if it is correct 8)) allows | |
155 | * to handle super-low mtu links fairly. | |
156 | */ | |
157 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && | |
aa8223c7 | 158 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { |
1da177e4 LT |
159 | /* Subtract also invariant (if peer is RFC compliant), |
160 | * tcp header plus fixed timestamp option length. | |
161 | * Resulting "len" is MSS free of SACK jitter. | |
162 | */ | |
463c84b9 ACM |
163 | len -= tcp_sk(sk)->tcp_header_len; |
164 | icsk->icsk_ack.last_seg_size = len; | |
1da177e4 | 165 | if (len == lss) { |
463c84b9 | 166 | icsk->icsk_ack.rcv_mss = len; |
1da177e4 LT |
167 | return; |
168 | } | |
169 | } | |
1ef9696c AK |
170 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) |
171 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; | |
463c84b9 | 172 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
1da177e4 LT |
173 | } |
174 | } | |
175 | ||
463c84b9 | 176 | static void tcp_incr_quickack(struct sock *sk) |
1da177e4 | 177 | { |
463c84b9 | 178 | struct inet_connection_sock *icsk = inet_csk(sk); |
95c96174 | 179 | unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); |
1da177e4 | 180 | |
056834d9 IJ |
181 | if (quickacks == 0) |
182 | quickacks = 2; | |
463c84b9 ACM |
183 | if (quickacks > icsk->icsk_ack.quick) |
184 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); | |
1da177e4 LT |
185 | } |
186 | ||
1b9f4092 | 187 | static void tcp_enter_quickack_mode(struct sock *sk) |
1da177e4 | 188 | { |
463c84b9 ACM |
189 | struct inet_connection_sock *icsk = inet_csk(sk); |
190 | tcp_incr_quickack(sk); | |
191 | icsk->icsk_ack.pingpong = 0; | |
192 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 LT |
193 | } |
194 | ||
195 | /* Send ACKs quickly, if "quick" count is not exhausted | |
196 | * and the session is not interactive. | |
197 | */ | |
198 | ||
a2a385d6 | 199 | static inline bool tcp_in_quickack_mode(const struct sock *sk) |
1da177e4 | 200 | { |
463c84b9 | 201 | const struct inet_connection_sock *icsk = inet_csk(sk); |
a2a385d6 | 202 | |
463c84b9 | 203 | return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; |
1da177e4 LT |
204 | } |
205 | ||
735d3831 | 206 | static void tcp_ecn_queue_cwr(struct tcp_sock *tp) |
bdf1ee5d | 207 | { |
056834d9 | 208 | if (tp->ecn_flags & TCP_ECN_OK) |
bdf1ee5d IJ |
209 | tp->ecn_flags |= TCP_ECN_QUEUE_CWR; |
210 | } | |
211 | ||
735d3831 | 212 | static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb) |
bdf1ee5d IJ |
213 | { |
214 | if (tcp_hdr(skb)->cwr) | |
215 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
216 | } | |
217 | ||
735d3831 | 218 | static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp) |
bdf1ee5d IJ |
219 | { |
220 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
221 | } | |
222 | ||
735d3831 | 223 | static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) |
bdf1ee5d | 224 | { |
b82d1bb4 | 225 | switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { |
7a269ffa | 226 | case INET_ECN_NOT_ECT: |
bdf1ee5d | 227 | /* Funny extension: if ECT is not set on a segment, |
7a269ffa ED |
228 | * and we already seen ECT on a previous segment, |
229 | * it is probably a retransmit. | |
230 | */ | |
231 | if (tp->ecn_flags & TCP_ECN_SEEN) | |
bdf1ee5d | 232 | tcp_enter_quickack_mode((struct sock *)tp); |
7a269ffa ED |
233 | break; |
234 | case INET_ECN_CE: | |
9890092e FW |
235 | if (tcp_ca_needs_ecn((struct sock *)tp)) |
236 | tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE); | |
237 | ||
aae06bf5 ED |
238 | if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { |
239 | /* Better not delay acks, sender can have a very low cwnd */ | |
240 | tcp_enter_quickack_mode((struct sock *)tp); | |
241 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; | |
242 | } | |
9890092e FW |
243 | tp->ecn_flags |= TCP_ECN_SEEN; |
244 | break; | |
7a269ffa | 245 | default: |
9890092e FW |
246 | if (tcp_ca_needs_ecn((struct sock *)tp)) |
247 | tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE); | |
7a269ffa | 248 | tp->ecn_flags |= TCP_ECN_SEEN; |
9890092e | 249 | break; |
bdf1ee5d IJ |
250 | } |
251 | } | |
252 | ||
735d3831 FW |
253 | static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) |
254 | { | |
255 | if (tp->ecn_flags & TCP_ECN_OK) | |
256 | __tcp_ecn_check_ce(tp, skb); | |
257 | } | |
258 | ||
259 | static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) | |
bdf1ee5d | 260 | { |
056834d9 | 261 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) |
bdf1ee5d IJ |
262 | tp->ecn_flags &= ~TCP_ECN_OK; |
263 | } | |
264 | ||
735d3831 | 265 | static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) |
bdf1ee5d | 266 | { |
056834d9 | 267 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) |
bdf1ee5d IJ |
268 | tp->ecn_flags &= ~TCP_ECN_OK; |
269 | } | |
270 | ||
735d3831 | 271 | static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) |
bdf1ee5d | 272 | { |
056834d9 | 273 | if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) |
a2a385d6 ED |
274 | return true; |
275 | return false; | |
bdf1ee5d IJ |
276 | } |
277 | ||
1da177e4 LT |
278 | /* Buffer size and advertised window tuning. |
279 | * | |
280 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. | |
281 | */ | |
282 | ||
6ae70532 | 283 | static void tcp_sndbuf_expand(struct sock *sk) |
1da177e4 | 284 | { |
6ae70532 ED |
285 | const struct tcp_sock *tp = tcp_sk(sk); |
286 | int sndmem, per_mss; | |
287 | u32 nr_segs; | |
288 | ||
289 | /* Worst case is non GSO/TSO : each frame consumes one skb | |
290 | * and skb->head is kmalloced using power of two area of memory | |
291 | */ | |
292 | per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + | |
293 | MAX_TCP_HEADER + | |
294 | SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); | |
295 | ||
296 | per_mss = roundup_pow_of_two(per_mss) + | |
297 | SKB_DATA_ALIGN(sizeof(struct sk_buff)); | |
298 | ||
299 | nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd); | |
300 | nr_segs = max_t(u32, nr_segs, tp->reordering + 1); | |
301 | ||
302 | /* Fast Recovery (RFC 5681 3.2) : | |
303 | * Cubic needs 1.7 factor, rounded to 2 to include | |
304 | * extra cushion (application might react slowly to POLLOUT) | |
305 | */ | |
306 | sndmem = 2 * nr_segs * per_mss; | |
1da177e4 | 307 | |
06a59ecb ED |
308 | if (sk->sk_sndbuf < sndmem) |
309 | sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); | |
1da177e4 LT |
310 | } |
311 | ||
312 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) | |
313 | * | |
314 | * All tcp_full_space() is split to two parts: "network" buffer, allocated | |
315 | * forward and advertised in receiver window (tp->rcv_wnd) and | |
316 | * "application buffer", required to isolate scheduling/application | |
317 | * latencies from network. | |
318 | * window_clamp is maximal advertised window. It can be less than | |
319 | * tcp_full_space(), in this case tcp_full_space() - window_clamp | |
320 | * is reserved for "application" buffer. The less window_clamp is | |
321 | * the smoother our behaviour from viewpoint of network, but the lower | |
322 | * throughput and the higher sensitivity of the connection to losses. 8) | |
323 | * | |
324 | * rcv_ssthresh is more strict window_clamp used at "slow start" | |
325 | * phase to predict further behaviour of this connection. | |
326 | * It is used for two goals: | |
327 | * - to enforce header prediction at sender, even when application | |
328 | * requires some significant "application buffer". It is check #1. | |
329 | * - to prevent pruning of receive queue because of misprediction | |
330 | * of receiver window. Check #2. | |
331 | * | |
332 | * The scheme does not work when sender sends good segments opening | |
caa20d9a | 333 | * window and then starts to feed us spaghetti. But it should work |
1da177e4 LT |
334 | * in common situations. Otherwise, we have to rely on queue collapsing. |
335 | */ | |
336 | ||
337 | /* Slow part of check#2. */ | |
9e412ba7 | 338 | static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 339 | { |
9e412ba7 | 340 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 341 | /* Optimize this! */ |
dfd4f0ae ED |
342 | int truesize = tcp_win_from_space(skb->truesize) >> 1; |
343 | int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; | |
1da177e4 LT |
344 | |
345 | while (tp->rcv_ssthresh <= window) { | |
346 | if (truesize <= skb->len) | |
463c84b9 | 347 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; |
1da177e4 LT |
348 | |
349 | truesize >>= 1; | |
350 | window >>= 1; | |
351 | } | |
352 | return 0; | |
353 | } | |
354 | ||
cf533ea5 | 355 | static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 356 | { |
9e412ba7 IJ |
357 | struct tcp_sock *tp = tcp_sk(sk); |
358 | ||
1da177e4 LT |
359 | /* Check #1 */ |
360 | if (tp->rcv_ssthresh < tp->window_clamp && | |
361 | (int)tp->rcv_ssthresh < tcp_space(sk) && | |
b8da51eb | 362 | !tcp_under_memory_pressure(sk)) { |
1da177e4 LT |
363 | int incr; |
364 | ||
365 | /* Check #2. Increase window, if skb with such overhead | |
366 | * will fit to rcvbuf in future. | |
367 | */ | |
368 | if (tcp_win_from_space(skb->truesize) <= skb->len) | |
056834d9 | 369 | incr = 2 * tp->advmss; |
1da177e4 | 370 | else |
9e412ba7 | 371 | incr = __tcp_grow_window(sk, skb); |
1da177e4 LT |
372 | |
373 | if (incr) { | |
4d846f02 | 374 | incr = max_t(int, incr, 2 * skb->len); |
056834d9 IJ |
375 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, |
376 | tp->window_clamp); | |
463c84b9 | 377 | inet_csk(sk)->icsk_ack.quick |= 1; |
1da177e4 LT |
378 | } |
379 | } | |
380 | } | |
381 | ||
382 | /* 3. Tuning rcvbuf, when connection enters established state. */ | |
1da177e4 LT |
383 | static void tcp_fixup_rcvbuf(struct sock *sk) |
384 | { | |
e9266a02 | 385 | u32 mss = tcp_sk(sk)->advmss; |
e9266a02 | 386 | int rcvmem; |
1da177e4 | 387 | |
85f16525 YC |
388 | rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) * |
389 | tcp_default_init_rwnd(mss); | |
e9266a02 | 390 | |
b0983d3c ED |
391 | /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency |
392 | * Allow enough cushion so that sender is not limited by our window | |
393 | */ | |
394 | if (sysctl_tcp_moderate_rcvbuf) | |
395 | rcvmem <<= 2; | |
396 | ||
e9266a02 ED |
397 | if (sk->sk_rcvbuf < rcvmem) |
398 | sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]); | |
1da177e4 LT |
399 | } |
400 | ||
caa20d9a | 401 | /* 4. Try to fixup all. It is made immediately after connection enters |
1da177e4 LT |
402 | * established state. |
403 | */ | |
10467163 | 404 | void tcp_init_buffer_space(struct sock *sk) |
1da177e4 LT |
405 | { |
406 | struct tcp_sock *tp = tcp_sk(sk); | |
407 | int maxwin; | |
408 | ||
409 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) | |
410 | tcp_fixup_rcvbuf(sk); | |
411 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) | |
6ae70532 | 412 | tcp_sndbuf_expand(sk); |
1da177e4 LT |
413 | |
414 | tp->rcvq_space.space = tp->rcv_wnd; | |
b0983d3c ED |
415 | tp->rcvq_space.time = tcp_time_stamp; |
416 | tp->rcvq_space.seq = tp->copied_seq; | |
1da177e4 LT |
417 | |
418 | maxwin = tcp_full_space(sk); | |
419 | ||
420 | if (tp->window_clamp >= maxwin) { | |
421 | tp->window_clamp = maxwin; | |
422 | ||
423 | if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) | |
424 | tp->window_clamp = max(maxwin - | |
425 | (maxwin >> sysctl_tcp_app_win), | |
426 | 4 * tp->advmss); | |
427 | } | |
428 | ||
429 | /* Force reservation of one segment. */ | |
430 | if (sysctl_tcp_app_win && | |
431 | tp->window_clamp > 2 * tp->advmss && | |
432 | tp->window_clamp + tp->advmss > maxwin) | |
433 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); | |
434 | ||
435 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); | |
436 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
437 | } | |
438 | ||
1da177e4 | 439 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ |
9e412ba7 | 440 | static void tcp_clamp_window(struct sock *sk) |
1da177e4 | 441 | { |
9e412ba7 | 442 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 443 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 444 | |
6687e988 | 445 | icsk->icsk_ack.quick = 0; |
1da177e4 | 446 | |
326f36e9 JH |
447 | if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && |
448 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && | |
b8da51eb | 449 | !tcp_under_memory_pressure(sk) && |
180d8cd9 | 450 | sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { |
326f36e9 JH |
451 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), |
452 | sysctl_tcp_rmem[2]); | |
1da177e4 | 453 | } |
326f36e9 | 454 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) |
056834d9 | 455 | tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); |
1da177e4 LT |
456 | } |
457 | ||
40efc6fa SH |
458 | /* Initialize RCV_MSS value. |
459 | * RCV_MSS is an our guess about MSS used by the peer. | |
460 | * We haven't any direct information about the MSS. | |
461 | * It's better to underestimate the RCV_MSS rather than overestimate. | |
462 | * Overestimations make us ACKing less frequently than needed. | |
463 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). | |
464 | */ | |
465 | void tcp_initialize_rcv_mss(struct sock *sk) | |
466 | { | |
cf533ea5 | 467 | const struct tcp_sock *tp = tcp_sk(sk); |
40efc6fa SH |
468 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); |
469 | ||
056834d9 | 470 | hint = min(hint, tp->rcv_wnd / 2); |
bee7ca9e | 471 | hint = min(hint, TCP_MSS_DEFAULT); |
40efc6fa SH |
472 | hint = max(hint, TCP_MIN_MSS); |
473 | ||
474 | inet_csk(sk)->icsk_ack.rcv_mss = hint; | |
475 | } | |
4bc2f18b | 476 | EXPORT_SYMBOL(tcp_initialize_rcv_mss); |
40efc6fa | 477 | |
1da177e4 LT |
478 | /* Receiver "autotuning" code. |
479 | * | |
480 | * The algorithm for RTT estimation w/o timestamps is based on | |
481 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. | |
631dd1a8 | 482 | * <http://public.lanl.gov/radiant/pubs.html#DRS> |
1da177e4 LT |
483 | * |
484 | * More detail on this code can be found at | |
631dd1a8 | 485 | * <http://staff.psc.edu/jheffner/>, |
1da177e4 LT |
486 | * though this reference is out of date. A new paper |
487 | * is pending. | |
488 | */ | |
489 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) | |
490 | { | |
491 | u32 new_sample = tp->rcv_rtt_est.rtt; | |
492 | long m = sample; | |
493 | ||
494 | if (m == 0) | |
495 | m = 1; | |
496 | ||
497 | if (new_sample != 0) { | |
498 | /* If we sample in larger samples in the non-timestamp | |
499 | * case, we could grossly overestimate the RTT especially | |
500 | * with chatty applications or bulk transfer apps which | |
501 | * are stalled on filesystem I/O. | |
502 | * | |
503 | * Also, since we are only going for a minimum in the | |
31f34269 | 504 | * non-timestamp case, we do not smooth things out |
caa20d9a | 505 | * else with timestamps disabled convergence takes too |
1da177e4 LT |
506 | * long. |
507 | */ | |
508 | if (!win_dep) { | |
509 | m -= (new_sample >> 3); | |
510 | new_sample += m; | |
18a223e0 NC |
511 | } else { |
512 | m <<= 3; | |
513 | if (m < new_sample) | |
514 | new_sample = m; | |
515 | } | |
1da177e4 | 516 | } else { |
caa20d9a | 517 | /* No previous measure. */ |
1da177e4 LT |
518 | new_sample = m << 3; |
519 | } | |
520 | ||
521 | if (tp->rcv_rtt_est.rtt != new_sample) | |
522 | tp->rcv_rtt_est.rtt = new_sample; | |
523 | } | |
524 | ||
525 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) | |
526 | { | |
527 | if (tp->rcv_rtt_est.time == 0) | |
528 | goto new_measure; | |
529 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) | |
530 | return; | |
651913ce | 531 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1); |
1da177e4 LT |
532 | |
533 | new_measure: | |
534 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; | |
535 | tp->rcv_rtt_est.time = tcp_time_stamp; | |
536 | } | |
537 | ||
056834d9 IJ |
538 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, |
539 | const struct sk_buff *skb) | |
1da177e4 | 540 | { |
463c84b9 | 541 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
542 | if (tp->rx_opt.rcv_tsecr && |
543 | (TCP_SKB_CB(skb)->end_seq - | |
463c84b9 | 544 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) |
1da177e4 LT |
545 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); |
546 | } | |
547 | ||
548 | /* | |
549 | * This function should be called every time data is copied to user space. | |
550 | * It calculates the appropriate TCP receive buffer space. | |
551 | */ | |
552 | void tcp_rcv_space_adjust(struct sock *sk) | |
553 | { | |
554 | struct tcp_sock *tp = tcp_sk(sk); | |
555 | int time; | |
b0983d3c | 556 | int copied; |
e905a9ed | 557 | |
1da177e4 | 558 | time = tcp_time_stamp - tp->rcvq_space.time; |
056834d9 | 559 | if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) |
1da177e4 | 560 | return; |
e905a9ed | 561 | |
b0983d3c ED |
562 | /* Number of bytes copied to user in last RTT */ |
563 | copied = tp->copied_seq - tp->rcvq_space.seq; | |
564 | if (copied <= tp->rcvq_space.space) | |
565 | goto new_measure; | |
566 | ||
567 | /* A bit of theory : | |
568 | * copied = bytes received in previous RTT, our base window | |
569 | * To cope with packet losses, we need a 2x factor | |
570 | * To cope with slow start, and sender growing its cwin by 100 % | |
571 | * every RTT, we need a 4x factor, because the ACK we are sending | |
572 | * now is for the next RTT, not the current one : | |
573 | * <prev RTT . ><current RTT .. ><next RTT .... > | |
574 | */ | |
575 | ||
576 | if (sysctl_tcp_moderate_rcvbuf && | |
577 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { | |
578 | int rcvwin, rcvmem, rcvbuf; | |
1da177e4 | 579 | |
b0983d3c ED |
580 | /* minimal window to cope with packet losses, assuming |
581 | * steady state. Add some cushion because of small variations. | |
582 | */ | |
583 | rcvwin = (copied << 1) + 16 * tp->advmss; | |
1da177e4 | 584 | |
b0983d3c ED |
585 | /* If rate increased by 25%, |
586 | * assume slow start, rcvwin = 3 * copied | |
587 | * If rate increased by 50%, | |
588 | * assume sender can use 2x growth, rcvwin = 4 * copied | |
589 | */ | |
590 | if (copied >= | |
591 | tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) { | |
592 | if (copied >= | |
593 | tp->rcvq_space.space + (tp->rcvq_space.space >> 1)) | |
594 | rcvwin <<= 1; | |
595 | else | |
596 | rcvwin += (rcvwin >> 1); | |
597 | } | |
1da177e4 | 598 | |
b0983d3c ED |
599 | rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER); |
600 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
601 | rcvmem += 128; | |
1da177e4 | 602 | |
b0983d3c ED |
603 | rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]); |
604 | if (rcvbuf > sk->sk_rcvbuf) { | |
605 | sk->sk_rcvbuf = rcvbuf; | |
1da177e4 | 606 | |
b0983d3c ED |
607 | /* Make the window clamp follow along. */ |
608 | tp->window_clamp = rcvwin; | |
1da177e4 LT |
609 | } |
610 | } | |
b0983d3c | 611 | tp->rcvq_space.space = copied; |
e905a9ed | 612 | |
1da177e4 LT |
613 | new_measure: |
614 | tp->rcvq_space.seq = tp->copied_seq; | |
615 | tp->rcvq_space.time = tcp_time_stamp; | |
616 | } | |
617 | ||
618 | /* There is something which you must keep in mind when you analyze the | |
619 | * behavior of the tp->ato delayed ack timeout interval. When a | |
620 | * connection starts up, we want to ack as quickly as possible. The | |
621 | * problem is that "good" TCP's do slow start at the beginning of data | |
622 | * transmission. The means that until we send the first few ACK's the | |
623 | * sender will sit on his end and only queue most of his data, because | |
624 | * he can only send snd_cwnd unacked packets at any given time. For | |
625 | * each ACK we send, he increments snd_cwnd and transmits more of his | |
626 | * queue. -DaveM | |
627 | */ | |
9e412ba7 | 628 | static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 629 | { |
9e412ba7 | 630 | struct tcp_sock *tp = tcp_sk(sk); |
463c84b9 | 631 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
632 | u32 now; |
633 | ||
463c84b9 | 634 | inet_csk_schedule_ack(sk); |
1da177e4 | 635 | |
463c84b9 | 636 | tcp_measure_rcv_mss(sk, skb); |
1da177e4 LT |
637 | |
638 | tcp_rcv_rtt_measure(tp); | |
e905a9ed | 639 | |
1da177e4 LT |
640 | now = tcp_time_stamp; |
641 | ||
463c84b9 | 642 | if (!icsk->icsk_ack.ato) { |
1da177e4 LT |
643 | /* The _first_ data packet received, initialize |
644 | * delayed ACK engine. | |
645 | */ | |
463c84b9 ACM |
646 | tcp_incr_quickack(sk); |
647 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 | 648 | } else { |
463c84b9 | 649 | int m = now - icsk->icsk_ack.lrcvtime; |
1da177e4 | 650 | |
056834d9 | 651 | if (m <= TCP_ATO_MIN / 2) { |
1da177e4 | 652 | /* The fastest case is the first. */ |
463c84b9 ACM |
653 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; |
654 | } else if (m < icsk->icsk_ack.ato) { | |
655 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; | |
656 | if (icsk->icsk_ack.ato > icsk->icsk_rto) | |
657 | icsk->icsk_ack.ato = icsk->icsk_rto; | |
658 | } else if (m > icsk->icsk_rto) { | |
caa20d9a | 659 | /* Too long gap. Apparently sender failed to |
1da177e4 LT |
660 | * restart window, so that we send ACKs quickly. |
661 | */ | |
463c84b9 | 662 | tcp_incr_quickack(sk); |
3ab224be | 663 | sk_mem_reclaim(sk); |
1da177e4 LT |
664 | } |
665 | } | |
463c84b9 | 666 | icsk->icsk_ack.lrcvtime = now; |
1da177e4 | 667 | |
735d3831 | 668 | tcp_ecn_check_ce(tp, skb); |
1da177e4 LT |
669 | |
670 | if (skb->len >= 128) | |
9e412ba7 | 671 | tcp_grow_window(sk, skb); |
1da177e4 LT |
672 | } |
673 | ||
1da177e4 LT |
674 | /* Called to compute a smoothed rtt estimate. The data fed to this |
675 | * routine either comes from timestamps, or from segments that were | |
676 | * known _not_ to have been retransmitted [see Karn/Partridge | |
677 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 | |
678 | * piece by Van Jacobson. | |
679 | * NOTE: the next three routines used to be one big routine. | |
680 | * To save cycles in the RFC 1323 implementation it was better to break | |
681 | * it up into three procedures. -- erics | |
682 | */ | |
740b0f18 | 683 | static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) |
1da177e4 | 684 | { |
6687e988 | 685 | struct tcp_sock *tp = tcp_sk(sk); |
740b0f18 ED |
686 | long m = mrtt_us; /* RTT */ |
687 | u32 srtt = tp->srtt_us; | |
1da177e4 | 688 | |
1da177e4 LT |
689 | /* The following amusing code comes from Jacobson's |
690 | * article in SIGCOMM '88. Note that rtt and mdev | |
691 | * are scaled versions of rtt and mean deviation. | |
e905a9ed | 692 | * This is designed to be as fast as possible |
1da177e4 LT |
693 | * m stands for "measurement". |
694 | * | |
695 | * On a 1990 paper the rto value is changed to: | |
696 | * RTO = rtt + 4 * mdev | |
697 | * | |
698 | * Funny. This algorithm seems to be very broken. | |
699 | * These formulae increase RTO, when it should be decreased, increase | |
31f34269 | 700 | * too slowly, when it should be increased quickly, decrease too quickly |
1da177e4 LT |
701 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely |
702 | * does not matter how to _calculate_ it. Seems, it was trap | |
703 | * that VJ failed to avoid. 8) | |
704 | */ | |
4a5ab4e2 ED |
705 | if (srtt != 0) { |
706 | m -= (srtt >> 3); /* m is now error in rtt est */ | |
707 | srtt += m; /* rtt = 7/8 rtt + 1/8 new */ | |
1da177e4 LT |
708 | if (m < 0) { |
709 | m = -m; /* m is now abs(error) */ | |
740b0f18 | 710 | m -= (tp->mdev_us >> 2); /* similar update on mdev */ |
1da177e4 LT |
711 | /* This is similar to one of Eifel findings. |
712 | * Eifel blocks mdev updates when rtt decreases. | |
713 | * This solution is a bit different: we use finer gain | |
714 | * for mdev in this case (alpha*beta). | |
715 | * Like Eifel it also prevents growth of rto, | |
716 | * but also it limits too fast rto decreases, | |
717 | * happening in pure Eifel. | |
718 | */ | |
719 | if (m > 0) | |
720 | m >>= 3; | |
721 | } else { | |
740b0f18 | 722 | m -= (tp->mdev_us >> 2); /* similar update on mdev */ |
1da177e4 | 723 | } |
740b0f18 ED |
724 | tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ |
725 | if (tp->mdev_us > tp->mdev_max_us) { | |
726 | tp->mdev_max_us = tp->mdev_us; | |
727 | if (tp->mdev_max_us > tp->rttvar_us) | |
728 | tp->rttvar_us = tp->mdev_max_us; | |
1da177e4 LT |
729 | } |
730 | if (after(tp->snd_una, tp->rtt_seq)) { | |
740b0f18 ED |
731 | if (tp->mdev_max_us < tp->rttvar_us) |
732 | tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; | |
1da177e4 | 733 | tp->rtt_seq = tp->snd_nxt; |
740b0f18 | 734 | tp->mdev_max_us = tcp_rto_min_us(sk); |
1da177e4 LT |
735 | } |
736 | } else { | |
737 | /* no previous measure. */ | |
4a5ab4e2 | 738 | srtt = m << 3; /* take the measured time to be rtt */ |
740b0f18 ED |
739 | tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ |
740 | tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); | |
741 | tp->mdev_max_us = tp->rttvar_us; | |
1da177e4 LT |
742 | tp->rtt_seq = tp->snd_nxt; |
743 | } | |
740b0f18 | 744 | tp->srtt_us = max(1U, srtt); |
1da177e4 LT |
745 | } |
746 | ||
95bd09eb ED |
747 | /* Set the sk_pacing_rate to allow proper sizing of TSO packets. |
748 | * Note: TCP stack does not yet implement pacing. | |
749 | * FQ packet scheduler can be used to implement cheap but effective | |
750 | * TCP pacing, to smooth the burst on large writes when packets | |
751 | * in flight is significantly lower than cwnd (or rwin) | |
752 | */ | |
753 | static void tcp_update_pacing_rate(struct sock *sk) | |
754 | { | |
755 | const struct tcp_sock *tp = tcp_sk(sk); | |
756 | u64 rate; | |
757 | ||
758 | /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ | |
740b0f18 | 759 | rate = (u64)tp->mss_cache * 2 * (USEC_PER_SEC << 3); |
95bd09eb ED |
760 | |
761 | rate *= max(tp->snd_cwnd, tp->packets_out); | |
762 | ||
740b0f18 ED |
763 | if (likely(tp->srtt_us)) |
764 | do_div(rate, tp->srtt_us); | |
95bd09eb | 765 | |
ba537427 ED |
766 | /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate |
767 | * without any lock. We want to make sure compiler wont store | |
768 | * intermediate values in this location. | |
769 | */ | |
770 | ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate, | |
771 | sk->sk_max_pacing_rate); | |
95bd09eb ED |
772 | } |
773 | ||
1da177e4 LT |
774 | /* Calculate rto without backoff. This is the second half of Van Jacobson's |
775 | * routine referred to above. | |
776 | */ | |
f7e56a76 | 777 | static void tcp_set_rto(struct sock *sk) |
1da177e4 | 778 | { |
463c84b9 | 779 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
780 | /* Old crap is replaced with new one. 8) |
781 | * | |
782 | * More seriously: | |
783 | * 1. If rtt variance happened to be less 50msec, it is hallucination. | |
784 | * It cannot be less due to utterly erratic ACK generation made | |
785 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ | |
786 | * to do with delayed acks, because at cwnd>2 true delack timeout | |
787 | * is invisible. Actually, Linux-2.4 also generates erratic | |
caa20d9a | 788 | * ACKs in some circumstances. |
1da177e4 | 789 | */ |
f1ecd5d9 | 790 | inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); |
1da177e4 LT |
791 | |
792 | /* 2. Fixups made earlier cannot be right. | |
793 | * If we do not estimate RTO correctly without them, | |
794 | * all the algo is pure shit and should be replaced | |
caa20d9a | 795 | * with correct one. It is exactly, which we pretend to do. |
1da177e4 | 796 | */ |
1da177e4 | 797 | |
ee6aac59 IJ |
798 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo |
799 | * guarantees that rto is higher. | |
800 | */ | |
f1ecd5d9 | 801 | tcp_bound_rto(sk); |
1da177e4 LT |
802 | } |
803 | ||
cf533ea5 | 804 | __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) |
1da177e4 LT |
805 | { |
806 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); | |
807 | ||
22b71c8f | 808 | if (!cwnd) |
442b9635 | 809 | cwnd = TCP_INIT_CWND; |
1da177e4 LT |
810 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); |
811 | } | |
812 | ||
e60402d0 IJ |
813 | /* |
814 | * Packet counting of FACK is based on in-order assumptions, therefore TCP | |
815 | * disables it when reordering is detected | |
816 | */ | |
4aabd8ef | 817 | void tcp_disable_fack(struct tcp_sock *tp) |
e60402d0 | 818 | { |
85cc391c IJ |
819 | /* RFC3517 uses different metric in lost marker => reset on change */ |
820 | if (tcp_is_fack(tp)) | |
821 | tp->lost_skb_hint = NULL; | |
ab56222a | 822 | tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED; |
e60402d0 IJ |
823 | } |
824 | ||
564262c1 | 825 | /* Take a notice that peer is sending D-SACKs */ |
e60402d0 IJ |
826 | static void tcp_dsack_seen(struct tcp_sock *tp) |
827 | { | |
ab56222a | 828 | tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; |
e60402d0 IJ |
829 | } |
830 | ||
6687e988 ACM |
831 | static void tcp_update_reordering(struct sock *sk, const int metric, |
832 | const int ts) | |
1da177e4 | 833 | { |
6687e988 | 834 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 835 | if (metric > tp->reordering) { |
40b215e5 PE |
836 | int mib_idx; |
837 | ||
dca145ff | 838 | tp->reordering = min(sysctl_tcp_max_reordering, metric); |
1da177e4 LT |
839 | |
840 | /* This exciting event is worth to be remembered. 8) */ | |
841 | if (ts) | |
40b215e5 | 842 | mib_idx = LINUX_MIB_TCPTSREORDER; |
e60402d0 | 843 | else if (tcp_is_reno(tp)) |
40b215e5 | 844 | mib_idx = LINUX_MIB_TCPRENOREORDER; |
e60402d0 | 845 | else if (tcp_is_fack(tp)) |
40b215e5 | 846 | mib_idx = LINUX_MIB_TCPFACKREORDER; |
1da177e4 | 847 | else |
40b215e5 PE |
848 | mib_idx = LINUX_MIB_TCPSACKREORDER; |
849 | ||
de0744af | 850 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
1da177e4 | 851 | #if FASTRETRANS_DEBUG > 1 |
91df42be JP |
852 | pr_debug("Disorder%d %d %u f%u s%u rr%d\n", |
853 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, | |
854 | tp->reordering, | |
855 | tp->fackets_out, | |
856 | tp->sacked_out, | |
857 | tp->undo_marker ? tp->undo_retrans : 0); | |
1da177e4 | 858 | #endif |
e60402d0 | 859 | tcp_disable_fack(tp); |
1da177e4 | 860 | } |
eed530b6 YC |
861 | |
862 | if (metric > 0) | |
863 | tcp_disable_early_retrans(tp); | |
1da177e4 LT |
864 | } |
865 | ||
006f582c | 866 | /* This must be called before lost_out is incremented */ |
c8c213f2 IJ |
867 | static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) |
868 | { | |
51456b29 | 869 | if (!tp->retransmit_skb_hint || |
c8c213f2 IJ |
870 | before(TCP_SKB_CB(skb)->seq, |
871 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) | |
006f582c IJ |
872 | tp->retransmit_skb_hint = skb; |
873 | ||
874 | if (!tp->lost_out || | |
875 | after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) | |
876 | tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; | |
c8c213f2 IJ |
877 | } |
878 | ||
41ea36e3 IJ |
879 | static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) |
880 | { | |
881 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
882 | tcp_verify_retransmit_hint(tp, skb); | |
883 | ||
884 | tp->lost_out += tcp_skb_pcount(skb); | |
885 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
886 | } | |
887 | } | |
888 | ||
e1aa680f IJ |
889 | static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, |
890 | struct sk_buff *skb) | |
006f582c IJ |
891 | { |
892 | tcp_verify_retransmit_hint(tp, skb); | |
893 | ||
894 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
895 | tp->lost_out += tcp_skb_pcount(skb); | |
896 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
897 | } | |
898 | } | |
899 | ||
1da177e4 LT |
900 | /* This procedure tags the retransmission queue when SACKs arrive. |
901 | * | |
902 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). | |
903 | * Packets in queue with these bits set are counted in variables | |
904 | * sacked_out, retrans_out and lost_out, correspondingly. | |
905 | * | |
906 | * Valid combinations are: | |
907 | * Tag InFlight Description | |
908 | * 0 1 - orig segment is in flight. | |
909 | * S 0 - nothing flies, orig reached receiver. | |
910 | * L 0 - nothing flies, orig lost by net. | |
911 | * R 2 - both orig and retransmit are in flight. | |
912 | * L|R 1 - orig is lost, retransmit is in flight. | |
913 | * S|R 1 - orig reached receiver, retrans is still in flight. | |
914 | * (L|S|R is logically valid, it could occur when L|R is sacked, | |
915 | * but it is equivalent to plain S and code short-curcuits it to S. | |
916 | * L|S is logically invalid, it would mean -1 packet in flight 8)) | |
917 | * | |
918 | * These 6 states form finite state machine, controlled by the following events: | |
919 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) | |
920 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) | |
974c1236 | 921 | * 3. Loss detection event of two flavors: |
1da177e4 LT |
922 | * A. Scoreboard estimator decided the packet is lost. |
923 | * A'. Reno "three dupacks" marks head of queue lost. | |
974c1236 YC |
924 | * A''. Its FACK modification, head until snd.fack is lost. |
925 | * B. SACK arrives sacking SND.NXT at the moment, when the | |
1da177e4 LT |
926 | * segment was retransmitted. |
927 | * 4. D-SACK added new rule: D-SACK changes any tag to S. | |
928 | * | |
929 | * It is pleasant to note, that state diagram turns out to be commutative, | |
930 | * so that we are allowed not to be bothered by order of our actions, | |
931 | * when multiple events arrive simultaneously. (see the function below). | |
932 | * | |
933 | * Reordering detection. | |
934 | * -------------------- | |
935 | * Reordering metric is maximal distance, which a packet can be displaced | |
936 | * in packet stream. With SACKs we can estimate it: | |
937 | * | |
938 | * 1. SACK fills old hole and the corresponding segment was not | |
939 | * ever retransmitted -> reordering. Alas, we cannot use it | |
940 | * when segment was retransmitted. | |
941 | * 2. The last flaw is solved with D-SACK. D-SACK arrives | |
942 | * for retransmitted and already SACKed segment -> reordering.. | |
943 | * Both of these heuristics are not used in Loss state, when we cannot | |
944 | * account for retransmits accurately. | |
5b3c9882 IJ |
945 | * |
946 | * SACK block validation. | |
947 | * ---------------------- | |
948 | * | |
949 | * SACK block range validation checks that the received SACK block fits to | |
950 | * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. | |
951 | * Note that SND.UNA is not included to the range though being valid because | |
0e835331 IJ |
952 | * it means that the receiver is rather inconsistent with itself reporting |
953 | * SACK reneging when it should advance SND.UNA. Such SACK block this is | |
954 | * perfectly valid, however, in light of RFC2018 which explicitly states | |
955 | * that "SACK block MUST reflect the newest segment. Even if the newest | |
956 | * segment is going to be discarded ...", not that it looks very clever | |
957 | * in case of head skb. Due to potentional receiver driven attacks, we | |
958 | * choose to avoid immediate execution of a walk in write queue due to | |
959 | * reneging and defer head skb's loss recovery to standard loss recovery | |
960 | * procedure that will eventually trigger (nothing forbids us doing this). | |
5b3c9882 IJ |
961 | * |
962 | * Implements also blockage to start_seq wrap-around. Problem lies in the | |
963 | * fact that though start_seq (s) is before end_seq (i.e., not reversed), | |
964 | * there's no guarantee that it will be before snd_nxt (n). The problem | |
965 | * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt | |
966 | * wrap (s_w): | |
967 | * | |
968 | * <- outs wnd -> <- wrapzone -> | |
969 | * u e n u_w e_w s n_w | |
970 | * | | | | | | | | |
971 | * |<------------+------+----- TCP seqno space --------------+---------->| | |
972 | * ...-- <2^31 ->| |<--------... | |
973 | * ...---- >2^31 ------>| |<--------... | |
974 | * | |
975 | * Current code wouldn't be vulnerable but it's better still to discard such | |
976 | * crazy SACK blocks. Doing this check for start_seq alone closes somewhat | |
977 | * similar case (end_seq after snd_nxt wrap) as earlier reversed check in | |
978 | * snd_nxt wrap -> snd_una region will then become "well defined", i.e., | |
979 | * equal to the ideal case (infinite seqno space without wrap caused issues). | |
980 | * | |
981 | * With D-SACK the lower bound is extended to cover sequence space below | |
982 | * SND.UNA down to undo_marker, which is the last point of interest. Yet | |
564262c1 | 983 | * again, D-SACK block must not to go across snd_una (for the same reason as |
5b3c9882 IJ |
984 | * for the normal SACK blocks, explained above). But there all simplicity |
985 | * ends, TCP might receive valid D-SACKs below that. As long as they reside | |
986 | * fully below undo_marker they do not affect behavior in anyway and can | |
987 | * therefore be safely ignored. In rare cases (which are more or less | |
988 | * theoretical ones), the D-SACK will nicely cross that boundary due to skb | |
989 | * fragmentation and packet reordering past skb's retransmission. To consider | |
990 | * them correctly, the acceptable range must be extended even more though | |
991 | * the exact amount is rather hard to quantify. However, tp->max_window can | |
992 | * be used as an exaggerated estimate. | |
1da177e4 | 993 | */ |
a2a385d6 ED |
994 | static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, |
995 | u32 start_seq, u32 end_seq) | |
5b3c9882 IJ |
996 | { |
997 | /* Too far in future, or reversed (interpretation is ambiguous) */ | |
998 | if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) | |
a2a385d6 | 999 | return false; |
5b3c9882 IJ |
1000 | |
1001 | /* Nasty start_seq wrap-around check (see comments above) */ | |
1002 | if (!before(start_seq, tp->snd_nxt)) | |
a2a385d6 | 1003 | return false; |
5b3c9882 | 1004 | |
564262c1 | 1005 | /* In outstanding window? ...This is valid exit for D-SACKs too. |
5b3c9882 IJ |
1006 | * start_seq == snd_una is non-sensical (see comments above) |
1007 | */ | |
1008 | if (after(start_seq, tp->snd_una)) | |
a2a385d6 | 1009 | return true; |
5b3c9882 IJ |
1010 | |
1011 | if (!is_dsack || !tp->undo_marker) | |
a2a385d6 | 1012 | return false; |
5b3c9882 IJ |
1013 | |
1014 | /* ...Then it's D-SACK, and must reside below snd_una completely */ | |
f779b2d6 | 1015 | if (after(end_seq, tp->snd_una)) |
a2a385d6 | 1016 | return false; |
5b3c9882 IJ |
1017 | |
1018 | if (!before(start_seq, tp->undo_marker)) | |
a2a385d6 | 1019 | return true; |
5b3c9882 IJ |
1020 | |
1021 | /* Too old */ | |
1022 | if (!after(end_seq, tp->undo_marker)) | |
a2a385d6 | 1023 | return false; |
5b3c9882 IJ |
1024 | |
1025 | /* Undo_marker boundary crossing (overestimates a lot). Known already: | |
1026 | * start_seq < undo_marker and end_seq >= undo_marker. | |
1027 | */ | |
1028 | return !before(start_seq, end_seq - tp->max_window); | |
1029 | } | |
1030 | ||
1c1e87ed | 1031 | /* Check for lost retransmit. This superb idea is borrowed from "ratehalving". |
974c1236 | 1032 | * Event "B". Later note: FACK people cheated me again 8), we have to account |
1c1e87ed | 1033 | * for reordering! Ugly, but should help. |
f785a8e2 IJ |
1034 | * |
1035 | * Search retransmitted skbs from write_queue that were sent when snd_nxt was | |
1036 | * less than what is now known to be received by the other end (derived from | |
9f58f3b7 IJ |
1037 | * highest SACK block). Also calculate the lowest snd_nxt among the remaining |
1038 | * retransmitted skbs to avoid some costly processing per ACKs. | |
1c1e87ed | 1039 | */ |
407ef1de | 1040 | static void tcp_mark_lost_retrans(struct sock *sk) |
1c1e87ed | 1041 | { |
9f58f3b7 | 1042 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1c1e87ed IJ |
1043 | struct tcp_sock *tp = tcp_sk(sk); |
1044 | struct sk_buff *skb; | |
f785a8e2 | 1045 | int cnt = 0; |
df2e014b | 1046 | u32 new_low_seq = tp->snd_nxt; |
6859d494 | 1047 | u32 received_upto = tcp_highest_sack_seq(tp); |
9f58f3b7 IJ |
1048 | |
1049 | if (!tcp_is_fack(tp) || !tp->retrans_out || | |
1050 | !after(received_upto, tp->lost_retrans_low) || | |
1051 | icsk->icsk_ca_state != TCP_CA_Recovery) | |
407ef1de | 1052 | return; |
1c1e87ed IJ |
1053 | |
1054 | tcp_for_write_queue(skb, sk) { | |
1055 | u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; | |
1056 | ||
1057 | if (skb == tcp_send_head(sk)) | |
1058 | break; | |
f785a8e2 | 1059 | if (cnt == tp->retrans_out) |
1c1e87ed IJ |
1060 | break; |
1061 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1062 | continue; | |
1063 | ||
f785a8e2 IJ |
1064 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) |
1065 | continue; | |
1066 | ||
d0af4160 IJ |
1067 | /* TODO: We would like to get rid of tcp_is_fack(tp) only |
1068 | * constraint here (see above) but figuring out that at | |
1069 | * least tp->reordering SACK blocks reside between ack_seq | |
1070 | * and received_upto is not easy task to do cheaply with | |
1071 | * the available datastructures. | |
1072 | * | |
1073 | * Whether FACK should check here for tp->reordering segs | |
1074 | * in-between one could argue for either way (it would be | |
1075 | * rather simple to implement as we could count fack_count | |
1076 | * during the walk and do tp->fackets_out - fack_count). | |
1077 | */ | |
1078 | if (after(received_upto, ack_seq)) { | |
1c1e87ed IJ |
1079 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
1080 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1081 | ||
006f582c | 1082 | tcp_skb_mark_lost_uncond_verify(tp, skb); |
de0744af | 1083 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT); |
f785a8e2 | 1084 | } else { |
df2e014b | 1085 | if (before(ack_seq, new_low_seq)) |
b08d6cb2 | 1086 | new_low_seq = ack_seq; |
f785a8e2 | 1087 | cnt += tcp_skb_pcount(skb); |
1c1e87ed IJ |
1088 | } |
1089 | } | |
b08d6cb2 IJ |
1090 | |
1091 | if (tp->retrans_out) | |
1092 | tp->lost_retrans_low = new_low_seq; | |
1c1e87ed | 1093 | } |
5b3c9882 | 1094 | |
a2a385d6 ED |
1095 | static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, |
1096 | struct tcp_sack_block_wire *sp, int num_sacks, | |
1097 | u32 prior_snd_una) | |
d06e021d | 1098 | { |
1ed83465 | 1099 | struct tcp_sock *tp = tcp_sk(sk); |
d3e2ce3b HH |
1100 | u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); |
1101 | u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); | |
a2a385d6 | 1102 | bool dup_sack = false; |
d06e021d DM |
1103 | |
1104 | if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { | |
a2a385d6 | 1105 | dup_sack = true; |
e60402d0 | 1106 | tcp_dsack_seen(tp); |
de0744af | 1107 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV); |
d06e021d | 1108 | } else if (num_sacks > 1) { |
d3e2ce3b HH |
1109 | u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); |
1110 | u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); | |
d06e021d DM |
1111 | |
1112 | if (!after(end_seq_0, end_seq_1) && | |
1113 | !before(start_seq_0, start_seq_1)) { | |
a2a385d6 | 1114 | dup_sack = true; |
e60402d0 | 1115 | tcp_dsack_seen(tp); |
de0744af PE |
1116 | NET_INC_STATS_BH(sock_net(sk), |
1117 | LINUX_MIB_TCPDSACKOFORECV); | |
d06e021d DM |
1118 | } |
1119 | } | |
1120 | ||
1121 | /* D-SACK for already forgotten data... Do dumb counting. */ | |
6e08d5e3 | 1122 | if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 && |
d06e021d DM |
1123 | !after(end_seq_0, prior_snd_una) && |
1124 | after(end_seq_0, tp->undo_marker)) | |
1125 | tp->undo_retrans--; | |
1126 | ||
1127 | return dup_sack; | |
1128 | } | |
1129 | ||
a1197f5a | 1130 | struct tcp_sacktag_state { |
740b0f18 ED |
1131 | int reord; |
1132 | int fack_count; | |
31231a8a KKJ |
1133 | /* Timestamps for earliest and latest never-retransmitted segment |
1134 | * that was SACKed. RTO needs the earliest RTT to stay conservative, | |
1135 | * but congestion control should still get an accurate delay signal. | |
1136 | */ | |
1137 | struct skb_mstamp first_sackt; | |
1138 | struct skb_mstamp last_sackt; | |
740b0f18 | 1139 | int flag; |
a1197f5a IJ |
1140 | }; |
1141 | ||
d1935942 IJ |
1142 | /* Check if skb is fully within the SACK block. In presence of GSO skbs, |
1143 | * the incoming SACK may not exactly match but we can find smaller MSS | |
1144 | * aligned portion of it that matches. Therefore we might need to fragment | |
1145 | * which may fail and creates some hassle (caller must handle error case | |
1146 | * returns). | |
832d11c5 IJ |
1147 | * |
1148 | * FIXME: this could be merged to shift decision code | |
d1935942 | 1149 | */ |
0f79efdc | 1150 | static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, |
a2a385d6 | 1151 | u32 start_seq, u32 end_seq) |
d1935942 | 1152 | { |
a2a385d6 ED |
1153 | int err; |
1154 | bool in_sack; | |
d1935942 | 1155 | unsigned int pkt_len; |
adb92db8 | 1156 | unsigned int mss; |
d1935942 IJ |
1157 | |
1158 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1159 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1160 | ||
1161 | if (tcp_skb_pcount(skb) > 1 && !in_sack && | |
1162 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { | |
adb92db8 | 1163 | mss = tcp_skb_mss(skb); |
d1935942 IJ |
1164 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); |
1165 | ||
adb92db8 | 1166 | if (!in_sack) { |
d1935942 | 1167 | pkt_len = start_seq - TCP_SKB_CB(skb)->seq; |
adb92db8 IJ |
1168 | if (pkt_len < mss) |
1169 | pkt_len = mss; | |
1170 | } else { | |
d1935942 | 1171 | pkt_len = end_seq - TCP_SKB_CB(skb)->seq; |
adb92db8 IJ |
1172 | if (pkt_len < mss) |
1173 | return -EINVAL; | |
1174 | } | |
1175 | ||
1176 | /* Round if necessary so that SACKs cover only full MSSes | |
1177 | * and/or the remaining small portion (if present) | |
1178 | */ | |
1179 | if (pkt_len > mss) { | |
1180 | unsigned int new_len = (pkt_len / mss) * mss; | |
1181 | if (!in_sack && new_len < pkt_len) { | |
1182 | new_len += mss; | |
2cd0d743 | 1183 | if (new_len >= skb->len) |
adb92db8 IJ |
1184 | return 0; |
1185 | } | |
1186 | pkt_len = new_len; | |
1187 | } | |
6cc55e09 | 1188 | err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC); |
d1935942 IJ |
1189 | if (err < 0) |
1190 | return err; | |
1191 | } | |
1192 | ||
1193 | return in_sack; | |
1194 | } | |
1195 | ||
cc9a672e NC |
1196 | /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ |
1197 | static u8 tcp_sacktag_one(struct sock *sk, | |
1198 | struct tcp_sacktag_state *state, u8 sacked, | |
1199 | u32 start_seq, u32 end_seq, | |
740b0f18 ED |
1200 | int dup_sack, int pcount, |
1201 | const struct skb_mstamp *xmit_time) | |
9e10c47c | 1202 | { |
6859d494 | 1203 | struct tcp_sock *tp = tcp_sk(sk); |
a1197f5a | 1204 | int fack_count = state->fack_count; |
9e10c47c IJ |
1205 | |
1206 | /* Account D-SACK for retransmitted packet. */ | |
1207 | if (dup_sack && (sacked & TCPCB_RETRANS)) { | |
6e08d5e3 | 1208 | if (tp->undo_marker && tp->undo_retrans > 0 && |
cc9a672e | 1209 | after(end_seq, tp->undo_marker)) |
9e10c47c | 1210 | tp->undo_retrans--; |
ede9f3b1 | 1211 | if (sacked & TCPCB_SACKED_ACKED) |
a1197f5a | 1212 | state->reord = min(fack_count, state->reord); |
9e10c47c IJ |
1213 | } |
1214 | ||
1215 | /* Nothing to do; acked frame is about to be dropped (was ACKed). */ | |
cc9a672e | 1216 | if (!after(end_seq, tp->snd_una)) |
a1197f5a | 1217 | return sacked; |
9e10c47c IJ |
1218 | |
1219 | if (!(sacked & TCPCB_SACKED_ACKED)) { | |
1220 | if (sacked & TCPCB_SACKED_RETRANS) { | |
1221 | /* If the segment is not tagged as lost, | |
1222 | * we do not clear RETRANS, believing | |
1223 | * that retransmission is still in flight. | |
1224 | */ | |
1225 | if (sacked & TCPCB_LOST) { | |
a1197f5a | 1226 | sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); |
f58b22fd IJ |
1227 | tp->lost_out -= pcount; |
1228 | tp->retrans_out -= pcount; | |
9e10c47c IJ |
1229 | } |
1230 | } else { | |
1231 | if (!(sacked & TCPCB_RETRANS)) { | |
1232 | /* New sack for not retransmitted frame, | |
1233 | * which was in hole. It is reordering. | |
1234 | */ | |
cc9a672e | 1235 | if (before(start_seq, |
9e10c47c | 1236 | tcp_highest_sack_seq(tp))) |
a1197f5a IJ |
1237 | state->reord = min(fack_count, |
1238 | state->reord); | |
e33099f9 YC |
1239 | if (!after(end_seq, tp->high_seq)) |
1240 | state->flag |= FLAG_ORIG_SACK_ACKED; | |
31231a8a KKJ |
1241 | if (state->first_sackt.v64 == 0) |
1242 | state->first_sackt = *xmit_time; | |
1243 | state->last_sackt = *xmit_time; | |
9e10c47c IJ |
1244 | } |
1245 | ||
1246 | if (sacked & TCPCB_LOST) { | |
a1197f5a | 1247 | sacked &= ~TCPCB_LOST; |
f58b22fd | 1248 | tp->lost_out -= pcount; |
9e10c47c IJ |
1249 | } |
1250 | } | |
1251 | ||
a1197f5a IJ |
1252 | sacked |= TCPCB_SACKED_ACKED; |
1253 | state->flag |= FLAG_DATA_SACKED; | |
f58b22fd | 1254 | tp->sacked_out += pcount; |
9e10c47c | 1255 | |
f58b22fd | 1256 | fack_count += pcount; |
9e10c47c IJ |
1257 | |
1258 | /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ | |
00db4124 | 1259 | if (!tcp_is_fack(tp) && tp->lost_skb_hint && |
cc9a672e | 1260 | before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) |
f58b22fd | 1261 | tp->lost_cnt_hint += pcount; |
9e10c47c IJ |
1262 | |
1263 | if (fack_count > tp->fackets_out) | |
1264 | tp->fackets_out = fack_count; | |
9e10c47c IJ |
1265 | } |
1266 | ||
1267 | /* D-SACK. We can detect redundant retransmission in S|R and plain R | |
1268 | * frames and clear it. undo_retrans is decreased above, L|R frames | |
1269 | * are accounted above as well. | |
1270 | */ | |
a1197f5a IJ |
1271 | if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { |
1272 | sacked &= ~TCPCB_SACKED_RETRANS; | |
f58b22fd | 1273 | tp->retrans_out -= pcount; |
9e10c47c IJ |
1274 | } |
1275 | ||
a1197f5a | 1276 | return sacked; |
9e10c47c IJ |
1277 | } |
1278 | ||
daef52ba NC |
1279 | /* Shift newly-SACKed bytes from this skb to the immediately previous |
1280 | * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. | |
1281 | */ | |
a2a385d6 ED |
1282 | static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, |
1283 | struct tcp_sacktag_state *state, | |
1284 | unsigned int pcount, int shifted, int mss, | |
1285 | bool dup_sack) | |
832d11c5 IJ |
1286 | { |
1287 | struct tcp_sock *tp = tcp_sk(sk); | |
50133161 | 1288 | struct sk_buff *prev = tcp_write_queue_prev(sk, skb); |
daef52ba NC |
1289 | u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ |
1290 | u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ | |
832d11c5 IJ |
1291 | |
1292 | BUG_ON(!pcount); | |
1293 | ||
4c90d3b3 NC |
1294 | /* Adjust counters and hints for the newly sacked sequence |
1295 | * range but discard the return value since prev is already | |
1296 | * marked. We must tag the range first because the seq | |
1297 | * advancement below implicitly advances | |
1298 | * tcp_highest_sack_seq() when skb is highest_sack. | |
1299 | */ | |
1300 | tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, | |
59c9af42 | 1301 | start_seq, end_seq, dup_sack, pcount, |
740b0f18 | 1302 | &skb->skb_mstamp); |
4c90d3b3 NC |
1303 | |
1304 | if (skb == tp->lost_skb_hint) | |
0af2a0d0 NC |
1305 | tp->lost_cnt_hint += pcount; |
1306 | ||
832d11c5 IJ |
1307 | TCP_SKB_CB(prev)->end_seq += shifted; |
1308 | TCP_SKB_CB(skb)->seq += shifted; | |
1309 | ||
cd7d8498 ED |
1310 | tcp_skb_pcount_add(prev, pcount); |
1311 | BUG_ON(tcp_skb_pcount(skb) < pcount); | |
1312 | tcp_skb_pcount_add(skb, -pcount); | |
832d11c5 IJ |
1313 | |
1314 | /* When we're adding to gso_segs == 1, gso_size will be zero, | |
1315 | * in theory this shouldn't be necessary but as long as DSACK | |
1316 | * code can come after this skb later on it's better to keep | |
1317 | * setting gso_size to something. | |
1318 | */ | |
f69ad292 ED |
1319 | if (!TCP_SKB_CB(prev)->tcp_gso_size) |
1320 | TCP_SKB_CB(prev)->tcp_gso_size = mss; | |
832d11c5 IJ |
1321 | |
1322 | /* CHECKME: To clear or not to clear? Mimics normal skb currently */ | |
51466a75 | 1323 | if (tcp_skb_pcount(skb) <= 1) |
f69ad292 | 1324 | TCP_SKB_CB(skb)->tcp_gso_size = 0; |
832d11c5 | 1325 | |
832d11c5 IJ |
1326 | /* Difference in this won't matter, both ACKed by the same cumul. ACK */ |
1327 | TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); | |
1328 | ||
832d11c5 IJ |
1329 | if (skb->len > 0) { |
1330 | BUG_ON(!tcp_skb_pcount(skb)); | |
111cc8b9 | 1331 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED); |
a2a385d6 | 1332 | return false; |
832d11c5 IJ |
1333 | } |
1334 | ||
1335 | /* Whole SKB was eaten :-) */ | |
1336 | ||
92ee76b6 IJ |
1337 | if (skb == tp->retransmit_skb_hint) |
1338 | tp->retransmit_skb_hint = prev; | |
92ee76b6 IJ |
1339 | if (skb == tp->lost_skb_hint) { |
1340 | tp->lost_skb_hint = prev; | |
1341 | tp->lost_cnt_hint -= tcp_skb_pcount(prev); | |
1342 | } | |
1343 | ||
5e8a402f ED |
1344 | TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
1345 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) | |
1346 | TCP_SKB_CB(prev)->end_seq++; | |
1347 | ||
832d11c5 IJ |
1348 | if (skb == tcp_highest_sack(sk)) |
1349 | tcp_advance_highest_sack(sk, skb); | |
1350 | ||
1351 | tcp_unlink_write_queue(skb, sk); | |
1352 | sk_wmem_free_skb(sk, skb); | |
1353 | ||
111cc8b9 IJ |
1354 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED); |
1355 | ||
a2a385d6 | 1356 | return true; |
832d11c5 IJ |
1357 | } |
1358 | ||
1359 | /* I wish gso_size would have a bit more sane initialization than | |
1360 | * something-or-zero which complicates things | |
1361 | */ | |
cf533ea5 | 1362 | static int tcp_skb_seglen(const struct sk_buff *skb) |
832d11c5 | 1363 | { |
775ffabf | 1364 | return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); |
832d11c5 IJ |
1365 | } |
1366 | ||
1367 | /* Shifting pages past head area doesn't work */ | |
cf533ea5 | 1368 | static int skb_can_shift(const struct sk_buff *skb) |
832d11c5 IJ |
1369 | { |
1370 | return !skb_headlen(skb) && skb_is_nonlinear(skb); | |
1371 | } | |
1372 | ||
1373 | /* Try collapsing SACK blocks spanning across multiple skbs to a single | |
1374 | * skb. | |
1375 | */ | |
1376 | static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, | |
a1197f5a | 1377 | struct tcp_sacktag_state *state, |
832d11c5 | 1378 | u32 start_seq, u32 end_seq, |
a2a385d6 | 1379 | bool dup_sack) |
832d11c5 IJ |
1380 | { |
1381 | struct tcp_sock *tp = tcp_sk(sk); | |
1382 | struct sk_buff *prev; | |
1383 | int mss; | |
1384 | int pcount = 0; | |
1385 | int len; | |
1386 | int in_sack; | |
1387 | ||
1388 | if (!sk_can_gso(sk)) | |
1389 | goto fallback; | |
1390 | ||
1391 | /* Normally R but no L won't result in plain S */ | |
1392 | if (!dup_sack && | |
9969ca5f | 1393 | (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) |
832d11c5 IJ |
1394 | goto fallback; |
1395 | if (!skb_can_shift(skb)) | |
1396 | goto fallback; | |
1397 | /* This frame is about to be dropped (was ACKed). */ | |
1398 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1399 | goto fallback; | |
1400 | ||
1401 | /* Can only happen with delayed DSACK + discard craziness */ | |
1402 | if (unlikely(skb == tcp_write_queue_head(sk))) | |
1403 | goto fallback; | |
1404 | prev = tcp_write_queue_prev(sk, skb); | |
1405 | ||
1406 | if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) | |
1407 | goto fallback; | |
1408 | ||
1409 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1410 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1411 | ||
1412 | if (in_sack) { | |
1413 | len = skb->len; | |
1414 | pcount = tcp_skb_pcount(skb); | |
775ffabf | 1415 | mss = tcp_skb_seglen(skb); |
832d11c5 IJ |
1416 | |
1417 | /* TODO: Fix DSACKs to not fragment already SACKed and we can | |
1418 | * drop this restriction as unnecessary | |
1419 | */ | |
775ffabf | 1420 | if (mss != tcp_skb_seglen(prev)) |
832d11c5 IJ |
1421 | goto fallback; |
1422 | } else { | |
1423 | if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) | |
1424 | goto noop; | |
1425 | /* CHECKME: This is non-MSS split case only?, this will | |
1426 | * cause skipped skbs due to advancing loop btw, original | |
1427 | * has that feature too | |
1428 | */ | |
1429 | if (tcp_skb_pcount(skb) <= 1) | |
1430 | goto noop; | |
1431 | ||
1432 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); | |
1433 | if (!in_sack) { | |
1434 | /* TODO: head merge to next could be attempted here | |
1435 | * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), | |
1436 | * though it might not be worth of the additional hassle | |
1437 | * | |
1438 | * ...we can probably just fallback to what was done | |
1439 | * previously. We could try merging non-SACKed ones | |
1440 | * as well but it probably isn't going to buy off | |
1441 | * because later SACKs might again split them, and | |
1442 | * it would make skb timestamp tracking considerably | |
1443 | * harder problem. | |
1444 | */ | |
1445 | goto fallback; | |
1446 | } | |
1447 | ||
1448 | len = end_seq - TCP_SKB_CB(skb)->seq; | |
1449 | BUG_ON(len < 0); | |
1450 | BUG_ON(len > skb->len); | |
1451 | ||
1452 | /* MSS boundaries should be honoured or else pcount will | |
1453 | * severely break even though it makes things bit trickier. | |
1454 | * Optimize common case to avoid most of the divides | |
1455 | */ | |
1456 | mss = tcp_skb_mss(skb); | |
1457 | ||
1458 | /* TODO: Fix DSACKs to not fragment already SACKed and we can | |
1459 | * drop this restriction as unnecessary | |
1460 | */ | |
775ffabf | 1461 | if (mss != tcp_skb_seglen(prev)) |
832d11c5 IJ |
1462 | goto fallback; |
1463 | ||
1464 | if (len == mss) { | |
1465 | pcount = 1; | |
1466 | } else if (len < mss) { | |
1467 | goto noop; | |
1468 | } else { | |
1469 | pcount = len / mss; | |
1470 | len = pcount * mss; | |
1471 | } | |
1472 | } | |
1473 | ||
4648dc97 NC |
1474 | /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ |
1475 | if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) | |
1476 | goto fallback; | |
1477 | ||
832d11c5 IJ |
1478 | if (!skb_shift(prev, skb, len)) |
1479 | goto fallback; | |
9ec06ff5 | 1480 | if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack)) |
832d11c5 IJ |
1481 | goto out; |
1482 | ||
1483 | /* Hole filled allows collapsing with the next as well, this is very | |
1484 | * useful when hole on every nth skb pattern happens | |
1485 | */ | |
1486 | if (prev == tcp_write_queue_tail(sk)) | |
1487 | goto out; | |
1488 | skb = tcp_write_queue_next(sk, prev); | |
1489 | ||
f0bc52f3 IJ |
1490 | if (!skb_can_shift(skb) || |
1491 | (skb == tcp_send_head(sk)) || | |
1492 | ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || | |
775ffabf | 1493 | (mss != tcp_skb_seglen(skb))) |
832d11c5 IJ |
1494 | goto out; |
1495 | ||
1496 | len = skb->len; | |
1497 | if (skb_shift(prev, skb, len)) { | |
1498 | pcount += tcp_skb_pcount(skb); | |
9ec06ff5 | 1499 | tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0); |
832d11c5 IJ |
1500 | } |
1501 | ||
1502 | out: | |
a1197f5a | 1503 | state->fack_count += pcount; |
832d11c5 IJ |
1504 | return prev; |
1505 | ||
1506 | noop: | |
1507 | return skb; | |
1508 | ||
1509 | fallback: | |
111cc8b9 | 1510 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); |
832d11c5 IJ |
1511 | return NULL; |
1512 | } | |
1513 | ||
68f8353b IJ |
1514 | static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, |
1515 | struct tcp_sack_block *next_dup, | |
a1197f5a | 1516 | struct tcp_sacktag_state *state, |
68f8353b | 1517 | u32 start_seq, u32 end_seq, |
a2a385d6 | 1518 | bool dup_sack_in) |
68f8353b | 1519 | { |
832d11c5 IJ |
1520 | struct tcp_sock *tp = tcp_sk(sk); |
1521 | struct sk_buff *tmp; | |
1522 | ||
68f8353b IJ |
1523 | tcp_for_write_queue_from(skb, sk) { |
1524 | int in_sack = 0; | |
a2a385d6 | 1525 | bool dup_sack = dup_sack_in; |
68f8353b IJ |
1526 | |
1527 | if (skb == tcp_send_head(sk)) | |
1528 | break; | |
1529 | ||
1530 | /* queue is in-order => we can short-circuit the walk early */ | |
1531 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) | |
1532 | break; | |
1533 | ||
00db4124 | 1534 | if (next_dup && |
68f8353b IJ |
1535 | before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { |
1536 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1537 | next_dup->start_seq, | |
1538 | next_dup->end_seq); | |
1539 | if (in_sack > 0) | |
a2a385d6 | 1540 | dup_sack = true; |
68f8353b IJ |
1541 | } |
1542 | ||
832d11c5 IJ |
1543 | /* skb reference here is a bit tricky to get right, since |
1544 | * shifting can eat and free both this skb and the next, | |
1545 | * so not even _safe variant of the loop is enough. | |
1546 | */ | |
1547 | if (in_sack <= 0) { | |
a1197f5a IJ |
1548 | tmp = tcp_shift_skb_data(sk, skb, state, |
1549 | start_seq, end_seq, dup_sack); | |
00db4124 | 1550 | if (tmp) { |
832d11c5 IJ |
1551 | if (tmp != skb) { |
1552 | skb = tmp; | |
1553 | continue; | |
1554 | } | |
1555 | ||
1556 | in_sack = 0; | |
1557 | } else { | |
1558 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1559 | start_seq, | |
1560 | end_seq); | |
1561 | } | |
1562 | } | |
1563 | ||
68f8353b IJ |
1564 | if (unlikely(in_sack < 0)) |
1565 | break; | |
1566 | ||
832d11c5 | 1567 | if (in_sack) { |
cc9a672e NC |
1568 | TCP_SKB_CB(skb)->sacked = |
1569 | tcp_sacktag_one(sk, | |
1570 | state, | |
1571 | TCP_SKB_CB(skb)->sacked, | |
1572 | TCP_SKB_CB(skb)->seq, | |
1573 | TCP_SKB_CB(skb)->end_seq, | |
1574 | dup_sack, | |
59c9af42 | 1575 | tcp_skb_pcount(skb), |
740b0f18 | 1576 | &skb->skb_mstamp); |
68f8353b | 1577 | |
832d11c5 IJ |
1578 | if (!before(TCP_SKB_CB(skb)->seq, |
1579 | tcp_highest_sack_seq(tp))) | |
1580 | tcp_advance_highest_sack(sk, skb); | |
1581 | } | |
1582 | ||
a1197f5a | 1583 | state->fack_count += tcp_skb_pcount(skb); |
68f8353b IJ |
1584 | } |
1585 | return skb; | |
1586 | } | |
1587 | ||
1588 | /* Avoid all extra work that is being done by sacktag while walking in | |
1589 | * a normal way | |
1590 | */ | |
1591 | static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, | |
a1197f5a IJ |
1592 | struct tcp_sacktag_state *state, |
1593 | u32 skip_to_seq) | |
68f8353b IJ |
1594 | { |
1595 | tcp_for_write_queue_from(skb, sk) { | |
1596 | if (skb == tcp_send_head(sk)) | |
1597 | break; | |
1598 | ||
e8bae275 | 1599 | if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) |
68f8353b | 1600 | break; |
d152a7d8 | 1601 | |
a1197f5a | 1602 | state->fack_count += tcp_skb_pcount(skb); |
68f8353b IJ |
1603 | } |
1604 | return skb; | |
1605 | } | |
1606 | ||
1607 | static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, | |
1608 | struct sock *sk, | |
1609 | struct tcp_sack_block *next_dup, | |
a1197f5a IJ |
1610 | struct tcp_sacktag_state *state, |
1611 | u32 skip_to_seq) | |
68f8353b | 1612 | { |
51456b29 | 1613 | if (!next_dup) |
68f8353b IJ |
1614 | return skb; |
1615 | ||
1616 | if (before(next_dup->start_seq, skip_to_seq)) { | |
a1197f5a IJ |
1617 | skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); |
1618 | skb = tcp_sacktag_walk(skb, sk, NULL, state, | |
1619 | next_dup->start_seq, next_dup->end_seq, | |
1620 | 1); | |
68f8353b IJ |
1621 | } |
1622 | ||
1623 | return skb; | |
1624 | } | |
1625 | ||
cf533ea5 | 1626 | static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) |
68f8353b IJ |
1627 | { |
1628 | return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); | |
1629 | } | |
1630 | ||
1da177e4 | 1631 | static int |
cf533ea5 | 1632 | tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, |
196da974 | 1633 | u32 prior_snd_una, struct tcp_sacktag_state *state) |
1da177e4 LT |
1634 | { |
1635 | struct tcp_sock *tp = tcp_sk(sk); | |
cf533ea5 ED |
1636 | const unsigned char *ptr = (skb_transport_header(ack_skb) + |
1637 | TCP_SKB_CB(ack_skb)->sacked); | |
fd6dad61 | 1638 | struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); |
4389dded | 1639 | struct tcp_sack_block sp[TCP_NUM_SACKS]; |
68f8353b IJ |
1640 | struct tcp_sack_block *cache; |
1641 | struct sk_buff *skb; | |
4389dded | 1642 | int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); |
fd6dad61 | 1643 | int used_sacks; |
a2a385d6 | 1644 | bool found_dup_sack = false; |
68f8353b | 1645 | int i, j; |
fda03fbb | 1646 | int first_sack_index; |
1da177e4 | 1647 | |
196da974 KKJ |
1648 | state->flag = 0; |
1649 | state->reord = tp->packets_out; | |
a1197f5a | 1650 | |
d738cd8f | 1651 | if (!tp->sacked_out) { |
de83c058 IJ |
1652 | if (WARN_ON(tp->fackets_out)) |
1653 | tp->fackets_out = 0; | |
6859d494 | 1654 | tcp_highest_sack_reset(sk); |
d738cd8f | 1655 | } |
1da177e4 | 1656 | |
1ed83465 | 1657 | found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, |
d06e021d DM |
1658 | num_sacks, prior_snd_una); |
1659 | if (found_dup_sack) | |
196da974 | 1660 | state->flag |= FLAG_DSACKING_ACK; |
6f74651a BE |
1661 | |
1662 | /* Eliminate too old ACKs, but take into | |
1663 | * account more or less fresh ones, they can | |
1664 | * contain valid SACK info. | |
1665 | */ | |
1666 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) | |
1667 | return 0; | |
1668 |