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; |
f6722583 | 98 | int sysctl_tcp_min_rtt_wlen __read_mostly = 300; |
1da177e4 | 99 | |
7e380175 AP |
100 | int sysctl_tcp_thin_dupack __read_mostly; |
101 | ||
ab32ea5d | 102 | int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; |
6ba8a3b1 | 103 | int sysctl_tcp_early_retrans __read_mostly = 3; |
032ee423 | 104 | int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2; |
1da177e4 | 105 | |
1da177e4 LT |
106 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ |
107 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ | |
108 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ | |
109 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ | |
110 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ | |
111 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ | |
112 | #define FLAG_ECE 0x40 /* ECE in this ACK */ | |
291a00d1 | 113 | #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */ |
1da177e4 | 114 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ |
e33099f9 | 115 | #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */ |
2e605294 | 116 | #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ |
564262c1 | 117 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ |
cadbd031 | 118 | #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ |
12fb3dd9 | 119 | #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */ |
1da177e4 LT |
120 | |
121 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) | |
122 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) | |
123 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) | |
124 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) | |
125 | ||
1da177e4 | 126 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) |
bdf1ee5d | 127 | #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) |
1da177e4 | 128 | |
e662ca40 YC |
129 | #define REXMIT_NONE 0 /* no loss recovery to do */ |
130 | #define REXMIT_LOST 1 /* retransmit packets marked lost */ | |
131 | #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */ | |
132 | ||
e905a9ed | 133 | /* Adapt the MSS value used to make delayed ack decision to the |
1da177e4 | 134 | * real world. |
e905a9ed | 135 | */ |
056834d9 | 136 | static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 137 | { |
463c84b9 | 138 | struct inet_connection_sock *icsk = inet_csk(sk); |
e905a9ed | 139 | const unsigned int lss = icsk->icsk_ack.last_seg_size; |
463c84b9 | 140 | unsigned int len; |
1da177e4 | 141 | |
e905a9ed | 142 | icsk->icsk_ack.last_seg_size = 0; |
1da177e4 LT |
143 | |
144 | /* skb->len may jitter because of SACKs, even if peer | |
145 | * sends good full-sized frames. | |
146 | */ | |
056834d9 | 147 | len = skb_shinfo(skb)->gso_size ? : skb->len; |
463c84b9 ACM |
148 | if (len >= icsk->icsk_ack.rcv_mss) { |
149 | icsk->icsk_ack.rcv_mss = len; | |
1da177e4 LT |
150 | } else { |
151 | /* Otherwise, we make more careful check taking into account, | |
152 | * that SACKs block is variable. | |
153 | * | |
154 | * "len" is invariant segment length, including TCP header. | |
155 | */ | |
9c70220b | 156 | len += skb->data - skb_transport_header(skb); |
bee7ca9e | 157 | if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || |
1da177e4 LT |
158 | /* If PSH is not set, packet should be |
159 | * full sized, provided peer TCP is not badly broken. | |
160 | * This observation (if it is correct 8)) allows | |
161 | * to handle super-low mtu links fairly. | |
162 | */ | |
163 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && | |
aa8223c7 | 164 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { |
1da177e4 LT |
165 | /* Subtract also invariant (if peer is RFC compliant), |
166 | * tcp header plus fixed timestamp option length. | |
167 | * Resulting "len" is MSS free of SACK jitter. | |
168 | */ | |
463c84b9 ACM |
169 | len -= tcp_sk(sk)->tcp_header_len; |
170 | icsk->icsk_ack.last_seg_size = len; | |
1da177e4 | 171 | if (len == lss) { |
463c84b9 | 172 | icsk->icsk_ack.rcv_mss = len; |
1da177e4 LT |
173 | return; |
174 | } | |
175 | } | |
1ef9696c AK |
176 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) |
177 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; | |
463c84b9 | 178 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
1da177e4 LT |
179 | } |
180 | } | |
181 | ||
463c84b9 | 182 | static void tcp_incr_quickack(struct sock *sk) |
1da177e4 | 183 | { |
463c84b9 | 184 | struct inet_connection_sock *icsk = inet_csk(sk); |
95c96174 | 185 | unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); |
1da177e4 | 186 | |
056834d9 IJ |
187 | if (quickacks == 0) |
188 | quickacks = 2; | |
463c84b9 ACM |
189 | if (quickacks > icsk->icsk_ack.quick) |
190 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); | |
1da177e4 LT |
191 | } |
192 | ||
1b9f4092 | 193 | static void tcp_enter_quickack_mode(struct sock *sk) |
1da177e4 | 194 | { |
463c84b9 ACM |
195 | struct inet_connection_sock *icsk = inet_csk(sk); |
196 | tcp_incr_quickack(sk); | |
197 | icsk->icsk_ack.pingpong = 0; | |
198 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 LT |
199 | } |
200 | ||
201 | /* Send ACKs quickly, if "quick" count is not exhausted | |
202 | * and the session is not interactive. | |
203 | */ | |
204 | ||
2251ae46 | 205 | static bool tcp_in_quickack_mode(struct sock *sk) |
1da177e4 | 206 | { |
463c84b9 | 207 | const struct inet_connection_sock *icsk = inet_csk(sk); |
2251ae46 | 208 | const struct dst_entry *dst = __sk_dst_get(sk); |
a2a385d6 | 209 | |
2251ae46 JM |
210 | return (dst && dst_metric(dst, RTAX_QUICKACK)) || |
211 | (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong); | |
1da177e4 LT |
212 | } |
213 | ||
735d3831 | 214 | static void tcp_ecn_queue_cwr(struct tcp_sock *tp) |
bdf1ee5d | 215 | { |
056834d9 | 216 | if (tp->ecn_flags & TCP_ECN_OK) |
bdf1ee5d IJ |
217 | tp->ecn_flags |= TCP_ECN_QUEUE_CWR; |
218 | } | |
219 | ||
735d3831 | 220 | static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb) |
bdf1ee5d IJ |
221 | { |
222 | if (tcp_hdr(skb)->cwr) | |
223 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
224 | } | |
225 | ||
735d3831 | 226 | static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp) |
bdf1ee5d IJ |
227 | { |
228 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
229 | } | |
230 | ||
735d3831 | 231 | static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) |
bdf1ee5d | 232 | { |
b82d1bb4 | 233 | switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { |
7a269ffa | 234 | case INET_ECN_NOT_ECT: |
bdf1ee5d | 235 | /* Funny extension: if ECT is not set on a segment, |
7a269ffa ED |
236 | * and we already seen ECT on a previous segment, |
237 | * it is probably a retransmit. | |
238 | */ | |
239 | if (tp->ecn_flags & TCP_ECN_SEEN) | |
bdf1ee5d | 240 | tcp_enter_quickack_mode((struct sock *)tp); |
7a269ffa ED |
241 | break; |
242 | case INET_ECN_CE: | |
9890092e FW |
243 | if (tcp_ca_needs_ecn((struct sock *)tp)) |
244 | tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE); | |
245 | ||
aae06bf5 ED |
246 | if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { |
247 | /* Better not delay acks, sender can have a very low cwnd */ | |
248 | tcp_enter_quickack_mode((struct sock *)tp); | |
249 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; | |
250 | } | |
9890092e FW |
251 | tp->ecn_flags |= TCP_ECN_SEEN; |
252 | break; | |
7a269ffa | 253 | default: |
9890092e FW |
254 | if (tcp_ca_needs_ecn((struct sock *)tp)) |
255 | tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE); | |
7a269ffa | 256 | tp->ecn_flags |= TCP_ECN_SEEN; |
9890092e | 257 | break; |
bdf1ee5d IJ |
258 | } |
259 | } | |
260 | ||
735d3831 FW |
261 | static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) |
262 | { | |
263 | if (tp->ecn_flags & TCP_ECN_OK) | |
264 | __tcp_ecn_check_ce(tp, skb); | |
265 | } | |
266 | ||
267 | static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) | |
bdf1ee5d | 268 | { |
056834d9 | 269 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) |
bdf1ee5d IJ |
270 | tp->ecn_flags &= ~TCP_ECN_OK; |
271 | } | |
272 | ||
735d3831 | 273 | static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) |
bdf1ee5d | 274 | { |
056834d9 | 275 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) |
bdf1ee5d IJ |
276 | tp->ecn_flags &= ~TCP_ECN_OK; |
277 | } | |
278 | ||
735d3831 | 279 | static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) |
bdf1ee5d | 280 | { |
056834d9 | 281 | if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) |
a2a385d6 ED |
282 | return true; |
283 | return false; | |
bdf1ee5d IJ |
284 | } |
285 | ||
1da177e4 LT |
286 | /* Buffer size and advertised window tuning. |
287 | * | |
288 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. | |
289 | */ | |
290 | ||
6ae70532 | 291 | static void tcp_sndbuf_expand(struct sock *sk) |
1da177e4 | 292 | { |
6ae70532 ED |
293 | const struct tcp_sock *tp = tcp_sk(sk); |
294 | int sndmem, per_mss; | |
295 | u32 nr_segs; | |
296 | ||
297 | /* Worst case is non GSO/TSO : each frame consumes one skb | |
298 | * and skb->head is kmalloced using power of two area of memory | |
299 | */ | |
300 | per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + | |
301 | MAX_TCP_HEADER + | |
302 | SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); | |
303 | ||
304 | per_mss = roundup_pow_of_two(per_mss) + | |
305 | SKB_DATA_ALIGN(sizeof(struct sk_buff)); | |
306 | ||
307 | nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd); | |
308 | nr_segs = max_t(u32, nr_segs, tp->reordering + 1); | |
309 | ||
310 | /* Fast Recovery (RFC 5681 3.2) : | |
311 | * Cubic needs 1.7 factor, rounded to 2 to include | |
312 | * extra cushion (application might react slowly to POLLOUT) | |
313 | */ | |
314 | sndmem = 2 * nr_segs * per_mss; | |
1da177e4 | 315 | |
06a59ecb ED |
316 | if (sk->sk_sndbuf < sndmem) |
317 | sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); | |
1da177e4 LT |
318 | } |
319 | ||
320 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) | |
321 | * | |
322 | * All tcp_full_space() is split to two parts: "network" buffer, allocated | |
323 | * forward and advertised in receiver window (tp->rcv_wnd) and | |
324 | * "application buffer", required to isolate scheduling/application | |
325 | * latencies from network. | |
326 | * window_clamp is maximal advertised window. It can be less than | |
327 | * tcp_full_space(), in this case tcp_full_space() - window_clamp | |
328 | * is reserved for "application" buffer. The less window_clamp is | |
329 | * the smoother our behaviour from viewpoint of network, but the lower | |
330 | * throughput and the higher sensitivity of the connection to losses. 8) | |
331 | * | |
332 | * rcv_ssthresh is more strict window_clamp used at "slow start" | |
333 | * phase to predict further behaviour of this connection. | |
334 | * It is used for two goals: | |
335 | * - to enforce header prediction at sender, even when application | |
336 | * requires some significant "application buffer". It is check #1. | |
337 | * - to prevent pruning of receive queue because of misprediction | |
338 | * of receiver window. Check #2. | |
339 | * | |
340 | * The scheme does not work when sender sends good segments opening | |
caa20d9a | 341 | * window and then starts to feed us spaghetti. But it should work |
1da177e4 LT |
342 | * in common situations. Otherwise, we have to rely on queue collapsing. |
343 | */ | |
344 | ||
345 | /* Slow part of check#2. */ | |
9e412ba7 | 346 | static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 347 | { |
9e412ba7 | 348 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 349 | /* Optimize this! */ |
dfd4f0ae ED |
350 | int truesize = tcp_win_from_space(skb->truesize) >> 1; |
351 | int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; | |
1da177e4 LT |
352 | |
353 | while (tp->rcv_ssthresh <= window) { | |
354 | if (truesize <= skb->len) | |
463c84b9 | 355 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; |
1da177e4 LT |
356 | |
357 | truesize >>= 1; | |
358 | window >>= 1; | |
359 | } | |
360 | return 0; | |
361 | } | |
362 | ||
cf533ea5 | 363 | static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 364 | { |
9e412ba7 IJ |
365 | struct tcp_sock *tp = tcp_sk(sk); |
366 | ||
1da177e4 LT |
367 | /* Check #1 */ |
368 | if (tp->rcv_ssthresh < tp->window_clamp && | |
369 | (int)tp->rcv_ssthresh < tcp_space(sk) && | |
b8da51eb | 370 | !tcp_under_memory_pressure(sk)) { |
1da177e4 LT |
371 | int incr; |
372 | ||
373 | /* Check #2. Increase window, if skb with such overhead | |
374 | * will fit to rcvbuf in future. | |
375 | */ | |
376 | if (tcp_win_from_space(skb->truesize) <= skb->len) | |
056834d9 | 377 | incr = 2 * tp->advmss; |
1da177e4 | 378 | else |
9e412ba7 | 379 | incr = __tcp_grow_window(sk, skb); |
1da177e4 LT |
380 | |
381 | if (incr) { | |
4d846f02 | 382 | incr = max_t(int, incr, 2 * skb->len); |
056834d9 IJ |
383 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, |
384 | tp->window_clamp); | |
463c84b9 | 385 | inet_csk(sk)->icsk_ack.quick |= 1; |
1da177e4 LT |
386 | } |
387 | } | |
388 | } | |
389 | ||
390 | /* 3. Tuning rcvbuf, when connection enters established state. */ | |
1da177e4 LT |
391 | static void tcp_fixup_rcvbuf(struct sock *sk) |
392 | { | |
e9266a02 | 393 | u32 mss = tcp_sk(sk)->advmss; |
e9266a02 | 394 | int rcvmem; |
1da177e4 | 395 | |
85f16525 YC |
396 | rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) * |
397 | tcp_default_init_rwnd(mss); | |
e9266a02 | 398 | |
b0983d3c ED |
399 | /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency |
400 | * Allow enough cushion so that sender is not limited by our window | |
401 | */ | |
402 | if (sysctl_tcp_moderate_rcvbuf) | |
403 | rcvmem <<= 2; | |
404 | ||
e9266a02 ED |
405 | if (sk->sk_rcvbuf < rcvmem) |
406 | sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]); | |
1da177e4 LT |
407 | } |
408 | ||
caa20d9a | 409 | /* 4. Try to fixup all. It is made immediately after connection enters |
1da177e4 LT |
410 | * established state. |
411 | */ | |
10467163 | 412 | void tcp_init_buffer_space(struct sock *sk) |
1da177e4 LT |
413 | { |
414 | struct tcp_sock *tp = tcp_sk(sk); | |
415 | int maxwin; | |
416 | ||
417 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) | |
418 | tcp_fixup_rcvbuf(sk); | |
419 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) | |
6ae70532 | 420 | tcp_sndbuf_expand(sk); |
1da177e4 LT |
421 | |
422 | tp->rcvq_space.space = tp->rcv_wnd; | |
b0983d3c ED |
423 | tp->rcvq_space.time = tcp_time_stamp; |
424 | tp->rcvq_space.seq = tp->copied_seq; | |
1da177e4 LT |
425 | |
426 | maxwin = tcp_full_space(sk); | |
427 | ||
428 | if (tp->window_clamp >= maxwin) { | |
429 | tp->window_clamp = maxwin; | |
430 | ||
431 | if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) | |
432 | tp->window_clamp = max(maxwin - | |
433 | (maxwin >> sysctl_tcp_app_win), | |
434 | 4 * tp->advmss); | |
435 | } | |
436 | ||
437 | /* Force reservation of one segment. */ | |
438 | if (sysctl_tcp_app_win && | |
439 | tp->window_clamp > 2 * tp->advmss && | |
440 | tp->window_clamp + tp->advmss > maxwin) | |
441 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); | |
442 | ||
443 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); | |
444 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
445 | } | |
446 | ||
1da177e4 | 447 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ |
9e412ba7 | 448 | static void tcp_clamp_window(struct sock *sk) |
1da177e4 | 449 | { |
9e412ba7 | 450 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 451 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 452 | |
6687e988 | 453 | icsk->icsk_ack.quick = 0; |
1da177e4 | 454 | |
326f36e9 JH |
455 | if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && |
456 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && | |
b8da51eb | 457 | !tcp_under_memory_pressure(sk) && |
180d8cd9 | 458 | sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { |
326f36e9 JH |
459 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), |
460 | sysctl_tcp_rmem[2]); | |
1da177e4 | 461 | } |
326f36e9 | 462 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) |
056834d9 | 463 | tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); |
1da177e4 LT |
464 | } |
465 | ||
40efc6fa SH |
466 | /* Initialize RCV_MSS value. |
467 | * RCV_MSS is an our guess about MSS used by the peer. | |
468 | * We haven't any direct information about the MSS. | |
469 | * It's better to underestimate the RCV_MSS rather than overestimate. | |
470 | * Overestimations make us ACKing less frequently than needed. | |
471 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). | |
472 | */ | |
473 | void tcp_initialize_rcv_mss(struct sock *sk) | |
474 | { | |
cf533ea5 | 475 | const struct tcp_sock *tp = tcp_sk(sk); |
40efc6fa SH |
476 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); |
477 | ||
056834d9 | 478 | hint = min(hint, tp->rcv_wnd / 2); |
bee7ca9e | 479 | hint = min(hint, TCP_MSS_DEFAULT); |
40efc6fa SH |
480 | hint = max(hint, TCP_MIN_MSS); |
481 | ||
482 | inet_csk(sk)->icsk_ack.rcv_mss = hint; | |
483 | } | |
4bc2f18b | 484 | EXPORT_SYMBOL(tcp_initialize_rcv_mss); |
40efc6fa | 485 | |
1da177e4 LT |
486 | /* Receiver "autotuning" code. |
487 | * | |
488 | * The algorithm for RTT estimation w/o timestamps is based on | |
489 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. | |
631dd1a8 | 490 | * <http://public.lanl.gov/radiant/pubs.html#DRS> |
1da177e4 LT |
491 | * |
492 | * More detail on this code can be found at | |
631dd1a8 | 493 | * <http://staff.psc.edu/jheffner/>, |
1da177e4 LT |
494 | * though this reference is out of date. A new paper |
495 | * is pending. | |
496 | */ | |
497 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) | |
498 | { | |
499 | u32 new_sample = tp->rcv_rtt_est.rtt; | |
500 | long m = sample; | |
501 | ||
502 | if (m == 0) | |
503 | m = 1; | |
504 | ||
505 | if (new_sample != 0) { | |
506 | /* If we sample in larger samples in the non-timestamp | |
507 | * case, we could grossly overestimate the RTT especially | |
508 | * with chatty applications or bulk transfer apps which | |
509 | * are stalled on filesystem I/O. | |
510 | * | |
511 | * Also, since we are only going for a minimum in the | |
31f34269 | 512 | * non-timestamp case, we do not smooth things out |
caa20d9a | 513 | * else with timestamps disabled convergence takes too |
1da177e4 LT |
514 | * long. |
515 | */ | |
516 | if (!win_dep) { | |
517 | m -= (new_sample >> 3); | |
518 | new_sample += m; | |
18a223e0 NC |
519 | } else { |
520 | m <<= 3; | |
521 | if (m < new_sample) | |
522 | new_sample = m; | |
523 | } | |
1da177e4 | 524 | } else { |
caa20d9a | 525 | /* No previous measure. */ |
1da177e4 LT |
526 | new_sample = m << 3; |
527 | } | |
528 | ||
529 | if (tp->rcv_rtt_est.rtt != new_sample) | |
530 | tp->rcv_rtt_est.rtt = new_sample; | |
531 | } | |
532 | ||
533 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) | |
534 | { | |
535 | if (tp->rcv_rtt_est.time == 0) | |
536 | goto new_measure; | |
537 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) | |
538 | return; | |
651913ce | 539 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1); |
1da177e4 LT |
540 | |
541 | new_measure: | |
542 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; | |
543 | tp->rcv_rtt_est.time = tcp_time_stamp; | |
544 | } | |
545 | ||
056834d9 IJ |
546 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, |
547 | const struct sk_buff *skb) | |
1da177e4 | 548 | { |
463c84b9 | 549 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
550 | if (tp->rx_opt.rcv_tsecr && |
551 | (TCP_SKB_CB(skb)->end_seq - | |
463c84b9 | 552 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) |
1da177e4 LT |
553 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); |
554 | } | |
555 | ||
556 | /* | |
557 | * This function should be called every time data is copied to user space. | |
558 | * It calculates the appropriate TCP receive buffer space. | |
559 | */ | |
560 | void tcp_rcv_space_adjust(struct sock *sk) | |
561 | { | |
562 | struct tcp_sock *tp = tcp_sk(sk); | |
563 | int time; | |
b0983d3c | 564 | int copied; |
e905a9ed | 565 | |
1da177e4 | 566 | time = tcp_time_stamp - tp->rcvq_space.time; |
056834d9 | 567 | if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) |
1da177e4 | 568 | return; |
e905a9ed | 569 | |
b0983d3c ED |
570 | /* Number of bytes copied to user in last RTT */ |
571 | copied = tp->copied_seq - tp->rcvq_space.seq; | |
572 | if (copied <= tp->rcvq_space.space) | |
573 | goto new_measure; | |
574 | ||
575 | /* A bit of theory : | |
576 | * copied = bytes received in previous RTT, our base window | |
577 | * To cope with packet losses, we need a 2x factor | |
578 | * To cope with slow start, and sender growing its cwin by 100 % | |
579 | * every RTT, we need a 4x factor, because the ACK we are sending | |
580 | * now is for the next RTT, not the current one : | |
581 | * <prev RTT . ><current RTT .. ><next RTT .... > | |
582 | */ | |
583 | ||
584 | if (sysctl_tcp_moderate_rcvbuf && | |
585 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { | |
586 | int rcvwin, rcvmem, rcvbuf; | |
1da177e4 | 587 | |
b0983d3c ED |
588 | /* minimal window to cope with packet losses, assuming |
589 | * steady state. Add some cushion because of small variations. | |
590 | */ | |
591 | rcvwin = (copied << 1) + 16 * tp->advmss; | |
1da177e4 | 592 | |
b0983d3c ED |
593 | /* If rate increased by 25%, |
594 | * assume slow start, rcvwin = 3 * copied | |
595 | * If rate increased by 50%, | |
596 | * assume sender can use 2x growth, rcvwin = 4 * copied | |
597 | */ | |
598 | if (copied >= | |
599 | tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) { | |
600 | if (copied >= | |
601 | tp->rcvq_space.space + (tp->rcvq_space.space >> 1)) | |
602 | rcvwin <<= 1; | |
603 | else | |
604 | rcvwin += (rcvwin >> 1); | |
605 | } | |
1da177e4 | 606 | |
b0983d3c ED |
607 | rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER); |
608 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
609 | rcvmem += 128; | |
1da177e4 | 610 | |
b0983d3c ED |
611 | rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]); |
612 | if (rcvbuf > sk->sk_rcvbuf) { | |
613 | sk->sk_rcvbuf = rcvbuf; | |
1da177e4 | 614 | |
b0983d3c ED |
615 | /* Make the window clamp follow along. */ |
616 | tp->window_clamp = rcvwin; | |
1da177e4 LT |
617 | } |
618 | } | |
b0983d3c | 619 | tp->rcvq_space.space = copied; |
e905a9ed | 620 | |
1da177e4 LT |
621 | new_measure: |
622 | tp->rcvq_space.seq = tp->copied_seq; | |
623 | tp->rcvq_space.time = tcp_time_stamp; | |
624 | } | |
625 | ||
626 | /* There is something which you must keep in mind when you analyze the | |
627 | * behavior of the tp->ato delayed ack timeout interval. When a | |
628 | * connection starts up, we want to ack as quickly as possible. The | |
629 | * problem is that "good" TCP's do slow start at the beginning of data | |
630 | * transmission. The means that until we send the first few ACK's the | |
631 | * sender will sit on his end and only queue most of his data, because | |
632 | * he can only send snd_cwnd unacked packets at any given time. For | |
633 | * each ACK we send, he increments snd_cwnd and transmits more of his | |
634 | * queue. -DaveM | |
635 | */ | |
9e412ba7 | 636 | static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 637 | { |
9e412ba7 | 638 | struct tcp_sock *tp = tcp_sk(sk); |
463c84b9 | 639 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
640 | u32 now; |
641 | ||
463c84b9 | 642 | inet_csk_schedule_ack(sk); |
1da177e4 | 643 | |
463c84b9 | 644 | tcp_measure_rcv_mss(sk, skb); |
1da177e4 LT |
645 | |
646 | tcp_rcv_rtt_measure(tp); | |
e905a9ed | 647 | |
1da177e4 LT |
648 | now = tcp_time_stamp; |
649 | ||
463c84b9 | 650 | if (!icsk->icsk_ack.ato) { |
1da177e4 LT |
651 | /* The _first_ data packet received, initialize |
652 | * delayed ACK engine. | |
653 | */ | |
463c84b9 ACM |
654 | tcp_incr_quickack(sk); |
655 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 | 656 | } else { |
463c84b9 | 657 | int m = now - icsk->icsk_ack.lrcvtime; |
1da177e4 | 658 | |
056834d9 | 659 | if (m <= TCP_ATO_MIN / 2) { |
1da177e4 | 660 | /* The fastest case is the first. */ |
463c84b9 ACM |
661 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; |
662 | } else if (m < icsk->icsk_ack.ato) { | |
663 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; | |
664 | if (icsk->icsk_ack.ato > icsk->icsk_rto) | |
665 | icsk->icsk_ack.ato = icsk->icsk_rto; | |
666 | } else if (m > icsk->icsk_rto) { | |
caa20d9a | 667 | /* Too long gap. Apparently sender failed to |
1da177e4 LT |
668 | * restart window, so that we send ACKs quickly. |
669 | */ | |
463c84b9 | 670 | tcp_incr_quickack(sk); |
3ab224be | 671 | sk_mem_reclaim(sk); |
1da177e4 LT |
672 | } |
673 | } | |
463c84b9 | 674 | icsk->icsk_ack.lrcvtime = now; |
1da177e4 | 675 | |
735d3831 | 676 | tcp_ecn_check_ce(tp, skb); |
1da177e4 LT |
677 | |
678 | if (skb->len >= 128) | |
9e412ba7 | 679 | tcp_grow_window(sk, skb); |
1da177e4 LT |
680 | } |
681 | ||
1da177e4 LT |
682 | /* Called to compute a smoothed rtt estimate. The data fed to this |
683 | * routine either comes from timestamps, or from segments that were | |
684 | * known _not_ to have been retransmitted [see Karn/Partridge | |
685 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 | |
686 | * piece by Van Jacobson. | |
687 | * NOTE: the next three routines used to be one big routine. | |
688 | * To save cycles in the RFC 1323 implementation it was better to break | |
689 | * it up into three procedures. -- erics | |
690 | */ | |
740b0f18 | 691 | static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) |
1da177e4 | 692 | { |
6687e988 | 693 | struct tcp_sock *tp = tcp_sk(sk); |
740b0f18 ED |
694 | long m = mrtt_us; /* RTT */ |
695 | u32 srtt = tp->srtt_us; | |
1da177e4 | 696 | |
1da177e4 LT |
697 | /* The following amusing code comes from Jacobson's |
698 | * article in SIGCOMM '88. Note that rtt and mdev | |
699 | * are scaled versions of rtt and mean deviation. | |
e905a9ed | 700 | * This is designed to be as fast as possible |
1da177e4 LT |
701 | * m stands for "measurement". |
702 | * | |
703 | * On a 1990 paper the rto value is changed to: | |
704 | * RTO = rtt + 4 * mdev | |
705 | * | |
706 | * Funny. This algorithm seems to be very broken. | |
707 | * These formulae increase RTO, when it should be decreased, increase | |
31f34269 | 708 | * too slowly, when it should be increased quickly, decrease too quickly |
1da177e4 LT |
709 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely |
710 | * does not matter how to _calculate_ it. Seems, it was trap | |
711 | * that VJ failed to avoid. 8) | |
712 | */ | |
4a5ab4e2 ED |
713 | if (srtt != 0) { |
714 | m -= (srtt >> 3); /* m is now error in rtt est */ | |
715 | srtt += m; /* rtt = 7/8 rtt + 1/8 new */ | |
1da177e4 LT |
716 | if (m < 0) { |
717 | m = -m; /* m is now abs(error) */ | |
740b0f18 | 718 | m -= (tp->mdev_us >> 2); /* similar update on mdev */ |
1da177e4 LT |
719 | /* This is similar to one of Eifel findings. |
720 | * Eifel blocks mdev updates when rtt decreases. | |
721 | * This solution is a bit different: we use finer gain | |
722 | * for mdev in this case (alpha*beta). | |
723 | * Like Eifel it also prevents growth of rto, | |
724 | * but also it limits too fast rto decreases, | |
725 | * happening in pure Eifel. | |
726 | */ | |
727 | if (m > 0) | |
728 | m >>= 3; | |
729 | } else { | |
740b0f18 | 730 | m -= (tp->mdev_us >> 2); /* similar update on mdev */ |
1da177e4 | 731 | } |
740b0f18 ED |
732 | tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ |
733 | if (tp->mdev_us > tp->mdev_max_us) { | |
734 | tp->mdev_max_us = tp->mdev_us; | |
735 | if (tp->mdev_max_us > tp->rttvar_us) | |
736 | tp->rttvar_us = tp->mdev_max_us; | |
1da177e4 LT |
737 | } |
738 | if (after(tp->snd_una, tp->rtt_seq)) { | |
740b0f18 ED |
739 | if (tp->mdev_max_us < tp->rttvar_us) |
740 | tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; | |
1da177e4 | 741 | tp->rtt_seq = tp->snd_nxt; |
740b0f18 | 742 | tp->mdev_max_us = tcp_rto_min_us(sk); |
1da177e4 LT |
743 | } |
744 | } else { | |
745 | /* no previous measure. */ | |
4a5ab4e2 | 746 | srtt = m << 3; /* take the measured time to be rtt */ |
740b0f18 ED |
747 | tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ |
748 | tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); | |
749 | tp->mdev_max_us = tp->rttvar_us; | |
1da177e4 LT |
750 | tp->rtt_seq = tp->snd_nxt; |
751 | } | |
740b0f18 | 752 | tp->srtt_us = max(1U, srtt); |
1da177e4 LT |
753 | } |
754 | ||
95bd09eb ED |
755 | /* Set the sk_pacing_rate to allow proper sizing of TSO packets. |
756 | * Note: TCP stack does not yet implement pacing. | |
757 | * FQ packet scheduler can be used to implement cheap but effective | |
758 | * TCP pacing, to smooth the burst on large writes when packets | |
759 | * in flight is significantly lower than cwnd (or rwin) | |
760 | */ | |
43e122b0 ED |
761 | int sysctl_tcp_pacing_ss_ratio __read_mostly = 200; |
762 | int sysctl_tcp_pacing_ca_ratio __read_mostly = 120; | |
763 | ||
95bd09eb ED |
764 | static void tcp_update_pacing_rate(struct sock *sk) |
765 | { | |
766 | const struct tcp_sock *tp = tcp_sk(sk); | |
767 | u64 rate; | |
768 | ||
769 | /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ | |
43e122b0 ED |
770 | rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); |
771 | ||
772 | /* current rate is (cwnd * mss) / srtt | |
773 | * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. | |
774 | * In Congestion Avoidance phase, set it to 120 % the current rate. | |
775 | * | |
776 | * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) | |
777 | * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching | |
778 | * end of slow start and should slow down. | |
779 | */ | |
780 | if (tp->snd_cwnd < tp->snd_ssthresh / 2) | |
781 | rate *= sysctl_tcp_pacing_ss_ratio; | |
782 | else | |
783 | rate *= sysctl_tcp_pacing_ca_ratio; | |
95bd09eb ED |
784 | |
785 | rate *= max(tp->snd_cwnd, tp->packets_out); | |
786 | ||
740b0f18 ED |
787 | if (likely(tp->srtt_us)) |
788 | do_div(rate, tp->srtt_us); | |
95bd09eb | 789 | |
ba537427 ED |
790 | /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate |
791 | * without any lock. We want to make sure compiler wont store | |
792 | * intermediate values in this location. | |
793 | */ | |
794 | ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate, | |
795 | sk->sk_max_pacing_rate); | |
95bd09eb ED |
796 | } |
797 | ||
1da177e4 LT |
798 | /* Calculate rto without backoff. This is the second half of Van Jacobson's |
799 | * routine referred to above. | |
800 | */ | |
f7e56a76 | 801 | static void tcp_set_rto(struct sock *sk) |
1da177e4 | 802 | { |
463c84b9 | 803 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
804 | /* Old crap is replaced with new one. 8) |
805 | * | |
806 | * More seriously: | |
807 | * 1. If rtt variance happened to be less 50msec, it is hallucination. | |
808 | * It cannot be less due to utterly erratic ACK generation made | |
809 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ | |
810 | * to do with delayed acks, because at cwnd>2 true delack timeout | |
811 | * is invisible. Actually, Linux-2.4 also generates erratic | |
caa20d9a | 812 | * ACKs in some circumstances. |
1da177e4 | 813 | */ |
f1ecd5d9 | 814 | inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); |
1da177e4 LT |
815 | |
816 | /* 2. Fixups made earlier cannot be right. | |
817 | * If we do not estimate RTO correctly without them, | |
818 | * all the algo is pure shit and should be replaced | |
caa20d9a | 819 | * with correct one. It is exactly, which we pretend to do. |
1da177e4 | 820 | */ |
1da177e4 | 821 | |
ee6aac59 IJ |
822 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo |
823 | * guarantees that rto is higher. | |
824 | */ | |
f1ecd5d9 | 825 | tcp_bound_rto(sk); |
1da177e4 LT |
826 | } |
827 | ||
cf533ea5 | 828 | __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) |
1da177e4 LT |
829 | { |
830 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); | |
831 | ||
22b71c8f | 832 | if (!cwnd) |
442b9635 | 833 | cwnd = TCP_INIT_CWND; |
1da177e4 LT |
834 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); |
835 | } | |
836 | ||
e60402d0 IJ |
837 | /* |
838 | * Packet counting of FACK is based on in-order assumptions, therefore TCP | |
839 | * disables it when reordering is detected | |
840 | */ | |
4aabd8ef | 841 | void tcp_disable_fack(struct tcp_sock *tp) |
e60402d0 | 842 | { |
85cc391c IJ |
843 | /* RFC3517 uses different metric in lost marker => reset on change */ |
844 | if (tcp_is_fack(tp)) | |
845 | tp->lost_skb_hint = NULL; | |
ab56222a | 846 | tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED; |
e60402d0 IJ |
847 | } |
848 | ||
564262c1 | 849 | /* Take a notice that peer is sending D-SACKs */ |
e60402d0 IJ |
850 | static void tcp_dsack_seen(struct tcp_sock *tp) |
851 | { | |
ab56222a | 852 | tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; |
e60402d0 IJ |
853 | } |
854 | ||
6687e988 ACM |
855 | static void tcp_update_reordering(struct sock *sk, const int metric, |
856 | const int ts) | |
1da177e4 | 857 | { |
6687e988 | 858 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 859 | if (metric > tp->reordering) { |
40b215e5 PE |
860 | int mib_idx; |
861 | ||
dca145ff | 862 | tp->reordering = min(sysctl_tcp_max_reordering, metric); |
1da177e4 LT |
863 | |
864 | /* This exciting event is worth to be remembered. 8) */ | |
865 | if (ts) | |
40b215e5 | 866 | mib_idx = LINUX_MIB_TCPTSREORDER; |
e60402d0 | 867 | else if (tcp_is_reno(tp)) |
40b215e5 | 868 | mib_idx = LINUX_MIB_TCPRENOREORDER; |
e60402d0 | 869 | else if (tcp_is_fack(tp)) |
40b215e5 | 870 | mib_idx = LINUX_MIB_TCPFACKREORDER; |
1da177e4 | 871 | else |
40b215e5 PE |
872 | mib_idx = LINUX_MIB_TCPSACKREORDER; |
873 | ||
de0744af | 874 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
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, | |
879 | tp->fackets_out, | |
880 | tp->sacked_out, | |
881 | tp->undo_marker ? tp->undo_retrans : 0); | |
1da177e4 | 882 | #endif |
e60402d0 | 883 | tcp_disable_fack(tp); |
1da177e4 | 884 | } |
eed530b6 YC |
885 | |
886 | if (metric > 0) | |
887 | tcp_disable_early_retrans(tp); | |
4f41b1c5 | 888 | tp->rack.reord = 1; |
1da177e4 LT |
889 | } |
890 | ||
006f582c | 891 | /* This must be called before lost_out is incremented */ |
c8c213f2 IJ |
892 | static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) |
893 | { | |
51456b29 | 894 | if (!tp->retransmit_skb_hint || |
c8c213f2 IJ |
895 | before(TCP_SKB_CB(skb)->seq, |
896 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) | |
006f582c IJ |
897 | tp->retransmit_skb_hint = skb; |
898 | ||
899 | if (!tp->lost_out || | |
900 | after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) | |
901 | tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; | |
c8c213f2 IJ |
902 | } |
903 | ||
41ea36e3 IJ |
904 | static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) |
905 | { | |
906 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
907 | tcp_verify_retransmit_hint(tp, skb); | |
908 | ||
909 | tp->lost_out += tcp_skb_pcount(skb); | |
910 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
911 | } | |
912 | } | |
913 | ||
4f41b1c5 | 914 | void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb) |
006f582c IJ |
915 | { |
916 | tcp_verify_retransmit_hint(tp, skb); | |
917 | ||
918 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
919 | tp->lost_out += tcp_skb_pcount(skb); | |
920 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
921 | } | |
922 | } | |
923 | ||
1da177e4 LT |
924 | /* This procedure tags the retransmission queue when SACKs arrive. |
925 | * | |
926 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). | |
927 | * Packets in queue with these bits set are counted in variables | |
928 | * sacked_out, retrans_out and lost_out, correspondingly. | |
929 | * | |
930 | * Valid combinations are: | |
931 | * Tag InFlight Description | |
932 | * 0 1 - orig segment is in flight. | |
933 | * S 0 - nothing flies, orig reached receiver. | |
934 | * L 0 - nothing flies, orig lost by net. | |
935 | * R 2 - both orig and retransmit are in flight. | |
936 | * L|R 1 - orig is lost, retransmit is in flight. | |
937 | * S|R 1 - orig reached receiver, retrans is still in flight. | |
938 | * (L|S|R is logically valid, it could occur when L|R is sacked, | |
939 | * but it is equivalent to plain S and code short-curcuits it to S. | |
940 | * L|S is logically invalid, it would mean -1 packet in flight 8)) | |
941 | * | |
942 | * These 6 states form finite state machine, controlled by the following events: | |
943 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) | |
944 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) | |
974c1236 | 945 | * 3. Loss detection event of two flavors: |
1da177e4 LT |
946 | * A. Scoreboard estimator decided the packet is lost. |
947 | * A'. Reno "three dupacks" marks head of queue lost. | |
974c1236 YC |
948 | * A''. Its FACK modification, head until snd.fack is lost. |
949 | * B. SACK arrives sacking SND.NXT at the moment, when the | |
1da177e4 LT |
950 | * segment was retransmitted. |
951 | * 4. D-SACK added new rule: D-SACK changes any tag to S. | |
952 | * | |
953 | * It is pleasant to note, that state diagram turns out to be commutative, | |
954 | * so that we are allowed not to be bothered by order of our actions, | |
955 | * when multiple events arrive simultaneously. (see the function below). | |
956 | * | |
957 | * Reordering detection. | |
958 | * -------------------- | |
959 | * Reordering metric is maximal distance, which a packet can be displaced | |
960 | * in packet stream. With SACKs we can estimate it: | |
961 | * | |
962 | * 1. SACK fills old hole and the corresponding segment was not | |
963 | * ever retransmitted -> reordering. Alas, we cannot use it | |
964 | * when segment was retransmitted. | |
965 | * 2. The last flaw is solved with D-SACK. D-SACK arrives | |
966 | * for retransmitted and already SACKed segment -> reordering.. | |
967 | * Both of these heuristics are not used in Loss state, when we cannot | |
968 | * account for retransmits accurately. | |
5b3c9882 IJ |
969 | * |
970 | * SACK block validation. | |
971 | * ---------------------- | |
972 | * | |
973 | * SACK block range validation checks that the received SACK block fits to | |
974 | * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. | |
975 | * Note that SND.UNA is not included to the range though being valid because | |
0e835331 IJ |
976 | * it means that the receiver is rather inconsistent with itself reporting |
977 | * SACK reneging when it should advance SND.UNA. Such SACK block this is | |
978 | * perfectly valid, however, in light of RFC2018 which explicitly states | |
979 | * that "SACK block MUST reflect the newest segment. Even if the newest | |
980 | * segment is going to be discarded ...", not that it looks very clever | |
981 | * in case of head skb. Due to potentional receiver driven attacks, we | |
982 | * choose to avoid immediate execution of a walk in write queue due to | |
983 | * reneging and defer head skb's loss recovery to standard loss recovery | |
984 | * procedure that will eventually trigger (nothing forbids us doing this). | |
5b3c9882 IJ |
985 | * |
986 | * Implements also blockage to start_seq wrap-around. Problem lies in the | |
987 | * fact that though start_seq (s) is before end_seq (i.e., not reversed), | |
988 | * there's no guarantee that it will be before snd_nxt (n). The problem | |
989 | * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt | |
990 | * wrap (s_w): | |
991 | * | |
992 | * <- outs wnd -> <- wrapzone -> | |
993 | * u e n u_w e_w s n_w | |
994 | * | | | | | | | | |
995 | * |<------------+------+----- TCP seqno space --------------+---------->| | |
996 | * ...-- <2^31 ->| |<--------... | |
997 | * ...---- >2^31 ------>| |<--------... | |
998 | * | |
999 | * Current code wouldn't be vulnerable but it's better still to discard such | |
1000 | * crazy SACK blocks. Doing this check for start_seq alone closes somewhat | |
1001 | * similar case (end_seq after snd_nxt wrap) as earlier reversed check in | |
1002 | * snd_nxt wrap -> snd_una region will then become "well defined", i.e., | |
1003 | * equal to the ideal case (infinite seqno space without wrap caused issues). | |
1004 | * | |
1005 | * With D-SACK the lower bound is extended to cover sequence space below | |
1006 | * SND.UNA down to undo_marker, which is the last point of interest. Yet | |
564262c1 | 1007 | * again, D-SACK block must not to go across snd_una (for the same reason as |
5b3c9882 IJ |
1008 | * for the normal SACK blocks, explained above). But there all simplicity |
1009 | * ends, TCP might receive valid D-SACKs below that. As long as they reside | |
1010 | * fully below undo_marker they do not affect behavior in anyway and can | |
1011 | * therefore be safely ignored. In rare cases (which are more or less | |
1012 | * theoretical ones), the D-SACK will nicely cross that boundary due to skb | |
1013 | * fragmentation and packet reordering past skb's retransmission. To consider | |
1014 | * them correctly, the acceptable range must be extended even more though | |
1015 | * the exact amount is rather hard to quantify. However, tp->max_window can | |
1016 | * be used as an exaggerated estimate. | |
1da177e4 | 1017 | */ |
a2a385d6 ED |
1018 | static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, |
1019 | u32 start_seq, u32 end_seq) | |
5b3c9882 IJ |
1020 | { |
1021 | /* Too far in future, or reversed (interpretation is ambiguous) */ | |
1022 | if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) | |
a2a385d6 | 1023 | return false; |
5b3c9882 IJ |
1024 | |
1025 | /* Nasty start_seq wrap-around check (see comments above) */ | |
1026 | if (!before(start_seq, tp->snd_nxt)) | |
a2a385d6 | 1027 | return false; |
5b3c9882 | 1028 | |
564262c1 | 1029 | /* In outstanding window? ...This is valid exit for D-SACKs too. |
5b3c9882 IJ |
1030 | * start_seq == snd_una is non-sensical (see comments above) |
1031 | */ | |
1032 | if (after(start_seq, tp->snd_una)) | |
a2a385d6 | 1033 | return true; |
5b3c9882 IJ |
1034 | |
1035 | if (!is_dsack || !tp->undo_marker) | |
a2a385d6 | 1036 | return false; |
5b3c9882 IJ |
1037 | |
1038 | /* ...Then it's D-SACK, and must reside below snd_una completely */ | |
f779b2d6 | 1039 | if (after(end_seq, tp->snd_una)) |
a2a385d6 | 1040 | return false; |
5b3c9882 IJ |
1041 | |
1042 | if (!before(start_seq, tp->undo_marker)) | |
a2a385d6 | 1043 | return true; |
5b3c9882 IJ |
1044 | |
1045 | /* Too old */ | |
1046 | if (!after(end_seq, tp->undo_marker)) | |
a2a385d6 | 1047 | return false; |
5b3c9882 IJ |
1048 | |
1049 | /* Undo_marker boundary crossing (overestimates a lot). Known already: | |
1050 | * start_seq < undo_marker and end_seq >= undo_marker. | |
1051 | */ | |
1052 | return !before(start_seq, end_seq - tp->max_window); | |
1053 | } | |
1054 | ||
a2a385d6 ED |
1055 | static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, |
1056 | struct tcp_sack_block_wire *sp, int num_sacks, | |
1057 | u32 prior_snd_una) | |
d06e021d | 1058 | { |
1ed83465 | 1059 | struct tcp_sock *tp = tcp_sk(sk); |
d3e2ce3b HH |
1060 | u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); |
1061 | u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); | |
a2a385d6 | 1062 | bool dup_sack = false; |
d06e021d DM |
1063 | |
1064 | if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { | |
a2a385d6 | 1065 | dup_sack = true; |
e60402d0 | 1066 | tcp_dsack_seen(tp); |
de0744af | 1067 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV); |
d06e021d | 1068 | } else if (num_sacks > 1) { |
d3e2ce3b HH |
1069 | u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); |
1070 | u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); | |
d06e021d DM |
1071 | |
1072 | if (!after(end_seq_0, end_seq_1) && | |
1073 | !before(start_seq_0, start_seq_1)) { | |
a2a385d6 | 1074 | dup_sack = true; |
e60402d0 | 1075 | tcp_dsack_seen(tp); |
de0744af PE |
1076 | NET_INC_STATS_BH(sock_net(sk), |
1077 | LINUX_MIB_TCPDSACKOFORECV); | |
d06e021d DM |
1078 | } |
1079 | } | |
1080 | ||
1081 | /* D-SACK for already forgotten data... Do dumb counting. */ | |
6e08d5e3 | 1082 | if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 && |
d06e021d DM |
1083 | !after(end_seq_0, prior_snd_una) && |
1084 | after(end_seq_0, tp->undo_marker)) | |
1085 | tp->undo_retrans--; | |
1086 | ||
1087 | return dup_sack; | |
1088 | } | |
1089 | ||
a1197f5a | 1090 | struct tcp_sacktag_state { |
740b0f18 ED |
1091 | int reord; |
1092 | int fack_count; | |
31231a8a KKJ |
1093 | /* Timestamps for earliest and latest never-retransmitted segment |
1094 | * that was SACKed. RTO needs the earliest RTT to stay conservative, | |
1095 | * but congestion control should still get an accurate delay signal. | |
1096 | */ | |
1097 | struct skb_mstamp first_sackt; | |
1098 | struct skb_mstamp last_sackt; | |
740b0f18 | 1099 | int flag; |
a1197f5a IJ |
1100 | }; |
1101 | ||
d1935942 IJ |
1102 | /* Check if skb is fully within the SACK block. In presence of GSO skbs, |
1103 | * the incoming SACK may not exactly match but we can find smaller MSS | |
1104 | * aligned portion of it that matches. Therefore we might need to fragment | |
1105 | * which may fail and creates some hassle (caller must handle error case | |
1106 | * returns). | |
832d11c5 IJ |
1107 | * |
1108 | * FIXME: this could be merged to shift decision code | |
d1935942 | 1109 | */ |
0f79efdc | 1110 | static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, |
a2a385d6 | 1111 | u32 start_seq, u32 end_seq) |
d1935942 | 1112 | { |
a2a385d6 ED |
1113 | int err; |
1114 | bool in_sack; | |
d1935942 | 1115 | unsigned int pkt_len; |
adb92db8 | 1116 | unsigned int mss; |
d1935942 IJ |
1117 | |
1118 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1119 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1120 | ||
1121 | if (tcp_skb_pcount(skb) > 1 && !in_sack && | |
1122 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { | |
adb92db8 | 1123 | mss = tcp_skb_mss(skb); |
d1935942 IJ |
1124 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); |
1125 | ||
adb92db8 | 1126 | if (!in_sack) { |
d1935942 | 1127 | pkt_len = start_seq - TCP_SKB_CB(skb)->seq; |
adb92db8 IJ |
1128 | if (pkt_len < mss) |
1129 | pkt_len = mss; | |
1130 | } else { | |
d1935942 | 1131 | pkt_len = end_seq - TCP_SKB_CB(skb)->seq; |
adb92db8 IJ |
1132 | if (pkt_len < mss) |
1133 | return -EINVAL; | |
1134 | } | |
1135 | ||
1136 | /* Round if necessary so that SACKs cover only full MSSes | |
1137 | * and/or the remaining small portion (if present) | |
1138 | */ | |
1139 | if (pkt_len > mss) { | |
1140 | unsigned int new_len = (pkt_len / mss) * mss; | |
1141 | if (!in_sack && new_len < pkt_len) { | |
1142 | new_len += mss; | |
2cd0d743 | 1143 | if (new_len >= skb->len) |
adb92db8 IJ |
1144 | return 0; |
1145 | } | |
1146 | pkt_len = new_len; | |
1147 | } | |
6cc55e09 | 1148 | err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC); |
d1935942 IJ |
1149 | if (err < 0) |
1150 | return err; | |
1151 | } | |
1152 | ||
1153 | return in_sack; | |
1154 | } | |
1155 | ||
cc9a672e NC |
1156 | /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ |
1157 | static u8 tcp_sacktag_one(struct sock *sk, | |
1158 | struct tcp_sacktag_state *state, u8 sacked, | |
1159 | u32 start_seq, u32 end_seq, | |
740b0f18 ED |
1160 | int dup_sack, int pcount, |
1161 | const struct skb_mstamp *xmit_time) | |
9e10c47c | 1162 | { |
6859d494 | 1163 | struct tcp_sock *tp = tcp_sk(sk); |
a1197f5a | 1164 | int fack_count = state->fack_count; |
9e10c47c IJ |
1165 | |
1166 | /* Account D-SACK for retransmitted packet. */ | |
1167 | if (dup_sack && (sacked & TCPCB_RETRANS)) { | |
6e08d5e3 | 1168 | if (tp->undo_marker && tp->undo_retrans > 0 && |
cc9a672e | 1169 | after(end_seq, tp->undo_marker)) |
9e10c47c | 1170 | tp->undo_retrans--; |
ede9f3b1 | 1171 | if (sacked & TCPCB_SACKED_ACKED) |
a1197f5a | 1172 | state->reord = min(fack_count, state->reord); |
9e10c47c IJ |
1173 | } |
1174 | ||
1175 | /* Nothing to do; acked frame is about to be dropped (was ACKed). */ | |
cc9a672e | 1176 | if (!after(end_seq, tp->snd_una)) |
a1197f5a | 1177 | return sacked; |
9e10c47c IJ |
1178 | |
1179 | if (!(sacked & TCPCB_SACKED_ACKED)) { | |
659a8ad5 YC |
1180 | tcp_rack_advance(tp, xmit_time, sacked); |
1181 | ||
9e10c47c IJ |
1182 | if (sacked & TCPCB_SACKED_RETRANS) { |
1183 | /* If the segment is not tagged as lost, | |
1184 | * we do not clear RETRANS, believing | |
1185 | * that retransmission is still in flight. | |
1186 | */ | |
1187 | if (sacked & TCPCB_LOST) { | |
a1197f5a | 1188 | sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); |
f58b22fd IJ |
1189 | tp->lost_out -= pcount; |
1190 | tp->retrans_out -= pcount; | |
9e10c47c IJ |
1191 | } |
1192 | } else { | |
1193 | if (!(sacked & TCPCB_RETRANS)) { | |
1194 | /* New sack for not retransmitted frame, | |
1195 | * which was in hole. It is reordering. | |
1196 | */ | |
cc9a672e | 1197 | if (before(start_seq, |
9e10c47c | 1198 | tcp_highest_sack_seq(tp))) |
a1197f5a IJ |
1199 | state->reord = min(fack_count, |
1200 | state->reord); | |
e33099f9 YC |
1201 | if (!after(end_seq, tp->high_seq)) |
1202 | state->flag |= FLAG_ORIG_SACK_ACKED; | |
31231a8a KKJ |
1203 | if (state->first_sackt.v64 == 0) |
1204 | state->first_sackt = *xmit_time; | |
1205 | state->last_sackt = *xmit_time; | |
9e10c47c IJ |
1206 | } |
1207 | ||
1208 | if (sacked & TCPCB_LOST) { | |
a1197f5a | 1209 | sacked &= ~TCPCB_LOST; |
f58b22fd | 1210 | tp->lost_out -= pcount; |
9e10c47c IJ |
1211 | } |
1212 | } | |
1213 | ||
a1197f5a IJ |
1214 | sacked |= TCPCB_SACKED_ACKED; |
1215 | state->flag |= FLAG_DATA_SACKED; | |
f58b22fd | 1216 | tp->sacked_out += pcount; |
ddf1af6f | 1217 | tp->delivered += pcount; /* Out-of-order packets delivered */ |
9e10c47c | 1218 | |
f58b22fd | 1219 | fack_count += pcount; |
9e10c47c IJ |
1220 | |
1221 | /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ | |
00db4124 | 1222 | if (!tcp_is_fack(tp) && tp->lost_skb_hint && |
cc9a672e | 1223 | before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) |
f58b22fd | 1224 | tp->lost_cnt_hint += pcount; |
9e10c47c IJ |
1225 | |
1226 | if (fack_count > tp->fackets_out) | |
1227 | tp->fackets_out = fack_count; | |
9e10c47c IJ |
1228 | } |
1229 | ||
1230 | /* D-SACK. We can detect redundant retransmission in S|R and plain R | |
1231 | * frames and clear it. undo_retrans is decreased above, L|R frames | |
1232 | * are accounted above as well. | |
1233 | */ | |
a1197f5a IJ |
1234 | if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { |
1235 | sacked &= ~TCPCB_SACKED_RETRANS; | |
f58b22fd | 1236 | tp->retrans_out -= pcount; |
9e10c47c IJ |
1237 | } |
1238 | ||
a1197f5a | 1239 | return sacked; |
9e10c47c IJ |
1240 | } |
1241 | ||
daef52ba NC |
1242 | /* Shift newly-SACKed bytes from this skb to the immediately previous |
1243 | * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. | |
1244 | */ | |
a2a385d6 ED |
1245 | static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, |
1246 | struct tcp_sacktag_state *state, | |
1247 | unsigned int pcount, int shifted, int mss, | |
1248 | bool dup_sack) | |
832d11c5 IJ |
1249 | { |
1250 | struct tcp_sock *tp = tcp_sk(sk); | |
50133161 | 1251 | struct sk_buff *prev = tcp_write_queue_prev(sk, skb); |
daef52ba NC |
1252 | u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ |
1253 | u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ | |
832d11c5 IJ |
1254 | |
1255 | BUG_ON(!pcount); | |
1256 | ||
4c90d3b3 NC |
1257 | /* Adjust counters and hints for the newly sacked sequence |
1258 | * range but discard the return value since prev is already | |
1259 | * marked. We must tag the range first because the seq | |
1260 | * advancement below implicitly advances | |
1261 | * tcp_highest_sack_seq() when skb is highest_sack. | |
1262 | */ | |
1263 | tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, | |
59c9af42 | 1264 | start_seq, end_seq, dup_sack, pcount, |
740b0f18 | 1265 | &skb->skb_mstamp); |
4c90d3b3 NC |
1266 | |
1267 | if (skb == tp->lost_skb_hint) | |
0af2a0d0 NC |
1268 | tp->lost_cnt_hint += pcount; |
1269 | ||
832d11c5 IJ |
1270 | TCP_SKB_CB(prev)->end_seq += shifted; |
1271 | TCP_SKB_CB(skb)->seq += shifted; | |
1272 | ||
cd7d8498 ED |
1273 | tcp_skb_pcount_add(prev, pcount); |
1274 | BUG_ON(tcp_skb_pcount(skb) < pcount); | |
1275 | tcp_skb_pcount_add(skb, -pcount); | |
832d11c5 IJ |
1276 | |
1277 | /* When we're adding to gso_segs == 1, gso_size will be zero, | |
1278 | * in theory this shouldn't be necessary but as long as DSACK | |
1279 | * code can come after this skb later on it's better to keep | |
1280 | * setting gso_size to something. | |
1281 | */ | |
f69ad292 ED |
1282 | if (!TCP_SKB_CB(prev)->tcp_gso_size) |
1283 | TCP_SKB_CB(prev)->tcp_gso_size = mss; | |
832d11c5 IJ |
1284 | |
1285 | /* CHECKME: To clear or not to clear? Mimics normal skb currently */ | |
51466a75 | 1286 | if (tcp_skb_pcount(skb) <= 1) |
f69ad292 | 1287 | TCP_SKB_CB(skb)->tcp_gso_size = 0; |
832d11c5 | 1288 | |
832d11c5 IJ |
1289 | /* Difference in this won't matter, both ACKed by the same cumul. ACK */ |
1290 | TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); | |
1291 | ||
832d11c5 IJ |
1292 | if (skb->len > 0) { |
1293 | BUG_ON(!tcp_skb_pcount(skb)); | |
111cc8b9 | 1294 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED); |
a2a385d6 | 1295 | return false; |
832d11c5 IJ |
1296 | } |
1297 | ||
1298 | /* Whole SKB was eaten :-) */ | |
1299 | ||
92ee76b6 IJ |
1300 | if (skb == tp->retransmit_skb_hint) |
1301 | tp->retransmit_skb_hint = prev; | |
92ee76b6 IJ |
1302 | if (skb == tp->lost_skb_hint) { |
1303 | tp->lost_skb_hint = prev; | |
1304 | tp->lost_cnt_hint -= tcp_skb_pcount(prev); | |
1305 | } | |
1306 | ||
5e8a402f ED |
1307 | TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
1308 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) | |
1309 | TCP_SKB_CB(prev)->end_seq++; | |
1310 | ||
832d11c5 IJ |
1311 | if (skb == tcp_highest_sack(sk)) |
1312 | tcp_advance_highest_sack(sk, skb); | |
1313 | ||
1314 | tcp_unlink_write_queue(skb, sk); | |
1315 | sk_wmem_free_skb(sk, skb); | |
1316 | ||
111cc8b9 IJ |
1317 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED); |
1318 | ||
a2a385d6 | 1319 | return true; |
832d11c5 IJ |
1320 | } |
1321 | ||
1322 | /* I wish gso_size would have a bit more sane initialization than | |
1323 | * something-or-zero which complicates things | |
1324 | */ | |
cf533ea5 | 1325 | static int tcp_skb_seglen(const struct sk_buff *skb) |
832d11c5 | 1326 | { |
775ffabf | 1327 | return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); |
832d11c5 IJ |
1328 | } |
1329 | ||
1330 | /* Shifting pages past head area doesn't work */ | |
cf533ea5 | 1331 | static int skb_can_shift(const struct sk_buff *skb) |
832d11c5 IJ |
1332 | { |
1333 | return !skb_headlen(skb) && skb_is_nonlinear(skb); | |
1334 | } | |
1335 | ||
1336 | /* Try collapsing SACK blocks spanning across multiple skbs to a single | |
1337 | * skb. | |
1338 | */ | |
1339 | static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, | |
a1197f5a | 1340 | struct tcp_sacktag_state *state, |
832d11c5 | 1341 | u32 start_seq, u32 end_seq, |
a2a385d6 | 1342 | bool dup_sack) |
832d11c5 IJ |
1343 | { |
1344 | struct tcp_sock *tp = tcp_sk(sk); | |
1345 | struct sk_buff *prev; | |
1346 | int mss; | |
1347 | int pcount = 0; | |
1348 | int len; | |
1349 | int in_sack; | |
1350 | ||
1351 | if (!sk_can_gso(sk)) | |
1352 | goto fallback; | |
1353 | ||
1354 | /* Normally R but no L won't result in plain S */ | |
1355 | if (!dup_sack && | |
9969ca5f | 1356 | (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) |
832d11c5 IJ |
1357 | goto fallback; |
1358 | if (!skb_can_shift(skb)) | |
1359 | goto fallback; | |
1360 | /* This frame is about to be dropped (was ACKed). */ | |
1361 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1362 | goto fallback; | |
1363 | ||
1364 | /* Can only happen with delayed DSACK + discard craziness */ | |
1365 | if (unlikely(skb == tcp_write_queue_head(sk))) | |
1366 | goto fallback; | |
1367 | prev = tcp_write_queue_prev(sk, skb); | |
1368 | ||
1369 | if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) | |
1370 | goto fallback; | |
1371 | ||
1372 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1373 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1374 | ||
1375 | if (in_sack) { | |
1376 | len = skb->len; | |
1377 | pcount = tcp_skb_pcount(skb); | |
775ffabf | 1378 | mss = tcp_skb_seglen(skb); |
832d11c5 IJ |
1379 | |
1380 | /* TODO: Fix DSACKs to not fragment already SACKed and we can | |
1381 | * drop this restriction as unnecessary | |
1382 | */ | |
775ffabf | 1383 | if (mss != tcp_skb_seglen(prev)) |
832d11c5 IJ |
1384 | goto fallback; |
1385 | } else { | |
1386 | if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) | |
1387 | goto noop; | |
1388 | /* CHECKME: This is non-MSS split case only?, this will | |
1389 | * cause skipped skbs due to advancing loop btw, original | |
1390 | * has that feature too | |
1391 | */ | |
1392 | if (tcp_skb_pcount(skb) <= 1) | |
1393 | goto noop; | |
1394 | ||
1395 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); | |
1396 | if (!in_sack) { | |
1397 | /* TODO: head merge to next could be attempted here | |
1398 | * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), | |
1399 | * though it might not be worth of the additional hassle | |
1400 | * | |
1401 | * ...we can probably just fallback to what was done | |
1402 | * previously. We could try merging non-SACKed ones | |
1403 | * as well but it probably isn't going to buy off | |
1404 | * because later SACKs might again split them, and | |
1405 | * it would make skb timestamp tracking considerably | |
1406 | * harder problem. | |
1407 | */ | |
1408 | goto fallback; | |
1409 | } | |
1410 | ||
1411 | len = end_seq - TCP_SKB_CB(skb)->seq; | |
1412 | BUG_ON(len < 0); | |
1413 | BUG_ON(len > skb->len); | |
1414 | ||
1415 | /* MSS boundaries should be honoured or else pcount will | |
1416 | * severely break even though it makes things bit trickier. | |
1417 | * Optimize common case to avoid most of the divides | |
1418 | */ | |
1419 | mss = tcp_skb_mss(skb); | |
1420 | ||
1421 | /* TODO: Fix DSACKs to not fragment already SACKed and we can | |
1422 | * drop this restriction as unnecessary | |
1423 | */ | |
775ffabf | 1424 | if (mss != tcp_skb_seglen(prev)) |
832d11c5 IJ |
1425 | goto fallback; |
1426 | ||
1427 | if (len == mss) { | |
1428 | pcount = 1; | |
1429 | } else if (len < mss) { | |
1430 | goto noop; | |
1431 | } else { | |
1432 | pcount = len / mss; | |
1433 | len = pcount * mss; | |
1434 | } | |
1435 | } | |
1436 | ||
4648dc97 NC |
1437 | /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ |
1438 | if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) | |
1439 | goto fallback; | |
1440 | ||
832d11c5 IJ |
1441 | if (!skb_shift(prev, skb, len)) |
1442 | goto fallback; | |
9ec06ff5 | 1443 | if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack)) |
832d11c5 IJ |
1444 | goto out; |
1445 | ||
1446 | /* Hole filled allows collapsing with the next as well, this is very | |
1447 | * useful when hole on every nth skb pattern happens | |
1448 | */ | |
1449 | if (prev == tcp_write_queue_tail(sk)) | |
1450 | goto out; | |
1451 | skb = tcp_write_queue_next(sk, prev); | |
1452 | ||
f0bc52f3 IJ |
1453 | if (!skb_can_shift(skb) || |
1454 | (skb == tcp_send_head(sk)) || | |
1455 | ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || | |
775ffabf | 1456 | (mss != tcp_skb_seglen(skb))) |
832d11c5 IJ |
1457 | goto out; |
1458 | ||
1459 | len = skb->len; | |
1460 | if (skb_shift(prev, skb, len)) { | |
1461 | pcount += tcp_skb_pcount(skb); | |
9ec06ff5 | 1462 | tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0); |
832d11c5 IJ |
1463 | } |
1464 | ||
1465 | out: | |
a1197f5a | 1466 | state->fack_count += pcount; |
832d11c5 IJ |
1467 | return prev; |
1468 | ||
1469 | noop: | |
1470 | return skb; | |
1471 | ||
1472 | fallback: | |
111cc8b9 | 1473 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); |
832d11c5 IJ |
1474 | return NULL; |
1475 | } | |
1476 | ||
68f8353b IJ |
1477 | static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, |
1478 | struct tcp_sack_block *next_dup, | |
a1197f5a | 1479 | struct tcp_sacktag_state *state, |
68f8353b | 1480 | u32 start_seq, u32 end_seq, |
a2a385d6 | 1481 | bool dup_sack_in) |
68f8353b | 1482 | { |
832d11c5 IJ |
1483 | struct tcp_sock *tp = tcp_sk(sk); |
1484 | struct sk_buff *tmp; | |
1485 | ||
68f8353b IJ |
1486 | tcp_for_write_queue_from(skb, sk) { |
1487 | int in_sack = 0; | |
a2a385d6 | 1488 | bool dup_sack = dup_sack_in; |
68f8353b IJ |
1489 | |
1490 | if (skb == tcp_send_head(sk)) | |
1491 | break; | |
1492 | ||
1493 | /* queue is in-order => we can short-circuit the walk early */ | |
1494 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) | |
1495 | break; | |
1496 | ||
00db4124 | 1497 | if (next_dup && |
68f8353b IJ |
1498 | before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { |
1499 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1500 | next_dup->start_seq, | |
1501 | next_dup->end_seq); | |
1502 | if (in_sack > 0) | |
a2a385d6 | 1503 | dup_sack = true; |
68f8353b IJ |
1504 | } |
1505 | ||
832d11c5 IJ |
1506 | /* skb reference here is a bit tricky to get right, since |
1507 | * shifting can eat and free both this skb and the next, | |
1508 | * so not even _safe variant of the loop is enough. | |
1509 | */ | |
1510 | if (in_sack <= 0) { | |
a1197f5a IJ |
1511 | tmp = tcp_shift_skb_data(sk, skb, state, |
1512 | start_seq, end_seq, dup_sack); | |
00db4124 | 1513 | if (tmp) { |
832d11c5 IJ |
1514 | if (tmp != skb) { |
1515 | skb = tmp; | |
1516 | continue; | |
1517 | } | |
1518 | ||
1519 | in_sack = 0; | |
1520 | } else { | |
1521 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1522 | start_seq, | |
1523 | end_seq); | |
1524 | } | |
1525 | } | |
1526 | ||
68f8353b IJ |
1527 | if (unlikely(in_sack < 0)) |
1528 | break; | |
1529 | ||
832d11c5 | 1530 | if (in_sack) { |
cc9a672e NC |
1531 | TCP_SKB_CB(skb)->sacked = |
1532 | tcp_sacktag_one(sk, | |
1533 | state, | |
1534 | TCP_SKB_CB(skb)->sacked, | |
1535 | TCP_SKB_CB(skb)->seq, | |
1536 | TCP_SKB_CB(skb)->end_seq, | |
1537 | dup_sack, | |
59c9af42 | 1538 | tcp_skb_pcount(skb), |
740b0f18 | 1539 | &skb->skb_mstamp); |
68f8353b | 1540 | |
832d11c5 IJ |
1541 | if (!before(TCP_SKB_CB(skb)->seq, |
1542 | tcp_highest_sack_seq(tp))) | |
1543 | tcp_advance_highest_sack(sk, skb); | |
1544 | } | |
1545 | ||
a1197f5a | 1546 | state->fack_count += tcp_skb_pcount(skb); |
68f8353b IJ |
1547 | } |
1548 | return skb; | |
1549 | } | |
1550 | ||
1551 | /* Avoid all extra work that is being done by sacktag while walking in | |
1552 | * a normal way | |
1553 | */ | |
1554 | static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, | |
a1197f5a IJ |
1555 | struct tcp_sacktag_state *state, |
1556 | u32 skip_to_seq) | |
68f8353b IJ |
1557 | { |
1558 | tcp_for_write_queue_from(skb, sk) { | |
1559 | if (skb == tcp_send_head(sk)) | |
1560 | break; | |
1561 | ||
e8bae275 | 1562 | if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) |
68f8353b | 1563 | break; |
d152a7d8 | 1564 | |
a1197f5a | 1565 | state->fack_count += tcp_skb_pcount(skb); |
68f8353b IJ |
1566 | } |
1567 | return skb; | |
1568 | } | |
1569 | ||
1570 | static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, | |
1571 | struct sock *sk, | |
1572 | struct tcp_sack_block *next_dup, | |
a1197f5a IJ |
1573 | struct tcp_sacktag_state *state, |
1574 | u32 skip_to_seq) | |
68f8353b | 1575 | { |
51456b29 | 1576 | if (!next_dup) |
68f8353b IJ |
1577 | return skb; |
1578 | ||
1579 | if (before(next_dup->start_seq, skip_to_seq)) { | |
a1197f5a IJ |
1580 | skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); |
1581 | skb = tcp_sacktag_walk(skb, sk, NULL, state, | |
1582 | next_dup->start_seq, next_dup->end_seq, | |
1583 | 1); | |
68f8353b IJ |
1584 | } |
1585 | ||
1586 | return skb; | |
1587 | } | |
1588 | ||
cf533ea5 | 1589 | static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) |
68f8353b IJ |
1590 | { |
1591 | return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); | |
1592 | } | |
1593 | ||
1da177e4 | 1594 | static int |
cf533ea5 | 1595 | tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, |
196da974 | 1596 | u32 prior_snd_una, struct tcp_sacktag_state *state) |
1da177e4 LT |
1597 | { |
1598 | struct tcp_sock *tp = tcp_sk(sk); | |
cf533ea5 ED |
1599 | const unsigned char *ptr = (skb_transport_header(ack_skb) + |
1600 | TCP_SKB_CB(ack_skb)->sacked); | |
fd6dad61 | 1601 | struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); |
4389dded | 1602 | struct tcp_sack_block sp[TCP_NUM_SACKS]; |
68f8353b IJ |
1603 | struct tcp_sack_block *cache; |
1604 | struct sk_buff *skb; | |
4389dded | 1605 | int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); |
fd6dad61 | 1606 | int used_sacks; |
a2a385d6 | 1607 | bool found_dup_sack = false; |
68f8353b | 1608 | int i, j; |
fda03fbb | 1609 | int first_sack_index; |
1da177e4 | 1610 | |
196da974 KKJ |
1611 | state->flag = 0; |
1612 | state->reord = tp->packets_out; | |
a1197f5a | 1613 | |
d738cd8f | 1614 | if (!tp->sacked_out) { |
de83c058 IJ |
1615 | if (WARN_ON(tp->fackets_out)) |
1616 | tp->fackets_out = 0; | |
6859d494 | 1617 | tcp_highest_sack_reset(sk); |
d738cd8f | 1618 | } |
1da177e4 | 1619 | |
1ed83465 | 1620 | found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, |
d06e021d DM |
1621 | num_sacks, prior_snd_una); |
1622 | if (found_dup_sack) | |
196da974 | 1623 | state->flag |= FLAG_DSACKING_ACK; |
6f74651a BE |
1624 | |
1625 | /* Eliminate too old ACKs, but take into | |
1626 | * account more or less fresh ones, they can | |
1627 | * contain valid SACK info. | |
1628 | */ | |
1629 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) | |
1630 | return 0; | |
1631 |