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; | |
4bc2f18b | 84 | EXPORT_SYMBOL(sysctl_tcp_reordering); |
ab32ea5d BH |
85 | int sysctl_tcp_dsack __read_mostly = 1; |
86 | int sysctl_tcp_app_win __read_mostly = 31; | |
b49960a0 | 87 | int sysctl_tcp_adv_win_scale __read_mostly = 1; |
4bc2f18b | 88 | EXPORT_SYMBOL(sysctl_tcp_adv_win_scale); |
1da177e4 | 89 | |
282f23c6 ED |
90 | /* rfc5961 challenge ack rate limiting */ |
91 | int sysctl_tcp_challenge_ack_limit = 100; | |
92 | ||
ab32ea5d BH |
93 | int sysctl_tcp_stdurg __read_mostly; |
94 | int sysctl_tcp_rfc1337 __read_mostly; | |
95 | int sysctl_tcp_max_orphans __read_mostly = NR_FILE; | |
c96fd3d4 | 96 | int sysctl_tcp_frto __read_mostly = 2; |
1da177e4 | 97 | |
7e380175 AP |
98 | int sysctl_tcp_thin_dupack __read_mostly; |
99 | ||
ab32ea5d | 100 | int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; |
6ba8a3b1 | 101 | int sysctl_tcp_early_retrans __read_mostly = 3; |
1da177e4 | 102 | |
1da177e4 LT |
103 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ |
104 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ | |
105 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ | |
106 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ | |
107 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ | |
108 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ | |
109 | #define FLAG_ECE 0x40 /* ECE in this ACK */ | |
1da177e4 | 110 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ |
e33099f9 | 111 | #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */ |
2e605294 | 112 | #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ |
564262c1 | 113 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ |
cadbd031 | 114 | #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ |
12fb3dd9 | 115 | #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */ |
1da177e4 LT |
116 | |
117 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) | |
118 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) | |
119 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) | |
120 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) | |
121 | ||
1da177e4 | 122 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) |
bdf1ee5d | 123 | #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) |
1da177e4 | 124 | |
e905a9ed | 125 | /* Adapt the MSS value used to make delayed ack decision to the |
1da177e4 | 126 | * real world. |
e905a9ed | 127 | */ |
056834d9 | 128 | static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 129 | { |
463c84b9 | 130 | struct inet_connection_sock *icsk = inet_csk(sk); |
e905a9ed | 131 | const unsigned int lss = icsk->icsk_ack.last_seg_size; |
463c84b9 | 132 | unsigned int len; |
1da177e4 | 133 | |
e905a9ed | 134 | icsk->icsk_ack.last_seg_size = 0; |
1da177e4 LT |
135 | |
136 | /* skb->len may jitter because of SACKs, even if peer | |
137 | * sends good full-sized frames. | |
138 | */ | |
056834d9 | 139 | len = skb_shinfo(skb)->gso_size ? : skb->len; |
463c84b9 ACM |
140 | if (len >= icsk->icsk_ack.rcv_mss) { |
141 | icsk->icsk_ack.rcv_mss = len; | |
1da177e4 LT |
142 | } else { |
143 | /* Otherwise, we make more careful check taking into account, | |
144 | * that SACKs block is variable. | |
145 | * | |
146 | * "len" is invariant segment length, including TCP header. | |
147 | */ | |
9c70220b | 148 | len += skb->data - skb_transport_header(skb); |
bee7ca9e | 149 | if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || |
1da177e4 LT |
150 | /* If PSH is not set, packet should be |
151 | * full sized, provided peer TCP is not badly broken. | |
152 | * This observation (if it is correct 8)) allows | |
153 | * to handle super-low mtu links fairly. | |
154 | */ | |
155 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && | |
aa8223c7 | 156 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { |
1da177e4 LT |
157 | /* Subtract also invariant (if peer is RFC compliant), |
158 | * tcp header plus fixed timestamp option length. | |
159 | * Resulting "len" is MSS free of SACK jitter. | |
160 | */ | |
463c84b9 ACM |
161 | len -= tcp_sk(sk)->tcp_header_len; |
162 | icsk->icsk_ack.last_seg_size = len; | |
1da177e4 | 163 | if (len == lss) { |
463c84b9 | 164 | icsk->icsk_ack.rcv_mss = len; |
1da177e4 LT |
165 | return; |
166 | } | |
167 | } | |
1ef9696c AK |
168 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) |
169 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; | |
463c84b9 | 170 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
1da177e4 LT |
171 | } |
172 | } | |
173 | ||
463c84b9 | 174 | static void tcp_incr_quickack(struct sock *sk) |
1da177e4 | 175 | { |
463c84b9 | 176 | struct inet_connection_sock *icsk = inet_csk(sk); |
95c96174 | 177 | unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); |
1da177e4 | 178 | |
056834d9 IJ |
179 | if (quickacks == 0) |
180 | quickacks = 2; | |
463c84b9 ACM |
181 | if (quickacks > icsk->icsk_ack.quick) |
182 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); | |
1da177e4 LT |
183 | } |
184 | ||
1b9f4092 | 185 | static void tcp_enter_quickack_mode(struct sock *sk) |
1da177e4 | 186 | { |
463c84b9 ACM |
187 | struct inet_connection_sock *icsk = inet_csk(sk); |
188 | tcp_incr_quickack(sk); | |
189 | icsk->icsk_ack.pingpong = 0; | |
190 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 LT |
191 | } |
192 | ||
193 | /* Send ACKs quickly, if "quick" count is not exhausted | |
194 | * and the session is not interactive. | |
195 | */ | |
196 | ||
a2a385d6 | 197 | static inline bool tcp_in_quickack_mode(const struct sock *sk) |
1da177e4 | 198 | { |
463c84b9 | 199 | const struct inet_connection_sock *icsk = inet_csk(sk); |
a2a385d6 | 200 | |
463c84b9 | 201 | return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; |
1da177e4 LT |
202 | } |
203 | ||
735d3831 | 204 | static void tcp_ecn_queue_cwr(struct tcp_sock *tp) |
bdf1ee5d | 205 | { |
056834d9 | 206 | if (tp->ecn_flags & TCP_ECN_OK) |
bdf1ee5d IJ |
207 | tp->ecn_flags |= TCP_ECN_QUEUE_CWR; |
208 | } | |
209 | ||
735d3831 | 210 | static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb) |
bdf1ee5d IJ |
211 | { |
212 | if (tcp_hdr(skb)->cwr) | |
213 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
214 | } | |
215 | ||
735d3831 | 216 | static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp) |
bdf1ee5d IJ |
217 | { |
218 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
219 | } | |
220 | ||
735d3831 | 221 | static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) |
bdf1ee5d | 222 | { |
b82d1bb4 | 223 | switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { |
7a269ffa | 224 | case INET_ECN_NOT_ECT: |
bdf1ee5d | 225 | /* Funny extension: if ECT is not set on a segment, |
7a269ffa ED |
226 | * and we already seen ECT on a previous segment, |
227 | * it is probably a retransmit. | |
228 | */ | |
229 | if (tp->ecn_flags & TCP_ECN_SEEN) | |
bdf1ee5d | 230 | tcp_enter_quickack_mode((struct sock *)tp); |
7a269ffa ED |
231 | break; |
232 | case INET_ECN_CE: | |
9890092e FW |
233 | if (tcp_ca_needs_ecn((struct sock *)tp)) |
234 | tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE); | |
235 | ||
aae06bf5 ED |
236 | if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { |
237 | /* Better not delay acks, sender can have a very low cwnd */ | |
238 | tcp_enter_quickack_mode((struct sock *)tp); | |
239 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; | |
240 | } | |
9890092e FW |
241 | tp->ecn_flags |= TCP_ECN_SEEN; |
242 | break; | |
7a269ffa | 243 | default: |
9890092e FW |
244 | if (tcp_ca_needs_ecn((struct sock *)tp)) |
245 | tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE); | |
7a269ffa | 246 | tp->ecn_flags |= TCP_ECN_SEEN; |
9890092e | 247 | break; |
bdf1ee5d IJ |
248 | } |
249 | } | |
250 | ||
735d3831 FW |
251 | static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) |
252 | { | |
253 | if (tp->ecn_flags & TCP_ECN_OK) | |
254 | __tcp_ecn_check_ce(tp, skb); | |
255 | } | |
256 | ||
257 | static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) | |
bdf1ee5d | 258 | { |
056834d9 | 259 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) |
bdf1ee5d IJ |
260 | tp->ecn_flags &= ~TCP_ECN_OK; |
261 | } | |
262 | ||
735d3831 | 263 | static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) |
bdf1ee5d | 264 | { |
056834d9 | 265 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) |
bdf1ee5d IJ |
266 | tp->ecn_flags &= ~TCP_ECN_OK; |
267 | } | |
268 | ||
735d3831 | 269 | static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) |
bdf1ee5d | 270 | { |
056834d9 | 271 | if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) |
a2a385d6 ED |
272 | return true; |
273 | return false; | |
bdf1ee5d IJ |
274 | } |
275 | ||
1da177e4 LT |
276 | /* Buffer size and advertised window tuning. |
277 | * | |
278 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. | |
279 | */ | |
280 | ||
6ae70532 | 281 | static void tcp_sndbuf_expand(struct sock *sk) |
1da177e4 | 282 | { |
6ae70532 ED |
283 | const struct tcp_sock *tp = tcp_sk(sk); |
284 | int sndmem, per_mss; | |
285 | u32 nr_segs; | |
286 | ||
287 | /* Worst case is non GSO/TSO : each frame consumes one skb | |
288 | * and skb->head is kmalloced using power of two area of memory | |
289 | */ | |
290 | per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + | |
291 | MAX_TCP_HEADER + | |
292 | SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); | |
293 | ||
294 | per_mss = roundup_pow_of_two(per_mss) + | |
295 | SKB_DATA_ALIGN(sizeof(struct sk_buff)); | |
296 | ||
297 | nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd); | |
298 | nr_segs = max_t(u32, nr_segs, tp->reordering + 1); | |
299 | ||
300 | /* Fast Recovery (RFC 5681 3.2) : | |
301 | * Cubic needs 1.7 factor, rounded to 2 to include | |
302 | * extra cushion (application might react slowly to POLLOUT) | |
303 | */ | |
304 | sndmem = 2 * nr_segs * per_mss; | |
1da177e4 | 305 | |
06a59ecb ED |
306 | if (sk->sk_sndbuf < sndmem) |
307 | sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); | |
1da177e4 LT |
308 | } |
309 | ||
310 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) | |
311 | * | |
312 | * All tcp_full_space() is split to two parts: "network" buffer, allocated | |
313 | * forward and advertised in receiver window (tp->rcv_wnd) and | |
314 | * "application buffer", required to isolate scheduling/application | |
315 | * latencies from network. | |
316 | * window_clamp is maximal advertised window. It can be less than | |
317 | * tcp_full_space(), in this case tcp_full_space() - window_clamp | |
318 | * is reserved for "application" buffer. The less window_clamp is | |
319 | * the smoother our behaviour from viewpoint of network, but the lower | |
320 | * throughput and the higher sensitivity of the connection to losses. 8) | |
321 | * | |
322 | * rcv_ssthresh is more strict window_clamp used at "slow start" | |
323 | * phase to predict further behaviour of this connection. | |
324 | * It is used for two goals: | |
325 | * - to enforce header prediction at sender, even when application | |
326 | * requires some significant "application buffer". It is check #1. | |
327 | * - to prevent pruning of receive queue because of misprediction | |
328 | * of receiver window. Check #2. | |
329 | * | |
330 | * The scheme does not work when sender sends good segments opening | |
caa20d9a | 331 | * window and then starts to feed us spaghetti. But it should work |
1da177e4 LT |
332 | * in common situations. Otherwise, we have to rely on queue collapsing. |
333 | */ | |
334 | ||
335 | /* Slow part of check#2. */ | |
9e412ba7 | 336 | static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 337 | { |
9e412ba7 | 338 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 339 | /* Optimize this! */ |
dfd4f0ae ED |
340 | int truesize = tcp_win_from_space(skb->truesize) >> 1; |
341 | int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; | |
1da177e4 LT |
342 | |
343 | while (tp->rcv_ssthresh <= window) { | |
344 | if (truesize <= skb->len) | |
463c84b9 | 345 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; |
1da177e4 LT |
346 | |
347 | truesize >>= 1; | |
348 | window >>= 1; | |
349 | } | |
350 | return 0; | |
351 | } | |
352 | ||
cf533ea5 | 353 | static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 354 | { |
9e412ba7 IJ |
355 | struct tcp_sock *tp = tcp_sk(sk); |
356 | ||
1da177e4 LT |
357 | /* Check #1 */ |
358 | if (tp->rcv_ssthresh < tp->window_clamp && | |
359 | (int)tp->rcv_ssthresh < tcp_space(sk) && | |
180d8cd9 | 360 | !sk_under_memory_pressure(sk)) { |
1da177e4 LT |
361 | int incr; |
362 | ||
363 | /* Check #2. Increase window, if skb with such overhead | |
364 | * will fit to rcvbuf in future. | |
365 | */ | |
366 | if (tcp_win_from_space(skb->truesize) <= skb->len) | |
056834d9 | 367 | incr = 2 * tp->advmss; |
1da177e4 | 368 | else |
9e412ba7 | 369 | incr = __tcp_grow_window(sk, skb); |
1da177e4 LT |
370 | |
371 | if (incr) { | |
4d846f02 | 372 | incr = max_t(int, incr, 2 * skb->len); |
056834d9 IJ |
373 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, |
374 | tp->window_clamp); | |
463c84b9 | 375 | inet_csk(sk)->icsk_ack.quick |= 1; |
1da177e4 LT |
376 | } |
377 | } | |
378 | } | |
379 | ||
380 | /* 3. Tuning rcvbuf, when connection enters established state. */ | |
1da177e4 LT |
381 | static void tcp_fixup_rcvbuf(struct sock *sk) |
382 | { | |
e9266a02 | 383 | u32 mss = tcp_sk(sk)->advmss; |
e9266a02 | 384 | int rcvmem; |
1da177e4 | 385 | |
85f16525 YC |
386 | rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) * |
387 | tcp_default_init_rwnd(mss); | |
e9266a02 | 388 | |
b0983d3c ED |
389 | /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency |
390 | * Allow enough cushion so that sender is not limited by our window | |
391 | */ | |
392 | if (sysctl_tcp_moderate_rcvbuf) | |
393 | rcvmem <<= 2; | |
394 | ||
e9266a02 ED |
395 | if (sk->sk_rcvbuf < rcvmem) |
396 | sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]); | |
1da177e4 LT |
397 | } |
398 | ||
caa20d9a | 399 | /* 4. Try to fixup all. It is made immediately after connection enters |
1da177e4 LT |
400 | * established state. |
401 | */ | |
10467163 | 402 | void tcp_init_buffer_space(struct sock *sk) |
1da177e4 LT |
403 | { |
404 | struct tcp_sock *tp = tcp_sk(sk); | |
405 | int maxwin; | |
406 | ||
407 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) | |
408 | tcp_fixup_rcvbuf(sk); | |
409 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) | |
6ae70532 | 410 | tcp_sndbuf_expand(sk); |
1da177e4 LT |
411 | |
412 | tp->rcvq_space.space = tp->rcv_wnd; | |
b0983d3c ED |
413 | tp->rcvq_space.time = tcp_time_stamp; |
414 | tp->rcvq_space.seq = tp->copied_seq; | |
1da177e4 LT |
415 | |
416 | maxwin = tcp_full_space(sk); | |
417 | ||
418 | if (tp->window_clamp >= maxwin) { | |
419 | tp->window_clamp = maxwin; | |
420 | ||
421 | if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) | |
422 | tp->window_clamp = max(maxwin - | |
423 | (maxwin >> sysctl_tcp_app_win), | |
424 | 4 * tp->advmss); | |
425 | } | |
426 | ||
427 | /* Force reservation of one segment. */ | |
428 | if (sysctl_tcp_app_win && | |
429 | tp->window_clamp > 2 * tp->advmss && | |
430 | tp->window_clamp + tp->advmss > maxwin) | |
431 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); | |
432 | ||
433 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); | |
434 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
435 | } | |
436 | ||
1da177e4 | 437 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ |
9e412ba7 | 438 | static void tcp_clamp_window(struct sock *sk) |
1da177e4 | 439 | { |
9e412ba7 | 440 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 441 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 442 | |
6687e988 | 443 | icsk->icsk_ack.quick = 0; |
1da177e4 | 444 | |
326f36e9 JH |
445 | if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && |
446 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && | |
180d8cd9 GC |
447 | !sk_under_memory_pressure(sk) && |
448 | sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { | |
326f36e9 JH |
449 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), |
450 | sysctl_tcp_rmem[2]); | |
1da177e4 | 451 | } |
326f36e9 | 452 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) |
056834d9 | 453 | tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); |
1da177e4 LT |
454 | } |
455 | ||
40efc6fa SH |
456 | /* Initialize RCV_MSS value. |
457 | * RCV_MSS is an our guess about MSS used by the peer. | |
458 | * We haven't any direct information about the MSS. | |
459 | * It's better to underestimate the RCV_MSS rather than overestimate. | |
460 | * Overestimations make us ACKing less frequently than needed. | |
461 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). | |
462 | */ | |
463 | void tcp_initialize_rcv_mss(struct sock *sk) | |
464 | { | |
cf533ea5 | 465 | const struct tcp_sock *tp = tcp_sk(sk); |
40efc6fa SH |
466 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); |
467 | ||
056834d9 | 468 | hint = min(hint, tp->rcv_wnd / 2); |
bee7ca9e | 469 | hint = min(hint, TCP_MSS_DEFAULT); |
40efc6fa SH |
470 | hint = max(hint, TCP_MIN_MSS); |
471 | ||
472 | inet_csk(sk)->icsk_ack.rcv_mss = hint; | |
473 | } | |
4bc2f18b | 474 | EXPORT_SYMBOL(tcp_initialize_rcv_mss); |
40efc6fa | 475 | |
1da177e4 LT |
476 | /* Receiver "autotuning" code. |
477 | * | |
478 | * The algorithm for RTT estimation w/o timestamps is based on | |
479 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. | |
631dd1a8 | 480 | * <http://public.lanl.gov/radiant/pubs.html#DRS> |
1da177e4 LT |
481 | * |
482 | * More detail on this code can be found at | |
631dd1a8 | 483 | * <http://staff.psc.edu/jheffner/>, |
1da177e4 LT |
484 | * though this reference is out of date. A new paper |
485 | * is pending. | |
486 | */ | |
487 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) | |
488 | { | |
489 | u32 new_sample = tp->rcv_rtt_est.rtt; | |
490 | long m = sample; | |
491 | ||
492 | if (m == 0) | |
493 | m = 1; | |
494 | ||
495 | if (new_sample != 0) { | |
496 | /* If we sample in larger samples in the non-timestamp | |
497 | * case, we could grossly overestimate the RTT especially | |
498 | * with chatty applications or bulk transfer apps which | |
499 | * are stalled on filesystem I/O. | |
500 | * | |
501 | * Also, since we are only going for a minimum in the | |
31f34269 | 502 | * non-timestamp case, we do not smooth things out |
caa20d9a | 503 | * else with timestamps disabled convergence takes too |
1da177e4 LT |
504 | * long. |
505 | */ | |
506 | if (!win_dep) { | |
507 | m -= (new_sample >> 3); | |
508 | new_sample += m; | |
18a223e0 NC |
509 | } else { |
510 | m <<= 3; | |
511 | if (m < new_sample) | |
512 | new_sample = m; | |
513 | } | |
1da177e4 | 514 | } else { |
caa20d9a | 515 | /* No previous measure. */ |
1da177e4 LT |
516 | new_sample = m << 3; |
517 | } | |
518 | ||
519 | if (tp->rcv_rtt_est.rtt != new_sample) | |
520 | tp->rcv_rtt_est.rtt = new_sample; | |
521 | } | |
522 | ||
523 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) | |
524 | { | |
525 | if (tp->rcv_rtt_est.time == 0) | |
526 | goto new_measure; | |
527 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) | |
528 | return; | |
651913ce | 529 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1); |
1da177e4 LT |
530 | |
531 | new_measure: | |
532 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; | |
533 | tp->rcv_rtt_est.time = tcp_time_stamp; | |
534 | } | |
535 | ||
056834d9 IJ |
536 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, |
537 | const struct sk_buff *skb) | |
1da177e4 | 538 | { |
463c84b9 | 539 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
540 | if (tp->rx_opt.rcv_tsecr && |
541 | (TCP_SKB_CB(skb)->end_seq - | |
463c84b9 | 542 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) |
1da177e4 LT |
543 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); |
544 | } | |
545 | ||
546 | /* | |
547 | * This function should be called every time data is copied to user space. | |
548 | * It calculates the appropriate TCP receive buffer space. | |
549 | */ | |
550 | void tcp_rcv_space_adjust(struct sock *sk) | |
551 | { | |
552 | struct tcp_sock *tp = tcp_sk(sk); | |
553 | int time; | |
b0983d3c | 554 | int copied; |
e905a9ed | 555 | |
1da177e4 | 556 | time = tcp_time_stamp - tp->rcvq_space.time; |
056834d9 | 557 | if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) |
1da177e4 | 558 | return; |
e905a9ed | 559 | |
b0983d3c ED |
560 | /* Number of bytes copied to user in last RTT */ |
561 | copied = tp->copied_seq - tp->rcvq_space.seq; | |
562 | if (copied <= tp->rcvq_space.space) | |
563 | goto new_measure; | |
564 | ||
565 | /* A bit of theory : | |
566 | * copied = bytes received in previous RTT, our base window | |
567 | * To cope with packet losses, we need a 2x factor | |
568 | * To cope with slow start, and sender growing its cwin by 100 % | |
569 | * every RTT, we need a 4x factor, because the ACK we are sending | |
570 | * now is for the next RTT, not the current one : | |
571 | * <prev RTT . ><current RTT .. ><next RTT .... > | |
572 | */ | |
573 | ||
574 | if (sysctl_tcp_moderate_rcvbuf && | |
575 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { | |
576 | int rcvwin, rcvmem, rcvbuf; | |
1da177e4 | 577 | |
b0983d3c ED |
578 | /* minimal window to cope with packet losses, assuming |
579 | * steady state. Add some cushion because of small variations. | |
580 | */ | |
581 | rcvwin = (copied << 1) + 16 * tp->advmss; | |
1da177e4 | 582 | |
b0983d3c ED |
583 | /* If rate increased by 25%, |
584 | * assume slow start, rcvwin = 3 * copied | |
585 | * If rate increased by 50%, | |
586 | * assume sender can use 2x growth, rcvwin = 4 * copied | |
587 | */ | |
588 | if (copied >= | |
589 | tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) { | |
590 | if (copied >= | |
591 | tp->rcvq_space.space + (tp->rcvq_space.space >> 1)) | |
592 | rcvwin <<= 1; | |
593 | else | |
594 | rcvwin += (rcvwin >> 1); | |
595 | } | |
1da177e4 | 596 | |
b0983d3c ED |
597 | rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER); |
598 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
599 | rcvmem += 128; | |
1da177e4 | 600 | |
b0983d3c ED |
601 | rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]); |
602 | if (rcvbuf > sk->sk_rcvbuf) { | |
603 | sk->sk_rcvbuf = rcvbuf; | |
1da177e4 | 604 | |
b0983d3c ED |
605 | /* Make the window clamp follow along. */ |
606 | tp->window_clamp = rcvwin; | |
1da177e4 LT |
607 | } |
608 | } | |
b0983d3c | 609 | tp->rcvq_space.space = copied; |
e905a9ed | 610 | |
1da177e4 LT |
611 | new_measure: |
612 | tp->rcvq_space.seq = tp->copied_seq; | |
613 | tp->rcvq_space.time = tcp_time_stamp; | |
614 | } | |
615 | ||
616 | /* There is something which you must keep in mind when you analyze the | |
617 | * behavior of the tp->ato delayed ack timeout interval. When a | |
618 | * connection starts up, we want to ack as quickly as possible. The | |
619 | * problem is that "good" TCP's do slow start at the beginning of data | |
620 | * transmission. The means that until we send the first few ACK's the | |
621 | * sender will sit on his end and only queue most of his data, because | |
622 | * he can only send snd_cwnd unacked packets at any given time. For | |
623 | * each ACK we send, he increments snd_cwnd and transmits more of his | |
624 | * queue. -DaveM | |
625 | */ | |
9e412ba7 | 626 | static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 627 | { |
9e412ba7 | 628 | struct tcp_sock *tp = tcp_sk(sk); |
463c84b9 | 629 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
630 | u32 now; |
631 | ||
463c84b9 | 632 | inet_csk_schedule_ack(sk); |
1da177e4 | 633 | |
463c84b9 | 634 | tcp_measure_rcv_mss(sk, skb); |
1da177e4 LT |
635 | |
636 | tcp_rcv_rtt_measure(tp); | |
e905a9ed | 637 | |
1da177e4 LT |
638 | now = tcp_time_stamp; |
639 | ||
463c84b9 | 640 | if (!icsk->icsk_ack.ato) { |
1da177e4 LT |
641 | /* The _first_ data packet received, initialize |
642 | * delayed ACK engine. | |
643 | */ | |
463c84b9 ACM |
644 | tcp_incr_quickack(sk); |
645 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 | 646 | } else { |
463c84b9 | 647 | int m = now - icsk->icsk_ack.lrcvtime; |
1da177e4 | 648 | |
056834d9 | 649 | if (m <= TCP_ATO_MIN / 2) { |
1da177e4 | 650 | /* The fastest case is the first. */ |
463c84b9 ACM |
651 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; |
652 | } else if (m < icsk->icsk_ack.ato) { | |
653 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; | |
654 | if (icsk->icsk_ack.ato > icsk->icsk_rto) | |
655 | icsk->icsk_ack.ato = icsk->icsk_rto; | |
656 | } else if (m > icsk->icsk_rto) { | |
caa20d9a | 657 | /* Too long gap. Apparently sender failed to |
1da177e4 LT |
658 | * restart window, so that we send ACKs quickly. |
659 | */ | |
463c84b9 | 660 | tcp_incr_quickack(sk); |
3ab224be | 661 | sk_mem_reclaim(sk); |
1da177e4 LT |
662 | } |
663 | } | |
463c84b9 | 664 | icsk->icsk_ack.lrcvtime = now; |
1da177e4 | 665 | |
735d3831 | 666 | tcp_ecn_check_ce(tp, skb); |
1da177e4 LT |
667 | |
668 | if (skb->len >= 128) | |
9e412ba7 | 669 | tcp_grow_window(sk, skb); |
1da177e4 LT |
670 | } |
671 | ||
1da177e4 LT |
672 | /* Called to compute a smoothed rtt estimate. The data fed to this |
673 | * routine either comes from timestamps, or from segments that were | |
674 | * known _not_ to have been retransmitted [see Karn/Partridge | |
675 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 | |
676 | * piece by Van Jacobson. | |
677 | * NOTE: the next three routines used to be one big routine. | |
678 | * To save cycles in the RFC 1323 implementation it was better to break | |
679 | * it up into three procedures. -- erics | |
680 | */ | |
740b0f18 | 681 | static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) |
1da177e4 | 682 | { |
6687e988 | 683 | struct tcp_sock *tp = tcp_sk(sk); |
740b0f18 ED |
684 | long m = mrtt_us; /* RTT */ |
685 | u32 srtt = tp->srtt_us; | |
1da177e4 | 686 | |
1da177e4 LT |
687 | /* The following amusing code comes from Jacobson's |
688 | * article in SIGCOMM '88. Note that rtt and mdev | |
689 | * are scaled versions of rtt and mean deviation. | |
e905a9ed | 690 | * This is designed to be as fast as possible |
1da177e4 LT |
691 | * m stands for "measurement". |
692 | * | |
693 | * On a 1990 paper the rto value is changed to: | |
694 | * RTO = rtt + 4 * mdev | |
695 | * | |
696 | * Funny. This algorithm seems to be very broken. | |
697 | * These formulae increase RTO, when it should be decreased, increase | |
31f34269 | 698 | * too slowly, when it should be increased quickly, decrease too quickly |
1da177e4 LT |
699 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely |
700 | * does not matter how to _calculate_ it. Seems, it was trap | |
701 | * that VJ failed to avoid. 8) | |
702 | */ | |
4a5ab4e2 ED |
703 | if (srtt != 0) { |
704 | m -= (srtt >> 3); /* m is now error in rtt est */ | |
705 | srtt += m; /* rtt = 7/8 rtt + 1/8 new */ | |
1da177e4 LT |
706 | if (m < 0) { |
707 | m = -m; /* m is now abs(error) */ | |
740b0f18 | 708 | m -= (tp->mdev_us >> 2); /* similar update on mdev */ |
1da177e4 LT |
709 | /* This is similar to one of Eifel findings. |
710 | * Eifel blocks mdev updates when rtt decreases. | |
711 | * This solution is a bit different: we use finer gain | |
712 | * for mdev in this case (alpha*beta). | |
713 | * Like Eifel it also prevents growth of rto, | |
714 | * but also it limits too fast rto decreases, | |
715 | * happening in pure Eifel. | |
716 | */ | |
717 | if (m > 0) | |
718 | m >>= 3; | |
719 | } else { | |
740b0f18 | 720 | m -= (tp->mdev_us >> 2); /* similar update on mdev */ |
1da177e4 | 721 | } |
740b0f18 ED |
722 | tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ |
723 | if (tp->mdev_us > tp->mdev_max_us) { | |
724 | tp->mdev_max_us = tp->mdev_us; | |
725 | if (tp->mdev_max_us > tp->rttvar_us) | |
726 | tp->rttvar_us = tp->mdev_max_us; | |
1da177e4 LT |
727 | } |
728 | if (after(tp->snd_una, tp->rtt_seq)) { | |
740b0f18 ED |
729 | if (tp->mdev_max_us < tp->rttvar_us) |
730 | tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; | |
1da177e4 | 731 | tp->rtt_seq = tp->snd_nxt; |
740b0f18 | 732 | tp->mdev_max_us = tcp_rto_min_us(sk); |
1da177e4 LT |
733 | } |
734 | } else { | |
735 | /* no previous measure. */ | |
4a5ab4e2 | 736 | srtt = m << 3; /* take the measured time to be rtt */ |
740b0f18 ED |
737 | tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ |
738 | tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); | |
739 | tp->mdev_max_us = tp->rttvar_us; | |
1da177e4 LT |
740 | tp->rtt_seq = tp->snd_nxt; |
741 | } | |
740b0f18 | 742 | tp->srtt_us = max(1U, srtt); |
1da177e4 LT |
743 | } |
744 | ||
95bd09eb ED |
745 | /* Set the sk_pacing_rate to allow proper sizing of TSO packets. |
746 | * Note: TCP stack does not yet implement pacing. | |
747 | * FQ packet scheduler can be used to implement cheap but effective | |
748 | * TCP pacing, to smooth the burst on large writes when packets | |
749 | * in flight is significantly lower than cwnd (or rwin) | |
750 | */ | |
751 | static void tcp_update_pacing_rate(struct sock *sk) | |
752 | { | |
753 | const struct tcp_sock *tp = tcp_sk(sk); | |
754 | u64 rate; | |
755 | ||
756 | /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ | |
740b0f18 | 757 | rate = (u64)tp->mss_cache * 2 * (USEC_PER_SEC << 3); |
95bd09eb ED |
758 | |
759 | rate *= max(tp->snd_cwnd, tp->packets_out); | |
760 | ||
740b0f18 ED |
761 | if (likely(tp->srtt_us)) |
762 | do_div(rate, tp->srtt_us); | |
95bd09eb | 763 | |
ba537427 ED |
764 | /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate |
765 | * without any lock. We want to make sure compiler wont store | |
766 | * intermediate values in this location. | |
767 | */ | |
768 | ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate, | |
769 | sk->sk_max_pacing_rate); | |
95bd09eb ED |
770 | } |
771 | ||
1da177e4 LT |
772 | /* Calculate rto without backoff. This is the second half of Van Jacobson's |
773 | * routine referred to above. | |
774 | */ | |
f7e56a76 | 775 | static void tcp_set_rto(struct sock *sk) |
1da177e4 | 776 | { |
463c84b9 | 777 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
778 | /* Old crap is replaced with new one. 8) |
779 | * | |
780 | * More seriously: | |
781 | * 1. If rtt variance happened to be less 50msec, it is hallucination. | |
782 | * It cannot be less due to utterly erratic ACK generation made | |
783 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ | |
784 | * to do with delayed acks, because at cwnd>2 true delack timeout | |
785 | * is invisible. Actually, Linux-2.4 also generates erratic | |
caa20d9a | 786 | * ACKs in some circumstances. |
1da177e4 | 787 | */ |
f1ecd5d9 | 788 | inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); |
1da177e4 LT |
789 | |
790 | /* 2. Fixups made earlier cannot be right. | |
791 | * If we do not estimate RTO correctly without them, | |
792 | * all the algo is pure shit and should be replaced | |
caa20d9a | 793 | * with correct one. It is exactly, which we pretend to do. |
1da177e4 | 794 | */ |
1da177e4 | 795 | |
ee6aac59 IJ |
796 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo |
797 | * guarantees that rto is higher. | |
798 | */ | |
f1ecd5d9 | 799 | tcp_bound_rto(sk); |
1da177e4 LT |
800 | } |
801 | ||
cf533ea5 | 802 | __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) |
1da177e4 LT |
803 | { |
804 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); | |
805 | ||
22b71c8f | 806 | if (!cwnd) |
442b9635 | 807 | cwnd = TCP_INIT_CWND; |
1da177e4 LT |
808 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); |
809 | } | |
810 | ||
e60402d0 IJ |
811 | /* |
812 | * Packet counting of FACK is based on in-order assumptions, therefore TCP | |
813 | * disables it when reordering is detected | |
814 | */ | |
4aabd8ef | 815 | void tcp_disable_fack(struct tcp_sock *tp) |
e60402d0 | 816 | { |
85cc391c IJ |
817 | /* RFC3517 uses different metric in lost marker => reset on change */ |
818 | if (tcp_is_fack(tp)) | |
819 | tp->lost_skb_hint = NULL; | |
ab56222a | 820 | tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED; |
e60402d0 IJ |
821 | } |
822 | ||
564262c1 | 823 | /* Take a notice that peer is sending D-SACKs */ |
e60402d0 IJ |
824 | static void tcp_dsack_seen(struct tcp_sock *tp) |
825 | { | |
ab56222a | 826 | tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; |
e60402d0 IJ |
827 | } |
828 | ||
6687e988 ACM |
829 | static void tcp_update_reordering(struct sock *sk, const int metric, |
830 | const int ts) | |
1da177e4 | 831 | { |
6687e988 | 832 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 833 | if (metric > tp->reordering) { |
40b215e5 PE |
834 | int mib_idx; |
835 | ||
1da177e4 LT |
836 | tp->reordering = min(TCP_MAX_REORDERING, metric); |
837 | ||
838 | /* This exciting event is worth to be remembered. 8) */ | |
839 | if (ts) | |
40b215e5 | 840 | mib_idx = LINUX_MIB_TCPTSREORDER; |
e60402d0 | 841 | else if (tcp_is_reno(tp)) |
40b215e5 | 842 | mib_idx = LINUX_MIB_TCPRENOREORDER; |
e60402d0 | 843 | else if (tcp_is_fack(tp)) |
40b215e5 | 844 | mib_idx = LINUX_MIB_TCPFACKREORDER; |
1da177e4 | 845 | else |
40b215e5 PE |
846 | mib_idx = LINUX_MIB_TCPSACKREORDER; |
847 | ||
de0744af | 848 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
1da177e4 | 849 | #if FASTRETRANS_DEBUG > 1 |
91df42be JP |
850 | pr_debug("Disorder%d %d %u f%u s%u rr%d\n", |
851 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, | |
852 | tp->reordering, | |
853 | tp->fackets_out, | |
854 | tp->sacked_out, | |
855 | tp->undo_marker ? tp->undo_retrans : 0); | |
1da177e4 | 856 | #endif |
e60402d0 | 857 | tcp_disable_fack(tp); |
1da177e4 | 858 | } |
eed530b6 YC |
859 | |
860 | if (metric > 0) | |
861 | tcp_disable_early_retrans(tp); | |
1da177e4 LT |
862 | } |
863 | ||
006f582c | 864 | /* This must be called before lost_out is incremented */ |
c8c213f2 IJ |
865 | static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) |
866 | { | |
006f582c | 867 | if ((tp->retransmit_skb_hint == NULL) || |
c8c213f2 IJ |
868 | before(TCP_SKB_CB(skb)->seq, |
869 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) | |
006f582c IJ |
870 | tp->retransmit_skb_hint = skb; |
871 | ||
872 | if (!tp->lost_out || | |
873 | after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) | |
874 | tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; | |
c8c213f2 IJ |
875 | } |
876 | ||
41ea36e3 IJ |
877 | static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) |
878 | { | |
879 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
880 | tcp_verify_retransmit_hint(tp, skb); | |
881 | ||
882 | tp->lost_out += tcp_skb_pcount(skb); | |
883 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
884 | } | |
885 | } | |
886 | ||
e1aa680f IJ |
887 | static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, |
888 | struct sk_buff *skb) | |
006f582c IJ |
889 | { |
890 | tcp_verify_retransmit_hint(tp, skb); | |
891 | ||
892 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
893 | tp->lost_out += tcp_skb_pcount(skb); | |
894 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
895 | } | |
896 | } | |
897 | ||
1da177e4 LT |
898 | /* This procedure tags the retransmission queue when SACKs arrive. |
899 | * | |
900 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). | |
901 | * Packets in queue with these bits set are counted in variables | |
902 | * sacked_out, retrans_out and lost_out, correspondingly. | |
903 | * | |
904 | * Valid combinations are: | |
905 | * Tag InFlight Description | |
906 | * 0 1 - orig segment is in flight. | |
907 | * S 0 - nothing flies, orig reached receiver. | |
908 | * L 0 - nothing flies, orig lost by net. | |
909 | * R 2 - both orig and retransmit are in flight. | |
910 | * L|R 1 - orig is lost, retransmit is in flight. | |
911 | * S|R 1 - orig reached receiver, retrans is still in flight. | |
912 | * (L|S|R is logically valid, it could occur when L|R is sacked, | |
913 | * but it is equivalent to plain S and code short-curcuits it to S. | |
914 | * L|S is logically invalid, it would mean -1 packet in flight 8)) | |
915 | * | |
916 | * These 6 states form finite state machine, controlled by the following events: | |
917 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) | |
918 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) | |
974c1236 | 919 | * 3. Loss detection event of two flavors: |
1da177e4 LT |
920 | * A. Scoreboard estimator decided the packet is lost. |
921 | * A'. Reno "three dupacks" marks head of queue lost. | |
974c1236 YC |
922 | * A''. Its FACK modification, head until snd.fack is lost. |
923 | * B. SACK arrives sacking SND.NXT at the moment, when the | |
1da177e4 LT |
924 | * segment was retransmitted. |
925 | * 4. D-SACK added new rule: D-SACK changes any tag to S. | |
926 | * | |
927 | * It is pleasant to note, that state diagram turns out to be commutative, | |
928 | * so that we are allowed not to be bothered by order of our actions, | |
929 | * when multiple events arrive simultaneously. (see the function below). | |
930 | * | |
931 | * Reordering detection. | |
932 | * -------------------- | |
933 | * Reordering metric is maximal distance, which a packet can be displaced | |
934 | * in packet stream. With SACKs we can estimate it: | |
935 | * | |
936 | * 1. SACK fills old hole and the corresponding segment was not | |
937 | * ever retransmitted -> reordering. Alas, we cannot use it | |
938 | * when segment was retransmitted. | |
939 | * 2. The last flaw is solved with D-SACK. D-SACK arrives | |
940 | * for retransmitted and already SACKed segment -> reordering.. | |
941 | * Both of these heuristics are not used in Loss state, when we cannot | |
942 | * account for retransmits accurately. | |
5b3c9882 IJ |
943 | * |
944 | * SACK block validation. | |
945 | * ---------------------- | |
946 | * | |
947 | * SACK block range validation checks that the received SACK block fits to | |
948 | * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. | |
949 | * Note that SND.UNA is not included to the range though being valid because | |
0e835331 IJ |
950 | * it means that the receiver is rather inconsistent with itself reporting |
951 | * SACK reneging when it should advance SND.UNA. Such SACK block this is | |
952 | * perfectly valid, however, in light of RFC2018 which explicitly states | |
953 | * that "SACK block MUST reflect the newest segment. Even if the newest | |
954 | * segment is going to be discarded ...", not that it looks very clever | |
955 | * in case of head skb. Due to potentional receiver driven attacks, we | |
956 | * choose to avoid immediate execution of a walk in write queue due to | |
957 | * reneging and defer head skb's loss recovery to standard loss recovery | |
958 | * procedure that will eventually trigger (nothing forbids us doing this). | |
5b3c9882 IJ |
959 | * |
960 | * Implements also blockage to start_seq wrap-around. Problem lies in the | |
961 | * fact that though start_seq (s) is before end_seq (i.e., not reversed), | |
962 | * there's no guarantee that it will be before snd_nxt (n). The problem | |
963 | * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt | |
964 | * wrap (s_w): | |
965 | * | |
966 | * <- outs wnd -> <- wrapzone -> | |
967 | * u e n u_w e_w s n_w | |
968 | * | | | | | | | | |
969 | * |<------------+------+----- TCP seqno space --------------+---------->| | |
970 | * ...-- <2^31 ->| |<--------... | |
971 | * ...---- >2^31 ------>| |<--------... | |
972 | * | |
973 | * Current code wouldn't be vulnerable but it's better still to discard such | |
974 | * crazy SACK blocks. Doing this check for start_seq alone closes somewhat | |
975 | * similar case (end_seq after snd_nxt wrap) as earlier reversed check in | |
976 | * snd_nxt wrap -> snd_una region will then become "well defined", i.e., | |
977 | * equal to the ideal case (infinite seqno space without wrap caused issues). | |
978 | * | |
979 | * With D-SACK the lower bound is extended to cover sequence space below | |
980 | * SND.UNA down to undo_marker, which is the last point of interest. Yet | |
564262c1 | 981 | * again, D-SACK block must not to go across snd_una (for the same reason as |
5b3c9882 IJ |
982 | * for the normal SACK blocks, explained above). But there all simplicity |
983 | * ends, TCP might receive valid D-SACKs below that. As long as they reside | |
984 | * fully below undo_marker they do not affect behavior in anyway and can | |
985 | * therefore be safely ignored. In rare cases (which are more or less | |
986 | * theoretical ones), the D-SACK will nicely cross that boundary due to skb | |
987 | * fragmentation and packet reordering past skb's retransmission. To consider | |
988 | * them correctly, the acceptable range must be extended even more though | |
989 | * the exact amount is rather hard to quantify. However, tp->max_window can | |
990 | * be used as an exaggerated estimate. | |
1da177e4 | 991 | */ |
a2a385d6 ED |
992 | static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, |
993 | u32 start_seq, u32 end_seq) | |
5b3c9882 IJ |
994 | { |
995 | /* Too far in future, or reversed (interpretation is ambiguous) */ | |
996 | if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) | |
a2a385d6 | 997 | return false; |
5b3c9882 IJ |
998 | |
999 | /* Nasty start_seq wrap-around check (see comments above) */ | |
1000 | if (!before(start_seq, tp->snd_nxt)) | |
a2a385d6 | 1001 | return false; |
5b3c9882 | 1002 | |
564262c1 | 1003 | /* In outstanding window? ...This is valid exit for D-SACKs too. |
5b3c9882 IJ |
1004 | * start_seq == snd_una is non-sensical (see comments above) |
1005 | */ | |
1006 | if (after(start_seq, tp->snd_una)) | |
a2a385d6 | 1007 | return true; |
5b3c9882 IJ |
1008 | |
1009 | if (!is_dsack || !tp->undo_marker) | |
a2a385d6 | 1010 | return false; |
5b3c9882 IJ |
1011 | |
1012 | /* ...Then it's D-SACK, and must reside below snd_una completely */ | |
f779b2d6 | 1013 | if (after(end_seq, tp->snd_una)) |
a2a385d6 | 1014 | return false; |
5b3c9882 IJ |
1015 | |
1016 | if (!before(start_seq, tp->undo_marker)) | |
a2a385d6 | 1017 | return true; |
5b3c9882 IJ |
1018 | |
1019 | /* Too old */ | |
1020 | if (!after(end_seq, tp->undo_marker)) | |
a2a385d6 | 1021 | return false; |
5b3c9882 IJ |
1022 | |
1023 | /* Undo_marker boundary crossing (overestimates a lot). Known already: | |
1024 | * start_seq < undo_marker and end_seq >= undo_marker. | |
1025 | */ | |
1026 | return !before(start_seq, end_seq - tp->max_window); | |
1027 | } | |
1028 | ||
1c1e87ed | 1029 | /* Check for lost retransmit. This superb idea is borrowed from "ratehalving". |
974c1236 | 1030 | * Event "B". Later note: FACK people cheated me again 8), we have to account |
1c1e87ed | 1031 | * for reordering! Ugly, but should help. |
f785a8e2 IJ |
1032 | * |
1033 | * Search retransmitted skbs from write_queue that were sent when snd_nxt was | |
1034 | * less than what is now known to be received by the other end (derived from | |
9f58f3b7 IJ |
1035 | * highest SACK block). Also calculate the lowest snd_nxt among the remaining |
1036 | * retransmitted skbs to avoid some costly processing per ACKs. | |
1c1e87ed | 1037 | */ |
407ef1de | 1038 | static void tcp_mark_lost_retrans(struct sock *sk) |
1c1e87ed | 1039 | { |
9f58f3b7 | 1040 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1c1e87ed IJ |
1041 | struct tcp_sock *tp = tcp_sk(sk); |
1042 | struct sk_buff *skb; | |
f785a8e2 | 1043 | int cnt = 0; |
df2e014b | 1044 | u32 new_low_seq = tp->snd_nxt; |
6859d494 | 1045 | u32 received_upto = tcp_highest_sack_seq(tp); |
9f58f3b7 IJ |
1046 | |
1047 | if (!tcp_is_fack(tp) || !tp->retrans_out || | |
1048 | !after(received_upto, tp->lost_retrans_low) || | |
1049 | icsk->icsk_ca_state != TCP_CA_Recovery) | |
407ef1de | 1050 | return; |
1c1e87ed IJ |
1051 | |
1052 | tcp_for_write_queue(skb, sk) { | |
1053 | u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; | |
1054 | ||
1055 | if (skb == tcp_send_head(sk)) | |
1056 | break; | |
f785a8e2 | 1057 | if (cnt == tp->retrans_out) |
1c1e87ed IJ |
1058 | break; |
1059 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1060 | continue; | |
1061 | ||
f785a8e2 IJ |
1062 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) |
1063 | continue; | |
1064 | ||
d0af4160 IJ |
1065 | /* TODO: We would like to get rid of tcp_is_fack(tp) only |
1066 | * constraint here (see above) but figuring out that at | |
1067 | * least tp->reordering SACK blocks reside between ack_seq | |
1068 | * and received_upto is not easy task to do cheaply with | |
1069 | * the available datastructures. | |
1070 | * | |
1071 | * Whether FACK should check here for tp->reordering segs | |
1072 | * in-between one could argue for either way (it would be | |
1073 | * rather simple to implement as we could count fack_count | |
1074 | * during the walk and do tp->fackets_out - fack_count). | |
1075 | */ | |
1076 | if (after(received_upto, ack_seq)) { | |
1c1e87ed IJ |
1077 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
1078 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1079 | ||
006f582c | 1080 | tcp_skb_mark_lost_uncond_verify(tp, skb); |
de0744af | 1081 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT); |
f785a8e2 | 1082 | } else { |
df2e014b | 1083 | if (before(ack_seq, new_low_seq)) |
b08d6cb2 | 1084 | new_low_seq = ack_seq; |
f785a8e2 | 1085 | cnt += tcp_skb_pcount(skb); |
1c1e87ed IJ |
1086 | } |
1087 | } | |
b08d6cb2 IJ |
1088 | |
1089 | if (tp->retrans_out) | |
1090 | tp->lost_retrans_low = new_low_seq; | |
1c1e87ed | 1091 | } |
5b3c9882 | 1092 | |
a2a385d6 ED |
1093 | static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, |
1094 | struct tcp_sack_block_wire *sp, int num_sacks, | |
1095 | u32 prior_snd_una) | |
d06e021d | 1096 | { |
1ed83465 | 1097 | struct tcp_sock *tp = tcp_sk(sk); |
d3e2ce3b HH |
1098 | u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); |
1099 | u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); | |
a2a385d6 | 1100 | bool dup_sack = false; |
d06e021d DM |
1101 | |
1102 | if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { | |
a2a385d6 | 1103 | dup_sack = true; |
e60402d0 | 1104 | tcp_dsack_seen(tp); |
de0744af | 1105 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV); |
d06e021d | 1106 | } else if (num_sacks > 1) { |
d3e2ce3b HH |
1107 | u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); |
1108 | u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); | |
d06e021d DM |
1109 | |
1110 | if (!after(end_seq_0, end_seq_1) && | |
1111 | !before(start_seq_0, start_seq_1)) { | |
a2a385d6 | 1112 | dup_sack = true; |
e60402d0 | 1113 | tcp_dsack_seen(tp); |
de0744af PE |
1114 | NET_INC_STATS_BH(sock_net(sk), |
1115 | LINUX_MIB_TCPDSACKOFORECV); | |
d06e021d DM |
1116 | } |
1117 | } | |
1118 | ||
1119 | /* D-SACK for already forgotten data... Do dumb counting. */ | |
6e08d5e3 | 1120 | if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 && |
d06e021d DM |
1121 | !after(end_seq_0, prior_snd_una) && |
1122 | after(end_seq_0, tp->undo_marker)) | |
1123 | tp->undo_retrans--; | |
1124 | ||
1125 | return dup_sack; | |
1126 | } | |
1127 | ||
a1197f5a | 1128 | struct tcp_sacktag_state { |
740b0f18 ED |
1129 | int reord; |
1130 | int fack_count; | |
1131 | long rtt_us; /* RTT measured by SACKing never-retransmitted data */ | |
1132 | int flag; | |
a1197f5a IJ |
1133 | }; |
1134 | ||
d1935942 IJ |
1135 | /* Check if skb is fully within the SACK block. In presence of GSO skbs, |
1136 | * the incoming SACK may not exactly match but we can find smaller MSS | |
1137 | * aligned portion of it that matches. Therefore we might need to fragment | |
1138 | * which may fail and creates some hassle (caller must handle error case | |
1139 | * returns). | |
832d11c5 IJ |
1140 | * |
1141 | * FIXME: this could be merged to shift decision code | |
d1935942 | 1142 | */ |
0f79efdc | 1143 | static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, |
a2a385d6 | 1144 | u32 start_seq, u32 end_seq) |
d1935942 | 1145 | { |
a2a385d6 ED |
1146 | int err; |
1147 | bool in_sack; | |
d1935942 | 1148 | unsigned int pkt_len; |
adb92db8 | 1149 | unsigned int mss; |
d1935942 IJ |
1150 | |
1151 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1152 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1153 | ||
1154 | if (tcp_skb_pcount(skb) > 1 && !in_sack && | |
1155 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { | |
adb92db8 | 1156 | mss = tcp_skb_mss(skb); |
d1935942 IJ |
1157 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); |
1158 | ||
adb92db8 | 1159 | if (!in_sack) { |
d1935942 | 1160 | pkt_len = start_seq - TCP_SKB_CB(skb)->seq; |
adb92db8 IJ |
1161 | if (pkt_len < mss) |
1162 | pkt_len = mss; | |
1163 | } else { | |
d1935942 | 1164 | pkt_len = end_seq - TCP_SKB_CB(skb)->seq; |
adb92db8 IJ |
1165 | if (pkt_len < mss) |
1166 | return -EINVAL; | |
1167 | } | |
1168 | ||
1169 | /* Round if necessary so that SACKs cover only full MSSes | |
1170 | * and/or the remaining small portion (if present) | |
1171 | */ | |
1172 | if (pkt_len > mss) { | |
1173 | unsigned int new_len = (pkt_len / mss) * mss; | |
1174 | if (!in_sack && new_len < pkt_len) { | |
1175 | new_len += mss; | |
2cd0d743 | 1176 | if (new_len >= skb->len) |
adb92db8 IJ |
1177 | return 0; |
1178 | } | |
1179 | pkt_len = new_len; | |
1180 | } | |
6cc55e09 | 1181 | err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC); |
d1935942 IJ |
1182 | if (err < 0) |
1183 | return err; | |
1184 | } | |
1185 | ||
1186 | return in_sack; | |
1187 | } | |
1188 | ||
cc9a672e NC |
1189 | /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ |
1190 | static u8 tcp_sacktag_one(struct sock *sk, | |
1191 | struct tcp_sacktag_state *state, u8 sacked, | |
1192 | u32 start_seq, u32 end_seq, | |
740b0f18 ED |
1193 | int dup_sack, int pcount, |
1194 | const struct skb_mstamp *xmit_time) | |
9e10c47c | 1195 | { |
6859d494 | 1196 | struct tcp_sock *tp = tcp_sk(sk); |
a1197f5a | 1197 | int fack_count = state->fack_count; |
9e10c47c IJ |
1198 | |
1199 | /* Account D-SACK for retransmitted packet. */ | |
1200 | if (dup_sack && (sacked & TCPCB_RETRANS)) { | |
6e08d5e3 | 1201 | if (tp->undo_marker && tp->undo_retrans > 0 && |
cc9a672e | 1202 | after(end_seq, tp->undo_marker)) |
9e10c47c | 1203 | tp->undo_retrans--; |
ede9f3b1 | 1204 | if (sacked & TCPCB_SACKED_ACKED) |
a1197f5a | 1205 | state->reord = min(fack_count, state->reord); |
9e10c47c IJ |
1206 | } |
1207 | ||
1208 | /* Nothing to do; acked frame is about to be dropped (was ACKed). */ | |
cc9a672e | 1209 | if (!after(end_seq, tp->snd_una)) |
a1197f5a | 1210 | return sacked; |
9e10c47c IJ |
1211 | |
1212 | if (!(sacked & TCPCB_SACKED_ACKED)) { | |
1213 | if (sacked & TCPCB_SACKED_RETRANS) { | |
1214 | /* If the segment is not tagged as lost, | |
1215 | * we do not clear RETRANS, believing | |
1216 | * that retransmission is still in flight. | |
1217 | */ | |
1218 | if (sacked & TCPCB_LOST) { | |
a1197f5a | 1219 | sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); |
f58b22fd IJ |
1220 | tp->lost_out -= pcount; |
1221 | tp->retrans_out -= pcount; | |
9e10c47c IJ |
1222 | } |
1223 | } else { | |
1224 | if (!(sacked & TCPCB_RETRANS)) { | |
1225 | /* New sack for not retransmitted frame, | |
1226 | * which was in hole. It is reordering. | |
1227 | */ | |
cc9a672e | 1228 | if (before(start_seq, |
9e10c47c | 1229 | tcp_highest_sack_seq(tp))) |
a1197f5a IJ |
1230 | state->reord = min(fack_count, |
1231 | state->reord); | |
e33099f9 YC |
1232 | if (!after(end_seq, tp->high_seq)) |
1233 | state->flag |= FLAG_ORIG_SACK_ACKED; | |
59c9af42 | 1234 | /* Pick the earliest sequence sacked for RTT */ |
740b0f18 ED |
1235 | if (state->rtt_us < 0) { |
1236 | struct skb_mstamp now; | |
1237 | ||
1238 | skb_mstamp_get(&now); | |
1239 | state->rtt_us = skb_mstamp_us_delta(&now, | |
1240 | xmit_time); | |
1241 | } | |
9e10c47c IJ |
1242 | } |
1243 | ||
1244 | if (sacked & TCPCB_LOST) { | |
a1197f5a | 1245 | sacked &= ~TCPCB_LOST; |
f58b22fd | 1246 | tp->lost_out -= pcount; |
9e10c47c IJ |
1247 | } |
1248 | } | |
1249 | ||
a1197f5a IJ |
1250 | sacked |= TCPCB_SACKED_ACKED; |
1251 | state->flag |= FLAG_DATA_SACKED; | |
f58b22fd | 1252 | tp->sacked_out += pcount; |
9e10c47c | 1253 | |
f58b22fd | 1254 | fack_count += pcount; |
9e10c47c IJ |
1255 | |
1256 | /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ | |
1257 | if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && | |
cc9a672e | 1258 | before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) |
f58b22fd | 1259 | tp->lost_cnt_hint += pcount; |
9e10c47c IJ |
1260 | |
1261 | if (fack_count > tp->fackets_out) | |
1262 | tp->fackets_out = fack_count; | |
9e10c47c IJ |
1263 | } |
1264 | ||
1265 | /* D-SACK. We can detect redundant retransmission in S|R and plain R | |
1266 | * frames and clear it. undo_retrans is decreased above, L|R frames | |
1267 | * are accounted above as well. | |
1268 | */ | |
a1197f5a IJ |
1269 | if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { |
1270 | sacked &= ~TCPCB_SACKED_RETRANS; | |
f58b22fd | 1271 | tp->retrans_out -= pcount; |
9e10c47c IJ |
1272 | } |
1273 | ||
a1197f5a | 1274 | return sacked; |
9e10c47c IJ |
1275 | } |
1276 | ||
daef52ba NC |
1277 | /* Shift newly-SACKed bytes from this skb to the immediately previous |
1278 | * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. | |
1279 | */ | |
a2a385d6 ED |
1280 | static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, |
1281 | struct tcp_sacktag_state *state, | |
1282 | unsigned int pcount, int shifted, int mss, | |
1283 | bool dup_sack) | |
832d11c5 IJ |
1284 | { |
1285 | struct tcp_sock *tp = tcp_sk(sk); | |
50133161 | 1286 | struct sk_buff *prev = tcp_write_queue_prev(sk, skb); |
daef52ba NC |
1287 | u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ |
1288 | u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ | |
832d11c5 IJ |
1289 | |
1290 | BUG_ON(!pcount); | |
1291 | ||
4c90d3b3 NC |
1292 | /* Adjust counters and hints for the newly sacked sequence |
1293 | * range but discard the return value since prev is already | |
1294 | * marked. We must tag the range first because the seq | |
1295 | * advancement below implicitly advances | |
1296 | * tcp_highest_sack_seq() when skb is highest_sack. | |
1297 | */ | |
1298 | tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, | |
59c9af42 | 1299 | start_seq, end_seq, dup_sack, pcount, |
740b0f18 | 1300 | &skb->skb_mstamp); |
4c90d3b3 NC |
1301 | |
1302 | if (skb == tp->lost_skb_hint) | |
0af2a0d0 NC |
1303 | tp->lost_cnt_hint += pcount; |
1304 | ||
832d11c5 IJ |
1305 | TCP_SKB_CB(prev)->end_seq += shifted; |
1306 | TCP_SKB_CB(skb)->seq += shifted; | |
1307 | ||
cd7d8498 ED |
1308 | tcp_skb_pcount_add(prev, pcount); |
1309 | BUG_ON(tcp_skb_pcount(skb) < pcount); | |
1310 | tcp_skb_pcount_add(skb, -pcount); | |
832d11c5 IJ |
1311 | |
1312 | /* When we're adding to gso_segs == 1, gso_size will be zero, | |
1313 | * in theory this shouldn't be necessary but as long as DSACK | |
1314 | * code can come after this skb later on it's better to keep | |
1315 | * setting gso_size to something. | |
1316 | */ | |
1317 | if (!skb_shinfo(prev)->gso_size) { | |
1318 | skb_shinfo(prev)->gso_size = mss; | |
c9af6db4 | 1319 | skb_shinfo(prev)->gso_type = sk->sk_gso_type; |
832d11c5 IJ |
1320 | } |
1321 | ||
1322 | /* CHECKME: To clear or not to clear? Mimics normal skb currently */ | |
cd7d8498 | 1323 | if (tcp_skb_pcount(skb) <= 1) { |
832d11c5 | 1324 | skb_shinfo(skb)->gso_size = 0; |
c9af6db4 | 1325 | skb_shinfo(skb)->gso_type = 0; |
832d11c5 IJ |
1326 | } |
1327 | ||
832d11c5 IJ |
1328 | /* Difference in this won't matter, both ACKed by the same cumul. ACK */ |
1329 | TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); | |
1330 | ||
832d11c5 IJ |
1331 | if (skb->len > 0) { |
1332 | BUG_ON(!tcp_skb_pcount(skb)); | |
111cc8b9 | 1333 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED); |
a2a385d6 | 1334 | return false; |
832d11c5 IJ |
1335 | } |
1336 | ||
1337 | /* Whole SKB was eaten :-) */ | |
1338 | ||
92ee76b6 IJ |
1339 | if (skb == tp->retransmit_skb_hint) |
1340 | tp->retransmit_skb_hint = prev; | |
92ee76b6 IJ |
1341 | if (skb == tp->lost_skb_hint) { |
1342 | tp->lost_skb_hint = prev; | |
1343 | tp->lost_cnt_hint -= tcp_skb_pcount(prev); | |
1344 | } | |
1345 | ||
5e8a402f ED |
1346 | TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
1347 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) | |
1348 | TCP_SKB_CB(prev)->end_seq++; | |
1349 | ||
832d11c5 IJ |
1350 | if (skb == tcp_highest_sack(sk)) |
1351 | tcp_advance_highest_sack(sk, skb); | |
1352 | ||
1353 | tcp_unlink_write_queue(skb, sk); | |
1354 | sk_wmem_free_skb(sk, skb); | |
1355 | ||
111cc8b9 IJ |
1356 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED); |
1357 | ||
a2a385d6 | 1358 | return true; |
832d11c5 IJ |
1359 | } |
1360 | ||
1361 | /* I wish gso_size would have a bit more sane initialization than | |
1362 | * something-or-zero which complicates things | |
1363 | */ | |
cf533ea5 | 1364 | static int tcp_skb_seglen(const struct sk_buff *skb) |
832d11c5 | 1365 | { |
775ffabf | 1366 | return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); |
832d11c5 IJ |
1367 | } |
1368 | ||
1369 | /* Shifting pages past head area doesn't work */ | |
cf533ea5 | 1370 | static int skb_can_shift(const struct sk_buff *skb) |
832d11c5 IJ |
1371 | { |
1372 | return !skb_headlen(skb) && skb_is_nonlinear(skb); | |
1373 | } | |
1374 | ||
1375 | /* Try collapsing SACK blocks spanning across multiple skbs to a single | |
1376 | * skb. | |
1377 | */ | |
1378 | static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, | |
a1197f5a | 1379 | struct tcp_sacktag_state *state, |
832d11c5 | 1380 | u32 start_seq, u32 end_seq, |
a2a385d6 | 1381 | bool dup_sack) |
832d11c5 IJ |
1382 | { |
1383 | struct tcp_sock *tp = tcp_sk(sk); | |
1384 | struct sk_buff *prev; | |
1385 | int mss; | |
1386 | int pcount = 0; | |
1387 | int len; | |
1388 | int in_sack; | |
1389 | ||
1390 | if (!sk_can_gso(sk)) | |
1391 | goto fallback; | |
1392 | ||
1393 | /* Normally R but no L won't result in plain S */ | |
1394 | if (!dup_sack && | |
9969ca5f | 1395 | (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) |
832d11c5 IJ |
1396 | goto fallback; |
1397 | if (!skb_can_shift(skb)) | |
1398 | goto fallback; | |
1399 | /* This frame is about to be dropped (was ACKed). */ | |
1400 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1401 | goto fallback; | |
1402 | ||
1403 | /* Can only happen with delayed DSACK + discard craziness */ | |
1404 | if (unlikely(skb == tcp_write_queue_head(sk))) | |
1405 | goto fallback; | |
1406 | prev = tcp_write_queue_prev(sk, skb); | |
1407 | ||
1408 | if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) | |
1409 | goto fallback; | |
1410 | ||
1411 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1412 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1413 | ||
1414 | if (in_sack) { | |
1415 | len = skb->len; | |
1416 | pcount = tcp_skb_pcount(skb); | |
775ffabf | 1417 | mss = tcp_skb_seglen(skb); |
832d11c5 IJ |
1418 | |
1419 | /* TODO: Fix DSACKs to not fragment already SACKed and we can | |
1420 | * drop this restriction as unnecessary | |
1421 | */ | |
775ffabf | 1422 | if (mss != tcp_skb_seglen(prev)) |
832d11c5 IJ |
1423 | goto fallback; |
1424 | } else { | |
1425 | if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) | |
1426 | goto noop; | |
1427 | /* CHECKME: This is non-MSS split case only?, this will | |
1428 | * cause skipped skbs due to advancing loop btw, original | |
1429 | * has that feature too | |
1430 | */ | |
1431 | if (tcp_skb_pcount(skb) <= 1) | |
1432 | goto noop; | |
1433 | ||
1434 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); | |
1435 | if (!in_sack) { | |
1436 | /* TODO: head merge to next could be attempted here | |
1437 | * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), | |
1438 | * though it might not be worth of the additional hassle | |
1439 | * | |
1440 | * ...we can probably just fallback to what was done | |
1441 | * previously. We could try merging non-SACKed ones | |
1442 | * as well but it probably isn't going to buy off | |
1443 | * because later SACKs might again split them, and | |
1444 | * it would make skb timestamp tracking considerably | |
1445 | * harder problem. | |
1446 | */ | |
1447 | goto fallback; | |
1448 | } | |
1449 | ||
1450 | len = end_seq - TCP_SKB_CB(skb)->seq; | |
1451 | BUG_ON(len < 0); | |
1452 | BUG_ON(len > skb->len); | |
1453 | ||
1454 | /* MSS boundaries should be honoured or else pcount will | |
1455 | * severely break even though it makes things bit trickier. | |
1456 | * Optimize common case to avoid most of the divides | |
1457 | */ | |
1458 | mss = tcp_skb_mss(skb); | |
1459 | ||
1460 | /* TODO: Fix DSACKs to not fragment already SACKed and we can | |
1461 | * drop this restriction as unnecessary | |
1462 | */ | |
775ffabf | 1463 | if (mss != tcp_skb_seglen(prev)) |
832d11c5 IJ |
1464 | goto fallback; |
1465 | ||
1466 | if (len == mss) { | |
1467 | pcount = 1; | |
1468 | } else if (len < mss) { | |
1469 | goto noop; | |
1470 | } else { | |
1471 | pcount = len / mss; | |
1472 | len = pcount * mss; | |
1473 | } | |
1474 | } | |
1475 | ||
4648dc97 NC |
1476 | /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ |
1477 | if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) | |
1478 | goto fallback; | |
1479 | ||
832d11c5 IJ |
1480 | if (!skb_shift(prev, skb, len)) |
1481 | goto fallback; | |
9ec06ff5 | 1482 | if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack)) |
832d11c5 IJ |
1483 | goto out; |
1484 | ||
1485 | /* Hole filled allows collapsing with the next as well, this is very | |
1486 | * useful when hole on every nth skb pattern happens | |
1487 | */ | |
1488 | if (prev == tcp_write_queue_tail(sk)) | |
1489 | goto out; | |
1490 | skb = tcp_write_queue_next(sk, prev); | |
1491 | ||
f0bc52f3 IJ |
1492 | if (!skb_can_shift(skb) || |
1493 | (skb == tcp_send_head(sk)) || | |
1494 | ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || | |
775ffabf | 1495 | (mss != tcp_skb_seglen(skb))) |
832d11c5 IJ |
1496 | goto out; |
1497 | ||
1498 | len = skb->len; | |
1499 | if (skb_shift(prev, skb, len)) { | |
1500 | pcount += tcp_skb_pcount(skb); | |
9ec06ff5 | 1501 | tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0); |
832d11c5 IJ |
1502 | } |
1503 | ||
1504 | out: | |
a1197f5a | 1505 | state->fack_count += pcount; |
832d11c5 IJ |
1506 | return prev; |
1507 | ||
1508 | noop: | |
1509 | return skb; | |
1510 | ||
1511 | fallback: | |
111cc8b9 | 1512 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); |
832d11c5 IJ |
1513 | return NULL; |
1514 | } | |
1515 | ||
68f8353b IJ |
1516 | static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, |
1517 | struct tcp_sack_block *next_dup, | |
a1197f5a | 1518 | struct tcp_sacktag_state *state, |
68f8353b | 1519 | u32 start_seq, u32 end_seq, |
a2a385d6 | 1520 | bool dup_sack_in) |
68f8353b | 1521 | { |
832d11c5 IJ |
1522 | struct tcp_sock *tp = tcp_sk(sk); |
1523 | struct sk_buff *tmp; | |
1524 | ||
68f8353b IJ |
1525 | tcp_for_write_queue_from(skb, sk) { |
1526 | int in_sack = 0; | |
a2a385d6 | 1527 | bool dup_sack = dup_sack_in; |
68f8353b IJ |
1528 | |
1529 | if (skb == tcp_send_head(sk)) | |
1530 | break; | |
1531 | ||
1532 | /* queue is in-order => we can short-circuit the walk early */ | |
1533 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) | |
1534 | break; | |
1535 | ||
1536 | if ((next_dup != NULL) && | |
1537 | before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { | |
1538 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1539 | next_dup->start_seq, | |
1540 | next_dup->end_seq); | |
1541 | if (in_sack > 0) | |
a2a385d6 | 1542 | dup_sack = true; |
68f8353b IJ |
1543 | } |
1544 | ||
832d11c5 IJ |
1545 | /* skb reference here is a bit tricky to get right, since |
1546 | * shifting can eat and free both this skb and the next, | |
1547 | * so not even _safe variant of the loop is enough. | |
1548 | */ | |
1549 | if (in_sack <= 0) { | |
a1197f5a IJ |
1550 | tmp = tcp_shift_skb_data(sk, skb, state, |
1551 | start_seq, end_seq, dup_sack); | |
832d11c5 IJ |
1552 | if (tmp != NULL) { |
1553 | if (tmp != skb) { | |
1554 | skb = tmp; | |
1555 | continue; | |
1556 | } | |
1557 | ||
1558 | in_sack = 0; | |
1559 | } else { | |
1560 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1561 | start_seq, | |
1562 | end_seq); | |
1563 | } | |
1564 | } | |
1565 | ||
68f8353b IJ |
1566 | if (unlikely(in_sack < 0)) |
1567 | break; | |
1568 | ||
832d11c5 | 1569 | if (in_sack) { |
cc9a672e NC |
1570 | TCP_SKB_CB(skb)->sacked = |
1571 | tcp_sacktag_one(sk, | |
1572 | state, | |
1573 | TCP_SKB_CB(skb)->sacked, | |
1574 | TCP_SKB_CB(skb)->seq, | |
1575 | TCP_SKB_CB(skb)->end_seq, | |
1576 | dup_sack, | |
59c9af42 | 1577 | tcp_skb_pcount(skb), |
740b0f18 | 1578 | &skb->skb_mstamp); |
68f8353b | 1579 | |
832d11c5 IJ |
1580 | if (!before(TCP_SKB_CB(skb)->seq, |
1581 | tcp_highest_sack_seq(tp))) | |
1582 | tcp_advance_highest_sack(sk, skb); | |
1583 | } | |
1584 | ||
a1197f5a | 1585 | state->fack_count += tcp_skb_pcount(skb); |
68f8353b IJ |
1586 | } |
1587 | return skb; | |
1588 | } | |
1589 | ||
1590 | /* Avoid all extra work that is being done by sacktag while walking in | |
1591 | * a normal way | |
1592 | */ | |
1593 | static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, | |
a1197f5a IJ |
1594 | struct tcp_sacktag_state *state, |
1595 | u32 skip_to_seq) | |
68f8353b IJ |
1596 | { |
1597 | tcp_for_write_queue_from(skb, sk) { | |
1598 | if (skb == tcp_send_head(sk)) | |
1599 | break; | |
1600 | ||
e8bae275 | 1601 | if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) |
68f8353b | 1602 | break; |
d152a7d8 | 1603 | |
a1197f5a | 1604 | state->fack_count += tcp_skb_pcount(skb); |
68f8353b IJ |
1605 | } |
1606 | return skb; | |
1607 | } | |
1608 | ||
1609 | static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, | |
1610 | struct sock *sk, | |
1611 | struct tcp_sack_block *next_dup, | |
a1197f5a IJ |
1612 | struct tcp_sacktag_state *state, |
1613 | u32 skip_to_seq) | |
68f8353b IJ |
1614 | { |
1615 | if (next_dup == NULL) | |
1616 | return skb; | |
1617 | ||
1618 | if (before(next_dup->start_seq, skip_to_seq)) { | |
a1197f5a IJ |
1619 | skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); |
1620 | skb = tcp_sacktag_walk(skb, sk, NULL, state, | |
1621 | next_dup->start_seq, next_dup->end_seq, | |
1622 | 1); | |
68f8353b IJ |
1623 | } |
1624 | ||
1625 | return skb; | |
1626 | } | |
1627 | ||
cf533ea5 | 1628 | static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) |
68f8353b IJ |
1629 | { |
1630 | return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); | |
1631 | } | |
1632 | ||
1da177e4 | 1633 | static int |
cf533ea5 | 1634 | tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, |
740b0f18 | 1635 | u32 prior_snd_una, long *sack_rtt_us) |
1da177e4 LT |
1636 | { |
1637 | struct tcp_sock *tp = tcp_sk(sk); | |
cf533ea5 ED |
1638 | const unsigned char *ptr = (skb_transport_header(ack_skb) + |
1639 | TCP_SKB_CB(ack_skb)->sacked); | |
fd6dad61 | 1640 | struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); |
4389dded | 1641 | struct tcp_sack_block sp[TCP_NUM_SACKS]; |
68f8353b | 1642 | struct tcp_sack_block *cache; |
a1197f5a | 1643 | struct tcp_sacktag_state state; |
68f8353b | 1644 | struct sk_buff *skb; |
4389dded | 1645 | int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); |
fd6dad61 | 1646 | int used_sacks; |
a2a385d6 | 1647 | bool found_dup_sack = false; |
68f8353b | 1648 | int i, j; |
fda03fbb | 1649 | int first_sack_index; |
1da177e4 | 1650 | |
a1197f5a IJ |
1651 | state.flag = 0; |
1652 | state.reord = tp->packets_out; | |
740b0f18 | 1653 | state.rtt_us = -1L; |
a1197f5a | 1654 | |
d738cd8f | 1655 | if (!tp->sacked_out) { |
de83c058 IJ |
1656 | if (WARN_ON(tp->fackets_out)) |
1657 | tp->fackets_out = 0; | |
6859d494 | 1658 | tcp_highest_sack_reset(sk); |
d738cd8f | 1659 | } |
1da177e4 | 1660 | |
1ed83465 | 1661 | found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, |
d06e021d DM |
1662 | num_sacks, prior_snd_una); |
1663 | if (found_dup_sack) | |
a1197f5a | 1664 | state.flag |= FLAG_DSACKING_ACK; |
6f74651a BE |
1665 | |
1666 | /* Eliminate too old ACKs, but take into | |
1667 | * account more or less fresh ones, they can | |
1668 | * contain valid SACK info. | |
1669 | */ | |
1670 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) | |
1671 | return 0; | |
1672 |