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> |
5ffc02a1 | 71 | #include <net/dst.h> |
1da177e4 LT |
72 | #include <net/tcp.h> |
73 | #include <net/inet_common.h> | |
74 | #include <linux/ipsec.h> | |
75 | #include <asm/unaligned.h> | |
1a2449a8 | 76 | #include <net/netdma.h> |
1da177e4 | 77 | |
ab32ea5d BH |
78 | int sysctl_tcp_timestamps __read_mostly = 1; |
79 | int sysctl_tcp_window_scaling __read_mostly = 1; | |
80 | int sysctl_tcp_sack __read_mostly = 1; | |
81 | int sysctl_tcp_fack __read_mostly = 1; | |
82 | int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; | |
4bc2f18b | 83 | EXPORT_SYMBOL(sysctl_tcp_reordering); |
ab32ea5d BH |
84 | int sysctl_tcp_dsack __read_mostly = 1; |
85 | int sysctl_tcp_app_win __read_mostly = 31; | |
b49960a0 | 86 | int sysctl_tcp_adv_win_scale __read_mostly = 1; |
4bc2f18b | 87 | EXPORT_SYMBOL(sysctl_tcp_adv_win_scale); |
1da177e4 | 88 | |
282f23c6 ED |
89 | /* rfc5961 challenge ack rate limiting */ |
90 | int sysctl_tcp_challenge_ack_limit = 100; | |
91 | ||
ab32ea5d BH |
92 | int sysctl_tcp_stdurg __read_mostly; |
93 | int sysctl_tcp_rfc1337 __read_mostly; | |
94 | int sysctl_tcp_max_orphans __read_mostly = NR_FILE; | |
c96fd3d4 | 95 | int sysctl_tcp_frto __read_mostly = 2; |
1da177e4 | 96 | |
7e380175 AP |
97 | int sysctl_tcp_thin_dupack __read_mostly; |
98 | ||
ab32ea5d | 99 | int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; |
6ba8a3b1 | 100 | int sysctl_tcp_early_retrans __read_mostly = 3; |
1da177e4 | 101 | |
1da177e4 LT |
102 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ |
103 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ | |
104 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ | |
105 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ | |
106 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ | |
107 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ | |
108 | #define FLAG_ECE 0x40 /* ECE in this ACK */ | |
1da177e4 | 109 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ |
e33099f9 | 110 | #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */ |
2e605294 | 111 | #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ |
564262c1 | 112 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ |
cadbd031 | 113 | #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ |
12fb3dd9 | 114 | #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */ |
1da177e4 LT |
115 | |
116 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) | |
117 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) | |
118 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) | |
119 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) | |
120 | ||
1da177e4 | 121 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) |
bdf1ee5d | 122 | #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) |
1da177e4 | 123 | |
e905a9ed | 124 | /* Adapt the MSS value used to make delayed ack decision to the |
1da177e4 | 125 | * real world. |
e905a9ed | 126 | */ |
056834d9 | 127 | static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 128 | { |
463c84b9 | 129 | struct inet_connection_sock *icsk = inet_csk(sk); |
e905a9ed | 130 | const unsigned int lss = icsk->icsk_ack.last_seg_size; |
463c84b9 | 131 | unsigned int len; |
1da177e4 | 132 | |
e905a9ed | 133 | icsk->icsk_ack.last_seg_size = 0; |
1da177e4 LT |
134 | |
135 | /* skb->len may jitter because of SACKs, even if peer | |
136 | * sends good full-sized frames. | |
137 | */ | |
056834d9 | 138 | len = skb_shinfo(skb)->gso_size ? : skb->len; |
463c84b9 ACM |
139 | if (len >= icsk->icsk_ack.rcv_mss) { |
140 | icsk->icsk_ack.rcv_mss = len; | |
1da177e4 LT |
141 | } else { |
142 | /* Otherwise, we make more careful check taking into account, | |
143 | * that SACKs block is variable. | |
144 | * | |
145 | * "len" is invariant segment length, including TCP header. | |
146 | */ | |
9c70220b | 147 | len += skb->data - skb_transport_header(skb); |
bee7ca9e | 148 | if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || |
1da177e4 LT |
149 | /* If PSH is not set, packet should be |
150 | * full sized, provided peer TCP is not badly broken. | |
151 | * This observation (if it is correct 8)) allows | |
152 | * to handle super-low mtu links fairly. | |
153 | */ | |
154 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && | |
aa8223c7 | 155 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { |
1da177e4 LT |
156 | /* Subtract also invariant (if peer is RFC compliant), |
157 | * tcp header plus fixed timestamp option length. | |
158 | * Resulting "len" is MSS free of SACK jitter. | |
159 | */ | |
463c84b9 ACM |
160 | len -= tcp_sk(sk)->tcp_header_len; |
161 | icsk->icsk_ack.last_seg_size = len; | |
1da177e4 | 162 | if (len == lss) { |
463c84b9 | 163 | icsk->icsk_ack.rcv_mss = len; |
1da177e4 LT |
164 | return; |
165 | } | |
166 | } | |
1ef9696c AK |
167 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) |
168 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; | |
463c84b9 | 169 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
1da177e4 LT |
170 | } |
171 | } | |
172 | ||
463c84b9 | 173 | static void tcp_incr_quickack(struct sock *sk) |
1da177e4 | 174 | { |
463c84b9 | 175 | struct inet_connection_sock *icsk = inet_csk(sk); |
95c96174 | 176 | unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); |
1da177e4 | 177 | |
056834d9 IJ |
178 | if (quickacks == 0) |
179 | quickacks = 2; | |
463c84b9 ACM |
180 | if (quickacks > icsk->icsk_ack.quick) |
181 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); | |
1da177e4 LT |
182 | } |
183 | ||
1b9f4092 | 184 | static void tcp_enter_quickack_mode(struct sock *sk) |
1da177e4 | 185 | { |
463c84b9 ACM |
186 | struct inet_connection_sock *icsk = inet_csk(sk); |
187 | tcp_incr_quickack(sk); | |
188 | icsk->icsk_ack.pingpong = 0; | |
189 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 LT |
190 | } |
191 | ||
192 | /* Send ACKs quickly, if "quick" count is not exhausted | |
193 | * and the session is not interactive. | |
194 | */ | |
195 | ||
a2a385d6 | 196 | static inline bool tcp_in_quickack_mode(const struct sock *sk) |
1da177e4 | 197 | { |
463c84b9 | 198 | const struct inet_connection_sock *icsk = inet_csk(sk); |
a2a385d6 | 199 | |
463c84b9 | 200 | return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; |
1da177e4 LT |
201 | } |
202 | ||
bdf1ee5d IJ |
203 | static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) |
204 | { | |
056834d9 | 205 | if (tp->ecn_flags & TCP_ECN_OK) |
bdf1ee5d IJ |
206 | tp->ecn_flags |= TCP_ECN_QUEUE_CWR; |
207 | } | |
208 | ||
cf533ea5 | 209 | static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb) |
bdf1ee5d IJ |
210 | { |
211 | if (tcp_hdr(skb)->cwr) | |
212 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
213 | } | |
214 | ||
215 | static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) | |
216 | { | |
217 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
218 | } | |
219 | ||
7a269ffa | 220 | static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) |
bdf1ee5d | 221 | { |
7a269ffa ED |
222 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
223 | return; | |
224 | ||
b82d1bb4 | 225 | switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { |
7a269ffa | 226 | case INET_ECN_NOT_ECT: |
bdf1ee5d | 227 | /* Funny extension: if ECT is not set on a segment, |
7a269ffa ED |
228 | * and we already seen ECT on a previous segment, |
229 | * it is probably a retransmit. | |
230 | */ | |
231 | if (tp->ecn_flags & TCP_ECN_SEEN) | |
bdf1ee5d | 232 | tcp_enter_quickack_mode((struct sock *)tp); |
7a269ffa ED |
233 | break; |
234 | case INET_ECN_CE: | |
aae06bf5 ED |
235 | if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { |
236 | /* Better not delay acks, sender can have a very low cwnd */ | |
237 | tcp_enter_quickack_mode((struct sock *)tp); | |
238 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; | |
239 | } | |
7a269ffa ED |
240 | /* fallinto */ |
241 | default: | |
242 | tp->ecn_flags |= TCP_ECN_SEEN; | |
bdf1ee5d IJ |
243 | } |
244 | } | |
245 | ||
cf533ea5 | 246 | static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) |
bdf1ee5d | 247 | { |
056834d9 | 248 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) |
bdf1ee5d IJ |
249 | tp->ecn_flags &= ~TCP_ECN_OK; |
250 | } | |
251 | ||
cf533ea5 | 252 | static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) |
bdf1ee5d | 253 | { |
056834d9 | 254 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) |
bdf1ee5d IJ |
255 | tp->ecn_flags &= ~TCP_ECN_OK; |
256 | } | |
257 | ||
a2a385d6 | 258 | static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) |
bdf1ee5d | 259 | { |
056834d9 | 260 | if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) |
a2a385d6 ED |
261 | return true; |
262 | return false; | |
bdf1ee5d IJ |
263 | } |
264 | ||
1da177e4 LT |
265 | /* Buffer size and advertised window tuning. |
266 | * | |
267 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. | |
268 | */ | |
269 | ||
270 | static void tcp_fixup_sndbuf(struct sock *sk) | |
271 | { | |
87fb4b7b | 272 | int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER); |
1da177e4 | 273 | |
06a59ecb ED |
274 | sndmem *= TCP_INIT_CWND; |
275 | if (sk->sk_sndbuf < sndmem) | |
276 | sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); | |
1da177e4 LT |
277 | } |
278 | ||
279 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) | |
280 | * | |
281 | * All tcp_full_space() is split to two parts: "network" buffer, allocated | |
282 | * forward and advertised in receiver window (tp->rcv_wnd) and | |
283 | * "application buffer", required to isolate scheduling/application | |
284 | * latencies from network. | |
285 | * window_clamp is maximal advertised window. It can be less than | |
286 | * tcp_full_space(), in this case tcp_full_space() - window_clamp | |
287 | * is reserved for "application" buffer. The less window_clamp is | |
288 | * the smoother our behaviour from viewpoint of network, but the lower | |
289 | * throughput and the higher sensitivity of the connection to losses. 8) | |
290 | * | |
291 | * rcv_ssthresh is more strict window_clamp used at "slow start" | |
292 | * phase to predict further behaviour of this connection. | |
293 | * It is used for two goals: | |
294 | * - to enforce header prediction at sender, even when application | |
295 | * requires some significant "application buffer". It is check #1. | |
296 | * - to prevent pruning of receive queue because of misprediction | |
297 | * of receiver window. Check #2. | |
298 | * | |
299 | * The scheme does not work when sender sends good segments opening | |
caa20d9a | 300 | * window and then starts to feed us spaghetti. But it should work |
1da177e4 LT |
301 | * in common situations. Otherwise, we have to rely on queue collapsing. |
302 | */ | |
303 | ||
304 | /* Slow part of check#2. */ | |
9e412ba7 | 305 | static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 306 | { |
9e412ba7 | 307 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 308 | /* Optimize this! */ |
dfd4f0ae ED |
309 | int truesize = tcp_win_from_space(skb->truesize) >> 1; |
310 | int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; | |
1da177e4 LT |
311 | |
312 | while (tp->rcv_ssthresh <= window) { | |
313 | if (truesize <= skb->len) | |
463c84b9 | 314 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; |
1da177e4 LT |
315 | |
316 | truesize >>= 1; | |
317 | window >>= 1; | |
318 | } | |
319 | return 0; | |
320 | } | |
321 | ||
cf533ea5 | 322 | static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 323 | { |
9e412ba7 IJ |
324 | struct tcp_sock *tp = tcp_sk(sk); |
325 | ||
1da177e4 LT |
326 | /* Check #1 */ |
327 | if (tp->rcv_ssthresh < tp->window_clamp && | |
328 | (int)tp->rcv_ssthresh < tcp_space(sk) && | |
180d8cd9 | 329 | !sk_under_memory_pressure(sk)) { |
1da177e4 LT |
330 | int incr; |
331 | ||
332 | /* Check #2. Increase window, if skb with such overhead | |
333 | * will fit to rcvbuf in future. | |
334 | */ | |
335 | if (tcp_win_from_space(skb->truesize) <= skb->len) | |
056834d9 | 336 | incr = 2 * tp->advmss; |
1da177e4 | 337 | else |
9e412ba7 | 338 | incr = __tcp_grow_window(sk, skb); |
1da177e4 LT |
339 | |
340 | if (incr) { | |
4d846f02 | 341 | incr = max_t(int, incr, 2 * skb->len); |
056834d9 IJ |
342 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, |
343 | tp->window_clamp); | |
463c84b9 | 344 | inet_csk(sk)->icsk_ack.quick |= 1; |
1da177e4 LT |
345 | } |
346 | } | |
347 | } | |
348 | ||
349 | /* 3. Tuning rcvbuf, when connection enters established state. */ | |
1da177e4 LT |
350 | static void tcp_fixup_rcvbuf(struct sock *sk) |
351 | { | |
e9266a02 | 352 | u32 mss = tcp_sk(sk)->advmss; |
e9266a02 | 353 | int rcvmem; |
1da177e4 | 354 | |
85f16525 YC |
355 | rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) * |
356 | tcp_default_init_rwnd(mss); | |
e9266a02 ED |
357 | |
358 | if (sk->sk_rcvbuf < rcvmem) | |
359 | sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]); | |
1da177e4 LT |
360 | } |
361 | ||
caa20d9a | 362 | /* 4. Try to fixup all. It is made immediately after connection enters |
1da177e4 LT |
363 | * established state. |
364 | */ | |
10467163 | 365 | void tcp_init_buffer_space(struct sock *sk) |
1da177e4 LT |
366 | { |
367 | struct tcp_sock *tp = tcp_sk(sk); | |
368 | int maxwin; | |
369 | ||
370 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) | |
371 | tcp_fixup_rcvbuf(sk); | |
372 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) | |
373 | tcp_fixup_sndbuf(sk); | |
374 | ||
375 | tp->rcvq_space.space = tp->rcv_wnd; | |
376 | ||
377 | maxwin = tcp_full_space(sk); | |
378 | ||
379 | if (tp->window_clamp >= maxwin) { | |
380 | tp->window_clamp = maxwin; | |
381 | ||
382 | if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) | |
383 | tp->window_clamp = max(maxwin - | |
384 | (maxwin >> sysctl_tcp_app_win), | |
385 | 4 * tp->advmss); | |
386 | } | |
387 | ||
388 | /* Force reservation of one segment. */ | |
389 | if (sysctl_tcp_app_win && | |
390 | tp->window_clamp > 2 * tp->advmss && | |
391 | tp->window_clamp + tp->advmss > maxwin) | |
392 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); | |
393 | ||
394 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); | |
395 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
396 | } | |
397 | ||
1da177e4 | 398 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ |
9e412ba7 | 399 | static void tcp_clamp_window(struct sock *sk) |
1da177e4 | 400 | { |
9e412ba7 | 401 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 402 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 403 | |
6687e988 | 404 | icsk->icsk_ack.quick = 0; |
1da177e4 | 405 | |
326f36e9 JH |
406 | if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && |
407 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && | |
180d8cd9 GC |
408 | !sk_under_memory_pressure(sk) && |
409 | sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { | |
326f36e9 JH |
410 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), |
411 | sysctl_tcp_rmem[2]); | |
1da177e4 | 412 | } |
326f36e9 | 413 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) |
056834d9 | 414 | tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); |
1da177e4 LT |
415 | } |
416 | ||
40efc6fa SH |
417 | /* Initialize RCV_MSS value. |
418 | * RCV_MSS is an our guess about MSS used by the peer. | |
419 | * We haven't any direct information about the MSS. | |
420 | * It's better to underestimate the RCV_MSS rather than overestimate. | |
421 | * Overestimations make us ACKing less frequently than needed. | |
422 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). | |
423 | */ | |
424 | void tcp_initialize_rcv_mss(struct sock *sk) | |
425 | { | |
cf533ea5 | 426 | const struct tcp_sock *tp = tcp_sk(sk); |
40efc6fa SH |
427 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); |
428 | ||
056834d9 | 429 | hint = min(hint, tp->rcv_wnd / 2); |
bee7ca9e | 430 | hint = min(hint, TCP_MSS_DEFAULT); |
40efc6fa SH |
431 | hint = max(hint, TCP_MIN_MSS); |
432 | ||
433 | inet_csk(sk)->icsk_ack.rcv_mss = hint; | |
434 | } | |
4bc2f18b | 435 | EXPORT_SYMBOL(tcp_initialize_rcv_mss); |
40efc6fa | 436 | |
1da177e4 LT |
437 | /* Receiver "autotuning" code. |
438 | * | |
439 | * The algorithm for RTT estimation w/o timestamps is based on | |
440 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. | |
631dd1a8 | 441 | * <http://public.lanl.gov/radiant/pubs.html#DRS> |
1da177e4 LT |
442 | * |
443 | * More detail on this code can be found at | |
631dd1a8 | 444 | * <http://staff.psc.edu/jheffner/>, |
1da177e4 LT |
445 | * though this reference is out of date. A new paper |
446 | * is pending. | |
447 | */ | |
448 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) | |
449 | { | |
450 | u32 new_sample = tp->rcv_rtt_est.rtt; | |
451 | long m = sample; | |
452 | ||
453 | if (m == 0) | |
454 | m = 1; | |
455 | ||
456 | if (new_sample != 0) { | |
457 | /* If we sample in larger samples in the non-timestamp | |
458 | * case, we could grossly overestimate the RTT especially | |
459 | * with chatty applications or bulk transfer apps which | |
460 | * are stalled on filesystem I/O. | |
461 | * | |
462 | * Also, since we are only going for a minimum in the | |
31f34269 | 463 | * non-timestamp case, we do not smooth things out |
caa20d9a | 464 | * else with timestamps disabled convergence takes too |
1da177e4 LT |
465 | * long. |
466 | */ | |
467 | if (!win_dep) { | |
468 | m -= (new_sample >> 3); | |
469 | new_sample += m; | |
18a223e0 NC |
470 | } else { |
471 | m <<= 3; | |
472 | if (m < new_sample) | |
473 | new_sample = m; | |
474 | } | |
1da177e4 | 475 | } else { |
caa20d9a | 476 | /* No previous measure. */ |
1da177e4 LT |
477 | new_sample = m << 3; |
478 | } | |
479 | ||
480 | if (tp->rcv_rtt_est.rtt != new_sample) | |
481 | tp->rcv_rtt_est.rtt = new_sample; | |
482 | } | |
483 | ||
484 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) | |
485 | { | |
486 | if (tp->rcv_rtt_est.time == 0) | |
487 | goto new_measure; | |
488 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) | |
489 | return; | |
651913ce | 490 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1); |
1da177e4 LT |
491 | |
492 | new_measure: | |
493 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; | |
494 | tp->rcv_rtt_est.time = tcp_time_stamp; | |
495 | } | |
496 | ||
056834d9 IJ |
497 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, |
498 | const struct sk_buff *skb) | |
1da177e4 | 499 | { |
463c84b9 | 500 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
501 | if (tp->rx_opt.rcv_tsecr && |
502 | (TCP_SKB_CB(skb)->end_seq - | |
463c84b9 | 503 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) |
1da177e4 LT |
504 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); |
505 | } | |
506 | ||
507 | /* | |
508 | * This function should be called every time data is copied to user space. | |
509 | * It calculates the appropriate TCP receive buffer space. | |
510 | */ | |
511 | void tcp_rcv_space_adjust(struct sock *sk) | |
512 | { | |
513 | struct tcp_sock *tp = tcp_sk(sk); | |
514 | int time; | |
515 | int space; | |
e905a9ed | 516 | |
1da177e4 LT |
517 | if (tp->rcvq_space.time == 0) |
518 | goto new_measure; | |
e905a9ed | 519 | |
1da177e4 | 520 | time = tcp_time_stamp - tp->rcvq_space.time; |
056834d9 | 521 | if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) |
1da177e4 | 522 | return; |
e905a9ed | 523 | |
1da177e4 LT |
524 | space = 2 * (tp->copied_seq - tp->rcvq_space.seq); |
525 | ||
526 | space = max(tp->rcvq_space.space, space); | |
527 | ||
528 | if (tp->rcvq_space.space != space) { | |
529 | int rcvmem; | |
530 | ||
531 | tp->rcvq_space.space = space; | |
532 | ||
6fcf9412 JH |
533 | if (sysctl_tcp_moderate_rcvbuf && |
534 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { | |
1da177e4 LT |
535 | int new_clamp = space; |
536 | ||
537 | /* Receive space grows, normalize in order to | |
538 | * take into account packet headers and sk_buff | |
539 | * structure overhead. | |
540 | */ | |
541 | space /= tp->advmss; | |
542 | if (!space) | |
543 | space = 1; | |
87fb4b7b | 544 | rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER); |
1da177e4 LT |
545 | while (tcp_win_from_space(rcvmem) < tp->advmss) |
546 | rcvmem += 128; | |
547 | space *= rcvmem; | |
548 | space = min(space, sysctl_tcp_rmem[2]); | |
549 | if (space > sk->sk_rcvbuf) { | |
550 | sk->sk_rcvbuf = space; | |
551 | ||
552 | /* Make the window clamp follow along. */ | |
553 | tp->window_clamp = new_clamp; | |
554 | } | |
555 | } | |
556 | } | |
e905a9ed | 557 | |
1da177e4 LT |
558 | new_measure: |
559 | tp->rcvq_space.seq = tp->copied_seq; | |
560 | tp->rcvq_space.time = tcp_time_stamp; | |
561 | } | |
562 | ||
563 | /* There is something which you must keep in mind when you analyze the | |
564 | * behavior of the tp->ato delayed ack timeout interval. When a | |
565 | * connection starts up, we want to ack as quickly as possible. The | |
566 | * problem is that "good" TCP's do slow start at the beginning of data | |
567 | * transmission. The means that until we send the first few ACK's the | |
568 | * sender will sit on his end and only queue most of his data, because | |
569 | * he can only send snd_cwnd unacked packets at any given time. For | |
570 | * each ACK we send, he increments snd_cwnd and transmits more of his | |
571 | * queue. -DaveM | |
572 | */ | |
9e412ba7 | 573 | static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 574 | { |
9e412ba7 | 575 | struct tcp_sock *tp = tcp_sk(sk); |
463c84b9 | 576 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
577 | u32 now; |
578 | ||
463c84b9 | 579 | inet_csk_schedule_ack(sk); |
1da177e4 | 580 | |
463c84b9 | 581 | tcp_measure_rcv_mss(sk, skb); |
1da177e4 LT |
582 | |
583 | tcp_rcv_rtt_measure(tp); | |
e905a9ed | 584 | |
1da177e4 LT |
585 | now = tcp_time_stamp; |
586 | ||
463c84b9 | 587 | if (!icsk->icsk_ack.ato) { |
1da177e4 LT |
588 | /* The _first_ data packet received, initialize |
589 | * delayed ACK engine. | |
590 | */ | |
463c84b9 ACM |
591 | tcp_incr_quickack(sk); |
592 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 | 593 | } else { |
463c84b9 | 594 | int m = now - icsk->icsk_ack.lrcvtime; |
1da177e4 | 595 | |
056834d9 | 596 | if (m <= TCP_ATO_MIN / 2) { |
1da177e4 | 597 | /* The fastest case is the first. */ |
463c84b9 ACM |
598 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; |
599 | } else if (m < icsk->icsk_ack.ato) { | |
600 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; | |
601 | if (icsk->icsk_ack.ato > icsk->icsk_rto) | |
602 | icsk->icsk_ack.ato = icsk->icsk_rto; | |
603 | } else if (m > icsk->icsk_rto) { | |
caa20d9a | 604 | /* Too long gap. Apparently sender failed to |
1da177e4 LT |
605 | * restart window, so that we send ACKs quickly. |
606 | */ | |
463c84b9 | 607 | tcp_incr_quickack(sk); |
3ab224be | 608 | sk_mem_reclaim(sk); |
1da177e4 LT |
609 | } |
610 | } | |
463c84b9 | 611 | icsk->icsk_ack.lrcvtime = now; |
1da177e4 LT |
612 | |
613 | TCP_ECN_check_ce(tp, skb); | |
614 | ||
615 | if (skb->len >= 128) | |
9e412ba7 | 616 | tcp_grow_window(sk, skb); |
1da177e4 LT |
617 | } |
618 | ||
1da177e4 LT |
619 | /* Called to compute a smoothed rtt estimate. The data fed to this |
620 | * routine either comes from timestamps, or from segments that were | |
621 | * known _not_ to have been retransmitted [see Karn/Partridge | |
622 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 | |
623 | * piece by Van Jacobson. | |
624 | * NOTE: the next three routines used to be one big routine. | |
625 | * To save cycles in the RFC 1323 implementation it was better to break | |
626 | * it up into three procedures. -- erics | |
627 | */ | |
2d2abbab | 628 | static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) |
1da177e4 | 629 | { |
6687e988 | 630 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
631 | long m = mrtt; /* RTT */ |
632 | ||
1da177e4 LT |
633 | /* The following amusing code comes from Jacobson's |
634 | * article in SIGCOMM '88. Note that rtt and mdev | |
635 | * are scaled versions of rtt and mean deviation. | |
e905a9ed | 636 | * This is designed to be as fast as possible |
1da177e4 LT |
637 | * m stands for "measurement". |
638 | * | |
639 | * On a 1990 paper the rto value is changed to: | |
640 | * RTO = rtt + 4 * mdev | |
641 | * | |
642 | * Funny. This algorithm seems to be very broken. | |
643 | * These formulae increase RTO, when it should be decreased, increase | |
31f34269 | 644 | * too slowly, when it should be increased quickly, decrease too quickly |
1da177e4 LT |
645 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely |
646 | * does not matter how to _calculate_ it. Seems, it was trap | |
647 | * that VJ failed to avoid. 8) | |
648 | */ | |
2de979bd | 649 | if (m == 0) |
1da177e4 LT |
650 | m = 1; |
651 | if (tp->srtt != 0) { | |
652 | m -= (tp->srtt >> 3); /* m is now error in rtt est */ | |
653 | tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ | |
654 | if (m < 0) { | |
655 | m = -m; /* m is now abs(error) */ | |
656 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
657 | /* This is similar to one of Eifel findings. | |
658 | * Eifel blocks mdev updates when rtt decreases. | |
659 | * This solution is a bit different: we use finer gain | |
660 | * for mdev in this case (alpha*beta). | |
661 | * Like Eifel it also prevents growth of rto, | |
662 | * but also it limits too fast rto decreases, | |
663 | * happening in pure Eifel. | |
664 | */ | |
665 | if (m > 0) | |
666 | m >>= 3; | |
667 | } else { | |
668 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
669 | } | |
670 | tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ | |
671 | if (tp->mdev > tp->mdev_max) { | |
672 | tp->mdev_max = tp->mdev; | |
673 | if (tp->mdev_max > tp->rttvar) | |
674 | tp->rttvar = tp->mdev_max; | |
675 | } | |
676 | if (after(tp->snd_una, tp->rtt_seq)) { | |
677 | if (tp->mdev_max < tp->rttvar) | |
056834d9 | 678 | tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2; |
1da177e4 | 679 | tp->rtt_seq = tp->snd_nxt; |
05bb1fad | 680 | tp->mdev_max = tcp_rto_min(sk); |
1da177e4 LT |
681 | } |
682 | } else { | |
683 | /* no previous measure. */ | |
056834d9 IJ |
684 | tp->srtt = m << 3; /* take the measured time to be rtt */ |
685 | tp->mdev = m << 1; /* make sure rto = 3*rtt */ | |
05bb1fad | 686 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); |
1da177e4 LT |
687 | tp->rtt_seq = tp->snd_nxt; |
688 | } | |
1da177e4 LT |
689 | } |
690 | ||
691 | /* Calculate rto without backoff. This is the second half of Van Jacobson's | |
692 | * routine referred to above. | |
693 | */ | |
4aabd8ef | 694 | void tcp_set_rto(struct sock *sk) |
1da177e4 | 695 | { |
463c84b9 | 696 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
697 | /* Old crap is replaced with new one. 8) |
698 | * | |
699 | * More seriously: | |
700 | * 1. If rtt variance happened to be less 50msec, it is hallucination. | |
701 | * It cannot be less due to utterly erratic ACK generation made | |
702 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ | |
703 | * to do with delayed acks, because at cwnd>2 true delack timeout | |
704 | * is invisible. Actually, Linux-2.4 also generates erratic | |
caa20d9a | 705 | * ACKs in some circumstances. |
1da177e4 | 706 | */ |
f1ecd5d9 | 707 | inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); |
1da177e4 LT |
708 | |
709 | /* 2. Fixups made earlier cannot be right. | |
710 | * If we do not estimate RTO correctly without them, | |
711 | * all the algo is pure shit and should be replaced | |
caa20d9a | 712 | * with correct one. It is exactly, which we pretend to do. |
1da177e4 | 713 | */ |
1da177e4 | 714 | |
ee6aac59 IJ |
715 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo |
716 | * guarantees that rto is higher. | |
717 | */ | |
f1ecd5d9 | 718 | tcp_bound_rto(sk); |
1da177e4 LT |
719 | } |
720 | ||
cf533ea5 | 721 | __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) |
1da177e4 LT |
722 | { |
723 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); | |
724 | ||
22b71c8f | 725 | if (!cwnd) |
442b9635 | 726 | cwnd = TCP_INIT_CWND; |
1da177e4 LT |
727 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); |
728 | } | |
729 | ||
e60402d0 IJ |
730 | /* |
731 | * Packet counting of FACK is based on in-order assumptions, therefore TCP | |
732 | * disables it when reordering is detected | |
733 | */ | |
4aabd8ef | 734 | void tcp_disable_fack(struct tcp_sock *tp) |
e60402d0 | 735 | { |
85cc391c IJ |
736 | /* RFC3517 uses different metric in lost marker => reset on change */ |
737 | if (tcp_is_fack(tp)) | |
738 | tp->lost_skb_hint = NULL; | |
ab56222a | 739 | tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED; |
e60402d0 IJ |
740 | } |
741 | ||
564262c1 | 742 | /* Take a notice that peer is sending D-SACKs */ |
e60402d0 IJ |
743 | static void tcp_dsack_seen(struct tcp_sock *tp) |
744 | { | |
ab56222a | 745 | tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; |
e60402d0 IJ |
746 | } |
747 | ||
6687e988 ACM |
748 | static void tcp_update_reordering(struct sock *sk, const int metric, |
749 | const int ts) | |
1da177e4 | 750 | { |
6687e988 | 751 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 752 | if (metric > tp->reordering) { |
40b215e5 PE |
753 | int mib_idx; |
754 | ||
1da177e4 LT |
755 | tp->reordering = min(TCP_MAX_REORDERING, metric); |
756 | ||
757 | /* This exciting event is worth to be remembered. 8) */ | |
758 | if (ts) | |
40b215e5 | 759 | mib_idx = LINUX_MIB_TCPTSREORDER; |
e60402d0 | 760 | else if (tcp_is_reno(tp)) |
40b215e5 | 761 | mib_idx = LINUX_MIB_TCPRENOREORDER; |
e60402d0 | 762 | else if (tcp_is_fack(tp)) |
40b215e5 | 763 | mib_idx = LINUX_MIB_TCPFACKREORDER; |
1da177e4 | 764 | else |
40b215e5 PE |
765 | mib_idx = LINUX_MIB_TCPSACKREORDER; |
766 | ||
de0744af | 767 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
1da177e4 | 768 | #if FASTRETRANS_DEBUG > 1 |
91df42be JP |
769 | pr_debug("Disorder%d %d %u f%u s%u rr%d\n", |
770 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, | |
771 | tp->reordering, | |
772 | tp->fackets_out, | |
773 | tp->sacked_out, | |
774 | tp->undo_marker ? tp->undo_retrans : 0); | |
1da177e4 | 775 | #endif |
e60402d0 | 776 | tcp_disable_fack(tp); |
1da177e4 | 777 | } |
eed530b6 YC |
778 | |
779 | if (metric > 0) | |
780 | tcp_disable_early_retrans(tp); | |
1da177e4 LT |
781 | } |
782 | ||
006f582c | 783 | /* This must be called before lost_out is incremented */ |
c8c213f2 IJ |
784 | static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) |
785 | { | |
006f582c | 786 | if ((tp->retransmit_skb_hint == NULL) || |
c8c213f2 IJ |
787 | before(TCP_SKB_CB(skb)->seq, |
788 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) | |
006f582c IJ |
789 | tp->retransmit_skb_hint = skb; |
790 | ||
791 | if (!tp->lost_out || | |
792 | after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) | |
793 | tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; | |
c8c213f2 IJ |
794 | } |
795 | ||
41ea36e3 IJ |
796 | static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) |
797 | { | |
798 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
799 | tcp_verify_retransmit_hint(tp, skb); | |
800 | ||
801 | tp->lost_out += tcp_skb_pcount(skb); | |
802 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
803 | } | |
804 | } | |
805 | ||
e1aa680f IJ |
806 | static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, |
807 | struct sk_buff *skb) | |
006f582c IJ |
808 | { |
809 | tcp_verify_retransmit_hint(tp, skb); | |
810 | ||
811 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
812 | tp->lost_out += tcp_skb_pcount(skb); | |
813 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
814 | } | |
815 | } | |
816 | ||
1da177e4 LT |
817 | /* This procedure tags the retransmission queue when SACKs arrive. |
818 | * | |
819 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). | |
820 | * Packets in queue with these bits set are counted in variables | |
821 | * sacked_out, retrans_out and lost_out, correspondingly. | |
822 | * | |
823 | * Valid combinations are: | |
824 | * Tag InFlight Description | |
825 | * 0 1 - orig segment is in flight. | |
826 | * S 0 - nothing flies, orig reached receiver. | |
827 | * L 0 - nothing flies, orig lost by net. | |
828 | * R 2 - both orig and retransmit are in flight. | |
829 | * L|R 1 - orig is lost, retransmit is in flight. | |
830 | * S|R 1 - orig reached receiver, retrans is still in flight. | |
831 | * (L|S|R is logically valid, it could occur when L|R is sacked, | |
832 | * but it is equivalent to plain S and code short-curcuits it to S. | |
833 | * L|S is logically invalid, it would mean -1 packet in flight 8)) | |
834 | * | |
835 | * These 6 states form finite state machine, controlled by the following events: | |
836 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) | |
837 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) | |
974c1236 | 838 | * 3. Loss detection event of two flavors: |
1da177e4 LT |
839 | * A. Scoreboard estimator decided the packet is lost. |
840 | * A'. Reno "three dupacks" marks head of queue lost. | |
974c1236 YC |
841 | * A''. Its FACK modification, head until snd.fack is lost. |
842 | * B. SACK arrives sacking SND.NXT at the moment, when the | |
1da177e4 LT |
843 | * segment was retransmitted. |
844 | * 4. D-SACK added new rule: D-SACK changes any tag to S. | |
845 | * | |
846 | * It is pleasant to note, that state diagram turns out to be commutative, | |
847 | * so that we are allowed not to be bothered by order of our actions, | |
848 | * when multiple events arrive simultaneously. (see the function below). | |
849 | * | |
850 | * Reordering detection. | |
851 | * -------------------- | |
852 | * Reordering metric is maximal distance, which a packet can be displaced | |
853 | * in packet stream. With SACKs we can estimate it: | |
854 | * | |
855 | * 1. SACK fills old hole and the corresponding segment was not | |
856 | * ever retransmitted -> reordering. Alas, we cannot use it | |
857 | * when segment was retransmitted. | |
858 | * 2. The last flaw is solved with D-SACK. D-SACK arrives | |
859 | * for retransmitted and already SACKed segment -> reordering.. | |
860 | * Both of these heuristics are not used in Loss state, when we cannot | |
861 | * account for retransmits accurately. | |
5b3c9882 IJ |
862 | * |
863 | * SACK block validation. | |
864 | * ---------------------- | |
865 | * | |
866 | * SACK block range validation checks that the received SACK block fits to | |
867 | * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. | |
868 | * Note that SND.UNA is not included to the range though being valid because | |
0e835331 IJ |
869 | * it means that the receiver is rather inconsistent with itself reporting |
870 | * SACK reneging when it should advance SND.UNA. Such SACK block this is | |
871 | * perfectly valid, however, in light of RFC2018 which explicitly states | |
872 | * that "SACK block MUST reflect the newest segment. Even if the newest | |
873 | * segment is going to be discarded ...", not that it looks very clever | |
874 | * in case of head skb. Due to potentional receiver driven attacks, we | |
875 | * choose to avoid immediate execution of a walk in write queue due to | |
876 | * reneging and defer head skb's loss recovery to standard loss recovery | |
877 | * procedure that will eventually trigger (nothing forbids us doing this). | |
5b3c9882 IJ |
878 | * |
879 | * Implements also blockage to start_seq wrap-around. Problem lies in the | |
880 | * fact that though start_seq (s) is before end_seq (i.e., not reversed), | |
881 | * there's no guarantee that it will be before snd_nxt (n). The problem | |
882 | * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt | |
883 | * wrap (s_w): | |
884 | * | |
885 | * <- outs wnd -> <- wrapzone -> | |
886 | * u e n u_w e_w s n_w | |
887 | * | | | | | | | | |
888 | * |<------------+------+----- TCP seqno space --------------+---------->| | |
889 | * ...-- <2^31 ->| |<--------... | |
890 | * ...---- >2^31 ------>| |<--------... | |
891 | * | |
892 | * Current code wouldn't be vulnerable but it's better still to discard such | |
893 | * crazy SACK blocks. Doing this check for start_seq alone closes somewhat | |
894 | * similar case (end_seq after snd_nxt wrap) as earlier reversed check in | |
895 | * snd_nxt wrap -> snd_una region will then become "well defined", i.e., | |
896 | * equal to the ideal case (infinite seqno space without wrap caused issues). | |
897 | * | |
898 | * With D-SACK the lower bound is extended to cover sequence space below | |
899 | * SND.UNA down to undo_marker, which is the last point of interest. Yet | |
564262c1 | 900 | * again, D-SACK block must not to go across snd_una (for the same reason as |
5b3c9882 IJ |
901 | * for the normal SACK blocks, explained above). But there all simplicity |
902 | * ends, TCP might receive valid D-SACKs below that. As long as they reside | |
903 | * fully below undo_marker they do not affect behavior in anyway and can | |
904 | * therefore be safely ignored. In rare cases (which are more or less | |
905 | * theoretical ones), the D-SACK will nicely cross that boundary due to skb | |
906 | * fragmentation and packet reordering past skb's retransmission. To consider | |
907 | * them correctly, the acceptable range must be extended even more though | |
908 | * the exact amount is rather hard to quantify. However, tp->max_window can | |
909 | * be used as an exaggerated estimate. | |
1da177e4 | 910 | */ |
a2a385d6 ED |
911 | static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, |
912 | u32 start_seq, u32 end_seq) | |
5b3c9882 IJ |
913 | { |
914 | /* Too far in future, or reversed (interpretation is ambiguous) */ | |
915 | if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) | |
a2a385d6 | 916 | return false; |
5b3c9882 IJ |
917 | |
918 | /* Nasty start_seq wrap-around check (see comments above) */ | |
919 | if (!before(start_seq, tp->snd_nxt)) | |
a2a385d6 | 920 | return false; |
5b3c9882 | 921 | |
564262c1 | 922 | /* In outstanding window? ...This is valid exit for D-SACKs too. |
5b3c9882 IJ |
923 | * start_seq == snd_una is non-sensical (see comments above) |
924 | */ | |
925 | if (after(start_seq, tp->snd_una)) | |
a2a385d6 | 926 | return true; |
5b3c9882 IJ |
927 | |
928 | if (!is_dsack || !tp->undo_marker) | |
a2a385d6 | 929 | return false; |
5b3c9882 IJ |
930 | |
931 | /* ...Then it's D-SACK, and must reside below snd_una completely */ | |
f779b2d6 | 932 | if (after(end_seq, tp->snd_una)) |
a2a385d6 | 933 | return false; |
5b3c9882 IJ |
934 | |
935 | if (!before(start_seq, tp->undo_marker)) | |
a2a385d6 | 936 | return true; |
5b3c9882 IJ |
937 | |
938 | /* Too old */ | |
939 | if (!after(end_seq, tp->undo_marker)) | |
a2a385d6 | 940 | return false; |
5b3c9882 IJ |
941 | |
942 | /* Undo_marker boundary crossing (overestimates a lot). Known already: | |
943 | * start_seq < undo_marker and end_seq >= undo_marker. | |
944 | */ | |
945 | return !before(start_seq, end_seq - tp->max_window); | |
946 | } | |
947 | ||
1c1e87ed | 948 | /* Check for lost retransmit. This superb idea is borrowed from "ratehalving". |
974c1236 | 949 | * Event "B". Later note: FACK people cheated me again 8), we have to account |
1c1e87ed | 950 | * for reordering! Ugly, but should help. |
f785a8e2 IJ |
951 | * |
952 | * Search retransmitted skbs from write_queue that were sent when snd_nxt was | |
953 | * less than what is now known to be received by the other end (derived from | |
9f58f3b7 IJ |
954 | * highest SACK block). Also calculate the lowest snd_nxt among the remaining |
955 | * retransmitted skbs to avoid some costly processing per ACKs. | |
1c1e87ed | 956 | */ |
407ef1de | 957 | static void tcp_mark_lost_retrans(struct sock *sk) |
1c1e87ed | 958 | { |
9f58f3b7 | 959 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1c1e87ed IJ |
960 | struct tcp_sock *tp = tcp_sk(sk); |
961 | struct sk_buff *skb; | |
f785a8e2 | 962 | int cnt = 0; |
df2e014b | 963 | u32 new_low_seq = tp->snd_nxt; |
6859d494 | 964 | u32 received_upto = tcp_highest_sack_seq(tp); |
9f58f3b7 IJ |
965 | |
966 | if (!tcp_is_fack(tp) || !tp->retrans_out || | |
967 | !after(received_upto, tp->lost_retrans_low) || | |
968 | icsk->icsk_ca_state != TCP_CA_Recovery) | |
407ef1de | 969 | return; |
1c1e87ed IJ |
970 | |
971 | tcp_for_write_queue(skb, sk) { | |
972 | u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; | |
973 | ||
974 | if (skb == tcp_send_head(sk)) | |
975 | break; | |
f785a8e2 | 976 | if (cnt == tp->retrans_out) |
1c1e87ed IJ |
977 | break; |
978 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
979 | continue; | |
980 | ||
f785a8e2 IJ |
981 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) |
982 | continue; | |
983 | ||
d0af4160 IJ |
984 | /* TODO: We would like to get rid of tcp_is_fack(tp) only |
985 | * constraint here (see above) but figuring out that at | |
986 | * least tp->reordering SACK blocks reside between ack_seq | |
987 | * and received_upto is not easy task to do cheaply with | |
988 | * the available datastructures. | |
989 | * | |
990 | * Whether FACK should check here for tp->reordering segs | |
991 | * in-between one could argue for either way (it would be | |
992 | * rather simple to implement as we could count fack_count | |
993 | * during the walk and do tp->fackets_out - fack_count). | |
994 | */ | |
995 | if (after(received_upto, ack_seq)) { | |
1c1e87ed IJ |
996 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
997 | tp->retrans_out -= tcp_skb_pcount(skb); | |
998 | ||
006f582c | 999 | tcp_skb_mark_lost_uncond_verify(tp, skb); |
de0744af | 1000 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT); |
f785a8e2 | 1001 | } else { |
df2e014b | 1002 | if (before(ack_seq, new_low_seq)) |
b08d6cb2 | 1003 | new_low_seq = ack_seq; |
f785a8e2 | 1004 | cnt += tcp_skb_pcount(skb); |
1c1e87ed IJ |
1005 | } |
1006 | } | |
b08d6cb2 IJ |
1007 | |
1008 | if (tp->retrans_out) | |
1009 | tp->lost_retrans_low = new_low_seq; | |
1c1e87ed | 1010 | } |
5b3c9882 | 1011 | |
a2a385d6 ED |
1012 | static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, |
1013 | struct tcp_sack_block_wire *sp, int num_sacks, | |
1014 | u32 prior_snd_una) | |
d06e021d | 1015 | { |
1ed83465 | 1016 | struct tcp_sock *tp = tcp_sk(sk); |
d3e2ce3b HH |
1017 | u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); |
1018 | u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); | |
a2a385d6 | 1019 | bool dup_sack = false; |
d06e021d DM |
1020 | |
1021 | if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { | |
a2a385d6 | 1022 | dup_sack = true; |
e60402d0 | 1023 | tcp_dsack_seen(tp); |
de0744af | 1024 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV); |
d06e021d | 1025 | } else if (num_sacks > 1) { |
d3e2ce3b HH |
1026 | u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); |
1027 | u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); | |
d06e021d DM |
1028 | |
1029 | if (!after(end_seq_0, end_seq_1) && | |
1030 | !before(start_seq_0, start_seq_1)) { | |
a2a385d6 | 1031 | dup_sack = true; |
e60402d0 | 1032 | tcp_dsack_seen(tp); |
de0744af PE |
1033 | NET_INC_STATS_BH(sock_net(sk), |
1034 | LINUX_MIB_TCPDSACKOFORECV); | |
d06e021d DM |
1035 | } |
1036 | } | |
1037 | ||
1038 | /* D-SACK for already forgotten data... Do dumb counting. */ | |
c24f691b | 1039 | if (dup_sack && tp->undo_marker && tp->undo_retrans && |
d06e021d DM |
1040 | !after(end_seq_0, prior_snd_una) && |
1041 | after(end_seq_0, tp->undo_marker)) | |
1042 | tp->undo_retrans--; | |
1043 | ||
1044 | return dup_sack; | |
1045 | } | |
1046 | ||
a1197f5a IJ |
1047 | struct tcp_sacktag_state { |
1048 | int reord; | |
1049 | int fack_count; | |
1050 | int flag; | |
59c9af42 | 1051 | s32 rtt; /* RTT measured by SACKing never-retransmitted data */ |
a1197f5a IJ |
1052 | }; |
1053 | ||
d1935942 IJ |
1054 | /* Check if skb is fully within the SACK block. In presence of GSO skbs, |
1055 | * the incoming SACK may not exactly match but we can find smaller MSS | |
1056 | * aligned portion of it that matches. Therefore we might need to fragment | |
1057 | * which may fail and creates some hassle (caller must handle error case | |
1058 | * returns). | |
832d11c5 IJ |
1059 | * |
1060 | * FIXME: this could be merged to shift decision code | |
d1935942 | 1061 | */ |
0f79efdc | 1062 | static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, |
a2a385d6 | 1063 | u32 start_seq, u32 end_seq) |
d1935942 | 1064 | { |
a2a385d6 ED |
1065 | int err; |
1066 | bool in_sack; | |
d1935942 | 1067 | unsigned int pkt_len; |
adb92db8 | 1068 | unsigned int mss; |
d1935942 IJ |
1069 | |
1070 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1071 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1072 | ||
1073 | if (tcp_skb_pcount(skb) > 1 && !in_sack && | |
1074 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { | |
adb92db8 | 1075 | mss = tcp_skb_mss(skb); |
d1935942 IJ |
1076 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); |
1077 | ||
adb92db8 | 1078 | if (!in_sack) { |
d1935942 | 1079 | pkt_len = start_seq - TCP_SKB_CB(skb)->seq; |
adb92db8 IJ |
1080 | if (pkt_len < mss) |
1081 | pkt_len = mss; | |
1082 | } else { | |
d1935942 | 1083 | pkt_len = end_seq - TCP_SKB_CB(skb)->seq; |
adb92db8 IJ |
1084 | if (pkt_len < mss) |
1085 | return -EINVAL; | |
1086 | } | |
1087 | ||
1088 | /* Round if necessary so that SACKs cover only full MSSes | |
1089 | * and/or the remaining small portion (if present) | |
1090 | */ | |
1091 | if (pkt_len > mss) { | |
1092 | unsigned int new_len = (pkt_len / mss) * mss; | |
1093 | if (!in_sack && new_len < pkt_len) { | |
1094 | new_len += mss; | |
1095 | if (new_len > skb->len) | |
1096 | return 0; | |
1097 | } | |
1098 | pkt_len = new_len; | |
1099 | } | |
1100 | err = tcp_fragment(sk, skb, pkt_len, mss); | |
d1935942 IJ |
1101 | if (err < 0) |
1102 | return err; | |
1103 | } | |
1104 | ||
1105 | return in_sack; | |
1106 | } | |
1107 | ||
cc9a672e NC |
1108 | /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ |
1109 | static u8 tcp_sacktag_one(struct sock *sk, | |
1110 | struct tcp_sacktag_state *state, u8 sacked, | |
1111 | u32 start_seq, u32 end_seq, | |
59c9af42 | 1112 | int dup_sack, int pcount, u32 xmit_time) |
9e10c47c | 1113 | { |
6859d494 | 1114 | struct tcp_sock *tp = tcp_sk(sk); |
a1197f5a | 1115 | int fack_count = state->fack_count; |
9e10c47c IJ |
1116 | |
1117 | /* Account D-SACK for retransmitted packet. */ | |
1118 | if (dup_sack && (sacked & TCPCB_RETRANS)) { | |
c24f691b | 1119 | if (tp->undo_marker && tp->undo_retrans && |
cc9a672e | 1120 | after(end_seq, tp->undo_marker)) |
9e10c47c | 1121 | tp->undo_retrans--; |
ede9f3b1 | 1122 | if (sacked & TCPCB_SACKED_ACKED) |
a1197f5a | 1123 | state->reord = min(fack_count, state->reord); |
9e10c47c IJ |
1124 | } |
1125 | ||
1126 | /* Nothing to do; acked frame is about to be dropped (was ACKed). */ | |
cc9a672e | 1127 | if (!after(end_seq, tp->snd_una)) |
a1197f5a | 1128 | return sacked; |
9e10c47c IJ |
1129 | |
1130 | if (!(sacked & TCPCB_SACKED_ACKED)) { | |
1131 | if (sacked & TCPCB_SACKED_RETRANS) { | |
1132 | /* If the segment is not tagged as lost, | |
1133 | * we do not clear RETRANS, believing | |
1134 | * that retransmission is still in flight. | |
1135 | */ | |
1136 | if (sacked & TCPCB_LOST) { | |
a1197f5a | 1137 | sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); |
f58b22fd IJ |
1138 | tp->lost_out -= pcount; |
1139 | tp->retrans_out -= pcount; | |
9e10c47c IJ |
1140 | } |
1141 | } else { | |
1142 | if (!(sacked & TCPCB_RETRANS)) { | |
1143 | /* New sack for not retransmitted frame, | |
1144 | * which was in hole. It is reordering. | |
1145 | */ | |
cc9a672e | 1146 | if (before(start_seq, |
9e10c47c | 1147 | tcp_highest_sack_seq(tp))) |
a1197f5a IJ |
1148 | state->reord = min(fack_count, |
1149 | state->reord); | |
e33099f9 YC |
1150 | if (!after(end_seq, tp->high_seq)) |
1151 | state->flag |= FLAG_ORIG_SACK_ACKED; | |
59c9af42 YC |
1152 | /* Pick the earliest sequence sacked for RTT */ |
1153 | if (state->rtt < 0) | |
1154 | state->rtt = tcp_time_stamp - xmit_time; | |
9e10c47c IJ |
1155 | } |
1156 | ||
1157 | if (sacked & TCPCB_LOST) { | |
a1197f5a | 1158 | sacked &= ~TCPCB_LOST; |
f58b22fd | 1159 | tp->lost_out -= pcount; |
9e10c47c IJ |
1160 | } |
1161 | } | |
1162 | ||
a1197f5a IJ |
1163 | sacked |= TCPCB_SACKED_ACKED; |
1164 | state->flag |= FLAG_DATA_SACKED; | |
f58b22fd | 1165 | tp->sacked_out += pcount; |
9e10c47c | 1166 | |
f58b22fd | 1167 | fack_count += pcount; |
9e10c47c IJ |
1168 | |
1169 | /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ | |
1170 | if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && | |
cc9a672e | 1171 | before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) |
f58b22fd | 1172 | tp->lost_cnt_hint += pcount; |
9e10c47c IJ |
1173 | |
1174 | if (fack_count > tp->fackets_out) | |
1175 | tp->fackets_out = fack_count; | |
9e10c47c IJ |
1176 | } |
1177 | ||
1178 | /* D-SACK. We can detect redundant retransmission in S|R and plain R | |
1179 | * frames and clear it. undo_retrans is decreased above, L|R frames | |
1180 | * are accounted above as well. | |
1181 | */ | |
a1197f5a IJ |
1182 | if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { |
1183 | sacked &= ~TCPCB_SACKED_RETRANS; | |
f58b22fd | 1184 | tp->retrans_out -= pcount; |
9e10c47c IJ |
1185 | } |
1186 | ||
a1197f5a | 1187 | return sacked; |
9e10c47c IJ |
1188 | } |
1189 | ||
daef52ba NC |
1190 | /* Shift newly-SACKed bytes from this skb to the immediately previous |
1191 | * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. | |
1192 | */ | |
a2a385d6 ED |
1193 | static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, |
1194 | struct tcp_sacktag_state *state, | |
1195 | unsigned int pcount, int shifted, int mss, | |
1196 | bool dup_sack) | |
832d11c5 IJ |
1197 | { |
1198 | struct tcp_sock *tp = tcp_sk(sk); | |
50133161 | 1199 | struct sk_buff *prev = tcp_write_queue_prev(sk, skb); |
daef52ba NC |
1200 | u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ |
1201 | u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ | |
832d11c5 IJ |
1202 | |
1203 | BUG_ON(!pcount); | |
1204 | ||
4c90d3b3 NC |
1205 | /* Adjust counters and hints for the newly sacked sequence |
1206 | * range but discard the return value since prev is already | |
1207 | * marked. We must tag the range first because the seq | |
1208 | * advancement below implicitly advances | |
1209 | * tcp_highest_sack_seq() when skb is highest_sack. | |
1210 | */ | |
1211 | tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, | |
59c9af42 YC |
1212 | start_seq, end_seq, dup_sack, pcount, |
1213 | TCP_SKB_CB(skb)->when); | |
4c90d3b3 NC |
1214 | |
1215 | if (skb == tp->lost_skb_hint) | |
0af2a0d0 NC |
1216 | tp->lost_cnt_hint += pcount; |
1217 | ||
832d11c5 IJ |
1218 | TCP_SKB_CB(prev)->end_seq += shifted; |
1219 | TCP_SKB_CB(skb)->seq += shifted; | |
1220 | ||
1221 | skb_shinfo(prev)->gso_segs += pcount; | |
1222 | BUG_ON(skb_shinfo(skb)->gso_segs < pcount); | |
1223 | skb_shinfo(skb)->gso_segs -= pcount; | |
1224 | ||
1225 | /* When we're adding to gso_segs == 1, gso_size will be zero, | |
1226 | * in theory this shouldn't be necessary but as long as DSACK | |
1227 | * code can come after this skb later on it's better to keep | |
1228 | * setting gso_size to something. | |
1229 | */ | |
1230 | if (!skb_shinfo(prev)->gso_size) { | |
1231 | skb_shinfo(prev)->gso_size = mss; | |
c9af6db4 | 1232 | skb_shinfo(prev)->gso_type = sk->sk_gso_type; |
832d11c5 IJ |
1233 | } |
1234 | ||
1235 | /* CHECKME: To clear or not to clear? Mimics normal skb currently */ | |
1236 | if (skb_shinfo(skb)->gso_segs <= 1) { | |
1237 | skb_shinfo(skb)->gso_size = 0; | |
c9af6db4 | 1238 | skb_shinfo(skb)->gso_type = 0; |
832d11c5 IJ |
1239 | } |
1240 | ||
832d11c5 IJ |
1241 | /* Difference in this won't matter, both ACKed by the same cumul. ACK */ |
1242 | TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); | |
1243 | ||
832d11c5 IJ |
1244 | if (skb->len > 0) { |
1245 | BUG_ON(!tcp_skb_pcount(skb)); | |
111cc8b9 | 1246 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED); |
a2a385d6 | 1247 | return false; |
832d11c5 IJ |
1248 | } |
1249 | ||
1250 | /* Whole SKB was eaten :-) */ | |
1251 | ||
92ee76b6 IJ |
1252 | if (skb == tp->retransmit_skb_hint) |
1253 | tp->retransmit_skb_hint = prev; | |
92ee76b6 IJ |
1254 | if (skb == tp->lost_skb_hint) { |
1255 | tp->lost_skb_hint = prev; | |
1256 | tp->lost_cnt_hint -= tcp_skb_pcount(prev); | |
1257 | } | |
1258 | ||
4de075e0 | 1259 | TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags; |
832d11c5 IJ |
1260 | if (skb == tcp_highest_sack(sk)) |
1261 | tcp_advance_highest_sack(sk, skb); | |
1262 | ||
1263 | tcp_unlink_write_queue(skb, sk); | |
1264 | sk_wmem_free_skb(sk, skb); | |
1265 | ||
111cc8b9 IJ |
1266 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED); |
1267 | ||
a2a385d6 | 1268 | return true; |
832d11c5 IJ |
1269 | } |
1270 | ||
1271 | /* I wish gso_size would have a bit more sane initialization than | |
1272 | * something-or-zero which complicates things | |
1273 | */ | |
cf533ea5 | 1274 | static int tcp_skb_seglen(const struct sk_buff *skb) |
832d11c5 | 1275 | { |
775ffabf | 1276 | return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); |
832d11c5 IJ |
1277 | } |
1278 | ||
1279 | /* Shifting pages past head area doesn't work */ | |
cf533ea5 | 1280 | static int skb_can_shift(const struct sk_buff *skb) |
832d11c5 IJ |
1281 | { |
1282 | return !skb_headlen(skb) && skb_is_nonlinear(skb); | |
1283 | } | |
1284 | ||
1285 | /* Try collapsing SACK blocks spanning across multiple skbs to a single | |
1286 | * skb. | |
1287 | */ | |
1288 | static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, | |
a1197f5a | 1289 | struct tcp_sacktag_state *state, |
832d11c5 | 1290 | u32 start_seq, u32 end_seq, |
a2a385d6 | 1291 | bool dup_sack) |
832d11c5 IJ |
1292 | { |
1293 | struct tcp_sock *tp = tcp_sk(sk); | |
1294 | struct sk_buff *prev; | |
1295 | int mss; | |
1296 | int pcount = 0; | |
1297 | int len; | |
1298 | int in_sack; | |
1299 | ||
1300 | if (!sk_can_gso(sk)) | |
1301 | goto fallback; | |
1302 | ||
1303 | /* Normally R but no L won't result in plain S */ | |
1304 | if (!dup_sack && | |
9969ca5f | 1305 | (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) |
832d11c5 IJ |
1306 | goto fallback; |
1307 | if (!skb_can_shift(skb)) | |
1308 | goto fallback; | |
1309 | /* This frame is about to be dropped (was ACKed). */ | |
1310 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1311 | goto fallback; | |
1312 | ||
1313 | /* Can only happen with delayed DSACK + discard craziness */ | |
1314 | if (unlikely(skb == tcp_write_queue_head(sk))) | |
1315 | goto fallback; | |
1316 | prev = tcp_write_queue_prev(sk, skb); | |
1317 | ||
1318 | if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) | |
1319 | goto fallback; | |
1320 | ||
1321 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1322 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1323 | ||
1324 | if (in_sack) { | |
1325 | len = skb->len; | |
1326 | pcount = tcp_skb_pcount(skb); | |
775ffabf | 1327 | mss = tcp_skb_seglen(skb); |
832d11c5 IJ |
1328 | |
1329 | /* TODO: Fix DSACKs to not fragment already SACKed and we can | |
1330 | * drop this restriction as unnecessary | |
1331 | */ | |
775ffabf | 1332 | if (mss != tcp_skb_seglen(prev)) |
832d11c5 IJ |
1333 | goto fallback; |
1334 | } else { | |
1335 | if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) | |
1336 | goto noop; | |
1337 | /* CHECKME: This is non-MSS split case only?, this will | |
1338 | * cause skipped skbs due to advancing loop btw, original | |
1339 | * has that feature too | |
1340 | */ | |
1341 | if (tcp_skb_pcount(skb) <= 1) | |
1342 | goto noop; | |
1343 | ||
1344 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); | |
1345 | if (!in_sack) { | |
1346 | /* TODO: head merge to next could be attempted here | |
1347 | * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), | |
1348 | * though it might not be worth of the additional hassle | |
1349 | * | |
1350 | * ...we can probably just fallback to what was done | |
1351 | * previously. We could try merging non-SACKed ones | |
1352 | * as well but it probably isn't going to buy off | |
1353 | * because later SACKs might again split them, and | |
1354 | * it would make skb timestamp tracking considerably | |
1355 | * harder problem. | |
1356 | */ | |
1357 | goto fallback; | |
1358 | } | |
1359 | ||
1360 | len = end_seq - TCP_SKB_CB(skb)->seq; | |
1361 | BUG_ON(len < 0); | |
1362 | BUG_ON(len > skb->len); | |
1363 | ||
1364 | /* MSS boundaries should be honoured or else pcount will | |
1365 | * severely break even though it makes things bit trickier. | |
1366 | * Optimize common case to avoid most of the divides | |
1367 | */ | |
1368 | mss = tcp_skb_mss(skb); | |
1369 | ||
1370 | /* TODO: Fix DSACKs to not fragment already SACKed and we can | |
1371 | * drop this restriction as unnecessary | |
1372 | */ | |
775ffabf | 1373 | if (mss != tcp_skb_seglen(prev)) |
832d11c5 IJ |
1374 | goto fallback; |
1375 | ||
1376 | if (len == mss) { | |
1377 | pcount = 1; | |
1378 | } else if (len < mss) { | |
1379 | goto noop; | |
1380 | } else { | |
1381 | pcount = len / mss; | |
1382 | len = pcount * mss; | |
1383 | } | |
1384 | } | |
1385 | ||
4648dc97 NC |
1386 | /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ |
1387 | if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) | |
1388 | goto fallback; | |
1389 | ||
832d11c5 IJ |
1390 | if (!skb_shift(prev, skb, len)) |
1391 | goto fallback; | |
9ec06ff5 | 1392 | if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack)) |
832d11c5 IJ |
1393 | goto out; |
1394 | ||
1395 | /* Hole filled allows collapsing with the next as well, this is very | |
1396 | * useful when hole on every nth skb pattern happens | |
1397 | */ | |
1398 | if (prev == tcp_write_queue_tail(sk)) | |
1399 | goto out; | |
1400 | skb = tcp_write_queue_next(sk, prev); | |
1401 | ||
f0bc52f3 IJ |
1402 | if (!skb_can_shift(skb) || |
1403 | (skb == tcp_send_head(sk)) || | |
1404 | ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || | |
775ffabf | 1405 | (mss != tcp_skb_seglen(skb))) |
832d11c5 IJ |
1406 | goto out; |
1407 | ||
1408 | len = skb->len; | |
1409 | if (skb_shift(prev, skb, len)) { | |
1410 | pcount += tcp_skb_pcount(skb); | |
9ec06ff5 | 1411 | tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0); |
832d11c5 IJ |
1412 | } |
1413 | ||
1414 | out: | |
a1197f5a | 1415 | state->fack_count += pcount; |
832d11c5 IJ |
1416 | return prev; |
1417 | ||
1418 | noop: | |
1419 | return skb; | |
1420 | ||
1421 | fallback: | |
111cc8b9 | 1422 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); |
832d11c5 IJ |
1423 | return NULL; |
1424 | } | |
1425 | ||
68f8353b IJ |
1426 | static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, |
1427 | struct tcp_sack_block *next_dup, | |
a1197f5a | 1428 | struct tcp_sacktag_state *state, |
68f8353b | 1429 | u32 start_seq, u32 end_seq, |
a2a385d6 | 1430 | bool dup_sack_in) |
68f8353b | 1431 | { |
832d11c5 IJ |
1432 | struct tcp_sock *tp = tcp_sk(sk); |
1433 | struct sk_buff *tmp; | |
1434 | ||
68f8353b IJ |
1435 | tcp_for_write_queue_from(skb, sk) { |
1436 | int in_sack = 0; | |
a2a385d6 | 1437 | bool dup_sack = dup_sack_in; |
68f8353b IJ |
1438 | |
1439 | if (skb == tcp_send_head(sk)) | |
1440 | break; | |
1441 | ||
1442 | /* queue is in-order => we can short-circuit the walk early */ | |
1443 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) | |
1444 | break; | |
1445 | ||
1446 | if ((next_dup != NULL) && | |
1447 | before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { | |
1448 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1449 | next_dup->start_seq, | |
1450 | next_dup->end_seq); | |
1451 | if (in_sack > 0) | |
a2a385d6 | 1452 | dup_sack = true; |
68f8353b IJ |
1453 | } |
1454 | ||
832d11c5 IJ |
1455 | /* skb reference here is a bit tricky to get right, since |
1456 | * shifting can eat and free both this skb and the next, | |
1457 | * so not even _safe variant of the loop is enough. | |
1458 | */ | |
1459 | if (in_sack <= 0) { | |
a1197f5a IJ |
1460 | tmp = tcp_shift_skb_data(sk, skb, state, |
1461 | start_seq, end_seq, dup_sack); | |
832d11c5 IJ |
1462 | if (tmp != NULL) { |
1463 | if (tmp != skb) { | |
1464 | skb = tmp; | |
1465 | continue; | |
1466 | } | |
1467 | ||
1468 | in_sack = 0; | |
1469 | } else { | |
1470 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1471 | start_seq, | |
1472 | end_seq); | |
1473 | } | |
1474 | } | |
1475 | ||
68f8353b IJ |
1476 | if (unlikely(in_sack < 0)) |
1477 | break; | |
1478 | ||
832d11c5 | 1479 | if (in_sack) { |
cc9a672e NC |
1480 | TCP_SKB_CB(skb)->sacked = |
1481 | tcp_sacktag_one(sk, | |
1482 | state, | |
1483 | TCP_SKB_CB(skb)->sacked, | |
1484 | TCP_SKB_CB(skb)->seq, | |
1485 | TCP_SKB_CB(skb)->end_seq, | |
1486 | dup_sack, | |
59c9af42 YC |
1487 | tcp_skb_pcount(skb), |
1488 | TCP_SKB_CB(skb)->when); | |
68f8353b | 1489 | |
832d11c5 IJ |
1490 | if (!before(TCP_SKB_CB(skb)->seq, |
1491 | tcp_highest_sack_seq(tp))) | |
1492 | tcp_advance_highest_sack(sk, skb); | |
1493 | } | |
1494 | ||
a1197f5a | 1495 | state->fack_count += tcp_skb_pcount(skb); |
68f8353b IJ |
1496 | } |
1497 | return skb; | |
1498 | } | |
1499 | ||
1500 | /* Avoid all extra work that is being done by sacktag while walking in | |
1501 | * a normal way | |
1502 | */ | |
1503 | static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, | |
a1197f5a IJ |
1504 | struct tcp_sacktag_state *state, |
1505 | u32 skip_to_seq) | |
68f8353b IJ |
1506 | { |
1507 | tcp_for_write_queue_from(skb, sk) { | |
1508 | if (skb == tcp_send_head(sk)) | |
1509 | break; | |
1510 | ||
e8bae275 | 1511 | if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) |
68f8353b | 1512 | break; |
d152a7d8 | 1513 | |
a1197f5a | 1514 | state->fack_count += tcp_skb_pcount(skb); |
68f8353b IJ |
1515 | } |
1516 | return skb; | |
1517 | } | |
1518 | ||
1519 | static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, | |
1520 | struct sock *sk, | |
1521 | struct tcp_sack_block *next_dup, | |
a1197f5a IJ |
1522 | struct tcp_sacktag_state *state, |
1523 | u32 skip_to_seq) | |
68f8353b IJ |
1524 | { |
1525 | if (next_dup == NULL) | |
1526 | return skb; | |
1527 | ||
1528 | if (before(next_dup->start_seq, skip_to_seq)) { | |
a1197f5a IJ |
1529 | skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); |
1530 | skb = tcp_sacktag_walk(skb, sk, NULL, state, | |
1531 | next_dup->start_seq, next_dup->end_seq, | |
1532 | 1); | |
68f8353b IJ |
1533 | } |
1534 | ||
1535 | return skb; | |
1536 | } | |
1537 | ||
cf533ea5 | 1538 | static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) |
68f8353b IJ |
1539 | { |
1540 | return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); | |
1541 | } | |
1542 | ||
1da177e4 | 1543 | static int |
cf533ea5 | 1544 | tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, |
59c9af42 | 1545 | u32 prior_snd_una, s32 *sack_rtt) |
1da177e4 LT |
1546 | { |
1547 | struct tcp_sock *tp = tcp_sk(sk); | |
cf533ea5 ED |
1548 | const unsigned char *ptr = (skb_transport_header(ack_skb) + |
1549 | TCP_SKB_CB(ack_skb)->sacked); | |
fd6dad61 | 1550 | struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); |
4389dded | 1551 | struct tcp_sack_block sp[TCP_NUM_SACKS]; |
68f8353b | 1552 | struct tcp_sack_block *cache; |
a1197f5a | 1553 | struct tcp_sacktag_state state; |
68f8353b | 1554 | struct sk_buff *skb; |
4389dded | 1555 | int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); |
fd6dad61 | 1556 | int used_sacks; |
a2a385d6 | 1557 | bool found_dup_sack = false; |
68f8353b | 1558 | int i, j; |
fda03fbb | 1559 | int first_sack_index; |
1da177e4 | 1560 | |
a1197f5a IJ |
1561 | state.flag = 0; |
1562 | state.reord = tp->packets_out; | |
59c9af42 | 1563 | state.rtt = -1; |
a1197f5a | 1564 | |
d738cd8f | 1565 | if (!tp->sacked_out) { |
de83c058 IJ |
1566 | if (WARN_ON(tp->fackets_out)) |
1567 | tp->fackets_out = 0; | |
6859d494 | 1568 | tcp_highest_sack_reset(sk); |
d738cd8f | 1569 | } |
1da177e4 | 1570 | |
1ed83465 | 1571 | found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, |
d06e021d DM |
1572 | num_sacks, prior_snd_una); |
1573 | if (found_dup_sack) | |
a1197f5a | 1574 | state.flag |= FLAG_DSACKING_ACK; |
6f74651a BE |
1575 | |
1576 | /* Eliminate too old ACKs, but take into | |
1577 | * account more or less fresh ones, they can | |
1578 | * contain valid SACK info. | |
1579 | */ | |
1580 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) | |
1581 | return 0; | |
1582 | ||