]>
Commit | Line | Data |
---|---|---|
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 | * | |
8 | * Authors: Ross Biro | |
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: Pedro Roque : Retransmit queue handled by TCP. | |
23 | * : Fragmentation on mtu decrease | |
24 | * : Segment collapse on retransmit | |
25 | * : AF independence | |
26 | * | |
27 | * Linus Torvalds : send_delayed_ack | |
28 | * David S. Miller : Charge memory using the right skb | |
29 | * during syn/ack processing. | |
30 | * David S. Miller : Output engine completely rewritten. | |
31 | * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. | |
32 | * Cacophonix Gaul : draft-minshall-nagle-01 | |
33 | * J Hadi Salim : ECN support | |
34 | * | |
35 | */ | |
36 | ||
37 | #include <net/tcp.h> | |
38 | ||
39 | #include <linux/compiler.h> | |
40 | #include <linux/module.h> | |
41 | ||
42 | /* People can turn this off for buggy TCP's found in printers etc. */ | |
43 | int sysctl_tcp_retrans_collapse __read_mostly = 1; | |
44 | ||
45 | /* People can turn this on to work with those rare, broken TCPs that | |
46 | * interpret the window field as a signed quantity. | |
47 | */ | |
48 | int sysctl_tcp_workaround_signed_windows __read_mostly = 0; | |
49 | ||
50 | /* This limits the percentage of the congestion window which we | |
51 | * will allow a single TSO frame to consume. Building TSO frames | |
52 | * which are too large can cause TCP streams to be bursty. | |
53 | */ | |
54 | int sysctl_tcp_tso_win_divisor __read_mostly = 3; | |
55 | ||
56 | int sysctl_tcp_mtu_probing __read_mostly = 0; | |
57 | int sysctl_tcp_base_mss __read_mostly = 512; | |
58 | ||
59 | /* By default, RFC2861 behavior. */ | |
60 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; | |
61 | ||
62 | /* Account for new data that has been sent to the network. */ | |
63 | static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) | |
64 | { | |
65 | struct tcp_sock *tp = tcp_sk(sk); | |
66 | unsigned int prior_packets = tp->packets_out; | |
67 | ||
68 | tcp_advance_send_head(sk, skb); | |
69 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; | |
70 | ||
71 | /* Don't override Nagle indefinately with F-RTO */ | |
72 | if (tp->frto_counter == 2) | |
73 | tp->frto_counter = 3; | |
74 | ||
75 | tp->packets_out += tcp_skb_pcount(skb); | |
76 | if (!prior_packets) | |
77 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
78 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
79 | } | |
80 | ||
81 | /* SND.NXT, if window was not shrunk. | |
82 | * If window has been shrunk, what should we make? It is not clear at all. | |
83 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
84 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
85 | * invalid. OK, let's make this for now: | |
86 | */ | |
87 | static inline __u32 tcp_acceptable_seq(struct sock *sk) | |
88 | { | |
89 | struct tcp_sock *tp = tcp_sk(sk); | |
90 | ||
91 | if (!before(tcp_wnd_end(tp), tp->snd_nxt)) | |
92 | return tp->snd_nxt; | |
93 | else | |
94 | return tcp_wnd_end(tp); | |
95 | } | |
96 | ||
97 | /* Calculate mss to advertise in SYN segment. | |
98 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
99 | * | |
100 | * 1. It is independent of path mtu. | |
101 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
102 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
103 | * attached devices, because some buggy hosts are confused by | |
104 | * large MSS. | |
105 | * 4. We do not make 3, we advertise MSS, calculated from first | |
106 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
107 | * This may be overridden via information stored in routing table. | |
108 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
109 | * probably even Jumbo". | |
110 | */ | |
111 | static __u16 tcp_advertise_mss(struct sock *sk) | |
112 | { | |
113 | struct tcp_sock *tp = tcp_sk(sk); | |
114 | struct dst_entry *dst = __sk_dst_get(sk); | |
115 | int mss = tp->advmss; | |
116 | ||
117 | if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { | |
118 | mss = dst_metric(dst, RTAX_ADVMSS); | |
119 | tp->advmss = mss; | |
120 | } | |
121 | ||
122 | return (__u16)mss; | |
123 | } | |
124 | ||
125 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
126 | * This is the first part of cwnd validation mechanism. */ | |
127 | static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) | |
128 | { | |
129 | struct tcp_sock *tp = tcp_sk(sk); | |
130 | s32 delta = tcp_time_stamp - tp->lsndtime; | |
131 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); | |
132 | u32 cwnd = tp->snd_cwnd; | |
133 | ||
134 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); | |
135 | ||
136 | tp->snd_ssthresh = tcp_current_ssthresh(sk); | |
137 | restart_cwnd = min(restart_cwnd, cwnd); | |
138 | ||
139 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) | |
140 | cwnd >>= 1; | |
141 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
142 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
143 | tp->snd_cwnd_used = 0; | |
144 | } | |
145 | ||
146 | /* Congestion state accounting after a packet has been sent. */ | |
147 | static void tcp_event_data_sent(struct tcp_sock *tp, | |
148 | struct sk_buff *skb, struct sock *sk) | |
149 | { | |
150 | struct inet_connection_sock *icsk = inet_csk(sk); | |
151 | const u32 now = tcp_time_stamp; | |
152 | ||
153 | if (sysctl_tcp_slow_start_after_idle && | |
154 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) | |
155 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); | |
156 | ||
157 | tp->lsndtime = now; | |
158 | ||
159 | /* If it is a reply for ato after last received | |
160 | * packet, enter pingpong mode. | |
161 | */ | |
162 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) | |
163 | icsk->icsk_ack.pingpong = 1; | |
164 | } | |
165 | ||
166 | /* Account for an ACK we sent. */ | |
167 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) | |
168 | { | |
169 | tcp_dec_quickack_mode(sk, pkts); | |
170 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
171 | } | |
172 | ||
173 | /* Determine a window scaling and initial window to offer. | |
174 | * Based on the assumption that the given amount of space | |
175 | * will be offered. Store the results in the tp structure. | |
176 | * NOTE: for smooth operation initial space offering should | |
177 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
178 | * This MUST be enforced by all callers. | |
179 | */ | |
180 | void tcp_select_initial_window(int __space, __u32 mss, | |
181 | __u32 *rcv_wnd, __u32 *window_clamp, | |
182 | int wscale_ok, __u8 *rcv_wscale) | |
183 | { | |
184 | unsigned int space = (__space < 0 ? 0 : __space); | |
185 | ||
186 | /* If no clamp set the clamp to the max possible scaled window */ | |
187 | if (*window_clamp == 0) | |
188 | (*window_clamp) = (65535 << 14); | |
189 | space = min(*window_clamp, space); | |
190 | ||
191 | /* Quantize space offering to a multiple of mss if possible. */ | |
192 | if (space > mss) | |
193 | space = (space / mss) * mss; | |
194 | ||
195 | /* NOTE: offering an initial window larger than 32767 | |
196 | * will break some buggy TCP stacks. If the admin tells us | |
197 | * it is likely we could be speaking with such a buggy stack | |
198 | * we will truncate our initial window offering to 32K-1 | |
199 | * unless the remote has sent us a window scaling option, | |
200 | * which we interpret as a sign the remote TCP is not | |
201 | * misinterpreting the window field as a signed quantity. | |
202 | */ | |
203 | if (sysctl_tcp_workaround_signed_windows) | |
204 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
205 | else | |
206 | (*rcv_wnd) = space; | |
207 | ||
208 | (*rcv_wscale) = 0; | |
209 | if (wscale_ok) { | |
210 | /* Set window scaling on max possible window | |
211 | * See RFC1323 for an explanation of the limit to 14 | |
212 | */ | |
213 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); | |
214 | space = min_t(u32, space, *window_clamp); | |
215 | while (space > 65535 && (*rcv_wscale) < 14) { | |
216 | space >>= 1; | |
217 | (*rcv_wscale)++; | |
218 | } | |
219 | } | |
220 | ||
221 | /* Set initial window to value enough for senders, | |
222 | * following RFC2414. Senders, not following this RFC, | |
223 | * will be satisfied with 2. | |
224 | */ | |
225 | if (mss > (1 << *rcv_wscale)) { | |
226 | int init_cwnd = 4; | |
227 | if (mss > 1460 * 3) | |
228 | init_cwnd = 2; | |
229 | else if (mss > 1460) | |
230 | init_cwnd = 3; | |
231 | if (*rcv_wnd > init_cwnd * mss) | |
232 | *rcv_wnd = init_cwnd * mss; | |
233 | } | |
234 | ||
235 | /* Set the clamp no higher than max representable value */ | |
236 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); | |
237 | } | |
238 | ||
239 | /* Chose a new window to advertise, update state in tcp_sock for the | |
240 | * socket, and return result with RFC1323 scaling applied. The return | |
241 | * value can be stuffed directly into th->window for an outgoing | |
242 | * frame. | |
243 | */ | |
244 | static u16 tcp_select_window(struct sock *sk) | |
245 | { | |
246 | struct tcp_sock *tp = tcp_sk(sk); | |
247 | u32 cur_win = tcp_receive_window(tp); | |
248 | u32 new_win = __tcp_select_window(sk); | |
249 | ||
250 | /* Never shrink the offered window */ | |
251 | if (new_win < cur_win) { | |
252 | /* Danger Will Robinson! | |
253 | * Don't update rcv_wup/rcv_wnd here or else | |
254 | * we will not be able to advertise a zero | |
255 | * window in time. --DaveM | |
256 | * | |
257 | * Relax Will Robinson. | |
258 | */ | |
259 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); | |
260 | } | |
261 | tp->rcv_wnd = new_win; | |
262 | tp->rcv_wup = tp->rcv_nxt; | |
263 | ||
264 | /* Make sure we do not exceed the maximum possible | |
265 | * scaled window. | |
266 | */ | |
267 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) | |
268 | new_win = min(new_win, MAX_TCP_WINDOW); | |
269 | else | |
270 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
271 | ||
272 | /* RFC1323 scaling applied */ | |
273 | new_win >>= tp->rx_opt.rcv_wscale; | |
274 | ||
275 | /* If we advertise zero window, disable fast path. */ | |
276 | if (new_win == 0) | |
277 | tp->pred_flags = 0; | |
278 | ||
279 | return new_win; | |
280 | } | |
281 | ||
282 | /* Packet ECN state for a SYN-ACK */ | |
283 | static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb) | |
284 | { | |
285 | TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR; | |
286 | if (!(tp->ecn_flags & TCP_ECN_OK)) | |
287 | TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE; | |
288 | } | |
289 | ||
290 | /* Packet ECN state for a SYN. */ | |
291 | static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) | |
292 | { | |
293 | struct tcp_sock *tp = tcp_sk(sk); | |
294 | ||
295 | tp->ecn_flags = 0; | |
296 | if (sysctl_tcp_ecn == 1) { | |
297 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR; | |
298 | tp->ecn_flags = TCP_ECN_OK; | |
299 | } | |
300 | } | |
301 | ||
302 | static __inline__ void | |
303 | TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) | |
304 | { | |
305 | if (inet_rsk(req)->ecn_ok) | |
306 | th->ece = 1; | |
307 | } | |
308 | ||
309 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to | |
310 | * be sent. | |
311 | */ | |
312 | static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, | |
313 | int tcp_header_len) | |
314 | { | |
315 | struct tcp_sock *tp = tcp_sk(sk); | |
316 | ||
317 | if (tp->ecn_flags & TCP_ECN_OK) { | |
318 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
319 | if (skb->len != tcp_header_len && | |
320 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
321 | INET_ECN_xmit(sk); | |
322 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { | |
323 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; | |
324 | tcp_hdr(skb)->cwr = 1; | |
325 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; | |
326 | } | |
327 | } else { | |
328 | /* ACK or retransmitted segment: clear ECT|CE */ | |
329 | INET_ECN_dontxmit(sk); | |
330 | } | |
331 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
332 | tcp_hdr(skb)->ece = 1; | |
333 | } | |
334 | } | |
335 | ||
336 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, | |
337 | * auto increment end seqno. | |
338 | */ | |
339 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) | |
340 | { | |
341 | skb->csum = 0; | |
342 | ||
343 | TCP_SKB_CB(skb)->flags = flags; | |
344 | TCP_SKB_CB(skb)->sacked = 0; | |
345 | ||
346 | skb_shinfo(skb)->gso_segs = 1; | |
347 | skb_shinfo(skb)->gso_size = 0; | |
348 | skb_shinfo(skb)->gso_type = 0; | |
349 | ||
350 | TCP_SKB_CB(skb)->seq = seq; | |
351 | if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN)) | |
352 | seq++; | |
353 | TCP_SKB_CB(skb)->end_seq = seq; | |
354 | } | |
355 | ||
356 | static inline int tcp_urg_mode(const struct tcp_sock *tp) | |
357 | { | |
358 | return tp->snd_una != tp->snd_up; | |
359 | } | |
360 | ||
361 | #define OPTION_SACK_ADVERTISE (1 << 0) | |
362 | #define OPTION_TS (1 << 1) | |
363 | #define OPTION_MD5 (1 << 2) | |
364 | #define OPTION_WSCALE (1 << 3) | |
365 | ||
366 | struct tcp_out_options { | |
367 | u8 options; /* bit field of OPTION_* */ | |
368 | u8 ws; /* window scale, 0 to disable */ | |
369 | u8 num_sack_blocks; /* number of SACK blocks to include */ | |
370 | u16 mss; /* 0 to disable */ | |
371 | __u32 tsval, tsecr; /* need to include OPTION_TS */ | |
372 | }; | |
373 | ||
374 | /* Write previously computed TCP options to the packet. | |
375 | * | |
376 | * Beware: Something in the Internet is very sensitive to the ordering of | |
377 | * TCP options, we learned this through the hard way, so be careful here. | |
378 | * Luckily we can at least blame others for their non-compliance but from | |
379 | * inter-operatibility perspective it seems that we're somewhat stuck with | |
380 | * the ordering which we have been using if we want to keep working with | |
381 | * those broken things (not that it currently hurts anybody as there isn't | |
382 | * particular reason why the ordering would need to be changed). | |
383 | * | |
384 | * At least SACK_PERM as the first option is known to lead to a disaster | |
385 | * (but it may well be that other scenarios fail similarly). | |
386 | */ | |
387 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, | |
388 | const struct tcp_out_options *opts, | |
389 | __u8 **md5_hash) { | |
390 | if (unlikely(OPTION_MD5 & opts->options)) { | |
391 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
392 | (TCPOPT_NOP << 16) | | |
393 | (TCPOPT_MD5SIG << 8) | | |
394 | TCPOLEN_MD5SIG); | |
395 | *md5_hash = (__u8 *)ptr; | |
396 | ptr += 4; | |
397 | } else { | |
398 | *md5_hash = NULL; | |
399 | } | |
400 | ||
401 | if (unlikely(opts->mss)) { | |
402 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
403 | (TCPOLEN_MSS << 16) | | |
404 | opts->mss); | |
405 | } | |
406 | ||
407 | if (likely(OPTION_TS & opts->options)) { | |
408 | if (unlikely(OPTION_SACK_ADVERTISE & opts->options)) { | |
409 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | | |
410 | (TCPOLEN_SACK_PERM << 16) | | |
411 | (TCPOPT_TIMESTAMP << 8) | | |
412 | TCPOLEN_TIMESTAMP); | |
413 | } else { | |
414 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
415 | (TCPOPT_NOP << 16) | | |
416 | (TCPOPT_TIMESTAMP << 8) | | |
417 | TCPOLEN_TIMESTAMP); | |
418 | } | |
419 | *ptr++ = htonl(opts->tsval); | |
420 | *ptr++ = htonl(opts->tsecr); | |
421 | } | |
422 | ||
423 | if (unlikely(OPTION_SACK_ADVERTISE & opts->options && | |
424 | !(OPTION_TS & opts->options))) { | |
425 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
426 | (TCPOPT_NOP << 16) | | |
427 | (TCPOPT_SACK_PERM << 8) | | |
428 | TCPOLEN_SACK_PERM); | |
429 | } | |
430 | ||
431 | if (unlikely(OPTION_WSCALE & opts->options)) { | |
432 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
433 | (TCPOPT_WINDOW << 16) | | |
434 | (TCPOLEN_WINDOW << 8) | | |
435 | opts->ws); | |
436 | } | |
437 | ||
438 | if (unlikely(opts->num_sack_blocks)) { | |
439 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
440 | tp->duplicate_sack : tp->selective_acks; | |
441 | int this_sack; | |
442 | ||
443 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
444 | (TCPOPT_NOP << 16) | | |
445 | (TCPOPT_SACK << 8) | | |
446 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * | |
447 | TCPOLEN_SACK_PERBLOCK))); | |
448 | ||
449 | for (this_sack = 0; this_sack < opts->num_sack_blocks; | |
450 | ++this_sack) { | |
451 | *ptr++ = htonl(sp[this_sack].start_seq); | |
452 | *ptr++ = htonl(sp[this_sack].end_seq); | |
453 | } | |
454 | ||
455 | tp->rx_opt.dsack = 0; | |
456 | } | |
457 | } | |
458 | ||
459 | /* Compute TCP options for SYN packets. This is not the final | |
460 | * network wire format yet. | |
461 | */ | |
462 | static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb, | |
463 | struct tcp_out_options *opts, | |
464 | struct tcp_md5sig_key **md5) { | |
465 | struct tcp_sock *tp = tcp_sk(sk); | |
466 | unsigned size = 0; | |
467 | struct dst_entry *dst = __sk_dst_get(sk); | |
468 | ||
469 | #ifdef CONFIG_TCP_MD5SIG | |
470 | *md5 = tp->af_specific->md5_lookup(sk, sk); | |
471 | if (*md5) { | |
472 | opts->options |= OPTION_MD5; | |
473 | size += TCPOLEN_MD5SIG_ALIGNED; | |
474 | } | |
475 | #else | |
476 | *md5 = NULL; | |
477 | #endif | |
478 | ||
479 | /* We always get an MSS option. The option bytes which will be seen in | |
480 | * normal data packets should timestamps be used, must be in the MSS | |
481 | * advertised. But we subtract them from tp->mss_cache so that | |
482 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
483 | * fact here if necessary. If we don't do this correctly, as a | |
484 | * receiver we won't recognize data packets as being full sized when we | |
485 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
486 | * SACKs don't matter, we never delay an ACK when we have any of those | |
487 | * going out. */ | |
488 | opts->mss = tcp_advertise_mss(sk); | |
489 | size += TCPOLEN_MSS_ALIGNED; | |
490 | ||
491 | if (likely(sysctl_tcp_timestamps && | |
492 | !dst_feature(dst, RTAX_FEATURE_NO_TSTAMP) && | |
493 | *md5 == NULL)) { | |
494 | opts->options |= OPTION_TS; | |
495 | opts->tsval = TCP_SKB_CB(skb)->when; | |
496 | opts->tsecr = tp->rx_opt.ts_recent; | |
497 | size += TCPOLEN_TSTAMP_ALIGNED; | |
498 | } | |
499 | if (likely(sysctl_tcp_window_scaling && | |
500 | !dst_feature(dst, RTAX_FEATURE_NO_WSCALE))) { | |
501 | opts->ws = tp->rx_opt.rcv_wscale; | |
502 | opts->options |= OPTION_WSCALE; | |
503 | size += TCPOLEN_WSCALE_ALIGNED; | |
504 | } | |
505 | if (likely(sysctl_tcp_sack && | |
506 | !dst_feature(dst, RTAX_FEATURE_NO_SACK))) { | |
507 | opts->options |= OPTION_SACK_ADVERTISE; | |
508 | if (unlikely(!(OPTION_TS & opts->options))) | |
509 | size += TCPOLEN_SACKPERM_ALIGNED; | |
510 | } | |
511 | ||
512 | return size; | |
513 | } | |
514 | ||
515 | /* Set up TCP options for SYN-ACKs. */ | |
516 | static unsigned tcp_synack_options(struct sock *sk, | |
517 | struct request_sock *req, | |
518 | unsigned mss, struct sk_buff *skb, | |
519 | struct tcp_out_options *opts, | |
520 | struct tcp_md5sig_key **md5) { | |
521 | unsigned size = 0; | |
522 | struct inet_request_sock *ireq = inet_rsk(req); | |
523 | char doing_ts; | |
524 | ||
525 | #ifdef CONFIG_TCP_MD5SIG | |
526 | *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); | |
527 | if (*md5) { | |
528 | opts->options |= OPTION_MD5; | |
529 | size += TCPOLEN_MD5SIG_ALIGNED; | |
530 | } | |
531 | #else | |
532 | *md5 = NULL; | |
533 | #endif | |
534 | ||
535 | /* we can't fit any SACK blocks in a packet with MD5 + TS | |
536 | options. There was discussion about disabling SACK rather than TS in | |
537 | order to fit in better with old, buggy kernels, but that was deemed | |
538 | to be unnecessary. */ | |
539 | doing_ts = ireq->tstamp_ok && !(*md5 && ireq->sack_ok); | |
540 | ||
541 | opts->mss = mss; | |
542 | size += TCPOLEN_MSS_ALIGNED; | |
543 | ||
544 | if (likely(ireq->wscale_ok)) { | |
545 | opts->ws = ireq->rcv_wscale; | |
546 | opts->options |= OPTION_WSCALE; | |
547 | size += TCPOLEN_WSCALE_ALIGNED; | |
548 | } | |
549 | if (likely(doing_ts)) { | |
550 | opts->options |= OPTION_TS; | |
551 | opts->tsval = TCP_SKB_CB(skb)->when; | |
552 | opts->tsecr = req->ts_recent; | |
553 | size += TCPOLEN_TSTAMP_ALIGNED; | |
554 | } | |
555 | if (likely(ireq->sack_ok)) { | |
556 | opts->options |= OPTION_SACK_ADVERTISE; | |
557 | if (unlikely(!doing_ts)) | |
558 | size += TCPOLEN_SACKPERM_ALIGNED; | |
559 | } | |
560 | ||
561 | return size; | |
562 | } | |
563 | ||
564 | /* Compute TCP options for ESTABLISHED sockets. This is not the | |
565 | * final wire format yet. | |
566 | */ | |
567 | static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb, | |
568 | struct tcp_out_options *opts, | |
569 | struct tcp_md5sig_key **md5) { | |
570 | struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; | |
571 | struct tcp_sock *tp = tcp_sk(sk); | |
572 | unsigned size = 0; | |
573 | unsigned int eff_sacks; | |
574 | ||
575 | #ifdef CONFIG_TCP_MD5SIG | |
576 | *md5 = tp->af_specific->md5_lookup(sk, sk); | |
577 | if (unlikely(*md5)) { | |
578 | opts->options |= OPTION_MD5; | |
579 | size += TCPOLEN_MD5SIG_ALIGNED; | |
580 | } | |
581 | #else | |
582 | *md5 = NULL; | |
583 | #endif | |
584 | ||
585 | if (likely(tp->rx_opt.tstamp_ok)) { | |
586 | opts->options |= OPTION_TS; | |
587 | opts->tsval = tcb ? tcb->when : 0; | |
588 | opts->tsecr = tp->rx_opt.ts_recent; | |
589 | size += TCPOLEN_TSTAMP_ALIGNED; | |
590 | } | |
591 | ||
592 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; | |
593 | if (unlikely(eff_sacks)) { | |
594 | const unsigned remaining = MAX_TCP_OPTION_SPACE - size; | |
595 | opts->num_sack_blocks = | |
596 | min_t(unsigned, eff_sacks, | |
597 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / | |
598 | TCPOLEN_SACK_PERBLOCK); | |
599 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
600 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
601 | } | |
602 | ||
603 | return size; | |
604 | } | |
605 | ||
606 | /* This routine actually transmits TCP packets queued in by | |
607 | * tcp_do_sendmsg(). This is used by both the initial | |
608 | * transmission and possible later retransmissions. | |
609 | * All SKB's seen here are completely headerless. It is our | |
610 | * job to build the TCP header, and pass the packet down to | |
611 | * IP so it can do the same plus pass the packet off to the | |
612 | * device. | |
613 | * | |
614 | * We are working here with either a clone of the original | |
615 | * SKB, or a fresh unique copy made by the retransmit engine. | |
616 | */ | |
617 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, | |
618 | gfp_t gfp_mask) | |
619 | { | |
620 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
621 | struct inet_sock *inet; | |
622 | struct tcp_sock *tp; | |
623 | struct tcp_skb_cb *tcb; | |
624 | struct tcp_out_options opts; | |
625 | unsigned tcp_options_size, tcp_header_size; | |
626 | struct tcp_md5sig_key *md5; | |
627 | __u8 *md5_hash_location; | |
628 | struct tcphdr *th; | |
629 | int err; | |
630 | ||
631 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
632 | ||
633 | /* If congestion control is doing timestamping, we must | |
634 | * take such a timestamp before we potentially clone/copy. | |
635 | */ | |
636 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) | |
637 | __net_timestamp(skb); | |
638 | ||
639 | if (likely(clone_it)) { | |
640 | if (unlikely(skb_cloned(skb))) | |
641 | skb = pskb_copy(skb, gfp_mask); | |
642 | else | |
643 | skb = skb_clone(skb, gfp_mask); | |
644 | if (unlikely(!skb)) | |
645 | return -ENOBUFS; | |
646 | } | |
647 | ||
648 | inet = inet_sk(sk); | |
649 | tp = tcp_sk(sk); | |
650 | tcb = TCP_SKB_CB(skb); | |
651 | memset(&opts, 0, sizeof(opts)); | |
652 | ||
653 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) | |
654 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); | |
655 | else | |
656 | tcp_options_size = tcp_established_options(sk, skb, &opts, | |
657 | &md5); | |
658 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); | |
659 | ||
660 | if (tcp_packets_in_flight(tp) == 0) | |
661 | tcp_ca_event(sk, CA_EVENT_TX_START); | |
662 | ||
663 | skb_push(skb, tcp_header_size); | |
664 | skb_reset_transport_header(skb); | |
665 | skb_set_owner_w(skb, sk); | |
666 | ||
667 | /* Build TCP header and checksum it. */ | |
668 | th = tcp_hdr(skb); | |
669 | th->source = inet->inet_sport; | |
670 | th->dest = inet->inet_dport; | |
671 | th->seq = htonl(tcb->seq); | |
672 | th->ack_seq = htonl(tp->rcv_nxt); | |
673 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | | |
674 | tcb->flags); | |
675 | ||
676 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { | |
677 | /* RFC1323: The window in SYN & SYN/ACK segments | |
678 | * is never scaled. | |
679 | */ | |
680 | th->window = htons(min(tp->rcv_wnd, 65535U)); | |
681 | } else { | |
682 | th->window = htons(tcp_select_window(sk)); | |
683 | } | |
684 | th->check = 0; | |
685 | th->urg_ptr = 0; | |
686 | ||
687 | /* The urg_mode check is necessary during a below snd_una win probe */ | |
688 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { | |
689 | if (before(tp->snd_up, tcb->seq + 0x10000)) { | |
690 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
691 | th->urg = 1; | |
692 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { | |
693 | th->urg_ptr = 0xFFFF; | |
694 | th->urg = 1; | |
695 | } | |
696 | } | |
697 | ||
698 | tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location); | |
699 | if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0)) | |
700 | TCP_ECN_send(sk, skb, tcp_header_size); | |
701 | ||
702 | #ifdef CONFIG_TCP_MD5SIG | |
703 | /* Calculate the MD5 hash, as we have all we need now */ | |
704 | if (md5) { | |
705 | sk->sk_route_caps &= ~NETIF_F_GSO_MASK; | |
706 | tp->af_specific->calc_md5_hash(md5_hash_location, | |
707 | md5, sk, NULL, skb); | |
708 | } | |
709 | #endif | |
710 | ||
711 | icsk->icsk_af_ops->send_check(sk, skb->len, skb); | |
712 | ||
713 | if (likely(tcb->flags & TCPCB_FLAG_ACK)) | |
714 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); | |
715 | ||
716 | if (skb->len != tcp_header_size) | |
717 | tcp_event_data_sent(tp, skb, sk); | |
718 | ||
719 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) | |
720 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); | |
721 | ||
722 | err = icsk->icsk_af_ops->queue_xmit(skb, 0); | |
723 | if (likely(err <= 0)) | |
724 | return err; | |
725 | ||
726 | tcp_enter_cwr(sk, 1); | |
727 | ||
728 | return net_xmit_eval(err); | |
729 | } | |
730 | ||
731 | /* This routine just queues the buffer for sending. | |
732 | * | |
733 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
734 | * otherwise socket can stall. | |
735 | */ | |
736 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
737 | { | |
738 | struct tcp_sock *tp = tcp_sk(sk); | |
739 | ||
740 | /* Advance write_seq and place onto the write_queue. */ | |
741 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
742 | skb_header_release(skb); | |
743 | tcp_add_write_queue_tail(sk, skb); | |
744 | sk->sk_wmem_queued += skb->truesize; | |
745 | sk_mem_charge(sk, skb->truesize); | |
746 | } | |
747 | ||
748 | /* Initialize TSO segments for a packet. */ | |
749 | static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, | |
750 | unsigned int mss_now) | |
751 | { | |
752 | if (skb->len <= mss_now || !sk_can_gso(sk) || | |
753 | skb->ip_summed == CHECKSUM_NONE) { | |
754 | /* Avoid the costly divide in the normal | |
755 | * non-TSO case. | |
756 | */ | |
757 | skb_shinfo(skb)->gso_segs = 1; | |
758 | skb_shinfo(skb)->gso_size = 0; | |
759 | skb_shinfo(skb)->gso_type = 0; | |
760 | } else { | |
761 | skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); | |
762 | skb_shinfo(skb)->gso_size = mss_now; | |
763 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; | |
764 | } | |
765 | } | |
766 | ||
767 | /* When a modification to fackets out becomes necessary, we need to check | |
768 | * skb is counted to fackets_out or not. | |
769 | */ | |
770 | static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, | |
771 | int decr) | |
772 | { | |
773 | struct tcp_sock *tp = tcp_sk(sk); | |
774 | ||
775 | if (!tp->sacked_out || tcp_is_reno(tp)) | |
776 | return; | |
777 | ||
778 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) | |
779 | tp->fackets_out -= decr; | |
780 | } | |
781 | ||
782 | /* Pcount in the middle of the write queue got changed, we need to do various | |
783 | * tweaks to fix counters | |
784 | */ | |
785 | static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr) | |
786 | { | |
787 | struct tcp_sock *tp = tcp_sk(sk); | |
788 | ||
789 | tp->packets_out -= decr; | |
790 | ||
791 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
792 | tp->sacked_out -= decr; | |
793 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
794 | tp->retrans_out -= decr; | |
795 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) | |
796 | tp->lost_out -= decr; | |
797 | ||
798 | /* Reno case is special. Sigh... */ | |
799 | if (tcp_is_reno(tp) && decr > 0) | |
800 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); | |
801 | ||
802 | tcp_adjust_fackets_out(sk, skb, decr); | |
803 | ||
804 | if (tp->lost_skb_hint && | |
805 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && | |
806 | (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) | |
807 | tp->lost_cnt_hint -= decr; | |
808 | ||
809 | tcp_verify_left_out(tp); | |
810 | } | |
811 | ||
812 | /* Function to create two new TCP segments. Shrinks the given segment | |
813 | * to the specified size and appends a new segment with the rest of the | |
814 | * packet to the list. This won't be called frequently, I hope. | |
815 | * Remember, these are still headerless SKBs at this point. | |
816 | */ | |
817 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, | |
818 | unsigned int mss_now) | |
819 | { | |
820 | struct tcp_sock *tp = tcp_sk(sk); | |
821 | struct sk_buff *buff; | |
822 | int nsize, old_factor; | |
823 | int nlen; | |
824 | u8 flags; | |
825 | ||
826 | BUG_ON(len > skb->len); | |
827 | ||
828 | nsize = skb_headlen(skb) - len; | |
829 | if (nsize < 0) | |
830 | nsize = 0; | |
831 | ||
832 | if (skb_cloned(skb) && | |
833 | skb_is_nonlinear(skb) && | |
834 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
835 | return -ENOMEM; | |
836 | ||
837 | /* Get a new skb... force flag on. */ | |
838 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); | |
839 | if (buff == NULL) | |
840 | return -ENOMEM; /* We'll just try again later. */ | |
841 | ||
842 | sk->sk_wmem_queued += buff->truesize; | |
843 | sk_mem_charge(sk, buff->truesize); | |
844 | nlen = skb->len - len - nsize; | |
845 | buff->truesize += nlen; | |
846 | skb->truesize -= nlen; | |
847 | ||
848 | /* Correct the sequence numbers. */ | |
849 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
850 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
851 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
852 | ||
853 | /* PSH and FIN should only be set in the second packet. */ | |
854 | flags = TCP_SKB_CB(skb)->flags; | |
855 | TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); | |
856 | TCP_SKB_CB(buff)->flags = flags; | |
857 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; | |
858 | ||
859 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { | |
860 | /* Copy and checksum data tail into the new buffer. */ | |
861 | buff->csum = csum_partial_copy_nocheck(skb->data + len, | |
862 | skb_put(buff, nsize), | |
863 | nsize, 0); | |
864 | ||
865 | skb_trim(skb, len); | |
866 | ||
867 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
868 | } else { | |
869 | skb->ip_summed = CHECKSUM_PARTIAL; | |
870 | skb_split(skb, buff, len); | |
871 | } | |
872 | ||
873 | buff->ip_summed = skb->ip_summed; | |
874 | ||
875 | /* Looks stupid, but our code really uses when of | |
876 | * skbs, which it never sent before. --ANK | |
877 | */ | |
878 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; | |
879 | buff->tstamp = skb->tstamp; | |
880 | ||
881 | old_factor = tcp_skb_pcount(skb); | |
882 | ||
883 | /* Fix up tso_factor for both original and new SKB. */ | |
884 | tcp_set_skb_tso_segs(sk, skb, mss_now); | |
885 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
886 | ||
887 | /* If this packet has been sent out already, we must | |
888 | * adjust the various packet counters. | |
889 | */ | |
890 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { | |
891 | int diff = old_factor - tcp_skb_pcount(skb) - | |
892 | tcp_skb_pcount(buff); | |
893 | ||
894 | if (diff) | |
895 | tcp_adjust_pcount(sk, skb, diff); | |
896 | } | |
897 | ||
898 | /* Link BUFF into the send queue. */ | |
899 | skb_header_release(buff); | |
900 | tcp_insert_write_queue_after(skb, buff, sk); | |
901 | ||
902 | return 0; | |
903 | } | |
904 | ||
905 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
906 | * eventually). The difference is that pulled data not copied, but | |
907 | * immediately discarded. | |
908 | */ | |
909 | static void __pskb_trim_head(struct sk_buff *skb, int len) | |
910 | { | |
911 | int i, k, eat; | |
912 | ||
913 | eat = len; | |
914 | k = 0; | |
915 | for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { | |
916 | if (skb_shinfo(skb)->frags[i].size <= eat) { | |
917 | put_page(skb_shinfo(skb)->frags[i].page); | |
918 | eat -= skb_shinfo(skb)->frags[i].size; | |
919 | } else { | |
920 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; | |
921 | if (eat) { | |
922 | skb_shinfo(skb)->frags[k].page_offset += eat; | |
923 | skb_shinfo(skb)->frags[k].size -= eat; | |
924 | eat = 0; | |
925 | } | |
926 | k++; | |
927 | } | |
928 | } | |
929 | skb_shinfo(skb)->nr_frags = k; | |
930 | ||
931 | skb_reset_tail_pointer(skb); | |
932 | skb->data_len -= len; | |
933 | skb->len = skb->data_len; | |
934 | } | |
935 | ||
936 | /* Remove acked data from a packet in the transmit queue. */ | |
937 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) | |
938 | { | |
939 | if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
940 | return -ENOMEM; | |
941 | ||
942 | /* If len == headlen, we avoid __skb_pull to preserve alignment. */ | |
943 | if (unlikely(len < skb_headlen(skb))) | |
944 | __skb_pull(skb, len); | |
945 | else | |
946 | __pskb_trim_head(skb, len - skb_headlen(skb)); | |
947 | ||
948 | TCP_SKB_CB(skb)->seq += len; | |
949 | skb->ip_summed = CHECKSUM_PARTIAL; | |
950 | ||
951 | skb->truesize -= len; | |
952 | sk->sk_wmem_queued -= len; | |
953 | sk_mem_uncharge(sk, len); | |
954 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); | |
955 | ||
956 | /* Any change of skb->len requires recalculation of tso | |
957 | * factor and mss. | |
958 | */ | |
959 | if (tcp_skb_pcount(skb) > 1) | |
960 | tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk)); | |
961 | ||
962 | return 0; | |
963 | } | |
964 | ||
965 | /* Calculate MSS. Not accounting for SACKs here. */ | |
966 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
967 | { | |
968 | struct tcp_sock *tp = tcp_sk(sk); | |
969 | struct inet_connection_sock *icsk = inet_csk(sk); | |
970 | int mss_now; | |
971 | ||
972 | /* Calculate base mss without TCP options: | |
973 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
974 | */ | |
975 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
976 | ||
977 | /* Clamp it (mss_clamp does not include tcp options) */ | |
978 | if (mss_now > tp->rx_opt.mss_clamp) | |
979 | mss_now = tp->rx_opt.mss_clamp; | |
980 | ||
981 | /* Now subtract optional transport overhead */ | |
982 | mss_now -= icsk->icsk_ext_hdr_len; | |
983 | ||
984 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
985 | if (mss_now < 48) | |
986 | mss_now = 48; | |
987 | ||
988 | /* Now subtract TCP options size, not including SACKs */ | |
989 | mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); | |
990 | ||
991 | return mss_now; | |
992 | } | |
993 | ||
994 | /* Inverse of above */ | |
995 | int tcp_mss_to_mtu(struct sock *sk, int mss) | |
996 | { | |
997 | struct tcp_sock *tp = tcp_sk(sk); | |
998 | struct inet_connection_sock *icsk = inet_csk(sk); | |
999 | int mtu; | |
1000 | ||
1001 | mtu = mss + | |
1002 | tp->tcp_header_len + | |
1003 | icsk->icsk_ext_hdr_len + | |
1004 | icsk->icsk_af_ops->net_header_len; | |
1005 | ||
1006 | return mtu; | |
1007 | } | |
1008 | ||
1009 | /* MTU probing init per socket */ | |
1010 | void tcp_mtup_init(struct sock *sk) | |
1011 | { | |
1012 | struct tcp_sock *tp = tcp_sk(sk); | |
1013 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1014 | ||
1015 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; | |
1016 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + | |
1017 | icsk->icsk_af_ops->net_header_len; | |
1018 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); | |
1019 | icsk->icsk_mtup.probe_size = 0; | |
1020 | } | |
1021 | ||
1022 | /* This function synchronize snd mss to current pmtu/exthdr set. | |
1023 | ||
1024 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
1025 | for TCP options, but includes only bare TCP header. | |
1026 | ||
1027 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
1028 | It is minimum of user_mss and mss received with SYN. | |
1029 | It also does not include TCP options. | |
1030 | ||
1031 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. | |
1032 | ||
1033 | tp->mss_cache is current effective sending mss, including | |
1034 | all tcp options except for SACKs. It is evaluated, | |
1035 | taking into account current pmtu, but never exceeds | |
1036 | tp->rx_opt.mss_clamp. | |
1037 | ||
1038 | NOTE1. rfc1122 clearly states that advertised MSS | |
1039 | DOES NOT include either tcp or ip options. | |
1040 | ||
1041 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache | |
1042 | are READ ONLY outside this function. --ANK (980731) | |
1043 | */ | |
1044 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) | |
1045 | { | |
1046 | struct tcp_sock *tp = tcp_sk(sk); | |
1047 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1048 | int mss_now; | |
1049 | ||
1050 | if (icsk->icsk_mtup.search_high > pmtu) | |
1051 | icsk->icsk_mtup.search_high = pmtu; | |
1052 | ||
1053 | mss_now = tcp_mtu_to_mss(sk, pmtu); | |
1054 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); | |
1055 | ||
1056 | /* And store cached results */ | |
1057 | icsk->icsk_pmtu_cookie = pmtu; | |
1058 | if (icsk->icsk_mtup.enabled) | |
1059 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
1060 | tp->mss_cache = mss_now; | |
1061 | ||
1062 | return mss_now; | |
1063 | } | |
1064 | ||
1065 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1066 | * and even PMTU discovery events into account. | |
1067 | */ | |
1068 | unsigned int tcp_current_mss(struct sock *sk) | |
1069 | { | |
1070 | struct tcp_sock *tp = tcp_sk(sk); | |
1071 | struct dst_entry *dst = __sk_dst_get(sk); | |
1072 | u32 mss_now; | |
1073 | unsigned header_len; | |
1074 | struct tcp_out_options opts; | |
1075 | struct tcp_md5sig_key *md5; | |
1076 | ||
1077 | mss_now = tp->mss_cache; | |
1078 | ||
1079 | if (dst) { | |
1080 | u32 mtu = dst_mtu(dst); | |
1081 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) | |
1082 | mss_now = tcp_sync_mss(sk, mtu); | |
1083 | } | |
1084 | ||
1085 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + | |
1086 | sizeof(struct tcphdr); | |
1087 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1088 | * some common options. If this is an odd packet (because we have SACK | |
1089 | * blocks etc) then our calculated header_len will be different, and | |
1090 | * we have to adjust mss_now correspondingly */ | |
1091 | if (header_len != tp->tcp_header_len) { | |
1092 | int delta = (int) header_len - tp->tcp_header_len; | |
1093 | mss_now -= delta; | |
1094 | } | |
1095 | ||
1096 | return mss_now; | |
1097 | } | |
1098 | ||
1099 | /* Congestion window validation. (RFC2861) */ | |
1100 | static void tcp_cwnd_validate(struct sock *sk) | |
1101 | { | |
1102 | struct tcp_sock *tp = tcp_sk(sk); | |
1103 | ||
1104 | if (tp->packets_out >= tp->snd_cwnd) { | |
1105 | /* Network is feed fully. */ | |
1106 | tp->snd_cwnd_used = 0; | |
1107 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1108 | } else { | |
1109 | /* Network starves. */ | |
1110 | if (tp->packets_out > tp->snd_cwnd_used) | |
1111 | tp->snd_cwnd_used = tp->packets_out; | |
1112 | ||
1113 | if (sysctl_tcp_slow_start_after_idle && | |
1114 | (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) | |
1115 | tcp_cwnd_application_limited(sk); | |
1116 | } | |
1117 | } | |
1118 | ||
1119 | /* Returns the portion of skb which can be sent right away without | |
1120 | * introducing MSS oddities to segment boundaries. In rare cases where | |
1121 | * mss_now != mss_cache, we will request caller to create a small skb | |
1122 | * per input skb which could be mostly avoided here (if desired). | |
1123 | * | |
1124 | * We explicitly want to create a request for splitting write queue tail | |
1125 | * to a small skb for Nagle purposes while avoiding unnecessary modulos, | |
1126 | * thus all the complexity (cwnd_len is always MSS multiple which we | |
1127 | * return whenever allowed by the other factors). Basically we need the | |
1128 | * modulo only when the receiver window alone is the limiting factor or | |
1129 | * when we would be allowed to send the split-due-to-Nagle skb fully. | |
1130 | */ | |
1131 | static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, | |
1132 | unsigned int mss_now, unsigned int cwnd) | |
1133 | { | |
1134 | struct tcp_sock *tp = tcp_sk(sk); | |
1135 | u32 needed, window, cwnd_len; | |
1136 | ||
1137 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; | |
1138 | cwnd_len = mss_now * cwnd; | |
1139 | ||
1140 | if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) | |
1141 | return cwnd_len; | |
1142 | ||
1143 | needed = min(skb->len, window); | |
1144 | ||
1145 | if (cwnd_len <= needed) | |
1146 | return cwnd_len; | |
1147 | ||
1148 | return needed - needed % mss_now; | |
1149 | } | |
1150 | ||
1151 | /* Can at least one segment of SKB be sent right now, according to the | |
1152 | * congestion window rules? If so, return how many segments are allowed. | |
1153 | */ | |
1154 | static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, | |
1155 | struct sk_buff *skb) | |
1156 | { | |
1157 | u32 in_flight, cwnd; | |
1158 | ||
1159 | /* Don't be strict about the congestion window for the final FIN. */ | |
1160 | if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && | |
1161 | tcp_skb_pcount(skb) == 1) | |
1162 | return 1; | |
1163 | ||
1164 | in_flight = tcp_packets_in_flight(tp); | |
1165 | cwnd = tp->snd_cwnd; | |
1166 | if (in_flight < cwnd) | |
1167 | return (cwnd - in_flight); | |
1168 | ||
1169 | return 0; | |
1170 | } | |
1171 | ||
1172 | /* Intialize TSO state of a skb. | |
1173 | * This must be invoked the first time we consider transmitting | |
1174 | * SKB onto the wire. | |
1175 | */ | |
1176 | static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, | |
1177 | unsigned int mss_now) | |
1178 | { | |
1179 | int tso_segs = tcp_skb_pcount(skb); | |
1180 | ||
1181 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { | |
1182 | tcp_set_skb_tso_segs(sk, skb, mss_now); | |
1183 | tso_segs = tcp_skb_pcount(skb); | |
1184 | } | |
1185 | return tso_segs; | |
1186 | } | |
1187 | ||
1188 | /* Minshall's variant of the Nagle send check. */ | |
1189 | static inline int tcp_minshall_check(const struct tcp_sock *tp) | |
1190 | { | |
1191 | return after(tp->snd_sml, tp->snd_una) && | |
1192 | !after(tp->snd_sml, tp->snd_nxt); | |
1193 | } | |
1194 | ||
1195 | /* Return 0, if packet can be sent now without violation Nagle's rules: | |
1196 | * 1. It is full sized. | |
1197 | * 2. Or it contains FIN. (already checked by caller) | |
1198 | * 3. Or TCP_NODELAY was set. | |
1199 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1200 | * With Minshall's modification: all sent small packets are ACKed. | |
1201 | */ | |
1202 | static inline int tcp_nagle_check(const struct tcp_sock *tp, | |
1203 | const struct sk_buff *skb, | |
1204 | unsigned mss_now, int nonagle) | |
1205 | { | |
1206 | return (skb->len < mss_now && | |
1207 | ((nonagle & TCP_NAGLE_CORK) || | |
1208 | (!nonagle && tp->packets_out && tcp_minshall_check(tp)))); | |
1209 | } | |
1210 | ||
1211 | /* Return non-zero if the Nagle test allows this packet to be | |
1212 | * sent now. | |
1213 | */ | |
1214 | static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, | |
1215 | unsigned int cur_mss, int nonagle) | |
1216 | { | |
1217 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1218 | * write_queue (they have no chances to get new data). | |
1219 | * | |
1220 | * This is implemented in the callers, where they modify the 'nonagle' | |
1221 | * argument based upon the location of SKB in the send queue. | |
1222 | */ | |
1223 | if (nonagle & TCP_NAGLE_PUSH) | |
1224 | return 1; | |
1225 | ||
1226 | /* Don't use the nagle rule for urgent data (or for the final FIN). | |
1227 | * Nagle can be ignored during F-RTO too (see RFC4138). | |
1228 | */ | |
1229 | if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || | |
1230 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) | |
1231 | return 1; | |
1232 | ||
1233 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) | |
1234 | return 1; | |
1235 | ||
1236 | return 0; | |
1237 | } | |
1238 | ||
1239 | /* Does at least the first segment of SKB fit into the send window? */ | |
1240 | static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, | |
1241 | unsigned int cur_mss) | |
1242 | { | |
1243 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1244 | ||
1245 | if (skb->len > cur_mss) | |
1246 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1247 | ||
1248 | return !after(end_seq, tcp_wnd_end(tp)); | |
1249 | } | |
1250 | ||
1251 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) | |
1252 | * should be put on the wire right now. If so, it returns the number of | |
1253 | * packets allowed by the congestion window. | |
1254 | */ | |
1255 | static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, | |
1256 | unsigned int cur_mss, int nonagle) | |
1257 | { | |
1258 | struct tcp_sock *tp = tcp_sk(sk); | |
1259 | unsigned int cwnd_quota; | |
1260 | ||
1261 | tcp_init_tso_segs(sk, skb, cur_mss); | |
1262 | ||
1263 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1264 | return 0; | |
1265 | ||
1266 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
1267 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) | |
1268 | cwnd_quota = 0; | |
1269 | ||
1270 | return cwnd_quota; | |
1271 | } | |
1272 | ||
1273 | /* Test if sending is allowed right now. */ | |
1274 | int tcp_may_send_now(struct sock *sk) | |
1275 | { | |
1276 | struct tcp_sock *tp = tcp_sk(sk); | |
1277 | struct sk_buff *skb = tcp_send_head(sk); | |
1278 | ||
1279 | return (skb && | |
1280 | tcp_snd_test(sk, skb, tcp_current_mss(sk), | |
1281 | (tcp_skb_is_last(sk, skb) ? | |
1282 | tp->nonagle : TCP_NAGLE_PUSH))); | |
1283 | } | |
1284 | ||
1285 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet | |
1286 | * which is put after SKB on the list. It is very much like | |
1287 | * tcp_fragment() except that it may make several kinds of assumptions | |
1288 | * in order to speed up the splitting operation. In particular, we | |
1289 | * know that all the data is in scatter-gather pages, and that the | |
1290 | * packet has never been sent out before (and thus is not cloned). | |
1291 | */ | |
1292 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, | |
1293 | unsigned int mss_now) | |
1294 | { | |
1295 | struct sk_buff *buff; | |
1296 | int nlen = skb->len - len; | |
1297 | u8 flags; | |
1298 | ||
1299 | /* All of a TSO frame must be composed of paged data. */ | |
1300 | if (skb->len != skb->data_len) | |
1301 | return tcp_fragment(sk, skb, len, mss_now); | |
1302 | ||
1303 | buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC); | |
1304 | if (unlikely(buff == NULL)) | |
1305 | return -ENOMEM; | |
1306 | ||
1307 | sk->sk_wmem_queued += buff->truesize; | |
1308 | sk_mem_charge(sk, buff->truesize); | |
1309 | buff->truesize += nlen; | |
1310 | skb->truesize -= nlen; | |
1311 | ||
1312 | /* Correct the sequence numbers. */ | |
1313 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1314 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1315 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1316 | ||
1317 | /* PSH and FIN should only be set in the second packet. */ | |
1318 | flags = TCP_SKB_CB(skb)->flags; | |
1319 | TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); | |
1320 | TCP_SKB_CB(buff)->flags = flags; | |
1321 | ||
1322 | /* This packet was never sent out yet, so no SACK bits. */ | |
1323 | TCP_SKB_CB(buff)->sacked = 0; | |
1324 | ||
1325 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; | |
1326 | skb_split(skb, buff, len); | |
1327 | ||
1328 | /* Fix up tso_factor for both original and new SKB. */ | |
1329 | tcp_set_skb_tso_segs(sk, skb, mss_now); | |
1330 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
1331 | ||
1332 | /* Link BUFF into the send queue. */ | |
1333 | skb_header_release(buff); | |
1334 | tcp_insert_write_queue_after(skb, buff, sk); | |
1335 | ||
1336 | return 0; | |
1337 | } | |
1338 | ||
1339 | /* Try to defer sending, if possible, in order to minimize the amount | |
1340 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
1341 | * | |
1342 | * This algorithm is from John Heffner. | |
1343 | */ | |
1344 | static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) | |
1345 | { | |
1346 | struct tcp_sock *tp = tcp_sk(sk); | |
1347 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
1348 | u32 send_win, cong_win, limit, in_flight; | |
1349 | ||
1350 | if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) | |
1351 | goto send_now; | |
1352 | ||
1353 | if (icsk->icsk_ca_state != TCP_CA_Open) | |
1354 | goto send_now; | |
1355 | ||
1356 | /* Defer for less than two clock ticks. */ | |
1357 | if (tp->tso_deferred && | |
1358 | (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) | |
1359 | goto send_now; | |
1360 | ||
1361 | in_flight = tcp_packets_in_flight(tp); | |
1362 | ||
1363 | BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); | |
1364 | ||
1365 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; | |
1366 | ||
1367 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
1368 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
1369 | ||
1370 | limit = min(send_win, cong_win); | |
1371 | ||
1372 | /* If a full-sized TSO skb can be sent, do it. */ | |
1373 | if (limit >= sk->sk_gso_max_size) | |
1374 | goto send_now; | |
1375 | ||
1376 | /* Middle in queue won't get any more data, full sendable already? */ | |
1377 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) | |
1378 | goto send_now; | |
1379 | ||
1380 | if (sysctl_tcp_tso_win_divisor) { | |
1381 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); | |
1382 | ||
1383 | /* If at least some fraction of a window is available, | |
1384 | * just use it. | |
1385 | */ | |
1386 | chunk /= sysctl_tcp_tso_win_divisor; | |
1387 | if (limit >= chunk) | |
1388 | goto send_now; | |
1389 | } else { | |
1390 | /* Different approach, try not to defer past a single | |
1391 | * ACK. Receiver should ACK every other full sized | |
1392 | * frame, so if we have space for more than 3 frames | |
1393 | * then send now. | |
1394 | */ | |
1395 | if (limit > tcp_max_burst(tp) * tp->mss_cache) | |
1396 | goto send_now; | |
1397 | } | |
1398 | ||
1399 | /* Ok, it looks like it is advisable to defer. */ | |
1400 | tp->tso_deferred = 1 | (jiffies << 1); | |
1401 | ||
1402 | return 1; | |
1403 | ||
1404 | send_now: | |
1405 | tp->tso_deferred = 0; | |
1406 | return 0; | |
1407 | } | |
1408 | ||
1409 | /* Create a new MTU probe if we are ready. | |
1410 | * MTU probe is regularly attempting to increase the path MTU by | |
1411 | * deliberately sending larger packets. This discovers routing | |
1412 | * changes resulting in larger path MTUs. | |
1413 | * | |
1414 | * Returns 0 if we should wait to probe (no cwnd available), | |
1415 | * 1 if a probe was sent, | |
1416 | * -1 otherwise | |
1417 | */ | |
1418 | static int tcp_mtu_probe(struct sock *sk) | |
1419 | { | |
1420 | struct tcp_sock *tp = tcp_sk(sk); | |
1421 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1422 | struct sk_buff *skb, *nskb, *next; | |
1423 | int len; | |
1424 | int probe_size; | |
1425 | int size_needed; | |
1426 | int copy; | |
1427 | int mss_now; | |
1428 | ||
1429 | /* Not currently probing/verifying, | |
1430 | * not in recovery, | |
1431 | * have enough cwnd, and | |
1432 | * not SACKing (the variable headers throw things off) */ | |
1433 | if (!icsk->icsk_mtup.enabled || | |
1434 | icsk->icsk_mtup.probe_size || | |
1435 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
1436 | tp->snd_cwnd < 11 || | |
1437 | tp->rx_opt.num_sacks || tp->rx_opt.dsack) | |
1438 | return -1; | |
1439 | ||
1440 | /* Very simple search strategy: just double the MSS. */ | |
1441 | mss_now = tcp_current_mss(sk); | |
1442 | probe_size = 2 * tp->mss_cache; | |
1443 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; | |
1444 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { | |
1445 | /* TODO: set timer for probe_converge_event */ | |
1446 | return -1; | |
1447 | } | |
1448 | ||
1449 | /* Have enough data in the send queue to probe? */ | |
1450 | if (tp->write_seq - tp->snd_nxt < size_needed) | |
1451 | return -1; | |
1452 | ||
1453 | if (tp->snd_wnd < size_needed) | |
1454 | return -1; | |
1455 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) | |
1456 | return 0; | |
1457 | ||
1458 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ | |
1459 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { | |
1460 | if (!tcp_packets_in_flight(tp)) | |
1461 | return -1; | |
1462 | else | |
1463 | return 0; | |
1464 | } | |
1465 | ||
1466 | /* We're allowed to probe. Build it now. */ | |
1467 | if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) | |
1468 | return -1; | |
1469 | sk->sk_wmem_queued += nskb->truesize; | |
1470 | sk_mem_charge(sk, nskb->truesize); | |
1471 | ||
1472 | skb = tcp_send_head(sk); | |
1473 | ||
1474 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
1475 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
1476 | TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK; | |
1477 | TCP_SKB_CB(nskb)->sacked = 0; | |
1478 | nskb->csum = 0; | |
1479 | nskb->ip_summed = skb->ip_summed; | |
1480 | ||
1481 | tcp_insert_write_queue_before(nskb, skb, sk); | |
1482 | ||
1483 | len = 0; | |
1484 | tcp_for_write_queue_from_safe(skb, next, sk) { | |
1485 | copy = min_t(int, skb->len, probe_size - len); | |
1486 | if (nskb->ip_summed) | |
1487 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); | |
1488 | else | |
1489 | nskb->csum = skb_copy_and_csum_bits(skb, 0, | |
1490 | skb_put(nskb, copy), | |
1491 | copy, nskb->csum); | |
1492 | ||
1493 | if (skb->len <= copy) { | |
1494 | /* We've eaten all the data from this skb. | |
1495 | * Throw it away. */ | |
1496 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; | |
1497 | tcp_unlink_write_queue(skb, sk); | |
1498 | sk_wmem_free_skb(sk, skb); | |
1499 | } else { | |
1500 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & | |
1501 | ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); | |
1502 | if (!skb_shinfo(skb)->nr_frags) { | |
1503 | skb_pull(skb, copy); | |
1504 | if (skb->ip_summed != CHECKSUM_PARTIAL) | |
1505 | skb->csum = csum_partial(skb->data, | |
1506 | skb->len, 0); | |
1507 | } else { | |
1508 | __pskb_trim_head(skb, copy); | |
1509 | tcp_set_skb_tso_segs(sk, skb, mss_now); | |
1510 | } | |
1511 | TCP_SKB_CB(skb)->seq += copy; | |
1512 | } | |
1513 | ||
1514 | len += copy; | |
1515 | ||
1516 | if (len >= probe_size) | |
1517 | break; | |
1518 | } | |
1519 | tcp_init_tso_segs(sk, nskb, nskb->len); | |
1520 | ||
1521 | /* We're ready to send. If this fails, the probe will | |
1522 | * be resegmented into mss-sized pieces by tcp_write_xmit(). */ | |
1523 | TCP_SKB_CB(nskb)->when = tcp_time_stamp; | |
1524 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { | |
1525 | /* Decrement cwnd here because we are sending | |
1526 | * effectively two packets. */ | |
1527 | tp->snd_cwnd--; | |
1528 | tcp_event_new_data_sent(sk, nskb); | |
1529 | ||
1530 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
1531 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; | |
1532 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
1533 | ||
1534 | return 1; | |
1535 | } | |
1536 | ||
1537 | return -1; | |
1538 | } | |
1539 | ||
1540 | /* This routine writes packets to the network. It advances the | |
1541 | * send_head. This happens as incoming acks open up the remote | |
1542 | * window for us. | |
1543 | * | |
1544 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between | |
1545 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into | |
1546 | * account rare use of URG, this is not a big flaw. | |
1547 | * | |
1548 | * Returns 1, if no segments are in flight and we have queued segments, but | |
1549 | * cannot send anything now because of SWS or another problem. | |
1550 | */ | |
1551 | static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, | |
1552 | int push_one, gfp_t gfp) | |
1553 | { | |
1554 | struct tcp_sock *tp = tcp_sk(sk); | |
1555 | struct sk_buff *skb; | |
1556 | unsigned int tso_segs, sent_pkts; | |
1557 | int cwnd_quota; | |
1558 | int result; | |
1559 | ||
1560 | sent_pkts = 0; | |
1561 | ||
1562 | if (!push_one) { | |
1563 | /* Do MTU probing. */ | |
1564 | result = tcp_mtu_probe(sk); | |
1565 | if (!result) { | |
1566 | return 0; | |
1567 | } else if (result > 0) { | |
1568 | sent_pkts = 1; | |
1569 | } | |
1570 | } | |
1571 | ||
1572 | while ((skb = tcp_send_head(sk))) { | |
1573 | unsigned int limit; | |
1574 | ||
1575 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); | |
1576 | BUG_ON(!tso_segs); | |
1577 | ||
1578 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
1579 | if (!cwnd_quota) | |
1580 | break; | |
1581 | ||
1582 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) | |
1583 | break; | |
1584 | ||
1585 | if (tso_segs == 1) { | |
1586 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, | |
1587 | (tcp_skb_is_last(sk, skb) ? | |
1588 | nonagle : TCP_NAGLE_PUSH)))) | |
1589 | break; | |
1590 | } else { | |
1591 | if (!push_one && tcp_tso_should_defer(sk, skb)) | |
1592 | break; | |
1593 | } | |
1594 | ||
1595 | limit = mss_now; | |
1596 | if (tso_segs > 1 && !tcp_urg_mode(tp)) | |
1597 | limit = tcp_mss_split_point(sk, skb, mss_now, | |
1598 | cwnd_quota); | |
1599 | ||
1600 | if (skb->len > limit && | |
1601 | unlikely(tso_fragment(sk, skb, limit, mss_now))) | |
1602 | break; | |
1603 | ||
1604 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1605 | ||
1606 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) | |
1607 | break; | |
1608 | ||
1609 | /* Advance the send_head. This one is sent out. | |
1610 | * This call will increment packets_out. | |
1611 | */ | |
1612 | tcp_event_new_data_sent(sk, skb); | |
1613 | ||
1614 | tcp_minshall_update(tp, mss_now, skb); | |
1615 | sent_pkts++; | |
1616 | ||
1617 | if (push_one) | |
1618 | break; | |
1619 | } | |
1620 | ||
1621 | if (likely(sent_pkts)) { | |
1622 | tcp_cwnd_validate(sk); | |
1623 | return 0; | |
1624 | } | |
1625 | return !tp->packets_out && tcp_send_head(sk); | |
1626 | } | |
1627 | ||
1628 | /* Push out any pending frames which were held back due to | |
1629 | * TCP_CORK or attempt at coalescing tiny packets. | |
1630 | * The socket must be locked by the caller. | |
1631 | */ | |
1632 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, | |
1633 | int nonagle) | |
1634 | { | |
1635 | struct sk_buff *skb = tcp_send_head(sk); | |
1636 | ||
1637 | if (!skb) | |
1638 | return; | |
1639 | ||
1640 | /* If we are closed, the bytes will have to remain here. | |
1641 | * In time closedown will finish, we empty the write queue and | |
1642 | * all will be happy. | |
1643 | */ | |
1644 | if (unlikely(sk->sk_state == TCP_CLOSE)) | |
1645 | return; | |
1646 | ||
1647 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC)) | |
1648 | tcp_check_probe_timer(sk); | |
1649 | } | |
1650 | ||
1651 | /* Send _single_ skb sitting at the send head. This function requires | |
1652 | * true push pending frames to setup probe timer etc. | |
1653 | */ | |
1654 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
1655 | { | |
1656 | struct sk_buff *skb = tcp_send_head(sk); | |
1657 | ||
1658 | BUG_ON(!skb || skb->len < mss_now); | |
1659 | ||
1660 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); | |
1661 | } | |
1662 | ||
1663 | /* This function returns the amount that we can raise the | |
1664 | * usable window based on the following constraints | |
1665 | * | |
1666 | * 1. The window can never be shrunk once it is offered (RFC 793) | |
1667 | * 2. We limit memory per socket | |
1668 | * | |
1669 | * RFC 1122: | |
1670 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
1671 | * RECV.NEXT + RCV.WIN fixed until: | |
1672 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
1673 | * | |
1674 | * i.e. don't raise the right edge of the window until you can raise | |
1675 | * it at least MSS bytes. | |
1676 | * | |
1677 | * Unfortunately, the recommended algorithm breaks header prediction, | |
1678 | * since header prediction assumes th->window stays fixed. | |
1679 | * | |
1680 | * Strictly speaking, keeping th->window fixed violates the receiver | |
1681 | * side SWS prevention criteria. The problem is that under this rule | |
1682 | * a stream of single byte packets will cause the right side of the | |
1683 | * window to always advance by a single byte. | |
1684 | * | |
1685 | * Of course, if the sender implements sender side SWS prevention | |
1686 | * then this will not be a problem. | |
1687 | * | |
1688 | * BSD seems to make the following compromise: | |
1689 | * | |
1690 | * If the free space is less than the 1/4 of the maximum | |
1691 | * space available and the free space is less than 1/2 mss, | |
1692 | * then set the window to 0. | |
1693 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
1694 | * Otherwise, just prevent the window from shrinking | |
1695 | * and from being larger than the largest representable value. | |
1696 | * | |
1697 | * This prevents incremental opening of the window in the regime | |
1698 | * where TCP is limited by the speed of the reader side taking | |
1699 | * data out of the TCP receive queue. It does nothing about | |
1700 | * those cases where the window is constrained on the sender side | |
1701 | * because the pipeline is full. | |
1702 | * | |
1703 | * BSD also seems to "accidentally" limit itself to windows that are a | |
1704 | * multiple of MSS, at least until the free space gets quite small. | |
1705 | * This would appear to be a side effect of the mbuf implementation. | |
1706 | * Combining these two algorithms results in the observed behavior | |
1707 | * of having a fixed window size at almost all times. | |
1708 | * | |
1709 | * Below we obtain similar behavior by forcing the offered window to | |
1710 | * a multiple of the mss when it is feasible to do so. | |
1711 | * | |
1712 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
1713 | * Regular options like TIMESTAMP are taken into account. | |
1714 | */ | |
1715 | u32 __tcp_select_window(struct sock *sk) | |
1716 | { | |
1717 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1718 | struct tcp_sock *tp = tcp_sk(sk); | |
1719 | /* MSS for the peer's data. Previous versions used mss_clamp | |
1720 | * here. I don't know if the value based on our guesses | |
1721 | * of peer's MSS is better for the performance. It's more correct | |
1722 | * but may be worse for the performance because of rcv_mss | |
1723 | * fluctuations. --SAW 1998/11/1 | |
1724 | */ | |
1725 | int mss = icsk->icsk_ack.rcv_mss; | |
1726 | int free_space = tcp_space(sk); | |
1727 | int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); | |
1728 | int window; | |
1729 | ||
1730 | if (mss > full_space) | |
1731 | mss = full_space; | |
1732 | ||
1733 | if (free_space < (full_space >> 1)) { | |
1734 | icsk->icsk_ack.quick = 0; | |
1735 | ||
1736 | if (tcp_memory_pressure) | |
1737 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, | |
1738 | 4U * tp->advmss); | |
1739 | ||
1740 | if (free_space < mss) | |
1741 | return 0; | |
1742 | } | |
1743 | ||
1744 | if (free_space > tp->rcv_ssthresh) | |
1745 | free_space = tp->rcv_ssthresh; | |
1746 | ||
1747 | /* Don't do rounding if we are using window scaling, since the | |
1748 | * scaled window will not line up with the MSS boundary anyway. | |
1749 | */ | |
1750 | window = tp->rcv_wnd; | |
1751 | if (tp->rx_opt.rcv_wscale) { | |
1752 | window = free_space; | |
1753 | ||
1754 | /* Advertise enough space so that it won't get scaled away. | |
1755 | * Import case: prevent zero window announcement if | |
1756 | * 1<<rcv_wscale > mss. | |
1757 | */ | |
1758 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) | |
1759 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) | |
1760 | << tp->rx_opt.rcv_wscale); | |
1761 | } else { | |
1762 | /* Get the largest window that is a nice multiple of mss. | |
1763 | * Window clamp already applied above. | |
1764 | * If our current window offering is within 1 mss of the | |
1765 | * free space we just keep it. This prevents the divide | |
1766 | * and multiply from happening most of the time. | |
1767 | * We also don't do any window rounding when the free space | |
1768 | * is too small. | |
1769 | */ | |
1770 | if (window <= free_space - mss || window > free_space) | |
1771 | window = (free_space / mss) * mss; | |
1772 | else if (mss == full_space && | |
1773 | free_space > window + (full_space >> 1)) | |
1774 | window = free_space; | |
1775 | } | |
1776 | ||
1777 | return window; | |
1778 | } | |
1779 | ||
1780 | /* Collapses two adjacent SKB's during retransmission. */ | |
1781 | static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) | |
1782 | { | |
1783 | struct tcp_sock *tp = tcp_sk(sk); | |
1784 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); | |
1785 | int skb_size, next_skb_size; | |
1786 | ||
1787 | skb_size = skb->len; | |
1788 | next_skb_size = next_skb->len; | |
1789 | ||
1790 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); | |
1791 | ||
1792 | tcp_highest_sack_combine(sk, next_skb, skb); | |
1793 | ||
1794 | tcp_unlink_write_queue(next_skb, sk); | |
1795 | ||
1796 | skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), | |
1797 | next_skb_size); | |
1798 | ||
1799 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) | |
1800 | skb->ip_summed = CHECKSUM_PARTIAL; | |
1801 | ||
1802 | if (skb->ip_summed != CHECKSUM_PARTIAL) | |
1803 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); | |
1804 | ||
1805 | /* Update sequence range on original skb. */ | |
1806 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1807 | ||
1808 | /* Merge over control information. This moves PSH/FIN etc. over */ | |
1809 | TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags; | |
1810 | ||
1811 | /* All done, get rid of second SKB and account for it so | |
1812 | * packet counting does not break. | |
1813 | */ | |
1814 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; | |
1815 | ||
1816 | /* changed transmit queue under us so clear hints */ | |
1817 | tcp_clear_retrans_hints_partial(tp); | |
1818 | if (next_skb == tp->retransmit_skb_hint) | |
1819 | tp->retransmit_skb_hint = skb; | |
1820 | ||
1821 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); | |
1822 | ||
1823 | sk_wmem_free_skb(sk, next_skb); | |
1824 | } | |
1825 | ||
1826 | /* Check if coalescing SKBs is legal. */ | |
1827 | static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb) | |
1828 | { | |
1829 | if (tcp_skb_pcount(skb) > 1) | |
1830 | return 0; | |
1831 | /* TODO: SACK collapsing could be used to remove this condition */ | |
1832 | if (skb_shinfo(skb)->nr_frags != 0) | |
1833 | return 0; | |
1834 | if (skb_cloned(skb)) | |
1835 | return 0; | |
1836 | if (skb == tcp_send_head(sk)) | |
1837 | return 0; | |
1838 | /* Some heurestics for collapsing over SACK'd could be invented */ | |
1839 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
1840 | return 0; | |
1841 | ||
1842 | return 1; | |
1843 | } | |
1844 | ||
1845 | /* Collapse packets in the retransmit queue to make to create | |
1846 | * less packets on the wire. This is only done on retransmission. | |
1847 | */ | |
1848 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, | |
1849 | int space) | |
1850 | { | |
1851 | struct tcp_sock *tp = tcp_sk(sk); | |
1852 | struct sk_buff *skb = to, *tmp; | |
1853 | int first = 1; | |
1854 | ||
1855 | if (!sysctl_tcp_retrans_collapse) | |
1856 | return; | |
1857 | if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) | |
1858 | return; | |
1859 | ||
1860 | tcp_for_write_queue_from_safe(skb, tmp, sk) { | |
1861 | if (!tcp_can_collapse(sk, skb)) | |
1862 | break; | |
1863 | ||
1864 | space -= skb->len; | |
1865 | ||
1866 | if (first) { | |
1867 | first = 0; | |
1868 | continue; | |
1869 | } | |
1870 | ||
1871 | if (space < 0) | |
1872 | break; | |
1873 | /* Punt if not enough space exists in the first SKB for | |
1874 | * the data in the second | |
1875 | */ | |
1876 | if (skb->len > skb_tailroom(to)) | |
1877 | break; | |
1878 | ||
1879 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) | |
1880 | break; | |
1881 | ||
1882 | tcp_collapse_retrans(sk, to); | |
1883 | } | |
1884 | } | |
1885 | ||
1886 | /* This retransmits one SKB. Policy decisions and retransmit queue | |
1887 | * state updates are done by the caller. Returns non-zero if an | |
1888 | * error occurred which prevented the send. | |
1889 | */ | |
1890 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) | |
1891 | { | |
1892 | struct tcp_sock *tp = tcp_sk(sk); | |
1893 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1894 | unsigned int cur_mss; | |
1895 | int err; | |
1896 | ||
1897 | /* Inconslusive MTU probe */ | |
1898 | if (icsk->icsk_mtup.probe_size) { | |
1899 | icsk->icsk_mtup.probe_size = 0; | |
1900 | } | |
1901 | ||
1902 | /* Do not sent more than we queued. 1/4 is reserved for possible | |
1903 | * copying overhead: fragmentation, tunneling, mangling etc. | |
1904 | */ | |
1905 | if (atomic_read(&sk->sk_wmem_alloc) > | |
1906 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) | |
1907 | return -EAGAIN; | |
1908 | ||
1909 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { | |
1910 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1911 | BUG(); | |
1912 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) | |
1913 | return -ENOMEM; | |
1914 | } | |
1915 | ||
1916 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) | |
1917 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
1918 | ||
1919 | cur_mss = tcp_current_mss(sk); | |
1920 | ||
1921 | /* If receiver has shrunk his window, and skb is out of | |
1922 | * new window, do not retransmit it. The exception is the | |
1923 | * case, when window is shrunk to zero. In this case | |
1924 | * our retransmit serves as a zero window probe. | |
1925 | */ | |
1926 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && | |
1927 | TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1928 | return -EAGAIN; | |
1929 | ||
1930 | if (skb->len > cur_mss) { | |
1931 | if (tcp_fragment(sk, skb, cur_mss, cur_mss)) | |
1932 | return -ENOMEM; /* We'll try again later. */ | |
1933 | } else { | |
1934 | int oldpcount = tcp_skb_pcount(skb); | |
1935 | ||
1936 | if (unlikely(oldpcount > 1)) { | |
1937 | tcp_init_tso_segs(sk, skb, cur_mss); | |
1938 | tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); | |
1939 | } | |
1940 | } | |
1941 | ||
1942 | tcp_retrans_try_collapse(sk, skb, cur_mss); | |
1943 | ||
1944 | /* Some Solaris stacks overoptimize and ignore the FIN on a | |
1945 | * retransmit when old data is attached. So strip it off | |
1946 | * since it is cheap to do so and saves bytes on the network. | |
1947 | */ | |
1948 | if (skb->len > 0 && | |
1949 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && | |
1950 | tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { | |
1951 | if (!pskb_trim(skb, 0)) { | |
1952 | /* Reuse, even though it does some unnecessary work */ | |
1953 | tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, | |
1954 | TCP_SKB_CB(skb)->flags); | |
1955 | skb->ip_summed = CHECKSUM_NONE; | |
1956 | } | |
1957 | } | |
1958 | ||
1959 | /* Make a copy, if the first transmission SKB clone we made | |
1960 | * is still in somebody's hands, else make a clone. | |
1961 | */ | |
1962 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1963 | ||
1964 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); | |
1965 | ||
1966 | if (err == 0) { | |
1967 | /* Update global TCP statistics. */ | |
1968 | TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); | |
1969 | ||
1970 | tp->total_retrans++; | |
1971 | ||
1972 | #if FASTRETRANS_DEBUG > 0 | |
1973 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { | |
1974 | if (net_ratelimit()) | |
1975 | printk(KERN_DEBUG "retrans_out leaked.\n"); | |
1976 | } | |
1977 | #endif | |
1978 | if (!tp->retrans_out) | |
1979 | tp->lost_retrans_low = tp->snd_nxt; | |
1980 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; | |
1981 | tp->retrans_out += tcp_skb_pcount(skb); | |
1982 | ||
1983 | /* Save stamp of the first retransmit. */ | |
1984 | if (!tp->retrans_stamp) | |
1985 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; | |
1986 | ||
1987 | tp->undo_retrans++; | |
1988 | ||
1989 | /* snd_nxt is stored to detect loss of retransmitted segment, | |
1990 | * see tcp_input.c tcp_sacktag_write_queue(). | |
1991 | */ | |
1992 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; | |
1993 | } | |
1994 | return err; | |
1995 | } | |
1996 | ||
1997 | /* Check if we forward retransmits are possible in the current | |
1998 | * window/congestion state. | |
1999 | */ | |
2000 | static int tcp_can_forward_retransmit(struct sock *sk) | |
2001 | { | |
2002 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
2003 | struct tcp_sock *tp = tcp_sk(sk); | |
2004 | ||
2005 | /* Forward retransmissions are possible only during Recovery. */ | |
2006 | if (icsk->icsk_ca_state != TCP_CA_Recovery) | |
2007 | return 0; | |
2008 | ||
2009 | /* No forward retransmissions in Reno are possible. */ | |
2010 | if (tcp_is_reno(tp)) | |
2011 | return 0; | |
2012 | ||
2013 | /* Yeah, we have to make difficult choice between forward transmission | |
2014 | * and retransmission... Both ways have their merits... | |
2015 | * | |
2016 | * For now we do not retransmit anything, while we have some new | |
2017 | * segments to send. In the other cases, follow rule 3 for | |
2018 | * NextSeg() specified in RFC3517. | |
2019 | */ | |
2020 | ||
2021 | if (tcp_may_send_now(sk)) | |
2022 | return 0; | |
2023 | ||
2024 | return 1; | |
2025 | } | |
2026 | ||
2027 | /* This gets called after a retransmit timeout, and the initially | |
2028 | * retransmitted data is acknowledged. It tries to continue | |
2029 | * resending the rest of the retransmit queue, until either | |
2030 | * we've sent it all or the congestion window limit is reached. | |
2031 | * If doing SACK, the first ACK which comes back for a timeout | |
2032 | * based retransmit packet might feed us FACK information again. | |
2033 | * If so, we use it to avoid unnecessarily retransmissions. | |
2034 | */ | |
2035 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
2036 | { | |
2037 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
2038 | struct tcp_sock *tp = tcp_sk(sk); | |
2039 | struct sk_buff *skb; | |
2040 | struct sk_buff *hole = NULL; | |
2041 | u32 last_lost; | |
2042 | int mib_idx; | |
2043 | int fwd_rexmitting = 0; | |
2044 | ||
2045 | if (!tp->lost_out) | |
2046 | tp->retransmit_high = tp->snd_una; | |
2047 | ||
2048 | if (tp->retransmit_skb_hint) { | |
2049 | skb = tp->retransmit_skb_hint; | |
2050 | last_lost = TCP_SKB_CB(skb)->end_seq; | |
2051 | if (after(last_lost, tp->retransmit_high)) | |
2052 | last_lost = tp->retransmit_high; | |
2053 | } else { | |
2054 | skb = tcp_write_queue_head(sk); | |
2055 | last_lost = tp->snd_una; | |
2056 | } | |
2057 | ||
2058 | tcp_for_write_queue_from(skb, sk) { | |
2059 | __u8 sacked = TCP_SKB_CB(skb)->sacked; | |
2060 | ||
2061 | if (skb == tcp_send_head(sk)) | |
2062 | break; | |
2063 | /* we could do better than to assign each time */ | |
2064 | if (hole == NULL) | |
2065 | tp->retransmit_skb_hint = skb; | |
2066 | ||
2067 | /* Assume this retransmit will generate | |
2068 | * only one packet for congestion window | |
2069 | * calculation purposes. This works because | |
2070 | * tcp_retransmit_skb() will chop up the | |
2071 | * packet to be MSS sized and all the | |
2072 | * packet counting works out. | |
2073 | */ | |
2074 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) | |
2075 | return; | |
2076 | ||
2077 | if (fwd_rexmitting) { | |
2078 | begin_fwd: | |
2079 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) | |
2080 | break; | |
2081 | mib_idx = LINUX_MIB_TCPFORWARDRETRANS; | |
2082 | ||
2083 | } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { | |
2084 | tp->retransmit_high = last_lost; | |
2085 | if (!tcp_can_forward_retransmit(sk)) | |
2086 | break; | |
2087 | /* Backtrack if necessary to non-L'ed skb */ | |
2088 | if (hole != NULL) { | |
2089 | skb = hole; | |
2090 | hole = NULL; | |
2091 | } | |
2092 | fwd_rexmitting = 1; | |
2093 | goto begin_fwd; | |
2094 | ||
2095 | } else if (!(sacked & TCPCB_LOST)) { | |
2096 | if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) | |
2097 | hole = skb; | |
2098 | continue; | |
2099 | ||
2100 | } else { | |
2101 | last_lost = TCP_SKB_CB(skb)->end_seq; | |
2102 | if (icsk->icsk_ca_state != TCP_CA_Loss) | |
2103 | mib_idx = LINUX_MIB_TCPFASTRETRANS; | |
2104 | else | |
2105 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; | |
2106 | } | |
2107 | ||
2108 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) | |
2109 | continue; | |
2110 | ||
2111 | if (tcp_retransmit_skb(sk, skb)) | |
2112 | return; | |
2113 | NET_INC_STATS_BH(sock_net(sk), mib_idx); | |
2114 | ||
2115 | if (skb == tcp_write_queue_head(sk)) | |
2116 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
2117 | inet_csk(sk)->icsk_rto, | |
2118 | TCP_RTO_MAX); | |
2119 | } | |
2120 | } | |
2121 | ||
2122 | /* Send a fin. The caller locks the socket for us. This cannot be | |
2123 | * allowed to fail queueing a FIN frame under any circumstances. | |
2124 | */ | |
2125 | void tcp_send_fin(struct sock *sk) | |
2126 | { | |
2127 | struct tcp_sock *tp = tcp_sk(sk); | |
2128 | struct sk_buff *skb = tcp_write_queue_tail(sk); | |
2129 | int mss_now; | |
2130 | ||
2131 | /* Optimization, tack on the FIN if we have a queue of | |
2132 | * unsent frames. But be careful about outgoing SACKS | |
2133 | * and IP options. | |
2134 | */ | |
2135 | mss_now = tcp_current_mss(sk); | |
2136 | ||
2137 | if (tcp_send_head(sk) != NULL) { | |
2138 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; | |
2139 | TCP_SKB_CB(skb)->end_seq++; | |
2140 | tp->write_seq++; | |
2141 | } else { | |
2142 | /* Socket is locked, keep trying until memory is available. */ | |
2143 | for (;;) { | |
2144 | skb = alloc_skb_fclone(MAX_TCP_HEADER, | |
2145 | sk->sk_allocation); | |
2146 | if (skb) | |
2147 | break; | |
2148 | yield(); | |
2149 | } | |
2150 | ||
2151 | /* Reserve space for headers and prepare control bits. */ | |
2152 | skb_reserve(skb, MAX_TCP_HEADER); | |
2153 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ | |
2154 | tcp_init_nondata_skb(skb, tp->write_seq, | |
2155 | TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); | |
2156 | tcp_queue_skb(sk, skb); | |
2157 | } | |
2158 | __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); | |
2159 | } | |
2160 | ||
2161 | /* We get here when a process closes a file descriptor (either due to | |
2162 | * an explicit close() or as a byproduct of exit()'ing) and there | |
2163 | * was unread data in the receive queue. This behavior is recommended | |
2164 | * by RFC 2525, section 2.17. -DaveM | |
2165 | */ | |
2166 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) | |
2167 | { | |
2168 | struct sk_buff *skb; | |
2169 | ||
2170 | /* NOTE: No TCP options attached and we never retransmit this. */ | |
2171 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
2172 | if (!skb) { | |
2173 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); | |
2174 | return; | |
2175 | } | |
2176 | ||
2177 | /* Reserve space for headers and prepare control bits. */ | |
2178 | skb_reserve(skb, MAX_TCP_HEADER); | |
2179 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), | |
2180 | TCPCB_FLAG_ACK | TCPCB_FLAG_RST); | |
2181 | /* Send it off. */ | |
2182 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
2183 | if (tcp_transmit_skb(sk, skb, 0, priority)) | |
2184 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); | |
2185 | ||
2186 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); | |
2187 | } | |
2188 | ||
2189 | /* Send a crossed SYN-ACK during socket establishment. | |
2190 | * WARNING: This routine must only be called when we have already sent | |
2191 | * a SYN packet that crossed the incoming SYN that caused this routine | |
2192 | * to get called. If this assumption fails then the initial rcv_wnd | |
2193 | * and rcv_wscale values will not be correct. | |
2194 | */ | |
2195 | int tcp_send_synack(struct sock *sk) | |
2196 | { | |
2197 | struct sk_buff *skb; | |
2198 | ||
2199 | skb = tcp_write_queue_head(sk); | |
2200 | if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) { | |
2201 | printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); | |
2202 | return -EFAULT; | |
2203 | } | |
2204 | if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) { | |
2205 | if (skb_cloned(skb)) { | |
2206 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); | |
2207 | if (nskb == NULL) | |
2208 | return -ENOMEM; | |
2209 | tcp_unlink_write_queue(skb, sk); | |
2210 | skb_header_release(nskb); | |
2211 | __tcp_add_write_queue_head(sk, nskb); | |
2212 | sk_wmem_free_skb(sk, skb); | |
2213 | sk->sk_wmem_queued += nskb->truesize; | |
2214 | sk_mem_charge(sk, nskb->truesize); | |
2215 | skb = nskb; | |
2216 | } | |
2217 | ||
2218 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; | |
2219 | TCP_ECN_send_synack(tcp_sk(sk), skb); | |
2220 | } | |
2221 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
2222 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); | |
2223 | } | |
2224 | ||
2225 | /* Prepare a SYN-ACK. */ | |
2226 | struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, | |
2227 | struct request_sock *req, | |
2228 | struct request_values *rvp) | |
2229 | { | |
2230 | struct inet_request_sock *ireq = inet_rsk(req); | |
2231 | struct tcp_sock *tp = tcp_sk(sk); | |
2232 | struct tcphdr *th; | |
2233 | int tcp_header_size; | |
2234 | struct tcp_out_options opts; | |
2235 | struct sk_buff *skb; | |
2236 | struct tcp_md5sig_key *md5; | |
2237 | __u8 *md5_hash_location; | |
2238 | int mss; | |
2239 | ||
2240 | skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); | |
2241 | if (skb == NULL) | |
2242 | return NULL; | |
2243 | ||
2244 | /* Reserve space for headers. */ | |
2245 | skb_reserve(skb, MAX_TCP_HEADER); | |
2246 | ||
2247 | skb_dst_set(skb, dst_clone(dst)); | |
2248 | ||
2249 | mss = dst_metric(dst, RTAX_ADVMSS); | |
2250 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) | |
2251 | mss = tp->rx_opt.user_mss; | |
2252 | ||
2253 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ | |
2254 | __u8 rcv_wscale; | |
2255 | /* Set this up on the first call only */ | |
2256 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); | |
2257 | /* tcp_full_space because it is guaranteed to be the first packet */ | |
2258 | tcp_select_initial_window(tcp_full_space(sk), | |
2259 | mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), | |
2260 | &req->rcv_wnd, | |
2261 | &req->window_clamp, | |
2262 | ireq->wscale_ok, | |
2263 | &rcv_wscale); | |
2264 | ireq->rcv_wscale = rcv_wscale; | |
2265 | } | |
2266 | ||
2267 | memset(&opts, 0, sizeof(opts)); | |
2268 | #ifdef CONFIG_SYN_COOKIES | |
2269 | if (unlikely(req->cookie_ts)) | |
2270 | TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); | |
2271 | else | |
2272 | #endif | |
2273 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
2274 | tcp_header_size = tcp_synack_options(sk, req, mss, | |
2275 | skb, &opts, &md5) + | |
2276 | sizeof(struct tcphdr); | |
2277 | ||
2278 | skb_push(skb, tcp_header_size); | |
2279 | skb_reset_transport_header(skb); | |
2280 | ||
2281 | th = tcp_hdr(skb); | |
2282 | memset(th, 0, sizeof(struct tcphdr)); | |
2283 | th->syn = 1; | |
2284 | th->ack = 1; | |
2285 | TCP_ECN_make_synack(req, th); | |
2286 | th->source = ireq->loc_port; | |
2287 | th->dest = ireq->rmt_port; | |
2288 | /* Setting of flags are superfluous here for callers (and ECE is | |
2289 | * not even correctly set) | |
2290 | */ | |
2291 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, | |
2292 | TCPCB_FLAG_SYN | TCPCB_FLAG_ACK); | |
2293 | th->seq = htonl(TCP_SKB_CB(skb)->seq); | |
2294 | th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); | |
2295 | ||
2296 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
2297 | th->window = htons(min(req->rcv_wnd, 65535U)); | |
2298 | tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location); | |
2299 | th->doff = (tcp_header_size >> 2); | |
2300 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); | |
2301 | ||
2302 | #ifdef CONFIG_TCP_MD5SIG | |
2303 | /* Okay, we have all we need - do the md5 hash if needed */ | |
2304 | if (md5) { | |
2305 | tcp_rsk(req)->af_specific->calc_md5_hash(md5_hash_location, | |
2306 | md5, NULL, req, skb); | |
2307 | } | |
2308 | #endif | |
2309 | ||
2310 | return skb; | |
2311 | } | |
2312 | ||
2313 | /* Do all connect socket setups that can be done AF independent. */ | |
2314 | static void tcp_connect_init(struct sock *sk) | |
2315 | { | |
2316 | struct dst_entry *dst = __sk_dst_get(sk); | |
2317 | struct tcp_sock *tp = tcp_sk(sk); | |
2318 | __u8 rcv_wscale; | |
2319 | ||
2320 | /* We'll fix this up when we get a response from the other end. | |
2321 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
2322 | */ | |
2323 | tp->tcp_header_len = sizeof(struct tcphdr) + | |
2324 | (sysctl_tcp_timestamps && | |
2325 | (!dst_feature(dst, RTAX_FEATURE_NO_TSTAMP) ? | |
2326 | TCPOLEN_TSTAMP_ALIGNED : 0)); | |
2327 | ||
2328 | #ifdef CONFIG_TCP_MD5SIG | |
2329 | if (tp->af_specific->md5_lookup(sk, sk) != NULL) | |
2330 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; | |
2331 | #endif | |
2332 | ||
2333 | /* If user gave his TCP_MAXSEG, record it to clamp */ | |
2334 | if (tp->rx_opt.user_mss) | |
2335 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
2336 | tp->max_window = 0; | |
2337 | tcp_mtup_init(sk); | |
2338 | tcp_sync_mss(sk, dst_mtu(dst)); | |
2339 | ||
2340 | if (!tp->window_clamp) | |
2341 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
2342 | tp->advmss = dst_metric(dst, RTAX_ADVMSS); | |
2343 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) | |
2344 | tp->advmss = tp->rx_opt.user_mss; | |
2345 | ||
2346 | tcp_initialize_rcv_mss(sk); | |
2347 | ||
2348 | tcp_select_initial_window(tcp_full_space(sk), | |
2349 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), | |
2350 | &tp->rcv_wnd, | |
2351 | &tp->window_clamp, | |
2352 | (sysctl_tcp_window_scaling && | |
2353 | !dst_feature(dst, RTAX_FEATURE_NO_WSCALE)), | |
2354 | &rcv_wscale); | |
2355 | ||
2356 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
2357 | tp->rcv_ssthresh = tp->rcv_wnd; | |
2358 | ||
2359 | sk->sk_err = 0; | |
2360 | sock_reset_flag(sk, SOCK_DONE); | |
2361 | tp->snd_wnd = 0; | |
2362 | tcp_init_wl(tp, 0); | |
2363 | tp->snd_una = tp->write_seq; | |
2364 | tp->snd_sml = tp->write_seq; | |
2365 | tp->snd_up = tp->write_seq; | |
2366 | tp->rcv_nxt = 0; | |
2367 | tp->rcv_wup = 0; | |
2368 | tp->copied_seq = 0; | |
2369 | ||
2370 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; | |
2371 | inet_csk(sk)->icsk_retransmits = 0; | |
2372 | tcp_clear_retrans(tp); | |
2373 | } | |
2374 | ||
2375 | /* Build a SYN and send it off. */ | |
2376 | int tcp_connect(struct sock *sk) | |
2377 | { | |
2378 | struct tcp_sock *tp = tcp_sk(sk); | |
2379 | struct sk_buff *buff; | |
2380 | ||
2381 | tcp_connect_init(sk); | |
2382 | ||
2383 | buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); | |
2384 | if (unlikely(buff == NULL)) | |
2385 | return -ENOBUFS; | |
2386 | ||
2387 | /* Reserve space for headers. */ | |
2388 | skb_reserve(buff, MAX_TCP_HEADER); | |
2389 | ||
2390 | tp->snd_nxt = tp->write_seq; | |
2391 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN); | |
2392 | TCP_ECN_send_syn(sk, buff); | |
2393 | ||
2394 | /* Send it off. */ | |
2395 | TCP_SKB_CB(buff)->when = tcp_time_stamp; | |
2396 | tp->retrans_stamp = TCP_SKB_CB(buff)->when; | |
2397 | skb_header_release(buff); | |
2398 | __tcp_add_write_queue_tail(sk, buff); | |
2399 | sk->sk_wmem_queued += buff->truesize; | |
2400 | sk_mem_charge(sk, buff->truesize); | |
2401 | tp->packets_out += tcp_skb_pcount(buff); | |
2402 | tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); | |
2403 | ||
2404 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
2405 | * in order to make this packet get counted in tcpOutSegs. | |
2406 | */ | |
2407 | tp->snd_nxt = tp->write_seq; | |
2408 | tp->pushed_seq = tp->write_seq; | |
2409 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); | |
2410 | ||
2411 | /* Timer for repeating the SYN until an answer. */ | |
2412 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
2413 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
2414 | return 0; | |
2415 | } | |
2416 | ||
2417 | /* Send out a delayed ack, the caller does the policy checking | |
2418 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
2419 | * for details. | |
2420 | */ | |
2421 | void tcp_send_delayed_ack(struct sock *sk) | |
2422 | { | |
2423 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2424 | int ato = icsk->icsk_ack.ato; | |
2425 | unsigned long timeout; | |
2426 | ||
2427 | if (ato > TCP_DELACK_MIN) { | |
2428 | const struct tcp_sock *tp = tcp_sk(sk); | |
2429 | int max_ato = HZ / 2; | |
2430 | ||
2431 | if (icsk->icsk_ack.pingpong || | |
2432 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) | |
2433 | max_ato = TCP_DELACK_MAX; | |
2434 | ||
2435 | /* Slow path, intersegment interval is "high". */ | |
2436 | ||
2437 | /* If some rtt estimate is known, use it to bound delayed ack. | |
2438 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements | |
2439 | * directly. | |
2440 | */ | |
2441 | if (tp->srtt) { | |
2442 | int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); | |
2443 | ||
2444 | if (rtt < max_ato) | |
2445 | max_ato = rtt; | |
2446 | } | |
2447 | ||
2448 | ato = min(ato, max_ato); | |
2449 | } | |
2450 | ||
2451 | /* Stay within the limit we were given */ | |
2452 | timeout = jiffies + ato; | |
2453 | ||
2454 | /* Use new timeout only if there wasn't a older one earlier. */ | |
2455 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { | |
2456 | /* If delack timer was blocked or is about to expire, | |
2457 | * send ACK now. | |
2458 | */ | |
2459 | if (icsk->icsk_ack.blocked || | |
2460 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
2461 | tcp_send_ack(sk); | |
2462 | return; | |
2463 | } | |
2464 | ||
2465 | if (!time_before(timeout, icsk->icsk_ack.timeout)) | |
2466 | timeout = icsk->icsk_ack.timeout; | |
2467 | } | |
2468 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; | |
2469 | icsk->icsk_ack.timeout = timeout; | |
2470 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
2471 | } | |
2472 | ||
2473 | /* This routine sends an ack and also updates the window. */ | |
2474 | void tcp_send_ack(struct sock *sk) | |
2475 | { | |
2476 | struct sk_buff *buff; | |
2477 | ||
2478 | /* If we have been reset, we may not send again. */ | |
2479 | if (sk->sk_state == TCP_CLOSE) | |
2480 | return; | |
2481 | ||
2482 | /* We are not putting this on the write queue, so | |
2483 | * tcp_transmit_skb() will set the ownership to this | |
2484 | * sock. | |
2485 | */ | |
2486 | buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
2487 | if (buff == NULL) { | |
2488 | inet_csk_schedule_ack(sk); | |
2489 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; | |
2490 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, | |
2491 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
2492 | return; | |
2493 | } | |
2494 | ||
2495 | /* Reserve space for headers and prepare control bits. */ | |
2496 | skb_reserve(buff, MAX_TCP_HEADER); | |
2497 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK); | |
2498 | ||
2499 | /* Send it off, this clears delayed acks for us. */ | |
2500 | TCP_SKB_CB(buff)->when = tcp_time_stamp; | |
2501 | tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); | |
2502 | } | |
2503 | ||
2504 | /* This routine sends a packet with an out of date sequence | |
2505 | * number. It assumes the other end will try to ack it. | |
2506 | * | |
2507 | * Question: what should we make while urgent mode? | |
2508 | * 4.4BSD forces sending single byte of data. We cannot send | |
2509 | * out of window data, because we have SND.NXT==SND.MAX... | |
2510 | * | |
2511 | * Current solution: to send TWO zero-length segments in urgent mode: | |
2512 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
2513 | * out-of-date with SND.UNA-1 to probe window. | |
2514 | */ | |
2515 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) | |
2516 | { | |
2517 | struct tcp_sock *tp = tcp_sk(sk); | |
2518 | struct sk_buff *skb; | |
2519 | ||
2520 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
2521 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
2522 | if (skb == NULL) | |
2523 | return -1; | |
2524 | ||
2525 | /* Reserve space for headers and set control bits. */ | |
2526 | skb_reserve(skb, MAX_TCP_HEADER); | |
2527 | /* Use a previous sequence. This should cause the other | |
2528 | * end to send an ack. Don't queue or clone SKB, just | |
2529 | * send it. | |
2530 | */ | |
2531 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK); | |
2532 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
2533 | return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); | |
2534 | } | |
2535 | ||
2536 | /* Initiate keepalive or window probe from timer. */ | |
2537 | int tcp_write_wakeup(struct sock *sk) | |
2538 | { | |
2539 | struct tcp_sock *tp = tcp_sk(sk); | |
2540 | struct sk_buff *skb; | |
2541 | ||
2542 | if (sk->sk_state == TCP_CLOSE) | |
2543 | return -1; | |
2544 | ||
2545 | if ((skb = tcp_send_head(sk)) != NULL && | |
2546 | before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { | |
2547 | int err; | |
2548 | unsigned int mss = tcp_current_mss(sk); | |
2549 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; | |
2550 | ||
2551 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
2552 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
2553 | ||
2554 | /* We are probing the opening of a window | |
2555 | * but the window size is != 0 | |
2556 | * must have been a result SWS avoidance ( sender ) | |
2557 | */ | |
2558 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
2559 | skb->len > mss) { | |
2560 | seg_size = min(seg_size, mss); | |
2561 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; | |
2562 | if (tcp_fragment(sk, skb, seg_size, mss)) | |
2563 | return -1; | |
2564 | } else if (!tcp_skb_pcount(skb)) | |
2565 | tcp_set_skb_tso_segs(sk, skb, mss); | |
2566 | ||
2567 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; | |
2568 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
2569 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); | |
2570 | if (!err) | |
2571 | tcp_event_new_data_sent(sk, skb); | |
2572 | return err; | |
2573 | } else { | |
2574 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) | |
2575 | tcp_xmit_probe_skb(sk, 1); | |
2576 | return tcp_xmit_probe_skb(sk, 0); | |
2577 | } | |
2578 | } | |
2579 | ||
2580 | /* A window probe timeout has occurred. If window is not closed send | |
2581 | * a partial packet else a zero probe. | |
2582 | */ | |
2583 | void tcp_send_probe0(struct sock *sk) | |
2584 | { | |
2585 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2586 | struct tcp_sock *tp = tcp_sk(sk); | |
2587 | int err; | |
2588 | ||
2589 | err = tcp_write_wakeup(sk); | |
2590 | ||
2591 | if (tp->packets_out || !tcp_send_head(sk)) { | |
2592 | /* Cancel probe timer, if it is not required. */ | |
2593 | icsk->icsk_probes_out = 0; | |
2594 | icsk->icsk_backoff = 0; | |
2595 | return; | |
2596 | } | |
2597 | ||
2598 | if (err <= 0) { | |
2599 | if (icsk->icsk_backoff < sysctl_tcp_retries2) | |
2600 | icsk->icsk_backoff++; | |
2601 | icsk->icsk_probes_out++; | |
2602 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, | |
2603 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), | |
2604 | TCP_RTO_MAX); | |
2605 | } else { | |
2606 | /* If packet was not sent due to local congestion, | |
2607 | * do not backoff and do not remember icsk_probes_out. | |
2608 | * Let local senders to fight for local resources. | |
2609 | * | |
2610 | * Use accumulated backoff yet. | |
2611 | */ | |
2612 | if (!icsk->icsk_probes_out) | |
2613 | icsk->icsk_probes_out = 1; | |
2614 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, | |
2615 | min(icsk->icsk_rto << icsk->icsk_backoff, | |
2616 | TCP_RESOURCE_PROBE_INTERVAL), | |
2617 | TCP_RTO_MAX); | |
2618 | } | |
2619 | } | |
2620 | ||
2621 | EXPORT_SYMBOL(tcp_select_initial_window); | |
2622 | EXPORT_SYMBOL(tcp_connect); | |
2623 | EXPORT_SYMBOL(tcp_make_synack); | |
2624 | EXPORT_SYMBOL(tcp_simple_retransmit); | |
2625 | EXPORT_SYMBOL(tcp_sync_mss); | |
2626 | EXPORT_SYMBOL(tcp_mtup_init); |