]>
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 | #define pr_fmt(fmt) "TCP: " fmt | |
38 | ||
39 | #include <net/tcp.h> | |
40 | ||
41 | #include <linux/compiler.h> | |
42 | #include <linux/gfp.h> | |
43 | #include <linux/module.h> | |
44 | ||
45 | /* People can turn this off for buggy TCP's found in printers etc. */ | |
46 | int sysctl_tcp_retrans_collapse __read_mostly = 1; | |
47 | ||
48 | /* People can turn this on to work with those rare, broken TCPs that | |
49 | * interpret the window field as a signed quantity. | |
50 | */ | |
51 | int sysctl_tcp_workaround_signed_windows __read_mostly = 0; | |
52 | ||
53 | /* Default TSQ limit of four TSO segments */ | |
54 | int sysctl_tcp_limit_output_bytes __read_mostly = 262144; | |
55 | ||
56 | /* This limits the percentage of the congestion window which we | |
57 | * will allow a single TSO frame to consume. Building TSO frames | |
58 | * which are too large can cause TCP streams to be bursty. | |
59 | */ | |
60 | int sysctl_tcp_tso_win_divisor __read_mostly = 3; | |
61 | ||
62 | /* By default, RFC2861 behavior. */ | |
63 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; | |
64 | ||
65 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, | |
66 | int push_one, gfp_t gfp); | |
67 | ||
68 | /* Account for new data that has been sent to the network. */ | |
69 | static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) | |
70 | { | |
71 | struct inet_connection_sock *icsk = inet_csk(sk); | |
72 | struct tcp_sock *tp = tcp_sk(sk); | |
73 | unsigned int prior_packets = tp->packets_out; | |
74 | ||
75 | tcp_advance_send_head(sk, skb); | |
76 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; | |
77 | ||
78 | tp->packets_out += tcp_skb_pcount(skb); | |
79 | if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) | |
80 | tcp_rearm_rto(sk); | |
81 | ||
82 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, | |
83 | tcp_skb_pcount(skb)); | |
84 | } | |
85 | ||
86 | /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one | |
87 | * window scaling factor due to loss of precision. | |
88 | * If window has been shrunk, what should we make? It is not clear at all. | |
89 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
90 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
91 | * invalid. OK, let's make this for now: | |
92 | */ | |
93 | static inline __u32 tcp_acceptable_seq(const struct sock *sk) | |
94 | { | |
95 | const struct tcp_sock *tp = tcp_sk(sk); | |
96 | ||
97 | if (!before(tcp_wnd_end(tp), tp->snd_nxt) || | |
98 | (tp->rx_opt.wscale_ok && | |
99 | ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) | |
100 | return tp->snd_nxt; | |
101 | else | |
102 | return tcp_wnd_end(tp); | |
103 | } | |
104 | ||
105 | /* Calculate mss to advertise in SYN segment. | |
106 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
107 | * | |
108 | * 1. It is independent of path mtu. | |
109 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
110 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
111 | * attached devices, because some buggy hosts are confused by | |
112 | * large MSS. | |
113 | * 4. We do not make 3, we advertise MSS, calculated from first | |
114 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
115 | * This may be overridden via information stored in routing table. | |
116 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
117 | * probably even Jumbo". | |
118 | */ | |
119 | static __u16 tcp_advertise_mss(struct sock *sk) | |
120 | { | |
121 | struct tcp_sock *tp = tcp_sk(sk); | |
122 | const struct dst_entry *dst = __sk_dst_get(sk); | |
123 | int mss = tp->advmss; | |
124 | ||
125 | if (dst) { | |
126 | unsigned int metric = dst_metric_advmss(dst); | |
127 | ||
128 | if (metric < mss) { | |
129 | mss = metric; | |
130 | tp->advmss = mss; | |
131 | } | |
132 | } | |
133 | ||
134 | return (__u16)mss; | |
135 | } | |
136 | ||
137 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
138 | * This is the first part of cwnd validation mechanism. | |
139 | */ | |
140 | void tcp_cwnd_restart(struct sock *sk, s32 delta) | |
141 | { | |
142 | struct tcp_sock *tp = tcp_sk(sk); | |
143 | u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); | |
144 | u32 cwnd = tp->snd_cwnd; | |
145 | ||
146 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); | |
147 | ||
148 | tp->snd_ssthresh = tcp_current_ssthresh(sk); | |
149 | restart_cwnd = min(restart_cwnd, cwnd); | |
150 | ||
151 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) | |
152 | cwnd >>= 1; | |
153 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
154 | tp->snd_cwnd_stamp = tcp_jiffies32; | |
155 | tp->snd_cwnd_used = 0; | |
156 | } | |
157 | ||
158 | /* Congestion state accounting after a packet has been sent. */ | |
159 | static void tcp_event_data_sent(struct tcp_sock *tp, | |
160 | struct sock *sk) | |
161 | { | |
162 | struct inet_connection_sock *icsk = inet_csk(sk); | |
163 | const u32 now = tcp_jiffies32; | |
164 | ||
165 | if (tcp_packets_in_flight(tp) == 0) | |
166 | tcp_ca_event(sk, CA_EVENT_TX_START); | |
167 | ||
168 | tp->lsndtime = now; | |
169 | ||
170 | /* If it is a reply for ato after last received | |
171 | * packet, enter pingpong mode. | |
172 | */ | |
173 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) | |
174 | icsk->icsk_ack.pingpong = 1; | |
175 | } | |
176 | ||
177 | /* Account for an ACK we sent. */ | |
178 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) | |
179 | { | |
180 | tcp_dec_quickack_mode(sk, pkts); | |
181 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
182 | } | |
183 | ||
184 | ||
185 | u32 tcp_default_init_rwnd(u32 mss) | |
186 | { | |
187 | /* Initial receive window should be twice of TCP_INIT_CWND to | |
188 | * enable proper sending of new unsent data during fast recovery | |
189 | * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a | |
190 | * limit when mss is larger than 1460. | |
191 | */ | |
192 | u32 init_rwnd = TCP_INIT_CWND * 2; | |
193 | ||
194 | if (mss > 1460) | |
195 | init_rwnd = max((1460 * init_rwnd) / mss, 2U); | |
196 | return init_rwnd; | |
197 | } | |
198 | ||
199 | /* Determine a window scaling and initial window to offer. | |
200 | * Based on the assumption that the given amount of space | |
201 | * will be offered. Store the results in the tp structure. | |
202 | * NOTE: for smooth operation initial space offering should | |
203 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
204 | * This MUST be enforced by all callers. | |
205 | */ | |
206 | void tcp_select_initial_window(int __space, __u32 mss, | |
207 | __u32 *rcv_wnd, __u32 *window_clamp, | |
208 | int wscale_ok, __u8 *rcv_wscale, | |
209 | __u32 init_rcv_wnd) | |
210 | { | |
211 | unsigned int space = (__space < 0 ? 0 : __space); | |
212 | ||
213 | /* If no clamp set the clamp to the max possible scaled window */ | |
214 | if (*window_clamp == 0) | |
215 | (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); | |
216 | space = min(*window_clamp, space); | |
217 | ||
218 | /* Quantize space offering to a multiple of mss if possible. */ | |
219 | if (space > mss) | |
220 | space = rounddown(space, mss); | |
221 | ||
222 | /* NOTE: offering an initial window larger than 32767 | |
223 | * will break some buggy TCP stacks. If the admin tells us | |
224 | * it is likely we could be speaking with such a buggy stack | |
225 | * we will truncate our initial window offering to 32K-1 | |
226 | * unless the remote has sent us a window scaling option, | |
227 | * which we interpret as a sign the remote TCP is not | |
228 | * misinterpreting the window field as a signed quantity. | |
229 | */ | |
230 | if (sysctl_tcp_workaround_signed_windows) | |
231 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
232 | else | |
233 | (*rcv_wnd) = space; | |
234 | ||
235 | (*rcv_wscale) = 0; | |
236 | if (wscale_ok) { | |
237 | /* Set window scaling on max possible window */ | |
238 | space = max_t(u32, space, sysctl_tcp_rmem[2]); | |
239 | space = max_t(u32, space, sysctl_rmem_max); | |
240 | space = min_t(u32, space, *window_clamp); | |
241 | while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) { | |
242 | space >>= 1; | |
243 | (*rcv_wscale)++; | |
244 | } | |
245 | } | |
246 | ||
247 | if (mss > (1 << *rcv_wscale)) { | |
248 | if (!init_rcv_wnd) /* Use default unless specified otherwise */ | |
249 | init_rcv_wnd = tcp_default_init_rwnd(mss); | |
250 | *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); | |
251 | } | |
252 | ||
253 | /* Set the clamp no higher than max representable value */ | |
254 | (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); | |
255 | } | |
256 | EXPORT_SYMBOL(tcp_select_initial_window); | |
257 | ||
258 | /* Chose a new window to advertise, update state in tcp_sock for the | |
259 | * socket, and return result with RFC1323 scaling applied. The return | |
260 | * value can be stuffed directly into th->window for an outgoing | |
261 | * frame. | |
262 | */ | |
263 | static u16 tcp_select_window(struct sock *sk) | |
264 | { | |
265 | struct tcp_sock *tp = tcp_sk(sk); | |
266 | u32 old_win = tp->rcv_wnd; | |
267 | u32 cur_win = tcp_receive_window(tp); | |
268 | u32 new_win = __tcp_select_window(sk); | |
269 | ||
270 | /* Never shrink the offered window */ | |
271 | if (new_win < cur_win) { | |
272 | /* Danger Will Robinson! | |
273 | * Don't update rcv_wup/rcv_wnd here or else | |
274 | * we will not be able to advertise a zero | |
275 | * window in time. --DaveM | |
276 | * | |
277 | * Relax Will Robinson. | |
278 | */ | |
279 | if (new_win == 0) | |
280 | NET_INC_STATS(sock_net(sk), | |
281 | LINUX_MIB_TCPWANTZEROWINDOWADV); | |
282 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); | |
283 | } | |
284 | tp->rcv_wnd = new_win; | |
285 | tp->rcv_wup = tp->rcv_nxt; | |
286 | ||
287 | /* Make sure we do not exceed the maximum possible | |
288 | * scaled window. | |
289 | */ | |
290 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) | |
291 | new_win = min(new_win, MAX_TCP_WINDOW); | |
292 | else | |
293 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
294 | ||
295 | /* RFC1323 scaling applied */ | |
296 | new_win >>= tp->rx_opt.rcv_wscale; | |
297 | ||
298 | /* If we advertise zero window, disable fast path. */ | |
299 | if (new_win == 0) { | |
300 | tp->pred_flags = 0; | |
301 | if (old_win) | |
302 | NET_INC_STATS(sock_net(sk), | |
303 | LINUX_MIB_TCPTOZEROWINDOWADV); | |
304 | } else if (old_win == 0) { | |
305 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); | |
306 | } | |
307 | ||
308 | return new_win; | |
309 | } | |
310 | ||
311 | /* Packet ECN state for a SYN-ACK */ | |
312 | static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) | |
313 | { | |
314 | const struct tcp_sock *tp = tcp_sk(sk); | |
315 | ||
316 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; | |
317 | if (!(tp->ecn_flags & TCP_ECN_OK)) | |
318 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; | |
319 | else if (tcp_ca_needs_ecn(sk) || | |
320 | tcp_bpf_ca_needs_ecn(sk)) | |
321 | INET_ECN_xmit(sk); | |
322 | } | |
323 | ||
324 | /* Packet ECN state for a SYN. */ | |
325 | static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) | |
326 | { | |
327 | struct tcp_sock *tp = tcp_sk(sk); | |
328 | bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); | |
329 | bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || | |
330 | tcp_ca_needs_ecn(sk) || bpf_needs_ecn; | |
331 | ||
332 | if (!use_ecn) { | |
333 | const struct dst_entry *dst = __sk_dst_get(sk); | |
334 | ||
335 | if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) | |
336 | use_ecn = true; | |
337 | } | |
338 | ||
339 | tp->ecn_flags = 0; | |
340 | ||
341 | if (use_ecn) { | |
342 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; | |
343 | tp->ecn_flags = TCP_ECN_OK; | |
344 | if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) | |
345 | INET_ECN_xmit(sk); | |
346 | } | |
347 | } | |
348 | ||
349 | static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) | |
350 | { | |
351 | if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) | |
352 | /* tp->ecn_flags are cleared at a later point in time when | |
353 | * SYN ACK is ultimatively being received. | |
354 | */ | |
355 | TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); | |
356 | } | |
357 | ||
358 | static void | |
359 | tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) | |
360 | { | |
361 | if (inet_rsk(req)->ecn_ok) | |
362 | th->ece = 1; | |
363 | } | |
364 | ||
365 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to | |
366 | * be sent. | |
367 | */ | |
368 | static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, | |
369 | struct tcphdr *th, int tcp_header_len) | |
370 | { | |
371 | struct tcp_sock *tp = tcp_sk(sk); | |
372 | ||
373 | if (tp->ecn_flags & TCP_ECN_OK) { | |
374 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
375 | if (skb->len != tcp_header_len && | |
376 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
377 | INET_ECN_xmit(sk); | |
378 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { | |
379 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; | |
380 | th->cwr = 1; | |
381 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; | |
382 | } | |
383 | } else if (!tcp_ca_needs_ecn(sk)) { | |
384 | /* ACK or retransmitted segment: clear ECT|CE */ | |
385 | INET_ECN_dontxmit(sk); | |
386 | } | |
387 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
388 | th->ece = 1; | |
389 | } | |
390 | } | |
391 | ||
392 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, | |
393 | * auto increment end seqno. | |
394 | */ | |
395 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) | |
396 | { | |
397 | skb->ip_summed = CHECKSUM_PARTIAL; | |
398 | skb->csum = 0; | |
399 | ||
400 | TCP_SKB_CB(skb)->tcp_flags = flags; | |
401 | TCP_SKB_CB(skb)->sacked = 0; | |
402 | ||
403 | tcp_skb_pcount_set(skb, 1); | |
404 | ||
405 | TCP_SKB_CB(skb)->seq = seq; | |
406 | if (flags & (TCPHDR_SYN | TCPHDR_FIN)) | |
407 | seq++; | |
408 | TCP_SKB_CB(skb)->end_seq = seq; | |
409 | } | |
410 | ||
411 | static inline bool tcp_urg_mode(const struct tcp_sock *tp) | |
412 | { | |
413 | return tp->snd_una != tp->snd_up; | |
414 | } | |
415 | ||
416 | #define OPTION_SACK_ADVERTISE (1 << 0) | |
417 | #define OPTION_TS (1 << 1) | |
418 | #define OPTION_MD5 (1 << 2) | |
419 | #define OPTION_WSCALE (1 << 3) | |
420 | #define OPTION_FAST_OPEN_COOKIE (1 << 8) | |
421 | ||
422 | struct tcp_out_options { | |
423 | u16 options; /* bit field of OPTION_* */ | |
424 | u16 mss; /* 0 to disable */ | |
425 | u8 ws; /* window scale, 0 to disable */ | |
426 | u8 num_sack_blocks; /* number of SACK blocks to include */ | |
427 | u8 hash_size; /* bytes in hash_location */ | |
428 | __u8 *hash_location; /* temporary pointer, overloaded */ | |
429 | __u32 tsval, tsecr; /* need to include OPTION_TS */ | |
430 | struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ | |
431 | }; | |
432 | ||
433 | /* Write previously computed TCP options to the packet. | |
434 | * | |
435 | * Beware: Something in the Internet is very sensitive to the ordering of | |
436 | * TCP options, we learned this through the hard way, so be careful here. | |
437 | * Luckily we can at least blame others for their non-compliance but from | |
438 | * inter-operability perspective it seems that we're somewhat stuck with | |
439 | * the ordering which we have been using if we want to keep working with | |
440 | * those broken things (not that it currently hurts anybody as there isn't | |
441 | * particular reason why the ordering would need to be changed). | |
442 | * | |
443 | * At least SACK_PERM as the first option is known to lead to a disaster | |
444 | * (but it may well be that other scenarios fail similarly). | |
445 | */ | |
446 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, | |
447 | struct tcp_out_options *opts) | |
448 | { | |
449 | u16 options = opts->options; /* mungable copy */ | |
450 | ||
451 | if (unlikely(OPTION_MD5 & options)) { | |
452 | *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | | |
453 | (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); | |
454 | /* overload cookie hash location */ | |
455 | opts->hash_location = (__u8 *)ptr; | |
456 | ptr += 4; | |
457 | } | |
458 | ||
459 | if (unlikely(opts->mss)) { | |
460 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
461 | (TCPOLEN_MSS << 16) | | |
462 | opts->mss); | |
463 | } | |
464 | ||
465 | if (likely(OPTION_TS & options)) { | |
466 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
467 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | | |
468 | (TCPOLEN_SACK_PERM << 16) | | |
469 | (TCPOPT_TIMESTAMP << 8) | | |
470 | TCPOLEN_TIMESTAMP); | |
471 | options &= ~OPTION_SACK_ADVERTISE; | |
472 | } else { | |
473 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
474 | (TCPOPT_NOP << 16) | | |
475 | (TCPOPT_TIMESTAMP << 8) | | |
476 | TCPOLEN_TIMESTAMP); | |
477 | } | |
478 | *ptr++ = htonl(opts->tsval); | |
479 | *ptr++ = htonl(opts->tsecr); | |
480 | } | |
481 | ||
482 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
483 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
484 | (TCPOPT_NOP << 16) | | |
485 | (TCPOPT_SACK_PERM << 8) | | |
486 | TCPOLEN_SACK_PERM); | |
487 | } | |
488 | ||
489 | if (unlikely(OPTION_WSCALE & options)) { | |
490 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
491 | (TCPOPT_WINDOW << 16) | | |
492 | (TCPOLEN_WINDOW << 8) | | |
493 | opts->ws); | |
494 | } | |
495 | ||
496 | if (unlikely(opts->num_sack_blocks)) { | |
497 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
498 | tp->duplicate_sack : tp->selective_acks; | |
499 | int this_sack; | |
500 | ||
501 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
502 | (TCPOPT_NOP << 16) | | |
503 | (TCPOPT_SACK << 8) | | |
504 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * | |
505 | TCPOLEN_SACK_PERBLOCK))); | |
506 | ||
507 | for (this_sack = 0; this_sack < opts->num_sack_blocks; | |
508 | ++this_sack) { | |
509 | *ptr++ = htonl(sp[this_sack].start_seq); | |
510 | *ptr++ = htonl(sp[this_sack].end_seq); | |
511 | } | |
512 | ||
513 | tp->rx_opt.dsack = 0; | |
514 | } | |
515 | ||
516 | if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { | |
517 | struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; | |
518 | u8 *p = (u8 *)ptr; | |
519 | u32 len; /* Fast Open option length */ | |
520 | ||
521 | if (foc->exp) { | |
522 | len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; | |
523 | *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | | |
524 | TCPOPT_FASTOPEN_MAGIC); | |
525 | p += TCPOLEN_EXP_FASTOPEN_BASE; | |
526 | } else { | |
527 | len = TCPOLEN_FASTOPEN_BASE + foc->len; | |
528 | *p++ = TCPOPT_FASTOPEN; | |
529 | *p++ = len; | |
530 | } | |
531 | ||
532 | memcpy(p, foc->val, foc->len); | |
533 | if ((len & 3) == 2) { | |
534 | p[foc->len] = TCPOPT_NOP; | |
535 | p[foc->len + 1] = TCPOPT_NOP; | |
536 | } | |
537 | ptr += (len + 3) >> 2; | |
538 | } | |
539 | } | |
540 | ||
541 | /* Compute TCP options for SYN packets. This is not the final | |
542 | * network wire format yet. | |
543 | */ | |
544 | static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, | |
545 | struct tcp_out_options *opts, | |
546 | struct tcp_md5sig_key **md5) | |
547 | { | |
548 | struct tcp_sock *tp = tcp_sk(sk); | |
549 | unsigned int remaining = MAX_TCP_OPTION_SPACE; | |
550 | struct tcp_fastopen_request *fastopen = tp->fastopen_req; | |
551 | ||
552 | #ifdef CONFIG_TCP_MD5SIG | |
553 | *md5 = tp->af_specific->md5_lookup(sk, sk); | |
554 | if (*md5) { | |
555 | opts->options |= OPTION_MD5; | |
556 | remaining -= TCPOLEN_MD5SIG_ALIGNED; | |
557 | } | |
558 | #else | |
559 | *md5 = NULL; | |
560 | #endif | |
561 | ||
562 | /* We always get an MSS option. The option bytes which will be seen in | |
563 | * normal data packets should timestamps be used, must be in the MSS | |
564 | * advertised. But we subtract them from tp->mss_cache so that | |
565 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
566 | * fact here if necessary. If we don't do this correctly, as a | |
567 | * receiver we won't recognize data packets as being full sized when we | |
568 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
569 | * SACKs don't matter, we never delay an ACK when we have any of those | |
570 | * going out. */ | |
571 | opts->mss = tcp_advertise_mss(sk); | |
572 | remaining -= TCPOLEN_MSS_ALIGNED; | |
573 | ||
574 | if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { | |
575 | opts->options |= OPTION_TS; | |
576 | opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; | |
577 | opts->tsecr = tp->rx_opt.ts_recent; | |
578 | remaining -= TCPOLEN_TSTAMP_ALIGNED; | |
579 | } | |
580 | if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { | |
581 | opts->ws = tp->rx_opt.rcv_wscale; | |
582 | opts->options |= OPTION_WSCALE; | |
583 | remaining -= TCPOLEN_WSCALE_ALIGNED; | |
584 | } | |
585 | if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { | |
586 | opts->options |= OPTION_SACK_ADVERTISE; | |
587 | if (unlikely(!(OPTION_TS & opts->options))) | |
588 | remaining -= TCPOLEN_SACKPERM_ALIGNED; | |
589 | } | |
590 | ||
591 | if (fastopen && fastopen->cookie.len >= 0) { | |
592 | u32 need = fastopen->cookie.len; | |
593 | ||
594 | need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : | |
595 | TCPOLEN_FASTOPEN_BASE; | |
596 | need = (need + 3) & ~3U; /* Align to 32 bits */ | |
597 | if (remaining >= need) { | |
598 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
599 | opts->fastopen_cookie = &fastopen->cookie; | |
600 | remaining -= need; | |
601 | tp->syn_fastopen = 1; | |
602 | tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; | |
603 | } | |
604 | } | |
605 | ||
606 | return MAX_TCP_OPTION_SPACE - remaining; | |
607 | } | |
608 | ||
609 | /* Set up TCP options for SYN-ACKs. */ | |
610 | static unsigned int tcp_synack_options(struct request_sock *req, | |
611 | unsigned int mss, struct sk_buff *skb, | |
612 | struct tcp_out_options *opts, | |
613 | const struct tcp_md5sig_key *md5, | |
614 | struct tcp_fastopen_cookie *foc) | |
615 | { | |
616 | struct inet_request_sock *ireq = inet_rsk(req); | |
617 | unsigned int remaining = MAX_TCP_OPTION_SPACE; | |
618 | ||
619 | #ifdef CONFIG_TCP_MD5SIG | |
620 | if (md5) { | |
621 | opts->options |= OPTION_MD5; | |
622 | remaining -= TCPOLEN_MD5SIG_ALIGNED; | |
623 | ||
624 | /* We can't fit any SACK blocks in a packet with MD5 + TS | |
625 | * options. There was discussion about disabling SACK | |
626 | * rather than TS in order to fit in better with old, | |
627 | * buggy kernels, but that was deemed to be unnecessary. | |
628 | */ | |
629 | ireq->tstamp_ok &= !ireq->sack_ok; | |
630 | } | |
631 | #endif | |
632 | ||
633 | /* We always send an MSS option. */ | |
634 | opts->mss = mss; | |
635 | remaining -= TCPOLEN_MSS_ALIGNED; | |
636 | ||
637 | if (likely(ireq->wscale_ok)) { | |
638 | opts->ws = ireq->rcv_wscale; | |
639 | opts->options |= OPTION_WSCALE; | |
640 | remaining -= TCPOLEN_WSCALE_ALIGNED; | |
641 | } | |
642 | if (likely(ireq->tstamp_ok)) { | |
643 | opts->options |= OPTION_TS; | |
644 | opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; | |
645 | opts->tsecr = req->ts_recent; | |
646 | remaining -= TCPOLEN_TSTAMP_ALIGNED; | |
647 | } | |
648 | if (likely(ireq->sack_ok)) { | |
649 | opts->options |= OPTION_SACK_ADVERTISE; | |
650 | if (unlikely(!ireq->tstamp_ok)) | |
651 | remaining -= TCPOLEN_SACKPERM_ALIGNED; | |
652 | } | |
653 | if (foc != NULL && foc->len >= 0) { | |
654 | u32 need = foc->len; | |
655 | ||
656 | need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : | |
657 | TCPOLEN_FASTOPEN_BASE; | |
658 | need = (need + 3) & ~3U; /* Align to 32 bits */ | |
659 | if (remaining >= need) { | |
660 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
661 | opts->fastopen_cookie = foc; | |
662 | remaining -= need; | |
663 | } | |
664 | } | |
665 | ||
666 | return MAX_TCP_OPTION_SPACE - remaining; | |
667 | } | |
668 | ||
669 | /* Compute TCP options for ESTABLISHED sockets. This is not the | |
670 | * final wire format yet. | |
671 | */ | |
672 | static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, | |
673 | struct tcp_out_options *opts, | |
674 | struct tcp_md5sig_key **md5) | |
675 | { | |
676 | struct tcp_sock *tp = tcp_sk(sk); | |
677 | unsigned int size = 0; | |
678 | unsigned int eff_sacks; | |
679 | ||
680 | opts->options = 0; | |
681 | ||
682 | #ifdef CONFIG_TCP_MD5SIG | |
683 | *md5 = tp->af_specific->md5_lookup(sk, sk); | |
684 | if (unlikely(*md5)) { | |
685 | opts->options |= OPTION_MD5; | |
686 | size += TCPOLEN_MD5SIG_ALIGNED; | |
687 | } | |
688 | #else | |
689 | *md5 = NULL; | |
690 | #endif | |
691 | ||
692 | if (likely(tp->rx_opt.tstamp_ok)) { | |
693 | opts->options |= OPTION_TS; | |
694 | opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; | |
695 | opts->tsecr = tp->rx_opt.ts_recent; | |
696 | size += TCPOLEN_TSTAMP_ALIGNED; | |
697 | } | |
698 | ||
699 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; | |
700 | if (unlikely(eff_sacks)) { | |
701 | const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; | |
702 | opts->num_sack_blocks = | |
703 | min_t(unsigned int, eff_sacks, | |
704 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / | |
705 | TCPOLEN_SACK_PERBLOCK); | |
706 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
707 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
708 | } | |
709 | ||
710 | return size; | |
711 | } | |
712 | ||
713 | ||
714 | /* TCP SMALL QUEUES (TSQ) | |
715 | * | |
716 | * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) | |
717 | * to reduce RTT and bufferbloat. | |
718 | * We do this using a special skb destructor (tcp_wfree). | |
719 | * | |
720 | * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb | |
721 | * needs to be reallocated in a driver. | |
722 | * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc | |
723 | * | |
724 | * Since transmit from skb destructor is forbidden, we use a tasklet | |
725 | * to process all sockets that eventually need to send more skbs. | |
726 | * We use one tasklet per cpu, with its own queue of sockets. | |
727 | */ | |
728 | struct tsq_tasklet { | |
729 | struct tasklet_struct tasklet; | |
730 | struct list_head head; /* queue of tcp sockets */ | |
731 | }; | |
732 | static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); | |
733 | ||
734 | static void tcp_tsq_handler(struct sock *sk) | |
735 | { | |
736 | if ((1 << sk->sk_state) & | |
737 | (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | | |
738 | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { | |
739 | struct tcp_sock *tp = tcp_sk(sk); | |
740 | ||
741 | if (tp->lost_out > tp->retrans_out && | |
742 | tp->snd_cwnd > tcp_packets_in_flight(tp)) | |
743 | tcp_xmit_retransmit_queue(sk); | |
744 | ||
745 | tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, | |
746 | 0, GFP_ATOMIC); | |
747 | } | |
748 | } | |
749 | /* | |
750 | * One tasklet per cpu tries to send more skbs. | |
751 | * We run in tasklet context but need to disable irqs when | |
752 | * transferring tsq->head because tcp_wfree() might | |
753 | * interrupt us (non NAPI drivers) | |
754 | */ | |
755 | static void tcp_tasklet_func(unsigned long data) | |
756 | { | |
757 | struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; | |
758 | LIST_HEAD(list); | |
759 | unsigned long flags; | |
760 | struct list_head *q, *n; | |
761 | struct tcp_sock *tp; | |
762 | struct sock *sk; | |
763 | ||
764 | local_irq_save(flags); | |
765 | list_splice_init(&tsq->head, &list); | |
766 | local_irq_restore(flags); | |
767 | ||
768 | list_for_each_safe(q, n, &list) { | |
769 | tp = list_entry(q, struct tcp_sock, tsq_node); | |
770 | list_del(&tp->tsq_node); | |
771 | ||
772 | sk = (struct sock *)tp; | |
773 | smp_mb__before_atomic(); | |
774 | clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); | |
775 | ||
776 | if (!sk->sk_lock.owned && | |
777 | test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) { | |
778 | bh_lock_sock(sk); | |
779 | if (!sock_owned_by_user(sk)) { | |
780 | clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); | |
781 | tcp_tsq_handler(sk); | |
782 | } | |
783 | bh_unlock_sock(sk); | |
784 | } | |
785 | ||
786 | sk_free(sk); | |
787 | } | |
788 | } | |
789 | ||
790 | #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ | |
791 | TCPF_WRITE_TIMER_DEFERRED | \ | |
792 | TCPF_DELACK_TIMER_DEFERRED | \ | |
793 | TCPF_MTU_REDUCED_DEFERRED) | |
794 | /** | |
795 | * tcp_release_cb - tcp release_sock() callback | |
796 | * @sk: socket | |
797 | * | |
798 | * called from release_sock() to perform protocol dependent | |
799 | * actions before socket release. | |
800 | */ | |
801 | void tcp_release_cb(struct sock *sk) | |
802 | { | |
803 | unsigned long flags, nflags; | |
804 | ||
805 | /* perform an atomic operation only if at least one flag is set */ | |
806 | do { | |
807 | flags = sk->sk_tsq_flags; | |
808 | if (!(flags & TCP_DEFERRED_ALL)) | |
809 | return; | |
810 | nflags = flags & ~TCP_DEFERRED_ALL; | |
811 | } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); | |
812 | ||
813 | if (flags & TCPF_TSQ_DEFERRED) | |
814 | tcp_tsq_handler(sk); | |
815 | ||
816 | /* Here begins the tricky part : | |
817 | * We are called from release_sock() with : | |
818 | * 1) BH disabled | |
819 | * 2) sk_lock.slock spinlock held | |
820 | * 3) socket owned by us (sk->sk_lock.owned == 1) | |
821 | * | |
822 | * But following code is meant to be called from BH handlers, | |
823 | * so we should keep BH disabled, but early release socket ownership | |
824 | */ | |
825 | sock_release_ownership(sk); | |
826 | ||
827 | if (flags & TCPF_WRITE_TIMER_DEFERRED) { | |
828 | tcp_write_timer_handler(sk); | |
829 | __sock_put(sk); | |
830 | } | |
831 | if (flags & TCPF_DELACK_TIMER_DEFERRED) { | |
832 | tcp_delack_timer_handler(sk); | |
833 | __sock_put(sk); | |
834 | } | |
835 | if (flags & TCPF_MTU_REDUCED_DEFERRED) { | |
836 | inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); | |
837 | __sock_put(sk); | |
838 | } | |
839 | } | |
840 | EXPORT_SYMBOL(tcp_release_cb); | |
841 | ||
842 | void __init tcp_tasklet_init(void) | |
843 | { | |
844 | int i; | |
845 | ||
846 | for_each_possible_cpu(i) { | |
847 | struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); | |
848 | ||
849 | INIT_LIST_HEAD(&tsq->head); | |
850 | tasklet_init(&tsq->tasklet, | |
851 | tcp_tasklet_func, | |
852 | (unsigned long)tsq); | |
853 | } | |
854 | } | |
855 | ||
856 | /* | |
857 | * Write buffer destructor automatically called from kfree_skb. | |
858 | * We can't xmit new skbs from this context, as we might already | |
859 | * hold qdisc lock. | |
860 | */ | |
861 | void tcp_wfree(struct sk_buff *skb) | |
862 | { | |
863 | struct sock *sk = skb->sk; | |
864 | struct tcp_sock *tp = tcp_sk(sk); | |
865 | unsigned long flags, nval, oval; | |
866 | ||
867 | /* Keep one reference on sk_wmem_alloc. | |
868 | * Will be released by sk_free() from here or tcp_tasklet_func() | |
869 | */ | |
870 | WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); | |
871 | ||
872 | /* If this softirq is serviced by ksoftirqd, we are likely under stress. | |
873 | * Wait until our queues (qdisc + devices) are drained. | |
874 | * This gives : | |
875 | * - less callbacks to tcp_write_xmit(), reducing stress (batches) | |
876 | * - chance for incoming ACK (processed by another cpu maybe) | |
877 | * to migrate this flow (skb->ooo_okay will be eventually set) | |
878 | */ | |
879 | if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) | |
880 | goto out; | |
881 | ||
882 | for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { | |
883 | struct tsq_tasklet *tsq; | |
884 | bool empty; | |
885 | ||
886 | if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) | |
887 | goto out; | |
888 | ||
889 | nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; | |
890 | nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); | |
891 | if (nval != oval) | |
892 | continue; | |
893 | ||
894 | /* queue this socket to tasklet queue */ | |
895 | local_irq_save(flags); | |
896 | tsq = this_cpu_ptr(&tsq_tasklet); | |
897 | empty = list_empty(&tsq->head); | |
898 | list_add(&tp->tsq_node, &tsq->head); | |
899 | if (empty) | |
900 | tasklet_schedule(&tsq->tasklet); | |
901 | local_irq_restore(flags); | |
902 | return; | |
903 | } | |
904 | out: | |
905 | sk_free(sk); | |
906 | } | |
907 | ||
908 | /* Note: Called under hard irq. | |
909 | * We can not call TCP stack right away. | |
910 | */ | |
911 | enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) | |
912 | { | |
913 | struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); | |
914 | struct sock *sk = (struct sock *)tp; | |
915 | unsigned long nval, oval; | |
916 | ||
917 | for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { | |
918 | struct tsq_tasklet *tsq; | |
919 | bool empty; | |
920 | ||
921 | if (oval & TSQF_QUEUED) | |
922 | break; | |
923 | ||
924 | nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; | |
925 | nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); | |
926 | if (nval != oval) | |
927 | continue; | |
928 | ||
929 | if (!refcount_inc_not_zero(&sk->sk_wmem_alloc)) | |
930 | break; | |
931 | /* queue this socket to tasklet queue */ | |
932 | tsq = this_cpu_ptr(&tsq_tasklet); | |
933 | empty = list_empty(&tsq->head); | |
934 | list_add(&tp->tsq_node, &tsq->head); | |
935 | if (empty) | |
936 | tasklet_schedule(&tsq->tasklet); | |
937 | break; | |
938 | } | |
939 | return HRTIMER_NORESTART; | |
940 | } | |
941 | ||
942 | /* BBR congestion control needs pacing. | |
943 | * Same remark for SO_MAX_PACING_RATE. | |
944 | * sch_fq packet scheduler is efficiently handling pacing, | |
945 | * but is not always installed/used. | |
946 | * Return true if TCP stack should pace packets itself. | |
947 | */ | |
948 | static bool tcp_needs_internal_pacing(const struct sock *sk) | |
949 | { | |
950 | return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED; | |
951 | } | |
952 | ||
953 | static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb) | |
954 | { | |
955 | u64 len_ns; | |
956 | u32 rate; | |
957 | ||
958 | if (!tcp_needs_internal_pacing(sk)) | |
959 | return; | |
960 | rate = sk->sk_pacing_rate; | |
961 | if (!rate || rate == ~0U) | |
962 | return; | |
963 | ||
964 | /* Should account for header sizes as sch_fq does, | |
965 | * but lets make things simple. | |
966 | */ | |
967 | len_ns = (u64)skb->len * NSEC_PER_SEC; | |
968 | do_div(len_ns, rate); | |
969 | hrtimer_start(&tcp_sk(sk)->pacing_timer, | |
970 | ktime_add_ns(ktime_get(), len_ns), | |
971 | HRTIMER_MODE_ABS_PINNED); | |
972 | } | |
973 | ||
974 | /* This routine actually transmits TCP packets queued in by | |
975 | * tcp_do_sendmsg(). This is used by both the initial | |
976 | * transmission and possible later retransmissions. | |
977 | * All SKB's seen here are completely headerless. It is our | |
978 | * job to build the TCP header, and pass the packet down to | |
979 | * IP so it can do the same plus pass the packet off to the | |
980 | * device. | |
981 | * | |
982 | * We are working here with either a clone of the original | |
983 | * SKB, or a fresh unique copy made by the retransmit engine. | |
984 | */ | |
985 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, | |
986 | gfp_t gfp_mask) | |
987 | { | |
988 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
989 | struct inet_sock *inet; | |
990 | struct tcp_sock *tp; | |
991 | struct tcp_skb_cb *tcb; | |
992 | struct tcp_out_options opts; | |
993 | unsigned int tcp_options_size, tcp_header_size; | |
994 | struct tcp_md5sig_key *md5; | |
995 | struct tcphdr *th; | |
996 | int err; | |
997 | ||
998 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
999 | tp = tcp_sk(sk); | |
1000 | ||
1001 | skb->skb_mstamp = tp->tcp_mstamp; | |
1002 | if (clone_it) { | |
1003 | TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq | |
1004 | - tp->snd_una; | |
1005 | tcp_rate_skb_sent(sk, skb); | |
1006 | ||
1007 | if (unlikely(skb_cloned(skb))) | |
1008 | skb = pskb_copy(skb, gfp_mask); | |
1009 | else | |
1010 | skb = skb_clone(skb, gfp_mask); | |
1011 | if (unlikely(!skb)) | |
1012 | return -ENOBUFS; | |
1013 | } | |
1014 | ||
1015 | inet = inet_sk(sk); | |
1016 | tcb = TCP_SKB_CB(skb); | |
1017 | memset(&opts, 0, sizeof(opts)); | |
1018 | ||
1019 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) | |
1020 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); | |
1021 | else | |
1022 | tcp_options_size = tcp_established_options(sk, skb, &opts, | |
1023 | &md5); | |
1024 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); | |
1025 | ||
1026 | /* if no packet is in qdisc/device queue, then allow XPS to select | |
1027 | * another queue. We can be called from tcp_tsq_handler() | |
1028 | * which holds one reference to sk_wmem_alloc. | |
1029 | * | |
1030 | * TODO: Ideally, in-flight pure ACK packets should not matter here. | |
1031 | * One way to get this would be to set skb->truesize = 2 on them. | |
1032 | */ | |
1033 | skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); | |
1034 | ||
1035 | /* If we had to use memory reserve to allocate this skb, | |
1036 | * this might cause drops if packet is looped back : | |
1037 | * Other socket might not have SOCK_MEMALLOC. | |
1038 | * Packets not looped back do not care about pfmemalloc. | |
1039 | */ | |
1040 | skb->pfmemalloc = 0; | |
1041 | ||
1042 | skb_push(skb, tcp_header_size); | |
1043 | skb_reset_transport_header(skb); | |
1044 | ||
1045 | skb_orphan(skb); | |
1046 | skb->sk = sk; | |
1047 | skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; | |
1048 | skb_set_hash_from_sk(skb, sk); | |
1049 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); | |
1050 | ||
1051 | skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); | |
1052 | ||
1053 | /* Build TCP header and checksum it. */ | |
1054 | th = (struct tcphdr *)skb->data; | |
1055 | th->source = inet->inet_sport; | |
1056 | th->dest = inet->inet_dport; | |
1057 | th->seq = htonl(tcb->seq); | |
1058 | th->ack_seq = htonl(tp->rcv_nxt); | |
1059 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | | |
1060 | tcb->tcp_flags); | |
1061 | ||
1062 | th->check = 0; | |
1063 | th->urg_ptr = 0; | |
1064 | ||
1065 | /* The urg_mode check is necessary during a below snd_una win probe */ | |
1066 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { | |
1067 | if (before(tp->snd_up, tcb->seq + 0x10000)) { | |
1068 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
1069 | th->urg = 1; | |
1070 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { | |
1071 | th->urg_ptr = htons(0xFFFF); | |
1072 | th->urg = 1; | |
1073 | } | |
1074 | } | |
1075 | ||
1076 | tcp_options_write((__be32 *)(th + 1), tp, &opts); | |
1077 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; | |
1078 | if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { | |
1079 | th->window = htons(tcp_select_window(sk)); | |
1080 | tcp_ecn_send(sk, skb, th, tcp_header_size); | |
1081 | } else { | |
1082 | /* RFC1323: The window in SYN & SYN/ACK segments | |
1083 | * is never scaled. | |
1084 | */ | |
1085 | th->window = htons(min(tp->rcv_wnd, 65535U)); | |
1086 | } | |
1087 | #ifdef CONFIG_TCP_MD5SIG | |
1088 | /* Calculate the MD5 hash, as we have all we need now */ | |
1089 | if (md5) { | |
1090 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); | |
1091 | tp->af_specific->calc_md5_hash(opts.hash_location, | |
1092 | md5, sk, skb); | |
1093 | } | |
1094 | #endif | |
1095 | ||
1096 | icsk->icsk_af_ops->send_check(sk, skb); | |
1097 | ||
1098 | if (likely(tcb->tcp_flags & TCPHDR_ACK)) | |
1099 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); | |
1100 | ||
1101 | if (skb->len != tcp_header_size) { | |
1102 | tcp_event_data_sent(tp, sk); | |
1103 | tp->data_segs_out += tcp_skb_pcount(skb); | |
1104 | tcp_internal_pacing(sk, skb); | |
1105 | } | |
1106 | ||
1107 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) | |
1108 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, | |
1109 | tcp_skb_pcount(skb)); | |
1110 | ||
1111 | tp->segs_out += tcp_skb_pcount(skb); | |
1112 | /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ | |
1113 | skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); | |
1114 | skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); | |
1115 | ||
1116 | /* Our usage of tstamp should remain private */ | |
1117 | skb->tstamp = 0; | |
1118 | ||
1119 | /* Cleanup our debris for IP stacks */ | |
1120 | memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), | |
1121 | sizeof(struct inet6_skb_parm))); | |
1122 | ||
1123 | err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); | |
1124 | ||
1125 | if (likely(err <= 0)) | |
1126 | return err; | |
1127 | ||
1128 | tcp_enter_cwr(sk); | |
1129 | ||
1130 | return net_xmit_eval(err); | |
1131 | } | |
1132 | ||
1133 | /* This routine just queues the buffer for sending. | |
1134 | * | |
1135 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
1136 | * otherwise socket can stall. | |
1137 | */ | |
1138 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
1139 | { | |
1140 | struct tcp_sock *tp = tcp_sk(sk); | |
1141 | ||
1142 | /* Advance write_seq and place onto the write_queue. */ | |
1143 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
1144 | __skb_header_release(skb); | |
1145 | tcp_add_write_queue_tail(sk, skb); | |
1146 | sk->sk_wmem_queued += skb->truesize; | |
1147 | sk_mem_charge(sk, skb->truesize); | |
1148 | } | |
1149 | ||
1150 | /* Initialize TSO segments for a packet. */ | |
1151 | static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) | |
1152 | { | |
1153 | if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { | |
1154 | /* Avoid the costly divide in the normal | |
1155 | * non-TSO case. | |
1156 | */ | |
1157 | tcp_skb_pcount_set(skb, 1); | |
1158 | TCP_SKB_CB(skb)->tcp_gso_size = 0; | |
1159 | } else { | |
1160 | tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); | |
1161 | TCP_SKB_CB(skb)->tcp_gso_size = mss_now; | |
1162 | } | |
1163 | } | |
1164 | ||
1165 | /* When a modification to fackets out becomes necessary, we need to check | |
1166 | * skb is counted to fackets_out or not. | |
1167 | */ | |
1168 | static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, | |
1169 | int decr) | |
1170 | { | |
1171 | struct tcp_sock *tp = tcp_sk(sk); | |
1172 | ||
1173 | if (!tp->sacked_out || tcp_is_reno(tp)) | |
1174 | return; | |
1175 | ||
1176 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) | |
1177 | tp->fackets_out -= decr; | |
1178 | } | |
1179 | ||
1180 | /* Pcount in the middle of the write queue got changed, we need to do various | |
1181 | * tweaks to fix counters | |
1182 | */ | |
1183 | static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) | |
1184 | { | |
1185 | struct tcp_sock *tp = tcp_sk(sk); | |
1186 | ||
1187 | tp->packets_out -= decr; | |
1188 | ||
1189 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
1190 | tp->sacked_out -= decr; | |
1191 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
1192 | tp->retrans_out -= decr; | |
1193 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) | |
1194 | tp->lost_out -= decr; | |
1195 | ||
1196 | /* Reno case is special. Sigh... */ | |
1197 | if (tcp_is_reno(tp) && decr > 0) | |
1198 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); | |
1199 | ||
1200 | tcp_adjust_fackets_out(sk, skb, decr); | |
1201 | ||
1202 | if (tp->lost_skb_hint && | |
1203 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && | |
1204 | (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) | |
1205 | tp->lost_cnt_hint -= decr; | |
1206 | ||
1207 | tcp_verify_left_out(tp); | |
1208 | } | |
1209 | ||
1210 | static bool tcp_has_tx_tstamp(const struct sk_buff *skb) | |
1211 | { | |
1212 | return TCP_SKB_CB(skb)->txstamp_ack || | |
1213 | (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); | |
1214 | } | |
1215 | ||
1216 | static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) | |
1217 | { | |
1218 | struct skb_shared_info *shinfo = skb_shinfo(skb); | |
1219 | ||
1220 | if (unlikely(tcp_has_tx_tstamp(skb)) && | |
1221 | !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { | |
1222 | struct skb_shared_info *shinfo2 = skb_shinfo(skb2); | |
1223 | u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; | |
1224 | ||
1225 | shinfo->tx_flags &= ~tsflags; | |
1226 | shinfo2->tx_flags |= tsflags; | |
1227 | swap(shinfo->tskey, shinfo2->tskey); | |
1228 | TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; | |
1229 | TCP_SKB_CB(skb)->txstamp_ack = 0; | |
1230 | } | |
1231 | } | |
1232 | ||
1233 | static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) | |
1234 | { | |
1235 | TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; | |
1236 | TCP_SKB_CB(skb)->eor = 0; | |
1237 | } | |
1238 | ||
1239 | /* Function to create two new TCP segments. Shrinks the given segment | |
1240 | * to the specified size and appends a new segment with the rest of the | |
1241 | * packet to the list. This won't be called frequently, I hope. | |
1242 | * Remember, these are still headerless SKBs at this point. | |
1243 | */ | |
1244 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, | |
1245 | unsigned int mss_now, gfp_t gfp) | |
1246 | { | |
1247 | struct tcp_sock *tp = tcp_sk(sk); | |
1248 | struct sk_buff *buff; | |
1249 | int nsize, old_factor; | |
1250 | int nlen; | |
1251 | u8 flags; | |
1252 | ||
1253 | if (WARN_ON(len > skb->len)) | |
1254 | return -EINVAL; | |
1255 | ||
1256 | nsize = skb_headlen(skb) - len; | |
1257 | if (nsize < 0) | |
1258 | nsize = 0; | |
1259 | ||
1260 | if (skb_unclone(skb, gfp)) | |
1261 | return -ENOMEM; | |
1262 | ||
1263 | /* Get a new skb... force flag on. */ | |
1264 | buff = sk_stream_alloc_skb(sk, nsize, gfp, true); | |
1265 | if (!buff) | |
1266 | return -ENOMEM; /* We'll just try again later. */ | |
1267 | ||
1268 | sk->sk_wmem_queued += buff->truesize; | |
1269 | sk_mem_charge(sk, buff->truesize); | |
1270 | nlen = skb->len - len - nsize; | |
1271 | buff->truesize += nlen; | |
1272 | skb->truesize -= nlen; | |
1273 | ||
1274 | /* Correct the sequence numbers. */ | |
1275 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1276 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1277 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1278 | ||
1279 | /* PSH and FIN should only be set in the second packet. */ | |
1280 | flags = TCP_SKB_CB(skb)->tcp_flags; | |
1281 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1282 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
1283 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; | |
1284 | tcp_skb_fragment_eor(skb, buff); | |
1285 | ||
1286 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { | |
1287 | /* Copy and checksum data tail into the new buffer. */ | |
1288 | buff->csum = csum_partial_copy_nocheck(skb->data + len, | |
1289 | skb_put(buff, nsize), | |
1290 | nsize, 0); | |
1291 | ||
1292 | skb_trim(skb, len); | |
1293 | ||
1294 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
1295 | } else { | |
1296 | skb->ip_summed = CHECKSUM_PARTIAL; | |
1297 | skb_split(skb, buff, len); | |
1298 | } | |
1299 | ||
1300 | buff->ip_summed = skb->ip_summed; | |
1301 | ||
1302 | buff->tstamp = skb->tstamp; | |
1303 | tcp_fragment_tstamp(skb, buff); | |
1304 | ||
1305 | old_factor = tcp_skb_pcount(skb); | |
1306 | ||
1307 | /* Fix up tso_factor for both original and new SKB. */ | |
1308 | tcp_set_skb_tso_segs(skb, mss_now); | |
1309 | tcp_set_skb_tso_segs(buff, mss_now); | |
1310 | ||
1311 | /* Update delivered info for the new segment */ | |
1312 | TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; | |
1313 | ||
1314 | /* If this packet has been sent out already, we must | |
1315 | * adjust the various packet counters. | |
1316 | */ | |
1317 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { | |
1318 | int diff = old_factor - tcp_skb_pcount(skb) - | |
1319 | tcp_skb_pcount(buff); | |
1320 | ||
1321 | if (diff) | |
1322 | tcp_adjust_pcount(sk, skb, diff); | |
1323 | } | |
1324 | ||
1325 | /* Link BUFF into the send queue. */ | |
1326 | __skb_header_release(buff); | |
1327 | tcp_insert_write_queue_after(skb, buff, sk); | |
1328 | ||
1329 | return 0; | |
1330 | } | |
1331 | ||
1332 | /* This is similar to __pskb_pull_tail(). The difference is that pulled | |
1333 | * data is not copied, but immediately discarded. | |
1334 | */ | |
1335 | static int __pskb_trim_head(struct sk_buff *skb, int len) | |
1336 | { | |
1337 | struct skb_shared_info *shinfo; | |
1338 | int i, k, eat; | |
1339 | ||
1340 | eat = min_t(int, len, skb_headlen(skb)); | |
1341 | if (eat) { | |
1342 | __skb_pull(skb, eat); | |
1343 | len -= eat; | |
1344 | if (!len) | |
1345 | return 0; | |
1346 | } | |
1347 | eat = len; | |
1348 | k = 0; | |
1349 | shinfo = skb_shinfo(skb); | |
1350 | for (i = 0; i < shinfo->nr_frags; i++) { | |
1351 | int size = skb_frag_size(&shinfo->frags[i]); | |
1352 | ||
1353 | if (size <= eat) { | |
1354 | skb_frag_unref(skb, i); | |
1355 | eat -= size; | |
1356 | } else { | |
1357 | shinfo->frags[k] = shinfo->frags[i]; | |
1358 | if (eat) { | |
1359 | shinfo->frags[k].page_offset += eat; | |
1360 | skb_frag_size_sub(&shinfo->frags[k], eat); | |
1361 | eat = 0; | |
1362 | } | |
1363 | k++; | |
1364 | } | |
1365 | } | |
1366 | shinfo->nr_frags = k; | |
1367 | ||
1368 | skb->data_len -= len; | |
1369 | skb->len = skb->data_len; | |
1370 | return len; | |
1371 | } | |
1372 | ||
1373 | /* Remove acked data from a packet in the transmit queue. */ | |
1374 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) | |
1375 | { | |
1376 | u32 delta_truesize; | |
1377 | ||
1378 | if (skb_unclone(skb, GFP_ATOMIC)) | |
1379 | return -ENOMEM; | |
1380 | ||
1381 | delta_truesize = __pskb_trim_head(skb, len); | |
1382 | ||
1383 | TCP_SKB_CB(skb)->seq += len; | |
1384 | skb->ip_summed = CHECKSUM_PARTIAL; | |
1385 | ||
1386 | if (delta_truesize) { | |
1387 | skb->truesize -= delta_truesize; | |
1388 | sk->sk_wmem_queued -= delta_truesize; | |
1389 | sk_mem_uncharge(sk, delta_truesize); | |
1390 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); | |
1391 | } | |
1392 | ||
1393 | /* Any change of skb->len requires recalculation of tso factor. */ | |
1394 | if (tcp_skb_pcount(skb) > 1) | |
1395 | tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); | |
1396 | ||
1397 | return 0; | |
1398 | } | |
1399 | ||
1400 | /* Calculate MSS not accounting any TCP options. */ | |
1401 | static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
1402 | { | |
1403 | const struct tcp_sock *tp = tcp_sk(sk); | |
1404 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
1405 | int mss_now; | |
1406 | ||
1407 | /* Calculate base mss without TCP options: | |
1408 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
1409 | */ | |
1410 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
1411 | ||
1412 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ | |
1413 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1414 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1415 | ||
1416 | if (dst && dst_allfrag(dst)) | |
1417 | mss_now -= icsk->icsk_af_ops->net_frag_header_len; | |
1418 | } | |
1419 | ||
1420 | /* Clamp it (mss_clamp does not include tcp options) */ | |
1421 | if (mss_now > tp->rx_opt.mss_clamp) | |
1422 | mss_now = tp->rx_opt.mss_clamp; | |
1423 | ||
1424 | /* Now subtract optional transport overhead */ | |
1425 | mss_now -= icsk->icsk_ext_hdr_len; | |
1426 | ||
1427 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
1428 | if (mss_now < 48) | |
1429 | mss_now = 48; | |
1430 | return mss_now; | |
1431 | } | |
1432 | ||
1433 | /* Calculate MSS. Not accounting for SACKs here. */ | |
1434 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
1435 | { | |
1436 | /* Subtract TCP options size, not including SACKs */ | |
1437 | return __tcp_mtu_to_mss(sk, pmtu) - | |
1438 | (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); | |
1439 | } | |
1440 | ||
1441 | /* Inverse of above */ | |
1442 | int tcp_mss_to_mtu(struct sock *sk, int mss) | |
1443 | { | |
1444 | const struct tcp_sock *tp = tcp_sk(sk); | |
1445 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
1446 | int mtu; | |
1447 | ||
1448 | mtu = mss + | |
1449 | tp->tcp_header_len + | |
1450 | icsk->icsk_ext_hdr_len + | |
1451 | icsk->icsk_af_ops->net_header_len; | |
1452 | ||
1453 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ | |
1454 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1455 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1456 | ||
1457 | if (dst && dst_allfrag(dst)) | |
1458 | mtu += icsk->icsk_af_ops->net_frag_header_len; | |
1459 | } | |
1460 | return mtu; | |
1461 | } | |
1462 | EXPORT_SYMBOL(tcp_mss_to_mtu); | |
1463 | ||
1464 | /* MTU probing init per socket */ | |
1465 | void tcp_mtup_init(struct sock *sk) | |
1466 | { | |
1467 | struct tcp_sock *tp = tcp_sk(sk); | |
1468 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1469 | struct net *net = sock_net(sk); | |
1470 | ||
1471 | icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; | |
1472 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + | |
1473 | icsk->icsk_af_ops->net_header_len; | |
1474 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); | |
1475 | icsk->icsk_mtup.probe_size = 0; | |
1476 | if (icsk->icsk_mtup.enabled) | |
1477 | icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; | |
1478 | } | |
1479 | EXPORT_SYMBOL(tcp_mtup_init); | |
1480 | ||
1481 | /* This function synchronize snd mss to current pmtu/exthdr set. | |
1482 | ||
1483 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
1484 | for TCP options, but includes only bare TCP header. | |
1485 | ||
1486 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
1487 | It is minimum of user_mss and mss received with SYN. | |
1488 | It also does not include TCP options. | |
1489 | ||
1490 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. | |
1491 | ||
1492 | tp->mss_cache is current effective sending mss, including | |
1493 | all tcp options except for SACKs. It is evaluated, | |
1494 | taking into account current pmtu, but never exceeds | |
1495 | tp->rx_opt.mss_clamp. | |
1496 | ||
1497 | NOTE1. rfc1122 clearly states that advertised MSS | |
1498 | DOES NOT include either tcp or ip options. | |
1499 | ||
1500 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache | |
1501 | are READ ONLY outside this function. --ANK (980731) | |
1502 | */ | |
1503 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) | |
1504 | { | |
1505 | struct tcp_sock *tp = tcp_sk(sk); | |
1506 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1507 | int mss_now; | |
1508 | ||
1509 | if (icsk->icsk_mtup.search_high > pmtu) | |
1510 | icsk->icsk_mtup.search_high = pmtu; | |
1511 | ||
1512 | mss_now = tcp_mtu_to_mss(sk, pmtu); | |
1513 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); | |
1514 | ||
1515 | /* And store cached results */ | |
1516 | icsk->icsk_pmtu_cookie = pmtu; | |
1517 | if (icsk->icsk_mtup.enabled) | |
1518 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
1519 | tp->mss_cache = mss_now; | |
1520 | ||
1521 | return mss_now; | |
1522 | } | |
1523 | EXPORT_SYMBOL(tcp_sync_mss); | |
1524 | ||
1525 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1526 | * and even PMTU discovery events into account. | |
1527 | */ | |
1528 | unsigned int tcp_current_mss(struct sock *sk) | |
1529 | { | |
1530 | const struct tcp_sock *tp = tcp_sk(sk); | |
1531 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1532 | u32 mss_now; | |
1533 | unsigned int header_len; | |
1534 | struct tcp_out_options opts; | |
1535 | struct tcp_md5sig_key *md5; | |
1536 | ||
1537 | mss_now = tp->mss_cache; | |
1538 | ||
1539 | if (dst) { | |
1540 | u32 mtu = dst_mtu(dst); | |
1541 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) | |
1542 | mss_now = tcp_sync_mss(sk, mtu); | |
1543 | } | |
1544 | ||
1545 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + | |
1546 | sizeof(struct tcphdr); | |
1547 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1548 | * some common options. If this is an odd packet (because we have SACK | |
1549 | * blocks etc) then our calculated header_len will be different, and | |
1550 | * we have to adjust mss_now correspondingly */ | |
1551 | if (header_len != tp->tcp_header_len) { | |
1552 | int delta = (int) header_len - tp->tcp_header_len; | |
1553 | mss_now -= delta; | |
1554 | } | |
1555 | ||
1556 | return mss_now; | |
1557 | } | |
1558 | ||
1559 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. | |
1560 | * As additional protections, we do not touch cwnd in retransmission phases, | |
1561 | * and if application hit its sndbuf limit recently. | |
1562 | */ | |
1563 | static void tcp_cwnd_application_limited(struct sock *sk) | |
1564 | { | |
1565 | struct tcp_sock *tp = tcp_sk(sk); | |
1566 | ||
1567 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && | |
1568 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { | |
1569 | /* Limited by application or receiver window. */ | |
1570 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); | |
1571 | u32 win_used = max(tp->snd_cwnd_used, init_win); | |
1572 | if (win_used < tp->snd_cwnd) { | |
1573 | tp->snd_ssthresh = tcp_current_ssthresh(sk); | |
1574 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; | |
1575 | } | |
1576 | tp->snd_cwnd_used = 0; | |
1577 | } | |
1578 | tp->snd_cwnd_stamp = tcp_jiffies32; | |
1579 | } | |
1580 | ||
1581 | static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) | |
1582 | { | |
1583 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; | |
1584 | struct tcp_sock *tp = tcp_sk(sk); | |
1585 | ||
1586 | /* Track the maximum number of outstanding packets in each | |
1587 | * window, and remember whether we were cwnd-limited then. | |
1588 | */ | |
1589 | if (!before(tp->snd_una, tp->max_packets_seq) || | |
1590 | tp->packets_out > tp->max_packets_out) { | |
1591 | tp->max_packets_out = tp->packets_out; | |
1592 | tp->max_packets_seq = tp->snd_nxt; | |
1593 | tp->is_cwnd_limited = is_cwnd_limited; | |
1594 | } | |
1595 | ||
1596 | if (tcp_is_cwnd_limited(sk)) { | |
1597 | /* Network is feed fully. */ | |
1598 | tp->snd_cwnd_used = 0; | |
1599 | tp->snd_cwnd_stamp = tcp_jiffies32; | |
1600 | } else { | |
1601 | /* Network starves. */ | |
1602 | if (tp->packets_out > tp->snd_cwnd_used) | |
1603 | tp->snd_cwnd_used = tp->packets_out; | |
1604 | ||
1605 | if (sysctl_tcp_slow_start_after_idle && | |
1606 | (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && | |
1607 | !ca_ops->cong_control) | |
1608 | tcp_cwnd_application_limited(sk); | |
1609 | ||
1610 | /* The following conditions together indicate the starvation | |
1611 | * is caused by insufficient sender buffer: | |
1612 | * 1) just sent some data (see tcp_write_xmit) | |
1613 | * 2) not cwnd limited (this else condition) | |
1614 | * 3) no more data to send (null tcp_send_head ) | |
1615 | * 4) application is hitting buffer limit (SOCK_NOSPACE) | |
1616 | */ | |
1617 | if (!tcp_send_head(sk) && sk->sk_socket && | |
1618 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && | |
1619 | (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) | |
1620 | tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); | |
1621 | } | |
1622 | } | |
1623 | ||
1624 | /* Minshall's variant of the Nagle send check. */ | |
1625 | static bool tcp_minshall_check(const struct tcp_sock *tp) | |
1626 | { | |
1627 | return after(tp->snd_sml, tp->snd_una) && | |
1628 | !after(tp->snd_sml, tp->snd_nxt); | |
1629 | } | |
1630 | ||
1631 | /* Update snd_sml if this skb is under mss | |
1632 | * Note that a TSO packet might end with a sub-mss segment | |
1633 | * The test is really : | |
1634 | * if ((skb->len % mss) != 0) | |
1635 | * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; | |
1636 | * But we can avoid doing the divide again given we already have | |
1637 | * skb_pcount = skb->len / mss_now | |
1638 | */ | |
1639 | static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, | |
1640 | const struct sk_buff *skb) | |
1641 | { | |
1642 | if (skb->len < tcp_skb_pcount(skb) * mss_now) | |
1643 | tp->snd_sml = TCP_SKB_CB(skb)->end_seq; | |
1644 | } | |
1645 | ||
1646 | /* Return false, if packet can be sent now without violation Nagle's rules: | |
1647 | * 1. It is full sized. (provided by caller in %partial bool) | |
1648 | * 2. Or it contains FIN. (already checked by caller) | |
1649 | * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. | |
1650 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1651 | * With Minshall's modification: all sent small packets are ACKed. | |
1652 | */ | |
1653 | static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, | |
1654 | int nonagle) | |
1655 | { | |
1656 | return partial && | |
1657 | ((nonagle & TCP_NAGLE_CORK) || | |
1658 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); | |
1659 | } | |
1660 | ||
1661 | /* Return how many segs we'd like on a TSO packet, | |
1662 | * to send one TSO packet per ms | |
1663 | */ | |
1664 | u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, | |
1665 | int min_tso_segs) | |
1666 | { | |
1667 | u32 bytes, segs; | |
1668 | ||
1669 | bytes = min(sk->sk_pacing_rate >> 10, | |
1670 | sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); | |
1671 | ||
1672 | /* Goal is to send at least one packet per ms, | |
1673 | * not one big TSO packet every 100 ms. | |
1674 | * This preserves ACK clocking and is consistent | |
1675 | * with tcp_tso_should_defer() heuristic. | |
1676 | */ | |
1677 | segs = max_t(u32, bytes / mss_now, min_tso_segs); | |
1678 | ||
1679 | return min_t(u32, segs, sk->sk_gso_max_segs); | |
1680 | } | |
1681 | EXPORT_SYMBOL(tcp_tso_autosize); | |
1682 | ||
1683 | /* Return the number of segments we want in the skb we are transmitting. | |
1684 | * See if congestion control module wants to decide; otherwise, autosize. | |
1685 | */ | |
1686 | static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) | |
1687 | { | |
1688 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; | |
1689 | u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0; | |
1690 | ||
1691 | return tso_segs ? : | |
1692 | tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs); | |
1693 | } | |
1694 | ||
1695 | /* Returns the portion of skb which can be sent right away */ | |
1696 | static unsigned int tcp_mss_split_point(const struct sock *sk, | |
1697 | const struct sk_buff *skb, | |
1698 | unsigned int mss_now, | |
1699 | unsigned int max_segs, | |
1700 | int nonagle) | |
1701 | { | |
1702 | const struct tcp_sock *tp = tcp_sk(sk); | |
1703 | u32 partial, needed, window, max_len; | |
1704 | ||
1705 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; | |
1706 | max_len = mss_now * max_segs; | |
1707 | ||
1708 | if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) | |
1709 | return max_len; | |
1710 | ||
1711 | needed = min(skb->len, window); | |
1712 | ||
1713 | if (max_len <= needed) | |
1714 | return max_len; | |
1715 | ||
1716 | partial = needed % mss_now; | |
1717 | /* If last segment is not a full MSS, check if Nagle rules allow us | |
1718 | * to include this last segment in this skb. | |
1719 | * Otherwise, we'll split the skb at last MSS boundary | |
1720 | */ | |
1721 | if (tcp_nagle_check(partial != 0, tp, nonagle)) | |
1722 | return needed - partial; | |
1723 | ||
1724 | return needed; | |
1725 | } | |
1726 | ||
1727 | /* Can at least one segment of SKB be sent right now, according to the | |
1728 | * congestion window rules? If so, return how many segments are allowed. | |
1729 | */ | |
1730 | static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, | |
1731 | const struct sk_buff *skb) | |
1732 | { | |
1733 | u32 in_flight, cwnd, halfcwnd; | |
1734 | ||
1735 | /* Don't be strict about the congestion window for the final FIN. */ | |
1736 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && | |
1737 | tcp_skb_pcount(skb) == 1) | |
1738 | return 1; | |
1739 | ||
1740 | in_flight = tcp_packets_in_flight(tp); | |
1741 | cwnd = tp->snd_cwnd; | |
1742 | if (in_flight >= cwnd) | |
1743 | return 0; | |
1744 | ||
1745 | /* For better scheduling, ensure we have at least | |
1746 | * 2 GSO packets in flight. | |
1747 | */ | |
1748 | halfcwnd = max(cwnd >> 1, 1U); | |
1749 | return min(halfcwnd, cwnd - in_flight); | |
1750 | } | |
1751 | ||
1752 | /* Initialize TSO state of a skb. | |
1753 | * This must be invoked the first time we consider transmitting | |
1754 | * SKB onto the wire. | |
1755 | */ | |
1756 | static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) | |
1757 | { | |
1758 | int tso_segs = tcp_skb_pcount(skb); | |
1759 | ||
1760 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { | |
1761 | tcp_set_skb_tso_segs(skb, mss_now); | |
1762 | tso_segs = tcp_skb_pcount(skb); | |
1763 | } | |
1764 | return tso_segs; | |
1765 | } | |
1766 | ||
1767 | ||
1768 | /* Return true if the Nagle test allows this packet to be | |
1769 | * sent now. | |
1770 | */ | |
1771 | static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, | |
1772 | unsigned int cur_mss, int nonagle) | |
1773 | { | |
1774 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1775 | * write_queue (they have no chances to get new data). | |
1776 | * | |
1777 | * This is implemented in the callers, where they modify the 'nonagle' | |
1778 | * argument based upon the location of SKB in the send queue. | |
1779 | */ | |
1780 | if (nonagle & TCP_NAGLE_PUSH) | |
1781 | return true; | |
1782 | ||
1783 | /* Don't use the nagle rule for urgent data (or for the final FIN). */ | |
1784 | if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) | |
1785 | return true; | |
1786 | ||
1787 | if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) | |
1788 | return true; | |
1789 | ||
1790 | return false; | |
1791 | } | |
1792 | ||
1793 | /* Does at least the first segment of SKB fit into the send window? */ | |
1794 | static bool tcp_snd_wnd_test(const struct tcp_sock *tp, | |
1795 | const struct sk_buff *skb, | |
1796 | unsigned int cur_mss) | |
1797 | { | |
1798 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1799 | ||
1800 | if (skb->len > cur_mss) | |
1801 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1802 | ||
1803 | return !after(end_seq, tcp_wnd_end(tp)); | |
1804 | } | |
1805 | ||
1806 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) | |
1807 | * should be put on the wire right now. If so, it returns the number of | |
1808 | * packets allowed by the congestion window. | |
1809 | */ | |
1810 | static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, | |
1811 | unsigned int cur_mss, int nonagle) | |
1812 | { | |
1813 | const struct tcp_sock *tp = tcp_sk(sk); | |
1814 | unsigned int cwnd_quota; | |
1815 | ||
1816 | tcp_init_tso_segs(skb, cur_mss); | |
1817 | ||
1818 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1819 | return 0; | |
1820 | ||
1821 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
1822 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) | |
1823 | cwnd_quota = 0; | |
1824 | ||
1825 | return cwnd_quota; | |
1826 | } | |
1827 | ||
1828 | /* Test if sending is allowed right now. */ | |
1829 | bool tcp_may_send_now(struct sock *sk) | |
1830 | { | |
1831 | const struct tcp_sock *tp = tcp_sk(sk); | |
1832 | struct sk_buff *skb = tcp_send_head(sk); | |
1833 | ||
1834 | return skb && | |
1835 | tcp_snd_test(sk, skb, tcp_current_mss(sk), | |
1836 | (tcp_skb_is_last(sk, skb) ? | |
1837 | tp->nonagle : TCP_NAGLE_PUSH)); | |
1838 | } | |
1839 | ||
1840 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet | |
1841 | * which is put after SKB on the list. It is very much like | |
1842 | * tcp_fragment() except that it may make several kinds of assumptions | |
1843 | * in order to speed up the splitting operation. In particular, we | |
1844 | * know that all the data is in scatter-gather pages, and that the | |
1845 | * packet has never been sent out before (and thus is not cloned). | |
1846 | */ | |
1847 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, | |
1848 | unsigned int mss_now, gfp_t gfp) | |
1849 | { | |
1850 | struct sk_buff *buff; | |
1851 | int nlen = skb->len - len; | |
1852 | u8 flags; | |
1853 | ||
1854 | /* All of a TSO frame must be composed of paged data. */ | |
1855 | if (skb->len != skb->data_len) | |
1856 | return tcp_fragment(sk, skb, len, mss_now, gfp); | |
1857 | ||
1858 | buff = sk_stream_alloc_skb(sk, 0, gfp, true); | |
1859 | if (unlikely(!buff)) | |
1860 | return -ENOMEM; | |
1861 | ||
1862 | sk->sk_wmem_queued += buff->truesize; | |
1863 | sk_mem_charge(sk, buff->truesize); | |
1864 | buff->truesize += nlen; | |
1865 | skb->truesize -= nlen; | |
1866 | ||
1867 | /* Correct the sequence numbers. */ | |
1868 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1869 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1870 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1871 | ||
1872 | /* PSH and FIN should only be set in the second packet. */ | |
1873 | flags = TCP_SKB_CB(skb)->tcp_flags; | |
1874 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1875 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
1876 | ||
1877 | /* This packet was never sent out yet, so no SACK bits. */ | |
1878 | TCP_SKB_CB(buff)->sacked = 0; | |
1879 | ||
1880 | tcp_skb_fragment_eor(skb, buff); | |
1881 | ||
1882 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; | |
1883 | skb_split(skb, buff, len); | |
1884 | tcp_fragment_tstamp(skb, buff); | |
1885 | ||
1886 | /* Fix up tso_factor for both original and new SKB. */ | |
1887 | tcp_set_skb_tso_segs(skb, mss_now); | |
1888 | tcp_set_skb_tso_segs(buff, mss_now); | |
1889 | ||
1890 | /* Link BUFF into the send queue. */ | |
1891 | __skb_header_release(buff); | |
1892 | tcp_insert_write_queue_after(skb, buff, sk); | |
1893 | ||
1894 | return 0; | |
1895 | } | |
1896 | ||
1897 | /* Try to defer sending, if possible, in order to minimize the amount | |
1898 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
1899 | * | |
1900 | * This algorithm is from John Heffner. | |
1901 | */ | |
1902 | static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, | |
1903 | bool *is_cwnd_limited, u32 max_segs) | |
1904 | { | |
1905 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
1906 | u32 age, send_win, cong_win, limit, in_flight; | |
1907 | struct tcp_sock *tp = tcp_sk(sk); | |
1908 | struct sk_buff *head; | |
1909 | int win_divisor; | |
1910 | ||
1911 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) | |
1912 | goto send_now; | |
1913 | ||
1914 | if (icsk->icsk_ca_state >= TCP_CA_Recovery) | |
1915 | goto send_now; | |
1916 | ||
1917 | /* Avoid bursty behavior by allowing defer | |
1918 | * only if the last write was recent. | |
1919 | */ | |
1920 | if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0) | |
1921 | goto send_now; | |
1922 | ||
1923 | in_flight = tcp_packets_in_flight(tp); | |
1924 | ||
1925 | BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); | |
1926 | ||
1927 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; | |
1928 | ||
1929 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
1930 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
1931 | ||
1932 | limit = min(send_win, cong_win); | |
1933 | ||
1934 | /* If a full-sized TSO skb can be sent, do it. */ | |
1935 | if (limit >= max_segs * tp->mss_cache) | |
1936 | goto send_now; | |
1937 | ||
1938 | /* Middle in queue won't get any more data, full sendable already? */ | |
1939 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) | |
1940 | goto send_now; | |
1941 | ||
1942 | win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); | |
1943 | if (win_divisor) { | |
1944 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); | |
1945 | ||
1946 | /* If at least some fraction of a window is available, | |
1947 | * just use it. | |
1948 | */ | |
1949 | chunk /= win_divisor; | |
1950 | if (limit >= chunk) | |
1951 | goto send_now; | |
1952 | } else { | |
1953 | /* Different approach, try not to defer past a single | |
1954 | * ACK. Receiver should ACK every other full sized | |
1955 | * frame, so if we have space for more than 3 frames | |
1956 | * then send now. | |
1957 | */ | |
1958 | if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) | |
1959 | goto send_now; | |
1960 | } | |
1961 | ||
1962 | head = tcp_write_queue_head(sk); | |
1963 | ||
1964 | age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp); | |
1965 | /* If next ACK is likely to come too late (half srtt), do not defer */ | |
1966 | if (age < (tp->srtt_us >> 4)) | |
1967 | goto send_now; | |
1968 | ||
1969 | /* Ok, it looks like it is advisable to defer. */ | |
1970 | ||
1971 | if (cong_win < send_win && cong_win <= skb->len) | |
1972 | *is_cwnd_limited = true; | |
1973 | ||
1974 | return true; | |
1975 | ||
1976 | send_now: | |
1977 | return false; | |
1978 | } | |
1979 | ||
1980 | static inline void tcp_mtu_check_reprobe(struct sock *sk) | |
1981 | { | |
1982 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1983 | struct tcp_sock *tp = tcp_sk(sk); | |
1984 | struct net *net = sock_net(sk); | |
1985 | u32 interval; | |
1986 | s32 delta; | |
1987 | ||
1988 | interval = net->ipv4.sysctl_tcp_probe_interval; | |
1989 | delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; | |
1990 | if (unlikely(delta >= interval * HZ)) { | |
1991 | int mss = tcp_current_mss(sk); | |
1992 | ||
1993 | /* Update current search range */ | |
1994 | icsk->icsk_mtup.probe_size = 0; | |
1995 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + | |
1996 | sizeof(struct tcphdr) + | |
1997 | icsk->icsk_af_ops->net_header_len; | |
1998 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); | |
1999 | ||
2000 | /* Update probe time stamp */ | |
2001 | icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; | |
2002 | } | |
2003 | } | |
2004 | ||
2005 | /* Create a new MTU probe if we are ready. | |
2006 | * MTU probe is regularly attempting to increase the path MTU by | |
2007 | * deliberately sending larger packets. This discovers routing | |
2008 | * changes resulting in larger path MTUs. | |
2009 | * | |
2010 | * Returns 0 if we should wait to probe (no cwnd available), | |
2011 | * 1 if a probe was sent, | |
2012 | * -1 otherwise | |
2013 | */ | |
2014 | static int tcp_mtu_probe(struct sock *sk) | |
2015 | { | |
2016 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2017 | struct tcp_sock *tp = tcp_sk(sk); | |
2018 | struct sk_buff *skb, *nskb, *next; | |
2019 | struct net *net = sock_net(sk); | |
2020 | int probe_size; | |
2021 | int size_needed; | |
2022 | int copy, len; | |
2023 | int mss_now; | |
2024 | int interval; | |
2025 | ||
2026 | /* Not currently probing/verifying, | |
2027 | * not in recovery, | |
2028 | * have enough cwnd, and | |
2029 | * not SACKing (the variable headers throw things off) | |
2030 | */ | |
2031 | if (likely(!icsk->icsk_mtup.enabled || | |
2032 | icsk->icsk_mtup.probe_size || | |
2033 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
2034 | tp->snd_cwnd < 11 || | |
2035 | tp->rx_opt.num_sacks || tp->rx_opt.dsack)) | |
2036 | return -1; | |
2037 | ||
2038 | /* Use binary search for probe_size between tcp_mss_base, | |
2039 | * and current mss_clamp. if (search_high - search_low) | |
2040 | * smaller than a threshold, backoff from probing. | |
2041 | */ | |
2042 | mss_now = tcp_current_mss(sk); | |
2043 | probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + | |
2044 | icsk->icsk_mtup.search_low) >> 1); | |
2045 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; | |
2046 | interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; | |
2047 | /* When misfortune happens, we are reprobing actively, | |
2048 | * and then reprobe timer has expired. We stick with current | |
2049 | * probing process by not resetting search range to its orignal. | |
2050 | */ | |
2051 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || | |
2052 | interval < net->ipv4.sysctl_tcp_probe_threshold) { | |
2053 | /* Check whether enough time has elaplased for | |
2054 | * another round of probing. | |
2055 | */ | |
2056 | tcp_mtu_check_reprobe(sk); | |
2057 | return -1; | |
2058 | } | |
2059 | ||
2060 | /* Have enough data in the send queue to probe? */ | |
2061 | if (tp->write_seq - tp->snd_nxt < size_needed) | |
2062 | return -1; | |
2063 | ||
2064 | if (tp->snd_wnd < size_needed) | |
2065 | return -1; | |
2066 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) | |
2067 | return 0; | |
2068 | ||
2069 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ | |
2070 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { | |
2071 | if (!tcp_packets_in_flight(tp)) | |
2072 | return -1; | |
2073 | else | |
2074 | return 0; | |
2075 | } | |
2076 | ||
2077 | /* We're allowed to probe. Build it now. */ | |
2078 | nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); | |
2079 | if (!nskb) | |
2080 | return -1; | |
2081 | sk->sk_wmem_queued += nskb->truesize; | |
2082 | sk_mem_charge(sk, nskb->truesize); | |
2083 | ||
2084 | skb = tcp_send_head(sk); | |
2085 | ||
2086 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
2087 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
2088 | TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; | |
2089 | TCP_SKB_CB(nskb)->sacked = 0; | |
2090 | nskb->csum = 0; | |
2091 | nskb->ip_summed = skb->ip_summed; | |
2092 | ||
2093 | tcp_insert_write_queue_before(nskb, skb, sk); | |
2094 | ||
2095 | len = 0; | |
2096 | tcp_for_write_queue_from_safe(skb, next, sk) { | |
2097 | copy = min_t(int, skb->len, probe_size - len); | |
2098 | if (nskb->ip_summed) { | |
2099 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); | |
2100 | } else { | |
2101 | __wsum csum = skb_copy_and_csum_bits(skb, 0, | |
2102 | skb_put(nskb, copy), | |
2103 | copy, 0); | |
2104 | nskb->csum = csum_block_add(nskb->csum, csum, len); | |
2105 | } | |
2106 | ||
2107 | if (skb->len <= copy) { | |
2108 | /* We've eaten all the data from this skb. | |
2109 | * Throw it away. */ | |
2110 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; | |
2111 | tcp_unlink_write_queue(skb, sk); | |
2112 | sk_wmem_free_skb(sk, skb); | |
2113 | } else { | |
2114 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & | |
2115 | ~(TCPHDR_FIN|TCPHDR_PSH); | |
2116 | if (!skb_shinfo(skb)->nr_frags) { | |
2117 | skb_pull(skb, copy); | |
2118 | if (skb->ip_summed != CHECKSUM_PARTIAL) | |
2119 | skb->csum = csum_partial(skb->data, | |
2120 | skb->len, 0); | |
2121 | } else { | |
2122 | __pskb_trim_head(skb, copy); | |
2123 | tcp_set_skb_tso_segs(skb, mss_now); | |
2124 | } | |
2125 | TCP_SKB_CB(skb)->seq += copy; | |
2126 | } | |
2127 | ||
2128 | len += copy; | |
2129 | ||
2130 | if (len >= probe_size) | |
2131 | break; | |
2132 | } | |
2133 | tcp_init_tso_segs(nskb, nskb->len); | |
2134 | ||
2135 | /* We're ready to send. If this fails, the probe will | |
2136 | * be resegmented into mss-sized pieces by tcp_write_xmit(). | |
2137 | */ | |
2138 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { | |
2139 | /* Decrement cwnd here because we are sending | |
2140 | * effectively two packets. */ | |
2141 | tp->snd_cwnd--; | |
2142 | tcp_event_new_data_sent(sk, nskb); | |
2143 | ||
2144 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
2145 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; | |
2146 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
2147 | ||
2148 | return 1; | |
2149 | } | |
2150 | ||
2151 | return -1; | |
2152 | } | |
2153 | ||
2154 | static bool tcp_pacing_check(const struct sock *sk) | |
2155 | { | |
2156 | return tcp_needs_internal_pacing(sk) && | |
2157 | hrtimer_active(&tcp_sk(sk)->pacing_timer); | |
2158 | } | |
2159 | ||
2160 | /* TCP Small Queues : | |
2161 | * Control number of packets in qdisc/devices to two packets / or ~1 ms. | |
2162 | * (These limits are doubled for retransmits) | |
2163 | * This allows for : | |
2164 | * - better RTT estimation and ACK scheduling | |
2165 | * - faster recovery | |
2166 | * - high rates | |
2167 | * Alas, some drivers / subsystems require a fair amount | |
2168 | * of queued bytes to ensure line rate. | |
2169 | * One example is wifi aggregation (802.11 AMPDU) | |
2170 | */ | |
2171 | static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, | |
2172 | unsigned int factor) | |
2173 | { | |
2174 | unsigned int limit; | |
2175 | ||
2176 | limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); | |
2177 | limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); | |
2178 | limit <<= factor; | |
2179 | ||
2180 | if (refcount_read(&sk->sk_wmem_alloc) > limit) { | |
2181 | /* Always send the 1st or 2nd skb in write queue. | |
2182 | * No need to wait for TX completion to call us back, | |
2183 | * after softirq/tasklet schedule. | |
2184 | * This helps when TX completions are delayed too much. | |
2185 | */ | |
2186 | if (skb == sk->sk_write_queue.next || | |
2187 | skb->prev == sk->sk_write_queue.next) | |
2188 | return false; | |
2189 | ||
2190 | set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); | |
2191 | /* It is possible TX completion already happened | |
2192 | * before we set TSQ_THROTTLED, so we must | |
2193 | * test again the condition. | |
2194 | */ | |
2195 | smp_mb__after_atomic(); | |
2196 | if (refcount_read(&sk->sk_wmem_alloc) > limit) | |
2197 | return true; | |
2198 | } | |
2199 | return false; | |
2200 | } | |
2201 | ||
2202 | static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) | |
2203 | { | |
2204 | const u32 now = tcp_jiffies32; | |
2205 | enum tcp_chrono old = tp->chrono_type; | |
2206 | ||
2207 | if (old > TCP_CHRONO_UNSPEC) | |
2208 | tp->chrono_stat[old - 1] += now - tp->chrono_start; | |
2209 | tp->chrono_start = now; | |
2210 | tp->chrono_type = new; | |
2211 | } | |
2212 | ||
2213 | void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) | |
2214 | { | |
2215 | struct tcp_sock *tp = tcp_sk(sk); | |
2216 | ||
2217 | /* If there are multiple conditions worthy of tracking in a | |
2218 | * chronograph then the highest priority enum takes precedence | |
2219 | * over the other conditions. So that if something "more interesting" | |
2220 | * starts happening, stop the previous chrono and start a new one. | |
2221 | */ | |
2222 | if (type > tp->chrono_type) | |
2223 | tcp_chrono_set(tp, type); | |
2224 | } | |
2225 | ||
2226 | void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) | |
2227 | { | |
2228 | struct tcp_sock *tp = tcp_sk(sk); | |
2229 | ||
2230 | ||
2231 | /* There are multiple conditions worthy of tracking in a | |
2232 | * chronograph, so that the highest priority enum takes | |
2233 | * precedence over the other conditions (see tcp_chrono_start). | |
2234 | * If a condition stops, we only stop chrono tracking if | |
2235 | * it's the "most interesting" or current chrono we are | |
2236 | * tracking and starts busy chrono if we have pending data. | |
2237 | */ | |
2238 | if (tcp_write_queue_empty(sk)) | |
2239 | tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); | |
2240 | else if (type == tp->chrono_type) | |
2241 | tcp_chrono_set(tp, TCP_CHRONO_BUSY); | |
2242 | } | |
2243 | ||
2244 | /* This routine writes packets to the network. It advances the | |
2245 | * send_head. This happens as incoming acks open up the remote | |
2246 | * window for us. | |
2247 | * | |
2248 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between | |
2249 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into | |
2250 | * account rare use of URG, this is not a big flaw. | |
2251 | * | |
2252 | * Send at most one packet when push_one > 0. Temporarily ignore | |
2253 | * cwnd limit to force at most one packet out when push_one == 2. | |
2254 | ||
2255 | * Returns true, if no segments are in flight and we have queued segments, | |
2256 | * but cannot send anything now because of SWS or another problem. | |
2257 | */ | |
2258 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, | |
2259 | int push_one, gfp_t gfp) | |
2260 | { | |
2261 | struct tcp_sock *tp = tcp_sk(sk); | |
2262 | struct sk_buff *skb; | |
2263 | unsigned int tso_segs, sent_pkts; | |
2264 | int cwnd_quota; | |
2265 | int result; | |
2266 | bool is_cwnd_limited = false, is_rwnd_limited = false; | |
2267 | u32 max_segs; | |
2268 | ||
2269 | sent_pkts = 0; | |
2270 | ||
2271 | if (!push_one) { | |
2272 | /* Do MTU probing. */ | |
2273 | result = tcp_mtu_probe(sk); | |
2274 | if (!result) { | |
2275 | return false; | |
2276 | } else if (result > 0) { | |
2277 | sent_pkts = 1; | |
2278 | } | |
2279 | } | |
2280 | ||
2281 | max_segs = tcp_tso_segs(sk, mss_now); | |
2282 | tcp_mstamp_refresh(tp); | |
2283 | while ((skb = tcp_send_head(sk))) { | |
2284 | unsigned int limit; | |
2285 | ||
2286 | if (tcp_pacing_check(sk)) | |
2287 | break; | |
2288 | ||
2289 | tso_segs = tcp_init_tso_segs(skb, mss_now); | |
2290 | BUG_ON(!tso_segs); | |
2291 | ||
2292 | if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { | |
2293 | /* "skb_mstamp" is used as a start point for the retransmit timer */ | |
2294 | skb->skb_mstamp = tp->tcp_mstamp; | |
2295 | goto repair; /* Skip network transmission */ | |
2296 | } | |
2297 | ||
2298 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
2299 | if (!cwnd_quota) { | |
2300 | if (push_one == 2) | |
2301 | /* Force out a loss probe pkt. */ | |
2302 | cwnd_quota = 1; | |
2303 | else | |
2304 | break; | |
2305 | } | |
2306 | ||
2307 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { | |
2308 | is_rwnd_limited = true; | |
2309 | break; | |
2310 | } | |
2311 | ||
2312 | if (tso_segs == 1) { | |
2313 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, | |
2314 | (tcp_skb_is_last(sk, skb) ? | |
2315 | nonagle : TCP_NAGLE_PUSH)))) | |
2316 | break; | |
2317 | } else { | |
2318 | if (!push_one && | |
2319 | tcp_tso_should_defer(sk, skb, &is_cwnd_limited, | |
2320 | max_segs)) | |
2321 | break; | |
2322 | } | |
2323 | ||
2324 | limit = mss_now; | |
2325 | if (tso_segs > 1 && !tcp_urg_mode(tp)) | |
2326 | limit = tcp_mss_split_point(sk, skb, mss_now, | |
2327 | min_t(unsigned int, | |
2328 | cwnd_quota, | |
2329 | max_segs), | |
2330 | nonagle); | |
2331 | ||
2332 | if (skb->len > limit && | |
2333 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) | |
2334 | break; | |
2335 | ||
2336 | if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) | |
2337 | clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); | |
2338 | if (tcp_small_queue_check(sk, skb, 0)) | |
2339 | break; | |
2340 | ||
2341 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) | |
2342 | break; | |
2343 | ||
2344 | repair: | |
2345 | /* Advance the send_head. This one is sent out. | |
2346 | * This call will increment packets_out. | |
2347 | */ | |
2348 | tcp_event_new_data_sent(sk, skb); | |
2349 | ||
2350 | tcp_minshall_update(tp, mss_now, skb); | |
2351 | sent_pkts += tcp_skb_pcount(skb); | |
2352 | ||
2353 | if (push_one) | |
2354 | break; | |
2355 | } | |
2356 | ||
2357 | if (is_rwnd_limited) | |
2358 | tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); | |
2359 | else | |
2360 | tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); | |
2361 | ||
2362 | if (likely(sent_pkts)) { | |
2363 | if (tcp_in_cwnd_reduction(sk)) | |
2364 | tp->prr_out += sent_pkts; | |
2365 | ||
2366 | /* Send one loss probe per tail loss episode. */ | |
2367 | if (push_one != 2) | |
2368 | tcp_schedule_loss_probe(sk); | |
2369 | is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); | |
2370 | tcp_cwnd_validate(sk, is_cwnd_limited); | |
2371 | return false; | |
2372 | } | |
2373 | return !tp->packets_out && tcp_send_head(sk); | |
2374 | } | |
2375 | ||
2376 | bool tcp_schedule_loss_probe(struct sock *sk) | |
2377 | { | |
2378 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2379 | struct tcp_sock *tp = tcp_sk(sk); | |
2380 | u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); | |
2381 | u32 timeout, rto_delta_us; | |
2382 | ||
2383 | /* Don't do any loss probe on a Fast Open connection before 3WHS | |
2384 | * finishes. | |
2385 | */ | |
2386 | if (tp->fastopen_rsk) | |
2387 | return false; | |
2388 | ||
2389 | /* Schedule a loss probe in 2*RTT for SACK capable connections | |
2390 | * in Open state, that are either limited by cwnd or application. | |
2391 | */ | |
2392 | if ((sysctl_tcp_early_retrans != 3 && sysctl_tcp_early_retrans != 4) || | |
2393 | !tp->packets_out || !tcp_is_sack(tp) || | |
2394 | icsk->icsk_ca_state != TCP_CA_Open) | |
2395 | return false; | |
2396 | ||
2397 | if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && | |
2398 | tcp_send_head(sk)) | |
2399 | return false; | |
2400 | ||
2401 | /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account | |
2402 | * for delayed ack when there's one outstanding packet. If no RTT | |
2403 | * sample is available then probe after TCP_TIMEOUT_INIT. | |
2404 | */ | |
2405 | timeout = rtt << 1 ? : TCP_TIMEOUT_INIT; | |
2406 | if (tp->packets_out == 1) | |
2407 | timeout = max_t(u32, timeout, | |
2408 | (rtt + (rtt >> 1) + TCP_DELACK_MAX)); | |
2409 | timeout = max_t(u32, timeout, msecs_to_jiffies(10)); | |
2410 | ||
2411 | /* If the RTO formula yields an earlier time, then use that time. */ | |
2412 | rto_delta_us = tcp_rto_delta_us(sk); /* How far in future is RTO? */ | |
2413 | if (rto_delta_us > 0) | |
2414 | timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); | |
2415 | ||
2416 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, | |
2417 | TCP_RTO_MAX); | |
2418 | return true; | |
2419 | } | |
2420 | ||
2421 | /* Thanks to skb fast clones, we can detect if a prior transmit of | |
2422 | * a packet is still in a qdisc or driver queue. | |
2423 | * In this case, there is very little point doing a retransmit ! | |
2424 | */ | |
2425 | static bool skb_still_in_host_queue(const struct sock *sk, | |
2426 | const struct sk_buff *skb) | |
2427 | { | |
2428 | if (unlikely(skb_fclone_busy(sk, skb))) { | |
2429 | NET_INC_STATS(sock_net(sk), | |
2430 | LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); | |
2431 | return true; | |
2432 | } | |
2433 | return false; | |
2434 | } | |
2435 | ||
2436 | /* When probe timeout (PTO) fires, try send a new segment if possible, else | |
2437 | * retransmit the last segment. | |
2438 | */ | |
2439 | void tcp_send_loss_probe(struct sock *sk) | |
2440 | { | |
2441 | struct tcp_sock *tp = tcp_sk(sk); | |
2442 | struct sk_buff *skb; | |
2443 | int pcount; | |
2444 | int mss = tcp_current_mss(sk); | |
2445 | ||
2446 | skb = tcp_send_head(sk); | |
2447 | if (skb) { | |
2448 | if (tcp_snd_wnd_test(tp, skb, mss)) { | |
2449 | pcount = tp->packets_out; | |
2450 | tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); | |
2451 | if (tp->packets_out > pcount) | |
2452 | goto probe_sent; | |
2453 | goto rearm_timer; | |
2454 | } | |
2455 | skb = tcp_write_queue_prev(sk, skb); | |
2456 | } else { | |
2457 | skb = tcp_write_queue_tail(sk); | |
2458 | } | |
2459 | ||
2460 | /* At most one outstanding TLP retransmission. */ | |
2461 | if (tp->tlp_high_seq) | |
2462 | goto rearm_timer; | |
2463 | ||
2464 | /* Retransmit last segment. */ | |
2465 | if (WARN_ON(!skb)) | |
2466 | goto rearm_timer; | |
2467 | ||
2468 | if (skb_still_in_host_queue(sk, skb)) | |
2469 | goto rearm_timer; | |
2470 | ||
2471 | pcount = tcp_skb_pcount(skb); | |
2472 | if (WARN_ON(!pcount)) | |
2473 | goto rearm_timer; | |
2474 | ||
2475 | if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { | |
2476 | if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, | |
2477 | GFP_ATOMIC))) | |
2478 | goto rearm_timer; | |
2479 | skb = tcp_write_queue_next(sk, skb); | |
2480 | } | |
2481 | ||
2482 | if (WARN_ON(!skb || !tcp_skb_pcount(skb))) | |
2483 | goto rearm_timer; | |
2484 | ||
2485 | if (__tcp_retransmit_skb(sk, skb, 1)) | |
2486 | goto rearm_timer; | |
2487 | ||
2488 | /* Record snd_nxt for loss detection. */ | |
2489 | tp->tlp_high_seq = tp->snd_nxt; | |
2490 | ||
2491 | probe_sent: | |
2492 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); | |
2493 | /* Reset s.t. tcp_rearm_rto will restart timer from now */ | |
2494 | inet_csk(sk)->icsk_pending = 0; | |
2495 | rearm_timer: | |
2496 | tcp_rearm_rto(sk); | |
2497 | } | |
2498 | ||
2499 | /* Push out any pending frames which were held back due to | |
2500 | * TCP_CORK or attempt at coalescing tiny packets. | |
2501 | * The socket must be locked by the caller. | |
2502 | */ | |
2503 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, | |
2504 | int nonagle) | |
2505 | { | |
2506 | /* If we are closed, the bytes will have to remain here. | |
2507 | * In time closedown will finish, we empty the write queue and | |
2508 | * all will be happy. | |
2509 | */ | |
2510 | if (unlikely(sk->sk_state == TCP_CLOSE)) | |
2511 | return; | |
2512 | ||
2513 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, | |
2514 | sk_gfp_mask(sk, GFP_ATOMIC))) | |
2515 | tcp_check_probe_timer(sk); | |
2516 | } | |
2517 | ||
2518 | /* Send _single_ skb sitting at the send head. This function requires | |
2519 | * true push pending frames to setup probe timer etc. | |
2520 | */ | |
2521 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
2522 | { | |
2523 | struct sk_buff *skb = tcp_send_head(sk); | |
2524 | ||
2525 | BUG_ON(!skb || skb->len < mss_now); | |
2526 | ||
2527 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); | |
2528 | } | |
2529 | ||
2530 | /* This function returns the amount that we can raise the | |
2531 | * usable window based on the following constraints | |
2532 | * | |
2533 | * 1. The window can never be shrunk once it is offered (RFC 793) | |
2534 | * 2. We limit memory per socket | |
2535 | * | |
2536 | * RFC 1122: | |
2537 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
2538 | * RECV.NEXT + RCV.WIN fixed until: | |
2539 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
2540 | * | |
2541 | * i.e. don't raise the right edge of the window until you can raise | |
2542 | * it at least MSS bytes. | |
2543 | * | |
2544 | * Unfortunately, the recommended algorithm breaks header prediction, | |
2545 | * since header prediction assumes th->window stays fixed. | |
2546 | * | |
2547 | * Strictly speaking, keeping th->window fixed violates the receiver | |
2548 | * side SWS prevention criteria. The problem is that under this rule | |
2549 | * a stream of single byte packets will cause the right side of the | |
2550 | * window to always advance by a single byte. | |
2551 | * | |
2552 | * Of course, if the sender implements sender side SWS prevention | |
2553 | * then this will not be a problem. | |
2554 | * | |
2555 | * BSD seems to make the following compromise: | |
2556 | * | |
2557 | * If the free space is less than the 1/4 of the maximum | |
2558 | * space available and the free space is less than 1/2 mss, | |
2559 | * then set the window to 0. | |
2560 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
2561 | * Otherwise, just prevent the window from shrinking | |
2562 | * and from being larger than the largest representable value. | |
2563 | * | |
2564 | * This prevents incremental opening of the window in the regime | |
2565 | * where TCP is limited by the speed of the reader side taking | |
2566 | * data out of the TCP receive queue. It does nothing about | |
2567 | * those cases where the window is constrained on the sender side | |
2568 | * because the pipeline is full. | |
2569 | * | |
2570 | * BSD also seems to "accidentally" limit itself to windows that are a | |
2571 | * multiple of MSS, at least until the free space gets quite small. | |
2572 | * This would appear to be a side effect of the mbuf implementation. | |
2573 | * Combining these two algorithms results in the observed behavior | |
2574 | * of having a fixed window size at almost all times. | |
2575 | * | |
2576 | * Below we obtain similar behavior by forcing the offered window to | |
2577 | * a multiple of the mss when it is feasible to do so. | |
2578 | * | |
2579 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
2580 | * Regular options like TIMESTAMP are taken into account. | |
2581 | */ | |
2582 | u32 __tcp_select_window(struct sock *sk) | |
2583 | { | |
2584 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2585 | struct tcp_sock *tp = tcp_sk(sk); | |
2586 | /* MSS for the peer's data. Previous versions used mss_clamp | |
2587 | * here. I don't know if the value based on our guesses | |
2588 | * of peer's MSS is better for the performance. It's more correct | |
2589 | * but may be worse for the performance because of rcv_mss | |
2590 | * fluctuations. --SAW 1998/11/1 | |
2591 | */ | |
2592 | int mss = icsk->icsk_ack.rcv_mss; | |
2593 | int free_space = tcp_space(sk); | |
2594 | int allowed_space = tcp_full_space(sk); | |
2595 | int full_space = min_t(int, tp->window_clamp, allowed_space); | |
2596 | int window; | |
2597 | ||
2598 | if (unlikely(mss > full_space)) { | |
2599 | mss = full_space; | |
2600 | if (mss <= 0) | |
2601 | return 0; | |
2602 | } | |
2603 | if (free_space < (full_space >> 1)) { | |
2604 | icsk->icsk_ack.quick = 0; | |
2605 | ||
2606 | if (tcp_under_memory_pressure(sk)) | |
2607 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, | |
2608 | 4U * tp->advmss); | |
2609 | ||
2610 | /* free_space might become our new window, make sure we don't | |
2611 | * increase it due to wscale. | |
2612 | */ | |
2613 | free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); | |
2614 | ||
2615 | /* if free space is less than mss estimate, or is below 1/16th | |
2616 | * of the maximum allowed, try to move to zero-window, else | |
2617 | * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and | |
2618 | * new incoming data is dropped due to memory limits. | |
2619 | * With large window, mss test triggers way too late in order | |
2620 | * to announce zero window in time before rmem limit kicks in. | |
2621 | */ | |
2622 | if (free_space < (allowed_space >> 4) || free_space < mss) | |
2623 | return 0; | |
2624 | } | |
2625 | ||
2626 | if (free_space > tp->rcv_ssthresh) | |
2627 | free_space = tp->rcv_ssthresh; | |
2628 | ||
2629 | /* Don't do rounding if we are using window scaling, since the | |
2630 | * scaled window will not line up with the MSS boundary anyway. | |
2631 | */ | |
2632 | if (tp->rx_opt.rcv_wscale) { | |
2633 | window = free_space; | |
2634 | ||
2635 | /* Advertise enough space so that it won't get scaled away. | |
2636 | * Import case: prevent zero window announcement if | |
2637 | * 1<<rcv_wscale > mss. | |
2638 | */ | |
2639 | window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); | |
2640 | } else { | |
2641 | window = tp->rcv_wnd; | |
2642 | /* Get the largest window that is a nice multiple of mss. | |
2643 | * Window clamp already applied above. | |
2644 | * If our current window offering is within 1 mss of the | |
2645 | * free space we just keep it. This prevents the divide | |
2646 | * and multiply from happening most of the time. | |
2647 | * We also don't do any window rounding when the free space | |
2648 | * is too small. | |
2649 | */ | |
2650 | if (window <= free_space - mss || window > free_space) | |
2651 | window = rounddown(free_space, mss); | |
2652 | else if (mss == full_space && | |
2653 | free_space > window + (full_space >> 1)) | |
2654 | window = free_space; | |
2655 | } | |
2656 | ||
2657 | return window; | |
2658 | } | |
2659 | ||
2660 | void tcp_skb_collapse_tstamp(struct sk_buff *skb, | |
2661 | const struct sk_buff *next_skb) | |
2662 | { | |
2663 | if (unlikely(tcp_has_tx_tstamp(next_skb))) { | |
2664 | const struct skb_shared_info *next_shinfo = | |
2665 | skb_shinfo(next_skb); | |
2666 | struct skb_shared_info *shinfo = skb_shinfo(skb); | |
2667 | ||
2668 | shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; | |
2669 | shinfo->tskey = next_shinfo->tskey; | |
2670 | TCP_SKB_CB(skb)->txstamp_ack |= | |
2671 | TCP_SKB_CB(next_skb)->txstamp_ack; | |
2672 | } | |
2673 | } | |
2674 | ||
2675 | /* Collapses two adjacent SKB's during retransmission. */ | |
2676 | static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) | |
2677 | { | |
2678 | struct tcp_sock *tp = tcp_sk(sk); | |
2679 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); | |
2680 | int skb_size, next_skb_size; | |
2681 | ||
2682 | skb_size = skb->len; | |
2683 | next_skb_size = next_skb->len; | |
2684 | ||
2685 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); | |
2686 | ||
2687 | if (next_skb_size) { | |
2688 | if (next_skb_size <= skb_availroom(skb)) | |
2689 | skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), | |
2690 | next_skb_size); | |
2691 | else if (!skb_shift(skb, next_skb, next_skb_size)) | |
2692 | return false; | |
2693 | } | |
2694 | tcp_highest_sack_combine(sk, next_skb, skb); | |
2695 | ||
2696 | tcp_unlink_write_queue(next_skb, sk); | |
2697 | ||
2698 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) | |
2699 | skb->ip_summed = CHECKSUM_PARTIAL; | |
2700 | ||
2701 | if (skb->ip_summed != CHECKSUM_PARTIAL) | |
2702 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); | |
2703 | ||
2704 | /* Update sequence range on original skb. */ | |
2705 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
2706 | ||
2707 | /* Merge over control information. This moves PSH/FIN etc. over */ | |
2708 | TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; | |
2709 | ||
2710 | /* All done, get rid of second SKB and account for it so | |
2711 | * packet counting does not break. | |
2712 | */ | |
2713 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; | |
2714 | TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; | |
2715 | ||
2716 | /* changed transmit queue under us so clear hints */ | |
2717 | tcp_clear_retrans_hints_partial(tp); | |
2718 | if (next_skb == tp->retransmit_skb_hint) | |
2719 | tp->retransmit_skb_hint = skb; | |
2720 | ||
2721 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); | |
2722 | ||
2723 | tcp_skb_collapse_tstamp(skb, next_skb); | |
2724 | ||
2725 | sk_wmem_free_skb(sk, next_skb); | |
2726 | return true; | |
2727 | } | |
2728 | ||
2729 | /* Check if coalescing SKBs is legal. */ | |
2730 | static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) | |
2731 | { | |
2732 | if (tcp_skb_pcount(skb) > 1) | |
2733 | return false; | |
2734 | if (skb_cloned(skb)) | |
2735 | return false; | |
2736 | if (skb == tcp_send_head(sk)) | |
2737 | return false; | |
2738 | /* Some heuristics for collapsing over SACK'd could be invented */ | |
2739 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
2740 | return false; | |
2741 | ||
2742 | return true; | |
2743 | } | |
2744 | ||
2745 | /* Collapse packets in the retransmit queue to make to create | |
2746 | * less packets on the wire. This is only done on retransmission. | |
2747 | */ | |
2748 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, | |
2749 | int space) | |
2750 | { | |
2751 | struct tcp_sock *tp = tcp_sk(sk); | |
2752 | struct sk_buff *skb = to, *tmp; | |
2753 | bool first = true; | |
2754 | ||
2755 | if (!sysctl_tcp_retrans_collapse) | |
2756 | return; | |
2757 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) | |
2758 | return; | |
2759 | ||
2760 | tcp_for_write_queue_from_safe(skb, tmp, sk) { | |
2761 | if (!tcp_can_collapse(sk, skb)) | |
2762 | break; | |
2763 | ||
2764 | if (!tcp_skb_can_collapse_to(to)) | |
2765 | break; | |
2766 | ||
2767 | space -= skb->len; | |
2768 | ||
2769 | if (first) { | |
2770 | first = false; | |
2771 | continue; | |
2772 | } | |
2773 | ||
2774 | if (space < 0) | |
2775 | break; | |
2776 | ||
2777 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) | |
2778 | break; | |
2779 | ||
2780 | if (!tcp_collapse_retrans(sk, to)) | |
2781 | break; | |
2782 | } | |
2783 | } | |
2784 | ||
2785 | /* This retransmits one SKB. Policy decisions and retransmit queue | |
2786 | * state updates are done by the caller. Returns non-zero if an | |
2787 | * error occurred which prevented the send. | |
2788 | */ | |
2789 | int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) | |
2790 | { | |
2791 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2792 | struct tcp_sock *tp = tcp_sk(sk); | |
2793 | unsigned int cur_mss; | |
2794 | int diff, len, err; | |
2795 | ||
2796 | ||
2797 | /* Inconclusive MTU probe */ | |
2798 | if (icsk->icsk_mtup.probe_size) | |
2799 | icsk->icsk_mtup.probe_size = 0; | |
2800 | ||
2801 | /* Do not sent more than we queued. 1/4 is reserved for possible | |
2802 | * copying overhead: fragmentation, tunneling, mangling etc. | |
2803 | */ | |
2804 | if (refcount_read(&sk->sk_wmem_alloc) > | |
2805 | min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), | |
2806 | sk->sk_sndbuf)) | |
2807 | return -EAGAIN; | |
2808 | ||
2809 | if (skb_still_in_host_queue(sk, skb)) | |
2810 | return -EBUSY; | |
2811 | ||
2812 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { | |
2813 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
2814 | BUG(); | |
2815 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) | |
2816 | return -ENOMEM; | |
2817 | } | |
2818 | ||
2819 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) | |
2820 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
2821 | ||
2822 | cur_mss = tcp_current_mss(sk); | |
2823 | ||
2824 | /* If receiver has shrunk his window, and skb is out of | |
2825 | * new window, do not retransmit it. The exception is the | |
2826 | * case, when window is shrunk to zero. In this case | |
2827 | * our retransmit serves as a zero window probe. | |
2828 | */ | |
2829 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && | |
2830 | TCP_SKB_CB(skb)->seq != tp->snd_una) | |
2831 | return -EAGAIN; | |
2832 | ||
2833 | len = cur_mss * segs; | |
2834 | if (skb->len > len) { | |
2835 | if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC)) | |
2836 | return -ENOMEM; /* We'll try again later. */ | |
2837 | } else { | |
2838 | if (skb_unclone(skb, GFP_ATOMIC)) | |
2839 | return -ENOMEM; | |
2840 | ||
2841 | diff = tcp_skb_pcount(skb); | |
2842 | tcp_set_skb_tso_segs(skb, cur_mss); | |
2843 | diff -= tcp_skb_pcount(skb); | |
2844 | if (diff) | |
2845 | tcp_adjust_pcount(sk, skb, diff); | |
2846 | if (skb->len < cur_mss) | |
2847 | tcp_retrans_try_collapse(sk, skb, cur_mss); | |
2848 | } | |
2849 | ||
2850 | /* RFC3168, section 6.1.1.1. ECN fallback */ | |
2851 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) | |
2852 | tcp_ecn_clear_syn(sk, skb); | |
2853 | ||
2854 | /* Update global and local TCP statistics. */ | |
2855 | segs = tcp_skb_pcount(skb); | |
2856 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); | |
2857 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) | |
2858 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); | |
2859 | tp->total_retrans += segs; | |
2860 | ||
2861 | /* make sure skb->data is aligned on arches that require it | |
2862 | * and check if ack-trimming & collapsing extended the headroom | |
2863 | * beyond what csum_start can cover. | |
2864 | */ | |
2865 | if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || | |
2866 | skb_headroom(skb) >= 0xFFFF)) { | |
2867 | struct sk_buff *nskb; | |
2868 | ||
2869 | skb->skb_mstamp = tp->tcp_mstamp; | |
2870 | nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); | |
2871 | err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : | |
2872 | -ENOBUFS; | |
2873 | } else { | |
2874 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); | |
2875 | } | |
2876 | ||
2877 | if (likely(!err)) { | |
2878 | TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; | |
2879 | } else if (err != -EBUSY) { | |
2880 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); | |
2881 | } | |
2882 | return err; | |
2883 | } | |
2884 | ||
2885 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) | |
2886 | { | |
2887 | struct tcp_sock *tp = tcp_sk(sk); | |
2888 | int err = __tcp_retransmit_skb(sk, skb, segs); | |
2889 | ||
2890 | if (err == 0) { | |
2891 | #if FASTRETRANS_DEBUG > 0 | |
2892 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { | |
2893 | net_dbg_ratelimited("retrans_out leaked\n"); | |
2894 | } | |
2895 | #endif | |
2896 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; | |
2897 | tp->retrans_out += tcp_skb_pcount(skb); | |
2898 | ||
2899 | /* Save stamp of the first retransmit. */ | |
2900 | if (!tp->retrans_stamp) | |
2901 | tp->retrans_stamp = tcp_skb_timestamp(skb); | |
2902 | ||
2903 | } | |
2904 | ||
2905 | if (tp->undo_retrans < 0) | |
2906 | tp->undo_retrans = 0; | |
2907 | tp->undo_retrans += tcp_skb_pcount(skb); | |
2908 | return err; | |
2909 | } | |
2910 | ||
2911 | /* This gets called after a retransmit timeout, and the initially | |
2912 | * retransmitted data is acknowledged. It tries to continue | |
2913 | * resending the rest of the retransmit queue, until either | |
2914 | * we've sent it all or the congestion window limit is reached. | |
2915 | * If doing SACK, the first ACK which comes back for a timeout | |
2916 | * based retransmit packet might feed us FACK information again. | |
2917 | * If so, we use it to avoid unnecessarily retransmissions. | |
2918 | */ | |
2919 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
2920 | { | |
2921 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
2922 | struct tcp_sock *tp = tcp_sk(sk); | |
2923 | struct sk_buff *skb; | |
2924 | struct sk_buff *hole = NULL; | |
2925 | u32 max_segs; | |
2926 | int mib_idx; | |
2927 | ||
2928 | if (!tp->packets_out) | |
2929 | return; | |
2930 | ||
2931 | if (tp->retransmit_skb_hint) { | |
2932 | skb = tp->retransmit_skb_hint; | |
2933 | } else { | |
2934 | skb = tcp_write_queue_head(sk); | |
2935 | } | |
2936 | ||
2937 | max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); | |
2938 | tcp_for_write_queue_from(skb, sk) { | |
2939 | __u8 sacked; | |
2940 | int segs; | |
2941 | ||
2942 | if (skb == tcp_send_head(sk)) | |
2943 | break; | |
2944 | ||
2945 | if (tcp_pacing_check(sk)) | |
2946 | break; | |
2947 | ||
2948 | /* we could do better than to assign each time */ | |
2949 | if (!hole) | |
2950 | tp->retransmit_skb_hint = skb; | |
2951 | ||
2952 | segs = tp->snd_cwnd - tcp_packets_in_flight(tp); | |
2953 | if (segs <= 0) | |
2954 | return; | |
2955 | sacked = TCP_SKB_CB(skb)->sacked; | |
2956 | /* In case tcp_shift_skb_data() have aggregated large skbs, | |
2957 | * we need to make sure not sending too bigs TSO packets | |
2958 | */ | |
2959 | segs = min_t(int, segs, max_segs); | |
2960 | ||
2961 | if (tp->retrans_out >= tp->lost_out) { | |
2962 | break; | |
2963 | } else if (!(sacked & TCPCB_LOST)) { | |
2964 | if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) | |
2965 | hole = skb; | |
2966 | continue; | |
2967 | ||
2968 | } else { | |
2969 | if (icsk->icsk_ca_state != TCP_CA_Loss) | |
2970 | mib_idx = LINUX_MIB_TCPFASTRETRANS; | |
2971 | else | |
2972 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; | |
2973 | } | |
2974 | ||
2975 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) | |
2976 | continue; | |
2977 | ||
2978 | if (tcp_small_queue_check(sk, skb, 1)) | |
2979 | return; | |
2980 | ||
2981 | if (tcp_retransmit_skb(sk, skb, segs)) | |
2982 | return; | |
2983 | ||
2984 | NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); | |
2985 | ||
2986 | if (tcp_in_cwnd_reduction(sk)) | |
2987 | tp->prr_out += tcp_skb_pcount(skb); | |
2988 | ||
2989 | if (skb == tcp_write_queue_head(sk) && | |
2990 | icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) | |
2991 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
2992 | inet_csk(sk)->icsk_rto, | |
2993 | TCP_RTO_MAX); | |
2994 | } | |
2995 | } | |
2996 | ||
2997 | /* We allow to exceed memory limits for FIN packets to expedite | |
2998 | * connection tear down and (memory) recovery. | |
2999 | * Otherwise tcp_send_fin() could be tempted to either delay FIN | |
3000 | * or even be forced to close flow without any FIN. | |
3001 | * In general, we want to allow one skb per socket to avoid hangs | |
3002 | * with edge trigger epoll() | |
3003 | */ | |
3004 | void sk_forced_mem_schedule(struct sock *sk, int size) | |
3005 | { | |
3006 | int amt; | |
3007 | ||
3008 | if (size <= sk->sk_forward_alloc) | |
3009 | return; | |
3010 | amt = sk_mem_pages(size); | |
3011 | sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; | |
3012 | sk_memory_allocated_add(sk, amt); | |
3013 | ||
3014 | if (mem_cgroup_sockets_enabled && sk->sk_memcg) | |
3015 | mem_cgroup_charge_skmem(sk->sk_memcg, amt); | |
3016 | } | |
3017 | ||
3018 | /* Send a FIN. The caller locks the socket for us. | |
3019 | * We should try to send a FIN packet really hard, but eventually give up. | |
3020 | */ | |
3021 | void tcp_send_fin(struct sock *sk) | |
3022 | { | |
3023 | struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); | |
3024 | struct tcp_sock *tp = tcp_sk(sk); | |
3025 | ||
3026 | /* Optimization, tack on the FIN if we have one skb in write queue and | |
3027 | * this skb was not yet sent, or we are under memory pressure. | |
3028 | * Note: in the latter case, FIN packet will be sent after a timeout, | |
3029 | * as TCP stack thinks it has already been transmitted. | |
3030 | */ | |
3031 | if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) { | |
3032 | coalesce: | |
3033 | TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; | |
3034 | TCP_SKB_CB(tskb)->end_seq++; | |
3035 | tp->write_seq++; | |
3036 | if (!tcp_send_head(sk)) { | |
3037 | /* This means tskb was already sent. | |
3038 | * Pretend we included the FIN on previous transmit. | |
3039 | * We need to set tp->snd_nxt to the value it would have | |
3040 | * if FIN had been sent. This is because retransmit path | |
3041 | * does not change tp->snd_nxt. | |
3042 | */ | |
3043 | tp->snd_nxt++; | |
3044 | return; | |
3045 | } | |
3046 | } else { | |
3047 | skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); | |
3048 | if (unlikely(!skb)) { | |
3049 | if (tskb) | |
3050 | goto coalesce; | |
3051 | return; | |
3052 | } | |
3053 | skb_reserve(skb, MAX_TCP_HEADER); | |
3054 | sk_forced_mem_schedule(sk, skb->truesize); | |
3055 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ | |
3056 | tcp_init_nondata_skb(skb, tp->write_seq, | |
3057 | TCPHDR_ACK | TCPHDR_FIN); | |
3058 | tcp_queue_skb(sk, skb); | |
3059 | } | |
3060 | __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); | |
3061 | } | |
3062 | ||
3063 | /* We get here when a process closes a file descriptor (either due to | |
3064 | * an explicit close() or as a byproduct of exit()'ing) and there | |
3065 | * was unread data in the receive queue. This behavior is recommended | |
3066 | * by RFC 2525, section 2.17. -DaveM | |
3067 | */ | |
3068 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) | |
3069 | { | |
3070 | struct sk_buff *skb; | |
3071 | ||
3072 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); | |
3073 | ||
3074 | /* NOTE: No TCP options attached and we never retransmit this. */ | |
3075 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
3076 | if (!skb) { | |
3077 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); | |
3078 | return; | |
3079 | } | |
3080 | ||
3081 | /* Reserve space for headers and prepare control bits. */ | |
3082 | skb_reserve(skb, MAX_TCP_HEADER); | |
3083 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), | |
3084 | TCPHDR_ACK | TCPHDR_RST); | |
3085 | tcp_mstamp_refresh(tcp_sk(sk)); | |
3086 | /* Send it off. */ | |
3087 | if (tcp_transmit_skb(sk, skb, 0, priority)) | |
3088 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); | |
3089 | } | |
3090 | ||
3091 | /* Send a crossed SYN-ACK during socket establishment. | |
3092 | * WARNING: This routine must only be called when we have already sent | |
3093 | * a SYN packet that crossed the incoming SYN that caused this routine | |
3094 | * to get called. If this assumption fails then the initial rcv_wnd | |
3095 | * and rcv_wscale values will not be correct. | |
3096 | */ | |
3097 | int tcp_send_synack(struct sock *sk) | |
3098 | { | |
3099 | struct sk_buff *skb; | |
3100 | ||
3101 | skb = tcp_write_queue_head(sk); | |
3102 | if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { | |
3103 | pr_debug("%s: wrong queue state\n", __func__); | |
3104 | return -EFAULT; | |
3105 | } | |
3106 | if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { | |
3107 | if (skb_cloned(skb)) { | |
3108 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); | |
3109 | if (!nskb) | |
3110 | return -ENOMEM; | |
3111 | tcp_unlink_write_queue(skb, sk); | |
3112 | __skb_header_release(nskb); | |
3113 | __tcp_add_write_queue_head(sk, nskb); | |
3114 | sk_wmem_free_skb(sk, skb); | |
3115 | sk->sk_wmem_queued += nskb->truesize; | |
3116 | sk_mem_charge(sk, nskb->truesize); | |
3117 | skb = nskb; | |
3118 | } | |
3119 | ||
3120 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; | |
3121 | tcp_ecn_send_synack(sk, skb); | |
3122 | } | |
3123 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); | |
3124 | } | |
3125 | ||
3126 | /** | |
3127 | * tcp_make_synack - Prepare a SYN-ACK. | |
3128 | * sk: listener socket | |
3129 | * dst: dst entry attached to the SYNACK | |
3130 | * req: request_sock pointer | |
3131 | * | |
3132 | * Allocate one skb and build a SYNACK packet. | |
3133 | * @dst is consumed : Caller should not use it again. | |
3134 | */ | |
3135 | struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, | |
3136 | struct request_sock *req, | |
3137 | struct tcp_fastopen_cookie *foc, | |
3138 | enum tcp_synack_type synack_type) | |
3139 | { | |
3140 | struct inet_request_sock *ireq = inet_rsk(req); | |
3141 | const struct tcp_sock *tp = tcp_sk(sk); | |
3142 | struct tcp_md5sig_key *md5 = NULL; | |
3143 | struct tcp_out_options opts; | |
3144 | struct sk_buff *skb; | |
3145 | int tcp_header_size; | |
3146 | struct tcphdr *th; | |
3147 | int mss; | |
3148 | ||
3149 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
3150 | if (unlikely(!skb)) { | |
3151 | dst_release(dst); | |
3152 | return NULL; | |
3153 | } | |
3154 | /* Reserve space for headers. */ | |
3155 | skb_reserve(skb, MAX_TCP_HEADER); | |
3156 | ||
3157 | switch (synack_type) { | |
3158 | case TCP_SYNACK_NORMAL: | |
3159 | skb_set_owner_w(skb, req_to_sk(req)); | |
3160 | break; | |
3161 | case TCP_SYNACK_COOKIE: | |
3162 | /* Under synflood, we do not attach skb to a socket, | |
3163 | * to avoid false sharing. | |
3164 | */ | |
3165 | break; | |
3166 | case TCP_SYNACK_FASTOPEN: | |
3167 | /* sk is a const pointer, because we want to express multiple | |
3168 | * cpu might call us concurrently. | |
3169 | * sk->sk_wmem_alloc in an atomic, we can promote to rw. | |
3170 | */ | |
3171 | skb_set_owner_w(skb, (struct sock *)sk); | |
3172 | break; | |
3173 | } | |
3174 | skb_dst_set(skb, dst); | |
3175 | ||
3176 | mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); | |
3177 | ||
3178 | memset(&opts, 0, sizeof(opts)); | |
3179 | #ifdef CONFIG_SYN_COOKIES | |
3180 | if (unlikely(req->cookie_ts)) | |
3181 | skb->skb_mstamp = cookie_init_timestamp(req); | |
3182 | else | |
3183 | #endif | |
3184 | skb->skb_mstamp = tcp_clock_us(); | |
3185 | ||
3186 | #ifdef CONFIG_TCP_MD5SIG | |
3187 | rcu_read_lock(); | |
3188 | md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); | |
3189 | #endif | |
3190 | skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); | |
3191 | tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) + | |
3192 | sizeof(*th); | |
3193 | ||
3194 | skb_push(skb, tcp_header_size); | |
3195 | skb_reset_transport_header(skb); | |
3196 | ||
3197 | th = (struct tcphdr *)skb->data; | |
3198 | memset(th, 0, sizeof(struct tcphdr)); | |
3199 | th->syn = 1; | |
3200 | th->ack = 1; | |
3201 | tcp_ecn_make_synack(req, th); | |
3202 | th->source = htons(ireq->ir_num); | |
3203 | th->dest = ireq->ir_rmt_port; | |
3204 | skb->mark = ireq->ir_mark; | |
3205 | /* Setting of flags are superfluous here for callers (and ECE is | |
3206 | * not even correctly set) | |
3207 | */ | |
3208 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, | |
3209 | TCPHDR_SYN | TCPHDR_ACK); | |
3210 | ||
3211 | th->seq = htonl(TCP_SKB_CB(skb)->seq); | |
3212 | /* XXX data is queued and acked as is. No buffer/window check */ | |
3213 | th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); | |
3214 | ||
3215 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
3216 | th->window = htons(min(req->rsk_rcv_wnd, 65535U)); | |
3217 | tcp_options_write((__be32 *)(th + 1), NULL, &opts); | |
3218 | th->doff = (tcp_header_size >> 2); | |
3219 | __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); | |
3220 | ||
3221 | #ifdef CONFIG_TCP_MD5SIG | |
3222 | /* Okay, we have all we need - do the md5 hash if needed */ | |
3223 | if (md5) | |
3224 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, | |
3225 | md5, req_to_sk(req), skb); | |
3226 | rcu_read_unlock(); | |
3227 | #endif | |
3228 | ||
3229 | /* Do not fool tcpdump (if any), clean our debris */ | |
3230 | skb->tstamp = 0; | |
3231 | return skb; | |
3232 | } | |
3233 | EXPORT_SYMBOL(tcp_make_synack); | |
3234 | ||
3235 | static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) | |
3236 | { | |
3237 | struct inet_connection_sock *icsk = inet_csk(sk); | |
3238 | const struct tcp_congestion_ops *ca; | |
3239 | u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); | |
3240 | ||
3241 | if (ca_key == TCP_CA_UNSPEC) | |
3242 | return; | |
3243 | ||
3244 | rcu_read_lock(); | |
3245 | ca = tcp_ca_find_key(ca_key); | |
3246 | if (likely(ca && try_module_get(ca->owner))) { | |
3247 | module_put(icsk->icsk_ca_ops->owner); | |
3248 | icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); | |
3249 | icsk->icsk_ca_ops = ca; | |
3250 | } | |
3251 | rcu_read_unlock(); | |
3252 | } | |
3253 | ||
3254 | /* Do all connect socket setups that can be done AF independent. */ | |
3255 | static void tcp_connect_init(struct sock *sk) | |
3256 | { | |
3257 | const struct dst_entry *dst = __sk_dst_get(sk); | |
3258 | struct tcp_sock *tp = tcp_sk(sk); | |
3259 | __u8 rcv_wscale; | |
3260 | u32 rcv_wnd; | |
3261 | ||
3262 | /* We'll fix this up when we get a response from the other end. | |
3263 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
3264 | */ | |
3265 | tp->tcp_header_len = sizeof(struct tcphdr); | |
3266 | if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) | |
3267 | tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; | |
3268 | ||
3269 | #ifdef CONFIG_TCP_MD5SIG | |
3270 | if (tp->af_specific->md5_lookup(sk, sk)) | |
3271 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; | |
3272 | #endif | |
3273 | ||
3274 | /* If user gave his TCP_MAXSEG, record it to clamp */ | |
3275 | if (tp->rx_opt.user_mss) | |
3276 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
3277 | tp->max_window = 0; | |
3278 | tcp_mtup_init(sk); | |
3279 | tcp_sync_mss(sk, dst_mtu(dst)); | |
3280 | ||
3281 | tcp_ca_dst_init(sk, dst); | |
3282 | ||
3283 | if (!tp->window_clamp) | |
3284 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
3285 | tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); | |
3286 | ||
3287 | tcp_initialize_rcv_mss(sk); | |
3288 | ||
3289 | /* limit the window selection if the user enforce a smaller rx buffer */ | |
3290 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
3291 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) | |
3292 | tp->window_clamp = tcp_full_space(sk); | |
3293 | ||
3294 | rcv_wnd = tcp_rwnd_init_bpf(sk); | |
3295 | if (rcv_wnd == 0) | |
3296 | rcv_wnd = dst_metric(dst, RTAX_INITRWND); | |
3297 | ||
3298 | tcp_select_initial_window(tcp_full_space(sk), | |
3299 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), | |
3300 | &tp->rcv_wnd, | |
3301 | &tp->window_clamp, | |
3302 | sock_net(sk)->ipv4.sysctl_tcp_window_scaling, | |
3303 | &rcv_wscale, | |
3304 | rcv_wnd); | |
3305 | ||
3306 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
3307 | tp->rcv_ssthresh = tp->rcv_wnd; | |
3308 | ||
3309 | sk->sk_err = 0; | |
3310 | sock_reset_flag(sk, SOCK_DONE); | |
3311 | tp->snd_wnd = 0; | |
3312 | tcp_init_wl(tp, 0); | |
3313 | tp->snd_una = tp->write_seq; | |
3314 | tp->snd_sml = tp->write_seq; | |
3315 | tp->snd_up = tp->write_seq; | |
3316 | tp->snd_nxt = tp->write_seq; | |
3317 | ||
3318 | if (likely(!tp->repair)) | |
3319 | tp->rcv_nxt = 0; | |
3320 | else | |
3321 | tp->rcv_tstamp = tcp_jiffies32; | |
3322 | tp->rcv_wup = tp->rcv_nxt; | |
3323 | tp->copied_seq = tp->rcv_nxt; | |
3324 | ||
3325 | inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); | |
3326 | inet_csk(sk)->icsk_retransmits = 0; | |
3327 | tcp_clear_retrans(tp); | |
3328 | } | |
3329 | ||
3330 | static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) | |
3331 | { | |
3332 | struct tcp_sock *tp = tcp_sk(sk); | |
3333 | struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); | |
3334 | ||
3335 | tcb->end_seq += skb->len; | |
3336 | __skb_header_release(skb); | |
3337 | __tcp_add_write_queue_tail(sk, skb); | |
3338 | sk->sk_wmem_queued += skb->truesize; | |
3339 | sk_mem_charge(sk, skb->truesize); | |
3340 | tp->write_seq = tcb->end_seq; | |
3341 | tp->packets_out += tcp_skb_pcount(skb); | |
3342 | } | |
3343 | ||
3344 | /* Build and send a SYN with data and (cached) Fast Open cookie. However, | |
3345 | * queue a data-only packet after the regular SYN, such that regular SYNs | |
3346 | * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges | |
3347 | * only the SYN sequence, the data are retransmitted in the first ACK. | |
3348 | * If cookie is not cached or other error occurs, falls back to send a | |
3349 | * regular SYN with Fast Open cookie request option. | |
3350 | */ | |
3351 | static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) | |
3352 | { | |
3353 | struct tcp_sock *tp = tcp_sk(sk); | |
3354 | struct tcp_fastopen_request *fo = tp->fastopen_req; | |
3355 | int space, err = 0; | |
3356 | struct sk_buff *syn_data; | |
3357 | ||
3358 | tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ | |
3359 | if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) | |
3360 | goto fallback; | |
3361 | ||
3362 | /* MSS for SYN-data is based on cached MSS and bounded by PMTU and | |
3363 | * user-MSS. Reserve maximum option space for middleboxes that add | |
3364 | * private TCP options. The cost is reduced data space in SYN :( | |
3365 | */ | |
3366 | tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); | |
3367 | ||
3368 | space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - | |
3369 | MAX_TCP_OPTION_SPACE; | |
3370 | ||
3371 | space = min_t(size_t, space, fo->size); | |
3372 | ||
3373 | /* limit to order-0 allocations */ | |
3374 | space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); | |
3375 | ||
3376 | syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); | |
3377 | if (!syn_data) | |
3378 | goto fallback; | |
3379 | syn_data->ip_summed = CHECKSUM_PARTIAL; | |
3380 | memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); | |
3381 | if (space) { | |
3382 | int copied = copy_from_iter(skb_put(syn_data, space), space, | |
3383 | &fo->data->msg_iter); | |
3384 | if (unlikely(!copied)) { | |
3385 | kfree_skb(syn_data); | |
3386 | goto fallback; | |
3387 | } | |
3388 | if (copied != space) { | |
3389 | skb_trim(syn_data, copied); | |
3390 | space = copied; | |
3391 | } | |
3392 | } | |
3393 | /* No more data pending in inet_wait_for_connect() */ | |
3394 | if (space == fo->size) | |
3395 | fo->data = NULL; | |
3396 | fo->copied = space; | |
3397 | ||
3398 | tcp_connect_queue_skb(sk, syn_data); | |
3399 | if (syn_data->len) | |
3400 | tcp_chrono_start(sk, TCP_CHRONO_BUSY); | |
3401 | ||
3402 | err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); | |
3403 | ||
3404 | syn->skb_mstamp = syn_data->skb_mstamp; | |
3405 | ||
3406 | /* Now full SYN+DATA was cloned and sent (or not), | |
3407 | * remove the SYN from the original skb (syn_data) | |
3408 | * we keep in write queue in case of a retransmit, as we | |
3409 | * also have the SYN packet (with no data) in the same queue. | |
3410 | */ | |
3411 | TCP_SKB_CB(syn_data)->seq++; | |
3412 | TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; | |
3413 | if (!err) { | |
3414 | tp->syn_data = (fo->copied > 0); | |
3415 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); | |
3416 | goto done; | |
3417 | } | |
3418 | ||
3419 | fallback: | |
3420 | /* Send a regular SYN with Fast Open cookie request option */ | |
3421 | if (fo->cookie.len > 0) | |
3422 | fo->cookie.len = 0; | |
3423 | err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); | |
3424 | if (err) | |
3425 | tp->syn_fastopen = 0; | |
3426 | done: | |
3427 | fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ | |
3428 | return err; | |
3429 | } | |
3430 | ||
3431 | /* Build a SYN and send it off. */ | |
3432 | int tcp_connect(struct sock *sk) | |
3433 | { | |
3434 | struct tcp_sock *tp = tcp_sk(sk); | |
3435 | struct sk_buff *buff; | |
3436 | int err; | |
3437 | ||
3438 | tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB); | |
3439 | ||
3440 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) | |
3441 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
3442 | ||
3443 | tcp_connect_init(sk); | |
3444 | ||
3445 | if (unlikely(tp->repair)) { | |
3446 | tcp_finish_connect(sk, NULL); | |
3447 | return 0; | |
3448 | } | |
3449 | ||
3450 | buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); | |
3451 | if (unlikely(!buff)) | |
3452 | return -ENOBUFS; | |
3453 | ||
3454 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); | |
3455 | tcp_mstamp_refresh(tp); | |
3456 | tp->retrans_stamp = tcp_time_stamp(tp); | |
3457 | tcp_connect_queue_skb(sk, buff); | |
3458 | tcp_ecn_send_syn(sk, buff); | |
3459 | ||
3460 | /* Send off SYN; include data in Fast Open. */ | |
3461 | err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : | |
3462 | tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); | |
3463 | if (err == -ECONNREFUSED) | |
3464 | return err; | |
3465 | ||
3466 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
3467 | * in order to make this packet get counted in tcpOutSegs. | |
3468 | */ | |
3469 | tp->snd_nxt = tp->write_seq; | |
3470 | tp->pushed_seq = tp->write_seq; | |
3471 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); | |
3472 | ||
3473 | /* Timer for repeating the SYN until an answer. */ | |
3474 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
3475 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
3476 | return 0; | |
3477 | } | |
3478 | EXPORT_SYMBOL(tcp_connect); | |
3479 | ||
3480 | /* Send out a delayed ack, the caller does the policy checking | |
3481 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
3482 | * for details. | |
3483 | */ | |
3484 | void tcp_send_delayed_ack(struct sock *sk) | |
3485 | { | |
3486 | struct inet_connection_sock *icsk = inet_csk(sk); | |
3487 | int ato = icsk->icsk_ack.ato; | |
3488 | unsigned long timeout; | |
3489 | ||
3490 | tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); | |
3491 | ||
3492 | if (ato > TCP_DELACK_MIN) { | |
3493 | const struct tcp_sock *tp = tcp_sk(sk); | |
3494 | int max_ato = HZ / 2; | |
3495 | ||
3496 | if (icsk->icsk_ack.pingpong || | |
3497 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) | |
3498 | max_ato = TCP_DELACK_MAX; | |
3499 | ||
3500 | /* Slow path, intersegment interval is "high". */ | |
3501 | ||
3502 | /* If some rtt estimate is known, use it to bound delayed ack. | |
3503 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements | |
3504 | * directly. | |
3505 | */ | |
3506 | if (tp->srtt_us) { | |
3507 | int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), | |
3508 | TCP_DELACK_MIN); | |
3509 | ||
3510 | if (rtt < max_ato) | |
3511 | max_ato = rtt; | |
3512 | } | |
3513 | ||
3514 | ato = min(ato, max_ato); | |
3515 | } | |
3516 | ||
3517 | /* Stay within the limit we were given */ | |
3518 | timeout = jiffies + ato; | |
3519 | ||
3520 | /* Use new timeout only if there wasn't a older one earlier. */ | |
3521 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { | |
3522 | /* If delack timer was blocked or is about to expire, | |
3523 | * send ACK now. | |
3524 | */ | |
3525 | if (icsk->icsk_ack.blocked || | |
3526 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
3527 | tcp_send_ack(sk); | |
3528 | return; | |
3529 | } | |
3530 | ||
3531 | if (!time_before(timeout, icsk->icsk_ack.timeout)) | |
3532 | timeout = icsk->icsk_ack.timeout; | |
3533 | } | |
3534 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; | |
3535 | icsk->icsk_ack.timeout = timeout; | |
3536 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
3537 | } | |
3538 | ||
3539 | /* This routine sends an ack and also updates the window. */ | |
3540 | void tcp_send_ack(struct sock *sk) | |
3541 | { | |
3542 | struct sk_buff *buff; | |
3543 | ||
3544 | /* If we have been reset, we may not send again. */ | |
3545 | if (sk->sk_state == TCP_CLOSE) | |
3546 | return; | |
3547 | ||
3548 | tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); | |
3549 | ||
3550 | /* We are not putting this on the write queue, so | |
3551 | * tcp_transmit_skb() will set the ownership to this | |
3552 | * sock. | |
3553 | */ | |
3554 | buff = alloc_skb(MAX_TCP_HEADER, | |
3555 | sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); | |
3556 | if (unlikely(!buff)) { | |
3557 | inet_csk_schedule_ack(sk); | |
3558 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; | |
3559 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, | |
3560 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
3561 | return; | |
3562 | } | |
3563 | ||
3564 | /* Reserve space for headers and prepare control bits. */ | |
3565 | skb_reserve(buff, MAX_TCP_HEADER); | |
3566 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); | |
3567 | ||
3568 | /* We do not want pure acks influencing TCP Small Queues or fq/pacing | |
3569 | * too much. | |
3570 | * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 | |
3571 | */ | |
3572 | skb_set_tcp_pure_ack(buff); | |
3573 | ||
3574 | /* Send it off, this clears delayed acks for us. */ | |
3575 | tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); | |
3576 | } | |
3577 | EXPORT_SYMBOL_GPL(tcp_send_ack); | |
3578 | ||
3579 | /* This routine sends a packet with an out of date sequence | |
3580 | * number. It assumes the other end will try to ack it. | |
3581 | * | |
3582 | * Question: what should we make while urgent mode? | |
3583 | * 4.4BSD forces sending single byte of data. We cannot send | |
3584 | * out of window data, because we have SND.NXT==SND.MAX... | |
3585 | * | |
3586 | * Current solution: to send TWO zero-length segments in urgent mode: | |
3587 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
3588 | * out-of-date with SND.UNA-1 to probe window. | |
3589 | */ | |
3590 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) | |
3591 | { | |
3592 | struct tcp_sock *tp = tcp_sk(sk); | |
3593 | struct sk_buff *skb; | |
3594 | ||
3595 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
3596 | skb = alloc_skb(MAX_TCP_HEADER, | |
3597 | sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); | |
3598 | if (!skb) | |
3599 | return -1; | |
3600 | ||
3601 | /* Reserve space for headers and set control bits. */ | |
3602 | skb_reserve(skb, MAX_TCP_HEADER); | |
3603 | /* Use a previous sequence. This should cause the other | |
3604 | * end to send an ack. Don't queue or clone SKB, just | |
3605 | * send it. | |
3606 | */ | |
3607 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); | |
3608 | NET_INC_STATS(sock_net(sk), mib); | |
3609 | return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); | |
3610 | } | |
3611 | ||
3612 | /* Called from setsockopt( ... TCP_REPAIR ) */ | |
3613 | void tcp_send_window_probe(struct sock *sk) | |
3614 | { | |
3615 | if (sk->sk_state == TCP_ESTABLISHED) { | |
3616 | tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; | |
3617 | tcp_mstamp_refresh(tcp_sk(sk)); | |
3618 | tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); | |
3619 | } | |
3620 | } | |
3621 | ||
3622 | /* Initiate keepalive or window probe from timer. */ | |
3623 | int tcp_write_wakeup(struct sock *sk, int mib) | |
3624 | { | |
3625 | struct tcp_sock *tp = tcp_sk(sk); | |
3626 | struct sk_buff *skb; | |
3627 | ||
3628 | if (sk->sk_state == TCP_CLOSE) | |
3629 | return -1; | |
3630 | ||
3631 | skb = tcp_send_head(sk); | |
3632 | if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { | |
3633 | int err; | |
3634 | unsigned int mss = tcp_current_mss(sk); | |
3635 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; | |
3636 | ||
3637 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
3638 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
3639 | ||
3640 | /* We are probing the opening of a window | |
3641 | * but the window size is != 0 | |
3642 | * must have been a result SWS avoidance ( sender ) | |
3643 | */ | |
3644 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
3645 | skb->len > mss) { | |
3646 | seg_size = min(seg_size, mss); | |
3647 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; | |
3648 | if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) | |
3649 | return -1; | |
3650 | } else if (!tcp_skb_pcount(skb)) | |
3651 | tcp_set_skb_tso_segs(skb, mss); | |
3652 | ||
3653 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; | |
3654 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); | |
3655 | if (!err) | |
3656 | tcp_event_new_data_sent(sk, skb); | |
3657 | return err; | |
3658 | } else { | |
3659 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) | |
3660 | tcp_xmit_probe_skb(sk, 1, mib); | |
3661 | return tcp_xmit_probe_skb(sk, 0, mib); | |
3662 | } | |
3663 | } | |
3664 | ||
3665 | /* A window probe timeout has occurred. If window is not closed send | |
3666 | * a partial packet else a zero probe. | |
3667 | */ | |
3668 | void tcp_send_probe0(struct sock *sk) | |
3669 | { | |
3670 | struct inet_connection_sock *icsk = inet_csk(sk); | |
3671 | struct tcp_sock *tp = tcp_sk(sk); | |
3672 | struct net *net = sock_net(sk); | |
3673 | unsigned long probe_max; | |
3674 | int err; | |
3675 | ||
3676 | err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); | |
3677 | ||
3678 | if (tp->packets_out || !tcp_send_head(sk)) { | |
3679 | /* Cancel probe timer, if it is not required. */ | |
3680 | icsk->icsk_probes_out = 0; | |
3681 | icsk->icsk_backoff = 0; | |
3682 | return; | |
3683 | } | |
3684 | ||
3685 | if (err <= 0) { | |
3686 | if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) | |
3687 | icsk->icsk_backoff++; | |
3688 | icsk->icsk_probes_out++; | |
3689 | probe_max = TCP_RTO_MAX; | |
3690 | } else { | |
3691 | /* If packet was not sent due to local congestion, | |
3692 | * do not backoff and do not remember icsk_probes_out. | |
3693 | * Let local senders to fight for local resources. | |
3694 | * | |
3695 | * Use accumulated backoff yet. | |
3696 | */ | |
3697 | if (!icsk->icsk_probes_out) | |
3698 | icsk->icsk_probes_out = 1; | |
3699 | probe_max = TCP_RESOURCE_PROBE_INTERVAL; | |
3700 | } | |
3701 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, | |
3702 | tcp_probe0_when(sk, probe_max), | |
3703 | TCP_RTO_MAX); | |
3704 | } | |
3705 | ||
3706 | int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) | |
3707 | { | |
3708 | const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; | |
3709 | struct flowi fl; | |
3710 | int res; | |
3711 | ||
3712 | tcp_rsk(req)->txhash = net_tx_rndhash(); | |
3713 | res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); | |
3714 | if (!res) { | |
3715 | __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); | |
3716 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); | |
3717 | if (unlikely(tcp_passive_fastopen(sk))) | |
3718 | tcp_sk(sk)->total_retrans++; | |
3719 | } | |
3720 | return res; | |
3721 | } | |
3722 | EXPORT_SYMBOL(tcp_rtx_synack); |