]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * INET An implementation of the TCP/IP protocol suite for the LINUX | |
3 | * operating system. INET is implemented using the BSD Socket | |
4 | * interface as the means of communication with the user level. | |
5 | * | |
6 | * Implementation of the Transmission Control Protocol(TCP). | |
7 | * | |
02c30a84 | 8 | * Authors: Ross Biro |
1da177e4 LT |
9 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
10 | * Mark Evans, <evansmp@uhura.aston.ac.uk> | |
11 | * Corey Minyard <wf-rch!minyard@relay.EU.net> | |
12 | * Florian La Roche, <flla@stud.uni-sb.de> | |
13 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> | |
14 | * Linus Torvalds, <torvalds@cs.helsinki.fi> | |
15 | * Alan Cox, <gw4pts@gw4pts.ampr.org> | |
16 | * Matthew Dillon, <dillon@apollo.west.oic.com> | |
17 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> | |
18 | * Jorge Cwik, <jorge@laser.satlink.net> | |
19 | */ | |
20 | ||
21 | /* | |
22 | * Changes: 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 | ||
91df42be JP |
37 | #define pr_fmt(fmt) "TCP: " fmt |
38 | ||
1da177e4 LT |
39 | #include <net/tcp.h> |
40 | ||
41 | #include <linux/compiler.h> | |
5a0e3ad6 | 42 | #include <linux/gfp.h> |
1da177e4 | 43 | #include <linux/module.h> |
60e2a778 | 44 | #include <linux/static_key.h> |
1da177e4 | 45 | |
e086101b | 46 | #include <trace/events/tcp.h> |
35089bb2 | 47 | |
46d3ceab ED |
48 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
49 | int push_one, gfp_t gfp); | |
519855c5 | 50 | |
67edfef7 | 51 | /* Account for new data that has been sent to the network. */ |
75c119af | 52 | static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 53 | { |
6ba8a3b1 | 54 | struct inet_connection_sock *icsk = inet_csk(sk); |
9e412ba7 | 55 | struct tcp_sock *tp = tcp_sk(sk); |
66f5fe62 | 56 | unsigned int prior_packets = tp->packets_out; |
9e412ba7 | 57 | |
1da177e4 | 58 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
8512430e | 59 | |
75c119af ED |
60 | __skb_unlink(skb, &sk->sk_write_queue); |
61 | tcp_rbtree_insert(&sk->tcp_rtx_queue, skb); | |
62 | ||
66f5fe62 | 63 | tp->packets_out += tcp_skb_pcount(skb); |
bec41a11 | 64 | if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) |
750ea2ba | 65 | tcp_rearm_rto(sk); |
f19c29e3 | 66 | |
f7324acd DM |
67 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, |
68 | tcp_skb_pcount(skb)); | |
1da177e4 LT |
69 | } |
70 | ||
a4ecb15a CC |
71 | /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one |
72 | * window scaling factor due to loss of precision. | |
1da177e4 LT |
73 | * If window has been shrunk, what should we make? It is not clear at all. |
74 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
75 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
76 | * invalid. OK, let's make this for now: | |
77 | */ | |
cf533ea5 | 78 | static inline __u32 tcp_acceptable_seq(const struct sock *sk) |
1da177e4 | 79 | { |
cf533ea5 | 80 | const struct tcp_sock *tp = tcp_sk(sk); |
9e412ba7 | 81 | |
a4ecb15a CC |
82 | if (!before(tcp_wnd_end(tp), tp->snd_nxt) || |
83 | (tp->rx_opt.wscale_ok && | |
84 | ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) | |
1da177e4 LT |
85 | return tp->snd_nxt; |
86 | else | |
90840def | 87 | return tcp_wnd_end(tp); |
1da177e4 LT |
88 | } |
89 | ||
90 | /* Calculate mss to advertise in SYN segment. | |
91 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
92 | * | |
93 | * 1. It is independent of path mtu. | |
94 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
95 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
96 | * attached devices, because some buggy hosts are confused by | |
97 | * large MSS. | |
98 | * 4. We do not make 3, we advertise MSS, calculated from first | |
99 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
100 | * This may be overridden via information stored in routing table. | |
101 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
102 | * probably even Jumbo". | |
103 | */ | |
104 | static __u16 tcp_advertise_mss(struct sock *sk) | |
105 | { | |
106 | struct tcp_sock *tp = tcp_sk(sk); | |
cf533ea5 | 107 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
108 | int mss = tp->advmss; |
109 | ||
0dbaee3b DM |
110 | if (dst) { |
111 | unsigned int metric = dst_metric_advmss(dst); | |
112 | ||
113 | if (metric < mss) { | |
114 | mss = metric; | |
115 | tp->advmss = mss; | |
116 | } | |
1da177e4 LT |
117 | } |
118 | ||
119 | return (__u16)mss; | |
120 | } | |
121 | ||
122 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
6f021c62 ED |
123 | * This is the first part of cwnd validation mechanism. |
124 | */ | |
125 | void tcp_cwnd_restart(struct sock *sk, s32 delta) | |
1da177e4 | 126 | { |
463c84b9 | 127 | struct tcp_sock *tp = tcp_sk(sk); |
6f021c62 | 128 | u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); |
1da177e4 LT |
129 | u32 cwnd = tp->snd_cwnd; |
130 | ||
6687e988 | 131 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 132 | |
6687e988 | 133 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
134 | restart_cwnd = min(restart_cwnd, cwnd); |
135 | ||
463c84b9 | 136 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 LT |
137 | cwnd >>= 1; |
138 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
c2203cf7 | 139 | tp->snd_cwnd_stamp = tcp_jiffies32; |
1da177e4 LT |
140 | tp->snd_cwnd_used = 0; |
141 | } | |
142 | ||
67edfef7 | 143 | /* Congestion state accounting after a packet has been sent. */ |
40efc6fa | 144 | static void tcp_event_data_sent(struct tcp_sock *tp, |
cf533ea5 | 145 | struct sock *sk) |
1da177e4 | 146 | { |
463c84b9 | 147 | struct inet_connection_sock *icsk = inet_csk(sk); |
d635fbe2 | 148 | const u32 now = tcp_jiffies32; |
1da177e4 | 149 | |
05c5a46d NC |
150 | if (tcp_packets_in_flight(tp) == 0) |
151 | tcp_ca_event(sk, CA_EVENT_TX_START); | |
152 | ||
1da177e4 LT |
153 | tp->lsndtime = now; |
154 | ||
155 | /* If it is a reply for ato after last received | |
156 | * packet, enter pingpong mode. | |
157 | */ | |
2251ae46 JM |
158 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
159 | icsk->icsk_ack.pingpong = 1; | |
1da177e4 LT |
160 | } |
161 | ||
67edfef7 | 162 | /* Account for an ACK we sent. */ |
deb42350 YC |
163 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts, |
164 | u32 rcv_nxt) | |
1da177e4 | 165 | { |
deb42350 YC |
166 | struct tcp_sock *tp = tcp_sk(sk); |
167 | ||
168 | if (unlikely(rcv_nxt != tp->rcv_nxt)) | |
169 | return; /* Special ACK sent by DCTCP to reflect ECN */ | |
463c84b9 ACM |
170 | tcp_dec_quickack_mode(sk, pkts); |
171 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
1da177e4 LT |
172 | } |
173 | ||
174 | /* Determine a window scaling and initial window to offer. | |
175 | * Based on the assumption that the given amount of space | |
176 | * will be offered. Store the results in the tp structure. | |
177 | * NOTE: for smooth operation initial space offering should | |
178 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
179 | * This MUST be enforced by all callers. | |
180 | */ | |
ceef9ab6 | 181 | void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss, |
1da177e4 | 182 | __u32 *rcv_wnd, __u32 *window_clamp, |
31d12926 | 183 | int wscale_ok, __u8 *rcv_wscale, |
184 | __u32 init_rcv_wnd) | |
1da177e4 LT |
185 | { |
186 | unsigned int space = (__space < 0 ? 0 : __space); | |
187 | ||
188 | /* If no clamp set the clamp to the max possible scaled window */ | |
189 | if (*window_clamp == 0) | |
589c49cb | 190 | (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); |
1da177e4 LT |
191 | space = min(*window_clamp, space); |
192 | ||
193 | /* Quantize space offering to a multiple of mss if possible. */ | |
194 | if (space > mss) | |
589c49cb | 195 | space = rounddown(space, mss); |
1da177e4 LT |
196 | |
197 | /* NOTE: offering an initial window larger than 32767 | |
15d99e02 RJ |
198 | * will break some buggy TCP stacks. If the admin tells us |
199 | * it is likely we could be speaking with such a buggy stack | |
200 | * we will truncate our initial window offering to 32K-1 | |
201 | * unless the remote has sent us a window scaling option, | |
202 | * which we interpret as a sign the remote TCP is not | |
203 | * misinterpreting the window field as a signed quantity. | |
1da177e4 | 204 | */ |
ceef9ab6 | 205 | if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) |
15d99e02 RJ |
206 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); |
207 | else | |
3ed06b27 YC |
208 | (*rcv_wnd) = min_t(u32, space, U16_MAX); |
209 | ||
210 | if (init_rcv_wnd) | |
211 | *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); | |
15d99e02 | 212 | |
1da177e4 LT |
213 | (*rcv_wscale) = 0; |
214 | if (wscale_ok) { | |
589c49cb | 215 | /* Set window scaling on max possible window */ |
356d1833 | 216 | space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); |
f626300a | 217 | space = max_t(u32, space, sysctl_rmem_max); |
316c1592 | 218 | space = min_t(u32, space, *window_clamp); |
589c49cb | 219 | while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) { |
1da177e4 LT |
220 | space >>= 1; |
221 | (*rcv_wscale)++; | |
222 | } | |
223 | } | |
224 | ||
1da177e4 | 225 | /* Set the clamp no higher than max representable value */ |
589c49cb | 226 | (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); |
1da177e4 | 227 | } |
4bc2f18b | 228 | EXPORT_SYMBOL(tcp_select_initial_window); |
1da177e4 LT |
229 | |
230 | /* Chose a new window to advertise, update state in tcp_sock for the | |
231 | * socket, and return result with RFC1323 scaling applied. The return | |
232 | * value can be stuffed directly into th->window for an outgoing | |
233 | * frame. | |
234 | */ | |
40efc6fa | 235 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
236 | { |
237 | struct tcp_sock *tp = tcp_sk(sk); | |
8e165e20 | 238 | u32 old_win = tp->rcv_wnd; |
1da177e4 LT |
239 | u32 cur_win = tcp_receive_window(tp); |
240 | u32 new_win = __tcp_select_window(sk); | |
241 | ||
242 | /* Never shrink the offered window */ | |
2de979bd | 243 | if (new_win < cur_win) { |
1da177e4 LT |
244 | /* Danger Will Robinson! |
245 | * Don't update rcv_wup/rcv_wnd here or else | |
246 | * we will not be able to advertise a zero | |
247 | * window in time. --DaveM | |
248 | * | |
249 | * Relax Will Robinson. | |
250 | */ | |
8e165e20 FW |
251 | if (new_win == 0) |
252 | NET_INC_STATS(sock_net(sk), | |
253 | LINUX_MIB_TCPWANTZEROWINDOWADV); | |
607bfbf2 | 254 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
1da177e4 LT |
255 | } |
256 | tp->rcv_wnd = new_win; | |
257 | tp->rcv_wup = tp->rcv_nxt; | |
258 | ||
259 | /* Make sure we do not exceed the maximum possible | |
260 | * scaled window. | |
261 | */ | |
ceef9ab6 ED |
262 | if (!tp->rx_opt.rcv_wscale && |
263 | sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) | |
1da177e4 LT |
264 | new_win = min(new_win, MAX_TCP_WINDOW); |
265 | else | |
266 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
267 | ||
268 | /* RFC1323 scaling applied */ | |
269 | new_win >>= tp->rx_opt.rcv_wscale; | |
270 | ||
31770e34 | 271 | /* If we advertise zero window, disable fast path. */ |
8e165e20 | 272 | if (new_win == 0) { |
31770e34 | 273 | tp->pred_flags = 0; |
8e165e20 FW |
274 | if (old_win) |
275 | NET_INC_STATS(sock_net(sk), | |
276 | LINUX_MIB_TCPTOZEROWINDOWADV); | |
277 | } else if (old_win == 0) { | |
278 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); | |
279 | } | |
1da177e4 LT |
280 | |
281 | return new_win; | |
282 | } | |
283 | ||
67edfef7 | 284 | /* Packet ECN state for a SYN-ACK */ |
735d3831 | 285 | static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) |
bdf1ee5d | 286 | { |
30e502a3 DB |
287 | const struct tcp_sock *tp = tcp_sk(sk); |
288 | ||
4de075e0 | 289 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; |
056834d9 | 290 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
4de075e0 | 291 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; |
91b5b21c LB |
292 | else if (tcp_ca_needs_ecn(sk) || |
293 | tcp_bpf_ca_needs_ecn(sk)) | |
30e502a3 | 294 | INET_ECN_xmit(sk); |
bdf1ee5d IJ |
295 | } |
296 | ||
67edfef7 | 297 | /* Packet ECN state for a SYN. */ |
735d3831 | 298 | static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) |
bdf1ee5d IJ |
299 | { |
300 | struct tcp_sock *tp = tcp_sk(sk); | |
91b5b21c | 301 | bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); |
f7b3bec6 | 302 | bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || |
91b5b21c | 303 | tcp_ca_needs_ecn(sk) || bpf_needs_ecn; |
f7b3bec6 FW |
304 | |
305 | if (!use_ecn) { | |
306 | const struct dst_entry *dst = __sk_dst_get(sk); | |
307 | ||
308 | if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) | |
309 | use_ecn = true; | |
310 | } | |
bdf1ee5d IJ |
311 | |
312 | tp->ecn_flags = 0; | |
f7b3bec6 FW |
313 | |
314 | if (use_ecn) { | |
4de075e0 | 315 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; |
bdf1ee5d | 316 | tp->ecn_flags = TCP_ECN_OK; |
91b5b21c | 317 | if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) |
30e502a3 | 318 | INET_ECN_xmit(sk); |
bdf1ee5d IJ |
319 | } |
320 | } | |
321 | ||
49213555 DB |
322 | static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) |
323 | { | |
324 | if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) | |
325 | /* tp->ecn_flags are cleared at a later point in time when | |
326 | * SYN ACK is ultimatively being received. | |
327 | */ | |
328 | TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); | |
329 | } | |
330 | ||
735d3831 | 331 | static void |
6ac705b1 | 332 | tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) |
bdf1ee5d | 333 | { |
6ac705b1 | 334 | if (inet_rsk(req)->ecn_ok) |
bdf1ee5d IJ |
335 | th->ece = 1; |
336 | } | |
337 | ||
67edfef7 AK |
338 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to |
339 | * be sent. | |
340 | */ | |
735d3831 | 341 | static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, |
ea1627c2 | 342 | struct tcphdr *th, int tcp_header_len) |
bdf1ee5d IJ |
343 | { |
344 | struct tcp_sock *tp = tcp_sk(sk); | |
345 | ||
346 | if (tp->ecn_flags & TCP_ECN_OK) { | |
347 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
348 | if (skb->len != tcp_header_len && | |
349 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
350 | INET_ECN_xmit(sk); | |
056834d9 | 351 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
bdf1ee5d | 352 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
ea1627c2 | 353 | th->cwr = 1; |
bdf1ee5d IJ |
354 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; |
355 | } | |
30e502a3 | 356 | } else if (!tcp_ca_needs_ecn(sk)) { |
bdf1ee5d IJ |
357 | /* ACK or retransmitted segment: clear ECT|CE */ |
358 | INET_ECN_dontxmit(sk); | |
359 | } | |
360 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
ea1627c2 | 361 | th->ece = 1; |
bdf1ee5d IJ |
362 | } |
363 | } | |
364 | ||
e870a8ef IJ |
365 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
366 | * auto increment end seqno. | |
367 | */ | |
368 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) | |
369 | { | |
2e8e18ef | 370 | skb->ip_summed = CHECKSUM_PARTIAL; |
e870a8ef | 371 | |
4de075e0 | 372 | TCP_SKB_CB(skb)->tcp_flags = flags; |
e870a8ef IJ |
373 | TCP_SKB_CB(skb)->sacked = 0; |
374 | ||
cd7d8498 | 375 | tcp_skb_pcount_set(skb, 1); |
e870a8ef IJ |
376 | |
377 | TCP_SKB_CB(skb)->seq = seq; | |
a3433f35 | 378 | if (flags & (TCPHDR_SYN | TCPHDR_FIN)) |
e870a8ef IJ |
379 | seq++; |
380 | TCP_SKB_CB(skb)->end_seq = seq; | |
381 | } | |
382 | ||
a2a385d6 | 383 | static inline bool tcp_urg_mode(const struct tcp_sock *tp) |
33f5f57e IJ |
384 | { |
385 | return tp->snd_una != tp->snd_up; | |
386 | } | |
387 | ||
33ad798c AL |
388 | #define OPTION_SACK_ADVERTISE (1 << 0) |
389 | #define OPTION_TS (1 << 1) | |
390 | #define OPTION_MD5 (1 << 2) | |
89e95a61 | 391 | #define OPTION_WSCALE (1 << 3) |
2100c8d2 | 392 | #define OPTION_FAST_OPEN_COOKIE (1 << 8) |
60e2a778 UB |
393 | #define OPTION_SMC (1 << 9) |
394 | ||
395 | static void smc_options_write(__be32 *ptr, u16 *options) | |
396 | { | |
397 | #if IS_ENABLED(CONFIG_SMC) | |
398 | if (static_branch_unlikely(&tcp_have_smc)) { | |
399 | if (unlikely(OPTION_SMC & *options)) { | |
400 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
401 | (TCPOPT_NOP << 16) | | |
402 | (TCPOPT_EXP << 8) | | |
403 | (TCPOLEN_EXP_SMC_BASE)); | |
404 | *ptr++ = htonl(TCPOPT_SMC_MAGIC); | |
405 | } | |
406 | } | |
407 | #endif | |
408 | } | |
33ad798c AL |
409 | |
410 | struct tcp_out_options { | |
2100c8d2 YC |
411 | u16 options; /* bit field of OPTION_* */ |
412 | u16 mss; /* 0 to disable */ | |
33ad798c AL |
413 | u8 ws; /* window scale, 0 to disable */ |
414 | u8 num_sack_blocks; /* number of SACK blocks to include */ | |
bd0388ae | 415 | u8 hash_size; /* bytes in hash_location */ |
bd0388ae | 416 | __u8 *hash_location; /* temporary pointer, overloaded */ |
2100c8d2 YC |
417 | __u32 tsval, tsecr; /* need to include OPTION_TS */ |
418 | struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ | |
33ad798c AL |
419 | }; |
420 | ||
67edfef7 AK |
421 | /* Write previously computed TCP options to the packet. |
422 | * | |
423 | * Beware: Something in the Internet is very sensitive to the ordering of | |
fd6149d3 IJ |
424 | * TCP options, we learned this through the hard way, so be careful here. |
425 | * Luckily we can at least blame others for their non-compliance but from | |
8e3bff96 | 426 | * inter-operability perspective it seems that we're somewhat stuck with |
fd6149d3 IJ |
427 | * the ordering which we have been using if we want to keep working with |
428 | * those broken things (not that it currently hurts anybody as there isn't | |
429 | * particular reason why the ordering would need to be changed). | |
430 | * | |
431 | * At least SACK_PERM as the first option is known to lead to a disaster | |
432 | * (but it may well be that other scenarios fail similarly). | |
433 | */ | |
33ad798c | 434 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, |
bd0388ae WAS |
435 | struct tcp_out_options *opts) |
436 | { | |
2100c8d2 | 437 | u16 options = opts->options; /* mungable copy */ |
bd0388ae | 438 | |
bd0388ae | 439 | if (unlikely(OPTION_MD5 & options)) { |
1a2c6181 CP |
440 | *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
441 | (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); | |
bd0388ae WAS |
442 | /* overload cookie hash location */ |
443 | opts->hash_location = (__u8 *)ptr; | |
33ad798c | 444 | ptr += 4; |
40efc6fa | 445 | } |
33ad798c | 446 | |
fd6149d3 IJ |
447 | if (unlikely(opts->mss)) { |
448 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
449 | (TCPOLEN_MSS << 16) | | |
450 | opts->mss); | |
451 | } | |
452 | ||
bd0388ae WAS |
453 | if (likely(OPTION_TS & options)) { |
454 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
455 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
456 | (TCPOLEN_SACK_PERM << 16) | | |
457 | (TCPOPT_TIMESTAMP << 8) | | |
458 | TCPOLEN_TIMESTAMP); | |
bd0388ae | 459 | options &= ~OPTION_SACK_ADVERTISE; |
33ad798c AL |
460 | } else { |
461 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
462 | (TCPOPT_NOP << 16) | | |
463 | (TCPOPT_TIMESTAMP << 8) | | |
464 | TCPOLEN_TIMESTAMP); | |
465 | } | |
466 | *ptr++ = htonl(opts->tsval); | |
467 | *ptr++ = htonl(opts->tsecr); | |
468 | } | |
469 | ||
bd0388ae | 470 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { |
33ad798c AL |
471 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
472 | (TCPOPT_NOP << 16) | | |
473 | (TCPOPT_SACK_PERM << 8) | | |
474 | TCPOLEN_SACK_PERM); | |
475 | } | |
476 | ||
bd0388ae | 477 | if (unlikely(OPTION_WSCALE & options)) { |
33ad798c AL |
478 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
479 | (TCPOPT_WINDOW << 16) | | |
480 | (TCPOLEN_WINDOW << 8) | | |
481 | opts->ws); | |
482 | } | |
483 | ||
484 | if (unlikely(opts->num_sack_blocks)) { | |
485 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
486 | tp->duplicate_sack : tp->selective_acks; | |
40efc6fa SH |
487 | int this_sack; |
488 | ||
489 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
490 | (TCPOPT_NOP << 16) | | |
491 | (TCPOPT_SACK << 8) | | |
33ad798c | 492 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
40efc6fa | 493 | TCPOLEN_SACK_PERBLOCK))); |
2de979bd | 494 | |
33ad798c AL |
495 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
496 | ++this_sack) { | |
40efc6fa SH |
497 | *ptr++ = htonl(sp[this_sack].start_seq); |
498 | *ptr++ = htonl(sp[this_sack].end_seq); | |
499 | } | |
2de979bd | 500 | |
5861f8e5 | 501 | tp->rx_opt.dsack = 0; |
40efc6fa | 502 | } |
2100c8d2 YC |
503 | |
504 | if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { | |
505 | struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; | |
7f9b838b DL |
506 | u8 *p = (u8 *)ptr; |
507 | u32 len; /* Fast Open option length */ | |
508 | ||
509 | if (foc->exp) { | |
510 | len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; | |
511 | *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | | |
512 | TCPOPT_FASTOPEN_MAGIC); | |
513 | p += TCPOLEN_EXP_FASTOPEN_BASE; | |
514 | } else { | |
515 | len = TCPOLEN_FASTOPEN_BASE + foc->len; | |
516 | *p++ = TCPOPT_FASTOPEN; | |
517 | *p++ = len; | |
518 | } | |
2100c8d2 | 519 | |
7f9b838b DL |
520 | memcpy(p, foc->val, foc->len); |
521 | if ((len & 3) == 2) { | |
522 | p[foc->len] = TCPOPT_NOP; | |
523 | p[foc->len + 1] = TCPOPT_NOP; | |
2100c8d2 | 524 | } |
7f9b838b | 525 | ptr += (len + 3) >> 2; |
2100c8d2 | 526 | } |
60e2a778 UB |
527 | |
528 | smc_options_write(ptr, &options); | |
529 | } | |
530 | ||
531 | static void smc_set_option(const struct tcp_sock *tp, | |
532 | struct tcp_out_options *opts, | |
533 | unsigned int *remaining) | |
534 | { | |
535 | #if IS_ENABLED(CONFIG_SMC) | |
536 | if (static_branch_unlikely(&tcp_have_smc)) { | |
537 | if (tp->syn_smc) { | |
538 | if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { | |
539 | opts->options |= OPTION_SMC; | |
540 | *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; | |
541 | } | |
542 | } | |
543 | } | |
544 | #endif | |
545 | } | |
546 | ||
547 | static void smc_set_option_cond(const struct tcp_sock *tp, | |
548 | const struct inet_request_sock *ireq, | |
549 | struct tcp_out_options *opts, | |
550 | unsigned int *remaining) | |
551 | { | |
552 | #if IS_ENABLED(CONFIG_SMC) | |
553 | if (static_branch_unlikely(&tcp_have_smc)) { | |
554 | if (tp->syn_smc && ireq->smc_ok) { | |
555 | if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { | |
556 | opts->options |= OPTION_SMC; | |
557 | *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; | |
558 | } | |
559 | } | |
560 | } | |
561 | #endif | |
33ad798c AL |
562 | } |
563 | ||
67edfef7 AK |
564 | /* Compute TCP options for SYN packets. This is not the final |
565 | * network wire format yet. | |
566 | */ | |
95c96174 | 567 | static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 568 | struct tcp_out_options *opts, |
cf533ea5 ED |
569 | struct tcp_md5sig_key **md5) |
570 | { | |
33ad798c | 571 | struct tcp_sock *tp = tcp_sk(sk); |
95c96174 | 572 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
783237e8 | 573 | struct tcp_fastopen_request *fastopen = tp->fastopen_req; |
33ad798c | 574 | |
cfb6eeb4 | 575 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
576 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
577 | if (*md5) { | |
578 | opts->options |= OPTION_MD5; | |
bd0388ae | 579 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
cfb6eeb4 | 580 | } |
33ad798c AL |
581 | #else |
582 | *md5 = NULL; | |
cfb6eeb4 | 583 | #endif |
33ad798c AL |
584 | |
585 | /* We always get an MSS option. The option bytes which will be seen in | |
586 | * normal data packets should timestamps be used, must be in the MSS | |
587 | * advertised. But we subtract them from tp->mss_cache so that | |
588 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
589 | * fact here if necessary. If we don't do this correctly, as a | |
590 | * receiver we won't recognize data packets as being full sized when we | |
591 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
592 | * SACKs don't matter, we never delay an ACK when we have any of those | |
593 | * going out. */ | |
594 | opts->mss = tcp_advertise_mss(sk); | |
bd0388ae | 595 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c | 596 | |
5d2ed052 | 597 | if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { |
33ad798c | 598 | opts->options |= OPTION_TS; |
7faee5c0 | 599 | opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; |
33ad798c | 600 | opts->tsecr = tp->rx_opt.ts_recent; |
bd0388ae | 601 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c | 602 | } |
9bb37ef0 | 603 | if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { |
33ad798c | 604 | opts->ws = tp->rx_opt.rcv_wscale; |
89e95a61 | 605 | opts->options |= OPTION_WSCALE; |
bd0388ae | 606 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 607 | } |
f9301034 | 608 | if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { |
33ad798c | 609 | opts->options |= OPTION_SACK_ADVERTISE; |
b32d1310 | 610 | if (unlikely(!(OPTION_TS & opts->options))) |
bd0388ae | 611 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
612 | } |
613 | ||
783237e8 | 614 | if (fastopen && fastopen->cookie.len >= 0) { |
2646c831 DL |
615 | u32 need = fastopen->cookie.len; |
616 | ||
617 | need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : | |
618 | TCPOLEN_FASTOPEN_BASE; | |
783237e8 YC |
619 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
620 | if (remaining >= need) { | |
621 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
622 | opts->fastopen_cookie = &fastopen->cookie; | |
623 | remaining -= need; | |
624 | tp->syn_fastopen = 1; | |
2646c831 | 625 | tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; |
783237e8 YC |
626 | } |
627 | } | |
bd0388ae | 628 | |
60e2a778 UB |
629 | smc_set_option(tp, opts, &remaining); |
630 | ||
bd0388ae | 631 | return MAX_TCP_OPTION_SPACE - remaining; |
40efc6fa SH |
632 | } |
633 | ||
67edfef7 | 634 | /* Set up TCP options for SYN-ACKs. */ |
60e2a778 UB |
635 | static unsigned int tcp_synack_options(const struct sock *sk, |
636 | struct request_sock *req, | |
37bfbdda ED |
637 | unsigned int mss, struct sk_buff *skb, |
638 | struct tcp_out_options *opts, | |
639 | const struct tcp_md5sig_key *md5, | |
640 | struct tcp_fastopen_cookie *foc) | |
4957faad | 641 | { |
33ad798c | 642 | struct inet_request_sock *ireq = inet_rsk(req); |
95c96174 | 643 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
33ad798c | 644 | |
cfb6eeb4 | 645 | #ifdef CONFIG_TCP_MD5SIG |
80f03e27 | 646 | if (md5) { |
33ad798c | 647 | opts->options |= OPTION_MD5; |
4957faad WAS |
648 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
649 | ||
650 | /* We can't fit any SACK blocks in a packet with MD5 + TS | |
651 | * options. There was discussion about disabling SACK | |
652 | * rather than TS in order to fit in better with old, | |
653 | * buggy kernels, but that was deemed to be unnecessary. | |
654 | */ | |
de213e5e | 655 | ireq->tstamp_ok &= !ireq->sack_ok; |
cfb6eeb4 YH |
656 | } |
657 | #endif | |
33ad798c | 658 | |
4957faad | 659 | /* We always send an MSS option. */ |
33ad798c | 660 | opts->mss = mss; |
4957faad | 661 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c AL |
662 | |
663 | if (likely(ireq->wscale_ok)) { | |
664 | opts->ws = ireq->rcv_wscale; | |
89e95a61 | 665 | opts->options |= OPTION_WSCALE; |
4957faad | 666 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 667 | } |
de213e5e | 668 | if (likely(ireq->tstamp_ok)) { |
33ad798c | 669 | opts->options |= OPTION_TS; |
95a22cae | 670 | opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; |
33ad798c | 671 | opts->tsecr = req->ts_recent; |
4957faad | 672 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c AL |
673 | } |
674 | if (likely(ireq->sack_ok)) { | |
675 | opts->options |= OPTION_SACK_ADVERTISE; | |
de213e5e | 676 | if (unlikely(!ireq->tstamp_ok)) |
4957faad | 677 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c | 678 | } |
7f9b838b DL |
679 | if (foc != NULL && foc->len >= 0) { |
680 | u32 need = foc->len; | |
681 | ||
682 | need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : | |
683 | TCPOLEN_FASTOPEN_BASE; | |
8336886f JC |
684 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
685 | if (remaining >= need) { | |
686 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
687 | opts->fastopen_cookie = foc; | |
688 | remaining -= need; | |
689 | } | |
690 | } | |
1a2c6181 | 691 | |
60e2a778 UB |
692 | smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); |
693 | ||
4957faad | 694 | return MAX_TCP_OPTION_SPACE - remaining; |
33ad798c AL |
695 | } |
696 | ||
67edfef7 AK |
697 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
698 | * final wire format yet. | |
699 | */ | |
95c96174 | 700 | static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 701 | struct tcp_out_options *opts, |
cf533ea5 ED |
702 | struct tcp_md5sig_key **md5) |
703 | { | |
33ad798c | 704 | struct tcp_sock *tp = tcp_sk(sk); |
95c96174 | 705 | unsigned int size = 0; |
cabeccbd | 706 | unsigned int eff_sacks; |
33ad798c | 707 | |
5843ef42 AK |
708 | opts->options = 0; |
709 | ||
33ad798c AL |
710 | #ifdef CONFIG_TCP_MD5SIG |
711 | *md5 = tp->af_specific->md5_lookup(sk, sk); | |
712 | if (unlikely(*md5)) { | |
713 | opts->options |= OPTION_MD5; | |
714 | size += TCPOLEN_MD5SIG_ALIGNED; | |
715 | } | |
716 | #else | |
717 | *md5 = NULL; | |
718 | #endif | |
719 | ||
720 | if (likely(tp->rx_opt.tstamp_ok)) { | |
721 | opts->options |= OPTION_TS; | |
7faee5c0 | 722 | opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; |
33ad798c AL |
723 | opts->tsecr = tp->rx_opt.ts_recent; |
724 | size += TCPOLEN_TSTAMP_ALIGNED; | |
725 | } | |
726 | ||
cabeccbd IJ |
727 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
728 | if (unlikely(eff_sacks)) { | |
95c96174 | 729 | const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; |
33ad798c | 730 | opts->num_sack_blocks = |
95c96174 | 731 | min_t(unsigned int, eff_sacks, |
33ad798c AL |
732 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
733 | TCPOLEN_SACK_PERBLOCK); | |
2af7feb3 ED |
734 | if (likely(opts->num_sack_blocks)) |
735 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
736 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
33ad798c AL |
737 | } |
738 | ||
739 | return size; | |
40efc6fa | 740 | } |
1da177e4 | 741 | |
46d3ceab ED |
742 | |
743 | /* TCP SMALL QUEUES (TSQ) | |
744 | * | |
745 | * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) | |
746 | * to reduce RTT and bufferbloat. | |
747 | * We do this using a special skb destructor (tcp_wfree). | |
748 | * | |
749 | * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb | |
750 | * needs to be reallocated in a driver. | |
8e3bff96 | 751 | * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc |
46d3ceab ED |
752 | * |
753 | * Since transmit from skb destructor is forbidden, we use a tasklet | |
754 | * to process all sockets that eventually need to send more skbs. | |
755 | * We use one tasklet per cpu, with its own queue of sockets. | |
756 | */ | |
757 | struct tsq_tasklet { | |
758 | struct tasklet_struct tasklet; | |
759 | struct list_head head; /* queue of tcp sockets */ | |
760 | }; | |
761 | static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); | |
762 | ||
6f458dfb ED |
763 | static void tcp_tsq_handler(struct sock *sk) |
764 | { | |
765 | if ((1 << sk->sk_state) & | |
766 | (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | | |
f9616c35 ED |
767 | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { |
768 | struct tcp_sock *tp = tcp_sk(sk); | |
769 | ||
770 | if (tp->lost_out > tp->retrans_out && | |
3a91d29f KD |
771 | tp->snd_cwnd > tcp_packets_in_flight(tp)) { |
772 | tcp_mstamp_refresh(tp); | |
f9616c35 | 773 | tcp_xmit_retransmit_queue(sk); |
3a91d29f | 774 | } |
f9616c35 ED |
775 | |
776 | tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, | |
bf06200e | 777 | 0, GFP_ATOMIC); |
f9616c35 | 778 | } |
6f458dfb | 779 | } |
46d3ceab | 780 | /* |
8e3bff96 | 781 | * One tasklet per cpu tries to send more skbs. |
46d3ceab | 782 | * We run in tasklet context but need to disable irqs when |
8e3bff96 | 783 | * transferring tsq->head because tcp_wfree() might |
46d3ceab ED |
784 | * interrupt us (non NAPI drivers) |
785 | */ | |
786 | static void tcp_tasklet_func(unsigned long data) | |
787 | { | |
788 | struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; | |
789 | LIST_HEAD(list); | |
790 | unsigned long flags; | |
791 | struct list_head *q, *n; | |
792 | struct tcp_sock *tp; | |
793 | struct sock *sk; | |
794 | ||
795 | local_irq_save(flags); | |
796 | list_splice_init(&tsq->head, &list); | |
797 | local_irq_restore(flags); | |
798 | ||
799 | list_for_each_safe(q, n, &list) { | |
800 | tp = list_entry(q, struct tcp_sock, tsq_node); | |
801 | list_del(&tp->tsq_node); | |
802 | ||
803 | sk = (struct sock *)tp; | |
0a9648f1 | 804 | smp_mb__before_atomic(); |
7aa5470c ED |
805 | clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); |
806 | ||
b223feb9 | 807 | if (!sk->sk_lock.owned && |
7aa5470c | 808 | test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) { |
b223feb9 ED |
809 | bh_lock_sock(sk); |
810 | if (!sock_owned_by_user(sk)) { | |
7aa5470c | 811 | clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); |
b223feb9 ED |
812 | tcp_tsq_handler(sk); |
813 | } | |
814 | bh_unlock_sock(sk); | |
46d3ceab | 815 | } |
46d3ceab | 816 | |
46d3ceab ED |
817 | sk_free(sk); |
818 | } | |
819 | } | |
820 | ||
40fc3423 ED |
821 | #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ |
822 | TCPF_WRITE_TIMER_DEFERRED | \ | |
823 | TCPF_DELACK_TIMER_DEFERRED | \ | |
824 | TCPF_MTU_REDUCED_DEFERRED) | |
46d3ceab ED |
825 | /** |
826 | * tcp_release_cb - tcp release_sock() callback | |
827 | * @sk: socket | |
828 | * | |
829 | * called from release_sock() to perform protocol dependent | |
830 | * actions before socket release. | |
831 | */ | |
832 | void tcp_release_cb(struct sock *sk) | |
833 | { | |
6f458dfb | 834 | unsigned long flags, nflags; |
46d3ceab | 835 | |
6f458dfb ED |
836 | /* perform an atomic operation only if at least one flag is set */ |
837 | do { | |
7aa5470c | 838 | flags = sk->sk_tsq_flags; |
6f458dfb ED |
839 | if (!(flags & TCP_DEFERRED_ALL)) |
840 | return; | |
841 | nflags = flags & ~TCP_DEFERRED_ALL; | |
7aa5470c | 842 | } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); |
6f458dfb | 843 | |
40fc3423 | 844 | if (flags & TCPF_TSQ_DEFERRED) |
6f458dfb ED |
845 | tcp_tsq_handler(sk); |
846 | ||
c3f9b018 ED |
847 | /* Here begins the tricky part : |
848 | * We are called from release_sock() with : | |
849 | * 1) BH disabled | |
850 | * 2) sk_lock.slock spinlock held | |
851 | * 3) socket owned by us (sk->sk_lock.owned == 1) | |
852 | * | |
853 | * But following code is meant to be called from BH handlers, | |
854 | * so we should keep BH disabled, but early release socket ownership | |
855 | */ | |
856 | sock_release_ownership(sk); | |
857 | ||
40fc3423 | 858 | if (flags & TCPF_WRITE_TIMER_DEFERRED) { |
6f458dfb | 859 | tcp_write_timer_handler(sk); |
144d56e9 ED |
860 | __sock_put(sk); |
861 | } | |
40fc3423 | 862 | if (flags & TCPF_DELACK_TIMER_DEFERRED) { |
6f458dfb | 863 | tcp_delack_timer_handler(sk); |
144d56e9 ED |
864 | __sock_put(sk); |
865 | } | |
40fc3423 | 866 | if (flags & TCPF_MTU_REDUCED_DEFERRED) { |
4fab9071 | 867 | inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); |
144d56e9 ED |
868 | __sock_put(sk); |
869 | } | |
46d3ceab ED |
870 | } |
871 | EXPORT_SYMBOL(tcp_release_cb); | |
872 | ||
873 | void __init tcp_tasklet_init(void) | |
874 | { | |
875 | int i; | |
876 | ||
877 | for_each_possible_cpu(i) { | |
878 | struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); | |
879 | ||
880 | INIT_LIST_HEAD(&tsq->head); | |
881 | tasklet_init(&tsq->tasklet, | |
882 | tcp_tasklet_func, | |
883 | (unsigned long)tsq); | |
884 | } | |
885 | } | |
886 | ||
887 | /* | |
888 | * Write buffer destructor automatically called from kfree_skb. | |
8e3bff96 | 889 | * We can't xmit new skbs from this context, as we might already |
46d3ceab ED |
890 | * hold qdisc lock. |
891 | */ | |
d6a4a104 | 892 | void tcp_wfree(struct sk_buff *skb) |
46d3ceab ED |
893 | { |
894 | struct sock *sk = skb->sk; | |
895 | struct tcp_sock *tp = tcp_sk(sk); | |
408f0a6c | 896 | unsigned long flags, nval, oval; |
9b462d02 ED |
897 | |
898 | /* Keep one reference on sk_wmem_alloc. | |
899 | * Will be released by sk_free() from here or tcp_tasklet_func() | |
900 | */ | |
14afee4b | 901 | WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); |
9b462d02 ED |
902 | |
903 | /* If this softirq is serviced by ksoftirqd, we are likely under stress. | |
904 | * Wait until our queues (qdisc + devices) are drained. | |
905 | * This gives : | |
906 | * - less callbacks to tcp_write_xmit(), reducing stress (batches) | |
907 | * - chance for incoming ACK (processed by another cpu maybe) | |
908 | * to migrate this flow (skb->ooo_okay will be eventually set) | |
909 | */ | |
14afee4b | 910 | if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) |
9b462d02 | 911 | goto out; |
46d3ceab | 912 | |
7aa5470c | 913 | for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { |
46d3ceab | 914 | struct tsq_tasklet *tsq; |
a9b204d1 | 915 | bool empty; |
46d3ceab | 916 | |
408f0a6c ED |
917 | if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) |
918 | goto out; | |
919 | ||
b223feb9 | 920 | nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; |
7aa5470c | 921 | nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); |
408f0a6c ED |
922 | if (nval != oval) |
923 | continue; | |
924 | ||
46d3ceab ED |
925 | /* queue this socket to tasklet queue */ |
926 | local_irq_save(flags); | |
903ceff7 | 927 | tsq = this_cpu_ptr(&tsq_tasklet); |
a9b204d1 | 928 | empty = list_empty(&tsq->head); |
46d3ceab | 929 | list_add(&tp->tsq_node, &tsq->head); |
a9b204d1 ED |
930 | if (empty) |
931 | tasklet_schedule(&tsq->tasklet); | |
46d3ceab | 932 | local_irq_restore(flags); |
9b462d02 | 933 | return; |
46d3ceab | 934 | } |
9b462d02 ED |
935 | out: |
936 | sk_free(sk); | |
46d3ceab ED |
937 | } |
938 | ||
218af599 ED |
939 | /* Note: Called under hard irq. |
940 | * We can not call TCP stack right away. | |
941 | */ | |
942 | enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) | |
943 | { | |
944 | struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); | |
945 | struct sock *sk = (struct sock *)tp; | |
946 | unsigned long nval, oval; | |
947 | ||
948 | for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { | |
949 | struct tsq_tasklet *tsq; | |
950 | bool empty; | |
951 | ||
952 | if (oval & TSQF_QUEUED) | |
953 | break; | |
954 | ||
955 | nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; | |
956 | nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); | |
957 | if (nval != oval) | |
958 | continue; | |
959 | ||
14afee4b | 960 | if (!refcount_inc_not_zero(&sk->sk_wmem_alloc)) |
218af599 ED |
961 | break; |
962 | /* queue this socket to tasklet queue */ | |
963 | tsq = this_cpu_ptr(&tsq_tasklet); | |
964 | empty = list_empty(&tsq->head); | |
965 | list_add(&tp->tsq_node, &tsq->head); | |
966 | if (empty) | |
967 | tasklet_schedule(&tsq->tasklet); | |
968 | break; | |
969 | } | |
970 | return HRTIMER_NORESTART; | |
971 | } | |
972 | ||
973 | /* BBR congestion control needs pacing. | |
974 | * Same remark for SO_MAX_PACING_RATE. | |
975 | * sch_fq packet scheduler is efficiently handling pacing, | |
976 | * but is not always installed/used. | |
977 | * Return true if TCP stack should pace packets itself. | |
978 | */ | |
979 | static bool tcp_needs_internal_pacing(const struct sock *sk) | |
980 | { | |
981 | return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED; | |
982 | } | |
983 | ||
984 | static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb) | |
985 | { | |
986 | u64 len_ns; | |
987 | u32 rate; | |
988 | ||
989 | if (!tcp_needs_internal_pacing(sk)) | |
990 | return; | |
991 | rate = sk->sk_pacing_rate; | |
992 | if (!rate || rate == ~0U) | |
993 | return; | |
994 | ||
995 | /* Should account for header sizes as sch_fq does, | |
996 | * but lets make things simple. | |
997 | */ | |
998 | len_ns = (u64)skb->len * NSEC_PER_SEC; | |
999 | do_div(len_ns, rate); | |
1000 | hrtimer_start(&tcp_sk(sk)->pacing_timer, | |
1001 | ktime_add_ns(ktime_get(), len_ns), | |
1002 | HRTIMER_MODE_ABS_PINNED); | |
1003 | } | |
1004 | ||
e2080072 ED |
1005 | static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb) |
1006 | { | |
1007 | skb->skb_mstamp = tp->tcp_mstamp; | |
1008 | list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); | |
1009 | } | |
1010 | ||
1da177e4 LT |
1011 | /* This routine actually transmits TCP packets queued in by |
1012 | * tcp_do_sendmsg(). This is used by both the initial | |
1013 | * transmission and possible later retransmissions. | |
1014 | * All SKB's seen here are completely headerless. It is our | |
1015 | * job to build the TCP header, and pass the packet down to | |
1016 | * IP so it can do the same plus pass the packet off to the | |
1017 | * device. | |
1018 | * | |
1019 | * We are working here with either a clone of the original | |
1020 | * SKB, or a fresh unique copy made by the retransmit engine. | |
1021 | */ | |
5cf68c20 YC |
1022 | static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, |
1023 | int clone_it, gfp_t gfp_mask, u32 rcv_nxt) | |
1da177e4 | 1024 | { |
dfb4b9dc DM |
1025 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1026 | struct inet_sock *inet; | |
1027 | struct tcp_sock *tp; | |
1028 | struct tcp_skb_cb *tcb; | |
33ad798c | 1029 | struct tcp_out_options opts; |
95c96174 | 1030 | unsigned int tcp_options_size, tcp_header_size; |
8c72c65b | 1031 | struct sk_buff *oskb = NULL; |
cfb6eeb4 | 1032 | struct tcp_md5sig_key *md5; |
dfb4b9dc | 1033 | struct tcphdr *th; |
dfb4b9dc DM |
1034 | int err; |
1035 | ||
1036 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
6f094b9e | 1037 | tp = tcp_sk(sk); |
dfb4b9dc | 1038 | |
ccdbb6e9 | 1039 | if (clone_it) { |
6f094b9e LB |
1040 | TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq |
1041 | - tp->snd_una; | |
8c72c65b | 1042 | oskb = skb; |
e2080072 ED |
1043 | |
1044 | tcp_skb_tsorted_save(oskb) { | |
1045 | if (unlikely(skb_cloned(oskb))) | |
1046 | skb = pskb_copy(oskb, gfp_mask); | |
1047 | else | |
1048 | skb = skb_clone(oskb, gfp_mask); | |
1049 | } tcp_skb_tsorted_restore(oskb); | |
1050 | ||
dfb4b9dc DM |
1051 | if (unlikely(!skb)) |
1052 | return -ENOBUFS; | |
1053 | } | |
8c72c65b | 1054 | skb->skb_mstamp = tp->tcp_mstamp; |
1da177e4 | 1055 | |
dfb4b9dc | 1056 | inet = inet_sk(sk); |
dfb4b9dc | 1057 | tcb = TCP_SKB_CB(skb); |
33ad798c | 1058 | memset(&opts, 0, sizeof(opts)); |
1da177e4 | 1059 | |
4de075e0 | 1060 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) |
33ad798c AL |
1061 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); |
1062 | else | |
1063 | tcp_options_size = tcp_established_options(sk, skb, &opts, | |
1064 | &md5); | |
1065 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); | |
e905a9ed | 1066 | |
547669d4 | 1067 | /* if no packet is in qdisc/device queue, then allow XPS to select |
b2532eb9 ED |
1068 | * another queue. We can be called from tcp_tsq_handler() |
1069 | * which holds one reference to sk_wmem_alloc. | |
1070 | * | |
1071 | * TODO: Ideally, in-flight pure ACK packets should not matter here. | |
1072 | * One way to get this would be to set skb->truesize = 2 on them. | |
547669d4 | 1073 | */ |
b2532eb9 | 1074 | skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); |
dfb4b9dc | 1075 | |
38ab52e8 ED |
1076 | /* If we had to use memory reserve to allocate this skb, |
1077 | * this might cause drops if packet is looped back : | |
1078 | * Other socket might not have SOCK_MEMALLOC. | |
1079 | * Packets not looped back do not care about pfmemalloc. | |
1080 | */ | |
1081 | skb->pfmemalloc = 0; | |
1082 | ||
aa8223c7 ACM |
1083 | skb_push(skb, tcp_header_size); |
1084 | skb_reset_transport_header(skb); | |
46d3ceab ED |
1085 | |
1086 | skb_orphan(skb); | |
1087 | skb->sk = sk; | |
1d2077ac | 1088 | skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; |
b73c3d0e | 1089 | skb_set_hash_from_sk(skb, sk); |
14afee4b | 1090 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); |
dfb4b9dc | 1091 | |
c3a2e837 JA |
1092 | skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); |
1093 | ||
dfb4b9dc | 1094 | /* Build TCP header and checksum it. */ |
ea1627c2 | 1095 | th = (struct tcphdr *)skb->data; |
c720c7e8 ED |
1096 | th->source = inet->inet_sport; |
1097 | th->dest = inet->inet_dport; | |
dfb4b9dc | 1098 | th->seq = htonl(tcb->seq); |
5cf68c20 | 1099 | th->ack_seq = htonl(rcv_nxt); |
df7a3b07 | 1100 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
4de075e0 | 1101 | tcb->tcp_flags); |
dfb4b9dc | 1102 | |
dfb4b9dc DM |
1103 | th->check = 0; |
1104 | th->urg_ptr = 0; | |
1da177e4 | 1105 | |
33f5f57e | 1106 | /* The urg_mode check is necessary during a below snd_una win probe */ |
7691367d HX |
1107 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
1108 | if (before(tp->snd_up, tcb->seq + 0x10000)) { | |
1109 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
1110 | th->urg = 1; | |
1111 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { | |
0eae88f3 | 1112 | th->urg_ptr = htons(0xFFFF); |
7691367d HX |
1113 | th->urg = 1; |
1114 | } | |
dfb4b9dc | 1115 | } |
1da177e4 | 1116 | |
bd0388ae | 1117 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
51466a75 | 1118 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
ea1627c2 ED |
1119 | if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { |
1120 | th->window = htons(tcp_select_window(sk)); | |
1121 | tcp_ecn_send(sk, skb, th, tcp_header_size); | |
1122 | } else { | |
1123 | /* RFC1323: The window in SYN & SYN/ACK segments | |
1124 | * is never scaled. | |
1125 | */ | |
1126 | th->window = htons(min(tp->rcv_wnd, 65535U)); | |
1127 | } | |
cfb6eeb4 YH |
1128 | #ifdef CONFIG_TCP_MD5SIG |
1129 | /* Calculate the MD5 hash, as we have all we need now */ | |
1130 | if (md5) { | |
a465419b | 1131 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); |
bd0388ae | 1132 | tp->af_specific->calc_md5_hash(opts.hash_location, |
39f8e58e | 1133 | md5, sk, skb); |
cfb6eeb4 YH |
1134 | } |
1135 | #endif | |
1136 | ||
bb296246 | 1137 | icsk->icsk_af_ops->send_check(sk, skb); |
1da177e4 | 1138 | |
4de075e0 | 1139 | if (likely(tcb->tcp_flags & TCPHDR_ACK)) |
deb42350 | 1140 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt); |
1da177e4 | 1141 | |
a44d6eac | 1142 | if (skb->len != tcp_header_size) { |
cf533ea5 | 1143 | tcp_event_data_sent(tp, sk); |
a44d6eac | 1144 | tp->data_segs_out += tcp_skb_pcount(skb); |
218af599 | 1145 | tcp_internal_pacing(sk, skb); |
a44d6eac | 1146 | } |
1da177e4 | 1147 | |
bd37a088 | 1148 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
aa2ea058 TH |
1149 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
1150 | tcp_skb_pcount(skb)); | |
1da177e4 | 1151 | |
2efd055c | 1152 | tp->segs_out += tcp_skb_pcount(skb); |
f69ad292 | 1153 | /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ |
cd7d8498 | 1154 | skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); |
f69ad292 | 1155 | skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); |
cd7d8498 | 1156 | |
7faee5c0 | 1157 | /* Our usage of tstamp should remain private */ |
2456e855 | 1158 | skb->tstamp = 0; |
971f10ec ED |
1159 | |
1160 | /* Cleanup our debris for IP stacks */ | |
1161 | memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), | |
1162 | sizeof(struct inet6_skb_parm))); | |
1163 | ||
b0270e91 | 1164 | err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); |
7faee5c0 | 1165 | |
8c72c65b ED |
1166 | if (unlikely(err > 0)) { |
1167 | tcp_enter_cwr(sk); | |
1168 | err = net_xmit_eval(err); | |
1169 | } | |
fc225799 | 1170 | if (!err && oskb) { |
e2080072 | 1171 | tcp_update_skb_after_send(tp, oskb); |
fc225799 ED |
1172 | tcp_rate_skb_sent(sk, oskb); |
1173 | } | |
8c72c65b | 1174 | return err; |
1da177e4 LT |
1175 | } |
1176 | ||
5cf68c20 YC |
1177 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
1178 | gfp_t gfp_mask) | |
1179 | { | |
1180 | return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, | |
1181 | tcp_sk(sk)->rcv_nxt); | |
1182 | } | |
1183 | ||
67edfef7 | 1184 | /* This routine just queues the buffer for sending. |
1da177e4 LT |
1185 | * |
1186 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
1187 | * otherwise socket can stall. | |
1188 | */ | |
1189 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
1190 | { | |
1191 | struct tcp_sock *tp = tcp_sk(sk); | |
1192 | ||
1193 | /* Advance write_seq and place onto the write_queue. */ | |
1194 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
f4a775d1 | 1195 | __skb_header_release(skb); |
fe067e8a | 1196 | tcp_add_write_queue_tail(sk, skb); |
3ab224be HA |
1197 | sk->sk_wmem_queued += skb->truesize; |
1198 | sk_mem_charge(sk, skb->truesize); | |
1da177e4 LT |
1199 | } |
1200 | ||
67edfef7 | 1201 | /* Initialize TSO segments for a packet. */ |
5bbb432c | 1202 | static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) |
f6302d1d | 1203 | { |
8f26fb1c | 1204 | if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { |
f6302d1d DM |
1205 | /* Avoid the costly divide in the normal |
1206 | * non-TSO case. | |
1207 | */ | |
cd7d8498 | 1208 | tcp_skb_pcount_set(skb, 1); |
f69ad292 | 1209 | TCP_SKB_CB(skb)->tcp_gso_size = 0; |
f6302d1d | 1210 | } else { |
cd7d8498 | 1211 | tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); |
f69ad292 | 1212 | TCP_SKB_CB(skb)->tcp_gso_size = mss_now; |
1da177e4 LT |
1213 | } |
1214 | } | |
1215 | ||
797108d1 IJ |
1216 | /* Pcount in the middle of the write queue got changed, we need to do various |
1217 | * tweaks to fix counters | |
1218 | */ | |
cf533ea5 | 1219 | static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) |
797108d1 IJ |
1220 | { |
1221 | struct tcp_sock *tp = tcp_sk(sk); | |
1222 | ||
1223 | tp->packets_out -= decr; | |
1224 | ||
1225 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
1226 | tp->sacked_out -= decr; | |
1227 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
1228 | tp->retrans_out -= decr; | |
1229 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) | |
1230 | tp->lost_out -= decr; | |
1231 | ||
1232 | /* Reno case is special. Sigh... */ | |
1233 | if (tcp_is_reno(tp) && decr > 0) | |
1234 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); | |
1235 | ||
797108d1 IJ |
1236 | if (tp->lost_skb_hint && |
1237 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && | |
713bafea | 1238 | (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) |
797108d1 IJ |
1239 | tp->lost_cnt_hint -= decr; |
1240 | ||
1241 | tcp_verify_left_out(tp); | |
1242 | } | |
1243 | ||
0a2cf20c SHY |
1244 | static bool tcp_has_tx_tstamp(const struct sk_buff *skb) |
1245 | { | |
1246 | return TCP_SKB_CB(skb)->txstamp_ack || | |
1247 | (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); | |
1248 | } | |
1249 | ||
490cc7d0 WB |
1250 | static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) |
1251 | { | |
1252 | struct skb_shared_info *shinfo = skb_shinfo(skb); | |
1253 | ||
0a2cf20c | 1254 | if (unlikely(tcp_has_tx_tstamp(skb)) && |
490cc7d0 WB |
1255 | !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { |
1256 | struct skb_shared_info *shinfo2 = skb_shinfo(skb2); | |
1257 | u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; | |
1258 | ||
1259 | shinfo->tx_flags &= ~tsflags; | |
1260 | shinfo2->tx_flags |= tsflags; | |
1261 | swap(shinfo->tskey, shinfo2->tskey); | |
b51e13fa MKL |
1262 | TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; |
1263 | TCP_SKB_CB(skb)->txstamp_ack = 0; | |
490cc7d0 WB |
1264 | } |
1265 | } | |
1266 | ||
a166140e MKL |
1267 | static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) |
1268 | { | |
1269 | TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; | |
1270 | TCP_SKB_CB(skb)->eor = 0; | |
1271 | } | |
1272 | ||
75c119af ED |
1273 | /* Insert buff after skb on the write or rtx queue of sk. */ |
1274 | static void tcp_insert_write_queue_after(struct sk_buff *skb, | |
1275 | struct sk_buff *buff, | |
1276 | struct sock *sk, | |
1277 | enum tcp_queue tcp_queue) | |
1278 | { | |
1279 | if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) | |
1280 | __skb_queue_after(&sk->sk_write_queue, skb, buff); | |
1281 | else | |
1282 | tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); | |
1283 | } | |
1284 | ||
1da177e4 LT |
1285 | /* Function to create two new TCP segments. Shrinks the given segment |
1286 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 1287 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
1288 | * Remember, these are still headerless SKBs at this point. |
1289 | */ | |
75c119af ED |
1290 | int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, |
1291 | struct sk_buff *skb, u32 len, | |
6cc55e09 | 1292 | unsigned int mss_now, gfp_t gfp) |
1da177e4 LT |
1293 | { |
1294 | struct tcp_sock *tp = tcp_sk(sk); | |
1295 | struct sk_buff *buff; | |
6475be16 | 1296 | int nsize, old_factor; |
f5667634 | 1297 | long limit; |
b60b49ea | 1298 | int nlen; |
9ce01461 | 1299 | u8 flags; |
1da177e4 | 1300 | |
2fceec13 IJ |
1301 | if (WARN_ON(len > skb->len)) |
1302 | return -EINVAL; | |
6a438bbe | 1303 | |
1da177e4 LT |
1304 | nsize = skb_headlen(skb) - len; |
1305 | if (nsize < 0) | |
1306 | nsize = 0; | |
1307 | ||
f5667634 ED |
1308 | /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb. |
1309 | * We need some allowance to not penalize applications setting small | |
1310 | * SO_SNDBUF values. | |
1311 | * Also allow first and last skb in retransmit queue to be split. | |
1312 | */ | |
1313 | limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE); | |
1314 | if (unlikely((sk->sk_wmem_queued >> 1) > limit && | |
1315 | tcp_queue != TCP_FRAG_IN_WRITE_QUEUE && | |
1316 | skb != tcp_rtx_queue_head(sk) && | |
1317 | skb != tcp_rtx_queue_tail(sk))) { | |
94c02771 ED |
1318 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG); |
1319 | return -ENOMEM; | |
1320 | } | |
1321 | ||
6cc55e09 | 1322 | if (skb_unclone(skb, gfp)) |
1da177e4 LT |
1323 | return -ENOMEM; |
1324 | ||
1325 | /* Get a new skb... force flag on. */ | |
eb934478 | 1326 | buff = sk_stream_alloc_skb(sk, nsize, gfp, true); |
51456b29 | 1327 | if (!buff) |
1da177e4 | 1328 | return -ENOMEM; /* We'll just try again later. */ |
ef5cb973 | 1329 | |
3ab224be HA |
1330 | sk->sk_wmem_queued += buff->truesize; |
1331 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea HX |
1332 | nlen = skb->len - len - nsize; |
1333 | buff->truesize += nlen; | |
1334 | skb->truesize -= nlen; | |
1da177e4 LT |
1335 | |
1336 | /* Correct the sequence numbers. */ | |
1337 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1338 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1339 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1340 | ||
1341 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1342 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1343 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1344 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
e14c3caf | 1345 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
a166140e | 1346 | tcp_skb_fragment_eor(skb, buff); |
1da177e4 | 1347 | |
84fa7933 | 1348 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
1da177e4 | 1349 | /* Copy and checksum data tail into the new buffer. */ |
056834d9 IJ |
1350 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
1351 | skb_put(buff, nsize), | |
1da177e4 LT |
1352 | nsize, 0); |
1353 | ||
1354 | skb_trim(skb, len); | |
1355 | ||
1356 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
1357 | } else { | |
84fa7933 | 1358 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1359 | skb_split(skb, buff, len); |
1360 | } | |
1361 | ||
1362 | buff->ip_summed = skb->ip_summed; | |
1363 | ||
a61bbcf2 | 1364 | buff->tstamp = skb->tstamp; |
490cc7d0 | 1365 | tcp_fragment_tstamp(skb, buff); |
1da177e4 | 1366 | |
6475be16 DM |
1367 | old_factor = tcp_skb_pcount(skb); |
1368 | ||
1da177e4 | 1369 | /* Fix up tso_factor for both original and new SKB. */ |
5bbb432c ED |
1370 | tcp_set_skb_tso_segs(skb, mss_now); |
1371 | tcp_set_skb_tso_segs(buff, mss_now); | |
1da177e4 | 1372 | |
b9f64820 YC |
1373 | /* Update delivered info for the new segment */ |
1374 | TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; | |
1375 | ||
6475be16 DM |
1376 | /* If this packet has been sent out already, we must |
1377 | * adjust the various packet counters. | |
1378 | */ | |
cf0b450c | 1379 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
1380 | int diff = old_factor - tcp_skb_pcount(skb) - |
1381 | tcp_skb_pcount(buff); | |
1da177e4 | 1382 | |
797108d1 IJ |
1383 | if (diff) |
1384 | tcp_adjust_pcount(sk, skb, diff); | |
1da177e4 LT |
1385 | } |
1386 | ||
1387 | /* Link BUFF into the send queue. */ | |
f4a775d1 | 1388 | __skb_header_release(buff); |
75c119af | 1389 | tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); |
f67971e6 ED |
1390 | if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) |
1391 | list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); | |
1da177e4 LT |
1392 | |
1393 | return 0; | |
1394 | } | |
1395 | ||
f4d01666 ED |
1396 | /* This is similar to __pskb_pull_tail(). The difference is that pulled |
1397 | * data is not copied, but immediately discarded. | |
1da177e4 | 1398 | */ |
7162fb24 | 1399 | static int __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 | 1400 | { |
7b7fc97a | 1401 | struct skb_shared_info *shinfo; |
1da177e4 LT |
1402 | int i, k, eat; |
1403 | ||
4fa48bf3 ED |
1404 | eat = min_t(int, len, skb_headlen(skb)); |
1405 | if (eat) { | |
1406 | __skb_pull(skb, eat); | |
1407 | len -= eat; | |
1408 | if (!len) | |
7162fb24 | 1409 | return 0; |
4fa48bf3 | 1410 | } |
1da177e4 LT |
1411 | eat = len; |
1412 | k = 0; | |
7b7fc97a ED |
1413 | shinfo = skb_shinfo(skb); |
1414 | for (i = 0; i < shinfo->nr_frags; i++) { | |
1415 | int size = skb_frag_size(&shinfo->frags[i]); | |
9e903e08 ED |
1416 | |
1417 | if (size <= eat) { | |
aff65da0 | 1418 | skb_frag_unref(skb, i); |
9e903e08 | 1419 | eat -= size; |
1da177e4 | 1420 | } else { |
7b7fc97a | 1421 | shinfo->frags[k] = shinfo->frags[i]; |
1da177e4 | 1422 | if (eat) { |
7b7fc97a ED |
1423 | shinfo->frags[k].page_offset += eat; |
1424 | skb_frag_size_sub(&shinfo->frags[k], eat); | |
1da177e4 LT |
1425 | eat = 0; |
1426 | } | |
1427 | k++; | |
1428 | } | |
1429 | } | |
7b7fc97a | 1430 | shinfo->nr_frags = k; |
1da177e4 | 1431 | |
1da177e4 LT |
1432 | skb->data_len -= len; |
1433 | skb->len = skb->data_len; | |
7162fb24 | 1434 | return len; |
1da177e4 LT |
1435 | } |
1436 | ||
67edfef7 | 1437 | /* Remove acked data from a packet in the transmit queue. */ |
1da177e4 LT |
1438 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
1439 | { | |
7162fb24 ED |
1440 | u32 delta_truesize; |
1441 | ||
14bbd6a5 | 1442 | if (skb_unclone(skb, GFP_ATOMIC)) |
1da177e4 LT |
1443 | return -ENOMEM; |
1444 | ||
7162fb24 | 1445 | delta_truesize = __pskb_trim_head(skb, len); |
1da177e4 LT |
1446 | |
1447 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 1448 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 | 1449 | |
7162fb24 ED |
1450 | if (delta_truesize) { |
1451 | skb->truesize -= delta_truesize; | |
1452 | sk->sk_wmem_queued -= delta_truesize; | |
1453 | sk_mem_uncharge(sk, delta_truesize); | |
1454 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); | |
1455 | } | |
1da177e4 | 1456 | |
5b35e1e6 | 1457 | /* Any change of skb->len requires recalculation of tso factor. */ |
1da177e4 | 1458 | if (tcp_skb_pcount(skb) > 1) |
5bbb432c | 1459 | tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); |
1da177e4 LT |
1460 | |
1461 | return 0; | |
1462 | } | |
1463 | ||
1b63edd6 YC |
1464 | /* Calculate MSS not accounting any TCP options. */ |
1465 | static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
5d424d5a | 1466 | { |
cf533ea5 ED |
1467 | const struct tcp_sock *tp = tcp_sk(sk); |
1468 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1469 | int mss_now; |
1470 | ||
1471 | /* Calculate base mss without TCP options: | |
1472 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
1473 | */ | |
1474 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
1475 | ||
67469601 ED |
1476 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1477 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1478 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1479 | ||
1480 | if (dst && dst_allfrag(dst)) | |
1481 | mss_now -= icsk->icsk_af_ops->net_frag_header_len; | |
1482 | } | |
1483 | ||
5d424d5a JH |
1484 | /* Clamp it (mss_clamp does not include tcp options) */ |
1485 | if (mss_now > tp->rx_opt.mss_clamp) | |
1486 | mss_now = tp->rx_opt.mss_clamp; | |
1487 | ||
1488 | /* Now subtract optional transport overhead */ | |
1489 | mss_now -= icsk->icsk_ext_hdr_len; | |
1490 | ||
1491 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
be208663 | 1492 | mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss); |
5d424d5a JH |
1493 | return mss_now; |
1494 | } | |
1495 | ||
1b63edd6 YC |
1496 | /* Calculate MSS. Not accounting for SACKs here. */ |
1497 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
1498 | { | |
1499 | /* Subtract TCP options size, not including SACKs */ | |
1500 | return __tcp_mtu_to_mss(sk, pmtu) - | |
1501 | (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); | |
1502 | } | |
1503 | ||
5d424d5a | 1504 | /* Inverse of above */ |
67469601 | 1505 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
5d424d5a | 1506 | { |
cf533ea5 ED |
1507 | const struct tcp_sock *tp = tcp_sk(sk); |
1508 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1509 | int mtu; |
1510 | ||
1511 | mtu = mss + | |
1512 | tp->tcp_header_len + | |
1513 | icsk->icsk_ext_hdr_len + | |
1514 | icsk->icsk_af_ops->net_header_len; | |
1515 | ||
67469601 ED |
1516 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1517 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1518 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1519 | ||
1520 | if (dst && dst_allfrag(dst)) | |
1521 | mtu += icsk->icsk_af_ops->net_frag_header_len; | |
1522 | } | |
5d424d5a JH |
1523 | return mtu; |
1524 | } | |
556c6b46 | 1525 | EXPORT_SYMBOL(tcp_mss_to_mtu); |
5d424d5a | 1526 | |
67edfef7 | 1527 | /* MTU probing init per socket */ |
5d424d5a JH |
1528 | void tcp_mtup_init(struct sock *sk) |
1529 | { | |
1530 | struct tcp_sock *tp = tcp_sk(sk); | |
1531 | struct inet_connection_sock *icsk = inet_csk(sk); | |
b0f9ca53 | 1532 | struct net *net = sock_net(sk); |
5d424d5a | 1533 | |
b0f9ca53 | 1534 | icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; |
5d424d5a | 1535 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + |
e905a9ed | 1536 | icsk->icsk_af_ops->net_header_len; |
b0f9ca53 | 1537 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); |
5d424d5a | 1538 | icsk->icsk_mtup.probe_size = 0; |
05cbc0db | 1539 | if (icsk->icsk_mtup.enabled) |
c74df29a | 1540 | icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; |
5d424d5a | 1541 | } |
4bc2f18b | 1542 | EXPORT_SYMBOL(tcp_mtup_init); |
5d424d5a | 1543 | |
1da177e4 LT |
1544 | /* This function synchronize snd mss to current pmtu/exthdr set. |
1545 | ||
1546 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
1547 | for TCP options, but includes only bare TCP header. | |
1548 | ||
1549 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 1550 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
1551 | It also does not include TCP options. |
1552 | ||
d83d8461 | 1553 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
1554 | |
1555 | tp->mss_cache is current effective sending mss, including | |
1556 | all tcp options except for SACKs. It is evaluated, | |
1557 | taking into account current pmtu, but never exceeds | |
1558 | tp->rx_opt.mss_clamp. | |
1559 | ||
1560 | NOTE1. rfc1122 clearly states that advertised MSS | |
1561 | DOES NOT include either tcp or ip options. | |
1562 | ||
d83d8461 ACM |
1563 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
1564 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 | 1565 | */ |
1da177e4 LT |
1566 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
1567 | { | |
1568 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 1569 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 1570 | int mss_now; |
1da177e4 | 1571 | |
5d424d5a JH |
1572 | if (icsk->icsk_mtup.search_high > pmtu) |
1573 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 1574 | |
5d424d5a | 1575 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
409d22b4 | 1576 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
1da177e4 LT |
1577 | |
1578 | /* And store cached results */ | |
d83d8461 | 1579 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
1580 | if (icsk->icsk_mtup.enabled) |
1581 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 1582 | tp->mss_cache = mss_now; |
1da177e4 LT |
1583 | |
1584 | return mss_now; | |
1585 | } | |
4bc2f18b | 1586 | EXPORT_SYMBOL(tcp_sync_mss); |
1da177e4 LT |
1587 | |
1588 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1589 | * and even PMTU discovery events into account. | |
1da177e4 | 1590 | */ |
0c54b85f | 1591 | unsigned int tcp_current_mss(struct sock *sk) |
1da177e4 | 1592 | { |
cf533ea5 ED |
1593 | const struct tcp_sock *tp = tcp_sk(sk); |
1594 | const struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 | 1595 | u32 mss_now; |
95c96174 | 1596 | unsigned int header_len; |
33ad798c AL |
1597 | struct tcp_out_options opts; |
1598 | struct tcp_md5sig_key *md5; | |
c1b4a7e6 DM |
1599 | |
1600 | mss_now = tp->mss_cache; | |
1601 | ||
1da177e4 LT |
1602 | if (dst) { |
1603 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 1604 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
1605 | mss_now = tcp_sync_mss(sk, mtu); |
1606 | } | |
1607 | ||
33ad798c AL |
1608 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
1609 | sizeof(struct tcphdr); | |
1610 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1611 | * some common options. If this is an odd packet (because we have SACK | |
1612 | * blocks etc) then our calculated header_len will be different, and | |
1613 | * we have to adjust mss_now correspondingly */ | |
1614 | if (header_len != tp->tcp_header_len) { | |
1615 | int delta = (int) header_len - tp->tcp_header_len; | |
1616 | mss_now -= delta; | |
1617 | } | |
cfb6eeb4 | 1618 | |
1da177e4 LT |
1619 | return mss_now; |
1620 | } | |
1621 | ||
86fd14ad WP |
1622 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. |
1623 | * As additional protections, we do not touch cwnd in retransmission phases, | |
1624 | * and if application hit its sndbuf limit recently. | |
1625 | */ | |
1626 | static void tcp_cwnd_application_limited(struct sock *sk) | |
1627 | { | |
1628 | struct tcp_sock *tp = tcp_sk(sk); | |
1629 | ||
1630 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && | |
1631 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { | |
1632 | /* Limited by application or receiver window. */ | |
1633 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); | |
1634 | u32 win_used = max(tp->snd_cwnd_used, init_win); | |
1635 | if (win_used < tp->snd_cwnd) { | |
1636 | tp->snd_ssthresh = tcp_current_ssthresh(sk); | |
1637 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; | |
1638 | } | |
1639 | tp->snd_cwnd_used = 0; | |
1640 | } | |
c2203cf7 | 1641 | tp->snd_cwnd_stamp = tcp_jiffies32; |
86fd14ad WP |
1642 | } |
1643 | ||
ca8a2263 | 1644 | static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) |
a762a980 | 1645 | { |
1b1fc3fd | 1646 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; |
9e412ba7 | 1647 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 | 1648 | |
ca8a2263 NC |
1649 | /* Track the maximum number of outstanding packets in each |
1650 | * window, and remember whether we were cwnd-limited then. | |
1651 | */ | |
1652 | if (!before(tp->snd_una, tp->max_packets_seq) || | |
1653 | tp->packets_out > tp->max_packets_out) { | |
1654 | tp->max_packets_out = tp->packets_out; | |
1655 | tp->max_packets_seq = tp->snd_nxt; | |
1656 | tp->is_cwnd_limited = is_cwnd_limited; | |
1657 | } | |
e114a710 | 1658 | |
24901551 | 1659 | if (tcp_is_cwnd_limited(sk)) { |
a762a980 DM |
1660 | /* Network is feed fully. */ |
1661 | tp->snd_cwnd_used = 0; | |
c2203cf7 | 1662 | tp->snd_cwnd_stamp = tcp_jiffies32; |
a762a980 DM |
1663 | } else { |
1664 | /* Network starves. */ | |
1665 | if (tp->packets_out > tp->snd_cwnd_used) | |
1666 | tp->snd_cwnd_used = tp->packets_out; | |
1667 | ||
b510f0d2 | 1668 | if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && |
c2203cf7 | 1669 | (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && |
1b1fc3fd | 1670 | !ca_ops->cong_control) |
a762a980 | 1671 | tcp_cwnd_application_limited(sk); |
b0f71bd3 FY |
1672 | |
1673 | /* The following conditions together indicate the starvation | |
1674 | * is caused by insufficient sender buffer: | |
1675 | * 1) just sent some data (see tcp_write_xmit) | |
1676 | * 2) not cwnd limited (this else condition) | |
75c119af | 1677 | * 3) no more data to send (tcp_write_queue_empty()) |
b0f71bd3 FY |
1678 | * 4) application is hitting buffer limit (SOCK_NOSPACE) |
1679 | */ | |
75c119af | 1680 | if (tcp_write_queue_empty(sk) && sk->sk_socket && |
b0f71bd3 FY |
1681 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && |
1682 | (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) | |
1683 | tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); | |
a762a980 DM |
1684 | } |
1685 | } | |
1686 | ||
d4589926 ED |
1687 | /* Minshall's variant of the Nagle send check. */ |
1688 | static bool tcp_minshall_check(const struct tcp_sock *tp) | |
1689 | { | |
1690 | return after(tp->snd_sml, tp->snd_una) && | |
1691 | !after(tp->snd_sml, tp->snd_nxt); | |
1692 | } | |
1693 | ||
1694 | /* Update snd_sml if this skb is under mss | |
1695 | * Note that a TSO packet might end with a sub-mss segment | |
1696 | * The test is really : | |
1697 | * if ((skb->len % mss) != 0) | |
1698 | * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; | |
1699 | * But we can avoid doing the divide again given we already have | |
1700 | * skb_pcount = skb->len / mss_now | |
0e3a4803 | 1701 | */ |
d4589926 ED |
1702 | static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, |
1703 | const struct sk_buff *skb) | |
1704 | { | |
1705 | if (skb->len < tcp_skb_pcount(skb) * mss_now) | |
1706 | tp->snd_sml = TCP_SKB_CB(skb)->end_seq; | |
1707 | } | |
1708 | ||
1709 | /* Return false, if packet can be sent now without violation Nagle's rules: | |
1710 | * 1. It is full sized. (provided by caller in %partial bool) | |
1711 | * 2. Or it contains FIN. (already checked by caller) | |
1712 | * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. | |
1713 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1714 | * With Minshall's modification: all sent small packets are ACKed. | |
1715 | */ | |
1716 | static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, | |
cc93fc51 | 1717 | int nonagle) |
d4589926 ED |
1718 | { |
1719 | return partial && | |
1720 | ((nonagle & TCP_NAGLE_CORK) || | |
1721 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); | |
1722 | } | |
605ad7f1 ED |
1723 | |
1724 | /* Return how many segs we'd like on a TSO packet, | |
1725 | * to send one TSO packet per ms | |
1726 | */ | |
1b3878ca NC |
1727 | u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, |
1728 | int min_tso_segs) | |
605ad7f1 ED |
1729 | { |
1730 | u32 bytes, segs; | |
1731 | ||
3a9b76fd | 1732 | bytes = min(sk->sk_pacing_rate >> sk->sk_pacing_shift, |
605ad7f1 ED |
1733 | sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); |
1734 | ||
1735 | /* Goal is to send at least one packet per ms, | |
1736 | * not one big TSO packet every 100 ms. | |
1737 | * This preserves ACK clocking and is consistent | |
1738 | * with tcp_tso_should_defer() heuristic. | |
1739 | */ | |
1b3878ca | 1740 | segs = max_t(u32, bytes / mss_now, min_tso_segs); |
605ad7f1 | 1741 | |
37a597f5 | 1742 | return segs; |
605ad7f1 | 1743 | } |
1b3878ca | 1744 | EXPORT_SYMBOL(tcp_tso_autosize); |
605ad7f1 | 1745 | |
ed6e7268 NC |
1746 | /* Return the number of segments we want in the skb we are transmitting. |
1747 | * See if congestion control module wants to decide; otherwise, autosize. | |
1748 | */ | |
1749 | static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) | |
1750 | { | |
1751 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; | |
1752 | u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0; | |
1753 | ||
37a597f5 ED |
1754 | if (!tso_segs) |
1755 | tso_segs = tcp_tso_autosize(sk, mss_now, | |
1756 | sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs); | |
1757 | return min_t(u32, tso_segs, sk->sk_gso_max_segs); | |
ed6e7268 NC |
1758 | } |
1759 | ||
d4589926 ED |
1760 | /* Returns the portion of skb which can be sent right away */ |
1761 | static unsigned int tcp_mss_split_point(const struct sock *sk, | |
1762 | const struct sk_buff *skb, | |
1763 | unsigned int mss_now, | |
1764 | unsigned int max_segs, | |
1765 | int nonagle) | |
c1b4a7e6 | 1766 | { |
cf533ea5 | 1767 | const struct tcp_sock *tp = tcp_sk(sk); |
d4589926 | 1768 | u32 partial, needed, window, max_len; |
c1b4a7e6 | 1769 | |
90840def | 1770 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
1485348d | 1771 | max_len = mss_now * max_segs; |
0e3a4803 | 1772 | |
1485348d BH |
1773 | if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) |
1774 | return max_len; | |
0e3a4803 | 1775 | |
5ea3a748 IJ |
1776 | needed = min(skb->len, window); |
1777 | ||
1485348d BH |
1778 | if (max_len <= needed) |
1779 | return max_len; | |
0e3a4803 | 1780 | |
d4589926 ED |
1781 | partial = needed % mss_now; |
1782 | /* If last segment is not a full MSS, check if Nagle rules allow us | |
1783 | * to include this last segment in this skb. | |
1784 | * Otherwise, we'll split the skb at last MSS boundary | |
1785 | */ | |
cc93fc51 | 1786 | if (tcp_nagle_check(partial != 0, tp, nonagle)) |
d4589926 ED |
1787 | return needed - partial; |
1788 | ||
1789 | return needed; | |
c1b4a7e6 DM |
1790 | } |
1791 | ||
1792 | /* Can at least one segment of SKB be sent right now, according to the | |
1793 | * congestion window rules? If so, return how many segments are allowed. | |
1794 | */ | |
cf533ea5 ED |
1795 | static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, |
1796 | const struct sk_buff *skb) | |
c1b4a7e6 | 1797 | { |
d649a7a8 | 1798 | u32 in_flight, cwnd, halfcwnd; |
c1b4a7e6 DM |
1799 | |
1800 | /* Don't be strict about the congestion window for the final FIN. */ | |
4de075e0 ED |
1801 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
1802 | tcp_skb_pcount(skb) == 1) | |
c1b4a7e6 DM |
1803 | return 1; |
1804 | ||
1805 | in_flight = tcp_packets_in_flight(tp); | |
1806 | cwnd = tp->snd_cwnd; | |
d649a7a8 ED |
1807 | if (in_flight >= cwnd) |
1808 | return 0; | |
c1b4a7e6 | 1809 | |
d649a7a8 ED |
1810 | /* For better scheduling, ensure we have at least |
1811 | * 2 GSO packets in flight. | |
1812 | */ | |
1813 | halfcwnd = max(cwnd >> 1, 1U); | |
1814 | return min(halfcwnd, cwnd - in_flight); | |
c1b4a7e6 DM |
1815 | } |
1816 | ||
b595076a | 1817 | /* Initialize TSO state of a skb. |
67edfef7 | 1818 | * This must be invoked the first time we consider transmitting |
c1b4a7e6 DM |
1819 | * SKB onto the wire. |
1820 | */ | |
5bbb432c | 1821 | static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) |
c1b4a7e6 DM |
1822 | { |
1823 | int tso_segs = tcp_skb_pcount(skb); | |
1824 | ||
f8269a49 | 1825 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
5bbb432c | 1826 | tcp_set_skb_tso_segs(skb, mss_now); |
c1b4a7e6 DM |
1827 | tso_segs = tcp_skb_pcount(skb); |
1828 | } | |
1829 | return tso_segs; | |
1830 | } | |
1831 | ||
c1b4a7e6 | 1832 | |
a2a385d6 | 1833 | /* Return true if the Nagle test allows this packet to be |
c1b4a7e6 DM |
1834 | * sent now. |
1835 | */ | |
a2a385d6 ED |
1836 | static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, |
1837 | unsigned int cur_mss, int nonagle) | |
c1b4a7e6 DM |
1838 | { |
1839 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1840 | * write_queue (they have no chances to get new data). | |
1841 | * | |
1842 | * This is implemented in the callers, where they modify the 'nonagle' | |
1843 | * argument based upon the location of SKB in the send queue. | |
1844 | */ | |
1845 | if (nonagle & TCP_NAGLE_PUSH) | |
a2a385d6 | 1846 | return true; |
c1b4a7e6 | 1847 | |
9b44190d YC |
1848 | /* Don't use the nagle rule for urgent data (or for the final FIN). */ |
1849 | if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) | |
a2a385d6 | 1850 | return true; |
c1b4a7e6 | 1851 | |
cc93fc51 | 1852 | if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) |
a2a385d6 | 1853 | return true; |
c1b4a7e6 | 1854 | |
a2a385d6 | 1855 | return false; |
c1b4a7e6 DM |
1856 | } |
1857 | ||
1858 | /* Does at least the first segment of SKB fit into the send window? */ | |
a2a385d6 ED |
1859 | static bool tcp_snd_wnd_test(const struct tcp_sock *tp, |
1860 | const struct sk_buff *skb, | |
1861 | unsigned int cur_mss) | |
c1b4a7e6 DM |
1862 | { |
1863 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1864 | ||
1865 | if (skb->len > cur_mss) | |
1866 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1867 | ||
90840def | 1868 | return !after(end_seq, tcp_wnd_end(tp)); |
c1b4a7e6 DM |
1869 | } |
1870 | ||
c1b4a7e6 DM |
1871 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet |
1872 | * which is put after SKB on the list. It is very much like | |
1873 | * tcp_fragment() except that it may make several kinds of assumptions | |
1874 | * in order to speed up the splitting operation. In particular, we | |
1875 | * know that all the data is in scatter-gather pages, and that the | |
1876 | * packet has never been sent out before (and thus is not cloned). | |
1877 | */ | |
75c119af ED |
1878 | static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue, |
1879 | struct sk_buff *skb, unsigned int len, | |
c4ead4c5 | 1880 | unsigned int mss_now, gfp_t gfp) |
c1b4a7e6 DM |
1881 | { |
1882 | struct sk_buff *buff; | |
1883 | int nlen = skb->len - len; | |
9ce01461 | 1884 | u8 flags; |
c1b4a7e6 DM |
1885 | |
1886 | /* All of a TSO frame must be composed of paged data. */ | |
c8ac3774 | 1887 | if (skb->len != skb->data_len) |
75c119af | 1888 | return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp); |
c1b4a7e6 | 1889 | |
eb934478 | 1890 | buff = sk_stream_alloc_skb(sk, 0, gfp, true); |
51456b29 | 1891 | if (unlikely(!buff)) |
c1b4a7e6 DM |
1892 | return -ENOMEM; |
1893 | ||
3ab224be HA |
1894 | sk->sk_wmem_queued += buff->truesize; |
1895 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea | 1896 | buff->truesize += nlen; |
c1b4a7e6 DM |
1897 | skb->truesize -= nlen; |
1898 | ||
1899 | /* Correct the sequence numbers. */ | |
1900 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1901 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1902 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1903 | ||
1904 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1905 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1906 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1907 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
c1b4a7e6 DM |
1908 | |
1909 | /* This packet was never sent out yet, so no SACK bits. */ | |
1910 | TCP_SKB_CB(buff)->sacked = 0; | |
1911 | ||
a166140e MKL |
1912 | tcp_skb_fragment_eor(skb, buff); |
1913 | ||
84fa7933 | 1914 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
c1b4a7e6 | 1915 | skb_split(skb, buff, len); |
490cc7d0 | 1916 | tcp_fragment_tstamp(skb, buff); |
c1b4a7e6 DM |
1917 | |
1918 | /* Fix up tso_factor for both original and new SKB. */ | |
5bbb432c ED |
1919 | tcp_set_skb_tso_segs(skb, mss_now); |
1920 | tcp_set_skb_tso_segs(buff, mss_now); | |
c1b4a7e6 DM |
1921 | |
1922 | /* Link BUFF into the send queue. */ | |
f4a775d1 | 1923 | __skb_header_release(buff); |
75c119af | 1924 | tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); |
c1b4a7e6 DM |
1925 | |
1926 | return 0; | |
1927 | } | |
1928 | ||
1929 | /* Try to defer sending, if possible, in order to minimize the amount | |
1930 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
1931 | * | |
1932 | * This algorithm is from John Heffner. | |
1933 | */ | |
ca8a2263 | 1934 | static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, |
118c45be ED |
1935 | bool *is_cwnd_limited, |
1936 | bool *is_rwnd_limited, | |
1937 | u32 max_segs) | |
c1b4a7e6 | 1938 | { |
6687e988 | 1939 | const struct inet_connection_sock *icsk = inet_csk(sk); |
50c8339e ED |
1940 | u32 age, send_win, cong_win, limit, in_flight; |
1941 | struct tcp_sock *tp = tcp_sk(sk); | |
50c8339e | 1942 | struct sk_buff *head; |
ad9f4f50 | 1943 | int win_divisor; |
c1b4a7e6 | 1944 | |
99d7662a | 1945 | if (icsk->icsk_ca_state >= TCP_CA_Recovery) |
ae8064ac JH |
1946 | goto send_now; |
1947 | ||
5f852eb5 ED |
1948 | /* Avoid bursty behavior by allowing defer |
1949 | * only if the last write was recent. | |
1950 | */ | |
d635fbe2 | 1951 | if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0) |
ae8064ac | 1952 | goto send_now; |
908a75c1 | 1953 | |
c1b4a7e6 DM |
1954 | in_flight = tcp_packets_in_flight(tp); |
1955 | ||
056834d9 | 1956 | BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); |
c1b4a7e6 | 1957 | |
90840def | 1958 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 DM |
1959 | |
1960 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
1961 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
1962 | ||
1963 | limit = min(send_win, cong_win); | |
1964 | ||
ba244fe9 | 1965 | /* If a full-sized TSO skb can be sent, do it. */ |
605ad7f1 | 1966 | if (limit >= max_segs * tp->mss_cache) |
ae8064ac | 1967 | goto send_now; |
ba244fe9 | 1968 | |
62ad2761 IJ |
1969 | /* Middle in queue won't get any more data, full sendable already? */ |
1970 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) | |
1971 | goto send_now; | |
1972 | ||
5bbcc0f5 | 1973 | win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); |
ad9f4f50 | 1974 | if (win_divisor) { |
c1b4a7e6 DM |
1975 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
1976 | ||
1977 | /* If at least some fraction of a window is available, | |
1978 | * just use it. | |
1979 | */ | |
ad9f4f50 | 1980 | chunk /= win_divisor; |
c1b4a7e6 | 1981 | if (limit >= chunk) |
ae8064ac | 1982 | goto send_now; |
c1b4a7e6 DM |
1983 | } else { |
1984 | /* Different approach, try not to defer past a single | |
1985 | * ACK. Receiver should ACK every other full sized | |
1986 | * frame, so if we have space for more than 3 frames | |
1987 | * then send now. | |
1988 | */ | |
6b5a5c0d | 1989 | if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) |
ae8064ac | 1990 | goto send_now; |
c1b4a7e6 DM |
1991 | } |
1992 | ||
75c119af ED |
1993 | /* TODO : use tsorted_sent_queue ? */ |
1994 | head = tcp_rtx_queue_head(sk); | |
1995 | if (!head) | |
1996 | goto send_now; | |
9a568de4 | 1997 | age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp); |
50c8339e ED |
1998 | /* If next ACK is likely to come too late (half srtt), do not defer */ |
1999 | if (age < (tp->srtt_us >> 4)) | |
2000 | goto send_now; | |
2001 | ||
118c45be ED |
2002 | /* Ok, it looks like it is advisable to defer. |
2003 | * Three cases are tracked : | |
2004 | * 1) We are cwnd-limited | |
2005 | * 2) We are rwnd-limited | |
2006 | * 3) We are application limited. | |
2007 | */ | |
2008 | if (cong_win < send_win) { | |
2009 | if (cong_win <= skb->len) { | |
2010 | *is_cwnd_limited = true; | |
2011 | return true; | |
2012 | } | |
2013 | } else { | |
2014 | if (send_win <= skb->len) { | |
2015 | *is_rwnd_limited = true; | |
2016 | return true; | |
2017 | } | |
2018 | } | |
ae8064ac | 2019 | |
118c45be ED |
2020 | /* If this packet won't get more data, do not wait. */ |
2021 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) | |
2022 | goto send_now; | |
ca8a2263 | 2023 | |
a2a385d6 | 2024 | return true; |
ae8064ac JH |
2025 | |
2026 | send_now: | |
a2a385d6 | 2027 | return false; |
c1b4a7e6 DM |
2028 | } |
2029 | ||
05cbc0db FD |
2030 | static inline void tcp_mtu_check_reprobe(struct sock *sk) |
2031 | { | |
2032 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2033 | struct tcp_sock *tp = tcp_sk(sk); | |
2034 | struct net *net = sock_net(sk); | |
2035 | u32 interval; | |
2036 | s32 delta; | |
2037 | ||
2038 | interval = net->ipv4.sysctl_tcp_probe_interval; | |
c74df29a | 2039 | delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; |
05cbc0db FD |
2040 | if (unlikely(delta >= interval * HZ)) { |
2041 | int mss = tcp_current_mss(sk); | |
2042 | ||
2043 | /* Update current search range */ | |
2044 | icsk->icsk_mtup.probe_size = 0; | |
2045 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + | |
2046 | sizeof(struct tcphdr) + | |
2047 | icsk->icsk_af_ops->net_header_len; | |
2048 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); | |
2049 | ||
2050 | /* Update probe time stamp */ | |
c74df29a | 2051 | icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; |
05cbc0db FD |
2052 | } |
2053 | } | |
2054 | ||
fce441bb IL |
2055 | static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) |
2056 | { | |
2057 | struct sk_buff *skb, *next; | |
2058 | ||
2059 | skb = tcp_send_head(sk); | |
2060 | tcp_for_write_queue_from_safe(skb, next, sk) { | |
2061 | if (len <= skb->len) | |
2062 | break; | |
2063 | ||
9f83cdc3 | 2064 | if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb)) |
fce441bb IL |
2065 | return false; |
2066 | ||
2067 | len -= skb->len; | |
2068 | } | |
2069 | ||
2070 | return true; | |
2071 | } | |
2072 | ||
5d424d5a | 2073 | /* Create a new MTU probe if we are ready. |
67edfef7 AK |
2074 | * MTU probe is regularly attempting to increase the path MTU by |
2075 | * deliberately sending larger packets. This discovers routing | |
2076 | * changes resulting in larger path MTUs. | |
2077 | * | |
5d424d5a JH |
2078 | * Returns 0 if we should wait to probe (no cwnd available), |
2079 | * 1 if a probe was sent, | |
056834d9 IJ |
2080 | * -1 otherwise |
2081 | */ | |
5d424d5a JH |
2082 | static int tcp_mtu_probe(struct sock *sk) |
2083 | { | |
5d424d5a | 2084 | struct inet_connection_sock *icsk = inet_csk(sk); |
12a59abc | 2085 | struct tcp_sock *tp = tcp_sk(sk); |
5d424d5a | 2086 | struct sk_buff *skb, *nskb, *next; |
6b58e0a5 | 2087 | struct net *net = sock_net(sk); |
5d424d5a | 2088 | int probe_size; |
91cc17c0 | 2089 | int size_needed; |
12a59abc | 2090 | int copy, len; |
5d424d5a | 2091 | int mss_now; |
6b58e0a5 | 2092 | int interval; |
5d424d5a JH |
2093 | |
2094 | /* Not currently probing/verifying, | |
2095 | * not in recovery, | |
2096 | * have enough cwnd, and | |
12a59abc ED |
2097 | * not SACKing (the variable headers throw things off) |
2098 | */ | |
2099 | if (likely(!icsk->icsk_mtup.enabled || | |
2100 | icsk->icsk_mtup.probe_size || | |
2101 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
2102 | tp->snd_cwnd < 11 || | |
2103 | tp->rx_opt.num_sacks || tp->rx_opt.dsack)) | |
5d424d5a JH |
2104 | return -1; |
2105 | ||
6b58e0a5 FD |
2106 | /* Use binary search for probe_size between tcp_mss_base, |
2107 | * and current mss_clamp. if (search_high - search_low) | |
2108 | * smaller than a threshold, backoff from probing. | |
2109 | */ | |
0c54b85f | 2110 | mss_now = tcp_current_mss(sk); |
6b58e0a5 FD |
2111 | probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + |
2112 | icsk->icsk_mtup.search_low) >> 1); | |
91cc17c0 | 2113 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
6b58e0a5 | 2114 | interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; |
05cbc0db FD |
2115 | /* When misfortune happens, we are reprobing actively, |
2116 | * and then reprobe timer has expired. We stick with current | |
2117 | * probing process by not resetting search range to its orignal. | |
2118 | */ | |
6b58e0a5 | 2119 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || |
05cbc0db FD |
2120 | interval < net->ipv4.sysctl_tcp_probe_threshold) { |
2121 | /* Check whether enough time has elaplased for | |
2122 | * another round of probing. | |
2123 | */ | |
2124 | tcp_mtu_check_reprobe(sk); | |
5d424d5a JH |
2125 | return -1; |
2126 | } | |
2127 | ||
2128 | /* Have enough data in the send queue to probe? */ | |
7f9c33e5 | 2129 | if (tp->write_seq - tp->snd_nxt < size_needed) |
5d424d5a JH |
2130 | return -1; |
2131 | ||
91cc17c0 IJ |
2132 | if (tp->snd_wnd < size_needed) |
2133 | return -1; | |
90840def | 2134 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
91cc17c0 | 2135 | return 0; |
5d424d5a | 2136 | |
d67c58e9 IJ |
2137 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
2138 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { | |
2139 | if (!tcp_packets_in_flight(tp)) | |
5d424d5a JH |
2140 | return -1; |
2141 | else | |
2142 | return 0; | |
2143 | } | |
2144 | ||
fce441bb IL |
2145 | if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) |
2146 | return -1; | |
2147 | ||
5d424d5a | 2148 | /* We're allowed to probe. Build it now. */ |
eb934478 | 2149 | nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); |
51456b29 | 2150 | if (!nskb) |
5d424d5a | 2151 | return -1; |
3ab224be HA |
2152 | sk->sk_wmem_queued += nskb->truesize; |
2153 | sk_mem_charge(sk, nskb->truesize); | |
5d424d5a | 2154 | |
fe067e8a | 2155 | skb = tcp_send_head(sk); |
5d424d5a JH |
2156 | |
2157 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
2158 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
4de075e0 | 2159 | TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; |
5d424d5a JH |
2160 | TCP_SKB_CB(nskb)->sacked = 0; |
2161 | nskb->csum = 0; | |
84fa7933 | 2162 | nskb->ip_summed = skb->ip_summed; |
5d424d5a | 2163 | |
50c4817e | 2164 | tcp_insert_write_queue_before(nskb, skb, sk); |
2b7cda9c | 2165 | tcp_highest_sack_replace(sk, skb, nskb); |
50c4817e | 2166 | |
5d424d5a | 2167 | len = 0; |
234b6860 | 2168 | tcp_for_write_queue_from_safe(skb, next, sk) { |
5d424d5a | 2169 | copy = min_t(int, skb->len, probe_size - len); |
2fe664f1 | 2170 | if (nskb->ip_summed) { |
5d424d5a | 2171 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); |
2fe664f1 DCS |
2172 | } else { |
2173 | __wsum csum = skb_copy_and_csum_bits(skb, 0, | |
2174 | skb_put(nskb, copy), | |
2175 | copy, 0); | |
2176 | nskb->csum = csum_block_add(nskb->csum, csum, len); | |
2177 | } | |
5d424d5a JH |
2178 | |
2179 | if (skb->len <= copy) { | |
2180 | /* We've eaten all the data from this skb. | |
2181 | * Throw it away. */ | |
4de075e0 | 2182 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
fce441bb IL |
2183 | /* If this is the last SKB we copy and eor is set |
2184 | * we need to propagate it to the new skb. | |
2185 | */ | |
2186 | TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; | |
9f83cdc3 | 2187 | tcp_skb_collapse_tstamp(nskb, skb); |
fe067e8a | 2188 | tcp_unlink_write_queue(skb, sk); |
3ab224be | 2189 | sk_wmem_free_skb(sk, skb); |
5d424d5a | 2190 | } else { |
4de075e0 | 2191 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & |
a3433f35 | 2192 | ~(TCPHDR_FIN|TCPHDR_PSH); |
5d424d5a JH |
2193 | if (!skb_shinfo(skb)->nr_frags) { |
2194 | skb_pull(skb, copy); | |
84fa7933 | 2195 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
056834d9 IJ |
2196 | skb->csum = csum_partial(skb->data, |
2197 | skb->len, 0); | |
5d424d5a JH |
2198 | } else { |
2199 | __pskb_trim_head(skb, copy); | |
5bbb432c | 2200 | tcp_set_skb_tso_segs(skb, mss_now); |
5d424d5a JH |
2201 | } |
2202 | TCP_SKB_CB(skb)->seq += copy; | |
2203 | } | |
2204 | ||
2205 | len += copy; | |
234b6860 IJ |
2206 | |
2207 | if (len >= probe_size) | |
2208 | break; | |
5d424d5a | 2209 | } |
5bbb432c | 2210 | tcp_init_tso_segs(nskb, nskb->len); |
5d424d5a JH |
2211 | |
2212 | /* We're ready to send. If this fails, the probe will | |
7faee5c0 ED |
2213 | * be resegmented into mss-sized pieces by tcp_write_xmit(). |
2214 | */ | |
5d424d5a JH |
2215 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { |
2216 | /* Decrement cwnd here because we are sending | |
056834d9 | 2217 | * effectively two packets. */ |
5d424d5a | 2218 | tp->snd_cwnd--; |
66f5fe62 | 2219 | tcp_event_new_data_sent(sk, nskb); |
5d424d5a JH |
2220 | |
2221 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
0e7b1368 JH |
2222 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
2223 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
5d424d5a JH |
2224 | |
2225 | return 1; | |
2226 | } | |
2227 | ||
2228 | return -1; | |
2229 | } | |
2230 | ||
218af599 ED |
2231 | static bool tcp_pacing_check(const struct sock *sk) |
2232 | { | |
2233 | return tcp_needs_internal_pacing(sk) && | |
2234 | hrtimer_active(&tcp_sk(sk)->pacing_timer); | |
2235 | } | |
2236 | ||
f9616c35 ED |
2237 | /* TCP Small Queues : |
2238 | * Control number of packets in qdisc/devices to two packets / or ~1 ms. | |
2239 | * (These limits are doubled for retransmits) | |
2240 | * This allows for : | |
2241 | * - better RTT estimation and ACK scheduling | |
2242 | * - faster recovery | |
2243 | * - high rates | |
2244 | * Alas, some drivers / subsystems require a fair amount | |
2245 | * of queued bytes to ensure line rate. | |
2246 | * One example is wifi aggregation (802.11 AMPDU) | |
2247 | */ | |
2248 | static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, | |
2249 | unsigned int factor) | |
2250 | { | |
2251 | unsigned int limit; | |
2252 | ||
3a9b76fd | 2253 | limit = max(2 * skb->truesize, sk->sk_pacing_rate >> sk->sk_pacing_shift); |
9184d8bb ED |
2254 | limit = min_t(u32, limit, |
2255 | sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); | |
f9616c35 ED |
2256 | limit <<= factor; |
2257 | ||
14afee4b | 2258 | if (refcount_read(&sk->sk_wmem_alloc) > limit) { |
75c119af | 2259 | /* Always send skb if rtx queue is empty. |
75eefc6c ED |
2260 | * No need to wait for TX completion to call us back, |
2261 | * after softirq/tasklet schedule. | |
2262 | * This helps when TX completions are delayed too much. | |
2263 | */ | |
75c119af | 2264 | if (tcp_rtx_queue_empty(sk)) |
75eefc6c ED |
2265 | return false; |
2266 | ||
7aa5470c | 2267 | set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); |
f9616c35 ED |
2268 | /* It is possible TX completion already happened |
2269 | * before we set TSQ_THROTTLED, so we must | |
2270 | * test again the condition. | |
2271 | */ | |
2272 | smp_mb__after_atomic(); | |
14afee4b | 2273 | if (refcount_read(&sk->sk_wmem_alloc) > limit) |
f9616c35 ED |
2274 | return true; |
2275 | } | |
2276 | return false; | |
2277 | } | |
2278 | ||
05b055e8 FY |
2279 | static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) |
2280 | { | |
628174cc | 2281 | const u32 now = tcp_jiffies32; |
efe967cd | 2282 | enum tcp_chrono old = tp->chrono_type; |
05b055e8 | 2283 | |
efe967cd AB |
2284 | if (old > TCP_CHRONO_UNSPEC) |
2285 | tp->chrono_stat[old - 1] += now - tp->chrono_start; | |
05b055e8 FY |
2286 | tp->chrono_start = now; |
2287 | tp->chrono_type = new; | |
2288 | } | |
2289 | ||
2290 | void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) | |
2291 | { | |
2292 | struct tcp_sock *tp = tcp_sk(sk); | |
2293 | ||
2294 | /* If there are multiple conditions worthy of tracking in a | |
0f87230d FY |
2295 | * chronograph then the highest priority enum takes precedence |
2296 | * over the other conditions. So that if something "more interesting" | |
05b055e8 FY |
2297 | * starts happening, stop the previous chrono and start a new one. |
2298 | */ | |
2299 | if (type > tp->chrono_type) | |
2300 | tcp_chrono_set(tp, type); | |
2301 | } | |
2302 | ||
2303 | void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) | |
2304 | { | |
2305 | struct tcp_sock *tp = tcp_sk(sk); | |
2306 | ||
0f87230d FY |
2307 | |
2308 | /* There are multiple conditions worthy of tracking in a | |
2309 | * chronograph, so that the highest priority enum takes | |
2310 | * precedence over the other conditions (see tcp_chrono_start). | |
2311 | * If a condition stops, we only stop chrono tracking if | |
2312 | * it's the "most interesting" or current chrono we are | |
2313 | * tracking and starts busy chrono if we have pending data. | |
2314 | */ | |
75c119af | 2315 | if (tcp_rtx_and_write_queues_empty(sk)) |
0f87230d FY |
2316 | tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); |
2317 | else if (type == tp->chrono_type) | |
2318 | tcp_chrono_set(tp, TCP_CHRONO_BUSY); | |
05b055e8 FY |
2319 | } |
2320 | ||
1da177e4 LT |
2321 | /* This routine writes packets to the network. It advances the |
2322 | * send_head. This happens as incoming acks open up the remote | |
2323 | * window for us. | |
2324 | * | |
f8269a49 IJ |
2325 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
2326 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into | |
2327 | * account rare use of URG, this is not a big flaw. | |
2328 | * | |
6ba8a3b1 ND |
2329 | * Send at most one packet when push_one > 0. Temporarily ignore |
2330 | * cwnd limit to force at most one packet out when push_one == 2. | |
2331 | ||
a2a385d6 ED |
2332 | * Returns true, if no segments are in flight and we have queued segments, |
2333 | * but cannot send anything now because of SWS or another problem. | |
1da177e4 | 2334 | */ |
a2a385d6 ED |
2335 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
2336 | int push_one, gfp_t gfp) | |
1da177e4 LT |
2337 | { |
2338 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 2339 | struct sk_buff *skb; |
c1b4a7e6 DM |
2340 | unsigned int tso_segs, sent_pkts; |
2341 | int cwnd_quota; | |
5d424d5a | 2342 | int result; |
5615f886 | 2343 | bool is_cwnd_limited = false, is_rwnd_limited = false; |
605ad7f1 | 2344 | u32 max_segs; |
1da177e4 | 2345 | |
92df7b51 | 2346 | sent_pkts = 0; |
5d424d5a | 2347 | |
ee1836ae | 2348 | tcp_mstamp_refresh(tp); |
d5dd9175 IJ |
2349 | if (!push_one) { |
2350 | /* Do MTU probing. */ | |
2351 | result = tcp_mtu_probe(sk); | |
2352 | if (!result) { | |
a2a385d6 | 2353 | return false; |
d5dd9175 IJ |
2354 | } else if (result > 0) { |
2355 | sent_pkts = 1; | |
2356 | } | |
5d424d5a JH |
2357 | } |
2358 | ||
ed6e7268 | 2359 | max_segs = tcp_tso_segs(sk, mss_now); |
fe067e8a | 2360 | while ((skb = tcp_send_head(sk))) { |
c8ac3774 HX |
2361 | unsigned int limit; |
2362 | ||
218af599 ED |
2363 | if (tcp_pacing_check(sk)) |
2364 | break; | |
2365 | ||
5bbb432c | 2366 | tso_segs = tcp_init_tso_segs(skb, mss_now); |
c1b4a7e6 | 2367 | BUG_ON(!tso_segs); |
aa93466b | 2368 | |
9d186cac | 2369 | if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { |
7faee5c0 | 2370 | /* "skb_mstamp" is used as a start point for the retransmit timer */ |
e2080072 | 2371 | tcp_update_skb_after_send(tp, skb); |
ec342325 | 2372 | goto repair; /* Skip network transmission */ |
9d186cac | 2373 | } |
ec342325 | 2374 | |
b68e9f85 | 2375 | cwnd_quota = tcp_cwnd_test(tp, skb); |
6ba8a3b1 ND |
2376 | if (!cwnd_quota) { |
2377 | if (push_one == 2) | |
2378 | /* Force out a loss probe pkt. */ | |
2379 | cwnd_quota = 1; | |
2380 | else | |
2381 | break; | |
2382 | } | |
b68e9f85 | 2383 | |
5615f886 FY |
2384 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { |
2385 | is_rwnd_limited = true; | |
b68e9f85 | 2386 | break; |
5615f886 | 2387 | } |
b68e9f85 | 2388 | |
d6a4e26a | 2389 | if (tso_segs == 1) { |
c1b4a7e6 DM |
2390 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, |
2391 | (tcp_skb_is_last(sk, skb) ? | |
2392 | nonagle : TCP_NAGLE_PUSH)))) | |
2393 | break; | |
2394 | } else { | |
ca8a2263 | 2395 | if (!push_one && |
605ad7f1 | 2396 | tcp_tso_should_defer(sk, skb, &is_cwnd_limited, |
118c45be | 2397 | &is_rwnd_limited, max_segs)) |
c1b4a7e6 DM |
2398 | break; |
2399 | } | |
aa93466b | 2400 | |
605ad7f1 | 2401 | limit = mss_now; |
d6a4e26a | 2402 | if (tso_segs > 1 && !tcp_urg_mode(tp)) |
605ad7f1 ED |
2403 | limit = tcp_mss_split_point(sk, skb, mss_now, |
2404 | min_t(unsigned int, | |
2405 | cwnd_quota, | |
2406 | max_segs), | |
2407 | nonagle); | |
2408 | ||
2409 | if (skb->len > limit && | |
75c119af ED |
2410 | unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, |
2411 | skb, limit, mss_now, gfp))) | |
605ad7f1 ED |
2412 | break; |
2413 | ||
7aa5470c ED |
2414 | if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) |
2415 | clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); | |
f9616c35 ED |
2416 | if (tcp_small_queue_check(sk, skb, 0)) |
2417 | break; | |
c9eeec26 | 2418 | |
d5dd9175 | 2419 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
92df7b51 | 2420 | break; |
1da177e4 | 2421 | |
ec342325 | 2422 | repair: |
92df7b51 DM |
2423 | /* Advance the send_head. This one is sent out. |
2424 | * This call will increment packets_out. | |
2425 | */ | |
66f5fe62 | 2426 | tcp_event_new_data_sent(sk, skb); |
1da177e4 | 2427 | |
92df7b51 | 2428 | tcp_minshall_update(tp, mss_now, skb); |
a262f0cd | 2429 | sent_pkts += tcp_skb_pcount(skb); |
d5dd9175 IJ |
2430 | |
2431 | if (push_one) | |
2432 | break; | |
92df7b51 | 2433 | } |
1da177e4 | 2434 | |
5615f886 FY |
2435 | if (is_rwnd_limited) |
2436 | tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); | |
2437 | else | |
2438 | tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); | |
2439 | ||
aa93466b | 2440 | if (likely(sent_pkts)) { |
684bad11 YC |
2441 | if (tcp_in_cwnd_reduction(sk)) |
2442 | tp->prr_out += sent_pkts; | |
6ba8a3b1 ND |
2443 | |
2444 | /* Send one loss probe per tail loss episode. */ | |
2445 | if (push_one != 2) | |
ed66dfaf | 2446 | tcp_schedule_loss_probe(sk, false); |
d2e1339f | 2447 | is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); |
ca8a2263 | 2448 | tcp_cwnd_validate(sk, is_cwnd_limited); |
a2a385d6 | 2449 | return false; |
1da177e4 | 2450 | } |
75c119af | 2451 | return !tp->packets_out && !tcp_write_queue_empty(sk); |
6ba8a3b1 ND |
2452 | } |
2453 | ||
ed66dfaf | 2454 | bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) |
6ba8a3b1 ND |
2455 | { |
2456 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2457 | struct tcp_sock *tp = tcp_sk(sk); | |
a2815817 | 2458 | u32 timeout, rto_delta_us; |
2ae21cf5 | 2459 | int early_retrans; |
6ba8a3b1 | 2460 | |
6ba8a3b1 ND |
2461 | /* Don't do any loss probe on a Fast Open connection before 3WHS |
2462 | * finishes. | |
2463 | */ | |
f9b99582 | 2464 | if (tp->fastopen_rsk) |
6ba8a3b1 ND |
2465 | return false; |
2466 | ||
2ae21cf5 | 2467 | early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; |
6ba8a3b1 | 2468 | /* Schedule a loss probe in 2*RTT for SACK capable connections |
d09cb159 | 2469 | * not in loss recovery, that are either limited by cwnd or application. |
6ba8a3b1 | 2470 | */ |
2ae21cf5 | 2471 | if ((early_retrans != 3 && early_retrans != 4) || |
bec41a11 | 2472 | !tp->packets_out || !tcp_is_sack(tp) || |
d09cb159 NC |
2473 | (icsk->icsk_ca_state != TCP_CA_Open && |
2474 | icsk->icsk_ca_state != TCP_CA_CWR)) | |
6ba8a3b1 ND |
2475 | return false; |
2476 | ||
bb4d991a | 2477 | /* Probe timeout is 2*rtt. Add minimum RTO to account |
f9b99582 YC |
2478 | * for delayed ack when there's one outstanding packet. If no RTT |
2479 | * sample is available then probe after TCP_TIMEOUT_INIT. | |
6ba8a3b1 | 2480 | */ |
bb4d991a YC |
2481 | if (tp->srtt_us) { |
2482 | timeout = usecs_to_jiffies(tp->srtt_us >> 2); | |
2483 | if (tp->packets_out == 1) | |
2484 | timeout += TCP_RTO_MIN; | |
2485 | else | |
2486 | timeout += TCP_TIMEOUT_MIN; | |
2487 | } else { | |
2488 | timeout = TCP_TIMEOUT_INIT; | |
2489 | } | |
6ba8a3b1 | 2490 | |
a2815817 | 2491 | /* If the RTO formula yields an earlier time, then use that time. */ |
ed66dfaf NC |
2492 | rto_delta_us = advancing_rto ? |
2493 | jiffies_to_usecs(inet_csk(sk)->icsk_rto) : | |
2494 | tcp_rto_delta_us(sk); /* How far in future is RTO? */ | |
a2815817 NC |
2495 | if (rto_delta_us > 0) |
2496 | timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); | |
6ba8a3b1 ND |
2497 | |
2498 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, | |
2499 | TCP_RTO_MAX); | |
2500 | return true; | |
2501 | } | |
2502 | ||
1f3279ae ED |
2503 | /* Thanks to skb fast clones, we can detect if a prior transmit of |
2504 | * a packet is still in a qdisc or driver queue. | |
2505 | * In this case, there is very little point doing a retransmit ! | |
1f3279ae ED |
2506 | */ |
2507 | static bool skb_still_in_host_queue(const struct sock *sk, | |
2508 | const struct sk_buff *skb) | |
2509 | { | |
39bb5e62 | 2510 | if (unlikely(skb_fclone_busy(sk, skb))) { |
c10d9310 ED |
2511 | NET_INC_STATS(sock_net(sk), |
2512 | LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); | |
1f3279ae ED |
2513 | return true; |
2514 | } | |
2515 | return false; | |
2516 | } | |
2517 | ||
b340b264 | 2518 | /* When probe timeout (PTO) fires, try send a new segment if possible, else |
6ba8a3b1 ND |
2519 | * retransmit the last segment. |
2520 | */ | |
2521 | void tcp_send_loss_probe(struct sock *sk) | |
2522 | { | |
9b717a8d | 2523 | struct tcp_sock *tp = tcp_sk(sk); |
6ba8a3b1 ND |
2524 | struct sk_buff *skb; |
2525 | int pcount; | |
2526 | int mss = tcp_current_mss(sk); | |
6ba8a3b1 | 2527 | |
b340b264 | 2528 | skb = tcp_send_head(sk); |
75c119af ED |
2529 | if (skb && tcp_snd_wnd_test(tp, skb, mss)) { |
2530 | pcount = tp->packets_out; | |
2531 | tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); | |
2532 | if (tp->packets_out > pcount) | |
2533 | goto probe_sent; | |
2534 | goto rearm_timer; | |
6ba8a3b1 | 2535 | } |
75c119af | 2536 | skb = skb_rb_last(&sk->tcp_rtx_queue); |
6ba8a3b1 | 2537 | |
0e660025 YC |
2538 | if (unlikely(!skb)) { |
2539 | WARN_ONCE(tp->packets_out, | |
2540 | "invalid inflight: %u state %u cwnd %u mss %d\n", | |
2541 | tp->packets_out, sk->sk_state, tp->snd_cwnd, mss); | |
2542 | inet_csk(sk)->icsk_pending = 0; | |
2543 | return; | |
2544 | } | |
2545 | ||
9b717a8d ND |
2546 | /* At most one outstanding TLP retransmission. */ |
2547 | if (tp->tlp_high_seq) | |
2548 | goto rearm_timer; | |
2549 | ||
1f3279ae ED |
2550 | if (skb_still_in_host_queue(sk, skb)) |
2551 | goto rearm_timer; | |
2552 | ||
6ba8a3b1 ND |
2553 | pcount = tcp_skb_pcount(skb); |
2554 | if (WARN_ON(!pcount)) | |
2555 | goto rearm_timer; | |
2556 | ||
2557 | if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { | |
75c119af ED |
2558 | if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, |
2559 | (pcount - 1) * mss, mss, | |
6cc55e09 | 2560 | GFP_ATOMIC))) |
6ba8a3b1 | 2561 | goto rearm_timer; |
75c119af | 2562 | skb = skb_rb_next(skb); |
6ba8a3b1 ND |
2563 | } |
2564 | ||
2565 | if (WARN_ON(!skb || !tcp_skb_pcount(skb))) | |
2566 | goto rearm_timer; | |
2567 | ||
10d3be56 | 2568 | if (__tcp_retransmit_skb(sk, skb, 1)) |
b340b264 | 2569 | goto rearm_timer; |
6ba8a3b1 | 2570 | |
9b717a8d | 2571 | /* Record snd_nxt for loss detection. */ |
b340b264 | 2572 | tp->tlp_high_seq = tp->snd_nxt; |
9b717a8d | 2573 | |
b340b264 | 2574 | probe_sent: |
c10d9310 | 2575 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); |
b340b264 YC |
2576 | /* Reset s.t. tcp_rearm_rto will restart timer from now */ |
2577 | inet_csk(sk)->icsk_pending = 0; | |
6ba8a3b1 | 2578 | rearm_timer: |
fcd16c0a | 2579 | tcp_rearm_rto(sk); |
1da177e4 LT |
2580 | } |
2581 | ||
a762a980 DM |
2582 | /* Push out any pending frames which were held back due to |
2583 | * TCP_CORK or attempt at coalescing tiny packets. | |
2584 | * The socket must be locked by the caller. | |
2585 | */ | |
9e412ba7 IJ |
2586 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
2587 | int nonagle) | |
a762a980 | 2588 | { |
726e07a8 IJ |
2589 | /* If we are closed, the bytes will have to remain here. |
2590 | * In time closedown will finish, we empty the write queue and | |
2591 | * all will be happy. | |
2592 | */ | |
2593 | if (unlikely(sk->sk_state == TCP_CLOSE)) | |
2594 | return; | |
2595 | ||
99a1dec7 | 2596 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, |
7450aaf6 | 2597 | sk_gfp_mask(sk, GFP_ATOMIC))) |
726e07a8 | 2598 | tcp_check_probe_timer(sk); |
a762a980 DM |
2599 | } |
2600 | ||
c1b4a7e6 DM |
2601 | /* Send _single_ skb sitting at the send head. This function requires |
2602 | * true push pending frames to setup probe timer etc. | |
2603 | */ | |
2604 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
2605 | { | |
fe067e8a | 2606 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
2607 | |
2608 | BUG_ON(!skb || skb->len < mss_now); | |
2609 | ||
d5dd9175 | 2610 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); |
c1b4a7e6 DM |
2611 | } |
2612 | ||
1da177e4 LT |
2613 | /* This function returns the amount that we can raise the |
2614 | * usable window based on the following constraints | |
e905a9ed | 2615 | * |
1da177e4 LT |
2616 | * 1. The window can never be shrunk once it is offered (RFC 793) |
2617 | * 2. We limit memory per socket | |
2618 | * | |
2619 | * RFC 1122: | |
2620 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
2621 | * RECV.NEXT + RCV.WIN fixed until: | |
2622 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
2623 | * | |
2624 | * i.e. don't raise the right edge of the window until you can raise | |
2625 | * it at least MSS bytes. | |
2626 | * | |
2627 | * Unfortunately, the recommended algorithm breaks header prediction, | |
2628 | * since header prediction assumes th->window stays fixed. | |
2629 | * | |
2630 | * Strictly speaking, keeping th->window fixed violates the receiver | |
2631 | * side SWS prevention criteria. The problem is that under this rule | |
2632 | * a stream of single byte packets will cause the right side of the | |
2633 | * window to always advance by a single byte. | |
e905a9ed | 2634 | * |
1da177e4 LT |
2635 | * Of course, if the sender implements sender side SWS prevention |
2636 | * then this will not be a problem. | |
e905a9ed | 2637 | * |
1da177e4 | 2638 | * BSD seems to make the following compromise: |
e905a9ed | 2639 | * |
1da177e4 LT |
2640 | * If the free space is less than the 1/4 of the maximum |
2641 | * space available and the free space is less than 1/2 mss, | |
2642 | * then set the window to 0. | |
2643 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
2644 | * Otherwise, just prevent the window from shrinking | |
2645 | * and from being larger than the largest representable value. | |
2646 | * | |
2647 | * This prevents incremental opening of the window in the regime | |
2648 | * where TCP is limited by the speed of the reader side taking | |
2649 | * data out of the TCP receive queue. It does nothing about | |
2650 | * those cases where the window is constrained on the sender side | |
2651 | * because the pipeline is full. | |
2652 | * | |
2653 | * BSD also seems to "accidentally" limit itself to windows that are a | |
2654 | * multiple of MSS, at least until the free space gets quite small. | |
2655 | * This would appear to be a side effect of the mbuf implementation. | |
2656 | * Combining these two algorithms results in the observed behavior | |
2657 | * of having a fixed window size at almost all times. | |
2658 | * | |
2659 | * Below we obtain similar behavior by forcing the offered window to | |
2660 | * a multiple of the mss when it is feasible to do so. | |
2661 | * | |
2662 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
2663 | * Regular options like TIMESTAMP are taken into account. | |
2664 | */ | |
2665 | u32 __tcp_select_window(struct sock *sk) | |
2666 | { | |
463c84b9 | 2667 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 2668 | struct tcp_sock *tp = tcp_sk(sk); |
caa20d9a | 2669 | /* MSS for the peer's data. Previous versions used mss_clamp |
1da177e4 LT |
2670 | * here. I don't know if the value based on our guesses |
2671 | * of peer's MSS is better for the performance. It's more correct | |
2672 | * but may be worse for the performance because of rcv_mss | |
2673 | * fluctuations. --SAW 1998/11/1 | |
2674 | */ | |
463c84b9 | 2675 | int mss = icsk->icsk_ack.rcv_mss; |
1da177e4 | 2676 | int free_space = tcp_space(sk); |
86c1a045 FW |
2677 | int allowed_space = tcp_full_space(sk); |
2678 | int full_space = min_t(int, tp->window_clamp, allowed_space); | |
1da177e4 LT |
2679 | int window; |
2680 | ||
06425c30 | 2681 | if (unlikely(mss > full_space)) { |
e905a9ed | 2682 | mss = full_space; |
06425c30 ED |
2683 | if (mss <= 0) |
2684 | return 0; | |
2685 | } | |
b92edbe0 | 2686 | if (free_space < (full_space >> 1)) { |
463c84b9 | 2687 | icsk->icsk_ack.quick = 0; |
1da177e4 | 2688 | |
b8da51eb | 2689 | if (tcp_under_memory_pressure(sk)) |
056834d9 IJ |
2690 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
2691 | 4U * tp->advmss); | |
1da177e4 | 2692 | |
86c1a045 FW |
2693 | /* free_space might become our new window, make sure we don't |
2694 | * increase it due to wscale. | |
2695 | */ | |
2696 | free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); | |
2697 | ||
2698 | /* if free space is less than mss estimate, or is below 1/16th | |
2699 | * of the maximum allowed, try to move to zero-window, else | |
2700 | * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and | |
2701 | * new incoming data is dropped due to memory limits. | |
2702 | * With large window, mss test triggers way too late in order | |
2703 | * to announce zero window in time before rmem limit kicks in. | |
2704 | */ | |
2705 | if (free_space < (allowed_space >> 4) || free_space < mss) | |
1da177e4 LT |
2706 | return 0; |
2707 | } | |
2708 | ||
2709 | if (free_space > tp->rcv_ssthresh) | |
2710 | free_space = tp->rcv_ssthresh; | |
2711 | ||
2712 | /* Don't do rounding if we are using window scaling, since the | |
2713 | * scaled window will not line up with the MSS boundary anyway. | |
2714 | */ | |
1da177e4 LT |
2715 | if (tp->rx_opt.rcv_wscale) { |
2716 | window = free_space; | |
2717 | ||
2718 | /* Advertise enough space so that it won't get scaled away. | |
2719 | * Import case: prevent zero window announcement if | |
2720 | * 1<<rcv_wscale > mss. | |
2721 | */ | |
1935299d | 2722 | window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); |
1da177e4 | 2723 | } else { |
1935299d | 2724 | window = tp->rcv_wnd; |
1da177e4 LT |
2725 | /* Get the largest window that is a nice multiple of mss. |
2726 | * Window clamp already applied above. | |
2727 | * If our current window offering is within 1 mss of the | |
2728 | * free space we just keep it. This prevents the divide | |
2729 | * and multiply from happening most of the time. | |
2730 | * We also don't do any window rounding when the free space | |
2731 | * is too small. | |
2732 | */ | |
2733 | if (window <= free_space - mss || window > free_space) | |
1935299d | 2734 | window = rounddown(free_space, mss); |
84565070 | 2735 | else if (mss == full_space && |
b92edbe0 | 2736 | free_space > window + (full_space >> 1)) |
84565070 | 2737 | window = free_space; |
1da177e4 LT |
2738 | } |
2739 | ||
2740 | return window; | |
2741 | } | |
2742 | ||
cfea5a68 MKL |
2743 | void tcp_skb_collapse_tstamp(struct sk_buff *skb, |
2744 | const struct sk_buff *next_skb) | |
082ac2d5 | 2745 | { |
0a2cf20c SHY |
2746 | if (unlikely(tcp_has_tx_tstamp(next_skb))) { |
2747 | const struct skb_shared_info *next_shinfo = | |
2748 | skb_shinfo(next_skb); | |
082ac2d5 MKL |
2749 | struct skb_shared_info *shinfo = skb_shinfo(skb); |
2750 | ||
0a2cf20c | 2751 | shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; |
082ac2d5 | 2752 | shinfo->tskey = next_shinfo->tskey; |
2de8023e MKL |
2753 | TCP_SKB_CB(skb)->txstamp_ack |= |
2754 | TCP_SKB_CB(next_skb)->txstamp_ack; | |
082ac2d5 MKL |
2755 | } |
2756 | } | |
2757 | ||
4a17fc3a | 2758 | /* Collapses two adjacent SKB's during retransmission. */ |
f8071cde | 2759 | static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) |
1da177e4 LT |
2760 | { |
2761 | struct tcp_sock *tp = tcp_sk(sk); | |
75c119af | 2762 | struct sk_buff *next_skb = skb_rb_next(skb); |
058dc334 | 2763 | int skb_size, next_skb_size; |
1da177e4 | 2764 | |
058dc334 IJ |
2765 | skb_size = skb->len; |
2766 | next_skb_size = next_skb->len; | |
1da177e4 | 2767 | |
058dc334 | 2768 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); |
a6963a6b | 2769 | |
f8071cde ED |
2770 | if (next_skb_size) { |
2771 | if (next_skb_size <= skb_availroom(skb)) | |
2772 | skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), | |
2773 | next_skb_size); | |
196379e6 | 2774 | else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size)) |
f8071cde ED |
2775 | return false; |
2776 | } | |
2b7cda9c | 2777 | tcp_highest_sack_replace(sk, next_skb, skb); |
1da177e4 | 2778 | |
058dc334 IJ |
2779 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
2780 | skb->ip_summed = CHECKSUM_PARTIAL; | |
1da177e4 | 2781 | |
058dc334 IJ |
2782 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
2783 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); | |
1da177e4 | 2784 | |
058dc334 IJ |
2785 | /* Update sequence range on original skb. */ |
2786 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1da177e4 | 2787 | |
e6c7d085 | 2788 | /* Merge over control information. This moves PSH/FIN etc. over */ |
4de075e0 | 2789 | TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; |
058dc334 IJ |
2790 | |
2791 | /* All done, get rid of second SKB and account for it so | |
2792 | * packet counting does not break. | |
2793 | */ | |
2794 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; | |
a643b5d4 | 2795 | TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; |
058dc334 IJ |
2796 | |
2797 | /* changed transmit queue under us so clear hints */ | |
ef9da47c IJ |
2798 | tcp_clear_retrans_hints_partial(tp); |
2799 | if (next_skb == tp->retransmit_skb_hint) | |
2800 | tp->retransmit_skb_hint = skb; | |
058dc334 | 2801 | |
797108d1 IJ |
2802 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
2803 | ||
082ac2d5 MKL |
2804 | tcp_skb_collapse_tstamp(skb, next_skb); |
2805 | ||
75c119af | 2806 | tcp_rtx_queue_unlink_and_free(next_skb, sk); |
f8071cde | 2807 | return true; |
1da177e4 LT |
2808 | } |
2809 | ||
67edfef7 | 2810 | /* Check if coalescing SKBs is legal. */ |
a2a385d6 | 2811 | static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) |
4a17fc3a IJ |
2812 | { |
2813 | if (tcp_skb_pcount(skb) > 1) | |
a2a385d6 | 2814 | return false; |
4a17fc3a | 2815 | if (skb_cloned(skb)) |
a2a385d6 | 2816 | return false; |
2331ccc5 | 2817 | /* Some heuristics for collapsing over SACK'd could be invented */ |
4a17fc3a | 2818 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
a2a385d6 | 2819 | return false; |
4a17fc3a | 2820 | |
a2a385d6 | 2821 | return true; |
4a17fc3a IJ |
2822 | } |
2823 | ||
67edfef7 AK |
2824 | /* Collapse packets in the retransmit queue to make to create |
2825 | * less packets on the wire. This is only done on retransmission. | |
2826 | */ | |
4a17fc3a IJ |
2827 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
2828 | int space) | |
2829 | { | |
2830 | struct tcp_sock *tp = tcp_sk(sk); | |
2831 | struct sk_buff *skb = to, *tmp; | |
a2a385d6 | 2832 | bool first = true; |
4a17fc3a | 2833 | |
e0a1e5b5 | 2834 | if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) |
4a17fc3a | 2835 | return; |
4de075e0 | 2836 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
4a17fc3a IJ |
2837 | return; |
2838 | ||
75c119af | 2839 | skb_rbtree_walk_from_safe(skb, tmp) { |
4a17fc3a IJ |
2840 | if (!tcp_can_collapse(sk, skb)) |
2841 | break; | |
2842 | ||
a643b5d4 MKL |
2843 | if (!tcp_skb_can_collapse_to(to)) |
2844 | break; | |
2845 | ||
4a17fc3a IJ |
2846 | space -= skb->len; |
2847 | ||
2848 | if (first) { | |
a2a385d6 | 2849 | first = false; |
4a17fc3a IJ |
2850 | continue; |
2851 | } | |
2852 | ||
2853 | if (space < 0) | |
2854 | break; | |
4a17fc3a IJ |
2855 | |
2856 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) | |
2857 | break; | |
2858 | ||
f8071cde ED |
2859 | if (!tcp_collapse_retrans(sk, to)) |
2860 | break; | |
4a17fc3a IJ |
2861 | } |
2862 | } | |
2863 | ||
1da177e4 LT |
2864 | /* This retransmits one SKB. Policy decisions and retransmit queue |
2865 | * state updates are done by the caller. Returns non-zero if an | |
2866 | * error occurred which prevented the send. | |
2867 | */ | |
10d3be56 | 2868 | int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) |
1da177e4 | 2869 | { |
5d424d5a | 2870 | struct inet_connection_sock *icsk = inet_csk(sk); |
10d3be56 | 2871 | struct tcp_sock *tp = tcp_sk(sk); |
7d227cd2 | 2872 | unsigned int cur_mss; |
10d3be56 ED |
2873 | int diff, len, err; |
2874 | ||
1da177e4 | 2875 | |
10d3be56 ED |
2876 | /* Inconclusive MTU probe */ |
2877 | if (icsk->icsk_mtup.probe_size) | |
5d424d5a | 2878 | icsk->icsk_mtup.probe_size = 0; |
5d424d5a | 2879 | |
1da177e4 | 2880 | /* Do not sent more than we queued. 1/4 is reserved for possible |
caa20d9a | 2881 | * copying overhead: fragmentation, tunneling, mangling etc. |
1da177e4 | 2882 | */ |
14afee4b | 2883 | if (refcount_read(&sk->sk_wmem_alloc) > |
ffb4d6c8 ED |
2884 | min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), |
2885 | sk->sk_sndbuf)) | |
1da177e4 LT |
2886 | return -EAGAIN; |
2887 | ||
1f3279ae ED |
2888 | if (skb_still_in_host_queue(sk, skb)) |
2889 | return -EBUSY; | |
2890 | ||
1da177e4 | 2891 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { |
7619ac5c ED |
2892 | if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { |
2893 | WARN_ON_ONCE(1); | |
2894 | return -EINVAL; | |
2895 | } | |
1da177e4 LT |
2896 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
2897 | return -ENOMEM; | |
2898 | } | |
2899 | ||
7d227cd2 SS |
2900 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
2901 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
2902 | ||
0c54b85f | 2903 | cur_mss = tcp_current_mss(sk); |
7d227cd2 | 2904 | |
1da177e4 LT |
2905 | /* If receiver has shrunk his window, and skb is out of |
2906 | * new window, do not retransmit it. The exception is the | |
2907 | * case, when window is shrunk to zero. In this case | |
2908 | * our retransmit serves as a zero window probe. | |
2909 | */ | |
9d4fb27d JP |
2910 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && |
2911 | TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1da177e4 LT |
2912 | return -EAGAIN; |
2913 | ||
10d3be56 ED |
2914 | len = cur_mss * segs; |
2915 | if (skb->len > len) { | |
75c119af ED |
2916 | if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, |
2917 | cur_mss, GFP_ATOMIC)) | |
1da177e4 | 2918 | return -ENOMEM; /* We'll try again later. */ |
02276f3c | 2919 | } else { |
10d3be56 ED |
2920 | if (skb_unclone(skb, GFP_ATOMIC)) |
2921 | return -ENOMEM; | |
9eb9362e | 2922 | |
10d3be56 ED |
2923 | diff = tcp_skb_pcount(skb); |
2924 | tcp_set_skb_tso_segs(skb, cur_mss); | |
2925 | diff -= tcp_skb_pcount(skb); | |
2926 | if (diff) | |
2927 | tcp_adjust_pcount(sk, skb, diff); | |
2928 | if (skb->len < cur_mss) | |
2929 | tcp_retrans_try_collapse(sk, skb, cur_mss); | |
1da177e4 LT |
2930 | } |
2931 | ||
49213555 DB |
2932 | /* RFC3168, section 6.1.1.1. ECN fallback */ |
2933 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) | |
2934 | tcp_ecn_clear_syn(sk, skb); | |
2935 | ||
678550c6 YC |
2936 | /* Update global and local TCP statistics. */ |
2937 | segs = tcp_skb_pcount(skb); | |
2938 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); | |
2939 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) | |
2940 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); | |
2941 | tp->total_retrans += segs; | |
2942 | ||
50bceae9 TG |
2943 | /* make sure skb->data is aligned on arches that require it |
2944 | * and check if ack-trimming & collapsing extended the headroom | |
2945 | * beyond what csum_start can cover. | |
2946 | */ | |
2947 | if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || | |
2948 | skb_headroom(skb) >= 0xFFFF)) { | |
10a81980 ED |
2949 | struct sk_buff *nskb; |
2950 | ||
e2080072 ED |
2951 | tcp_skb_tsorted_save(skb) { |
2952 | nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); | |
2953 | err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : | |
2954 | -ENOBUFS; | |
2955 | } tcp_skb_tsorted_restore(skb); | |
2956 | ||
5889e2c0 | 2957 | if (!err) { |
e2080072 | 2958 | tcp_update_skb_after_send(tp, skb); |
5889e2c0 YS |
2959 | tcp_rate_skb_sent(sk, skb); |
2960 | } | |
117632e6 | 2961 | } else { |
c84a5711 | 2962 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
117632e6 | 2963 | } |
c84a5711 | 2964 | |
fc9f3501 | 2965 | if (likely(!err)) { |
c84a5711 | 2966 | TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; |
e086101b | 2967 | trace_tcp_retransmit_skb(sk, skb); |
678550c6 | 2968 | } else if (err != -EBUSY) { |
cb7720b0 | 2969 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); |
fc9f3501 | 2970 | } |
c84a5711 | 2971 | return err; |
93b174ad YC |
2972 | } |
2973 | ||
10d3be56 | 2974 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) |
93b174ad YC |
2975 | { |
2976 | struct tcp_sock *tp = tcp_sk(sk); | |
10d3be56 | 2977 | int err = __tcp_retransmit_skb(sk, skb, segs); |
1da177e4 LT |
2978 | |
2979 | if (err == 0) { | |
1da177e4 | 2980 | #if FASTRETRANS_DEBUG > 0 |
056834d9 | 2981 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
e87cc472 | 2982 | net_dbg_ratelimited("retrans_out leaked\n"); |
1da177e4 LT |
2983 | } |
2984 | #endif | |
2985 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; | |
2986 | tp->retrans_out += tcp_skb_pcount(skb); | |
2987 | ||
2988 | /* Save stamp of the first retransmit. */ | |
2989 | if (!tp->retrans_stamp) | |
7faee5c0 | 2990 | tp->retrans_stamp = tcp_skb_timestamp(skb); |
1da177e4 | 2991 | |
1da177e4 | 2992 | } |
6e08d5e3 YC |
2993 | |
2994 | if (tp->undo_retrans < 0) | |
2995 | tp->undo_retrans = 0; | |
2996 | tp->undo_retrans += tcp_skb_pcount(skb); | |
1da177e4 LT |
2997 | return err; |
2998 | } | |
2999 | ||
3000 | /* This gets called after a retransmit timeout, and the initially | |
3001 | * retransmitted data is acknowledged. It tries to continue | |
3002 | * resending the rest of the retransmit queue, until either | |
3003 | * we've sent it all or the congestion window limit is reached. | |
1da177e4 LT |
3004 | */ |
3005 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
3006 | { | |
6687e988 | 3007 | const struct inet_connection_sock *icsk = inet_csk(sk); |
b9f1f1ce | 3008 | struct sk_buff *skb, *rtx_head, *hole = NULL; |
1da177e4 | 3009 | struct tcp_sock *tp = tcp_sk(sk); |
840a3cbe | 3010 | u32 max_segs; |
61eb55f4 | 3011 | int mib_idx; |
6a438bbe | 3012 | |
45e77d31 IJ |
3013 | if (!tp->packets_out) |
3014 | return; | |
3015 | ||
b9f1f1ce ED |
3016 | rtx_head = tcp_rtx_queue_head(sk); |
3017 | skb = tp->retransmit_skb_hint ?: rtx_head; | |
ed6e7268 | 3018 | max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); |
75c119af | 3019 | skb_rbtree_walk_from(skb) { |
dca0aaf8 | 3020 | __u8 sacked; |
10d3be56 | 3021 | int segs; |
1da177e4 | 3022 | |
218af599 ED |
3023 | if (tcp_pacing_check(sk)) |
3024 | break; | |
3025 | ||
08ebd172 | 3026 | /* we could do better than to assign each time */ |
51456b29 | 3027 | if (!hole) |
0e1c54c2 | 3028 | tp->retransmit_skb_hint = skb; |
08ebd172 | 3029 | |
10d3be56 ED |
3030 | segs = tp->snd_cwnd - tcp_packets_in_flight(tp); |
3031 | if (segs <= 0) | |
08ebd172 | 3032 | return; |
dca0aaf8 | 3033 | sacked = TCP_SKB_CB(skb)->sacked; |
a3d2e9f8 ED |
3034 | /* In case tcp_shift_skb_data() have aggregated large skbs, |
3035 | * we need to make sure not sending too bigs TSO packets | |
3036 | */ | |
3037 | segs = min_t(int, segs, max_segs); | |
1da177e4 | 3038 | |
840a3cbe YC |
3039 | if (tp->retrans_out >= tp->lost_out) { |
3040 | break; | |
0e1c54c2 | 3041 | } else if (!(sacked & TCPCB_LOST)) { |
51456b29 | 3042 | if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
0e1c54c2 IJ |
3043 | hole = skb; |
3044 | continue; | |
1da177e4 | 3045 | |
0e1c54c2 IJ |
3046 | } else { |
3047 | if (icsk->icsk_ca_state != TCP_CA_Loss) | |
3048 | mib_idx = LINUX_MIB_TCPFASTRETRANS; | |
3049 | else | |
3050 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; | |
3051 | } | |
1da177e4 | 3052 | |
0e1c54c2 | 3053 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
1da177e4 LT |
3054 | continue; |
3055 | ||
f9616c35 ED |
3056 | if (tcp_small_queue_check(sk, skb, 1)) |
3057 | return; | |
3058 | ||
10d3be56 | 3059 | if (tcp_retransmit_skb(sk, skb, segs)) |
0e1c54c2 | 3060 | return; |
24ab6bec | 3061 | |
de1d6578 | 3062 | NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); |
1da177e4 | 3063 | |
684bad11 | 3064 | if (tcp_in_cwnd_reduction(sk)) |
a262f0cd ND |
3065 | tp->prr_out += tcp_skb_pcount(skb); |
3066 | ||
75c119af | 3067 | if (skb == rtx_head && |
57dde7f7 | 3068 | icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) |
3f421baa ACM |
3069 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
3070 | inet_csk(sk)->icsk_rto, | |
3071 | TCP_RTO_MAX); | |
1da177e4 LT |
3072 | } |
3073 | } | |
3074 | ||
d83769a5 ED |
3075 | /* We allow to exceed memory limits for FIN packets to expedite |
3076 | * connection tear down and (memory) recovery. | |
845704a5 ED |
3077 | * Otherwise tcp_send_fin() could be tempted to either delay FIN |
3078 | * or even be forced to close flow without any FIN. | |
a6c5ea4c ED |
3079 | * In general, we want to allow one skb per socket to avoid hangs |
3080 | * with edge trigger epoll() | |
d83769a5 | 3081 | */ |
a6c5ea4c | 3082 | void sk_forced_mem_schedule(struct sock *sk, int size) |
d83769a5 | 3083 | { |
e805605c | 3084 | int amt; |
d83769a5 ED |
3085 | |
3086 | if (size <= sk->sk_forward_alloc) | |
3087 | return; | |
3088 | amt = sk_mem_pages(size); | |
3089 | sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; | |
e805605c JW |
3090 | sk_memory_allocated_add(sk, amt); |
3091 | ||
baac50bb JW |
3092 | if (mem_cgroup_sockets_enabled && sk->sk_memcg) |
3093 | mem_cgroup_charge_skmem(sk->sk_memcg, amt); | |
d83769a5 ED |
3094 | } |
3095 | ||
845704a5 ED |
3096 | /* Send a FIN. The caller locks the socket for us. |
3097 | * We should try to send a FIN packet really hard, but eventually give up. | |
1da177e4 LT |
3098 | */ |
3099 | void tcp_send_fin(struct sock *sk) | |
3100 | { | |
845704a5 | 3101 | struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); |
e905a9ed | 3102 | struct tcp_sock *tp = tcp_sk(sk); |
e905a9ed | 3103 | |
845704a5 ED |
3104 | /* Optimization, tack on the FIN if we have one skb in write queue and |
3105 | * this skb was not yet sent, or we are under memory pressure. | |
3106 | * Note: in the latter case, FIN packet will be sent after a timeout, | |
3107 | * as TCP stack thinks it has already been transmitted. | |
1da177e4 | 3108 | */ |
75c119af ED |
3109 | if (!tskb && tcp_under_memory_pressure(sk)) |
3110 | tskb = skb_rb_last(&sk->tcp_rtx_queue); | |
3111 | ||
3112 | if (tskb) { | |
845704a5 ED |
3113 | coalesce: |
3114 | TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; | |
3115 | TCP_SKB_CB(tskb)->end_seq++; | |
1da177e4 | 3116 | tp->write_seq++; |
75c119af | 3117 | if (tcp_write_queue_empty(sk)) { |
845704a5 ED |
3118 | /* This means tskb was already sent. |
3119 | * Pretend we included the FIN on previous transmit. | |
3120 | * We need to set tp->snd_nxt to the value it would have | |
3121 | * if FIN had been sent. This is because retransmit path | |
3122 | * does not change tp->snd_nxt. | |
3123 | */ | |
3124 | tp->snd_nxt++; | |
3125 | return; | |
3126 | } | |
1da177e4 | 3127 | } else { |
845704a5 ED |
3128 | skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); |
3129 | if (unlikely(!skb)) { | |
3130 | if (tskb) | |
3131 | goto coalesce; | |
3132 | return; | |
1da177e4 | 3133 | } |
e2080072 | 3134 | INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); |
d83769a5 | 3135 | skb_reserve(skb, MAX_TCP_HEADER); |
a6c5ea4c | 3136 | sk_forced_mem_schedule(sk, skb->truesize); |
1da177e4 | 3137 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
e870a8ef | 3138 | tcp_init_nondata_skb(skb, tp->write_seq, |
a3433f35 | 3139 | TCPHDR_ACK | TCPHDR_FIN); |
1da177e4 LT |
3140 | tcp_queue_skb(sk, skb); |
3141 | } | |
845704a5 | 3142 | __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); |
1da177e4 LT |
3143 | } |
3144 | ||
3145 | /* We get here when a process closes a file descriptor (either due to | |
3146 | * an explicit close() or as a byproduct of exit()'ing) and there | |
3147 | * was unread data in the receive queue. This behavior is recommended | |
65bb723c | 3148 | * by RFC 2525, section 2.17. -DaveM |
1da177e4 | 3149 | */ |
dd0fc66f | 3150 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
1da177e4 | 3151 | { |
1da177e4 LT |
3152 | struct sk_buff *skb; |
3153 | ||
7cc2b043 GF |
3154 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
3155 | ||
1da177e4 LT |
3156 | /* NOTE: No TCP options attached and we never retransmit this. */ |
3157 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
3158 | if (!skb) { | |
4e673444 | 3159 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
1da177e4 LT |
3160 | return; |
3161 | } | |
3162 | ||
3163 | /* Reserve space for headers and prepare control bits. */ | |
3164 | skb_reserve(skb, MAX_TCP_HEADER); | |
e870a8ef | 3165 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
a3433f35 | 3166 | TCPHDR_ACK | TCPHDR_RST); |
9a568de4 | 3167 | tcp_mstamp_refresh(tcp_sk(sk)); |
1da177e4 | 3168 | /* Send it off. */ |
dfb4b9dc | 3169 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
4e673444 | 3170 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
c24b14c4 SL |
3171 | |
3172 | /* skb of trace_tcp_send_reset() keeps the skb that caused RST, | |
3173 | * skb here is different to the troublesome skb, so use NULL | |
3174 | */ | |
3175 | trace_tcp_send_reset(sk, NULL); | |
1da177e4 LT |
3176 | } |
3177 | ||
67edfef7 AK |
3178 | /* Send a crossed SYN-ACK during socket establishment. |
3179 | * WARNING: This routine must only be called when we have already sent | |
1da177e4 LT |
3180 | * a SYN packet that crossed the incoming SYN that caused this routine |
3181 | * to get called. If this assumption fails then the initial rcv_wnd | |
3182 | * and rcv_wscale values will not be correct. | |
3183 | */ | |
3184 | int tcp_send_synack(struct sock *sk) | |
3185 | { | |
056834d9 | 3186 | struct sk_buff *skb; |
1da177e4 | 3187 | |
75c119af | 3188 | skb = tcp_rtx_queue_head(sk); |
51456b29 | 3189 | if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { |
75c119af | 3190 | pr_err("%s: wrong queue state\n", __func__); |
1da177e4 LT |
3191 | return -EFAULT; |
3192 | } | |
4de075e0 | 3193 | if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { |
1da177e4 | 3194 | if (skb_cloned(skb)) { |
e2080072 ED |
3195 | struct sk_buff *nskb; |
3196 | ||
3197 | tcp_skb_tsorted_save(skb) { | |
3198 | nskb = skb_copy(skb, GFP_ATOMIC); | |
3199 | } tcp_skb_tsorted_restore(skb); | |
51456b29 | 3200 | if (!nskb) |
1da177e4 | 3201 | return -ENOMEM; |
e2080072 | 3202 | INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); |
75c119af | 3203 | tcp_rtx_queue_unlink_and_free(skb, sk); |
f4a775d1 | 3204 | __skb_header_release(nskb); |
75c119af | 3205 | tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); |
3ab224be HA |
3206 | sk->sk_wmem_queued += nskb->truesize; |
3207 | sk_mem_charge(sk, nskb->truesize); | |
1da177e4 LT |
3208 | skb = nskb; |
3209 | } | |
3210 | ||
4de075e0 | 3211 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; |
735d3831 | 3212 | tcp_ecn_send_synack(sk, skb); |
1da177e4 | 3213 | } |
dfb4b9dc | 3214 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
3215 | } |
3216 | ||
4aea39c1 ED |
3217 | /** |
3218 | * tcp_make_synack - Prepare a SYN-ACK. | |
3219 | * sk: listener socket | |
3220 | * dst: dst entry attached to the SYNACK | |
3221 | * req: request_sock pointer | |
4aea39c1 ED |
3222 | * |
3223 | * Allocate one skb and build a SYNACK packet. | |
3224 | * @dst is consumed : Caller should not use it again. | |
3225 | */ | |
5d062de7 | 3226 | struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, |
e6b4d113 | 3227 | struct request_sock *req, |
ca6fb065 | 3228 | struct tcp_fastopen_cookie *foc, |
b3d05147 | 3229 | enum tcp_synack_type synack_type) |
1da177e4 | 3230 | { |
2e6599cb | 3231 | struct inet_request_sock *ireq = inet_rsk(req); |
5d062de7 | 3232 | const struct tcp_sock *tp = tcp_sk(sk); |
80f03e27 | 3233 | struct tcp_md5sig_key *md5 = NULL; |
5d062de7 ED |
3234 | struct tcp_out_options opts; |
3235 | struct sk_buff *skb; | |
bd0388ae | 3236 | int tcp_header_size; |
5d062de7 | 3237 | struct tcphdr *th; |
f5fff5dc | 3238 | int mss; |
1da177e4 | 3239 | |
ca6fb065 | 3240 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); |
4aea39c1 ED |
3241 | if (unlikely(!skb)) { |
3242 | dst_release(dst); | |
1da177e4 | 3243 | return NULL; |
4aea39c1 | 3244 | } |
1da177e4 LT |
3245 | /* Reserve space for headers. */ |
3246 | skb_reserve(skb, MAX_TCP_HEADER); | |
3247 | ||
b3d05147 ED |
3248 | switch (synack_type) { |
3249 | case TCP_SYNACK_NORMAL: | |
9e17f8a4 | 3250 | skb_set_owner_w(skb, req_to_sk(req)); |
b3d05147 ED |
3251 | break; |
3252 | case TCP_SYNACK_COOKIE: | |
3253 | /* Under synflood, we do not attach skb to a socket, | |
3254 | * to avoid false sharing. | |
3255 | */ | |
3256 | break; | |
3257 | case TCP_SYNACK_FASTOPEN: | |
ca6fb065 ED |
3258 | /* sk is a const pointer, because we want to express multiple |
3259 | * cpu might call us concurrently. | |
3260 | * sk->sk_wmem_alloc in an atomic, we can promote to rw. | |
3261 | */ | |
3262 | skb_set_owner_w(skb, (struct sock *)sk); | |
b3d05147 | 3263 | break; |
ca6fb065 | 3264 | } |
4aea39c1 | 3265 | skb_dst_set(skb, dst); |
1da177e4 | 3266 | |
3541f9e8 | 3267 | mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); |
f5fff5dc | 3268 | |
33ad798c | 3269 | memset(&opts, 0, sizeof(opts)); |
8b5f12d0 FW |
3270 | #ifdef CONFIG_SYN_COOKIES |
3271 | if (unlikely(req->cookie_ts)) | |
9a568de4 | 3272 | skb->skb_mstamp = cookie_init_timestamp(req); |
8b5f12d0 FW |
3273 | else |
3274 | #endif | |
9a568de4 | 3275 | skb->skb_mstamp = tcp_clock_us(); |
80f03e27 ED |
3276 | |
3277 | #ifdef CONFIG_TCP_MD5SIG | |
3278 | rcu_read_lock(); | |
fd3a154a | 3279 | md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); |
80f03e27 | 3280 | #endif |
58d607d3 | 3281 | skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); |
60e2a778 UB |
3282 | tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, |
3283 | foc) + sizeof(*th); | |
cfb6eeb4 | 3284 | |
aa8223c7 ACM |
3285 | skb_push(skb, tcp_header_size); |
3286 | skb_reset_transport_header(skb); | |
1da177e4 | 3287 | |
ea1627c2 | 3288 | th = (struct tcphdr *)skb->data; |
1da177e4 LT |
3289 | memset(th, 0, sizeof(struct tcphdr)); |
3290 | th->syn = 1; | |
3291 | th->ack = 1; | |
6ac705b1 | 3292 | tcp_ecn_make_synack(req, th); |
b44084c2 | 3293 | th->source = htons(ireq->ir_num); |
634fb979 | 3294 | th->dest = ireq->ir_rmt_port; |
e05a90ec | 3295 | skb->mark = ireq->ir_mark; |
3b117750 ED |
3296 | skb->ip_summed = CHECKSUM_PARTIAL; |
3297 | th->seq = htonl(tcp_rsk(req)->snt_isn); | |
8336886f JC |
3298 | /* XXX data is queued and acked as is. No buffer/window check */ |
3299 | th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); | |
1da177e4 LT |
3300 | |
3301 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
ed53d0ab | 3302 | th->window = htons(min(req->rsk_rcv_wnd, 65535U)); |
5d062de7 | 3303 | tcp_options_write((__be32 *)(th + 1), NULL, &opts); |
1da177e4 | 3304 | th->doff = (tcp_header_size >> 2); |
90bbcc60 | 3305 | __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); |
cfb6eeb4 YH |
3306 | |
3307 | #ifdef CONFIG_TCP_MD5SIG | |
3308 | /* Okay, we have all we need - do the md5 hash if needed */ | |
80f03e27 | 3309 | if (md5) |
bd0388ae | 3310 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
39f8e58e | 3311 | md5, req_to_sk(req), skb); |
80f03e27 | 3312 | rcu_read_unlock(); |
cfb6eeb4 YH |
3313 | #endif |
3314 | ||
b50edd78 | 3315 | /* Do not fool tcpdump (if any), clean our debris */ |
2456e855 | 3316 | skb->tstamp = 0; |
1da177e4 LT |
3317 | return skb; |
3318 | } | |
4bc2f18b | 3319 | EXPORT_SYMBOL(tcp_make_synack); |
1da177e4 | 3320 | |
81164413 DB |
3321 | static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) |
3322 | { | |
3323 | struct inet_connection_sock *icsk = inet_csk(sk); | |
3324 | const struct tcp_congestion_ops *ca; | |
3325 | u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); | |
3326 | ||
3327 | if (ca_key == TCP_CA_UNSPEC) | |
3328 | return; | |
3329 | ||
3330 | rcu_read_lock(); | |
3331 | ca = tcp_ca_find_key(ca_key); | |
3332 | if (likely(ca && try_module_get(ca->owner))) { | |
3333 | module_put(icsk->icsk_ca_ops->owner); | |
3334 | icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); | |
3335 | icsk->icsk_ca_ops = ca; | |
3336 | } | |
3337 | rcu_read_unlock(); | |
3338 | } | |
3339 | ||
67edfef7 | 3340 | /* Do all connect socket setups that can be done AF independent. */ |
f7e56a76 | 3341 | static void tcp_connect_init(struct sock *sk) |
1da177e4 | 3342 | { |
cf533ea5 | 3343 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
3344 | struct tcp_sock *tp = tcp_sk(sk); |
3345 | __u8 rcv_wscale; | |
13d3b1eb | 3346 | u32 rcv_wnd; |
1da177e4 LT |
3347 | |
3348 | /* We'll fix this up when we get a response from the other end. | |
3349 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
3350 | */ | |
5d2ed052 ED |
3351 | tp->tcp_header_len = sizeof(struct tcphdr); |
3352 | if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) | |
3353 | tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; | |
1da177e4 | 3354 | |
cfb6eeb4 | 3355 | #ifdef CONFIG_TCP_MD5SIG |
00db4124 | 3356 | if (tp->af_specific->md5_lookup(sk, sk)) |
cfb6eeb4 YH |
3357 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; |
3358 | #endif | |
3359 | ||
1da177e4 LT |
3360 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
3361 | if (tp->rx_opt.user_mss) | |
3362 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
3363 | tp->max_window = 0; | |
5d424d5a | 3364 | tcp_mtup_init(sk); |
1da177e4 LT |
3365 | tcp_sync_mss(sk, dst_mtu(dst)); |
3366 | ||
81164413 DB |
3367 | tcp_ca_dst_init(sk, dst); |
3368 | ||
1da177e4 LT |
3369 | if (!tp->window_clamp) |
3370 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
3541f9e8 | 3371 | tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); |
f5fff5dc | 3372 | |
1da177e4 | 3373 | tcp_initialize_rcv_mss(sk); |
1da177e4 | 3374 | |
e88c64f0 HPP |
3375 | /* limit the window selection if the user enforce a smaller rx buffer */ |
3376 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
3377 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) | |
3378 | tp->window_clamp = tcp_full_space(sk); | |
3379 | ||
13d3b1eb LB |
3380 | rcv_wnd = tcp_rwnd_init_bpf(sk); |
3381 | if (rcv_wnd == 0) | |
3382 | rcv_wnd = dst_metric(dst, RTAX_INITRWND); | |
3383 | ||
ceef9ab6 | 3384 | tcp_select_initial_window(sk, tcp_full_space(sk), |
1da177e4 LT |
3385 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), |
3386 | &tp->rcv_wnd, | |
3387 | &tp->window_clamp, | |
9bb37ef0 | 3388 | sock_net(sk)->ipv4.sysctl_tcp_window_scaling, |
31d12926 | 3389 | &rcv_wscale, |
13d3b1eb | 3390 | rcv_wnd); |
1da177e4 LT |
3391 | |
3392 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
3393 | tp->rcv_ssthresh = tp->rcv_wnd; | |
3394 | ||
3395 | sk->sk_err = 0; | |
3396 | sock_reset_flag(sk, SOCK_DONE); | |
3397 | tp->snd_wnd = 0; | |
ee7537b6 | 3398 | tcp_init_wl(tp, 0); |
7619ac5c | 3399 | tcp_write_queue_purge(sk); |
1da177e4 LT |
3400 | tp->snd_una = tp->write_seq; |
3401 | tp->snd_sml = tp->write_seq; | |
33f5f57e | 3402 | tp->snd_up = tp->write_seq; |
370816ae | 3403 | tp->snd_nxt = tp->write_seq; |
ee995283 PE |
3404 | |
3405 | if (likely(!tp->repair)) | |
3406 | tp->rcv_nxt = 0; | |
c7781a6e | 3407 | else |
70eabf0e | 3408 | tp->rcv_tstamp = tcp_jiffies32; |
ee995283 PE |
3409 | tp->rcv_wup = tp->rcv_nxt; |
3410 | tp->copied_seq = tp->rcv_nxt; | |
1da177e4 | 3411 | |
8550f328 | 3412 | inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); |
463c84b9 | 3413 | inet_csk(sk)->icsk_retransmits = 0; |
1da177e4 LT |
3414 | tcp_clear_retrans(tp); |
3415 | } | |
3416 | ||
783237e8 YC |
3417 | static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) |
3418 | { | |
3419 | struct tcp_sock *tp = tcp_sk(sk); | |
3420 | struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); | |
3421 | ||
3422 | tcb->end_seq += skb->len; | |
f4a775d1 | 3423 | __skb_header_release(skb); |
783237e8 YC |
3424 | sk->sk_wmem_queued += skb->truesize; |
3425 | sk_mem_charge(sk, skb->truesize); | |
3426 | tp->write_seq = tcb->end_seq; | |
3427 | tp->packets_out += tcp_skb_pcount(skb); | |
3428 | } | |
3429 | ||
3430 | /* Build and send a SYN with data and (cached) Fast Open cookie. However, | |
3431 | * queue a data-only packet after the regular SYN, such that regular SYNs | |
3432 | * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges | |
3433 | * only the SYN sequence, the data are retransmitted in the first ACK. | |
3434 | * If cookie is not cached or other error occurs, falls back to send a | |
3435 | * regular SYN with Fast Open cookie request option. | |
3436 | */ | |
3437 | static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) | |
3438 | { | |
3439 | struct tcp_sock *tp = tcp_sk(sk); | |
3440 | struct tcp_fastopen_request *fo = tp->fastopen_req; | |
065263f4 | 3441 | int space, err = 0; |
355a901e | 3442 | struct sk_buff *syn_data; |
aab48743 | 3443 | |
67da22d2 | 3444 | tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ |
065263f4 | 3445 | if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) |
783237e8 YC |
3446 | goto fallback; |
3447 | ||
3448 | /* MSS for SYN-data is based on cached MSS and bounded by PMTU and | |
3449 | * user-MSS. Reserve maximum option space for middleboxes that add | |
3450 | * private TCP options. The cost is reduced data space in SYN :( | |
3451 | */ | |
3541f9e8 ED |
3452 | tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); |
3453 | ||
1b63edd6 | 3454 | space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - |
783237e8 YC |
3455 | MAX_TCP_OPTION_SPACE; |
3456 | ||
f5ddcbbb ED |
3457 | space = min_t(size_t, space, fo->size); |
3458 | ||
3459 | /* limit to order-0 allocations */ | |
3460 | space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); | |
3461 | ||
eb934478 | 3462 | syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); |
355a901e | 3463 | if (!syn_data) |
783237e8 | 3464 | goto fallback; |
355a901e ED |
3465 | syn_data->ip_summed = CHECKSUM_PARTIAL; |
3466 | memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); | |
07e100f9 ED |
3467 | if (space) { |
3468 | int copied = copy_from_iter(skb_put(syn_data, space), space, | |
3469 | &fo->data->msg_iter); | |
3470 | if (unlikely(!copied)) { | |
ba233b34 | 3471 | tcp_skb_tsorted_anchor_cleanup(syn_data); |
07e100f9 ED |
3472 | kfree_skb(syn_data); |
3473 | goto fallback; | |
3474 | } | |
3475 | if (copied != space) { | |
3476 | skb_trim(syn_data, copied); | |
3477 | space = copied; | |
3478 | } | |
57be5bda | 3479 | } |
355a901e ED |
3480 | /* No more data pending in inet_wait_for_connect() */ |
3481 | if (space == fo->size) | |
3482 | fo->data = NULL; | |
3483 | fo->copied = space; | |
783237e8 | 3484 | |
355a901e | 3485 | tcp_connect_queue_skb(sk, syn_data); |
0f87230d FY |
3486 | if (syn_data->len) |
3487 | tcp_chrono_start(sk, TCP_CHRONO_BUSY); | |
783237e8 | 3488 | |
355a901e | 3489 | err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); |
783237e8 | 3490 | |
355a901e | 3491 | syn->skb_mstamp = syn_data->skb_mstamp; |
431a9124 | 3492 | |
355a901e ED |
3493 | /* Now full SYN+DATA was cloned and sent (or not), |
3494 | * remove the SYN from the original skb (syn_data) | |
3495 | * we keep in write queue in case of a retransmit, as we | |
3496 | * also have the SYN packet (with no data) in the same queue. | |
3497 | */ | |
3498 | TCP_SKB_CB(syn_data)->seq++; | |
3499 | TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; | |
3500 | if (!err) { | |
67da22d2 | 3501 | tp->syn_data = (fo->copied > 0); |
75c119af | 3502 | tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); |
f19c29e3 | 3503 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); |
783237e8 YC |
3504 | goto done; |
3505 | } | |
783237e8 | 3506 | |
75c119af ED |
3507 | /* data was not sent, put it in write_queue */ |
3508 | __skb_queue_tail(&sk->sk_write_queue, syn_data); | |
b5b7db8d ED |
3509 | tp->packets_out -= tcp_skb_pcount(syn_data); |
3510 | ||
783237e8 YC |
3511 | fallback: |
3512 | /* Send a regular SYN with Fast Open cookie request option */ | |
3513 | if (fo->cookie.len > 0) | |
3514 | fo->cookie.len = 0; | |
3515 | err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); | |
3516 | if (err) | |
3517 | tp->syn_fastopen = 0; | |
783237e8 YC |
3518 | done: |
3519 | fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ | |
3520 | return err; | |
3521 | } | |
3522 | ||
67edfef7 | 3523 | /* Build a SYN and send it off. */ |
1da177e4 LT |
3524 | int tcp_connect(struct sock *sk) |
3525 | { | |
3526 | struct tcp_sock *tp = tcp_sk(sk); | |
3527 | struct sk_buff *buff; | |
ee586811 | 3528 | int err; |
1da177e4 | 3529 | |
9872a4bd | 3530 | tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB); |
8ba60924 ED |
3531 | |
3532 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) | |
3533 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
3534 | ||
1da177e4 LT |
3535 | tcp_connect_init(sk); |
3536 | ||
2b916477 AV |
3537 | if (unlikely(tp->repair)) { |
3538 | tcp_finish_connect(sk, NULL); | |
3539 | return 0; | |
3540 | } | |
3541 | ||
eb934478 | 3542 | buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); |
355a901e | 3543 | if (unlikely(!buff)) |
1da177e4 LT |
3544 | return -ENOBUFS; |
3545 | ||
a3433f35 | 3546 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
9a568de4 ED |
3547 | tcp_mstamp_refresh(tp); |
3548 | tp->retrans_stamp = tcp_time_stamp(tp); | |
783237e8 | 3549 | tcp_connect_queue_skb(sk, buff); |
735d3831 | 3550 | tcp_ecn_send_syn(sk, buff); |
75c119af | 3551 | tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); |
1da177e4 | 3552 | |
783237e8 YC |
3553 | /* Send off SYN; include data in Fast Open. */ |
3554 | err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : | |
3555 | tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); | |
ee586811 EP |
3556 | if (err == -ECONNREFUSED) |
3557 | return err; | |
bd37a088 WY |
3558 | |
3559 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
3560 | * in order to make this packet get counted in tcpOutSegs. | |
3561 | */ | |
3562 | tp->snd_nxt = tp->write_seq; | |
3563 | tp->pushed_seq = tp->write_seq; | |
b5b7db8d ED |
3564 | buff = tcp_send_head(sk); |
3565 | if (unlikely(buff)) { | |
3566 | tp->snd_nxt = TCP_SKB_CB(buff)->seq; | |
3567 | tp->pushed_seq = TCP_SKB_CB(buff)->seq; | |
3568 | } | |
81cc8a75 | 3569 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
1da177e4 LT |
3570 | |
3571 | /* Timer for repeating the SYN until an answer. */ | |
3f421baa ACM |
3572 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
3573 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
3574 | return 0; |
3575 | } | |
4bc2f18b | 3576 | EXPORT_SYMBOL(tcp_connect); |
1da177e4 LT |
3577 | |
3578 | /* Send out a delayed ack, the caller does the policy checking | |
3579 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
3580 | * for details. | |
3581 | */ | |
3582 | void tcp_send_delayed_ack(struct sock *sk) | |
3583 | { | |
463c84b9 ACM |
3584 | struct inet_connection_sock *icsk = inet_csk(sk); |
3585 | int ato = icsk->icsk_ack.ato; | |
1da177e4 LT |
3586 | unsigned long timeout; |
3587 | ||
3588 | if (ato > TCP_DELACK_MIN) { | |
463c84b9 | 3589 | const struct tcp_sock *tp = tcp_sk(sk); |
056834d9 | 3590 | int max_ato = HZ / 2; |
1da177e4 | 3591 | |
056834d9 IJ |
3592 | if (icsk->icsk_ack.pingpong || |
3593 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) | |
1da177e4 LT |
3594 | max_ato = TCP_DELACK_MAX; |
3595 | ||
3596 | /* Slow path, intersegment interval is "high". */ | |
3597 | ||
3598 | /* If some rtt estimate is known, use it to bound delayed ack. | |
463c84b9 | 3599 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
1da177e4 LT |
3600 | * directly. |
3601 | */ | |
740b0f18 ED |
3602 | if (tp->srtt_us) { |
3603 | int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), | |
3604 | TCP_DELACK_MIN); | |
1da177e4 LT |
3605 | |
3606 | if (rtt < max_ato) | |
3607 | max_ato = rtt; | |
3608 | } | |
3609 | ||
3610 | ato = min(ato, max_ato); | |
3611 | } | |
3612 | ||
3613 | /* Stay within the limit we were given */ | |
3614 | timeout = jiffies + ato; | |
3615 | ||
3616 | /* Use new timeout only if there wasn't a older one earlier. */ | |
463c84b9 | 3617 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
1da177e4 LT |
3618 | /* If delack timer was blocked or is about to expire, |
3619 | * send ACK now. | |
3620 | */ | |
463c84b9 ACM |
3621 | if (icsk->icsk_ack.blocked || |
3622 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
1da177e4 LT |
3623 | tcp_send_ack(sk); |
3624 | return; | |
3625 | } | |
3626 | ||
463c84b9 ACM |
3627 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
3628 | timeout = icsk->icsk_ack.timeout; | |
1da177e4 | 3629 | } |
463c84b9 ACM |
3630 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
3631 | icsk->icsk_ack.timeout = timeout; | |
3632 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
1da177e4 LT |
3633 | } |
3634 | ||
3635 | /* This routine sends an ack and also updates the window. */ | |
5cf68c20 | 3636 | void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) |
1da177e4 | 3637 | { |
058dc334 | 3638 | struct sk_buff *buff; |
1da177e4 | 3639 | |
058dc334 IJ |
3640 | /* If we have been reset, we may not send again. */ |
3641 | if (sk->sk_state == TCP_CLOSE) | |
3642 | return; | |
1da177e4 | 3643 | |
058dc334 IJ |
3644 | /* We are not putting this on the write queue, so |
3645 | * tcp_transmit_skb() will set the ownership to this | |
3646 | * sock. | |
3647 | */ | |
7450aaf6 ED |
3648 | buff = alloc_skb(MAX_TCP_HEADER, |
3649 | sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); | |
3650 | if (unlikely(!buff)) { | |
058dc334 IJ |
3651 | inet_csk_schedule_ack(sk); |
3652 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; | |
3653 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, | |
3654 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
3655 | return; | |
1da177e4 | 3656 | } |
058dc334 IJ |
3657 | |
3658 | /* Reserve space for headers and prepare control bits. */ | |
3659 | skb_reserve(buff, MAX_TCP_HEADER); | |
a3433f35 | 3660 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
058dc334 | 3661 | |
98781965 ED |
3662 | /* We do not want pure acks influencing TCP Small Queues or fq/pacing |
3663 | * too much. | |
3664 | * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 | |
98781965 ED |
3665 | */ |
3666 | skb_set_tcp_pure_ack(buff); | |
3667 | ||
058dc334 | 3668 | /* Send it off, this clears delayed acks for us. */ |
5cf68c20 YC |
3669 | __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); |
3670 | } | |
deb42350 | 3671 | EXPORT_SYMBOL_GPL(__tcp_send_ack); |
5cf68c20 YC |
3672 | |
3673 | void tcp_send_ack(struct sock *sk) | |
3674 | { | |
3675 | __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); | |
1da177e4 LT |
3676 | } |
3677 | ||
3678 | /* This routine sends a packet with an out of date sequence | |
3679 | * number. It assumes the other end will try to ack it. | |
3680 | * | |
3681 | * Question: what should we make while urgent mode? | |
3682 | * 4.4BSD forces sending single byte of data. We cannot send | |
3683 | * out of window data, because we have SND.NXT==SND.MAX... | |
3684 | * | |
3685 | * Current solution: to send TWO zero-length segments in urgent mode: | |
3686 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
3687 | * out-of-date with SND.UNA-1 to probe window. | |
3688 | */ | |
e520af48 | 3689 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) |
1da177e4 LT |
3690 | { |
3691 | struct tcp_sock *tp = tcp_sk(sk); | |
3692 | struct sk_buff *skb; | |
3693 | ||
3694 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
7450aaf6 ED |
3695 | skb = alloc_skb(MAX_TCP_HEADER, |
3696 | sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); | |
51456b29 | 3697 | if (!skb) |
1da177e4 LT |
3698 | return -1; |
3699 | ||
3700 | /* Reserve space for headers and set control bits. */ | |
3701 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 LT |
3702 | /* Use a previous sequence. This should cause the other |
3703 | * end to send an ack. Don't queue or clone SKB, just | |
3704 | * send it. | |
3705 | */ | |
a3433f35 | 3706 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
e2e8009f | 3707 | NET_INC_STATS(sock_net(sk), mib); |
7450aaf6 | 3708 | return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); |
1da177e4 LT |
3709 | } |
3710 | ||
385e2070 | 3711 | /* Called from setsockopt( ... TCP_REPAIR ) */ |
ee995283 PE |
3712 | void tcp_send_window_probe(struct sock *sk) |
3713 | { | |
3714 | if (sk->sk_state == TCP_ESTABLISHED) { | |
3715 | tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; | |
9a568de4 | 3716 | tcp_mstamp_refresh(tcp_sk(sk)); |
e520af48 | 3717 | tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); |
ee995283 PE |
3718 | } |
3719 | } | |
3720 | ||
67edfef7 | 3721 | /* Initiate keepalive or window probe from timer. */ |
e520af48 | 3722 | int tcp_write_wakeup(struct sock *sk, int mib) |
1da177e4 | 3723 | { |
058dc334 IJ |
3724 | struct tcp_sock *tp = tcp_sk(sk); |
3725 | struct sk_buff *skb; | |
1da177e4 | 3726 | |
058dc334 IJ |
3727 | if (sk->sk_state == TCP_CLOSE) |
3728 | return -1; | |
3729 | ||
00db4124 IM |
3730 | skb = tcp_send_head(sk); |
3731 | if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { | |
058dc334 | 3732 | int err; |
0c54b85f | 3733 | unsigned int mss = tcp_current_mss(sk); |
058dc334 IJ |
3734 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
3735 | ||
3736 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
3737 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
3738 | ||
3739 | /* We are probing the opening of a window | |
3740 | * but the window size is != 0 | |
3741 | * must have been a result SWS avoidance ( sender ) | |
3742 | */ | |
3743 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
3744 | skb->len > mss) { | |
3745 | seg_size = min(seg_size, mss); | |
4de075e0 | 3746 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
75c119af ED |
3747 | if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, |
3748 | skb, seg_size, mss, GFP_ATOMIC)) | |
058dc334 IJ |
3749 | return -1; |
3750 | } else if (!tcp_skb_pcount(skb)) | |
5bbb432c | 3751 | tcp_set_skb_tso_segs(skb, mss); |
058dc334 | 3752 | |
4de075e0 | 3753 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
058dc334 IJ |
3754 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
3755 | if (!err) | |
3756 | tcp_event_new_data_sent(sk, skb); | |
3757 | return err; | |
3758 | } else { | |
33f5f57e | 3759 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
e520af48 ED |
3760 | tcp_xmit_probe_skb(sk, 1, mib); |
3761 | return tcp_xmit_probe_skb(sk, 0, mib); | |
1da177e4 | 3762 | } |
1da177e4 LT |
3763 | } |
3764 | ||
3765 | /* A window probe timeout has occurred. If window is not closed send | |
3766 | * a partial packet else a zero probe. | |
3767 | */ | |
3768 | void tcp_send_probe0(struct sock *sk) | |
3769 | { | |
463c84b9 | 3770 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 3771 | struct tcp_sock *tp = tcp_sk(sk); |
c6214a97 | 3772 | struct net *net = sock_net(sk); |
fcdd1cf4 | 3773 | unsigned long probe_max; |
1da177e4 LT |
3774 | int err; |
3775 | ||
e520af48 | 3776 | err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); |
1da177e4 | 3777 | |
75c119af | 3778 | if (tp->packets_out || tcp_write_queue_empty(sk)) { |
1da177e4 | 3779 | /* Cancel probe timer, if it is not required. */ |
6687e988 | 3780 | icsk->icsk_probes_out = 0; |
463c84b9 | 3781 | icsk->icsk_backoff = 0; |
1da177e4 LT |
3782 | return; |
3783 | } | |
3784 | ||
3785 | if (err <= 0) { | |
c6214a97 | 3786 | if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) |
463c84b9 | 3787 | icsk->icsk_backoff++; |
6687e988 | 3788 | icsk->icsk_probes_out++; |
fcdd1cf4 | 3789 | probe_max = TCP_RTO_MAX; |
1da177e4 LT |
3790 | } else { |
3791 | /* If packet was not sent due to local congestion, | |
6687e988 | 3792 | * do not backoff and do not remember icsk_probes_out. |
1da177e4 LT |
3793 | * Let local senders to fight for local resources. |
3794 | * | |
3795 | * Use accumulated backoff yet. | |
3796 | */ | |
6687e988 ACM |
3797 | if (!icsk->icsk_probes_out) |
3798 | icsk->icsk_probes_out = 1; | |
fcdd1cf4 | 3799 | probe_max = TCP_RESOURCE_PROBE_INTERVAL; |
1da177e4 | 3800 | } |
fcdd1cf4 | 3801 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
21c8fe99 | 3802 | tcp_probe0_when(sk, probe_max), |
fcdd1cf4 | 3803 | TCP_RTO_MAX); |
1da177e4 | 3804 | } |
5db92c99 | 3805 | |
ea3bea3a | 3806 | int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) |
5db92c99 OP |
3807 | { |
3808 | const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; | |
3809 | struct flowi fl; | |
3810 | int res; | |
3811 | ||
58d607d3 | 3812 | tcp_rsk(req)->txhash = net_tx_rndhash(); |
b3d05147 | 3813 | res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); |
5db92c99 | 3814 | if (!res) { |
90bbcc60 | 3815 | __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
02a1d6e7 | 3816 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); |
7e32b443 YC |
3817 | if (unlikely(tcp_passive_fastopen(sk))) |
3818 | tcp_sk(sk)->total_retrans++; | |
cf34ce3d | 3819 | trace_tcp_retransmit_synack(sk, req); |
5db92c99 OP |
3820 | } |
3821 | return res; | |
3822 | } | |
3823 | EXPORT_SYMBOL(tcp_rtx_synack); |