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.
6 * Implementation of the Transmission Control Protocol(TCP).
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>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
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
39 #include <linux/compiler.h>
40 #include <linux/module.h>
42 /* People can turn this off for buggy TCP's found in printers etc. */
43 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
45 /* People can turn this on to work with those rare, broken TCPs that
46 * interpret the window field as a signed quantity.
48 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
50 /* This limits the percentage of the congestion window which we
51 * will allow a single TSO frame to consume. Building TSO frames
52 * which are too large can cause TCP streams to be bursty.
54 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
56 int sysctl_tcp_mtu_probing __read_mostly
= 0;
57 int sysctl_tcp_base_mss __read_mostly
= 512;
59 /* By default, RFC2861 behavior. */
60 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
62 static void tcp_event_new_data_sent(struct sock
*sk
, struct sk_buff
*skb
)
64 struct tcp_sock
*tp
= tcp_sk(sk
);
65 unsigned int prior_packets
= tp
->packets_out
;
67 tcp_advance_send_head(sk
, skb
);
68 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
70 /* Don't override Nagle indefinately with F-RTO */
71 if (tp
->frto_counter
== 2)
74 tp
->packets_out
+= tcp_skb_pcount(skb
);
76 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
77 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
80 /* SND.NXT, if window was not shrunk.
81 * If window has been shrunk, what should we make? It is not clear at all.
82 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
83 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
84 * invalid. OK, let's make this for now:
86 static inline __u32
tcp_acceptable_seq(struct sock
*sk
)
88 struct tcp_sock
*tp
= tcp_sk(sk
);
90 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
93 return tcp_wnd_end(tp
);
96 /* Calculate mss to advertise in SYN segment.
97 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
99 * 1. It is independent of path mtu.
100 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
101 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
102 * attached devices, because some buggy hosts are confused by
104 * 4. We do not make 3, we advertise MSS, calculated from first
105 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
106 * This may be overridden via information stored in routing table.
107 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
108 * probably even Jumbo".
110 static __u16
tcp_advertise_mss(struct sock
*sk
)
112 struct tcp_sock
*tp
= tcp_sk(sk
);
113 struct dst_entry
*dst
= __sk_dst_get(sk
);
114 int mss
= tp
->advmss
;
116 if (dst
&& dst_metric(dst
, RTAX_ADVMSS
) < mss
) {
117 mss
= dst_metric(dst
, RTAX_ADVMSS
);
124 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
125 * This is the first part of cwnd validation mechanism. */
126 static void tcp_cwnd_restart(struct sock
*sk
, struct dst_entry
*dst
)
128 struct tcp_sock
*tp
= tcp_sk(sk
);
129 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
130 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
131 u32 cwnd
= tp
->snd_cwnd
;
133 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
135 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
136 restart_cwnd
= min(restart_cwnd
, cwnd
);
138 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
140 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
141 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
142 tp
->snd_cwnd_used
= 0;
145 static void tcp_event_data_sent(struct tcp_sock
*tp
,
146 struct sk_buff
*skb
, struct sock
*sk
)
148 struct inet_connection_sock
*icsk
= inet_csk(sk
);
149 const u32 now
= tcp_time_stamp
;
151 if (sysctl_tcp_slow_start_after_idle
&&
152 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
153 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
157 /* If it is a reply for ato after last received
158 * packet, enter pingpong mode.
160 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
161 icsk
->icsk_ack
.pingpong
= 1;
164 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
166 tcp_dec_quickack_mode(sk
, pkts
);
167 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
170 /* Determine a window scaling and initial window to offer.
171 * Based on the assumption that the given amount of space
172 * will be offered. Store the results in the tp structure.
173 * NOTE: for smooth operation initial space offering should
174 * be a multiple of mss if possible. We assume here that mss >= 1.
175 * This MUST be enforced by all callers.
177 void tcp_select_initial_window(int __space
, __u32 mss
,
178 __u32
*rcv_wnd
, __u32
*window_clamp
,
179 int wscale_ok
, __u8
*rcv_wscale
)
181 unsigned int space
= (__space
< 0 ? 0 : __space
);
183 /* If no clamp set the clamp to the max possible scaled window */
184 if (*window_clamp
== 0)
185 (*window_clamp
) = (65535 << 14);
186 space
= min(*window_clamp
, space
);
188 /* Quantize space offering to a multiple of mss if possible. */
190 space
= (space
/ mss
) * mss
;
192 /* NOTE: offering an initial window larger than 32767
193 * will break some buggy TCP stacks. If the admin tells us
194 * it is likely we could be speaking with such a buggy stack
195 * we will truncate our initial window offering to 32K-1
196 * unless the remote has sent us a window scaling option,
197 * which we interpret as a sign the remote TCP is not
198 * misinterpreting the window field as a signed quantity.
200 if (sysctl_tcp_workaround_signed_windows
)
201 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
207 /* Set window scaling on max possible window
208 * See RFC1323 for an explanation of the limit to 14
210 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
211 space
= min_t(u32
, space
, *window_clamp
);
212 while (space
> 65535 && (*rcv_wscale
) < 14) {
218 /* Set initial window to value enough for senders,
219 * following RFC2414. Senders, not following this RFC,
220 * will be satisfied with 2.
222 if (mss
> (1 << *rcv_wscale
)) {
228 if (*rcv_wnd
> init_cwnd
* mss
)
229 *rcv_wnd
= init_cwnd
* mss
;
232 /* Set the clamp no higher than max representable value */
233 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
236 /* Chose a new window to advertise, update state in tcp_sock for the
237 * socket, and return result with RFC1323 scaling applied. The return
238 * value can be stuffed directly into th->window for an outgoing
241 static u16
tcp_select_window(struct sock
*sk
)
243 struct tcp_sock
*tp
= tcp_sk(sk
);
244 u32 cur_win
= tcp_receive_window(tp
);
245 u32 new_win
= __tcp_select_window(sk
);
247 /* Never shrink the offered window */
248 if (new_win
< cur_win
) {
249 /* Danger Will Robinson!
250 * Don't update rcv_wup/rcv_wnd here or else
251 * we will not be able to advertise a zero
252 * window in time. --DaveM
254 * Relax Will Robinson.
256 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
258 tp
->rcv_wnd
= new_win
;
259 tp
->rcv_wup
= tp
->rcv_nxt
;
261 /* Make sure we do not exceed the maximum possible
264 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
265 new_win
= min(new_win
, MAX_TCP_WINDOW
);
267 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
269 /* RFC1323 scaling applied */
270 new_win
>>= tp
->rx_opt
.rcv_wscale
;
272 /* If we advertise zero window, disable fast path. */
279 static inline void TCP_ECN_send_synack(struct tcp_sock
*tp
, struct sk_buff
*skb
)
281 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_CWR
;
282 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
283 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_ECE
;
286 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
288 struct tcp_sock
*tp
= tcp_sk(sk
);
291 if (sysctl_tcp_ecn
) {
292 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ECE
| TCPCB_FLAG_CWR
;
293 tp
->ecn_flags
= TCP_ECN_OK
;
297 static __inline__
void
298 TCP_ECN_make_synack(struct request_sock
*req
, struct tcphdr
*th
)
300 if (inet_rsk(req
)->ecn_ok
)
304 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
307 struct tcp_sock
*tp
= tcp_sk(sk
);
309 if (tp
->ecn_flags
& TCP_ECN_OK
) {
310 /* Not-retransmitted data segment: set ECT and inject CWR. */
311 if (skb
->len
!= tcp_header_len
&&
312 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
314 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
315 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
316 tcp_hdr(skb
)->cwr
= 1;
317 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
320 /* ACK or retransmitted segment: clear ECT|CE */
321 INET_ECN_dontxmit(sk
);
323 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
324 tcp_hdr(skb
)->ece
= 1;
328 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
329 * auto increment end seqno.
331 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
335 TCP_SKB_CB(skb
)->flags
= flags
;
336 TCP_SKB_CB(skb
)->sacked
= 0;
338 skb_shinfo(skb
)->gso_segs
= 1;
339 skb_shinfo(skb
)->gso_size
= 0;
340 skb_shinfo(skb
)->gso_type
= 0;
342 TCP_SKB_CB(skb
)->seq
= seq
;
343 if (flags
& (TCPCB_FLAG_SYN
| TCPCB_FLAG_FIN
))
345 TCP_SKB_CB(skb
)->end_seq
= seq
;
348 static void tcp_build_and_update_options(__be32
*ptr
, struct tcp_sock
*tp
,
349 __u32 tstamp
, __u8
**md5_hash
)
351 if (tp
->rx_opt
.tstamp_ok
) {
352 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
354 (TCPOPT_TIMESTAMP
<< 8) |
356 *ptr
++ = htonl(tstamp
);
357 *ptr
++ = htonl(tp
->rx_opt
.ts_recent
);
359 if (tp
->rx_opt
.eff_sacks
) {
360 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
? tp
->duplicate_sack
: tp
->selective_acks
;
363 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
366 (TCPOLEN_SACK_BASE
+ (tp
->rx_opt
.eff_sacks
*
367 TCPOLEN_SACK_PERBLOCK
)));
369 for (this_sack
= 0; this_sack
< tp
->rx_opt
.eff_sacks
; this_sack
++) {
370 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
371 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
374 if (tp
->rx_opt
.dsack
) {
375 tp
->rx_opt
.dsack
= 0;
376 tp
->rx_opt
.eff_sacks
--;
379 #ifdef CONFIG_TCP_MD5SIG
381 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
383 (TCPOPT_MD5SIG
<< 8) |
385 *md5_hash
= (__u8
*)ptr
;
390 /* Construct a tcp options header for a SYN or SYN_ACK packet.
391 * If this is every changed make sure to change the definition of
392 * MAX_SYN_SIZE to match the new maximum number of options that you
395 * Note - that with the RFC2385 TCP option, we make room for the
396 * 16 byte MD5 hash. This will be filled in later, so the pointer for the
397 * location to be filled is passed back up.
399 static void tcp_syn_build_options(__be32
*ptr
, int mss
, int ts
, int sack
,
400 int offer_wscale
, int wscale
, __u32 tstamp
,
401 __u32 ts_recent
, __u8
**md5_hash
)
403 /* We always get an MSS option.
404 * The option bytes which will be seen in normal data
405 * packets should timestamps be used, must be in the MSS
406 * advertised. But we subtract them from tp->mss_cache so
407 * that calculations in tcp_sendmsg are simpler etc.
408 * So account for this fact here if necessary. If we
409 * don't do this correctly, as a receiver we won't
410 * recognize data packets as being full sized when we
411 * should, and thus we won't abide by the delayed ACK
413 * SACKs don't matter, we never delay an ACK when we
414 * have any of those going out.
416 *ptr
++ = htonl((TCPOPT_MSS
<< 24) | (TCPOLEN_MSS
<< 16) | mss
);
419 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
420 (TCPOLEN_SACK_PERM
<< 16) |
421 (TCPOPT_TIMESTAMP
<< 8) |
424 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
426 (TCPOPT_TIMESTAMP
<< 8) |
428 *ptr
++ = htonl(tstamp
); /* TSVAL */
429 *ptr
++ = htonl(ts_recent
); /* TSECR */
431 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
433 (TCPOPT_SACK_PERM
<< 8) |
436 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
437 (TCPOPT_WINDOW
<< 16) |
438 (TCPOLEN_WINDOW
<< 8) |
440 #ifdef CONFIG_TCP_MD5SIG
442 * If MD5 is enabled, then we set the option, and include the size
443 * (always 18). The actual MD5 hash is added just before the
447 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
449 (TCPOPT_MD5SIG
<< 8) |
451 *md5_hash
= (__u8
*)ptr
;
456 /* This routine actually transmits TCP packets queued in by
457 * tcp_do_sendmsg(). This is used by both the initial
458 * transmission and possible later retransmissions.
459 * All SKB's seen here are completely headerless. It is our
460 * job to build the TCP header, and pass the packet down to
461 * IP so it can do the same plus pass the packet off to the
464 * We are working here with either a clone of the original
465 * SKB, or a fresh unique copy made by the retransmit engine.
467 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
470 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
471 struct inet_sock
*inet
;
473 struct tcp_skb_cb
*tcb
;
475 #ifdef CONFIG_TCP_MD5SIG
476 struct tcp_md5sig_key
*md5
;
477 __u8
*md5_hash_location
;
483 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
485 /* If congestion control is doing timestamping, we must
486 * take such a timestamp before we potentially clone/copy.
488 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
489 __net_timestamp(skb
);
491 if (likely(clone_it
)) {
492 if (unlikely(skb_cloned(skb
)))
493 skb
= pskb_copy(skb
, gfp_mask
);
495 skb
= skb_clone(skb
, gfp_mask
);
502 tcb
= TCP_SKB_CB(skb
);
503 tcp_header_size
= tp
->tcp_header_len
;
505 #define SYSCTL_FLAG_TSTAMPS 0x1
506 #define SYSCTL_FLAG_WSCALE 0x2
507 #define SYSCTL_FLAG_SACK 0x4
510 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
511 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
512 if (sysctl_tcp_timestamps
) {
513 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
514 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
516 if (sysctl_tcp_window_scaling
) {
517 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
518 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
520 if (sysctl_tcp_sack
) {
521 sysctl_flags
|= SYSCTL_FLAG_SACK
;
522 if (!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
523 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
525 } else if (unlikely(tp
->rx_opt
.eff_sacks
)) {
526 /* A SACK is 2 pad bytes, a 2 byte header, plus
527 * 2 32-bit sequence numbers for each SACK block.
529 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
530 (tp
->rx_opt
.eff_sacks
*
531 TCPOLEN_SACK_PERBLOCK
));
534 if (tcp_packets_in_flight(tp
) == 0)
535 tcp_ca_event(sk
, CA_EVENT_TX_START
);
537 #ifdef CONFIG_TCP_MD5SIG
539 * Are we doing MD5 on this segment? If so - make
542 md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
544 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
547 skb_push(skb
, tcp_header_size
);
548 skb_reset_transport_header(skb
);
549 skb_set_owner_w(skb
, sk
);
551 /* Build TCP header and checksum it. */
553 th
->source
= inet
->sport
;
554 th
->dest
= inet
->dport
;
555 th
->seq
= htonl(tcb
->seq
);
556 th
->ack_seq
= htonl(tp
->rcv_nxt
);
557 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
560 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
561 /* RFC1323: The window in SYN & SYN/ACK segments
564 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
566 th
->window
= htons(tcp_select_window(sk
));
571 if (unlikely(tp
->urg_mode
&&
572 between(tp
->snd_up
, tcb
->seq
+ 1, tcb
->seq
+ 0xFFFF))) {
573 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
577 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
578 tcp_syn_build_options((__be32
*)(th
+ 1),
579 tcp_advertise_mss(sk
),
580 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
581 (sysctl_flags
& SYSCTL_FLAG_SACK
),
582 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
583 tp
->rx_opt
.rcv_wscale
,
585 tp
->rx_opt
.ts_recent
,
587 #ifdef CONFIG_TCP_MD5SIG
588 md5
? &md5_hash_location
:
592 tcp_build_and_update_options((__be32
*)(th
+ 1),
594 #ifdef CONFIG_TCP_MD5SIG
595 md5
? &md5_hash_location
:
598 TCP_ECN_send(sk
, skb
, tcp_header_size
);
601 #ifdef CONFIG_TCP_MD5SIG
602 /* Calculate the MD5 hash, as we have all we need now */
604 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
612 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
614 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
615 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
617 if (skb
->len
!= tcp_header_size
)
618 tcp_event_data_sent(tp
, skb
, sk
);
620 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
621 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
623 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
624 if (likely(err
<= 0))
627 tcp_enter_cwr(sk
, 1);
629 return net_xmit_eval(err
);
631 #undef SYSCTL_FLAG_TSTAMPS
632 #undef SYSCTL_FLAG_WSCALE
633 #undef SYSCTL_FLAG_SACK
636 /* This routine just queue's the buffer
638 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
639 * otherwise socket can stall.
641 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
643 struct tcp_sock
*tp
= tcp_sk(sk
);
645 /* Advance write_seq and place onto the write_queue. */
646 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
647 skb_header_release(skb
);
648 tcp_add_write_queue_tail(sk
, skb
);
649 sk
->sk_wmem_queued
+= skb
->truesize
;
650 sk_mem_charge(sk
, skb
->truesize
);
653 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
654 unsigned int mss_now
)
656 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
)) {
657 /* Avoid the costly divide in the normal
660 skb_shinfo(skb
)->gso_segs
= 1;
661 skb_shinfo(skb
)->gso_size
= 0;
662 skb_shinfo(skb
)->gso_type
= 0;
664 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
665 skb_shinfo(skb
)->gso_size
= mss_now
;
666 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
670 /* When a modification to fackets out becomes necessary, we need to check
671 * skb is counted to fackets_out or not.
673 static void tcp_adjust_fackets_out(struct sock
*sk
, struct sk_buff
*skb
,
676 struct tcp_sock
*tp
= tcp_sk(sk
);
678 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
681 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
682 tp
->fackets_out
-= decr
;
685 /* Function to create two new TCP segments. Shrinks the given segment
686 * to the specified size and appends a new segment with the rest of the
687 * packet to the list. This won't be called frequently, I hope.
688 * Remember, these are still headerless SKBs at this point.
690 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
691 unsigned int mss_now
)
693 struct tcp_sock
*tp
= tcp_sk(sk
);
694 struct sk_buff
*buff
;
695 int nsize
, old_factor
;
699 BUG_ON(len
> skb
->len
);
701 tcp_clear_retrans_hints_partial(tp
);
702 nsize
= skb_headlen(skb
) - len
;
706 if (skb_cloned(skb
) &&
707 skb_is_nonlinear(skb
) &&
708 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
711 /* Get a new skb... force flag on. */
712 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
714 return -ENOMEM
; /* We'll just try again later. */
716 sk
->sk_wmem_queued
+= buff
->truesize
;
717 sk_mem_charge(sk
, buff
->truesize
);
718 nlen
= skb
->len
- len
- nsize
;
719 buff
->truesize
+= nlen
;
720 skb
->truesize
-= nlen
;
722 /* Correct the sequence numbers. */
723 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
724 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
725 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
727 /* PSH and FIN should only be set in the second packet. */
728 flags
= TCP_SKB_CB(skb
)->flags
;
729 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
730 TCP_SKB_CB(buff
)->flags
= flags
;
731 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
733 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
734 /* Copy and checksum data tail into the new buffer. */
735 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
736 skb_put(buff
, nsize
),
741 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
743 skb
->ip_summed
= CHECKSUM_PARTIAL
;
744 skb_split(skb
, buff
, len
);
747 buff
->ip_summed
= skb
->ip_summed
;
749 /* Looks stupid, but our code really uses when of
750 * skbs, which it never sent before. --ANK
752 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
753 buff
->tstamp
= skb
->tstamp
;
755 old_factor
= tcp_skb_pcount(skb
);
757 /* Fix up tso_factor for both original and new SKB. */
758 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
759 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
761 /* If this packet has been sent out already, we must
762 * adjust the various packet counters.
764 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
765 int diff
= old_factor
- tcp_skb_pcount(skb
) -
766 tcp_skb_pcount(buff
);
768 tp
->packets_out
-= diff
;
770 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
771 tp
->sacked_out
-= diff
;
772 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
773 tp
->retrans_out
-= diff
;
775 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
776 tp
->lost_out
-= diff
;
778 /* Adjust Reno SACK estimate. */
779 if (tcp_is_reno(tp
) && diff
> 0) {
780 tcp_dec_pcount_approx_int(&tp
->sacked_out
, diff
);
781 tcp_verify_left_out(tp
);
783 tcp_adjust_fackets_out(sk
, skb
, diff
);
786 /* Link BUFF into the send queue. */
787 skb_header_release(buff
);
788 tcp_insert_write_queue_after(skb
, buff
, sk
);
793 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
794 * eventually). The difference is that pulled data not copied, but
795 * immediately discarded.
797 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
803 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
804 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
805 put_page(skb_shinfo(skb
)->frags
[i
].page
);
806 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
808 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
810 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
811 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
817 skb_shinfo(skb
)->nr_frags
= k
;
819 skb_reset_tail_pointer(skb
);
820 skb
->data_len
-= len
;
821 skb
->len
= skb
->data_len
;
824 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
826 if (skb_cloned(skb
) && pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
829 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
830 if (unlikely(len
< skb_headlen(skb
)))
831 __skb_pull(skb
, len
);
833 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
835 TCP_SKB_CB(skb
)->seq
+= len
;
836 skb
->ip_summed
= CHECKSUM_PARTIAL
;
838 skb
->truesize
-= len
;
839 sk
->sk_wmem_queued
-= len
;
840 sk_mem_uncharge(sk
, len
);
841 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
843 /* Any change of skb->len requires recalculation of tso
846 if (tcp_skb_pcount(skb
) > 1)
847 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
852 /* Not accounting for SACKs here. */
853 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
855 struct tcp_sock
*tp
= tcp_sk(sk
);
856 struct inet_connection_sock
*icsk
= inet_csk(sk
);
859 /* Calculate base mss without TCP options:
860 It is MMS_S - sizeof(tcphdr) of rfc1122
862 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
864 /* Clamp it (mss_clamp does not include tcp options) */
865 if (mss_now
> tp
->rx_opt
.mss_clamp
)
866 mss_now
= tp
->rx_opt
.mss_clamp
;
868 /* Now subtract optional transport overhead */
869 mss_now
-= icsk
->icsk_ext_hdr_len
;
871 /* Then reserve room for full set of TCP options and 8 bytes of data */
875 /* Now subtract TCP options size, not including SACKs */
876 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
881 /* Inverse of above */
882 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
884 struct tcp_sock
*tp
= tcp_sk(sk
);
885 struct inet_connection_sock
*icsk
= inet_csk(sk
);
890 icsk
->icsk_ext_hdr_len
+
891 icsk
->icsk_af_ops
->net_header_len
;
896 void tcp_mtup_init(struct sock
*sk
)
898 struct tcp_sock
*tp
= tcp_sk(sk
);
899 struct inet_connection_sock
*icsk
= inet_csk(sk
);
901 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
902 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
903 icsk
->icsk_af_ops
->net_header_len
;
904 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
905 icsk
->icsk_mtup
.probe_size
= 0;
908 /* Bound MSS / TSO packet size with the half of the window */
909 static int tcp_bound_to_half_wnd(struct tcp_sock
*tp
, int pktsize
)
911 if (tp
->max_window
&& pktsize
> (tp
->max_window
>> 1))
912 return max(tp
->max_window
>> 1, 68U - tp
->tcp_header_len
);
917 /* This function synchronize snd mss to current pmtu/exthdr set.
919 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
920 for TCP options, but includes only bare TCP header.
922 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
923 It is minimum of user_mss and mss received with SYN.
924 It also does not include TCP options.
926 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
928 tp->mss_cache is current effective sending mss, including
929 all tcp options except for SACKs. It is evaluated,
930 taking into account current pmtu, but never exceeds
931 tp->rx_opt.mss_clamp.
933 NOTE1. rfc1122 clearly states that advertised MSS
934 DOES NOT include either tcp or ip options.
936 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
937 are READ ONLY outside this function. --ANK (980731)
939 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
941 struct tcp_sock
*tp
= tcp_sk(sk
);
942 struct inet_connection_sock
*icsk
= inet_csk(sk
);
945 if (icsk
->icsk_mtup
.search_high
> pmtu
)
946 icsk
->icsk_mtup
.search_high
= pmtu
;
948 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
949 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
951 /* And store cached results */
952 icsk
->icsk_pmtu_cookie
= pmtu
;
953 if (icsk
->icsk_mtup
.enabled
)
954 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
955 tp
->mss_cache
= mss_now
;
960 /* Compute the current effective MSS, taking SACKs and IP options,
961 * and even PMTU discovery events into account.
963 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
964 * cannot be large. However, taking into account rare use of URG, this
967 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
969 struct tcp_sock
*tp
= tcp_sk(sk
);
970 struct dst_entry
*dst
= __sk_dst_get(sk
);
975 mss_now
= tp
->mss_cache
;
977 if (large_allowed
&& sk_can_gso(sk
) && !tp
->urg_mode
)
981 u32 mtu
= dst_mtu(dst
);
982 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
983 mss_now
= tcp_sync_mss(sk
, mtu
);
986 if (tp
->rx_opt
.eff_sacks
)
987 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
988 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
990 #ifdef CONFIG_TCP_MD5SIG
991 if (tp
->af_specific
->md5_lookup(sk
, sk
))
992 mss_now
-= TCPOLEN_MD5SIG_ALIGNED
;
995 xmit_size_goal
= mss_now
;
998 xmit_size_goal
= ((sk
->sk_gso_max_size
- 1) -
999 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
1000 inet_csk(sk
)->icsk_ext_hdr_len
-
1001 tp
->tcp_header_len
);
1003 xmit_size_goal
= tcp_bound_to_half_wnd(tp
, xmit_size_goal
);
1004 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
1006 tp
->xmit_size_goal
= xmit_size_goal
;
1011 /* Congestion window validation. (RFC2861) */
1012 static void tcp_cwnd_validate(struct sock
*sk
)
1014 struct tcp_sock
*tp
= tcp_sk(sk
);
1016 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1017 /* Network is feed fully. */
1018 tp
->snd_cwnd_used
= 0;
1019 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1021 /* Network starves. */
1022 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1023 tp
->snd_cwnd_used
= tp
->packets_out
;
1025 if (sysctl_tcp_slow_start_after_idle
&&
1026 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1027 tcp_cwnd_application_limited(sk
);
1031 /* Returns the portion of skb which can be sent right away without
1032 * introducing MSS oddities to segment boundaries. In rare cases where
1033 * mss_now != mss_cache, we will request caller to create a small skb
1034 * per input skb which could be mostly avoided here (if desired).
1036 * We explicitly want to create a request for splitting write queue tail
1037 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1038 * thus all the complexity (cwnd_len is always MSS multiple which we
1039 * return whenever allowed by the other factors). Basically we need the
1040 * modulo only when the receiver window alone is the limiting factor or
1041 * when we would be allowed to send the split-due-to-Nagle skb fully.
1043 static unsigned int tcp_mss_split_point(struct sock
*sk
, struct sk_buff
*skb
,
1044 unsigned int mss_now
, unsigned int cwnd
)
1046 struct tcp_sock
*tp
= tcp_sk(sk
);
1047 u32 needed
, window
, cwnd_len
;
1049 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1050 cwnd_len
= mss_now
* cwnd
;
1052 if (likely(cwnd_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1055 needed
= min(skb
->len
, window
);
1057 if (cwnd_len
<= needed
)
1060 return needed
- needed
% mss_now
;
1063 /* Can at least one segment of SKB be sent right now, according to the
1064 * congestion window rules? If so, return how many segments are allowed.
1066 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
,
1067 struct sk_buff
*skb
)
1069 u32 in_flight
, cwnd
;
1071 /* Don't be strict about the congestion window for the final FIN. */
1072 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1073 tcp_skb_pcount(skb
) == 1)
1076 in_flight
= tcp_packets_in_flight(tp
);
1077 cwnd
= tp
->snd_cwnd
;
1078 if (in_flight
< cwnd
)
1079 return (cwnd
- in_flight
);
1084 /* This must be invoked the first time we consider transmitting
1085 * SKB onto the wire.
1087 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
1088 unsigned int mss_now
)
1090 int tso_segs
= tcp_skb_pcount(skb
);
1092 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1093 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1094 tso_segs
= tcp_skb_pcount(skb
);
1099 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1101 return after(tp
->snd_sml
,tp
->snd_una
) &&
1102 !after(tp
->snd_sml
, tp
->snd_nxt
);
1105 /* Return 0, if packet can be sent now without violation Nagle's rules:
1106 * 1. It is full sized.
1107 * 2. Or it contains FIN. (already checked by caller)
1108 * 3. Or TCP_NODELAY was set.
1109 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1110 * With Minshall's modification: all sent small packets are ACKed.
1112 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1113 const struct sk_buff
*skb
,
1114 unsigned mss_now
, int nonagle
)
1116 return (skb
->len
< mss_now
&&
1117 ((nonagle
& TCP_NAGLE_CORK
) ||
1118 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
))));
1121 /* Return non-zero if the Nagle test allows this packet to be
1124 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1125 unsigned int cur_mss
, int nonagle
)
1127 /* Nagle rule does not apply to frames, which sit in the middle of the
1128 * write_queue (they have no chances to get new data).
1130 * This is implemented in the callers, where they modify the 'nonagle'
1131 * argument based upon the location of SKB in the send queue.
1133 if (nonagle
& TCP_NAGLE_PUSH
)
1136 /* Don't use the nagle rule for urgent data (or for the final FIN).
1137 * Nagle can be ignored during F-RTO too (see RFC4138).
1139 if (tp
->urg_mode
|| (tp
->frto_counter
== 2) ||
1140 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1143 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1149 /* Does at least the first segment of SKB fit into the send window? */
1150 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1151 unsigned int cur_mss
)
1153 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1155 if (skb
->len
> cur_mss
)
1156 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1158 return !after(end_seq
, tcp_wnd_end(tp
));
1161 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1162 * should be put on the wire right now. If so, it returns the number of
1163 * packets allowed by the congestion window.
1165 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1166 unsigned int cur_mss
, int nonagle
)
1168 struct tcp_sock
*tp
= tcp_sk(sk
);
1169 unsigned int cwnd_quota
;
1171 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1173 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1176 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1177 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1183 int tcp_may_send_now(struct sock
*sk
)
1185 struct tcp_sock
*tp
= tcp_sk(sk
);
1186 struct sk_buff
*skb
= tcp_send_head(sk
);
1189 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1190 (tcp_skb_is_last(sk
, skb
) ?
1191 tp
->nonagle
: TCP_NAGLE_PUSH
)));
1194 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1195 * which is put after SKB on the list. It is very much like
1196 * tcp_fragment() except that it may make several kinds of assumptions
1197 * in order to speed up the splitting operation. In particular, we
1198 * know that all the data is in scatter-gather pages, and that the
1199 * packet has never been sent out before (and thus is not cloned).
1201 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1202 unsigned int mss_now
)
1204 struct sk_buff
*buff
;
1205 int nlen
= skb
->len
- len
;
1208 /* All of a TSO frame must be composed of paged data. */
1209 if (skb
->len
!= skb
->data_len
)
1210 return tcp_fragment(sk
, skb
, len
, mss_now
);
1212 buff
= sk_stream_alloc_skb(sk
, 0, GFP_ATOMIC
);
1213 if (unlikely(buff
== NULL
))
1216 sk
->sk_wmem_queued
+= buff
->truesize
;
1217 sk_mem_charge(sk
, buff
->truesize
);
1218 buff
->truesize
+= nlen
;
1219 skb
->truesize
-= nlen
;
1221 /* Correct the sequence numbers. */
1222 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1223 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1224 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1226 /* PSH and FIN should only be set in the second packet. */
1227 flags
= TCP_SKB_CB(skb
)->flags
;
1228 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
1229 TCP_SKB_CB(buff
)->flags
= flags
;
1231 /* This packet was never sent out yet, so no SACK bits. */
1232 TCP_SKB_CB(buff
)->sacked
= 0;
1234 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1235 skb_split(skb
, buff
, len
);
1237 /* Fix up tso_factor for both original and new SKB. */
1238 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1239 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1241 /* Link BUFF into the send queue. */
1242 skb_header_release(buff
);
1243 tcp_insert_write_queue_after(skb
, buff
, sk
);
1248 /* Try to defer sending, if possible, in order to minimize the amount
1249 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1251 * This algorithm is from John Heffner.
1253 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1255 struct tcp_sock
*tp
= tcp_sk(sk
);
1256 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1257 u32 send_win
, cong_win
, limit
, in_flight
;
1259 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1262 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1265 /* Defer for less than two clock ticks. */
1266 if (tp
->tso_deferred
&&
1267 ((jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1270 in_flight
= tcp_packets_in_flight(tp
);
1272 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1274 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1276 /* From in_flight test above, we know that cwnd > in_flight. */
1277 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1279 limit
= min(send_win
, cong_win
);
1281 /* If a full-sized TSO skb can be sent, do it. */
1282 if (limit
>= sk
->sk_gso_max_size
)
1285 if (sysctl_tcp_tso_win_divisor
) {
1286 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1288 /* If at least some fraction of a window is available,
1291 chunk
/= sysctl_tcp_tso_win_divisor
;
1295 /* Different approach, try not to defer past a single
1296 * ACK. Receiver should ACK every other full sized
1297 * frame, so if we have space for more than 3 frames
1300 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1304 /* Ok, it looks like it is advisable to defer. */
1305 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1310 tp
->tso_deferred
= 0;
1314 /* Create a new MTU probe if we are ready.
1315 * Returns 0 if we should wait to probe (no cwnd available),
1316 * 1 if a probe was sent,
1319 static int tcp_mtu_probe(struct sock
*sk
)
1321 struct tcp_sock
*tp
= tcp_sk(sk
);
1322 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1323 struct sk_buff
*skb
, *nskb
, *next
;
1330 /* Not currently probing/verifying,
1332 * have enough cwnd, and
1333 * not SACKing (the variable headers throw things off) */
1334 if (!icsk
->icsk_mtup
.enabled
||
1335 icsk
->icsk_mtup
.probe_size
||
1336 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1337 tp
->snd_cwnd
< 11 ||
1338 tp
->rx_opt
.eff_sacks
)
1341 /* Very simple search strategy: just double the MSS. */
1342 mss_now
= tcp_current_mss(sk
, 0);
1343 probe_size
= 2 * tp
->mss_cache
;
1344 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1345 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1346 /* TODO: set timer for probe_converge_event */
1350 /* Have enough data in the send queue to probe? */
1351 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1354 if (tp
->snd_wnd
< size_needed
)
1356 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1359 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1360 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1361 if (!tcp_packets_in_flight(tp
))
1367 /* We're allowed to probe. Build it now. */
1368 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1370 sk
->sk_wmem_queued
+= nskb
->truesize
;
1371 sk_mem_charge(sk
, nskb
->truesize
);
1373 skb
= tcp_send_head(sk
);
1375 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1376 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1377 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1378 TCP_SKB_CB(nskb
)->sacked
= 0;
1380 nskb
->ip_summed
= skb
->ip_summed
;
1382 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1385 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1386 copy
= min_t(int, skb
->len
, probe_size
- len
);
1387 if (nskb
->ip_summed
)
1388 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1390 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1391 skb_put(nskb
, copy
),
1394 if (skb
->len
<= copy
) {
1395 /* We've eaten all the data from this skb.
1397 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1398 tcp_unlink_write_queue(skb
, sk
);
1399 sk_wmem_free_skb(sk
, skb
);
1401 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1402 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1403 if (!skb_shinfo(skb
)->nr_frags
) {
1404 skb_pull(skb
, copy
);
1405 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1406 skb
->csum
= csum_partial(skb
->data
,
1409 __pskb_trim_head(skb
, copy
);
1410 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1412 TCP_SKB_CB(skb
)->seq
+= copy
;
1417 if (len
>= probe_size
)
1420 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1422 /* We're ready to send. If this fails, the probe will
1423 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1424 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1425 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1426 /* Decrement cwnd here because we are sending
1427 * effectively two packets. */
1429 tcp_event_new_data_sent(sk
, nskb
);
1431 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1432 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1433 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1441 /* This routine writes packets to the network. It advances the
1442 * send_head. This happens as incoming acks open up the remote
1445 * Returns 1, if no segments are in flight and we have queued segments, but
1446 * cannot send anything now because of SWS or another problem.
1448 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1450 struct tcp_sock
*tp
= tcp_sk(sk
);
1451 struct sk_buff
*skb
;
1452 unsigned int tso_segs
, sent_pkts
;
1456 /* If we are closed, the bytes will have to remain here.
1457 * In time closedown will finish, we empty the write queue and all
1460 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1465 /* Do MTU probing. */
1466 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1468 } else if (result
> 0) {
1472 while ((skb
= tcp_send_head(sk
))) {
1475 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1478 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1482 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1485 if (tso_segs
== 1) {
1486 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1487 (tcp_skb_is_last(sk
, skb
) ?
1488 nonagle
: TCP_NAGLE_PUSH
))))
1491 if (tcp_tso_should_defer(sk
, skb
))
1497 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1500 if (skb
->len
> limit
&&
1501 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1504 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1506 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1509 /* Advance the send_head. This one is sent out.
1510 * This call will increment packets_out.
1512 tcp_event_new_data_sent(sk
, skb
);
1514 tcp_minshall_update(tp
, mss_now
, skb
);
1518 if (likely(sent_pkts
)) {
1519 tcp_cwnd_validate(sk
);
1522 return !tp
->packets_out
&& tcp_send_head(sk
);
1525 /* Push out any pending frames which were held back due to
1526 * TCP_CORK or attempt at coalescing tiny packets.
1527 * The socket must be locked by the caller.
1529 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1532 struct sk_buff
*skb
= tcp_send_head(sk
);
1535 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1536 tcp_check_probe_timer(sk
);
1540 /* Send _single_ skb sitting at the send head. This function requires
1541 * true push pending frames to setup probe timer etc.
1543 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1545 struct sk_buff
*skb
= tcp_send_head(sk
);
1546 unsigned int tso_segs
, cwnd_quota
;
1548 BUG_ON(!skb
|| skb
->len
< mss_now
);
1550 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1551 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1553 if (likely(cwnd_quota
)) {
1560 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1563 if (skb
->len
> limit
&&
1564 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1567 /* Send it out now. */
1568 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1570 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1571 tcp_event_new_data_sent(sk
, skb
);
1572 tcp_cwnd_validate(sk
);
1578 /* This function returns the amount that we can raise the
1579 * usable window based on the following constraints
1581 * 1. The window can never be shrunk once it is offered (RFC 793)
1582 * 2. We limit memory per socket
1585 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1586 * RECV.NEXT + RCV.WIN fixed until:
1587 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1589 * i.e. don't raise the right edge of the window until you can raise
1590 * it at least MSS bytes.
1592 * Unfortunately, the recommended algorithm breaks header prediction,
1593 * since header prediction assumes th->window stays fixed.
1595 * Strictly speaking, keeping th->window fixed violates the receiver
1596 * side SWS prevention criteria. The problem is that under this rule
1597 * a stream of single byte packets will cause the right side of the
1598 * window to always advance by a single byte.
1600 * Of course, if the sender implements sender side SWS prevention
1601 * then this will not be a problem.
1603 * BSD seems to make the following compromise:
1605 * If the free space is less than the 1/4 of the maximum
1606 * space available and the free space is less than 1/2 mss,
1607 * then set the window to 0.
1608 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1609 * Otherwise, just prevent the window from shrinking
1610 * and from being larger than the largest representable value.
1612 * This prevents incremental opening of the window in the regime
1613 * where TCP is limited by the speed of the reader side taking
1614 * data out of the TCP receive queue. It does nothing about
1615 * those cases where the window is constrained on the sender side
1616 * because the pipeline is full.
1618 * BSD also seems to "accidentally" limit itself to windows that are a
1619 * multiple of MSS, at least until the free space gets quite small.
1620 * This would appear to be a side effect of the mbuf implementation.
1621 * Combining these two algorithms results in the observed behavior
1622 * of having a fixed window size at almost all times.
1624 * Below we obtain similar behavior by forcing the offered window to
1625 * a multiple of the mss when it is feasible to do so.
1627 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1628 * Regular options like TIMESTAMP are taken into account.
1630 u32
__tcp_select_window(struct sock
*sk
)
1632 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1633 struct tcp_sock
*tp
= tcp_sk(sk
);
1634 /* MSS for the peer's data. Previous versions used mss_clamp
1635 * here. I don't know if the value based on our guesses
1636 * of peer's MSS is better for the performance. It's more correct
1637 * but may be worse for the performance because of rcv_mss
1638 * fluctuations. --SAW 1998/11/1
1640 int mss
= icsk
->icsk_ack
.rcv_mss
;
1641 int free_space
= tcp_space(sk
);
1642 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1645 if (mss
> full_space
)
1648 if (free_space
< (full_space
>> 1)) {
1649 icsk
->icsk_ack
.quick
= 0;
1651 if (tcp_memory_pressure
)
1652 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
1655 if (free_space
< mss
)
1659 if (free_space
> tp
->rcv_ssthresh
)
1660 free_space
= tp
->rcv_ssthresh
;
1662 /* Don't do rounding if we are using window scaling, since the
1663 * scaled window will not line up with the MSS boundary anyway.
1665 window
= tp
->rcv_wnd
;
1666 if (tp
->rx_opt
.rcv_wscale
) {
1667 window
= free_space
;
1669 /* Advertise enough space so that it won't get scaled away.
1670 * Import case: prevent zero window announcement if
1671 * 1<<rcv_wscale > mss.
1673 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1674 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1675 << tp
->rx_opt
.rcv_wscale
);
1677 /* Get the largest window that is a nice multiple of mss.
1678 * Window clamp already applied above.
1679 * If our current window offering is within 1 mss of the
1680 * free space we just keep it. This prevents the divide
1681 * and multiply from happening most of the time.
1682 * We also don't do any window rounding when the free space
1685 if (window
<= free_space
- mss
|| window
> free_space
)
1686 window
= (free_space
/ mss
) * mss
;
1687 else if (mss
== full_space
&&
1688 free_space
> window
+ (full_space
>> 1))
1689 window
= free_space
;
1695 /* Attempt to collapse two adjacent SKB's during retransmission. */
1696 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
,
1699 struct tcp_sock
*tp
= tcp_sk(sk
);
1700 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1701 int skb_size
, next_skb_size
;
1704 /* The first test we must make is that neither of these two
1705 * SKB's are still referenced by someone else.
1707 if (skb_cloned(skb
) || skb_cloned(next_skb
))
1710 skb_size
= skb
->len
;
1711 next_skb_size
= next_skb
->len
;
1712 flags
= TCP_SKB_CB(skb
)->flags
;
1714 /* Also punt if next skb has been SACK'd. */
1715 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1718 /* Next skb is out of window. */
1719 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tcp_wnd_end(tp
)))
1722 /* Punt if not enough space exists in the first SKB for
1723 * the data in the second, or the total combined payload
1724 * would exceed the MSS.
1726 if ((next_skb_size
> skb_tailroom(skb
)) ||
1727 ((skb_size
+ next_skb_size
) > mss_now
))
1730 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
1732 tcp_highest_sack_combine(sk
, next_skb
, skb
);
1734 /* Ok. We will be able to collapse the packet. */
1735 tcp_unlink_write_queue(next_skb
, sk
);
1737 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
1740 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1741 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1743 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1744 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1746 /* Update sequence range on original skb. */
1747 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1749 /* Merge over control information. */
1750 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1751 TCP_SKB_CB(skb
)->flags
= flags
;
1753 /* All done, get rid of second SKB and account for it so
1754 * packet counting does not break.
1756 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
1757 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1758 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1759 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_LOST
)
1760 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1761 /* Reno case is special. Sigh... */
1762 if (tcp_is_reno(tp
) && tp
->sacked_out
)
1763 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1765 tcp_adjust_fackets_out(sk
, next_skb
, tcp_skb_pcount(next_skb
));
1766 tp
->packets_out
-= tcp_skb_pcount(next_skb
);
1768 /* changed transmit queue under us so clear hints */
1769 tcp_clear_retrans_hints_partial(tp
);
1771 sk_wmem_free_skb(sk
, next_skb
);
1774 /* Do a simple retransmit without using the backoff mechanisms in
1775 * tcp_timer. This is used for path mtu discovery.
1776 * The socket is already locked here.
1778 void tcp_simple_retransmit(struct sock
*sk
)
1780 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1781 struct tcp_sock
*tp
= tcp_sk(sk
);
1782 struct sk_buff
*skb
;
1783 unsigned int mss
= tcp_current_mss(sk
, 0);
1786 tcp_for_write_queue(skb
, sk
) {
1787 if (skb
== tcp_send_head(sk
))
1789 if (skb
->len
> mss
&&
1790 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1791 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1792 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1793 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1795 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)) {
1796 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1797 tp
->lost_out
+= tcp_skb_pcount(skb
);
1803 tcp_clear_all_retrans_hints(tp
);
1808 if (tcp_is_reno(tp
))
1809 tcp_limit_reno_sacked(tp
);
1811 tcp_verify_left_out(tp
);
1813 /* Don't muck with the congestion window here.
1814 * Reason is that we do not increase amount of _data_
1815 * in network, but units changed and effective
1816 * cwnd/ssthresh really reduced now.
1818 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1819 tp
->high_seq
= tp
->snd_nxt
;
1820 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1821 tp
->prior_ssthresh
= 0;
1822 tp
->undo_marker
= 0;
1823 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1825 tcp_xmit_retransmit_queue(sk
);
1828 /* This retransmits one SKB. Policy decisions and retransmit queue
1829 * state updates are done by the caller. Returns non-zero if an
1830 * error occurred which prevented the send.
1832 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1834 struct tcp_sock
*tp
= tcp_sk(sk
);
1835 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1836 unsigned int cur_mss
;
1839 /* Inconslusive MTU probe */
1840 if (icsk
->icsk_mtup
.probe_size
) {
1841 icsk
->icsk_mtup
.probe_size
= 0;
1844 /* Do not sent more than we queued. 1/4 is reserved for possible
1845 * copying overhead: fragmentation, tunneling, mangling etc.
1847 if (atomic_read(&sk
->sk_wmem_alloc
) >
1848 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1851 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1852 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1854 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1858 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1859 return -EHOSTUNREACH
; /* Routing failure or similar. */
1861 cur_mss
= tcp_current_mss(sk
, 0);
1863 /* If receiver has shrunk his window, and skb is out of
1864 * new window, do not retransmit it. The exception is the
1865 * case, when window is shrunk to zero. In this case
1866 * our retransmit serves as a zero window probe.
1868 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))
1869 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1872 if (skb
->len
> cur_mss
) {
1873 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1874 return -ENOMEM
; /* We'll try again later. */
1877 /* Collapse two adjacent packets if worthwhile and we can. */
1878 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1879 (skb
->len
< (cur_mss
>> 1)) &&
1880 (tcp_write_queue_next(sk
, skb
) != tcp_send_head(sk
)) &&
1881 (!tcp_skb_is_last(sk
, skb
)) &&
1882 (skb_shinfo(skb
)->nr_frags
== 0 &&
1883 skb_shinfo(tcp_write_queue_next(sk
, skb
))->nr_frags
== 0) &&
1884 (tcp_skb_pcount(skb
) == 1 &&
1885 tcp_skb_pcount(tcp_write_queue_next(sk
, skb
)) == 1) &&
1886 (sysctl_tcp_retrans_collapse
!= 0))
1887 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1889 /* Some Solaris stacks overoptimize and ignore the FIN on a
1890 * retransmit when old data is attached. So strip it off
1891 * since it is cheap to do so and saves bytes on the network.
1894 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1895 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1896 if (!pskb_trim(skb
, 0)) {
1897 /* Reuse, even though it does some unnecessary work */
1898 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
1899 TCP_SKB_CB(skb
)->flags
);
1900 skb
->ip_summed
= CHECKSUM_NONE
;
1904 /* Make a copy, if the first transmission SKB clone we made
1905 * is still in somebody's hands, else make a clone.
1907 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1909 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1912 /* Update global TCP statistics. */
1913 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
1915 tp
->total_retrans
++;
1917 #if FASTRETRANS_DEBUG > 0
1918 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1919 if (net_ratelimit())
1920 printk(KERN_DEBUG
"retrans_out leaked.\n");
1923 if (!tp
->retrans_out
)
1924 tp
->lost_retrans_low
= tp
->snd_nxt
;
1925 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1926 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1928 /* Save stamp of the first retransmit. */
1929 if (!tp
->retrans_stamp
)
1930 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1934 /* snd_nxt is stored to detect loss of retransmitted segment,
1935 * see tcp_input.c tcp_sacktag_write_queue().
1937 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1942 /* This gets called after a retransmit timeout, and the initially
1943 * retransmitted data is acknowledged. It tries to continue
1944 * resending the rest of the retransmit queue, until either
1945 * we've sent it all or the congestion window limit is reached.
1946 * If doing SACK, the first ACK which comes back for a timeout
1947 * based retransmit packet might feed us FACK information again.
1948 * If so, we use it to avoid unnecessarily retransmissions.
1950 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1952 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1953 struct tcp_sock
*tp
= tcp_sk(sk
);
1954 struct sk_buff
*skb
;
1957 if (tp
->retransmit_skb_hint
) {
1958 skb
= tp
->retransmit_skb_hint
;
1959 packet_cnt
= tp
->retransmit_cnt_hint
;
1961 skb
= tcp_write_queue_head(sk
);
1965 /* First pass: retransmit lost packets. */
1967 tcp_for_write_queue_from(skb
, sk
) {
1968 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1970 if (skb
== tcp_send_head(sk
))
1972 /* we could do better than to assign each time */
1973 tp
->retransmit_skb_hint
= skb
;
1974 tp
->retransmit_cnt_hint
= packet_cnt
;
1976 /* Assume this retransmit will generate
1977 * only one packet for congestion window
1978 * calculation purposes. This works because
1979 * tcp_retransmit_skb() will chop up the
1980 * packet to be MSS sized and all the
1981 * packet counting works out.
1983 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1986 if (sacked
& TCPCB_LOST
) {
1987 if (!(sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1990 if (tcp_retransmit_skb(sk
, skb
)) {
1991 tp
->retransmit_skb_hint
= NULL
;
1994 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1995 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
1997 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
1998 NET_INC_STATS_BH(mib_idx
);
2000 if (skb
== tcp_write_queue_head(sk
))
2001 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2002 inet_csk(sk
)->icsk_rto
,
2006 packet_cnt
+= tcp_skb_pcount(skb
);
2007 if (packet_cnt
>= tp
->lost_out
)
2013 /* OK, demanded retransmission is finished. */
2015 /* Forward retransmissions are possible only during Recovery. */
2016 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2019 /* No forward retransmissions in Reno are possible. */
2020 if (tcp_is_reno(tp
))
2023 /* Yeah, we have to make difficult choice between forward transmission
2024 * and retransmission... Both ways have their merits...
2026 * For now we do not retransmit anything, while we have some new
2027 * segments to send. In the other cases, follow rule 3 for
2028 * NextSeg() specified in RFC3517.
2031 if (tcp_may_send_now(sk
))
2034 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2035 if (!tp
->sacked_out
)
2038 if (tp
->forward_skb_hint
)
2039 skb
= tp
->forward_skb_hint
;
2041 skb
= tcp_write_queue_head(sk
);
2043 tcp_for_write_queue_from(skb
, sk
) {
2044 if (skb
== tcp_send_head(sk
))
2046 tp
->forward_skb_hint
= skb
;
2048 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2051 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2054 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
2057 /* Ok, retransmit it. */
2058 if (tcp_retransmit_skb(sk
, skb
)) {
2059 tp
->forward_skb_hint
= NULL
;
2063 if (skb
== tcp_write_queue_head(sk
))
2064 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2065 inet_csk(sk
)->icsk_rto
,
2068 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
2072 /* Send a fin. The caller locks the socket for us. This cannot be
2073 * allowed to fail queueing a FIN frame under any circumstances.
2075 void tcp_send_fin(struct sock
*sk
)
2077 struct tcp_sock
*tp
= tcp_sk(sk
);
2078 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2081 /* Optimization, tack on the FIN if we have a queue of
2082 * unsent frames. But be careful about outgoing SACKS
2085 mss_now
= tcp_current_mss(sk
, 1);
2087 if (tcp_send_head(sk
) != NULL
) {
2088 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2089 TCP_SKB_CB(skb
)->end_seq
++;
2092 /* Socket is locked, keep trying until memory is available. */
2094 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
2100 /* Reserve space for headers and prepare control bits. */
2101 skb_reserve(skb
, MAX_TCP_HEADER
);
2102 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2103 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2104 TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2105 tcp_queue_skb(sk
, skb
);
2107 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2110 /* We get here when a process closes a file descriptor (either due to
2111 * an explicit close() or as a byproduct of exit()'ing) and there
2112 * was unread data in the receive queue. This behavior is recommended
2113 * by RFC 2525, section 2.17. -DaveM
2115 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2117 struct sk_buff
*skb
;
2119 /* NOTE: No TCP options attached and we never retransmit this. */
2120 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2122 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2126 /* Reserve space for headers and prepare control bits. */
2127 skb_reserve(skb
, MAX_TCP_HEADER
);
2128 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2129 TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2131 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2132 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2133 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2135 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2138 /* WARNING: This routine must only be called when we have already sent
2139 * a SYN packet that crossed the incoming SYN that caused this routine
2140 * to get called. If this assumption fails then the initial rcv_wnd
2141 * and rcv_wscale values will not be correct.
2143 int tcp_send_synack(struct sock
*sk
)
2145 struct sk_buff
*skb
;
2147 skb
= tcp_write_queue_head(sk
);
2148 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
)) {
2149 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2152 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_ACK
)) {
2153 if (skb_cloned(skb
)) {
2154 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2157 tcp_unlink_write_queue(skb
, sk
);
2158 skb_header_release(nskb
);
2159 __tcp_add_write_queue_head(sk
, nskb
);
2160 sk_wmem_free_skb(sk
, skb
);
2161 sk
->sk_wmem_queued
+= nskb
->truesize
;
2162 sk_mem_charge(sk
, nskb
->truesize
);
2166 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2167 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2169 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2170 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2174 * Prepare a SYN-ACK.
2176 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2177 struct request_sock
*req
)
2179 struct inet_request_sock
*ireq
= inet_rsk(req
);
2180 struct tcp_sock
*tp
= tcp_sk(sk
);
2182 int tcp_header_size
;
2183 struct sk_buff
*skb
;
2184 #ifdef CONFIG_TCP_MD5SIG
2185 struct tcp_md5sig_key
*md5
;
2186 __u8
*md5_hash_location
;
2189 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2193 /* Reserve space for headers. */
2194 skb_reserve(skb
, MAX_TCP_HEADER
);
2196 skb
->dst
= dst_clone(dst
);
2198 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2199 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2200 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2201 /* SACK_PERM is in the place of NOP NOP of TS */
2202 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2204 #ifdef CONFIG_TCP_MD5SIG
2205 /* Are we doing MD5 on this segment? If so - make room for it */
2206 md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
2208 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
2210 skb_push(skb
, tcp_header_size
);
2211 skb_reset_transport_header(skb
);
2214 memset(th
, 0, sizeof(struct tcphdr
));
2217 TCP_ECN_make_synack(req
, th
);
2218 th
->source
= inet_sk(sk
)->sport
;
2219 th
->dest
= ireq
->rmt_port
;
2220 /* Setting of flags are superfluous here for callers (and ECE is
2221 * not even correctly set)
2223 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2224 TCPCB_FLAG_SYN
| TCPCB_FLAG_ACK
);
2225 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2226 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2227 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2229 /* Set this up on the first call only */
2230 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2231 /* tcp_full_space because it is guaranteed to be the first packet */
2232 tcp_select_initial_window(tcp_full_space(sk
),
2233 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2238 ireq
->rcv_wscale
= rcv_wscale
;
2241 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2242 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2243 #ifdef CONFIG_SYN_COOKIES
2244 if (unlikely(req
->cookie_ts
))
2245 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2248 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2249 tcp_syn_build_options((__be32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2250 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2251 TCP_SKB_CB(skb
)->when
,
2254 #ifdef CONFIG_TCP_MD5SIG
2255 md5
? &md5_hash_location
:
2260 th
->doff
= (tcp_header_size
>> 2);
2261 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
2263 #ifdef CONFIG_TCP_MD5SIG
2264 /* Okay, we have all we need - do the md5 hash if needed */
2266 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
2278 * Do all connect socket setups that can be done AF independent.
2280 static void tcp_connect_init(struct sock
*sk
)
2282 struct dst_entry
*dst
= __sk_dst_get(sk
);
2283 struct tcp_sock
*tp
= tcp_sk(sk
);
2286 /* We'll fix this up when we get a response from the other end.
2287 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2289 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2290 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2292 #ifdef CONFIG_TCP_MD5SIG
2293 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2294 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2297 /* If user gave his TCP_MAXSEG, record it to clamp */
2298 if (tp
->rx_opt
.user_mss
)
2299 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2302 tcp_sync_mss(sk
, dst_mtu(dst
));
2304 if (!tp
->window_clamp
)
2305 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2306 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2307 tcp_initialize_rcv_mss(sk
);
2309 tcp_select_initial_window(tcp_full_space(sk
),
2310 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2313 sysctl_tcp_window_scaling
,
2316 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2317 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2320 sock_reset_flag(sk
, SOCK_DONE
);
2322 tcp_init_wl(tp
, tp
->write_seq
, 0);
2323 tp
->snd_una
= tp
->write_seq
;
2324 tp
->snd_sml
= tp
->write_seq
;
2329 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2330 inet_csk(sk
)->icsk_retransmits
= 0;
2331 tcp_clear_retrans(tp
);
2335 * Build a SYN and send it off.
2337 int tcp_connect(struct sock
*sk
)
2339 struct tcp_sock
*tp
= tcp_sk(sk
);
2340 struct sk_buff
*buff
;
2342 tcp_connect_init(sk
);
2344 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2345 if (unlikely(buff
== NULL
))
2348 /* Reserve space for headers. */
2349 skb_reserve(buff
, MAX_TCP_HEADER
);
2351 tp
->snd_nxt
= tp
->write_seq
;
2352 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPCB_FLAG_SYN
);
2353 TCP_ECN_send_syn(sk
, buff
);
2356 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2357 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2358 skb_header_release(buff
);
2359 __tcp_add_write_queue_tail(sk
, buff
);
2360 sk
->sk_wmem_queued
+= buff
->truesize
;
2361 sk_mem_charge(sk
, buff
->truesize
);
2362 tp
->packets_out
+= tcp_skb_pcount(buff
);
2363 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2365 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2366 * in order to make this packet get counted in tcpOutSegs.
2368 tp
->snd_nxt
= tp
->write_seq
;
2369 tp
->pushed_seq
= tp
->write_seq
;
2370 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2372 /* Timer for repeating the SYN until an answer. */
2373 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2374 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2378 /* Send out a delayed ack, the caller does the policy checking
2379 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2382 void tcp_send_delayed_ack(struct sock
*sk
)
2384 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2385 int ato
= icsk
->icsk_ack
.ato
;
2386 unsigned long timeout
;
2388 if (ato
> TCP_DELACK_MIN
) {
2389 const struct tcp_sock
*tp
= tcp_sk(sk
);
2390 int max_ato
= HZ
/ 2;
2392 if (icsk
->icsk_ack
.pingpong
||
2393 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2394 max_ato
= TCP_DELACK_MAX
;
2396 /* Slow path, intersegment interval is "high". */
2398 /* If some rtt estimate is known, use it to bound delayed ack.
2399 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2403 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
2409 ato
= min(ato
, max_ato
);
2412 /* Stay within the limit we were given */
2413 timeout
= jiffies
+ ato
;
2415 /* Use new timeout only if there wasn't a older one earlier. */
2416 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2417 /* If delack timer was blocked or is about to expire,
2420 if (icsk
->icsk_ack
.blocked
||
2421 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2426 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2427 timeout
= icsk
->icsk_ack
.timeout
;
2429 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2430 icsk
->icsk_ack
.timeout
= timeout
;
2431 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2434 /* This routine sends an ack and also updates the window. */
2435 void tcp_send_ack(struct sock
*sk
)
2437 struct sk_buff
*buff
;
2439 /* If we have been reset, we may not send again. */
2440 if (sk
->sk_state
== TCP_CLOSE
)
2443 /* We are not putting this on the write queue, so
2444 * tcp_transmit_skb() will set the ownership to this
2447 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2449 inet_csk_schedule_ack(sk
);
2450 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2451 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2452 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2456 /* Reserve space for headers and prepare control bits. */
2457 skb_reserve(buff
, MAX_TCP_HEADER
);
2458 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPCB_FLAG_ACK
);
2460 /* Send it off, this clears delayed acks for us. */
2461 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2462 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2465 /* This routine sends a packet with an out of date sequence
2466 * number. It assumes the other end will try to ack it.
2468 * Question: what should we make while urgent mode?
2469 * 4.4BSD forces sending single byte of data. We cannot send
2470 * out of window data, because we have SND.NXT==SND.MAX...
2472 * Current solution: to send TWO zero-length segments in urgent mode:
2473 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2474 * out-of-date with SND.UNA-1 to probe window.
2476 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2478 struct tcp_sock
*tp
= tcp_sk(sk
);
2479 struct sk_buff
*skb
;
2481 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2482 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2486 /* Reserve space for headers and set control bits. */
2487 skb_reserve(skb
, MAX_TCP_HEADER
);
2488 /* Use a previous sequence. This should cause the other
2489 * end to send an ack. Don't queue or clone SKB, just
2492 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPCB_FLAG_ACK
);
2493 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2494 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2497 int tcp_write_wakeup(struct sock
*sk
)
2499 struct tcp_sock
*tp
= tcp_sk(sk
);
2500 struct sk_buff
*skb
;
2502 if (sk
->sk_state
== TCP_CLOSE
)
2505 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2506 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
2508 unsigned int mss
= tcp_current_mss(sk
, 0);
2509 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2511 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2512 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2514 /* We are probing the opening of a window
2515 * but the window size is != 0
2516 * must have been a result SWS avoidance ( sender )
2518 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2520 seg_size
= min(seg_size
, mss
);
2521 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2522 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2524 } else if (!tcp_skb_pcount(skb
))
2525 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2527 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2528 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2529 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2531 tcp_event_new_data_sent(sk
, skb
);
2535 between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
2536 tcp_xmit_probe_skb(sk
, 1);
2537 return tcp_xmit_probe_skb(sk
, 0);
2541 /* A window probe timeout has occurred. If window is not closed send
2542 * a partial packet else a zero probe.
2544 void tcp_send_probe0(struct sock
*sk
)
2546 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2547 struct tcp_sock
*tp
= tcp_sk(sk
);
2550 err
= tcp_write_wakeup(sk
);
2552 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2553 /* Cancel probe timer, if it is not required. */
2554 icsk
->icsk_probes_out
= 0;
2555 icsk
->icsk_backoff
= 0;
2560 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2561 icsk
->icsk_backoff
++;
2562 icsk
->icsk_probes_out
++;
2563 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2564 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2567 /* If packet was not sent due to local congestion,
2568 * do not backoff and do not remember icsk_probes_out.
2569 * Let local senders to fight for local resources.
2571 * Use accumulated backoff yet.
2573 if (!icsk
->icsk_probes_out
)
2574 icsk
->icsk_probes_out
= 1;
2575 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2576 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2577 TCP_RESOURCE_PROBE_INTERVAL
),
2582 EXPORT_SYMBOL(tcp_select_initial_window
);
2583 EXPORT_SYMBOL(tcp_connect
);
2584 EXPORT_SYMBOL(tcp_make_synack
);
2585 EXPORT_SYMBOL(tcp_simple_retransmit
);
2586 EXPORT_SYMBOL(tcp_sync_mss
);
2587 EXPORT_SYMBOL(tcp_mtup_init
);