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).
8 * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Changes: Pedro Roque : Retransmit queue handled by TCP.
25 * : Fragmentation on mtu decrease
26 * : Segment collapse on retransmit
29 * Linus Torvalds : send_delayed_ack
30 * David S. Miller : Charge memory using the right skb
31 * during syn/ack processing.
32 * David S. Miller : Output engine completely rewritten.
33 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
34 * Cacophonix Gaul : draft-minshall-nagle-01
35 * J Hadi Salim : ECN support
41 #include <linux/compiler.h>
42 #include <linux/module.h>
44 /* People can turn this off for buggy TCP's found in printers etc. */
45 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
47 /* People can turn this on to work with those rare, broken TCPs that
48 * interpret the window field as a signed quantity.
50 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
52 /* This limits the percentage of the congestion window which we
53 * will allow a single TSO frame to consume. Building TSO frames
54 * which are too large can cause TCP streams to be bursty.
56 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
58 int sysctl_tcp_mtu_probing __read_mostly
= 0;
59 int sysctl_tcp_base_mss __read_mostly
= 512;
61 /* By default, RFC2861 behavior. */
62 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
64 static void update_send_head(struct sock
*sk
, struct sk_buff
*skb
)
66 struct tcp_sock
*tp
= tcp_sk(sk
);
68 tcp_advance_send_head(sk
, skb
);
69 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
70 tcp_packets_out_inc(sk
, skb
);
73 /* SND.NXT, if window was not shrunk.
74 * If window has been shrunk, what should we make? It is not clear at all.
75 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
76 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
77 * invalid. OK, let's make this for now:
79 static inline __u32
tcp_acceptable_seq(struct sock
*sk
)
81 struct tcp_sock
*tp
= tcp_sk(sk
);
83 if (!before(tp
->snd_una
+tp
->snd_wnd
, tp
->snd_nxt
))
86 return tp
->snd_una
+tp
->snd_wnd
;
89 /* Calculate mss to advertise in SYN segment.
90 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
92 * 1. It is independent of path mtu.
93 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
94 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
95 * attached devices, because some buggy hosts are confused by
97 * 4. We do not make 3, we advertise MSS, calculated from first
98 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
99 * This may be overridden via information stored in routing table.
100 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
101 * probably even Jumbo".
103 static __u16
tcp_advertise_mss(struct sock
*sk
)
105 struct tcp_sock
*tp
= tcp_sk(sk
);
106 struct dst_entry
*dst
= __sk_dst_get(sk
);
107 int mss
= tp
->advmss
;
109 if (dst
&& dst_metric(dst
, RTAX_ADVMSS
) < mss
) {
110 mss
= dst_metric(dst
, RTAX_ADVMSS
);
117 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
118 * This is the first part of cwnd validation mechanism. */
119 static void tcp_cwnd_restart(struct sock
*sk
, struct dst_entry
*dst
)
121 struct tcp_sock
*tp
= tcp_sk(sk
);
122 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
123 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
124 u32 cwnd
= tp
->snd_cwnd
;
126 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
128 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
129 restart_cwnd
= min(restart_cwnd
, cwnd
);
131 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
133 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
134 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
135 tp
->snd_cwnd_used
= 0;
138 static void tcp_event_data_sent(struct tcp_sock
*tp
,
139 struct sk_buff
*skb
, struct sock
*sk
)
141 struct inet_connection_sock
*icsk
= inet_csk(sk
);
142 const u32 now
= tcp_time_stamp
;
144 if (sysctl_tcp_slow_start_after_idle
&&
145 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
146 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
150 /* If it is a reply for ato after last received
151 * packet, enter pingpong mode.
153 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
154 icsk
->icsk_ack
.pingpong
= 1;
157 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
159 tcp_dec_quickack_mode(sk
, pkts
);
160 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
163 /* Determine a window scaling and initial window to offer.
164 * Based on the assumption that the given amount of space
165 * will be offered. Store the results in the tp structure.
166 * NOTE: for smooth operation initial space offering should
167 * be a multiple of mss if possible. We assume here that mss >= 1.
168 * This MUST be enforced by all callers.
170 void tcp_select_initial_window(int __space
, __u32 mss
,
171 __u32
*rcv_wnd
, __u32
*window_clamp
,
172 int wscale_ok
, __u8
*rcv_wscale
)
174 unsigned int space
= (__space
< 0 ? 0 : __space
);
176 /* If no clamp set the clamp to the max possible scaled window */
177 if (*window_clamp
== 0)
178 (*window_clamp
) = (65535 << 14);
179 space
= min(*window_clamp
, space
);
181 /* Quantize space offering to a multiple of mss if possible. */
183 space
= (space
/ mss
) * mss
;
185 /* NOTE: offering an initial window larger than 32767
186 * will break some buggy TCP stacks. If the admin tells us
187 * it is likely we could be speaking with such a buggy stack
188 * we will truncate our initial window offering to 32K-1
189 * unless the remote has sent us a window scaling option,
190 * which we interpret as a sign the remote TCP is not
191 * misinterpreting the window field as a signed quantity.
193 if (sysctl_tcp_workaround_signed_windows
)
194 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
200 /* Set window scaling on max possible window
201 * See RFC1323 for an explanation of the limit to 14
203 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
204 space
= min_t(u32
, space
, *window_clamp
);
205 while (space
> 65535 && (*rcv_wscale
) < 14) {
211 /* Set initial window to value enough for senders,
212 * following RFC2414. Senders, not following this RFC,
213 * will be satisfied with 2.
215 if (mss
> (1<<*rcv_wscale
)) {
221 if (*rcv_wnd
> init_cwnd
*mss
)
222 *rcv_wnd
= init_cwnd
*mss
;
225 /* Set the clamp no higher than max representable value */
226 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
229 /* Chose a new window to advertise, update state in tcp_sock for the
230 * socket, and return result with RFC1323 scaling applied. The return
231 * value can be stuffed directly into th->window for an outgoing
234 static u16
tcp_select_window(struct sock
*sk
)
236 struct tcp_sock
*tp
= tcp_sk(sk
);
237 u32 cur_win
= tcp_receive_window(tp
);
238 u32 new_win
= __tcp_select_window(sk
);
240 /* Never shrink the offered window */
241 if (new_win
< cur_win
) {
242 /* Danger Will Robinson!
243 * Don't update rcv_wup/rcv_wnd here or else
244 * we will not be able to advertise a zero
245 * window in time. --DaveM
247 * Relax Will Robinson.
251 tp
->rcv_wnd
= new_win
;
252 tp
->rcv_wup
= tp
->rcv_nxt
;
254 /* Make sure we do not exceed the maximum possible
257 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
258 new_win
= min(new_win
, MAX_TCP_WINDOW
);
260 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
262 /* RFC1323 scaling applied */
263 new_win
>>= tp
->rx_opt
.rcv_wscale
;
265 /* If we advertise zero window, disable fast path. */
272 static inline void TCP_ECN_send_synack(struct tcp_sock
*tp
,
275 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_CWR
;
276 if (!(tp
->ecn_flags
&TCP_ECN_OK
))
277 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_ECE
;
280 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
282 struct tcp_sock
*tp
= tcp_sk(sk
);
285 if (sysctl_tcp_ecn
) {
286 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ECE
|TCPCB_FLAG_CWR
;
287 tp
->ecn_flags
= TCP_ECN_OK
;
291 static __inline__
void
292 TCP_ECN_make_synack(struct request_sock
*req
, struct tcphdr
*th
)
294 if (inet_rsk(req
)->ecn_ok
)
298 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
301 struct tcp_sock
*tp
= tcp_sk(sk
);
303 if (tp
->ecn_flags
& TCP_ECN_OK
) {
304 /* Not-retransmitted data segment: set ECT and inject CWR. */
305 if (skb
->len
!= tcp_header_len
&&
306 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
308 if (tp
->ecn_flags
&TCP_ECN_QUEUE_CWR
) {
309 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
310 tcp_hdr(skb
)->cwr
= 1;
311 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
314 /* ACK or retransmitted segment: clear ECT|CE */
315 INET_ECN_dontxmit(sk
);
317 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
318 tcp_hdr(skb
)->ece
= 1;
322 static void tcp_build_and_update_options(__be32
*ptr
, struct tcp_sock
*tp
,
323 __u32 tstamp
, __u8
**md5_hash
)
325 if (tp
->rx_opt
.tstamp_ok
) {
326 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
328 (TCPOPT_TIMESTAMP
<< 8) |
330 *ptr
++ = htonl(tstamp
);
331 *ptr
++ = htonl(tp
->rx_opt
.ts_recent
);
333 if (tp
->rx_opt
.eff_sacks
) {
334 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
? tp
->duplicate_sack
: tp
->selective_acks
;
337 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
340 (TCPOLEN_SACK_BASE
+ (tp
->rx_opt
.eff_sacks
*
341 TCPOLEN_SACK_PERBLOCK
)));
343 for (this_sack
= 0; this_sack
< tp
->rx_opt
.eff_sacks
; this_sack
++) {
344 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
345 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
348 if (tp
->rx_opt
.dsack
) {
349 tp
->rx_opt
.dsack
= 0;
350 tp
->rx_opt
.eff_sacks
--;
353 #ifdef CONFIG_TCP_MD5SIG
355 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
357 (TCPOPT_MD5SIG
<< 8) |
359 *md5_hash
= (__u8
*)ptr
;
364 /* Construct a tcp options header for a SYN or SYN_ACK packet.
365 * If this is every changed make sure to change the definition of
366 * MAX_SYN_SIZE to match the new maximum number of options that you
369 * Note - that with the RFC2385 TCP option, we make room for the
370 * 16 byte MD5 hash. This will be filled in later, so the pointer for the
371 * location to be filled is passed back up.
373 static void tcp_syn_build_options(__be32
*ptr
, int mss
, int ts
, int sack
,
374 int offer_wscale
, int wscale
, __u32 tstamp
,
375 __u32 ts_recent
, __u8
**md5_hash
)
377 /* We always get an MSS option.
378 * The option bytes which will be seen in normal data
379 * packets should timestamps be used, must be in the MSS
380 * advertised. But we subtract them from tp->mss_cache so
381 * that calculations in tcp_sendmsg are simpler etc.
382 * So account for this fact here if necessary. If we
383 * don't do this correctly, as a receiver we won't
384 * recognize data packets as being full sized when we
385 * should, and thus we won't abide by the delayed ACK
387 * SACKs don't matter, we never delay an ACK when we
388 * have any of those going out.
390 *ptr
++ = htonl((TCPOPT_MSS
<< 24) | (TCPOLEN_MSS
<< 16) | mss
);
393 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
394 (TCPOLEN_SACK_PERM
<< 16) |
395 (TCPOPT_TIMESTAMP
<< 8) |
398 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
400 (TCPOPT_TIMESTAMP
<< 8) |
402 *ptr
++ = htonl(tstamp
); /* TSVAL */
403 *ptr
++ = htonl(ts_recent
); /* TSECR */
405 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
407 (TCPOPT_SACK_PERM
<< 8) |
410 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
411 (TCPOPT_WINDOW
<< 16) |
412 (TCPOLEN_WINDOW
<< 8) |
414 #ifdef CONFIG_TCP_MD5SIG
416 * If MD5 is enabled, then we set the option, and include the size
417 * (always 18). The actual MD5 hash is added just before the
421 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
423 (TCPOPT_MD5SIG
<< 8) |
425 *md5_hash
= (__u8
*) ptr
;
430 /* This routine actually transmits TCP packets queued in by
431 * tcp_do_sendmsg(). This is used by both the initial
432 * transmission and possible later retransmissions.
433 * All SKB's seen here are completely headerless. It is our
434 * job to build the TCP header, and pass the packet down to
435 * IP so it can do the same plus pass the packet off to the
438 * We are working here with either a clone of the original
439 * SKB, or a fresh unique copy made by the retransmit engine.
441 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
, gfp_t gfp_mask
)
443 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
444 struct inet_sock
*inet
;
446 struct tcp_skb_cb
*tcb
;
448 #ifdef CONFIG_TCP_MD5SIG
449 struct tcp_md5sig_key
*md5
;
450 __u8
*md5_hash_location
;
456 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
458 /* If congestion control is doing timestamping, we must
459 * take such a timestamp before we potentially clone/copy.
461 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
462 __net_timestamp(skb
);
464 if (likely(clone_it
)) {
465 if (unlikely(skb_cloned(skb
)))
466 skb
= pskb_copy(skb
, gfp_mask
);
468 skb
= skb_clone(skb
, gfp_mask
);
475 tcb
= TCP_SKB_CB(skb
);
476 tcp_header_size
= tp
->tcp_header_len
;
478 #define SYSCTL_FLAG_TSTAMPS 0x1
479 #define SYSCTL_FLAG_WSCALE 0x2
480 #define SYSCTL_FLAG_SACK 0x4
483 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
484 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
485 if (sysctl_tcp_timestamps
) {
486 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
487 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
489 if (sysctl_tcp_window_scaling
) {
490 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
491 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
493 if (sysctl_tcp_sack
) {
494 sysctl_flags
|= SYSCTL_FLAG_SACK
;
495 if (!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
496 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
498 } else if (unlikely(tp
->rx_opt
.eff_sacks
)) {
499 /* A SACK is 2 pad bytes, a 2 byte header, plus
500 * 2 32-bit sequence numbers for each SACK block.
502 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
503 (tp
->rx_opt
.eff_sacks
*
504 TCPOLEN_SACK_PERBLOCK
));
507 if (tcp_packets_in_flight(tp
) == 0)
508 tcp_ca_event(sk
, CA_EVENT_TX_START
);
510 #ifdef CONFIG_TCP_MD5SIG
512 * Are we doing MD5 on this segment? If so - make
515 md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
517 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
520 skb_push(skb
, tcp_header_size
);
521 skb_reset_transport_header(skb
);
522 skb_set_owner_w(skb
, sk
);
524 /* Build TCP header and checksum it. */
526 th
->source
= inet
->sport
;
527 th
->dest
= inet
->dport
;
528 th
->seq
= htonl(tcb
->seq
);
529 th
->ack_seq
= htonl(tp
->rcv_nxt
);
530 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
533 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
534 /* RFC1323: The window in SYN & SYN/ACK segments
537 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
539 th
->window
= htons(tcp_select_window(sk
));
544 if (unlikely(tp
->urg_mode
&&
545 between(tp
->snd_up
, tcb
->seq
+1, tcb
->seq
+0xFFFF))) {
546 th
->urg_ptr
= htons(tp
->snd_up
-tcb
->seq
);
550 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
551 tcp_syn_build_options((__be32
*)(th
+ 1),
552 tcp_advertise_mss(sk
),
553 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
554 (sysctl_flags
& SYSCTL_FLAG_SACK
),
555 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
556 tp
->rx_opt
.rcv_wscale
,
558 tp
->rx_opt
.ts_recent
,
560 #ifdef CONFIG_TCP_MD5SIG
561 md5
? &md5_hash_location
:
565 tcp_build_and_update_options((__be32
*)(th
+ 1),
567 #ifdef CONFIG_TCP_MD5SIG
568 md5
? &md5_hash_location
:
571 TCP_ECN_send(sk
, skb
, tcp_header_size
);
574 #ifdef CONFIG_TCP_MD5SIG
575 /* Calculate the MD5 hash, as we have all we need now */
577 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
586 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
588 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
589 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
591 if (skb
->len
!= tcp_header_size
)
592 tcp_event_data_sent(tp
, skb
, sk
);
594 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
595 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
597 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
598 if (likely(err
<= 0))
601 tcp_enter_cwr(sk
, 1);
603 return net_xmit_eval(err
);
605 #undef SYSCTL_FLAG_TSTAMPS
606 #undef SYSCTL_FLAG_WSCALE
607 #undef SYSCTL_FLAG_SACK
611 /* This routine just queue's the buffer
613 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
614 * otherwise socket can stall.
616 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
618 struct tcp_sock
*tp
= tcp_sk(sk
);
620 /* Advance write_seq and place onto the write_queue. */
621 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
622 skb_header_release(skb
);
623 tcp_add_write_queue_tail(sk
, skb
);
624 sk_charge_skb(sk
, skb
);
627 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
629 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
)) {
630 /* Avoid the costly divide in the normal
633 skb_shinfo(skb
)->gso_segs
= 1;
634 skb_shinfo(skb
)->gso_size
= 0;
635 skb_shinfo(skb
)->gso_type
= 0;
639 factor
= skb
->len
+ (mss_now
- 1);
641 skb_shinfo(skb
)->gso_segs
= factor
;
642 skb_shinfo(skb
)->gso_size
= mss_now
;
643 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
647 /* Function to create two new TCP segments. Shrinks the given segment
648 * to the specified size and appends a new segment with the rest of the
649 * packet to the list. This won't be called frequently, I hope.
650 * Remember, these are still headerless SKBs at this point.
652 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
, unsigned int mss_now
)
654 struct tcp_sock
*tp
= tcp_sk(sk
);
655 struct sk_buff
*buff
;
656 int nsize
, old_factor
;
660 BUG_ON(len
> skb
->len
);
662 clear_all_retrans_hints(tp
);
663 nsize
= skb_headlen(skb
) - len
;
667 if (skb_cloned(skb
) &&
668 skb_is_nonlinear(skb
) &&
669 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
672 /* Get a new skb... force flag on. */
673 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
675 return -ENOMEM
; /* We'll just try again later. */
677 sk_charge_skb(sk
, buff
);
678 nlen
= skb
->len
- len
- nsize
;
679 buff
->truesize
+= nlen
;
680 skb
->truesize
-= nlen
;
682 /* Correct the sequence numbers. */
683 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
684 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
685 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
687 /* PSH and FIN should only be set in the second packet. */
688 flags
= TCP_SKB_CB(skb
)->flags
;
689 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
690 TCP_SKB_CB(buff
)->flags
= flags
;
691 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
692 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
694 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
695 /* Copy and checksum data tail into the new buffer. */
696 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
701 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
703 skb
->ip_summed
= CHECKSUM_PARTIAL
;
704 skb_split(skb
, buff
, len
);
707 buff
->ip_summed
= skb
->ip_summed
;
709 /* Looks stupid, but our code really uses when of
710 * skbs, which it never sent before. --ANK
712 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
713 buff
->tstamp
= skb
->tstamp
;
715 old_factor
= tcp_skb_pcount(skb
);
717 /* Fix up tso_factor for both original and new SKB. */
718 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
719 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
721 /* If this packet has been sent out already, we must
722 * adjust the various packet counters.
724 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
725 int diff
= old_factor
- tcp_skb_pcount(skb
) -
726 tcp_skb_pcount(buff
);
728 tp
->packets_out
-= diff
;
730 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
731 tp
->sacked_out
-= diff
;
732 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
733 tp
->retrans_out
-= diff
;
735 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
736 tp
->lost_out
-= diff
;
739 /* Adjust Reno SACK estimate. */
740 if (tcp_is_reno(tp
)) {
741 tcp_dec_pcount_approx_int(&tp
->sacked_out
, diff
);
742 tcp_verify_left_out(tp
);
745 tcp_dec_pcount_approx_int(&tp
->fackets_out
, diff
);
746 /* SACK fastpath might overwrite it unless dealt with */
747 if (tp
->fastpath_skb_hint
!= NULL
&&
748 after(TCP_SKB_CB(tp
->fastpath_skb_hint
)->seq
,
749 TCP_SKB_CB(skb
)->seq
)) {
750 tcp_dec_pcount_approx_int(&tp
->fastpath_cnt_hint
, diff
);
755 /* Link BUFF into the send queue. */
756 skb_header_release(buff
);
757 tcp_insert_write_queue_after(skb
, buff
, sk
);
762 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
763 * eventually). The difference is that pulled data not copied, but
764 * immediately discarded.
766 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
772 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
773 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
774 put_page(skb_shinfo(skb
)->frags
[i
].page
);
775 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
777 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
779 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
780 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
786 skb_shinfo(skb
)->nr_frags
= k
;
788 skb_reset_tail_pointer(skb
);
789 skb
->data_len
-= len
;
790 skb
->len
= skb
->data_len
;
793 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
795 if (skb_cloned(skb
) &&
796 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
799 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
800 if (unlikely(len
< skb_headlen(skb
)))
801 __skb_pull(skb
, len
);
803 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
805 TCP_SKB_CB(skb
)->seq
+= len
;
806 skb
->ip_summed
= CHECKSUM_PARTIAL
;
808 skb
->truesize
-= len
;
809 sk
->sk_wmem_queued
-= len
;
810 sk
->sk_forward_alloc
+= len
;
811 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
813 /* Any change of skb->len requires recalculation of tso
816 if (tcp_skb_pcount(skb
) > 1)
817 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
822 /* Not accounting for SACKs here. */
823 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
825 struct tcp_sock
*tp
= tcp_sk(sk
);
826 struct inet_connection_sock
*icsk
= inet_csk(sk
);
829 /* Calculate base mss without TCP options:
830 It is MMS_S - sizeof(tcphdr) of rfc1122
832 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
834 /* Clamp it (mss_clamp does not include tcp options) */
835 if (mss_now
> tp
->rx_opt
.mss_clamp
)
836 mss_now
= tp
->rx_opt
.mss_clamp
;
838 /* Now subtract optional transport overhead */
839 mss_now
-= icsk
->icsk_ext_hdr_len
;
841 /* Then reserve room for full set of TCP options and 8 bytes of data */
845 /* Now subtract TCP options size, not including SACKs */
846 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
851 /* Inverse of above */
852 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
854 struct tcp_sock
*tp
= tcp_sk(sk
);
855 struct inet_connection_sock
*icsk
= inet_csk(sk
);
860 icsk
->icsk_ext_hdr_len
+
861 icsk
->icsk_af_ops
->net_header_len
;
866 void tcp_mtup_init(struct sock
*sk
)
868 struct tcp_sock
*tp
= tcp_sk(sk
);
869 struct inet_connection_sock
*icsk
= inet_csk(sk
);
871 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
872 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
873 icsk
->icsk_af_ops
->net_header_len
;
874 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
875 icsk
->icsk_mtup
.probe_size
= 0;
878 /* This function synchronize snd mss to current pmtu/exthdr set.
880 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
881 for TCP options, but includes only bare TCP header.
883 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
884 It is minimum of user_mss and mss received with SYN.
885 It also does not include TCP options.
887 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
889 tp->mss_cache is current effective sending mss, including
890 all tcp options except for SACKs. It is evaluated,
891 taking into account current pmtu, but never exceeds
892 tp->rx_opt.mss_clamp.
894 NOTE1. rfc1122 clearly states that advertised MSS
895 DOES NOT include either tcp or ip options.
897 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
898 are READ ONLY outside this function. --ANK (980731)
901 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
903 struct tcp_sock
*tp
= tcp_sk(sk
);
904 struct inet_connection_sock
*icsk
= inet_csk(sk
);
907 if (icsk
->icsk_mtup
.search_high
> pmtu
)
908 icsk
->icsk_mtup
.search_high
= pmtu
;
910 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
912 /* Bound mss with half of window */
913 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
914 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
916 /* And store cached results */
917 icsk
->icsk_pmtu_cookie
= pmtu
;
918 if (icsk
->icsk_mtup
.enabled
)
919 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
920 tp
->mss_cache
= mss_now
;
925 /* Compute the current effective MSS, taking SACKs and IP options,
926 * and even PMTU discovery events into account.
928 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
929 * cannot be large. However, taking into account rare use of URG, this
932 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
934 struct tcp_sock
*tp
= tcp_sk(sk
);
935 struct dst_entry
*dst
= __sk_dst_get(sk
);
940 mss_now
= tp
->mss_cache
;
942 if (large_allowed
&& sk_can_gso(sk
) && !tp
->urg_mode
)
946 u32 mtu
= dst_mtu(dst
);
947 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
948 mss_now
= tcp_sync_mss(sk
, mtu
);
951 if (tp
->rx_opt
.eff_sacks
)
952 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
953 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
955 #ifdef CONFIG_TCP_MD5SIG
956 if (tp
->af_specific
->md5_lookup(sk
, sk
))
957 mss_now
-= TCPOLEN_MD5SIG_ALIGNED
;
960 xmit_size_goal
= mss_now
;
963 xmit_size_goal
= (65535 -
964 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
965 inet_csk(sk
)->icsk_ext_hdr_len
-
968 if (tp
->max_window
&&
969 (xmit_size_goal
> (tp
->max_window
>> 1)))
970 xmit_size_goal
= max((tp
->max_window
>> 1),
971 68U - tp
->tcp_header_len
);
973 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
975 tp
->xmit_size_goal
= xmit_size_goal
;
980 /* Congestion window validation. (RFC2861) */
982 static void tcp_cwnd_validate(struct sock
*sk
)
984 struct tcp_sock
*tp
= tcp_sk(sk
);
985 __u32 packets_out
= tp
->packets_out
;
987 if (packets_out
>= tp
->snd_cwnd
) {
988 /* Network is feed fully. */
989 tp
->snd_cwnd_used
= 0;
990 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
992 /* Network starves. */
993 if (tp
->packets_out
> tp
->snd_cwnd_used
)
994 tp
->snd_cwnd_used
= tp
->packets_out
;
996 if (sysctl_tcp_slow_start_after_idle
&&
997 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
998 tcp_cwnd_application_limited(sk
);
1002 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
1004 u32 window
, cwnd_len
;
1006 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
1007 cwnd_len
= mss_now
* cwnd
;
1008 return min(window
, cwnd_len
);
1011 /* Can at least one segment of SKB be sent right now, according to the
1012 * congestion window rules? If so, return how many segments are allowed.
1014 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
1016 u32 in_flight
, cwnd
;
1018 /* Don't be strict about the congestion window for the final FIN. */
1019 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1020 tcp_skb_pcount(skb
) == 1)
1023 in_flight
= tcp_packets_in_flight(tp
);
1024 cwnd
= tp
->snd_cwnd
;
1025 if (in_flight
< cwnd
)
1026 return (cwnd
- in_flight
);
1031 /* This must be invoked the first time we consider transmitting
1032 * SKB onto the wire.
1034 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
1036 int tso_segs
= tcp_skb_pcount(skb
);
1040 tcp_skb_mss(skb
) != mss_now
)) {
1041 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1042 tso_segs
= tcp_skb_pcount(skb
);
1047 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1049 return after(tp
->snd_sml
,tp
->snd_una
) &&
1050 !after(tp
->snd_sml
, tp
->snd_nxt
);
1053 /* Return 0, if packet can be sent now without violation Nagle's rules:
1054 * 1. It is full sized.
1055 * 2. Or it contains FIN. (already checked by caller)
1056 * 3. Or TCP_NODELAY was set.
1057 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1058 * With Minshall's modification: all sent small packets are ACKed.
1061 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1062 const struct sk_buff
*skb
,
1063 unsigned mss_now
, int nonagle
)
1065 return (skb
->len
< mss_now
&&
1066 ((nonagle
&TCP_NAGLE_CORK
) ||
1069 tcp_minshall_check(tp
))));
1072 /* Return non-zero if the Nagle test allows this packet to be
1075 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1076 unsigned int cur_mss
, int nonagle
)
1078 /* Nagle rule does not apply to frames, which sit in the middle of the
1079 * write_queue (they have no chances to get new data).
1081 * This is implemented in the callers, where they modify the 'nonagle'
1082 * argument based upon the location of SKB in the send queue.
1084 if (nonagle
& TCP_NAGLE_PUSH
)
1087 /* Don't use the nagle rule for urgent data (or for the final FIN).
1088 * Nagle can be ignored during F-RTO too (see RFC4138).
1090 if (tp
->urg_mode
|| (tp
->frto_counter
== 2) ||
1091 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1094 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1100 /* Does at least the first segment of SKB fit into the send window? */
1101 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
1103 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1105 if (skb
->len
> cur_mss
)
1106 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1108 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
1111 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1112 * should be put on the wire right now. If so, it returns the number of
1113 * packets allowed by the congestion window.
1115 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1116 unsigned int cur_mss
, int nonagle
)
1118 struct tcp_sock
*tp
= tcp_sk(sk
);
1119 unsigned int cwnd_quota
;
1121 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1123 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1126 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1128 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1134 int tcp_may_send_now(struct sock
*sk
)
1136 struct tcp_sock
*tp
= tcp_sk(sk
);
1137 struct sk_buff
*skb
= tcp_send_head(sk
);
1140 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1141 (tcp_skb_is_last(sk
, skb
) ?
1146 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1147 * which is put after SKB on the list. It is very much like
1148 * tcp_fragment() except that it may make several kinds of assumptions
1149 * in order to speed up the splitting operation. In particular, we
1150 * know that all the data is in scatter-gather pages, and that the
1151 * packet has never been sent out before (and thus is not cloned).
1153 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
1155 struct sk_buff
*buff
;
1156 int nlen
= skb
->len
- len
;
1159 /* All of a TSO frame must be composed of paged data. */
1160 if (skb
->len
!= skb
->data_len
)
1161 return tcp_fragment(sk
, skb
, len
, mss_now
);
1163 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
1164 if (unlikely(buff
== NULL
))
1167 sk_charge_skb(sk
, buff
);
1168 buff
->truesize
+= nlen
;
1169 skb
->truesize
-= nlen
;
1171 /* Correct the sequence numbers. */
1172 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1173 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1174 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1176 /* PSH and FIN should only be set in the second packet. */
1177 flags
= TCP_SKB_CB(skb
)->flags
;
1178 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1179 TCP_SKB_CB(buff
)->flags
= flags
;
1181 /* This packet was never sent out yet, so no SACK bits. */
1182 TCP_SKB_CB(buff
)->sacked
= 0;
1184 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1185 skb_split(skb
, buff
, len
);
1187 /* Fix up tso_factor for both original and new SKB. */
1188 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1189 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1191 /* Link BUFF into the send queue. */
1192 skb_header_release(buff
);
1193 tcp_insert_write_queue_after(skb
, buff
, sk
);
1198 /* Try to defer sending, if possible, in order to minimize the amount
1199 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1201 * This algorithm is from John Heffner.
1203 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1205 struct tcp_sock
*tp
= tcp_sk(sk
);
1206 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1207 u32 send_win
, cong_win
, limit
, in_flight
;
1209 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1212 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1215 /* Defer for less than two clock ticks. */
1216 if (!tp
->tso_deferred
&& ((jiffies
<<1)>>1) - (tp
->tso_deferred
>>1) > 1)
1219 in_flight
= tcp_packets_in_flight(tp
);
1221 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
1222 (tp
->snd_cwnd
<= in_flight
));
1224 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
1226 /* From in_flight test above, we know that cwnd > in_flight. */
1227 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1229 limit
= min(send_win
, cong_win
);
1231 /* If a full-sized TSO skb can be sent, do it. */
1235 if (sysctl_tcp_tso_win_divisor
) {
1236 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1238 /* If at least some fraction of a window is available,
1241 chunk
/= sysctl_tcp_tso_win_divisor
;
1245 /* Different approach, try not to defer past a single
1246 * ACK. Receiver should ACK every other full sized
1247 * frame, so if we have space for more than 3 frames
1250 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1254 /* Ok, it looks like it is advisable to defer. */
1255 tp
->tso_deferred
= 1 | (jiffies
<<1);
1260 tp
->tso_deferred
= 0;
1264 /* Create a new MTU probe if we are ready.
1265 * Returns 0 if we should wait to probe (no cwnd available),
1266 * 1 if a probe was sent,
1268 static int tcp_mtu_probe(struct sock
*sk
)
1270 struct tcp_sock
*tp
= tcp_sk(sk
);
1271 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1272 struct sk_buff
*skb
, *nskb
, *next
;
1279 /* Not currently probing/verifying,
1281 * have enough cwnd, and
1282 * not SACKing (the variable headers throw things off) */
1283 if (!icsk
->icsk_mtup
.enabled
||
1284 icsk
->icsk_mtup
.probe_size
||
1285 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1286 tp
->snd_cwnd
< 11 ||
1287 tp
->rx_opt
.eff_sacks
)
1290 /* Very simple search strategy: just double the MSS. */
1291 mss_now
= tcp_current_mss(sk
, 0);
1292 probe_size
= 2*tp
->mss_cache
;
1293 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1294 /* TODO: set timer for probe_converge_event */
1298 /* Have enough data in the send queue to probe? */
1300 if ((skb
= tcp_send_head(sk
)) == NULL
)
1302 while ((len
+= skb
->len
) < probe_size
&& !tcp_skb_is_last(sk
, skb
))
1303 skb
= tcp_write_queue_next(sk
, skb
);
1304 if (len
< probe_size
)
1307 /* Receive window check. */
1308 if (after(TCP_SKB_CB(skb
)->seq
+ probe_size
, tp
->snd_una
+ tp
->snd_wnd
)) {
1309 if (tp
->snd_wnd
< probe_size
)
1315 /* Do we need to wait to drain cwnd? */
1316 pif
= tcp_packets_in_flight(tp
);
1317 if (pif
+ 2 > tp
->snd_cwnd
) {
1318 /* With no packets in flight, don't stall. */
1325 /* We're allowed to probe. Build it now. */
1326 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1328 sk_charge_skb(sk
, nskb
);
1330 skb
= tcp_send_head(sk
);
1331 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1332 tcp_advance_send_head(sk
, skb
);
1334 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1335 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1336 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1337 TCP_SKB_CB(nskb
)->sacked
= 0;
1339 nskb
->ip_summed
= skb
->ip_summed
;
1342 while (len
< probe_size
) {
1343 next
= tcp_write_queue_next(sk
, skb
);
1345 copy
= min_t(int, skb
->len
, probe_size
- len
);
1346 if (nskb
->ip_summed
)
1347 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1349 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1350 skb_put(nskb
, copy
), copy
, nskb
->csum
);
1352 if (skb
->len
<= copy
) {
1353 /* We've eaten all the data from this skb.
1355 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1356 tcp_unlink_write_queue(skb
, sk
);
1357 sk_stream_free_skb(sk
, skb
);
1359 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1360 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1361 if (!skb_shinfo(skb
)->nr_frags
) {
1362 skb_pull(skb
, copy
);
1363 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1364 skb
->csum
= csum_partial(skb
->data
, skb
->len
, 0);
1366 __pskb_trim_head(skb
, copy
);
1367 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1369 TCP_SKB_CB(skb
)->seq
+= copy
;
1375 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1377 /* We're ready to send. If this fails, the probe will
1378 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1379 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1380 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1381 /* Decrement cwnd here because we are sending
1382 * effectively two packets. */
1384 update_send_head(sk
, nskb
);
1386 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1387 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1388 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1397 /* This routine writes packets to the network. It advances the
1398 * send_head. This happens as incoming acks open up the remote
1401 * Returns 1, if no segments are in flight and we have queued segments, but
1402 * cannot send anything now because of SWS or another problem.
1404 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1406 struct tcp_sock
*tp
= tcp_sk(sk
);
1407 struct sk_buff
*skb
;
1408 unsigned int tso_segs
, sent_pkts
;
1412 /* If we are closed, the bytes will have to remain here.
1413 * In time closedown will finish, we empty the write queue and all
1416 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1421 /* Do MTU probing. */
1422 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1424 } else if (result
> 0) {
1428 while ((skb
= tcp_send_head(sk
))) {
1431 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1434 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1438 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1441 if (tso_segs
== 1) {
1442 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1443 (tcp_skb_is_last(sk
, skb
) ?
1444 nonagle
: TCP_NAGLE_PUSH
))))
1447 if (tcp_tso_should_defer(sk
, skb
))
1453 limit
= tcp_window_allows(tp
, skb
,
1454 mss_now
, cwnd_quota
);
1456 if (skb
->len
< limit
) {
1457 unsigned int trim
= skb
->len
% mss_now
;
1460 limit
= skb
->len
- trim
;
1464 if (skb
->len
> limit
&&
1465 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1468 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1470 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1473 /* Advance the send_head. This one is sent out.
1474 * This call will increment packets_out.
1476 update_send_head(sk
, skb
);
1478 tcp_minshall_update(tp
, mss_now
, skb
);
1482 if (likely(sent_pkts
)) {
1483 tcp_cwnd_validate(sk
);
1486 return !tp
->packets_out
&& tcp_send_head(sk
);
1489 /* Push out any pending frames which were held back due to
1490 * TCP_CORK or attempt at coalescing tiny packets.
1491 * The socket must be locked by the caller.
1493 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1496 struct sk_buff
*skb
= tcp_send_head(sk
);
1499 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1500 tcp_check_probe_timer(sk
);
1504 /* Send _single_ skb sitting at the send head. This function requires
1505 * true push pending frames to setup probe timer etc.
1507 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1509 struct tcp_sock
*tp
= tcp_sk(sk
);
1510 struct sk_buff
*skb
= tcp_send_head(sk
);
1511 unsigned int tso_segs
, cwnd_quota
;
1513 BUG_ON(!skb
|| skb
->len
< mss_now
);
1515 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1516 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1518 if (likely(cwnd_quota
)) {
1525 limit
= tcp_window_allows(tp
, skb
,
1526 mss_now
, cwnd_quota
);
1528 if (skb
->len
< limit
) {
1529 unsigned int trim
= skb
->len
% mss_now
;
1532 limit
= skb
->len
- trim
;
1536 if (skb
->len
> limit
&&
1537 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1540 /* Send it out now. */
1541 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1543 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1544 update_send_head(sk
, skb
);
1545 tcp_cwnd_validate(sk
);
1551 /* This function returns the amount that we can raise the
1552 * usable window based on the following constraints
1554 * 1. The window can never be shrunk once it is offered (RFC 793)
1555 * 2. We limit memory per socket
1558 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1559 * RECV.NEXT + RCV.WIN fixed until:
1560 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1562 * i.e. don't raise the right edge of the window until you can raise
1563 * it at least MSS bytes.
1565 * Unfortunately, the recommended algorithm breaks header prediction,
1566 * since header prediction assumes th->window stays fixed.
1568 * Strictly speaking, keeping th->window fixed violates the receiver
1569 * side SWS prevention criteria. The problem is that under this rule
1570 * a stream of single byte packets will cause the right side of the
1571 * window to always advance by a single byte.
1573 * Of course, if the sender implements sender side SWS prevention
1574 * then this will not be a problem.
1576 * BSD seems to make the following compromise:
1578 * If the free space is less than the 1/4 of the maximum
1579 * space available and the free space is less than 1/2 mss,
1580 * then set the window to 0.
1581 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1582 * Otherwise, just prevent the window from shrinking
1583 * and from being larger than the largest representable value.
1585 * This prevents incremental opening of the window in the regime
1586 * where TCP is limited by the speed of the reader side taking
1587 * data out of the TCP receive queue. It does nothing about
1588 * those cases where the window is constrained on the sender side
1589 * because the pipeline is full.
1591 * BSD also seems to "accidentally" limit itself to windows that are a
1592 * multiple of MSS, at least until the free space gets quite small.
1593 * This would appear to be a side effect of the mbuf implementation.
1594 * Combining these two algorithms results in the observed behavior
1595 * of having a fixed window size at almost all times.
1597 * Below we obtain similar behavior by forcing the offered window to
1598 * a multiple of the mss when it is feasible to do so.
1600 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1601 * Regular options like TIMESTAMP are taken into account.
1603 u32
__tcp_select_window(struct sock
*sk
)
1605 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1606 struct tcp_sock
*tp
= tcp_sk(sk
);
1607 /* MSS for the peer's data. Previous versions used mss_clamp
1608 * here. I don't know if the value based on our guesses
1609 * of peer's MSS is better for the performance. It's more correct
1610 * but may be worse for the performance because of rcv_mss
1611 * fluctuations. --SAW 1998/11/1
1613 int mss
= icsk
->icsk_ack
.rcv_mss
;
1614 int free_space
= tcp_space(sk
);
1615 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1618 if (mss
> full_space
)
1621 if (free_space
< full_space
/2) {
1622 icsk
->icsk_ack
.quick
= 0;
1624 if (tcp_memory_pressure
)
1625 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1627 if (free_space
< mss
)
1631 if (free_space
> tp
->rcv_ssthresh
)
1632 free_space
= tp
->rcv_ssthresh
;
1634 /* Don't do rounding if we are using window scaling, since the
1635 * scaled window will not line up with the MSS boundary anyway.
1637 window
= tp
->rcv_wnd
;
1638 if (tp
->rx_opt
.rcv_wscale
) {
1639 window
= free_space
;
1641 /* Advertise enough space so that it won't get scaled away.
1642 * Import case: prevent zero window announcement if
1643 * 1<<rcv_wscale > mss.
1645 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1646 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1647 << tp
->rx_opt
.rcv_wscale
);
1649 /* Get the largest window that is a nice multiple of mss.
1650 * Window clamp already applied above.
1651 * If our current window offering is within 1 mss of the
1652 * free space we just keep it. This prevents the divide
1653 * and multiply from happening most of the time.
1654 * We also don't do any window rounding when the free space
1657 if (window
<= free_space
- mss
|| window
> free_space
)
1658 window
= (free_space
/mss
)*mss
;
1659 else if (mss
== full_space
&&
1660 free_space
> window
+ full_space
/2)
1661 window
= free_space
;
1667 /* Attempt to collapse two adjacent SKB's during retransmission. */
1668 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1670 struct tcp_sock
*tp
= tcp_sk(sk
);
1671 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1673 /* The first test we must make is that neither of these two
1674 * SKB's are still referenced by someone else.
1676 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1677 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1678 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1680 /* Also punt if next skb has been SACK'd. */
1681 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1684 /* Next skb is out of window. */
1685 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1688 /* Punt if not enough space exists in the first SKB for
1689 * the data in the second, or the total combined payload
1690 * would exceed the MSS.
1692 if ((next_skb_size
> skb_tailroom(skb
)) ||
1693 ((skb_size
+ next_skb_size
) > mss_now
))
1696 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1697 tcp_skb_pcount(next_skb
) != 1);
1699 /* changing transmit queue under us so clear hints */
1700 clear_all_retrans_hints(tp
);
1702 /* Ok. We will be able to collapse the packet. */
1703 tcp_unlink_write_queue(next_skb
, sk
);
1705 skb_copy_from_linear_data(next_skb
,
1706 skb_put(skb
, next_skb_size
),
1709 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1710 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1712 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1713 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1715 /* Update sequence range on original skb. */
1716 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1718 /* Merge over control information. */
1719 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1720 TCP_SKB_CB(skb
)->flags
= flags
;
1722 /* All done, get rid of second SKB and account for it so
1723 * packet counting does not break.
1725 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1726 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1727 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1728 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
)
1729 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1730 /* Reno case is special. Sigh... */
1731 if (tcp_is_reno(tp
) && tp
->sacked_out
)
1732 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1734 /* Not quite right: it can be > snd.fack, but
1735 * it is better to underestimate fackets.
1737 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1738 tp
->packets_out
-= tcp_skb_pcount(next_skb
);
1739 sk_stream_free_skb(sk
, next_skb
);
1743 /* Do a simple retransmit without using the backoff mechanisms in
1744 * tcp_timer. This is used for path mtu discovery.
1745 * The socket is already locked here.
1747 void tcp_simple_retransmit(struct sock
*sk
)
1749 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1750 struct tcp_sock
*tp
= tcp_sk(sk
);
1751 struct sk_buff
*skb
;
1752 unsigned int mss
= tcp_current_mss(sk
, 0);
1755 tcp_for_write_queue(skb
, sk
) {
1756 if (skb
== tcp_send_head(sk
))
1758 if (skb
->len
> mss
&&
1759 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1760 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1761 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1762 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1764 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1765 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1766 tp
->lost_out
+= tcp_skb_pcount(skb
);
1772 clear_all_retrans_hints(tp
);
1777 tcp_verify_left_out(tp
);
1779 /* Don't muck with the congestion window here.
1780 * Reason is that we do not increase amount of _data_
1781 * in network, but units changed and effective
1782 * cwnd/ssthresh really reduced now.
1784 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1785 tp
->high_seq
= tp
->snd_nxt
;
1786 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1787 tp
->prior_ssthresh
= 0;
1788 tp
->undo_marker
= 0;
1789 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1791 tcp_xmit_retransmit_queue(sk
);
1794 /* This retransmits one SKB. Policy decisions and retransmit queue
1795 * state updates are done by the caller. Returns non-zero if an
1796 * error occurred which prevented the send.
1798 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1800 struct tcp_sock
*tp
= tcp_sk(sk
);
1801 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1802 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1805 /* Inconslusive MTU probe */
1806 if (icsk
->icsk_mtup
.probe_size
) {
1807 icsk
->icsk_mtup
.probe_size
= 0;
1810 /* Do not sent more than we queued. 1/4 is reserved for possible
1811 * copying overhead: fragmentation, tunneling, mangling etc.
1813 if (atomic_read(&sk
->sk_wmem_alloc
) >
1814 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1817 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1818 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1820 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1824 /* If receiver has shrunk his window, and skb is out of
1825 * new window, do not retransmit it. The exception is the
1826 * case, when window is shrunk to zero. In this case
1827 * our retransmit serves as a zero window probe.
1829 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1830 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1833 if (skb
->len
> cur_mss
) {
1834 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1835 return -ENOMEM
; /* We'll try again later. */
1838 /* Collapse two adjacent packets if worthwhile and we can. */
1839 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1840 (skb
->len
< (cur_mss
>> 1)) &&
1841 (tcp_write_queue_next(sk
, skb
) != tcp_send_head(sk
)) &&
1842 (!tcp_skb_is_last(sk
, skb
)) &&
1843 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(tcp_write_queue_next(sk
, skb
))->nr_frags
== 0) &&
1844 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk
, skb
)) == 1) &&
1845 (sysctl_tcp_retrans_collapse
!= 0))
1846 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1848 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1849 return -EHOSTUNREACH
; /* Routing failure or similar. */
1851 /* Some Solaris stacks overoptimize and ignore the FIN on a
1852 * retransmit when old data is attached. So strip it off
1853 * since it is cheap to do so and saves bytes on the network.
1856 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1857 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1858 if (!pskb_trim(skb
, 0)) {
1859 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1860 skb_shinfo(skb
)->gso_segs
= 1;
1861 skb_shinfo(skb
)->gso_size
= 0;
1862 skb_shinfo(skb
)->gso_type
= 0;
1863 skb
->ip_summed
= CHECKSUM_NONE
;
1868 /* Make a copy, if the first transmission SKB clone we made
1869 * is still in somebody's hands, else make a clone.
1871 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1873 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1876 /* Update global TCP statistics. */
1877 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1879 tp
->total_retrans
++;
1881 #if FASTRETRANS_DEBUG > 0
1882 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1883 if (net_ratelimit())
1884 printk(KERN_DEBUG
"retrans_out leaked.\n");
1887 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1888 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1890 /* Save stamp of the first retransmit. */
1891 if (!tp
->retrans_stamp
)
1892 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1896 /* snd_nxt is stored to detect loss of retransmitted segment,
1897 * see tcp_input.c tcp_sacktag_write_queue().
1899 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1904 /* This gets called after a retransmit timeout, and the initially
1905 * retransmitted data is acknowledged. It tries to continue
1906 * resending the rest of the retransmit queue, until either
1907 * we've sent it all or the congestion window limit is reached.
1908 * If doing SACK, the first ACK which comes back for a timeout
1909 * based retransmit packet might feed us FACK information again.
1910 * If so, we use it to avoid unnecessarily retransmissions.
1912 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1914 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1915 struct tcp_sock
*tp
= tcp_sk(sk
);
1916 struct sk_buff
*skb
;
1919 if (tp
->retransmit_skb_hint
) {
1920 skb
= tp
->retransmit_skb_hint
;
1921 packet_cnt
= tp
->retransmit_cnt_hint
;
1923 skb
= tcp_write_queue_head(sk
);
1927 /* First pass: retransmit lost packets. */
1929 tcp_for_write_queue_from(skb
, sk
) {
1930 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1932 if (skb
== tcp_send_head(sk
))
1934 /* we could do better than to assign each time */
1935 tp
->retransmit_skb_hint
= skb
;
1936 tp
->retransmit_cnt_hint
= packet_cnt
;
1938 /* Assume this retransmit will generate
1939 * only one packet for congestion window
1940 * calculation purposes. This works because
1941 * tcp_retransmit_skb() will chop up the
1942 * packet to be MSS sized and all the
1943 * packet counting works out.
1945 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1948 if (sacked
& TCPCB_LOST
) {
1949 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1950 if (tcp_retransmit_skb(sk
, skb
)) {
1951 tp
->retransmit_skb_hint
= NULL
;
1954 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1955 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1957 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1959 if (skb
== tcp_write_queue_head(sk
))
1960 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1961 inet_csk(sk
)->icsk_rto
,
1965 packet_cnt
+= tcp_skb_pcount(skb
);
1966 if (packet_cnt
>= tp
->lost_out
)
1972 /* OK, demanded retransmission is finished. */
1974 /* Forward retransmissions are possible only during Recovery. */
1975 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1978 /* No forward retransmissions in Reno are possible. */
1979 if (tcp_is_reno(tp
))
1982 /* Yeah, we have to make difficult choice between forward transmission
1983 * and retransmission... Both ways have their merits...
1985 * For now we do not retransmit anything, while we have some new
1986 * segments to send. In the other cases, follow rule 3 for
1987 * NextSeg() specified in RFC3517.
1990 if (tcp_may_send_now(sk
))
1993 /* If nothing is SACKed, highest_sack in the loop won't be valid */
1994 if (!tp
->sacked_out
)
1997 if (tp
->forward_skb_hint
)
1998 skb
= tp
->forward_skb_hint
;
2000 skb
= tcp_write_queue_head(sk
);
2002 tcp_for_write_queue_from(skb
, sk
) {
2003 if (skb
== tcp_send_head(sk
))
2005 tp
->forward_skb_hint
= skb
;
2007 if (after(TCP_SKB_CB(skb
)->seq
, tp
->highest_sack
))
2010 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2013 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
2016 /* Ok, retransmit it. */
2017 if (tcp_retransmit_skb(sk
, skb
)) {
2018 tp
->forward_skb_hint
= NULL
;
2022 if (skb
== tcp_write_queue_head(sk
))
2023 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2024 inet_csk(sk
)->icsk_rto
,
2027 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
2032 /* Send a fin. The caller locks the socket for us. This cannot be
2033 * allowed to fail queueing a FIN frame under any circumstances.
2035 void tcp_send_fin(struct sock
*sk
)
2037 struct tcp_sock
*tp
= tcp_sk(sk
);
2038 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2041 /* Optimization, tack on the FIN if we have a queue of
2042 * unsent frames. But be careful about outgoing SACKS
2045 mss_now
= tcp_current_mss(sk
, 1);
2047 if (tcp_send_head(sk
) != NULL
) {
2048 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2049 TCP_SKB_CB(skb
)->end_seq
++;
2052 /* Socket is locked, keep trying until memory is available. */
2054 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
2060 /* Reserve space for headers and prepare control bits. */
2061 skb_reserve(skb
, MAX_TCP_HEADER
);
2063 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2064 TCP_SKB_CB(skb
)->sacked
= 0;
2065 skb_shinfo(skb
)->gso_segs
= 1;
2066 skb_shinfo(skb
)->gso_size
= 0;
2067 skb_shinfo(skb
)->gso_type
= 0;
2069 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2070 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
2071 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2072 tcp_queue_skb(sk
, skb
);
2074 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2077 /* We get here when a process closes a file descriptor (either due to
2078 * an explicit close() or as a byproduct of exit()'ing) and there
2079 * was unread data in the receive queue. This behavior is recommended
2080 * by RFC 2525, section 2.17. -DaveM
2082 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2084 struct sk_buff
*skb
;
2086 /* NOTE: No TCP options attached and we never retransmit this. */
2087 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2089 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2093 /* Reserve space for headers and prepare control bits. */
2094 skb_reserve(skb
, MAX_TCP_HEADER
);
2096 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2097 TCP_SKB_CB(skb
)->sacked
= 0;
2098 skb_shinfo(skb
)->gso_segs
= 1;
2099 skb_shinfo(skb
)->gso_size
= 0;
2100 skb_shinfo(skb
)->gso_type
= 0;
2103 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
);
2104 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2105 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2106 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2107 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2110 /* WARNING: This routine must only be called when we have already sent
2111 * a SYN packet that crossed the incoming SYN that caused this routine
2112 * to get called. If this assumption fails then the initial rcv_wnd
2113 * and rcv_wscale values will not be correct.
2115 int tcp_send_synack(struct sock
*sk
)
2117 struct sk_buff
* skb
;
2119 skb
= tcp_write_queue_head(sk
);
2120 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
2121 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2124 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
2125 if (skb_cloned(skb
)) {
2126 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2129 tcp_unlink_write_queue(skb
, sk
);
2130 skb_header_release(nskb
);
2131 __tcp_add_write_queue_head(sk
, nskb
);
2132 sk_stream_free_skb(sk
, skb
);
2133 sk_charge_skb(sk
, nskb
);
2137 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2138 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2140 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2141 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2145 * Prepare a SYN-ACK.
2147 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2148 struct request_sock
*req
)
2150 struct inet_request_sock
*ireq
= inet_rsk(req
);
2151 struct tcp_sock
*tp
= tcp_sk(sk
);
2153 int tcp_header_size
;
2154 struct sk_buff
*skb
;
2155 #ifdef CONFIG_TCP_MD5SIG
2156 struct tcp_md5sig_key
*md5
;
2157 __u8
*md5_hash_location
;
2160 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2164 /* Reserve space for headers. */
2165 skb_reserve(skb
, MAX_TCP_HEADER
);
2167 skb
->dst
= dst_clone(dst
);
2169 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2170 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2171 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2172 /* SACK_PERM is in the place of NOP NOP of TS */
2173 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2175 #ifdef CONFIG_TCP_MD5SIG
2176 /* Are we doing MD5 on this segment? If so - make room for it */
2177 md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
2179 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
2181 skb_push(skb
, tcp_header_size
);
2182 skb_reset_transport_header(skb
);
2185 memset(th
, 0, sizeof(struct tcphdr
));
2188 TCP_ECN_make_synack(req
, th
);
2189 th
->source
= inet_sk(sk
)->sport
;
2190 th
->dest
= ireq
->rmt_port
;
2191 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
2192 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2193 TCP_SKB_CB(skb
)->sacked
= 0;
2194 skb_shinfo(skb
)->gso_segs
= 1;
2195 skb_shinfo(skb
)->gso_size
= 0;
2196 skb_shinfo(skb
)->gso_type
= 0;
2197 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2198 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2199 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2201 /* Set this up on the first call only */
2202 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2203 /* tcp_full_space because it is guaranteed to be the first packet */
2204 tcp_select_initial_window(tcp_full_space(sk
),
2205 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2210 ireq
->rcv_wscale
= rcv_wscale
;
2213 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2214 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2216 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2217 tcp_syn_build_options((__be32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2218 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2219 TCP_SKB_CB(skb
)->when
,
2222 #ifdef CONFIG_TCP_MD5SIG
2223 md5
? &md5_hash_location
:
2229 th
->doff
= (tcp_header_size
>> 2);
2230 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
2232 #ifdef CONFIG_TCP_MD5SIG
2233 /* Okay, we have all we need - do the md5 hash if needed */
2235 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
2238 tcp_hdr(skb
), sk
->sk_protocol
,
2247 * Do all connect socket setups that can be done AF independent.
2249 static void tcp_connect_init(struct sock
*sk
)
2251 struct dst_entry
*dst
= __sk_dst_get(sk
);
2252 struct tcp_sock
*tp
= tcp_sk(sk
);
2255 /* We'll fix this up when we get a response from the other end.
2256 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2258 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2259 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2261 #ifdef CONFIG_TCP_MD5SIG
2262 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2263 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2266 /* If user gave his TCP_MAXSEG, record it to clamp */
2267 if (tp
->rx_opt
.user_mss
)
2268 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2271 tcp_sync_mss(sk
, dst_mtu(dst
));
2273 if (!tp
->window_clamp
)
2274 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2275 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2276 tcp_initialize_rcv_mss(sk
);
2278 tcp_select_initial_window(tcp_full_space(sk
),
2279 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2282 sysctl_tcp_window_scaling
,
2285 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2286 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2289 sock_reset_flag(sk
, SOCK_DONE
);
2291 tcp_init_wl(tp
, tp
->write_seq
, 0);
2292 tp
->snd_una
= tp
->write_seq
;
2293 tp
->snd_sml
= tp
->write_seq
;
2298 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2299 inet_csk(sk
)->icsk_retransmits
= 0;
2300 tcp_clear_retrans(tp
);
2304 * Build a SYN and send it off.
2306 int tcp_connect(struct sock
*sk
)
2308 struct tcp_sock
*tp
= tcp_sk(sk
);
2309 struct sk_buff
*buff
;
2311 tcp_connect_init(sk
);
2313 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2314 if (unlikely(buff
== NULL
))
2317 /* Reserve space for headers. */
2318 skb_reserve(buff
, MAX_TCP_HEADER
);
2320 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
2321 TCP_ECN_send_syn(sk
, buff
);
2322 TCP_SKB_CB(buff
)->sacked
= 0;
2323 skb_shinfo(buff
)->gso_segs
= 1;
2324 skb_shinfo(buff
)->gso_size
= 0;
2325 skb_shinfo(buff
)->gso_type
= 0;
2327 tp
->snd_nxt
= tp
->write_seq
;
2328 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
2329 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
2332 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2333 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2334 skb_header_release(buff
);
2335 __tcp_add_write_queue_tail(sk
, buff
);
2336 sk_charge_skb(sk
, buff
);
2337 tp
->packets_out
+= tcp_skb_pcount(buff
);
2338 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2340 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2341 * in order to make this packet get counted in tcpOutSegs.
2343 tp
->snd_nxt
= tp
->write_seq
;
2344 tp
->pushed_seq
= tp
->write_seq
;
2345 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
2347 /* Timer for repeating the SYN until an answer. */
2348 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2349 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2353 /* Send out a delayed ack, the caller does the policy checking
2354 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2357 void tcp_send_delayed_ack(struct sock
*sk
)
2359 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2360 int ato
= icsk
->icsk_ack
.ato
;
2361 unsigned long timeout
;
2363 if (ato
> TCP_DELACK_MIN
) {
2364 const struct tcp_sock
*tp
= tcp_sk(sk
);
2367 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2368 max_ato
= TCP_DELACK_MAX
;
2370 /* Slow path, intersegment interval is "high". */
2372 /* If some rtt estimate is known, use it to bound delayed ack.
2373 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2377 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
2383 ato
= min(ato
, max_ato
);
2386 /* Stay within the limit we were given */
2387 timeout
= jiffies
+ ato
;
2389 /* Use new timeout only if there wasn't a older one earlier. */
2390 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2391 /* If delack timer was blocked or is about to expire,
2394 if (icsk
->icsk_ack
.blocked
||
2395 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2400 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2401 timeout
= icsk
->icsk_ack
.timeout
;
2403 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2404 icsk
->icsk_ack
.timeout
= timeout
;
2405 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2408 /* This routine sends an ack and also updates the window. */
2409 void tcp_send_ack(struct sock
*sk
)
2411 /* If we have been reset, we may not send again. */
2412 if (sk
->sk_state
!= TCP_CLOSE
) {
2413 struct sk_buff
*buff
;
2415 /* We are not putting this on the write queue, so
2416 * tcp_transmit_skb() will set the ownership to this
2419 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2421 inet_csk_schedule_ack(sk
);
2422 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2423 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2424 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2428 /* Reserve space for headers and prepare control bits. */
2429 skb_reserve(buff
, MAX_TCP_HEADER
);
2431 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
2432 TCP_SKB_CB(buff
)->sacked
= 0;
2433 skb_shinfo(buff
)->gso_segs
= 1;
2434 skb_shinfo(buff
)->gso_size
= 0;
2435 skb_shinfo(buff
)->gso_type
= 0;
2437 /* Send it off, this clears delayed acks for us. */
2438 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
);
2439 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2440 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2444 /* This routine sends a packet with an out of date sequence
2445 * number. It assumes the other end will try to ack it.
2447 * Question: what should we make while urgent mode?
2448 * 4.4BSD forces sending single byte of data. We cannot send
2449 * out of window data, because we have SND.NXT==SND.MAX...
2451 * Current solution: to send TWO zero-length segments in urgent mode:
2452 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2453 * out-of-date with SND.UNA-1 to probe window.
2455 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2457 struct tcp_sock
*tp
= tcp_sk(sk
);
2458 struct sk_buff
*skb
;
2460 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2461 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2465 /* Reserve space for headers and set control bits. */
2466 skb_reserve(skb
, MAX_TCP_HEADER
);
2468 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2469 TCP_SKB_CB(skb
)->sacked
= urgent
;
2470 skb_shinfo(skb
)->gso_segs
= 1;
2471 skb_shinfo(skb
)->gso_size
= 0;
2472 skb_shinfo(skb
)->gso_type
= 0;
2474 /* Use a previous sequence. This should cause the other
2475 * end to send an ack. Don't queue or clone SKB, just
2478 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2479 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2480 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2481 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2484 int tcp_write_wakeup(struct sock
*sk
)
2486 if (sk
->sk_state
!= TCP_CLOSE
) {
2487 struct tcp_sock
*tp
= tcp_sk(sk
);
2488 struct sk_buff
*skb
;
2490 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2491 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
2493 unsigned int mss
= tcp_current_mss(sk
, 0);
2494 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
2496 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2497 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2499 /* We are probing the opening of a window
2500 * but the window size is != 0
2501 * must have been a result SWS avoidance ( sender )
2503 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2505 seg_size
= min(seg_size
, mss
);
2506 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2507 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2509 } else if (!tcp_skb_pcount(skb
))
2510 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2512 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2513 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2514 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2516 update_send_head(sk
, skb
);
2521 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2522 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2523 return tcp_xmit_probe_skb(sk
, 0);
2529 /* A window probe timeout has occurred. If window is not closed send
2530 * a partial packet else a zero probe.
2532 void tcp_send_probe0(struct sock
*sk
)
2534 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2535 struct tcp_sock
*tp
= tcp_sk(sk
);
2538 err
= tcp_write_wakeup(sk
);
2540 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2541 /* Cancel probe timer, if it is not required. */
2542 icsk
->icsk_probes_out
= 0;
2543 icsk
->icsk_backoff
= 0;
2548 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2549 icsk
->icsk_backoff
++;
2550 icsk
->icsk_probes_out
++;
2551 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2552 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2555 /* If packet was not sent due to local congestion,
2556 * do not backoff and do not remember icsk_probes_out.
2557 * Let local senders to fight for local resources.
2559 * Use accumulated backoff yet.
2561 if (!icsk
->icsk_probes_out
)
2562 icsk
->icsk_probes_out
= 1;
2563 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2564 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2565 TCP_RESOURCE_PROBE_INTERVAL
),
2570 EXPORT_SYMBOL(tcp_connect
);
2571 EXPORT_SYMBOL(tcp_make_synack
);
2572 EXPORT_SYMBOL(tcp_simple_retransmit
);
2573 EXPORT_SYMBOL(tcp_sync_mss
);
2574 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);
2575 EXPORT_SYMBOL(tcp_mtup_init
);