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
;
545 sk
->sk_route_caps
&= ~NETIF_F_GSO_MASK
;
549 skb_push(skb
, tcp_header_size
);
550 skb_reset_transport_header(skb
);
551 skb_set_owner_w(skb
, sk
);
553 /* Build TCP header and checksum it. */
555 th
->source
= inet
->sport
;
556 th
->dest
= inet
->dport
;
557 th
->seq
= htonl(tcb
->seq
);
558 th
->ack_seq
= htonl(tp
->rcv_nxt
);
559 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
562 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
563 /* RFC1323: The window in SYN & SYN/ACK segments
566 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
568 th
->window
= htons(tcp_select_window(sk
));
573 if (unlikely(tp
->urg_mode
&&
574 between(tp
->snd_up
, tcb
->seq
+ 1, tcb
->seq
+ 0xFFFF))) {
575 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
579 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
580 tcp_syn_build_options((__be32
*)(th
+ 1),
581 tcp_advertise_mss(sk
),
582 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
583 (sysctl_flags
& SYSCTL_FLAG_SACK
),
584 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
585 tp
->rx_opt
.rcv_wscale
,
587 tp
->rx_opt
.ts_recent
,
589 #ifdef CONFIG_TCP_MD5SIG
590 md5
? &md5_hash_location
:
594 tcp_build_and_update_options((__be32
*)(th
+ 1),
596 #ifdef CONFIG_TCP_MD5SIG
597 md5
? &md5_hash_location
:
600 TCP_ECN_send(sk
, skb
, tcp_header_size
);
603 #ifdef CONFIG_TCP_MD5SIG
604 /* Calculate the MD5 hash, as we have all we need now */
606 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
611 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
613 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
614 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
616 if (skb
->len
!= tcp_header_size
)
617 tcp_event_data_sent(tp
, skb
, sk
);
619 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
620 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
622 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
623 if (likely(err
<= 0))
626 tcp_enter_cwr(sk
, 1);
628 return net_xmit_eval(err
);
630 #undef SYSCTL_FLAG_TSTAMPS
631 #undef SYSCTL_FLAG_WSCALE
632 #undef SYSCTL_FLAG_SACK
635 /* This routine just queue's the buffer
637 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
638 * otherwise socket can stall.
640 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
642 struct tcp_sock
*tp
= tcp_sk(sk
);
644 /* Advance write_seq and place onto the write_queue. */
645 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
646 skb_header_release(skb
);
647 tcp_add_write_queue_tail(sk
, skb
);
648 sk
->sk_wmem_queued
+= skb
->truesize
;
649 sk_mem_charge(sk
, skb
->truesize
);
652 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
653 unsigned int mss_now
)
655 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
)) {
656 /* Avoid the costly divide in the normal
659 skb_shinfo(skb
)->gso_segs
= 1;
660 skb_shinfo(skb
)->gso_size
= 0;
661 skb_shinfo(skb
)->gso_type
= 0;
663 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
664 skb_shinfo(skb
)->gso_size
= mss_now
;
665 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
669 /* When a modification to fackets out becomes necessary, we need to check
670 * skb is counted to fackets_out or not.
672 static void tcp_adjust_fackets_out(struct sock
*sk
, struct sk_buff
*skb
,
675 struct tcp_sock
*tp
= tcp_sk(sk
);
677 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
680 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
681 tp
->fackets_out
-= decr
;
684 /* Function to create two new TCP segments. Shrinks the given segment
685 * to the specified size and appends a new segment with the rest of the
686 * packet to the list. This won't be called frequently, I hope.
687 * Remember, these are still headerless SKBs at this point.
689 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
690 unsigned int mss_now
)
692 struct tcp_sock
*tp
= tcp_sk(sk
);
693 struct sk_buff
*buff
;
694 int nsize
, old_factor
;
698 BUG_ON(len
> skb
->len
);
700 tcp_clear_retrans_hints_partial(tp
);
701 nsize
= skb_headlen(skb
) - len
;
705 if (skb_cloned(skb
) &&
706 skb_is_nonlinear(skb
) &&
707 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
710 /* Get a new skb... force flag on. */
711 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
713 return -ENOMEM
; /* We'll just try again later. */
715 sk
->sk_wmem_queued
+= buff
->truesize
;
716 sk_mem_charge(sk
, buff
->truesize
);
717 nlen
= skb
->len
- len
- nsize
;
718 buff
->truesize
+= nlen
;
719 skb
->truesize
-= nlen
;
721 /* Correct the sequence numbers. */
722 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
723 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
724 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
726 /* PSH and FIN should only be set in the second packet. */
727 flags
= TCP_SKB_CB(skb
)->flags
;
728 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
729 TCP_SKB_CB(buff
)->flags
= flags
;
730 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
732 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
733 /* Copy and checksum data tail into the new buffer. */
734 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
735 skb_put(buff
, nsize
),
740 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
742 skb
->ip_summed
= CHECKSUM_PARTIAL
;
743 skb_split(skb
, buff
, len
);
746 buff
->ip_summed
= skb
->ip_summed
;
748 /* Looks stupid, but our code really uses when of
749 * skbs, which it never sent before. --ANK
751 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
752 buff
->tstamp
= skb
->tstamp
;
754 old_factor
= tcp_skb_pcount(skb
);
756 /* Fix up tso_factor for both original and new SKB. */
757 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
758 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
760 /* If this packet has been sent out already, we must
761 * adjust the various packet counters.
763 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
764 int diff
= old_factor
- tcp_skb_pcount(skb
) -
765 tcp_skb_pcount(buff
);
767 tp
->packets_out
-= diff
;
769 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
770 tp
->sacked_out
-= diff
;
771 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
772 tp
->retrans_out
-= diff
;
774 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
775 tp
->lost_out
-= diff
;
777 /* Adjust Reno SACK estimate. */
778 if (tcp_is_reno(tp
) && diff
> 0) {
779 tcp_dec_pcount_approx_int(&tp
->sacked_out
, diff
);
780 tcp_verify_left_out(tp
);
782 tcp_adjust_fackets_out(sk
, skb
, diff
);
785 /* Link BUFF into the send queue. */
786 skb_header_release(buff
);
787 tcp_insert_write_queue_after(skb
, buff
, sk
);
792 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
793 * eventually). The difference is that pulled data not copied, but
794 * immediately discarded.
796 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
802 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
803 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
804 put_page(skb_shinfo(skb
)->frags
[i
].page
);
805 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
807 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
809 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
810 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
816 skb_shinfo(skb
)->nr_frags
= k
;
818 skb_reset_tail_pointer(skb
);
819 skb
->data_len
-= len
;
820 skb
->len
= skb
->data_len
;
823 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
825 if (skb_cloned(skb
) && pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
828 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
829 if (unlikely(len
< skb_headlen(skb
)))
830 __skb_pull(skb
, len
);
832 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
834 TCP_SKB_CB(skb
)->seq
+= len
;
835 skb
->ip_summed
= CHECKSUM_PARTIAL
;
837 skb
->truesize
-= len
;
838 sk
->sk_wmem_queued
-= len
;
839 sk_mem_uncharge(sk
, len
);
840 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
842 /* Any change of skb->len requires recalculation of tso
845 if (tcp_skb_pcount(skb
) > 1)
846 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
851 /* Not accounting for SACKs here. */
852 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
854 struct tcp_sock
*tp
= tcp_sk(sk
);
855 struct inet_connection_sock
*icsk
= inet_csk(sk
);
858 /* Calculate base mss without TCP options:
859 It is MMS_S - sizeof(tcphdr) of rfc1122
861 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
863 /* Clamp it (mss_clamp does not include tcp options) */
864 if (mss_now
> tp
->rx_opt
.mss_clamp
)
865 mss_now
= tp
->rx_opt
.mss_clamp
;
867 /* Now subtract optional transport overhead */
868 mss_now
-= icsk
->icsk_ext_hdr_len
;
870 /* Then reserve room for full set of TCP options and 8 bytes of data */
874 /* Now subtract TCP options size, not including SACKs */
875 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
880 /* Inverse of above */
881 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
883 struct tcp_sock
*tp
= tcp_sk(sk
);
884 struct inet_connection_sock
*icsk
= inet_csk(sk
);
889 icsk
->icsk_ext_hdr_len
+
890 icsk
->icsk_af_ops
->net_header_len
;
895 void tcp_mtup_init(struct sock
*sk
)
897 struct tcp_sock
*tp
= tcp_sk(sk
);
898 struct inet_connection_sock
*icsk
= inet_csk(sk
);
900 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
901 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
902 icsk
->icsk_af_ops
->net_header_len
;
903 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
904 icsk
->icsk_mtup
.probe_size
= 0;
907 /* Bound MSS / TSO packet size with the half of the window */
908 static int tcp_bound_to_half_wnd(struct tcp_sock
*tp
, int pktsize
)
910 if (tp
->max_window
&& pktsize
> (tp
->max_window
>> 1))
911 return max(tp
->max_window
>> 1, 68U - tp
->tcp_header_len
);
916 /* This function synchronize snd mss to current pmtu/exthdr set.
918 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
919 for TCP options, but includes only bare TCP header.
921 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
922 It is minimum of user_mss and mss received with SYN.
923 It also does not include TCP options.
925 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
927 tp->mss_cache is current effective sending mss, including
928 all tcp options except for SACKs. It is evaluated,
929 taking into account current pmtu, but never exceeds
930 tp->rx_opt.mss_clamp.
932 NOTE1. rfc1122 clearly states that advertised MSS
933 DOES NOT include either tcp or ip options.
935 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
936 are READ ONLY outside this function. --ANK (980731)
938 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
940 struct tcp_sock
*tp
= tcp_sk(sk
);
941 struct inet_connection_sock
*icsk
= inet_csk(sk
);
944 if (icsk
->icsk_mtup
.search_high
> pmtu
)
945 icsk
->icsk_mtup
.search_high
= pmtu
;
947 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
948 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
950 /* And store cached results */
951 icsk
->icsk_pmtu_cookie
= pmtu
;
952 if (icsk
->icsk_mtup
.enabled
)
953 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
954 tp
->mss_cache
= mss_now
;
959 /* Compute the current effective MSS, taking SACKs and IP options,
960 * and even PMTU discovery events into account.
962 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
963 * cannot be large. However, taking into account rare use of URG, this
966 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
968 struct tcp_sock
*tp
= tcp_sk(sk
);
969 struct dst_entry
*dst
= __sk_dst_get(sk
);
974 mss_now
= tp
->mss_cache
;
976 if (large_allowed
&& sk_can_gso(sk
) && !tp
->urg_mode
)
980 u32 mtu
= dst_mtu(dst
);
981 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
982 mss_now
= tcp_sync_mss(sk
, mtu
);
985 if (tp
->rx_opt
.eff_sacks
)
986 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
987 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
989 #ifdef CONFIG_TCP_MD5SIG
990 if (tp
->af_specific
->md5_lookup(sk
, sk
))
991 mss_now
-= TCPOLEN_MD5SIG_ALIGNED
;
994 xmit_size_goal
= mss_now
;
997 xmit_size_goal
= ((sk
->sk_gso_max_size
- 1) -
998 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
999 inet_csk(sk
)->icsk_ext_hdr_len
-
1000 tp
->tcp_header_len
);
1002 xmit_size_goal
= tcp_bound_to_half_wnd(tp
, xmit_size_goal
);
1003 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
1005 tp
->xmit_size_goal
= xmit_size_goal
;
1010 /* Congestion window validation. (RFC2861) */
1011 static void tcp_cwnd_validate(struct sock
*sk
)
1013 struct tcp_sock
*tp
= tcp_sk(sk
);
1015 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1016 /* Network is feed fully. */
1017 tp
->snd_cwnd_used
= 0;
1018 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1020 /* Network starves. */
1021 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1022 tp
->snd_cwnd_used
= tp
->packets_out
;
1024 if (sysctl_tcp_slow_start_after_idle
&&
1025 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1026 tcp_cwnd_application_limited(sk
);
1030 /* Returns the portion of skb which can be sent right away without
1031 * introducing MSS oddities to segment boundaries. In rare cases where
1032 * mss_now != mss_cache, we will request caller to create a small skb
1033 * per input skb which could be mostly avoided here (if desired).
1035 * We explicitly want to create a request for splitting write queue tail
1036 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1037 * thus all the complexity (cwnd_len is always MSS multiple which we
1038 * return whenever allowed by the other factors). Basically we need the
1039 * modulo only when the receiver window alone is the limiting factor or
1040 * when we would be allowed to send the split-due-to-Nagle skb fully.
1042 static unsigned int tcp_mss_split_point(struct sock
*sk
, struct sk_buff
*skb
,
1043 unsigned int mss_now
, unsigned int cwnd
)
1045 struct tcp_sock
*tp
= tcp_sk(sk
);
1046 u32 needed
, window
, cwnd_len
;
1048 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1049 cwnd_len
= mss_now
* cwnd
;
1051 if (likely(cwnd_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1054 needed
= min(skb
->len
, window
);
1056 if (cwnd_len
<= needed
)
1059 return needed
- needed
% mss_now
;
1062 /* Can at least one segment of SKB be sent right now, according to the
1063 * congestion window rules? If so, return how many segments are allowed.
1065 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
,
1066 struct sk_buff
*skb
)
1068 u32 in_flight
, cwnd
;
1070 /* Don't be strict about the congestion window for the final FIN. */
1071 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1072 tcp_skb_pcount(skb
) == 1)
1075 in_flight
= tcp_packets_in_flight(tp
);
1076 cwnd
= tp
->snd_cwnd
;
1077 if (in_flight
< cwnd
)
1078 return (cwnd
- in_flight
);
1083 /* This must be invoked the first time we consider transmitting
1084 * SKB onto the wire.
1086 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
1087 unsigned int mss_now
)
1089 int tso_segs
= tcp_skb_pcount(skb
);
1091 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1092 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1093 tso_segs
= tcp_skb_pcount(skb
);
1098 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1100 return after(tp
->snd_sml
,tp
->snd_una
) &&
1101 !after(tp
->snd_sml
, tp
->snd_nxt
);
1104 /* Return 0, if packet can be sent now without violation Nagle's rules:
1105 * 1. It is full sized.
1106 * 2. Or it contains FIN. (already checked by caller)
1107 * 3. Or TCP_NODELAY was set.
1108 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1109 * With Minshall's modification: all sent small packets are ACKed.
1111 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1112 const struct sk_buff
*skb
,
1113 unsigned mss_now
, int nonagle
)
1115 return (skb
->len
< mss_now
&&
1116 ((nonagle
& TCP_NAGLE_CORK
) ||
1117 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
))));
1120 /* Return non-zero if the Nagle test allows this packet to be
1123 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1124 unsigned int cur_mss
, int nonagle
)
1126 /* Nagle rule does not apply to frames, which sit in the middle of the
1127 * write_queue (they have no chances to get new data).
1129 * This is implemented in the callers, where they modify the 'nonagle'
1130 * argument based upon the location of SKB in the send queue.
1132 if (nonagle
& TCP_NAGLE_PUSH
)
1135 /* Don't use the nagle rule for urgent data (or for the final FIN).
1136 * Nagle can be ignored during F-RTO too (see RFC4138).
1138 if (tp
->urg_mode
|| (tp
->frto_counter
== 2) ||
1139 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1142 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1148 /* Does at least the first segment of SKB fit into the send window? */
1149 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1150 unsigned int cur_mss
)
1152 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1154 if (skb
->len
> cur_mss
)
1155 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1157 return !after(end_seq
, tcp_wnd_end(tp
));
1160 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1161 * should be put on the wire right now. If so, it returns the number of
1162 * packets allowed by the congestion window.
1164 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1165 unsigned int cur_mss
, int nonagle
)
1167 struct tcp_sock
*tp
= tcp_sk(sk
);
1168 unsigned int cwnd_quota
;
1170 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1172 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1175 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1176 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1182 int tcp_may_send_now(struct sock
*sk
)
1184 struct tcp_sock
*tp
= tcp_sk(sk
);
1185 struct sk_buff
*skb
= tcp_send_head(sk
);
1188 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1189 (tcp_skb_is_last(sk
, skb
) ?
1190 tp
->nonagle
: TCP_NAGLE_PUSH
)));
1193 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1194 * which is put after SKB on the list. It is very much like
1195 * tcp_fragment() except that it may make several kinds of assumptions
1196 * in order to speed up the splitting operation. In particular, we
1197 * know that all the data is in scatter-gather pages, and that the
1198 * packet has never been sent out before (and thus is not cloned).
1200 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1201 unsigned int mss_now
)
1203 struct sk_buff
*buff
;
1204 int nlen
= skb
->len
- len
;
1207 /* All of a TSO frame must be composed of paged data. */
1208 if (skb
->len
!= skb
->data_len
)
1209 return tcp_fragment(sk
, skb
, len
, mss_now
);
1211 buff
= sk_stream_alloc_skb(sk
, 0, GFP_ATOMIC
);
1212 if (unlikely(buff
== NULL
))
1215 sk
->sk_wmem_queued
+= buff
->truesize
;
1216 sk_mem_charge(sk
, buff
->truesize
);
1217 buff
->truesize
+= nlen
;
1218 skb
->truesize
-= nlen
;
1220 /* Correct the sequence numbers. */
1221 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1222 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1223 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1225 /* PSH and FIN should only be set in the second packet. */
1226 flags
= TCP_SKB_CB(skb
)->flags
;
1227 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
1228 TCP_SKB_CB(buff
)->flags
= flags
;
1230 /* This packet was never sent out yet, so no SACK bits. */
1231 TCP_SKB_CB(buff
)->sacked
= 0;
1233 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1234 skb_split(skb
, buff
, len
);
1236 /* Fix up tso_factor for both original and new SKB. */
1237 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1238 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1240 /* Link BUFF into the send queue. */
1241 skb_header_release(buff
);
1242 tcp_insert_write_queue_after(skb
, buff
, sk
);
1247 /* Try to defer sending, if possible, in order to minimize the amount
1248 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1250 * This algorithm is from John Heffner.
1252 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1254 struct tcp_sock
*tp
= tcp_sk(sk
);
1255 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1256 u32 send_win
, cong_win
, limit
, in_flight
;
1258 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1261 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1264 /* Defer for less than two clock ticks. */
1265 if (tp
->tso_deferred
&&
1266 ((jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1269 in_flight
= tcp_packets_in_flight(tp
);
1271 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1273 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1275 /* From in_flight test above, we know that cwnd > in_flight. */
1276 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1278 limit
= min(send_win
, cong_win
);
1280 /* If a full-sized TSO skb can be sent, do it. */
1281 if (limit
>= sk
->sk_gso_max_size
)
1284 if (sysctl_tcp_tso_win_divisor
) {
1285 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1287 /* If at least some fraction of a window is available,
1290 chunk
/= sysctl_tcp_tso_win_divisor
;
1294 /* Different approach, try not to defer past a single
1295 * ACK. Receiver should ACK every other full sized
1296 * frame, so if we have space for more than 3 frames
1299 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1303 /* Ok, it looks like it is advisable to defer. */
1304 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1309 tp
->tso_deferred
= 0;
1313 /* Create a new MTU probe if we are ready.
1314 * Returns 0 if we should wait to probe (no cwnd available),
1315 * 1 if a probe was sent,
1318 static int tcp_mtu_probe(struct sock
*sk
)
1320 struct tcp_sock
*tp
= tcp_sk(sk
);
1321 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1322 struct sk_buff
*skb
, *nskb
, *next
;
1329 /* Not currently probing/verifying,
1331 * have enough cwnd, and
1332 * not SACKing (the variable headers throw things off) */
1333 if (!icsk
->icsk_mtup
.enabled
||
1334 icsk
->icsk_mtup
.probe_size
||
1335 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1336 tp
->snd_cwnd
< 11 ||
1337 tp
->rx_opt
.eff_sacks
)
1340 /* Very simple search strategy: just double the MSS. */
1341 mss_now
= tcp_current_mss(sk
, 0);
1342 probe_size
= 2 * tp
->mss_cache
;
1343 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1344 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1345 /* TODO: set timer for probe_converge_event */
1349 /* Have enough data in the send queue to probe? */
1350 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1353 if (tp
->snd_wnd
< size_needed
)
1355 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1358 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1359 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1360 if (!tcp_packets_in_flight(tp
))
1366 /* We're allowed to probe. Build it now. */
1367 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1369 sk
->sk_wmem_queued
+= nskb
->truesize
;
1370 sk_mem_charge(sk
, nskb
->truesize
);
1372 skb
= tcp_send_head(sk
);
1374 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1375 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1376 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1377 TCP_SKB_CB(nskb
)->sacked
= 0;
1379 nskb
->ip_summed
= skb
->ip_summed
;
1381 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1384 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1385 copy
= min_t(int, skb
->len
, probe_size
- len
);
1386 if (nskb
->ip_summed
)
1387 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1389 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1390 skb_put(nskb
, copy
),
1393 if (skb
->len
<= copy
) {
1394 /* We've eaten all the data from this skb.
1396 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1397 tcp_unlink_write_queue(skb
, sk
);
1398 sk_wmem_free_skb(sk
, skb
);
1400 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1401 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1402 if (!skb_shinfo(skb
)->nr_frags
) {
1403 skb_pull(skb
, copy
);
1404 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1405 skb
->csum
= csum_partial(skb
->data
,
1408 __pskb_trim_head(skb
, copy
);
1409 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1411 TCP_SKB_CB(skb
)->seq
+= copy
;
1416 if (len
>= probe_size
)
1419 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1421 /* We're ready to send. If this fails, the probe will
1422 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1423 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1424 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1425 /* Decrement cwnd here because we are sending
1426 * effectively two packets. */
1428 tcp_event_new_data_sent(sk
, nskb
);
1430 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1431 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1432 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1440 /* This routine writes packets to the network. It advances the
1441 * send_head. This happens as incoming acks open up the remote
1444 * Returns 1, if no segments are in flight and we have queued segments, but
1445 * cannot send anything now because of SWS or another problem.
1447 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1449 struct tcp_sock
*tp
= tcp_sk(sk
);
1450 struct sk_buff
*skb
;
1451 unsigned int tso_segs
, sent_pkts
;
1455 /* If we are closed, the bytes will have to remain here.
1456 * In time closedown will finish, we empty the write queue and all
1459 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1464 /* Do MTU probing. */
1465 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1467 } else if (result
> 0) {
1471 while ((skb
= tcp_send_head(sk
))) {
1474 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1477 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1481 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1484 if (tso_segs
== 1) {
1485 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1486 (tcp_skb_is_last(sk
, skb
) ?
1487 nonagle
: TCP_NAGLE_PUSH
))))
1490 if (tcp_tso_should_defer(sk
, skb
))
1496 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1499 if (skb
->len
> limit
&&
1500 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1503 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1505 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1508 /* Advance the send_head. This one is sent out.
1509 * This call will increment packets_out.
1511 tcp_event_new_data_sent(sk
, skb
);
1513 tcp_minshall_update(tp
, mss_now
, skb
);
1517 if (likely(sent_pkts
)) {
1518 tcp_cwnd_validate(sk
);
1521 return !tp
->packets_out
&& tcp_send_head(sk
);
1524 /* Push out any pending frames which were held back due to
1525 * TCP_CORK or attempt at coalescing tiny packets.
1526 * The socket must be locked by the caller.
1528 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1531 struct sk_buff
*skb
= tcp_send_head(sk
);
1534 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1535 tcp_check_probe_timer(sk
);
1539 /* Send _single_ skb sitting at the send head. This function requires
1540 * true push pending frames to setup probe timer etc.
1542 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1544 struct sk_buff
*skb
= tcp_send_head(sk
);
1545 unsigned int tso_segs
, cwnd_quota
;
1547 BUG_ON(!skb
|| skb
->len
< mss_now
);
1549 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1550 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1552 if (likely(cwnd_quota
)) {
1559 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1562 if (skb
->len
> limit
&&
1563 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1566 /* Send it out now. */
1567 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1569 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1570 tcp_event_new_data_sent(sk
, skb
);
1571 tcp_cwnd_validate(sk
);
1577 /* This function returns the amount that we can raise the
1578 * usable window based on the following constraints
1580 * 1. The window can never be shrunk once it is offered (RFC 793)
1581 * 2. We limit memory per socket
1584 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1585 * RECV.NEXT + RCV.WIN fixed until:
1586 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1588 * i.e. don't raise the right edge of the window until you can raise
1589 * it at least MSS bytes.
1591 * Unfortunately, the recommended algorithm breaks header prediction,
1592 * since header prediction assumes th->window stays fixed.
1594 * Strictly speaking, keeping th->window fixed violates the receiver
1595 * side SWS prevention criteria. The problem is that under this rule
1596 * a stream of single byte packets will cause the right side of the
1597 * window to always advance by a single byte.
1599 * Of course, if the sender implements sender side SWS prevention
1600 * then this will not be a problem.
1602 * BSD seems to make the following compromise:
1604 * If the free space is less than the 1/4 of the maximum
1605 * space available and the free space is less than 1/2 mss,
1606 * then set the window to 0.
1607 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1608 * Otherwise, just prevent the window from shrinking
1609 * and from being larger than the largest representable value.
1611 * This prevents incremental opening of the window in the regime
1612 * where TCP is limited by the speed of the reader side taking
1613 * data out of the TCP receive queue. It does nothing about
1614 * those cases where the window is constrained on the sender side
1615 * because the pipeline is full.
1617 * BSD also seems to "accidentally" limit itself to windows that are a
1618 * multiple of MSS, at least until the free space gets quite small.
1619 * This would appear to be a side effect of the mbuf implementation.
1620 * Combining these two algorithms results in the observed behavior
1621 * of having a fixed window size at almost all times.
1623 * Below we obtain similar behavior by forcing the offered window to
1624 * a multiple of the mss when it is feasible to do so.
1626 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1627 * Regular options like TIMESTAMP are taken into account.
1629 u32
__tcp_select_window(struct sock
*sk
)
1631 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1632 struct tcp_sock
*tp
= tcp_sk(sk
);
1633 /* MSS for the peer's data. Previous versions used mss_clamp
1634 * here. I don't know if the value based on our guesses
1635 * of peer's MSS is better for the performance. It's more correct
1636 * but may be worse for the performance because of rcv_mss
1637 * fluctuations. --SAW 1998/11/1
1639 int mss
= icsk
->icsk_ack
.rcv_mss
;
1640 int free_space
= tcp_space(sk
);
1641 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1644 if (mss
> full_space
)
1647 if (free_space
< (full_space
>> 1)) {
1648 icsk
->icsk_ack
.quick
= 0;
1650 if (tcp_memory_pressure
)
1651 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
1654 if (free_space
< mss
)
1658 if (free_space
> tp
->rcv_ssthresh
)
1659 free_space
= tp
->rcv_ssthresh
;
1661 /* Don't do rounding if we are using window scaling, since the
1662 * scaled window will not line up with the MSS boundary anyway.
1664 window
= tp
->rcv_wnd
;
1665 if (tp
->rx_opt
.rcv_wscale
) {
1666 window
= free_space
;
1668 /* Advertise enough space so that it won't get scaled away.
1669 * Import case: prevent zero window announcement if
1670 * 1<<rcv_wscale > mss.
1672 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1673 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1674 << tp
->rx_opt
.rcv_wscale
);
1676 /* Get the largest window that is a nice multiple of mss.
1677 * Window clamp already applied above.
1678 * If our current window offering is within 1 mss of the
1679 * free space we just keep it. This prevents the divide
1680 * and multiply from happening most of the time.
1681 * We also don't do any window rounding when the free space
1684 if (window
<= free_space
- mss
|| window
> free_space
)
1685 window
= (free_space
/ mss
) * mss
;
1686 else if (mss
== full_space
&&
1687 free_space
> window
+ (full_space
>> 1))
1688 window
= free_space
;
1694 /* Attempt to collapse two adjacent SKB's during retransmission. */
1695 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
,
1698 struct tcp_sock
*tp
= tcp_sk(sk
);
1699 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1700 int skb_size
, next_skb_size
;
1703 /* The first test we must make is that neither of these two
1704 * SKB's are still referenced by someone else.
1706 if (skb_cloned(skb
) || skb_cloned(next_skb
))
1709 skb_size
= skb
->len
;
1710 next_skb_size
= next_skb
->len
;
1711 flags
= TCP_SKB_CB(skb
)->flags
;
1713 /* Also punt if next skb has been SACK'd. */
1714 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1717 /* Next skb is out of window. */
1718 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tcp_wnd_end(tp
)))
1721 /* Punt if not enough space exists in the first SKB for
1722 * the data in the second, or the total combined payload
1723 * would exceed the MSS.
1725 if ((next_skb_size
> skb_tailroom(skb
)) ||
1726 ((skb_size
+ next_skb_size
) > mss_now
))
1729 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
1731 tcp_highest_sack_combine(sk
, next_skb
, skb
);
1733 /* Ok. We will be able to collapse the packet. */
1734 tcp_unlink_write_queue(next_skb
, sk
);
1736 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
1739 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1740 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1742 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1743 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1745 /* Update sequence range on original skb. */
1746 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1748 /* Merge over control information. */
1749 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1750 TCP_SKB_CB(skb
)->flags
= flags
;
1752 /* All done, get rid of second SKB and account for it so
1753 * packet counting does not break.
1755 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
1756 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1757 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1758 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_LOST
)
1759 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1760 /* Reno case is special. Sigh... */
1761 if (tcp_is_reno(tp
) && tp
->sacked_out
)
1762 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1764 tcp_adjust_fackets_out(sk
, next_skb
, tcp_skb_pcount(next_skb
));
1765 tp
->packets_out
-= tcp_skb_pcount(next_skb
);
1767 /* changed transmit queue under us so clear hints */
1768 tcp_clear_retrans_hints_partial(tp
);
1770 sk_wmem_free_skb(sk
, next_skb
);
1773 /* Do a simple retransmit without using the backoff mechanisms in
1774 * tcp_timer. This is used for path mtu discovery.
1775 * The socket is already locked here.
1777 void tcp_simple_retransmit(struct sock
*sk
)
1779 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1780 struct tcp_sock
*tp
= tcp_sk(sk
);
1781 struct sk_buff
*skb
;
1782 unsigned int mss
= tcp_current_mss(sk
, 0);
1785 tcp_for_write_queue(skb
, sk
) {
1786 if (skb
== tcp_send_head(sk
))
1788 if (skb
->len
> mss
&&
1789 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1790 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1791 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1792 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1794 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)) {
1795 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1796 tp
->lost_out
+= tcp_skb_pcount(skb
);
1802 tcp_clear_all_retrans_hints(tp
);
1807 if (tcp_is_reno(tp
))
1808 tcp_limit_reno_sacked(tp
);
1810 tcp_verify_left_out(tp
);
1812 /* Don't muck with the congestion window here.
1813 * Reason is that we do not increase amount of _data_
1814 * in network, but units changed and effective
1815 * cwnd/ssthresh really reduced now.
1817 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1818 tp
->high_seq
= tp
->snd_nxt
;
1819 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1820 tp
->prior_ssthresh
= 0;
1821 tp
->undo_marker
= 0;
1822 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1824 tcp_xmit_retransmit_queue(sk
);
1827 /* This retransmits one SKB. Policy decisions and retransmit queue
1828 * state updates are done by the caller. Returns non-zero if an
1829 * error occurred which prevented the send.
1831 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1833 struct tcp_sock
*tp
= tcp_sk(sk
);
1834 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1835 unsigned int cur_mss
;
1838 /* Inconslusive MTU probe */
1839 if (icsk
->icsk_mtup
.probe_size
) {
1840 icsk
->icsk_mtup
.probe_size
= 0;
1843 /* Do not sent more than we queued. 1/4 is reserved for possible
1844 * copying overhead: fragmentation, tunneling, mangling etc.
1846 if (atomic_read(&sk
->sk_wmem_alloc
) >
1847 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1850 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1851 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1853 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1857 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1858 return -EHOSTUNREACH
; /* Routing failure or similar. */
1860 cur_mss
= tcp_current_mss(sk
, 0);
1862 /* If receiver has shrunk his window, and skb is out of
1863 * new window, do not retransmit it. The exception is the
1864 * case, when window is shrunk to zero. In this case
1865 * our retransmit serves as a zero window probe.
1867 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))
1868 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1871 if (skb
->len
> cur_mss
) {
1872 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1873 return -ENOMEM
; /* We'll try again later. */
1876 /* Collapse two adjacent packets if worthwhile and we can. */
1877 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1878 (skb
->len
< (cur_mss
>> 1)) &&
1879 (tcp_write_queue_next(sk
, skb
) != tcp_send_head(sk
)) &&
1880 (!tcp_skb_is_last(sk
, skb
)) &&
1881 (skb_shinfo(skb
)->nr_frags
== 0 &&
1882 skb_shinfo(tcp_write_queue_next(sk
, skb
))->nr_frags
== 0) &&
1883 (tcp_skb_pcount(skb
) == 1 &&
1884 tcp_skb_pcount(tcp_write_queue_next(sk
, skb
)) == 1) &&
1885 (sysctl_tcp_retrans_collapse
!= 0))
1886 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1888 /* Some Solaris stacks overoptimize and ignore the FIN on a
1889 * retransmit when old data is attached. So strip it off
1890 * since it is cheap to do so and saves bytes on the network.
1893 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1894 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1895 if (!pskb_trim(skb
, 0)) {
1896 /* Reuse, even though it does some unnecessary work */
1897 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
1898 TCP_SKB_CB(skb
)->flags
);
1899 skb
->ip_summed
= CHECKSUM_NONE
;
1903 /* Make a copy, if the first transmission SKB clone we made
1904 * is still in somebody's hands, else make a clone.
1906 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1908 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1911 /* Update global TCP statistics. */
1912 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
1914 tp
->total_retrans
++;
1916 #if FASTRETRANS_DEBUG > 0
1917 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1918 if (net_ratelimit())
1919 printk(KERN_DEBUG
"retrans_out leaked.\n");
1922 if (!tp
->retrans_out
)
1923 tp
->lost_retrans_low
= tp
->snd_nxt
;
1924 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1925 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1927 /* Save stamp of the first retransmit. */
1928 if (!tp
->retrans_stamp
)
1929 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1933 /* snd_nxt is stored to detect loss of retransmitted segment,
1934 * see tcp_input.c tcp_sacktag_write_queue().
1936 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1941 /* This gets called after a retransmit timeout, and the initially
1942 * retransmitted data is acknowledged. It tries to continue
1943 * resending the rest of the retransmit queue, until either
1944 * we've sent it all or the congestion window limit is reached.
1945 * If doing SACK, the first ACK which comes back for a timeout
1946 * based retransmit packet might feed us FACK information again.
1947 * If so, we use it to avoid unnecessarily retransmissions.
1949 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1951 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1952 struct tcp_sock
*tp
= tcp_sk(sk
);
1953 struct sk_buff
*skb
;
1956 if (tp
->retransmit_skb_hint
) {
1957 skb
= tp
->retransmit_skb_hint
;
1958 packet_cnt
= tp
->retransmit_cnt_hint
;
1960 skb
= tcp_write_queue_head(sk
);
1964 /* First pass: retransmit lost packets. */
1966 tcp_for_write_queue_from(skb
, sk
) {
1967 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1969 if (skb
== tcp_send_head(sk
))
1971 /* we could do better than to assign each time */
1972 tp
->retransmit_skb_hint
= skb
;
1973 tp
->retransmit_cnt_hint
= packet_cnt
;
1975 /* Assume this retransmit will generate
1976 * only one packet for congestion window
1977 * calculation purposes. This works because
1978 * tcp_retransmit_skb() will chop up the
1979 * packet to be MSS sized and all the
1980 * packet counting works out.
1982 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1985 if (sacked
& TCPCB_LOST
) {
1986 if (!(sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1989 if (tcp_retransmit_skb(sk
, skb
)) {
1990 tp
->retransmit_skb_hint
= NULL
;
1993 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1994 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
1996 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
1997 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1999 if (skb
== tcp_write_queue_head(sk
))
2000 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2001 inet_csk(sk
)->icsk_rto
,
2005 packet_cnt
+= tcp_skb_pcount(skb
);
2006 if (packet_cnt
>= tp
->lost_out
)
2012 /* OK, demanded retransmission is finished. */
2014 /* Forward retransmissions are possible only during Recovery. */
2015 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2018 /* No forward retransmissions in Reno are possible. */
2019 if (tcp_is_reno(tp
))
2022 /* Yeah, we have to make difficult choice between forward transmission
2023 * and retransmission... Both ways have their merits...
2025 * For now we do not retransmit anything, while we have some new
2026 * segments to send. In the other cases, follow rule 3 for
2027 * NextSeg() specified in RFC3517.
2030 if (tcp_may_send_now(sk
))
2033 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2034 if (!tp
->sacked_out
)
2037 if (tp
->forward_skb_hint
)
2038 skb
= tp
->forward_skb_hint
;
2040 skb
= tcp_write_queue_head(sk
);
2042 tcp_for_write_queue_from(skb
, sk
) {
2043 if (skb
== tcp_send_head(sk
))
2045 tp
->forward_skb_hint
= skb
;
2047 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2050 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2053 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
2056 /* Ok, retransmit it. */
2057 if (tcp_retransmit_skb(sk
, skb
)) {
2058 tp
->forward_skb_hint
= NULL
;
2062 if (skb
== tcp_write_queue_head(sk
))
2063 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2064 inet_csk(sk
)->icsk_rto
,
2067 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPFORWARDRETRANS
);
2071 /* Send a fin. The caller locks the socket for us. This cannot be
2072 * allowed to fail queueing a FIN frame under any circumstances.
2074 void tcp_send_fin(struct sock
*sk
)
2076 struct tcp_sock
*tp
= tcp_sk(sk
);
2077 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2080 /* Optimization, tack on the FIN if we have a queue of
2081 * unsent frames. But be careful about outgoing SACKS
2084 mss_now
= tcp_current_mss(sk
, 1);
2086 if (tcp_send_head(sk
) != NULL
) {
2087 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2088 TCP_SKB_CB(skb
)->end_seq
++;
2091 /* Socket is locked, keep trying until memory is available. */
2093 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
2099 /* Reserve space for headers and prepare control bits. */
2100 skb_reserve(skb
, MAX_TCP_HEADER
);
2101 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2102 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2103 TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2104 tcp_queue_skb(sk
, skb
);
2106 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2109 /* We get here when a process closes a file descriptor (either due to
2110 * an explicit close() or as a byproduct of exit()'ing) and there
2111 * was unread data in the receive queue. This behavior is recommended
2112 * by RFC 2525, section 2.17. -DaveM
2114 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2116 struct sk_buff
*skb
;
2118 /* NOTE: No TCP options attached and we never retransmit this. */
2119 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2121 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2125 /* Reserve space for headers and prepare control bits. */
2126 skb_reserve(skb
, MAX_TCP_HEADER
);
2127 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2128 TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2130 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2131 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2132 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2134 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2137 /* WARNING: This routine must only be called when we have already sent
2138 * a SYN packet that crossed the incoming SYN that caused this routine
2139 * to get called. If this assumption fails then the initial rcv_wnd
2140 * and rcv_wscale values will not be correct.
2142 int tcp_send_synack(struct sock
*sk
)
2144 struct sk_buff
*skb
;
2146 skb
= tcp_write_queue_head(sk
);
2147 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
)) {
2148 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2151 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_ACK
)) {
2152 if (skb_cloned(skb
)) {
2153 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2156 tcp_unlink_write_queue(skb
, sk
);
2157 skb_header_release(nskb
);
2158 __tcp_add_write_queue_head(sk
, nskb
);
2159 sk_wmem_free_skb(sk
, skb
);
2160 sk
->sk_wmem_queued
+= nskb
->truesize
;
2161 sk_mem_charge(sk
, nskb
->truesize
);
2165 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2166 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2168 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2169 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2173 * Prepare a SYN-ACK.
2175 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2176 struct request_sock
*req
)
2178 struct inet_request_sock
*ireq
= inet_rsk(req
);
2179 struct tcp_sock
*tp
= tcp_sk(sk
);
2181 int tcp_header_size
;
2182 struct sk_buff
*skb
;
2183 #ifdef CONFIG_TCP_MD5SIG
2184 struct tcp_md5sig_key
*md5
;
2185 __u8
*md5_hash_location
;
2188 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2192 /* Reserve space for headers. */
2193 skb_reserve(skb
, MAX_TCP_HEADER
);
2195 skb
->dst
= dst_clone(dst
);
2197 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2198 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2199 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2200 /* SACK_PERM is in the place of NOP NOP of TS */
2201 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2203 #ifdef CONFIG_TCP_MD5SIG
2204 /* Are we doing MD5 on this segment? If so - make room for it */
2205 md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
2207 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
2209 skb_push(skb
, tcp_header_size
);
2210 skb_reset_transport_header(skb
);
2213 memset(th
, 0, sizeof(struct tcphdr
));
2216 TCP_ECN_make_synack(req
, th
);
2217 th
->source
= inet_sk(sk
)->sport
;
2218 th
->dest
= ireq
->rmt_port
;
2219 /* Setting of flags are superfluous here for callers (and ECE is
2220 * not even correctly set)
2222 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2223 TCPCB_FLAG_SYN
| TCPCB_FLAG_ACK
);
2224 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2225 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2226 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2228 /* Set this up on the first call only */
2229 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2230 /* tcp_full_space because it is guaranteed to be the first packet */
2231 tcp_select_initial_window(tcp_full_space(sk
),
2232 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2237 ireq
->rcv_wscale
= rcv_wscale
;
2240 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2241 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2242 #ifdef CONFIG_SYN_COOKIES
2243 if (unlikely(req
->cookie_ts
))
2244 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2247 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2248 tcp_syn_build_options((__be32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2249 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2250 TCP_SKB_CB(skb
)->when
,
2253 #ifdef CONFIG_TCP_MD5SIG
2254 md5
? &md5_hash_location
:
2259 th
->doff
= (tcp_header_size
>> 2);
2260 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
2262 #ifdef CONFIG_TCP_MD5SIG
2263 /* Okay, we have all we need - do the md5 hash if needed */
2265 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
2266 md5
, NULL
, req
, skb
);
2274 * Do all connect socket setups that can be done AF independent.
2276 static void tcp_connect_init(struct sock
*sk
)
2278 struct dst_entry
*dst
= __sk_dst_get(sk
);
2279 struct tcp_sock
*tp
= tcp_sk(sk
);
2282 /* We'll fix this up when we get a response from the other end.
2283 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2285 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2286 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2288 #ifdef CONFIG_TCP_MD5SIG
2289 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2290 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2293 /* If user gave his TCP_MAXSEG, record it to clamp */
2294 if (tp
->rx_opt
.user_mss
)
2295 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2298 tcp_sync_mss(sk
, dst_mtu(dst
));
2300 if (!tp
->window_clamp
)
2301 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2302 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2303 tcp_initialize_rcv_mss(sk
);
2305 tcp_select_initial_window(tcp_full_space(sk
),
2306 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2309 sysctl_tcp_window_scaling
,
2312 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2313 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2316 sock_reset_flag(sk
, SOCK_DONE
);
2318 tcp_init_wl(tp
, tp
->write_seq
, 0);
2319 tp
->snd_una
= tp
->write_seq
;
2320 tp
->snd_sml
= tp
->write_seq
;
2325 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2326 inet_csk(sk
)->icsk_retransmits
= 0;
2327 tcp_clear_retrans(tp
);
2331 * Build a SYN and send it off.
2333 int tcp_connect(struct sock
*sk
)
2335 struct tcp_sock
*tp
= tcp_sk(sk
);
2336 struct sk_buff
*buff
;
2338 tcp_connect_init(sk
);
2340 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2341 if (unlikely(buff
== NULL
))
2344 /* Reserve space for headers. */
2345 skb_reserve(buff
, MAX_TCP_HEADER
);
2347 tp
->snd_nxt
= tp
->write_seq
;
2348 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPCB_FLAG_SYN
);
2349 TCP_ECN_send_syn(sk
, buff
);
2352 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2353 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2354 skb_header_release(buff
);
2355 __tcp_add_write_queue_tail(sk
, buff
);
2356 sk
->sk_wmem_queued
+= buff
->truesize
;
2357 sk_mem_charge(sk
, buff
->truesize
);
2358 tp
->packets_out
+= tcp_skb_pcount(buff
);
2359 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2361 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2362 * in order to make this packet get counted in tcpOutSegs.
2364 tp
->snd_nxt
= tp
->write_seq
;
2365 tp
->pushed_seq
= tp
->write_seq
;
2366 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2368 /* Timer for repeating the SYN until an answer. */
2369 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2370 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2374 /* Send out a delayed ack, the caller does the policy checking
2375 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2378 void tcp_send_delayed_ack(struct sock
*sk
)
2380 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2381 int ato
= icsk
->icsk_ack
.ato
;
2382 unsigned long timeout
;
2384 if (ato
> TCP_DELACK_MIN
) {
2385 const struct tcp_sock
*tp
= tcp_sk(sk
);
2386 int max_ato
= HZ
/ 2;
2388 if (icsk
->icsk_ack
.pingpong
||
2389 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2390 max_ato
= TCP_DELACK_MAX
;
2392 /* Slow path, intersegment interval is "high". */
2394 /* If some rtt estimate is known, use it to bound delayed ack.
2395 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2399 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
2405 ato
= min(ato
, max_ato
);
2408 /* Stay within the limit we were given */
2409 timeout
= jiffies
+ ato
;
2411 /* Use new timeout only if there wasn't a older one earlier. */
2412 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2413 /* If delack timer was blocked or is about to expire,
2416 if (icsk
->icsk_ack
.blocked
||
2417 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2422 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2423 timeout
= icsk
->icsk_ack
.timeout
;
2425 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2426 icsk
->icsk_ack
.timeout
= timeout
;
2427 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2430 /* This routine sends an ack and also updates the window. */
2431 void tcp_send_ack(struct sock
*sk
)
2433 struct sk_buff
*buff
;
2435 /* If we have been reset, we may not send again. */
2436 if (sk
->sk_state
== TCP_CLOSE
)
2439 /* We are not putting this on the write queue, so
2440 * tcp_transmit_skb() will set the ownership to this
2443 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2445 inet_csk_schedule_ack(sk
);
2446 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2447 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2448 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2452 /* Reserve space for headers and prepare control bits. */
2453 skb_reserve(buff
, MAX_TCP_HEADER
);
2454 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPCB_FLAG_ACK
);
2456 /* Send it off, this clears delayed acks for us. */
2457 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2458 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2461 /* This routine sends a packet with an out of date sequence
2462 * number. It assumes the other end will try to ack it.
2464 * Question: what should we make while urgent mode?
2465 * 4.4BSD forces sending single byte of data. We cannot send
2466 * out of window data, because we have SND.NXT==SND.MAX...
2468 * Current solution: to send TWO zero-length segments in urgent mode:
2469 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2470 * out-of-date with SND.UNA-1 to probe window.
2472 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2474 struct tcp_sock
*tp
= tcp_sk(sk
);
2475 struct sk_buff
*skb
;
2477 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2478 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2482 /* Reserve space for headers and set control bits. */
2483 skb_reserve(skb
, MAX_TCP_HEADER
);
2484 /* Use a previous sequence. This should cause the other
2485 * end to send an ack. Don't queue or clone SKB, just
2488 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPCB_FLAG_ACK
);
2489 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2490 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2493 int tcp_write_wakeup(struct sock
*sk
)
2495 struct tcp_sock
*tp
= tcp_sk(sk
);
2496 struct sk_buff
*skb
;
2498 if (sk
->sk_state
== TCP_CLOSE
)
2501 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2502 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
2504 unsigned int mss
= tcp_current_mss(sk
, 0);
2505 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2507 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2508 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2510 /* We are probing the opening of a window
2511 * but the window size is != 0
2512 * must have been a result SWS avoidance ( sender )
2514 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2516 seg_size
= min(seg_size
, mss
);
2517 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2518 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2520 } else if (!tcp_skb_pcount(skb
))
2521 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2523 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2524 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2525 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2527 tcp_event_new_data_sent(sk
, skb
);
2531 between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
2532 tcp_xmit_probe_skb(sk
, 1);
2533 return tcp_xmit_probe_skb(sk
, 0);
2537 /* A window probe timeout has occurred. If window is not closed send
2538 * a partial packet else a zero probe.
2540 void tcp_send_probe0(struct sock
*sk
)
2542 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2543 struct tcp_sock
*tp
= tcp_sk(sk
);
2546 err
= tcp_write_wakeup(sk
);
2548 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2549 /* Cancel probe timer, if it is not required. */
2550 icsk
->icsk_probes_out
= 0;
2551 icsk
->icsk_backoff
= 0;
2556 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2557 icsk
->icsk_backoff
++;
2558 icsk
->icsk_probes_out
++;
2559 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2560 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2563 /* If packet was not sent due to local congestion,
2564 * do not backoff and do not remember icsk_probes_out.
2565 * Let local senders to fight for local resources.
2567 * Use accumulated backoff yet.
2569 if (!icsk
->icsk_probes_out
)
2570 icsk
->icsk_probes_out
= 1;
2571 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2572 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2573 TCP_RESOURCE_PROBE_INTERVAL
),
2578 EXPORT_SYMBOL(tcp_select_initial_window
);
2579 EXPORT_SYMBOL(tcp_connect
);
2580 EXPORT_SYMBOL(tcp_make_synack
);
2581 EXPORT_SYMBOL(tcp_simple_retransmit
);
2582 EXPORT_SYMBOL(tcp_sync_mss
);
2583 EXPORT_SYMBOL(tcp_mtup_init
);