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>
43 #include <linux/smp_lock.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
53 /* This limits the percentage of the congestion window which we
54 * will allow a single TSO frame to consume. Building TSO frames
55 * which are too large can cause TCP streams to be bursty.
57 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
59 int sysctl_tcp_mtu_probing __read_mostly
= 0;
60 int sysctl_tcp_base_mss __read_mostly
= 512;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
65 static void update_send_head(struct sock
*sk
, struct tcp_sock
*tp
,
68 sk
->sk_send_head
= skb
->next
;
69 if (sk
->sk_send_head
== (struct sk_buff
*)&sk
->sk_write_queue
)
70 sk
->sk_send_head
= NULL
;
71 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
72 tcp_packets_out_inc(sk
, tp
, skb
);
75 /* SND.NXT, if window was not shrunk.
76 * If window has been shrunk, what should we make? It is not clear at all.
77 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
78 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
79 * invalid. OK, let's make this for now:
81 static inline __u32
tcp_acceptable_seq(struct sock
*sk
, struct tcp_sock
*tp
)
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 void tcp_build_and_update_options(__be32
*ptr
, struct tcp_sock
*tp
,
275 if (tp
->rx_opt
.tstamp_ok
) {
276 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
278 (TCPOPT_TIMESTAMP
<< 8) |
280 *ptr
++ = htonl(tstamp
);
281 *ptr
++ = htonl(tp
->rx_opt
.ts_recent
);
283 if (tp
->rx_opt
.eff_sacks
) {
284 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
? tp
->duplicate_sack
: tp
->selective_acks
;
287 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
290 (TCPOLEN_SACK_BASE
+ (tp
->rx_opt
.eff_sacks
*
291 TCPOLEN_SACK_PERBLOCK
)));
292 for(this_sack
= 0; this_sack
< tp
->rx_opt
.eff_sacks
; this_sack
++) {
293 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
294 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
296 if (tp
->rx_opt
.dsack
) {
297 tp
->rx_opt
.dsack
= 0;
298 tp
->rx_opt
.eff_sacks
--;
303 /* Construct a tcp options header for a SYN or SYN_ACK packet.
304 * If this is every changed make sure to change the definition of
305 * MAX_SYN_SIZE to match the new maximum number of options that you
308 static void tcp_syn_build_options(__be32
*ptr
, int mss
, int ts
, int sack
,
309 int offer_wscale
, int wscale
, __u32 tstamp
,
312 /* We always get an MSS option.
313 * The option bytes which will be seen in normal data
314 * packets should timestamps be used, must be in the MSS
315 * advertised. But we subtract them from tp->mss_cache so
316 * that calculations in tcp_sendmsg are simpler etc.
317 * So account for this fact here if necessary. If we
318 * don't do this correctly, as a receiver we won't
319 * recognize data packets as being full sized when we
320 * should, and thus we won't abide by the delayed ACK
322 * SACKs don't matter, we never delay an ACK when we
323 * have any of those going out.
325 *ptr
++ = htonl((TCPOPT_MSS
<< 24) | (TCPOLEN_MSS
<< 16) | mss
);
328 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
329 (TCPOLEN_SACK_PERM
<< 16) |
330 (TCPOPT_TIMESTAMP
<< 8) |
333 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
335 (TCPOPT_TIMESTAMP
<< 8) |
337 *ptr
++ = htonl(tstamp
); /* TSVAL */
338 *ptr
++ = htonl(ts_recent
); /* TSECR */
340 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
342 (TCPOPT_SACK_PERM
<< 8) |
345 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
346 (TCPOPT_WINDOW
<< 16) |
347 (TCPOLEN_WINDOW
<< 8) |
351 /* This routine actually transmits TCP packets queued in by
352 * tcp_do_sendmsg(). This is used by both the initial
353 * transmission and possible later retransmissions.
354 * All SKB's seen here are completely headerless. It is our
355 * job to build the TCP header, and pass the packet down to
356 * IP so it can do the same plus pass the packet off to the
359 * We are working here with either a clone of the original
360 * SKB, or a fresh unique copy made by the retransmit engine.
362 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
, gfp_t gfp_mask
)
364 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
365 struct inet_sock
*inet
;
367 struct tcp_skb_cb
*tcb
;
373 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
375 /* If congestion control is doing timestamping, we must
376 * take such a timestamp before we potentially clone/copy.
378 if (icsk
->icsk_ca_ops
->rtt_sample
)
379 __net_timestamp(skb
);
381 if (likely(clone_it
)) {
382 if (unlikely(skb_cloned(skb
)))
383 skb
= pskb_copy(skb
, gfp_mask
);
385 skb
= skb_clone(skb
, gfp_mask
);
392 tcb
= TCP_SKB_CB(skb
);
393 tcp_header_size
= tp
->tcp_header_len
;
395 #define SYSCTL_FLAG_TSTAMPS 0x1
396 #define SYSCTL_FLAG_WSCALE 0x2
397 #define SYSCTL_FLAG_SACK 0x4
400 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
401 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
402 if(sysctl_tcp_timestamps
) {
403 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
404 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
406 if (sysctl_tcp_window_scaling
) {
407 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
408 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
410 if (sysctl_tcp_sack
) {
411 sysctl_flags
|= SYSCTL_FLAG_SACK
;
412 if (!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
413 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
415 } else if (unlikely(tp
->rx_opt
.eff_sacks
)) {
416 /* A SACK is 2 pad bytes, a 2 byte header, plus
417 * 2 32-bit sequence numbers for each SACK block.
419 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
420 (tp
->rx_opt
.eff_sacks
*
421 TCPOLEN_SACK_PERBLOCK
));
424 if (tcp_packets_in_flight(tp
) == 0)
425 tcp_ca_event(sk
, CA_EVENT_TX_START
);
427 th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
430 /* Build TCP header and checksum it. */
431 th
->source
= inet
->sport
;
432 th
->dest
= inet
->dport
;
433 th
->seq
= htonl(tcb
->seq
);
434 th
->ack_seq
= htonl(tp
->rcv_nxt
);
435 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
438 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
439 /* RFC1323: The window in SYN & SYN/ACK segments
442 th
->window
= htons(tp
->rcv_wnd
);
444 th
->window
= htons(tcp_select_window(sk
));
449 if (unlikely(tp
->urg_mode
&&
450 between(tp
->snd_up
, tcb
->seq
+1, tcb
->seq
+0xFFFF))) {
451 th
->urg_ptr
= htons(tp
->snd_up
-tcb
->seq
);
455 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
456 tcp_syn_build_options((__be32
*)(th
+ 1),
457 tcp_advertise_mss(sk
),
458 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
459 (sysctl_flags
& SYSCTL_FLAG_SACK
),
460 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
461 tp
->rx_opt
.rcv_wscale
,
463 tp
->rx_opt
.ts_recent
);
465 tcp_build_and_update_options((__be32
*)(th
+ 1),
467 TCP_ECN_send(sk
, tp
, skb
, tcp_header_size
);
470 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
472 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
473 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
475 if (skb
->len
!= tcp_header_size
)
476 tcp_event_data_sent(tp
, skb
, sk
);
478 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
479 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
481 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, sk
, 0);
482 if (likely(err
<= 0))
487 return net_xmit_eval(err
);
489 #undef SYSCTL_FLAG_TSTAMPS
490 #undef SYSCTL_FLAG_WSCALE
491 #undef SYSCTL_FLAG_SACK
495 /* This routine just queue's the buffer
497 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
498 * otherwise socket can stall.
500 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
502 struct tcp_sock
*tp
= tcp_sk(sk
);
504 /* Advance write_seq and place onto the write_queue. */
505 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
506 skb_header_release(skb
);
507 __skb_queue_tail(&sk
->sk_write_queue
, skb
);
508 sk_charge_skb(sk
, skb
);
510 /* Queue it, remembering where we must start sending. */
511 if (sk
->sk_send_head
== NULL
)
512 sk
->sk_send_head
= skb
;
515 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
517 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
)) {
518 /* Avoid the costly divide in the normal
521 skb_shinfo(skb
)->gso_segs
= 1;
522 skb_shinfo(skb
)->gso_size
= 0;
523 skb_shinfo(skb
)->gso_type
= 0;
527 factor
= skb
->len
+ (mss_now
- 1);
529 skb_shinfo(skb
)->gso_segs
= factor
;
530 skb_shinfo(skb
)->gso_size
= mss_now
;
531 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
535 /* Function to create two new TCP segments. Shrinks the given segment
536 * to the specified size and appends a new segment with the rest of the
537 * packet to the list. This won't be called frequently, I hope.
538 * Remember, these are still headerless SKBs at this point.
540 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
, unsigned int mss_now
)
542 struct tcp_sock
*tp
= tcp_sk(sk
);
543 struct sk_buff
*buff
;
544 int nsize
, old_factor
;
548 BUG_ON(len
> skb
->len
);
550 clear_all_retrans_hints(tp
);
551 nsize
= skb_headlen(skb
) - len
;
555 if (skb_cloned(skb
) &&
556 skb_is_nonlinear(skb
) &&
557 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
560 /* Get a new skb... force flag on. */
561 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
563 return -ENOMEM
; /* We'll just try again later. */
565 sk_charge_skb(sk
, buff
);
566 nlen
= skb
->len
- len
- nsize
;
567 buff
->truesize
+= nlen
;
568 skb
->truesize
-= nlen
;
570 /* Correct the sequence numbers. */
571 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
572 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
573 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
575 /* PSH and FIN should only be set in the second packet. */
576 flags
= TCP_SKB_CB(skb
)->flags
;
577 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
578 TCP_SKB_CB(buff
)->flags
= flags
;
579 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
580 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
582 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
583 /* Copy and checksum data tail into the new buffer. */
584 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
589 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
591 skb
->ip_summed
= CHECKSUM_PARTIAL
;
592 skb_split(skb
, buff
, len
);
595 buff
->ip_summed
= skb
->ip_summed
;
597 /* Looks stupid, but our code really uses when of
598 * skbs, which it never sent before. --ANK
600 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
601 buff
->tstamp
= skb
->tstamp
;
603 old_factor
= tcp_skb_pcount(skb
);
605 /* Fix up tso_factor for both original and new SKB. */
606 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
607 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
609 /* If this packet has been sent out already, we must
610 * adjust the various packet counters.
612 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
613 int diff
= old_factor
- tcp_skb_pcount(skb
) -
614 tcp_skb_pcount(buff
);
616 tp
->packets_out
-= diff
;
618 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
619 tp
->sacked_out
-= diff
;
620 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
621 tp
->retrans_out
-= diff
;
623 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
624 tp
->lost_out
-= diff
;
625 tp
->left_out
-= diff
;
629 /* Adjust Reno SACK estimate. */
630 if (!tp
->rx_opt
.sack_ok
) {
631 tp
->sacked_out
-= diff
;
632 if ((int)tp
->sacked_out
< 0)
634 tcp_sync_left_out(tp
);
637 tp
->fackets_out
-= diff
;
638 if ((int)tp
->fackets_out
< 0)
643 /* Link BUFF into the send queue. */
644 skb_header_release(buff
);
645 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
650 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
651 * eventually). The difference is that pulled data not copied, but
652 * immediately discarded.
654 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
660 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
661 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
662 put_page(skb_shinfo(skb
)->frags
[i
].page
);
663 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
665 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
667 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
668 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
674 skb_shinfo(skb
)->nr_frags
= k
;
676 skb
->tail
= skb
->data
;
677 skb
->data_len
-= len
;
678 skb
->len
= skb
->data_len
;
681 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
683 if (skb_cloned(skb
) &&
684 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
687 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
688 if (unlikely(len
< skb_headlen(skb
)))
689 __skb_pull(skb
, len
);
691 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
693 TCP_SKB_CB(skb
)->seq
+= len
;
694 skb
->ip_summed
= CHECKSUM_PARTIAL
;
696 skb
->truesize
-= len
;
697 sk
->sk_wmem_queued
-= len
;
698 sk
->sk_forward_alloc
+= len
;
699 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
701 /* Any change of skb->len requires recalculation of tso
704 if (tcp_skb_pcount(skb
) > 1)
705 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
710 /* Not accounting for SACKs here. */
711 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
713 struct tcp_sock
*tp
= tcp_sk(sk
);
714 struct inet_connection_sock
*icsk
= inet_csk(sk
);
717 /* Calculate base mss without TCP options:
718 It is MMS_S - sizeof(tcphdr) of rfc1122
720 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
722 /* Clamp it (mss_clamp does not include tcp options) */
723 if (mss_now
> tp
->rx_opt
.mss_clamp
)
724 mss_now
= tp
->rx_opt
.mss_clamp
;
726 /* Now subtract optional transport overhead */
727 mss_now
-= icsk
->icsk_ext_hdr_len
;
729 /* Then reserve room for full set of TCP options and 8 bytes of data */
733 /* Now subtract TCP options size, not including SACKs */
734 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
739 /* Inverse of above */
740 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
742 struct tcp_sock
*tp
= tcp_sk(sk
);
743 struct inet_connection_sock
*icsk
= inet_csk(sk
);
748 icsk
->icsk_ext_hdr_len
+
749 icsk
->icsk_af_ops
->net_header_len
;
754 void tcp_mtup_init(struct sock
*sk
)
756 struct tcp_sock
*tp
= tcp_sk(sk
);
757 struct inet_connection_sock
*icsk
= inet_csk(sk
);
759 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
760 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
761 icsk
->icsk_af_ops
->net_header_len
;
762 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
763 icsk
->icsk_mtup
.probe_size
= 0;
766 /* This function synchronize snd mss to current pmtu/exthdr set.
768 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
769 for TCP options, but includes only bare TCP header.
771 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
772 It is minimum of user_mss and mss received with SYN.
773 It also does not include TCP options.
775 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
777 tp->mss_cache is current effective sending mss, including
778 all tcp options except for SACKs. It is evaluated,
779 taking into account current pmtu, but never exceeds
780 tp->rx_opt.mss_clamp.
782 NOTE1. rfc1122 clearly states that advertised MSS
783 DOES NOT include either tcp or ip options.
785 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
786 are READ ONLY outside this function. --ANK (980731)
789 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
791 struct tcp_sock
*tp
= tcp_sk(sk
);
792 struct inet_connection_sock
*icsk
= inet_csk(sk
);
795 if (icsk
->icsk_mtup
.search_high
> pmtu
)
796 icsk
->icsk_mtup
.search_high
= pmtu
;
798 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
800 /* Bound mss with half of window */
801 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
802 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
804 /* And store cached results */
805 icsk
->icsk_pmtu_cookie
= pmtu
;
806 if (icsk
->icsk_mtup
.enabled
)
807 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
808 tp
->mss_cache
= mss_now
;
813 /* Compute the current effective MSS, taking SACKs and IP options,
814 * and even PMTU discovery events into account.
816 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
817 * cannot be large. However, taking into account rare use of URG, this
820 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
822 struct tcp_sock
*tp
= tcp_sk(sk
);
823 struct dst_entry
*dst
= __sk_dst_get(sk
);
828 mss_now
= tp
->mss_cache
;
830 if (large_allowed
&& sk_can_gso(sk
) && !tp
->urg_mode
)
834 u32 mtu
= dst_mtu(dst
);
835 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
836 mss_now
= tcp_sync_mss(sk
, mtu
);
839 if (tp
->rx_opt
.eff_sacks
)
840 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
841 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
843 xmit_size_goal
= mss_now
;
846 xmit_size_goal
= (65535 -
847 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
848 inet_csk(sk
)->icsk_ext_hdr_len
-
851 if (tp
->max_window
&&
852 (xmit_size_goal
> (tp
->max_window
>> 1)))
853 xmit_size_goal
= max((tp
->max_window
>> 1),
854 68U - tp
->tcp_header_len
);
856 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
858 tp
->xmit_size_goal
= xmit_size_goal
;
863 /* Congestion window validation. (RFC2861) */
865 static void tcp_cwnd_validate(struct sock
*sk
, struct tcp_sock
*tp
)
867 __u32 packets_out
= tp
->packets_out
;
869 if (packets_out
>= tp
->snd_cwnd
) {
870 /* Network is feed fully. */
871 tp
->snd_cwnd_used
= 0;
872 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
874 /* Network starves. */
875 if (tp
->packets_out
> tp
->snd_cwnd_used
)
876 tp
->snd_cwnd_used
= tp
->packets_out
;
878 if ((s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
879 tcp_cwnd_application_limited(sk
);
883 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
885 u32 window
, cwnd_len
;
887 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
888 cwnd_len
= mss_now
* cwnd
;
889 return min(window
, cwnd_len
);
892 /* Can at least one segment of SKB be sent right now, according to the
893 * congestion window rules? If so, return how many segments are allowed.
895 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
899 /* Don't be strict about the congestion window for the final FIN. */
900 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
903 in_flight
= tcp_packets_in_flight(tp
);
905 if (in_flight
< cwnd
)
906 return (cwnd
- in_flight
);
911 /* This must be invoked the first time we consider transmitting
914 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
916 int tso_segs
= tcp_skb_pcount(skb
);
920 tcp_skb_mss(skb
) != mss_now
)) {
921 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
922 tso_segs
= tcp_skb_pcount(skb
);
927 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
929 return after(tp
->snd_sml
,tp
->snd_una
) &&
930 !after(tp
->snd_sml
, tp
->snd_nxt
);
933 /* Return 0, if packet can be sent now without violation Nagle's rules:
934 * 1. It is full sized.
935 * 2. Or it contains FIN. (already checked by caller)
936 * 3. Or TCP_NODELAY was set.
937 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
938 * With Minshall's modification: all sent small packets are ACKed.
941 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
942 const struct sk_buff
*skb
,
943 unsigned mss_now
, int nonagle
)
945 return (skb
->len
< mss_now
&&
946 ((nonagle
&TCP_NAGLE_CORK
) ||
949 tcp_minshall_check(tp
))));
952 /* Return non-zero if the Nagle test allows this packet to be
955 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
956 unsigned int cur_mss
, int nonagle
)
958 /* Nagle rule does not apply to frames, which sit in the middle of the
959 * write_queue (they have no chances to get new data).
961 * This is implemented in the callers, where they modify the 'nonagle'
962 * argument based upon the location of SKB in the send queue.
964 if (nonagle
& TCP_NAGLE_PUSH
)
967 /* Don't use the nagle rule for urgent data (or for the final FIN). */
969 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
972 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
978 /* Does at least the first segment of SKB fit into the send window? */
979 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
981 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
983 if (skb
->len
> cur_mss
)
984 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
986 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
989 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
990 * should be put on the wire right now. If so, it returns the number of
991 * packets allowed by the congestion window.
993 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
994 unsigned int cur_mss
, int nonagle
)
996 struct tcp_sock
*tp
= tcp_sk(sk
);
997 unsigned int cwnd_quota
;
999 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1001 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1004 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1006 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1012 static inline int tcp_skb_is_last(const struct sock
*sk
,
1013 const struct sk_buff
*skb
)
1015 return skb
->next
== (struct sk_buff
*)&sk
->sk_write_queue
;
1018 int tcp_may_send_now(struct sock
*sk
, struct tcp_sock
*tp
)
1020 struct sk_buff
*skb
= sk
->sk_send_head
;
1023 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1024 (tcp_skb_is_last(sk
, skb
) ?
1029 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1030 * which is put after SKB on the list. It is very much like
1031 * tcp_fragment() except that it may make several kinds of assumptions
1032 * in order to speed up the splitting operation. In particular, we
1033 * know that all the data is in scatter-gather pages, and that the
1034 * packet has never been sent out before (and thus is not cloned).
1036 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
1038 struct sk_buff
*buff
;
1039 int nlen
= skb
->len
- len
;
1042 /* All of a TSO frame must be composed of paged data. */
1043 if (skb
->len
!= skb
->data_len
)
1044 return tcp_fragment(sk
, skb
, len
, mss_now
);
1046 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
1047 if (unlikely(buff
== NULL
))
1050 sk_charge_skb(sk
, buff
);
1051 buff
->truesize
+= nlen
;
1052 skb
->truesize
-= nlen
;
1054 /* Correct the sequence numbers. */
1055 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1056 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1057 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1059 /* PSH and FIN should only be set in the second packet. */
1060 flags
= TCP_SKB_CB(skb
)->flags
;
1061 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1062 TCP_SKB_CB(buff
)->flags
= flags
;
1064 /* This packet was never sent out yet, so no SACK bits. */
1065 TCP_SKB_CB(buff
)->sacked
= 0;
1067 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1068 skb_split(skb
, buff
, len
);
1070 /* Fix up tso_factor for both original and new SKB. */
1071 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1072 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1074 /* Link BUFF into the send queue. */
1075 skb_header_release(buff
);
1076 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
1081 /* Try to defer sending, if possible, in order to minimize the amount
1082 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1084 * This algorithm is from John Heffner.
1086 static int tcp_tso_should_defer(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
1088 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1089 u32 send_win
, cong_win
, limit
, in_flight
;
1091 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1094 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1097 /* Defer for less than two clock ticks. */
1098 if (!tp
->tso_deferred
&& ((jiffies
<<1)>>1) - (tp
->tso_deferred
>>1) > 1)
1101 in_flight
= tcp_packets_in_flight(tp
);
1103 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
1104 (tp
->snd_cwnd
<= in_flight
));
1106 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
1108 /* From in_flight test above, we know that cwnd > in_flight. */
1109 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1111 limit
= min(send_win
, cong_win
);
1113 /* If a full-sized TSO skb can be sent, do it. */
1117 if (sysctl_tcp_tso_win_divisor
) {
1118 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1120 /* If at least some fraction of a window is available,
1123 chunk
/= sysctl_tcp_tso_win_divisor
;
1127 /* Different approach, try not to defer past a single
1128 * ACK. Receiver should ACK every other full sized
1129 * frame, so if we have space for more than 3 frames
1132 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1136 /* Ok, it looks like it is advisable to defer. */
1137 tp
->tso_deferred
= 1 | (jiffies
<<1);
1142 tp
->tso_deferred
= 0;
1146 /* Create a new MTU probe if we are ready.
1147 * Returns 0 if we should wait to probe (no cwnd available),
1148 * 1 if a probe was sent,
1150 static int tcp_mtu_probe(struct sock
*sk
)
1152 struct tcp_sock
*tp
= tcp_sk(sk
);
1153 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1154 struct sk_buff
*skb
, *nskb
, *next
;
1161 /* Not currently probing/verifying,
1163 * have enough cwnd, and
1164 * not SACKing (the variable headers throw things off) */
1165 if (!icsk
->icsk_mtup
.enabled
||
1166 icsk
->icsk_mtup
.probe_size
||
1167 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1168 tp
->snd_cwnd
< 11 ||
1169 tp
->rx_opt
.eff_sacks
)
1172 /* Very simple search strategy: just double the MSS. */
1173 mss_now
= tcp_current_mss(sk
, 0);
1174 probe_size
= 2*tp
->mss_cache
;
1175 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1176 /* TODO: set timer for probe_converge_event */
1180 /* Have enough data in the send queue to probe? */
1182 if ((skb
= sk
->sk_send_head
) == NULL
)
1184 while ((len
+= skb
->len
) < probe_size
&& !tcp_skb_is_last(sk
, skb
))
1186 if (len
< probe_size
)
1189 /* Receive window check. */
1190 if (after(TCP_SKB_CB(skb
)->seq
+ probe_size
, tp
->snd_una
+ tp
->snd_wnd
)) {
1191 if (tp
->snd_wnd
< probe_size
)
1197 /* Do we need to wait to drain cwnd? */
1198 pif
= tcp_packets_in_flight(tp
);
1199 if (pif
+ 2 > tp
->snd_cwnd
) {
1200 /* With no packets in flight, don't stall. */
1207 /* We're allowed to probe. Build it now. */
1208 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1210 sk_charge_skb(sk
, nskb
);
1212 skb
= sk
->sk_send_head
;
1213 __skb_insert(nskb
, skb
->prev
, skb
, &sk
->sk_write_queue
);
1214 sk
->sk_send_head
= nskb
;
1216 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1217 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1218 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1219 TCP_SKB_CB(nskb
)->sacked
= 0;
1221 nskb
->ip_summed
= skb
->ip_summed
;
1224 while (len
< probe_size
) {
1227 copy
= min_t(int, skb
->len
, probe_size
- len
);
1228 if (nskb
->ip_summed
)
1229 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1231 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1232 skb_put(nskb
, copy
), copy
, nskb
->csum
);
1234 if (skb
->len
<= copy
) {
1235 /* We've eaten all the data from this skb.
1237 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1238 __skb_unlink(skb
, &sk
->sk_write_queue
);
1239 sk_stream_free_skb(sk
, skb
);
1241 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1242 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1243 if (!skb_shinfo(skb
)->nr_frags
) {
1244 skb_pull(skb
, copy
);
1245 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1246 skb
->csum
= csum_partial(skb
->data
, skb
->len
, 0);
1248 __pskb_trim_head(skb
, copy
);
1249 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1251 TCP_SKB_CB(skb
)->seq
+= copy
;
1257 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1259 /* We're ready to send. If this fails, the probe will
1260 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1261 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1262 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1263 /* Decrement cwnd here because we are sending
1264 * effectively two packets. */
1266 update_send_head(sk
, tp
, nskb
);
1268 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1269 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1270 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1279 /* This routine writes packets to the network. It advances the
1280 * send_head. This happens as incoming acks open up the remote
1283 * Returns 1, if no segments are in flight and we have queued segments, but
1284 * cannot send anything now because of SWS or another problem.
1286 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1288 struct tcp_sock
*tp
= tcp_sk(sk
);
1289 struct sk_buff
*skb
;
1290 unsigned int tso_segs
, sent_pkts
;
1294 /* If we are closed, the bytes will have to remain here.
1295 * In time closedown will finish, we empty the write queue and all
1298 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1303 /* Do MTU probing. */
1304 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1306 } else if (result
> 0) {
1310 while ((skb
= sk
->sk_send_head
)) {
1313 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1316 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1320 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1323 if (tso_segs
== 1) {
1324 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1325 (tcp_skb_is_last(sk
, skb
) ?
1326 nonagle
: TCP_NAGLE_PUSH
))))
1329 if (tcp_tso_should_defer(sk
, tp
, skb
))
1335 limit
= tcp_window_allows(tp
, skb
,
1336 mss_now
, cwnd_quota
);
1338 if (skb
->len
< limit
) {
1339 unsigned int trim
= skb
->len
% mss_now
;
1342 limit
= skb
->len
- trim
;
1346 if (skb
->len
> limit
&&
1347 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1350 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1352 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1355 /* Advance the send_head. This one is sent out.
1356 * This call will increment packets_out.
1358 update_send_head(sk
, tp
, skb
);
1360 tcp_minshall_update(tp
, mss_now
, skb
);
1364 if (likely(sent_pkts
)) {
1365 tcp_cwnd_validate(sk
, tp
);
1368 return !tp
->packets_out
&& sk
->sk_send_head
;
1371 /* Push out any pending frames which were held back due to
1372 * TCP_CORK or attempt at coalescing tiny packets.
1373 * The socket must be locked by the caller.
1375 void __tcp_push_pending_frames(struct sock
*sk
, struct tcp_sock
*tp
,
1376 unsigned int cur_mss
, int nonagle
)
1378 struct sk_buff
*skb
= sk
->sk_send_head
;
1381 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1382 tcp_check_probe_timer(sk
, tp
);
1386 /* Send _single_ skb sitting at the send head. This function requires
1387 * true push pending frames to setup probe timer etc.
1389 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1391 struct tcp_sock
*tp
= tcp_sk(sk
);
1392 struct sk_buff
*skb
= sk
->sk_send_head
;
1393 unsigned int tso_segs
, cwnd_quota
;
1395 BUG_ON(!skb
|| skb
->len
< mss_now
);
1397 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1398 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1400 if (likely(cwnd_quota
)) {
1407 limit
= tcp_window_allows(tp
, skb
,
1408 mss_now
, cwnd_quota
);
1410 if (skb
->len
< limit
) {
1411 unsigned int trim
= skb
->len
% mss_now
;
1414 limit
= skb
->len
- trim
;
1418 if (skb
->len
> limit
&&
1419 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1422 /* Send it out now. */
1423 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1425 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1426 update_send_head(sk
, tp
, skb
);
1427 tcp_cwnd_validate(sk
, tp
);
1433 /* This function returns the amount that we can raise the
1434 * usable window based on the following constraints
1436 * 1. The window can never be shrunk once it is offered (RFC 793)
1437 * 2. We limit memory per socket
1440 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1441 * RECV.NEXT + RCV.WIN fixed until:
1442 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1444 * i.e. don't raise the right edge of the window until you can raise
1445 * it at least MSS bytes.
1447 * Unfortunately, the recommended algorithm breaks header prediction,
1448 * since header prediction assumes th->window stays fixed.
1450 * Strictly speaking, keeping th->window fixed violates the receiver
1451 * side SWS prevention criteria. The problem is that under this rule
1452 * a stream of single byte packets will cause the right side of the
1453 * window to always advance by a single byte.
1455 * Of course, if the sender implements sender side SWS prevention
1456 * then this will not be a problem.
1458 * BSD seems to make the following compromise:
1460 * If the free space is less than the 1/4 of the maximum
1461 * space available and the free space is less than 1/2 mss,
1462 * then set the window to 0.
1463 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1464 * Otherwise, just prevent the window from shrinking
1465 * and from being larger than the largest representable value.
1467 * This prevents incremental opening of the window in the regime
1468 * where TCP is limited by the speed of the reader side taking
1469 * data out of the TCP receive queue. It does nothing about
1470 * those cases where the window is constrained on the sender side
1471 * because the pipeline is full.
1473 * BSD also seems to "accidentally" limit itself to windows that are a
1474 * multiple of MSS, at least until the free space gets quite small.
1475 * This would appear to be a side effect of the mbuf implementation.
1476 * Combining these two algorithms results in the observed behavior
1477 * of having a fixed window size at almost all times.
1479 * Below we obtain similar behavior by forcing the offered window to
1480 * a multiple of the mss when it is feasible to do so.
1482 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1483 * Regular options like TIMESTAMP are taken into account.
1485 u32
__tcp_select_window(struct sock
*sk
)
1487 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1488 struct tcp_sock
*tp
= tcp_sk(sk
);
1489 /* MSS for the peer's data. Previous versions used mss_clamp
1490 * here. I don't know if the value based on our guesses
1491 * of peer's MSS is better for the performance. It's more correct
1492 * but may be worse for the performance because of rcv_mss
1493 * fluctuations. --SAW 1998/11/1
1495 int mss
= icsk
->icsk_ack
.rcv_mss
;
1496 int free_space
= tcp_space(sk
);
1497 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1500 if (mss
> full_space
)
1503 if (free_space
< full_space
/2) {
1504 icsk
->icsk_ack
.quick
= 0;
1506 if (tcp_memory_pressure
)
1507 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1509 if (free_space
< mss
)
1513 if (free_space
> tp
->rcv_ssthresh
)
1514 free_space
= tp
->rcv_ssthresh
;
1516 /* Don't do rounding if we are using window scaling, since the
1517 * scaled window will not line up with the MSS boundary anyway.
1519 window
= tp
->rcv_wnd
;
1520 if (tp
->rx_opt
.rcv_wscale
) {
1521 window
= free_space
;
1523 /* Advertise enough space so that it won't get scaled away.
1524 * Import case: prevent zero window announcement if
1525 * 1<<rcv_wscale > mss.
1527 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1528 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1529 << tp
->rx_opt
.rcv_wscale
);
1531 /* Get the largest window that is a nice multiple of mss.
1532 * Window clamp already applied above.
1533 * If our current window offering is within 1 mss of the
1534 * free space we just keep it. This prevents the divide
1535 * and multiply from happening most of the time.
1536 * We also don't do any window rounding when the free space
1539 if (window
<= free_space
- mss
|| window
> free_space
)
1540 window
= (free_space
/mss
)*mss
;
1546 /* Attempt to collapse two adjacent SKB's during retransmission. */
1547 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1549 struct tcp_sock
*tp
= tcp_sk(sk
);
1550 struct sk_buff
*next_skb
= skb
->next
;
1552 /* The first test we must make is that neither of these two
1553 * SKB's are still referenced by someone else.
1555 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1556 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1557 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1559 /* Also punt if next skb has been SACK'd. */
1560 if(TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1563 /* Next skb is out of window. */
1564 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1567 /* Punt if not enough space exists in the first SKB for
1568 * the data in the second, or the total combined payload
1569 * would exceed the MSS.
1571 if ((next_skb_size
> skb_tailroom(skb
)) ||
1572 ((skb_size
+ next_skb_size
) > mss_now
))
1575 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1576 tcp_skb_pcount(next_skb
) != 1);
1578 /* changing transmit queue under us so clear hints */
1579 clear_all_retrans_hints(tp
);
1581 /* Ok. We will be able to collapse the packet. */
1582 __skb_unlink(next_skb
, &sk
->sk_write_queue
);
1584 memcpy(skb_put(skb
, next_skb_size
), next_skb
->data
, next_skb_size
);
1586 skb
->ip_summed
= next_skb
->ip_summed
;
1588 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1589 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1591 /* Update sequence range on original skb. */
1592 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1594 /* Merge over control information. */
1595 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1596 TCP_SKB_CB(skb
)->flags
= flags
;
1598 /* All done, get rid of second SKB and account for it so
1599 * packet counting does not break.
1601 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1602 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1603 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1604 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
) {
1605 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1606 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1608 /* Reno case is special. Sigh... */
1609 if (!tp
->rx_opt
.sack_ok
&& tp
->sacked_out
) {
1610 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1611 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1614 /* Not quite right: it can be > snd.fack, but
1615 * it is better to underestimate fackets.
1617 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1618 tcp_packets_out_dec(tp
, next_skb
);
1619 sk_stream_free_skb(sk
, next_skb
);
1623 /* Do a simple retransmit without using the backoff mechanisms in
1624 * tcp_timer. This is used for path mtu discovery.
1625 * The socket is already locked here.
1627 void tcp_simple_retransmit(struct sock
*sk
)
1629 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1630 struct tcp_sock
*tp
= tcp_sk(sk
);
1631 struct sk_buff
*skb
;
1632 unsigned int mss
= tcp_current_mss(sk
, 0);
1635 sk_stream_for_retrans_queue(skb
, sk
) {
1636 if (skb
->len
> mss
&&
1637 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1638 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1639 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1640 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1642 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1643 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1644 tp
->lost_out
+= tcp_skb_pcount(skb
);
1650 clear_all_retrans_hints(tp
);
1655 tcp_sync_left_out(tp
);
1657 /* Don't muck with the congestion window here.
1658 * Reason is that we do not increase amount of _data_
1659 * in network, but units changed and effective
1660 * cwnd/ssthresh really reduced now.
1662 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1663 tp
->high_seq
= tp
->snd_nxt
;
1664 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1665 tp
->prior_ssthresh
= 0;
1666 tp
->undo_marker
= 0;
1667 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1669 tcp_xmit_retransmit_queue(sk
);
1672 /* This retransmits one SKB. Policy decisions and retransmit queue
1673 * state updates are done by the caller. Returns non-zero if an
1674 * error occurred which prevented the send.
1676 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1678 struct tcp_sock
*tp
= tcp_sk(sk
);
1679 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1680 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1683 /* Inconslusive MTU probe */
1684 if (icsk
->icsk_mtup
.probe_size
) {
1685 icsk
->icsk_mtup
.probe_size
= 0;
1688 /* Do not sent more than we queued. 1/4 is reserved for possible
1689 * copying overhead: fragmentation, tunneling, mangling etc.
1691 if (atomic_read(&sk
->sk_wmem_alloc
) >
1692 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1695 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1696 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1698 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1702 /* If receiver has shrunk his window, and skb is out of
1703 * new window, do not retransmit it. The exception is the
1704 * case, when window is shrunk to zero. In this case
1705 * our retransmit serves as a zero window probe.
1707 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1708 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1711 if (skb
->len
> cur_mss
) {
1712 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1713 return -ENOMEM
; /* We'll try again later. */
1716 /* Collapse two adjacent packets if worthwhile and we can. */
1717 if(!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1718 (skb
->len
< (cur_mss
>> 1)) &&
1719 (skb
->next
!= sk
->sk_send_head
) &&
1720 (skb
->next
!= (struct sk_buff
*)&sk
->sk_write_queue
) &&
1721 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(skb
->next
)->nr_frags
== 0) &&
1722 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(skb
->next
) == 1) &&
1723 (sysctl_tcp_retrans_collapse
!= 0))
1724 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1726 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1727 return -EHOSTUNREACH
; /* Routing failure or similar. */
1729 /* Some Solaris stacks overoptimize and ignore the FIN on a
1730 * retransmit when old data is attached. So strip it off
1731 * since it is cheap to do so and saves bytes on the network.
1734 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1735 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1736 if (!pskb_trim(skb
, 0)) {
1737 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1738 skb_shinfo(skb
)->gso_segs
= 1;
1739 skb_shinfo(skb
)->gso_size
= 0;
1740 skb_shinfo(skb
)->gso_type
= 0;
1741 skb
->ip_summed
= CHECKSUM_NONE
;
1746 /* Make a copy, if the first transmission SKB clone we made
1747 * is still in somebody's hands, else make a clone.
1749 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1751 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1754 /* Update global TCP statistics. */
1755 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1757 tp
->total_retrans
++;
1759 #if FASTRETRANS_DEBUG > 0
1760 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1761 if (net_ratelimit())
1762 printk(KERN_DEBUG
"retrans_out leaked.\n");
1765 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1766 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1768 /* Save stamp of the first retransmit. */
1769 if (!tp
->retrans_stamp
)
1770 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1774 /* snd_nxt is stored to detect loss of retransmitted segment,
1775 * see tcp_input.c tcp_sacktag_write_queue().
1777 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1782 /* This gets called after a retransmit timeout, and the initially
1783 * retransmitted data is acknowledged. It tries to continue
1784 * resending the rest of the retransmit queue, until either
1785 * we've sent it all or the congestion window limit is reached.
1786 * If doing SACK, the first ACK which comes back for a timeout
1787 * based retransmit packet might feed us FACK information again.
1788 * If so, we use it to avoid unnecessarily retransmissions.
1790 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1792 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1793 struct tcp_sock
*tp
= tcp_sk(sk
);
1794 struct sk_buff
*skb
;
1797 if (tp
->retransmit_skb_hint
) {
1798 skb
= tp
->retransmit_skb_hint
;
1799 packet_cnt
= tp
->retransmit_cnt_hint
;
1801 skb
= sk
->sk_write_queue
.next
;
1805 /* First pass: retransmit lost packets. */
1807 sk_stream_for_retrans_queue_from(skb
, sk
) {
1808 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1810 /* we could do better than to assign each time */
1811 tp
->retransmit_skb_hint
= skb
;
1812 tp
->retransmit_cnt_hint
= packet_cnt
;
1814 /* Assume this retransmit will generate
1815 * only one packet for congestion window
1816 * calculation purposes. This works because
1817 * tcp_retransmit_skb() will chop up the
1818 * packet to be MSS sized and all the
1819 * packet counting works out.
1821 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1824 if (sacked
& TCPCB_LOST
) {
1825 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1826 if (tcp_retransmit_skb(sk
, skb
)) {
1827 tp
->retransmit_skb_hint
= NULL
;
1830 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1831 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1833 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1836 skb_peek(&sk
->sk_write_queue
))
1837 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1838 inet_csk(sk
)->icsk_rto
,
1842 packet_cnt
+= tcp_skb_pcount(skb
);
1843 if (packet_cnt
>= tp
->lost_out
)
1849 /* OK, demanded retransmission is finished. */
1851 /* Forward retransmissions are possible only during Recovery. */
1852 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1855 /* No forward retransmissions in Reno are possible. */
1856 if (!tp
->rx_opt
.sack_ok
)
1859 /* Yeah, we have to make difficult choice between forward transmission
1860 * and retransmission... Both ways have their merits...
1862 * For now we do not retransmit anything, while we have some new
1866 if (tcp_may_send_now(sk
, tp
))
1869 if (tp
->forward_skb_hint
) {
1870 skb
= tp
->forward_skb_hint
;
1871 packet_cnt
= tp
->forward_cnt_hint
;
1873 skb
= sk
->sk_write_queue
.next
;
1877 sk_stream_for_retrans_queue_from(skb
, sk
) {
1878 tp
->forward_cnt_hint
= packet_cnt
;
1879 tp
->forward_skb_hint
= skb
;
1881 /* Similar to the retransmit loop above we
1882 * can pretend that the retransmitted SKB
1883 * we send out here will be composed of one
1884 * real MSS sized packet because tcp_retransmit_skb()
1885 * will fragment it if necessary.
1887 if (++packet_cnt
> tp
->fackets_out
)
1890 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1893 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
1896 /* Ok, retransmit it. */
1897 if (tcp_retransmit_skb(sk
, skb
)) {
1898 tp
->forward_skb_hint
= NULL
;
1902 if (skb
== skb_peek(&sk
->sk_write_queue
))
1903 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1904 inet_csk(sk
)->icsk_rto
,
1907 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
1912 /* Send a fin. The caller locks the socket for us. This cannot be
1913 * allowed to fail queueing a FIN frame under any circumstances.
1915 void tcp_send_fin(struct sock
*sk
)
1917 struct tcp_sock
*tp
= tcp_sk(sk
);
1918 struct sk_buff
*skb
= skb_peek_tail(&sk
->sk_write_queue
);
1921 /* Optimization, tack on the FIN if we have a queue of
1922 * unsent frames. But be careful about outgoing SACKS
1925 mss_now
= tcp_current_mss(sk
, 1);
1927 if (sk
->sk_send_head
!= NULL
) {
1928 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
1929 TCP_SKB_CB(skb
)->end_seq
++;
1932 /* Socket is locked, keep trying until memory is available. */
1934 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
1940 /* Reserve space for headers and prepare control bits. */
1941 skb_reserve(skb
, MAX_TCP_HEADER
);
1943 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
1944 TCP_SKB_CB(skb
)->sacked
= 0;
1945 skb_shinfo(skb
)->gso_segs
= 1;
1946 skb_shinfo(skb
)->gso_size
= 0;
1947 skb_shinfo(skb
)->gso_type
= 0;
1949 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1950 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
1951 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
1952 tcp_queue_skb(sk
, skb
);
1954 __tcp_push_pending_frames(sk
, tp
, mss_now
, TCP_NAGLE_OFF
);
1957 /* We get here when a process closes a file descriptor (either due to
1958 * an explicit close() or as a byproduct of exit()'ing) and there
1959 * was unread data in the receive queue. This behavior is recommended
1960 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1962 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
1964 struct tcp_sock
*tp
= tcp_sk(sk
);
1965 struct sk_buff
*skb
;
1967 /* NOTE: No TCP options attached and we never retransmit this. */
1968 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
1970 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1974 /* Reserve space for headers and prepare control bits. */
1975 skb_reserve(skb
, MAX_TCP_HEADER
);
1977 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
1978 TCP_SKB_CB(skb
)->sacked
= 0;
1979 skb_shinfo(skb
)->gso_segs
= 1;
1980 skb_shinfo(skb
)->gso_size
= 0;
1981 skb_shinfo(skb
)->gso_type
= 0;
1984 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
, tp
);
1985 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
1986 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1987 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
1988 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
1991 /* WARNING: This routine must only be called when we have already sent
1992 * a SYN packet that crossed the incoming SYN that caused this routine
1993 * to get called. If this assumption fails then the initial rcv_wnd
1994 * and rcv_wscale values will not be correct.
1996 int tcp_send_synack(struct sock
*sk
)
1998 struct sk_buff
* skb
;
2000 skb
= skb_peek(&sk
->sk_write_queue
);
2001 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
2002 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2005 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
2006 if (skb_cloned(skb
)) {
2007 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2010 __skb_unlink(skb
, &sk
->sk_write_queue
);
2011 skb_header_release(nskb
);
2012 __skb_queue_head(&sk
->sk_write_queue
, nskb
);
2013 sk_stream_free_skb(sk
, skb
);
2014 sk_charge_skb(sk
, nskb
);
2018 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2019 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2021 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2022 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2026 * Prepare a SYN-ACK.
2028 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2029 struct request_sock
*req
)
2031 struct inet_request_sock
*ireq
= inet_rsk(req
);
2032 struct tcp_sock
*tp
= tcp_sk(sk
);
2034 int tcp_header_size
;
2035 struct sk_buff
*skb
;
2037 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2041 /* Reserve space for headers. */
2042 skb_reserve(skb
, MAX_TCP_HEADER
);
2044 skb
->dst
= dst_clone(dst
);
2046 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2047 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2048 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2049 /* SACK_PERM is in the place of NOP NOP of TS */
2050 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2051 skb
->h
.th
= th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
2053 memset(th
, 0, sizeof(struct tcphdr
));
2056 TCP_ECN_make_synack(req
, th
);
2057 th
->source
= inet_sk(sk
)->sport
;
2058 th
->dest
= ireq
->rmt_port
;
2059 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
2060 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2061 TCP_SKB_CB(skb
)->sacked
= 0;
2062 skb_shinfo(skb
)->gso_segs
= 1;
2063 skb_shinfo(skb
)->gso_size
= 0;
2064 skb_shinfo(skb
)->gso_type
= 0;
2065 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2066 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2067 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2069 /* Set this up on the first call only */
2070 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2071 /* tcp_full_space because it is guaranteed to be the first packet */
2072 tcp_select_initial_window(tcp_full_space(sk
),
2073 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2078 ireq
->rcv_wscale
= rcv_wscale
;
2081 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2082 th
->window
= htons(req
->rcv_wnd
);
2084 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2085 tcp_syn_build_options((__be32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2086 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2087 TCP_SKB_CB(skb
)->when
,
2091 th
->doff
= (tcp_header_size
>> 2);
2092 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
2097 * Do all connect socket setups that can be done AF independent.
2099 static void tcp_connect_init(struct sock
*sk
)
2101 struct dst_entry
*dst
= __sk_dst_get(sk
);
2102 struct tcp_sock
*tp
= tcp_sk(sk
);
2105 /* We'll fix this up when we get a response from the other end.
2106 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2108 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2109 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2111 /* If user gave his TCP_MAXSEG, record it to clamp */
2112 if (tp
->rx_opt
.user_mss
)
2113 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2116 tcp_sync_mss(sk
, dst_mtu(dst
));
2118 if (!tp
->window_clamp
)
2119 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2120 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2121 tcp_initialize_rcv_mss(sk
);
2123 tcp_select_initial_window(tcp_full_space(sk
),
2124 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2127 sysctl_tcp_window_scaling
,
2130 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2131 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2134 sock_reset_flag(sk
, SOCK_DONE
);
2136 tcp_init_wl(tp
, tp
->write_seq
, 0);
2137 tp
->snd_una
= tp
->write_seq
;
2138 tp
->snd_sml
= tp
->write_seq
;
2143 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2144 inet_csk(sk
)->icsk_retransmits
= 0;
2145 tcp_clear_retrans(tp
);
2149 * Build a SYN and send it off.
2151 int tcp_connect(struct sock
*sk
)
2153 struct tcp_sock
*tp
= tcp_sk(sk
);
2154 struct sk_buff
*buff
;
2156 tcp_connect_init(sk
);
2158 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2159 if (unlikely(buff
== NULL
))
2162 /* Reserve space for headers. */
2163 skb_reserve(buff
, MAX_TCP_HEADER
);
2165 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
2166 TCP_ECN_send_syn(sk
, tp
, buff
);
2167 TCP_SKB_CB(buff
)->sacked
= 0;
2168 skb_shinfo(buff
)->gso_segs
= 1;
2169 skb_shinfo(buff
)->gso_size
= 0;
2170 skb_shinfo(buff
)->gso_type
= 0;
2172 tp
->snd_nxt
= tp
->write_seq
;
2173 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
2174 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
2177 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2178 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2179 skb_header_release(buff
);
2180 __skb_queue_tail(&sk
->sk_write_queue
, buff
);
2181 sk_charge_skb(sk
, buff
);
2182 tp
->packets_out
+= tcp_skb_pcount(buff
);
2183 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2185 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2186 * in order to make this packet get counted in tcpOutSegs.
2188 tp
->snd_nxt
= tp
->write_seq
;
2189 tp
->pushed_seq
= tp
->write_seq
;
2190 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
2192 /* Timer for repeating the SYN until an answer. */
2193 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2194 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2198 /* Send out a delayed ack, the caller does the policy checking
2199 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2202 void tcp_send_delayed_ack(struct sock
*sk
)
2204 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2205 int ato
= icsk
->icsk_ack
.ato
;
2206 unsigned long timeout
;
2208 if (ato
> TCP_DELACK_MIN
) {
2209 const struct tcp_sock
*tp
= tcp_sk(sk
);
2212 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2213 max_ato
= TCP_DELACK_MAX
;
2215 /* Slow path, intersegment interval is "high". */
2217 /* If some rtt estimate is known, use it to bound delayed ack.
2218 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2222 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
2228 ato
= min(ato
, max_ato
);
2231 /* Stay within the limit we were given */
2232 timeout
= jiffies
+ ato
;
2234 /* Use new timeout only if there wasn't a older one earlier. */
2235 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2236 /* If delack timer was blocked or is about to expire,
2239 if (icsk
->icsk_ack
.blocked
||
2240 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2245 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2246 timeout
= icsk
->icsk_ack
.timeout
;
2248 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2249 icsk
->icsk_ack
.timeout
= timeout
;
2250 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2253 /* This routine sends an ack and also updates the window. */
2254 void tcp_send_ack(struct sock
*sk
)
2256 /* If we have been reset, we may not send again. */
2257 if (sk
->sk_state
!= TCP_CLOSE
) {
2258 struct tcp_sock
*tp
= tcp_sk(sk
);
2259 struct sk_buff
*buff
;
2261 /* We are not putting this on the write queue, so
2262 * tcp_transmit_skb() will set the ownership to this
2265 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2267 inet_csk_schedule_ack(sk
);
2268 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2269 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2270 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2274 /* Reserve space for headers and prepare control bits. */
2275 skb_reserve(buff
, MAX_TCP_HEADER
);
2277 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
2278 TCP_SKB_CB(buff
)->sacked
= 0;
2279 skb_shinfo(buff
)->gso_segs
= 1;
2280 skb_shinfo(buff
)->gso_size
= 0;
2281 skb_shinfo(buff
)->gso_type
= 0;
2283 /* Send it off, this clears delayed acks for us. */
2284 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
, tp
);
2285 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2286 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2290 /* This routine sends a packet with an out of date sequence
2291 * number. It assumes the other end will try to ack it.
2293 * Question: what should we make while urgent mode?
2294 * 4.4BSD forces sending single byte of data. We cannot send
2295 * out of window data, because we have SND.NXT==SND.MAX...
2297 * Current solution: to send TWO zero-length segments in urgent mode:
2298 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2299 * out-of-date with SND.UNA-1 to probe window.
2301 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2303 struct tcp_sock
*tp
= tcp_sk(sk
);
2304 struct sk_buff
*skb
;
2306 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2307 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2311 /* Reserve space for headers and set control bits. */
2312 skb_reserve(skb
, MAX_TCP_HEADER
);
2314 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2315 TCP_SKB_CB(skb
)->sacked
= urgent
;
2316 skb_shinfo(skb
)->gso_segs
= 1;
2317 skb_shinfo(skb
)->gso_size
= 0;
2318 skb_shinfo(skb
)->gso_type
= 0;
2320 /* Use a previous sequence. This should cause the other
2321 * end to send an ack. Don't queue or clone SKB, just
2324 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2325 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2326 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2327 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2330 int tcp_write_wakeup(struct sock
*sk
)
2332 if (sk
->sk_state
!= TCP_CLOSE
) {
2333 struct tcp_sock
*tp
= tcp_sk(sk
);
2334 struct sk_buff
*skb
;
2336 if ((skb
= sk
->sk_send_head
) != NULL
&&
2337 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
2339 unsigned int mss
= tcp_current_mss(sk
, 0);
2340 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
2342 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2343 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2345 /* We are probing the opening of a window
2346 * but the window size is != 0
2347 * must have been a result SWS avoidance ( sender )
2349 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2351 seg_size
= min(seg_size
, mss
);
2352 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2353 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2355 } else if (!tcp_skb_pcount(skb
))
2356 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2358 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2359 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2360 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2362 update_send_head(sk
, tp
, skb
);
2367 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2368 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2369 return tcp_xmit_probe_skb(sk
, 0);
2375 /* A window probe timeout has occurred. If window is not closed send
2376 * a partial packet else a zero probe.
2378 void tcp_send_probe0(struct sock
*sk
)
2380 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2381 struct tcp_sock
*tp
= tcp_sk(sk
);
2384 err
= tcp_write_wakeup(sk
);
2386 if (tp
->packets_out
|| !sk
->sk_send_head
) {
2387 /* Cancel probe timer, if it is not required. */
2388 icsk
->icsk_probes_out
= 0;
2389 icsk
->icsk_backoff
= 0;
2394 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2395 icsk
->icsk_backoff
++;
2396 icsk
->icsk_probes_out
++;
2397 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2398 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2401 /* If packet was not sent due to local congestion,
2402 * do not backoff and do not remember icsk_probes_out.
2403 * Let local senders to fight for local resources.
2405 * Use accumulated backoff yet.
2407 if (!icsk
->icsk_probes_out
)
2408 icsk
->icsk_probes_out
= 1;
2409 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2410 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2411 TCP_RESOURCE_PROBE_INTERVAL
),
2416 EXPORT_SYMBOL(tcp_connect
);
2417 EXPORT_SYMBOL(tcp_make_synack
);
2418 EXPORT_SYMBOL(tcp_simple_retransmit
);
2419 EXPORT_SYMBOL(tcp_sync_mss
);
2420 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);
2421 EXPORT_SYMBOL(tcp_mtup_init
);