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
,
273 __u32 tstamp
, __u8
**md5_hash
)
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
--;
301 #ifdef CONFIG_TCP_MD5SIG
303 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
305 (TCPOPT_MD5SIG
<< 8) |
307 *md5_hash
= (__u8
*)ptr
;
312 /* Construct a tcp options header for a SYN or SYN_ACK packet.
313 * If this is every changed make sure to change the definition of
314 * MAX_SYN_SIZE to match the new maximum number of options that you
317 * Note - that with the RFC2385 TCP option, we make room for the
318 * 16 byte MD5 hash. This will be filled in later, so the pointer for the
319 * location to be filled is passed back up.
321 static void tcp_syn_build_options(__be32
*ptr
, int mss
, int ts
, int sack
,
322 int offer_wscale
, int wscale
, __u32 tstamp
,
323 __u32 ts_recent
, __u8
**md5_hash
)
325 /* We always get an MSS option.
326 * The option bytes which will be seen in normal data
327 * packets should timestamps be used, must be in the MSS
328 * advertised. But we subtract them from tp->mss_cache so
329 * that calculations in tcp_sendmsg are simpler etc.
330 * So account for this fact here if necessary. If we
331 * don't do this correctly, as a receiver we won't
332 * recognize data packets as being full sized when we
333 * should, and thus we won't abide by the delayed ACK
335 * SACKs don't matter, we never delay an ACK when we
336 * have any of those going out.
338 *ptr
++ = htonl((TCPOPT_MSS
<< 24) | (TCPOLEN_MSS
<< 16) | mss
);
341 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
342 (TCPOLEN_SACK_PERM
<< 16) |
343 (TCPOPT_TIMESTAMP
<< 8) |
346 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
348 (TCPOPT_TIMESTAMP
<< 8) |
350 *ptr
++ = htonl(tstamp
); /* TSVAL */
351 *ptr
++ = htonl(ts_recent
); /* TSECR */
353 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
355 (TCPOPT_SACK_PERM
<< 8) |
358 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
359 (TCPOPT_WINDOW
<< 16) |
360 (TCPOLEN_WINDOW
<< 8) |
362 #ifdef CONFIG_TCP_MD5SIG
364 * If MD5 is enabled, then we set the option, and include the size
365 * (always 18). The actual MD5 hash is added just before the
369 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
371 (TCPOPT_MD5SIG
<< 8) |
373 *md5_hash
= (__u8
*) ptr
;
378 /* This routine actually transmits TCP packets queued in by
379 * tcp_do_sendmsg(). This is used by both the initial
380 * transmission and possible later retransmissions.
381 * All SKB's seen here are completely headerless. It is our
382 * job to build the TCP header, and pass the packet down to
383 * IP so it can do the same plus pass the packet off to the
386 * We are working here with either a clone of the original
387 * SKB, or a fresh unique copy made by the retransmit engine.
389 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
, gfp_t gfp_mask
)
391 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
392 struct inet_sock
*inet
;
394 struct tcp_skb_cb
*tcb
;
396 #ifdef CONFIG_TCP_MD5SIG
397 struct tcp_md5sig_key
*md5
;
398 __u8
*md5_hash_location
;
404 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
406 /* If congestion control is doing timestamping, we must
407 * take such a timestamp before we potentially clone/copy.
409 if (icsk
->icsk_ca_ops
->rtt_sample
)
410 __net_timestamp(skb
);
412 if (likely(clone_it
)) {
413 if (unlikely(skb_cloned(skb
)))
414 skb
= pskb_copy(skb
, gfp_mask
);
416 skb
= skb_clone(skb
, gfp_mask
);
423 tcb
= TCP_SKB_CB(skb
);
424 tcp_header_size
= tp
->tcp_header_len
;
426 #define SYSCTL_FLAG_TSTAMPS 0x1
427 #define SYSCTL_FLAG_WSCALE 0x2
428 #define SYSCTL_FLAG_SACK 0x4
431 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
432 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
433 if(sysctl_tcp_timestamps
) {
434 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
435 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
437 if (sysctl_tcp_window_scaling
) {
438 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
439 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
441 if (sysctl_tcp_sack
) {
442 sysctl_flags
|= SYSCTL_FLAG_SACK
;
443 if (!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
444 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
446 } else if (unlikely(tp
->rx_opt
.eff_sacks
)) {
447 /* A SACK is 2 pad bytes, a 2 byte header, plus
448 * 2 32-bit sequence numbers for each SACK block.
450 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
451 (tp
->rx_opt
.eff_sacks
*
452 TCPOLEN_SACK_PERBLOCK
));
455 if (tcp_packets_in_flight(tp
) == 0)
456 tcp_ca_event(sk
, CA_EVENT_TX_START
);
458 #ifdef CONFIG_TCP_MD5SIG
460 * Are we doing MD5 on this segment? If so - make
463 md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
465 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
468 th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
470 skb_set_owner_w(skb
, sk
);
472 /* Build TCP header and checksum it. */
473 th
->source
= inet
->sport
;
474 th
->dest
= inet
->dport
;
475 th
->seq
= htonl(tcb
->seq
);
476 th
->ack_seq
= htonl(tp
->rcv_nxt
);
477 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
480 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
481 /* RFC1323: The window in SYN & SYN/ACK segments
484 th
->window
= htons(tp
->rcv_wnd
);
486 th
->window
= htons(tcp_select_window(sk
));
491 if (unlikely(tp
->urg_mode
&&
492 between(tp
->snd_up
, tcb
->seq
+1, tcb
->seq
+0xFFFF))) {
493 th
->urg_ptr
= htons(tp
->snd_up
-tcb
->seq
);
497 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
498 tcp_syn_build_options((__be32
*)(th
+ 1),
499 tcp_advertise_mss(sk
),
500 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
501 (sysctl_flags
& SYSCTL_FLAG_SACK
),
502 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
503 tp
->rx_opt
.rcv_wscale
,
505 tp
->rx_opt
.ts_recent
,
507 #ifdef CONFIG_TCP_MD5SIG
508 md5
? &md5_hash_location
:
512 tcp_build_and_update_options((__be32
*)(th
+ 1),
514 #ifdef CONFIG_TCP_MD5SIG
515 md5
? &md5_hash_location
:
518 TCP_ECN_send(sk
, tp
, skb
, tcp_header_size
);
521 #ifdef CONFIG_TCP_MD5SIG
522 /* Calculate the MD5 hash, as we have all we need now */
524 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
533 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
535 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
536 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
538 if (skb
->len
!= tcp_header_size
)
539 tcp_event_data_sent(tp
, skb
, sk
);
541 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
542 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
544 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
545 if (likely(err
<= 0))
550 return net_xmit_eval(err
);
552 #undef SYSCTL_FLAG_TSTAMPS
553 #undef SYSCTL_FLAG_WSCALE
554 #undef SYSCTL_FLAG_SACK
558 /* This routine just queue's the buffer
560 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
561 * otherwise socket can stall.
563 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
565 struct tcp_sock
*tp
= tcp_sk(sk
);
567 /* Advance write_seq and place onto the write_queue. */
568 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
569 skb_header_release(skb
);
570 __skb_queue_tail(&sk
->sk_write_queue
, skb
);
571 sk_charge_skb(sk
, skb
);
573 /* Queue it, remembering where we must start sending. */
574 if (sk
->sk_send_head
== NULL
)
575 sk
->sk_send_head
= skb
;
578 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
580 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
)) {
581 /* Avoid the costly divide in the normal
584 skb_shinfo(skb
)->gso_segs
= 1;
585 skb_shinfo(skb
)->gso_size
= 0;
586 skb_shinfo(skb
)->gso_type
= 0;
590 factor
= skb
->len
+ (mss_now
- 1);
592 skb_shinfo(skb
)->gso_segs
= factor
;
593 skb_shinfo(skb
)->gso_size
= mss_now
;
594 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
598 /* Function to create two new TCP segments. Shrinks the given segment
599 * to the specified size and appends a new segment with the rest of the
600 * packet to the list. This won't be called frequently, I hope.
601 * Remember, these are still headerless SKBs at this point.
603 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
, unsigned int mss_now
)
605 struct tcp_sock
*tp
= tcp_sk(sk
);
606 struct sk_buff
*buff
;
607 int nsize
, old_factor
;
611 BUG_ON(len
> skb
->len
);
613 clear_all_retrans_hints(tp
);
614 nsize
= skb_headlen(skb
) - len
;
618 if (skb_cloned(skb
) &&
619 skb_is_nonlinear(skb
) &&
620 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
623 /* Get a new skb... force flag on. */
624 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
626 return -ENOMEM
; /* We'll just try again later. */
628 sk_charge_skb(sk
, buff
);
629 nlen
= skb
->len
- len
- nsize
;
630 buff
->truesize
+= nlen
;
631 skb
->truesize
-= nlen
;
633 /* Correct the sequence numbers. */
634 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
635 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
636 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
638 /* PSH and FIN should only be set in the second packet. */
639 flags
= TCP_SKB_CB(skb
)->flags
;
640 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
641 TCP_SKB_CB(buff
)->flags
= flags
;
642 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
643 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
645 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
646 /* Copy and checksum data tail into the new buffer. */
647 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
652 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
654 skb
->ip_summed
= CHECKSUM_PARTIAL
;
655 skb_split(skb
, buff
, len
);
658 buff
->ip_summed
= skb
->ip_summed
;
660 /* Looks stupid, but our code really uses when of
661 * skbs, which it never sent before. --ANK
663 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
664 buff
->tstamp
= skb
->tstamp
;
666 old_factor
= tcp_skb_pcount(skb
);
668 /* Fix up tso_factor for both original and new SKB. */
669 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
670 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
672 /* If this packet has been sent out already, we must
673 * adjust the various packet counters.
675 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
676 int diff
= old_factor
- tcp_skb_pcount(skb
) -
677 tcp_skb_pcount(buff
);
679 tp
->packets_out
-= diff
;
681 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
682 tp
->sacked_out
-= diff
;
683 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
684 tp
->retrans_out
-= diff
;
686 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
687 tp
->lost_out
-= diff
;
688 tp
->left_out
-= diff
;
692 /* Adjust Reno SACK estimate. */
693 if (!tp
->rx_opt
.sack_ok
) {
694 tp
->sacked_out
-= diff
;
695 if ((int)tp
->sacked_out
< 0)
697 tcp_sync_left_out(tp
);
700 tp
->fackets_out
-= diff
;
701 if ((int)tp
->fackets_out
< 0)
706 /* Link BUFF into the send queue. */
707 skb_header_release(buff
);
708 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
713 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
714 * eventually). The difference is that pulled data not copied, but
715 * immediately discarded.
717 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
723 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
724 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
725 put_page(skb_shinfo(skb
)->frags
[i
].page
);
726 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
728 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
730 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
731 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
737 skb_shinfo(skb
)->nr_frags
= k
;
739 skb
->tail
= skb
->data
;
740 skb
->data_len
-= len
;
741 skb
->len
= skb
->data_len
;
744 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
746 if (skb_cloned(skb
) &&
747 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
750 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
751 if (unlikely(len
< skb_headlen(skb
)))
752 __skb_pull(skb
, len
);
754 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
756 TCP_SKB_CB(skb
)->seq
+= len
;
757 skb
->ip_summed
= CHECKSUM_PARTIAL
;
759 skb
->truesize
-= len
;
760 sk
->sk_wmem_queued
-= len
;
761 sk
->sk_forward_alloc
+= len
;
762 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
764 /* Any change of skb->len requires recalculation of tso
767 if (tcp_skb_pcount(skb
) > 1)
768 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
773 /* Not accounting for SACKs here. */
774 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
776 struct tcp_sock
*tp
= tcp_sk(sk
);
777 struct inet_connection_sock
*icsk
= inet_csk(sk
);
780 /* Calculate base mss without TCP options:
781 It is MMS_S - sizeof(tcphdr) of rfc1122
783 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
785 /* Clamp it (mss_clamp does not include tcp options) */
786 if (mss_now
> tp
->rx_opt
.mss_clamp
)
787 mss_now
= tp
->rx_opt
.mss_clamp
;
789 /* Now subtract optional transport overhead */
790 mss_now
-= icsk
->icsk_ext_hdr_len
;
792 /* Then reserve room for full set of TCP options and 8 bytes of data */
796 /* Now subtract TCP options size, not including SACKs */
797 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
802 /* Inverse of above */
803 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
805 struct tcp_sock
*tp
= tcp_sk(sk
);
806 struct inet_connection_sock
*icsk
= inet_csk(sk
);
811 icsk
->icsk_ext_hdr_len
+
812 icsk
->icsk_af_ops
->net_header_len
;
817 void tcp_mtup_init(struct sock
*sk
)
819 struct tcp_sock
*tp
= tcp_sk(sk
);
820 struct inet_connection_sock
*icsk
= inet_csk(sk
);
822 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
823 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
824 icsk
->icsk_af_ops
->net_header_len
;
825 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
826 icsk
->icsk_mtup
.probe_size
= 0;
829 /* This function synchronize snd mss to current pmtu/exthdr set.
831 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
832 for TCP options, but includes only bare TCP header.
834 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
835 It is minimum of user_mss and mss received with SYN.
836 It also does not include TCP options.
838 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
840 tp->mss_cache is current effective sending mss, including
841 all tcp options except for SACKs. It is evaluated,
842 taking into account current pmtu, but never exceeds
843 tp->rx_opt.mss_clamp.
845 NOTE1. rfc1122 clearly states that advertised MSS
846 DOES NOT include either tcp or ip options.
848 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
849 are READ ONLY outside this function. --ANK (980731)
852 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
854 struct tcp_sock
*tp
= tcp_sk(sk
);
855 struct inet_connection_sock
*icsk
= inet_csk(sk
);
858 if (icsk
->icsk_mtup
.search_high
> pmtu
)
859 icsk
->icsk_mtup
.search_high
= pmtu
;
861 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
863 /* Bound mss with half of window */
864 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
865 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
867 /* And store cached results */
868 icsk
->icsk_pmtu_cookie
= pmtu
;
869 if (icsk
->icsk_mtup
.enabled
)
870 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
871 tp
->mss_cache
= mss_now
;
876 /* Compute the current effective MSS, taking SACKs and IP options,
877 * and even PMTU discovery events into account.
879 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
880 * cannot be large. However, taking into account rare use of URG, this
883 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
885 struct tcp_sock
*tp
= tcp_sk(sk
);
886 struct dst_entry
*dst
= __sk_dst_get(sk
);
891 mss_now
= tp
->mss_cache
;
893 if (large_allowed
&& sk_can_gso(sk
) && !tp
->urg_mode
)
897 u32 mtu
= dst_mtu(dst
);
898 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
899 mss_now
= tcp_sync_mss(sk
, mtu
);
902 if (tp
->rx_opt
.eff_sacks
)
903 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
904 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
906 #ifdef CONFIG_TCP_MD5SIG
907 if (tp
->af_specific
->md5_lookup(sk
, sk
))
908 mss_now
-= TCPOLEN_MD5SIG_ALIGNED
;
911 xmit_size_goal
= mss_now
;
914 xmit_size_goal
= (65535 -
915 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
916 inet_csk(sk
)->icsk_ext_hdr_len
-
919 if (tp
->max_window
&&
920 (xmit_size_goal
> (tp
->max_window
>> 1)))
921 xmit_size_goal
= max((tp
->max_window
>> 1),
922 68U - tp
->tcp_header_len
);
924 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
926 tp
->xmit_size_goal
= xmit_size_goal
;
931 /* Congestion window validation. (RFC2861) */
933 static void tcp_cwnd_validate(struct sock
*sk
, struct tcp_sock
*tp
)
935 __u32 packets_out
= tp
->packets_out
;
937 if (packets_out
>= tp
->snd_cwnd
) {
938 /* Network is feed fully. */
939 tp
->snd_cwnd_used
= 0;
940 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
942 /* Network starves. */
943 if (tp
->packets_out
> tp
->snd_cwnd_used
)
944 tp
->snd_cwnd_used
= tp
->packets_out
;
946 if ((s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
947 tcp_cwnd_application_limited(sk
);
951 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
953 u32 window
, cwnd_len
;
955 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
956 cwnd_len
= mss_now
* cwnd
;
957 return min(window
, cwnd_len
);
960 /* Can at least one segment of SKB be sent right now, according to the
961 * congestion window rules? If so, return how many segments are allowed.
963 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
967 /* Don't be strict about the congestion window for the final FIN. */
968 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
971 in_flight
= tcp_packets_in_flight(tp
);
973 if (in_flight
< cwnd
)
974 return (cwnd
- in_flight
);
979 /* This must be invoked the first time we consider transmitting
982 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
984 int tso_segs
= tcp_skb_pcount(skb
);
988 tcp_skb_mss(skb
) != mss_now
)) {
989 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
990 tso_segs
= tcp_skb_pcount(skb
);
995 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
997 return after(tp
->snd_sml
,tp
->snd_una
) &&
998 !after(tp
->snd_sml
, tp
->snd_nxt
);
1001 /* Return 0, if packet can be sent now without violation Nagle's rules:
1002 * 1. It is full sized.
1003 * 2. Or it contains FIN. (already checked by caller)
1004 * 3. Or TCP_NODELAY was set.
1005 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1006 * With Minshall's modification: all sent small packets are ACKed.
1009 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1010 const struct sk_buff
*skb
,
1011 unsigned mss_now
, int nonagle
)
1013 return (skb
->len
< mss_now
&&
1014 ((nonagle
&TCP_NAGLE_CORK
) ||
1017 tcp_minshall_check(tp
))));
1020 /* Return non-zero if the Nagle test allows this packet to be
1023 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1024 unsigned int cur_mss
, int nonagle
)
1026 /* Nagle rule does not apply to frames, which sit in the middle of the
1027 * write_queue (they have no chances to get new data).
1029 * This is implemented in the callers, where they modify the 'nonagle'
1030 * argument based upon the location of SKB in the send queue.
1032 if (nonagle
& TCP_NAGLE_PUSH
)
1035 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1037 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1040 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1046 /* Does at least the first segment of SKB fit into the send window? */
1047 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
1049 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1051 if (skb
->len
> cur_mss
)
1052 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1054 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
1057 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
1058 * should be put on the wire right now. If so, it returns the number of
1059 * packets allowed by the congestion window.
1061 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1062 unsigned int cur_mss
, int nonagle
)
1064 struct tcp_sock
*tp
= tcp_sk(sk
);
1065 unsigned int cwnd_quota
;
1067 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1069 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1072 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1074 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1080 static inline int tcp_skb_is_last(const struct sock
*sk
,
1081 const struct sk_buff
*skb
)
1083 return skb
->next
== (struct sk_buff
*)&sk
->sk_write_queue
;
1086 int tcp_may_send_now(struct sock
*sk
, struct tcp_sock
*tp
)
1088 struct sk_buff
*skb
= sk
->sk_send_head
;
1091 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1092 (tcp_skb_is_last(sk
, skb
) ?
1097 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1098 * which is put after SKB on the list. It is very much like
1099 * tcp_fragment() except that it may make several kinds of assumptions
1100 * in order to speed up the splitting operation. In particular, we
1101 * know that all the data is in scatter-gather pages, and that the
1102 * packet has never been sent out before (and thus is not cloned).
1104 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
1106 struct sk_buff
*buff
;
1107 int nlen
= skb
->len
- len
;
1110 /* All of a TSO frame must be composed of paged data. */
1111 if (skb
->len
!= skb
->data_len
)
1112 return tcp_fragment(sk
, skb
, len
, mss_now
);
1114 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
1115 if (unlikely(buff
== NULL
))
1118 sk_charge_skb(sk
, buff
);
1119 buff
->truesize
+= nlen
;
1120 skb
->truesize
-= nlen
;
1122 /* Correct the sequence numbers. */
1123 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1124 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1125 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1127 /* PSH and FIN should only be set in the second packet. */
1128 flags
= TCP_SKB_CB(skb
)->flags
;
1129 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1130 TCP_SKB_CB(buff
)->flags
= flags
;
1132 /* This packet was never sent out yet, so no SACK bits. */
1133 TCP_SKB_CB(buff
)->sacked
= 0;
1135 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1136 skb_split(skb
, buff
, len
);
1138 /* Fix up tso_factor for both original and new SKB. */
1139 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1140 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1142 /* Link BUFF into the send queue. */
1143 skb_header_release(buff
);
1144 __skb_append(skb
, buff
, &sk
->sk_write_queue
);
1149 /* Try to defer sending, if possible, in order to minimize the amount
1150 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1152 * This algorithm is from John Heffner.
1154 static int tcp_tso_should_defer(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
1156 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1157 u32 send_win
, cong_win
, limit
, in_flight
;
1159 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1162 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1165 /* Defer for less than two clock ticks. */
1166 if (!tp
->tso_deferred
&& ((jiffies
<<1)>>1) - (tp
->tso_deferred
>>1) > 1)
1169 in_flight
= tcp_packets_in_flight(tp
);
1171 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
1172 (tp
->snd_cwnd
<= in_flight
));
1174 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
1176 /* From in_flight test above, we know that cwnd > in_flight. */
1177 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1179 limit
= min(send_win
, cong_win
);
1181 /* If a full-sized TSO skb can be sent, do it. */
1185 if (sysctl_tcp_tso_win_divisor
) {
1186 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1188 /* If at least some fraction of a window is available,
1191 chunk
/= sysctl_tcp_tso_win_divisor
;
1195 /* Different approach, try not to defer past a single
1196 * ACK. Receiver should ACK every other full sized
1197 * frame, so if we have space for more than 3 frames
1200 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1204 /* Ok, it looks like it is advisable to defer. */
1205 tp
->tso_deferred
= 1 | (jiffies
<<1);
1210 tp
->tso_deferred
= 0;
1214 /* Create a new MTU probe if we are ready.
1215 * Returns 0 if we should wait to probe (no cwnd available),
1216 * 1 if a probe was sent,
1218 static int tcp_mtu_probe(struct sock
*sk
)
1220 struct tcp_sock
*tp
= tcp_sk(sk
);
1221 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1222 struct sk_buff
*skb
, *nskb
, *next
;
1229 /* Not currently probing/verifying,
1231 * have enough cwnd, and
1232 * not SACKing (the variable headers throw things off) */
1233 if (!icsk
->icsk_mtup
.enabled
||
1234 icsk
->icsk_mtup
.probe_size
||
1235 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1236 tp
->snd_cwnd
< 11 ||
1237 tp
->rx_opt
.eff_sacks
)
1240 /* Very simple search strategy: just double the MSS. */
1241 mss_now
= tcp_current_mss(sk
, 0);
1242 probe_size
= 2*tp
->mss_cache
;
1243 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1244 /* TODO: set timer for probe_converge_event */
1248 /* Have enough data in the send queue to probe? */
1250 if ((skb
= sk
->sk_send_head
) == NULL
)
1252 while ((len
+= skb
->len
) < probe_size
&& !tcp_skb_is_last(sk
, skb
))
1254 if (len
< probe_size
)
1257 /* Receive window check. */
1258 if (after(TCP_SKB_CB(skb
)->seq
+ probe_size
, tp
->snd_una
+ tp
->snd_wnd
)) {
1259 if (tp
->snd_wnd
< probe_size
)
1265 /* Do we need to wait to drain cwnd? */
1266 pif
= tcp_packets_in_flight(tp
);
1267 if (pif
+ 2 > tp
->snd_cwnd
) {
1268 /* With no packets in flight, don't stall. */
1275 /* We're allowed to probe. Build it now. */
1276 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1278 sk_charge_skb(sk
, nskb
);
1280 skb
= sk
->sk_send_head
;
1281 __skb_insert(nskb
, skb
->prev
, skb
, &sk
->sk_write_queue
);
1282 sk
->sk_send_head
= nskb
;
1284 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1285 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1286 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1287 TCP_SKB_CB(nskb
)->sacked
= 0;
1289 nskb
->ip_summed
= skb
->ip_summed
;
1292 while (len
< probe_size
) {
1295 copy
= min_t(int, skb
->len
, probe_size
- len
);
1296 if (nskb
->ip_summed
)
1297 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1299 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1300 skb_put(nskb
, copy
), copy
, nskb
->csum
);
1302 if (skb
->len
<= copy
) {
1303 /* We've eaten all the data from this skb.
1305 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1306 __skb_unlink(skb
, &sk
->sk_write_queue
);
1307 sk_stream_free_skb(sk
, skb
);
1309 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1310 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1311 if (!skb_shinfo(skb
)->nr_frags
) {
1312 skb_pull(skb
, copy
);
1313 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1314 skb
->csum
= csum_partial(skb
->data
, skb
->len
, 0);
1316 __pskb_trim_head(skb
, copy
);
1317 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1319 TCP_SKB_CB(skb
)->seq
+= copy
;
1325 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1327 /* We're ready to send. If this fails, the probe will
1328 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1329 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1330 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1331 /* Decrement cwnd here because we are sending
1332 * effectively two packets. */
1334 update_send_head(sk
, tp
, nskb
);
1336 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1337 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1338 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1347 /* This routine writes packets to the network. It advances the
1348 * send_head. This happens as incoming acks open up the remote
1351 * Returns 1, if no segments are in flight and we have queued segments, but
1352 * cannot send anything now because of SWS or another problem.
1354 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1356 struct tcp_sock
*tp
= tcp_sk(sk
);
1357 struct sk_buff
*skb
;
1358 unsigned int tso_segs
, sent_pkts
;
1362 /* If we are closed, the bytes will have to remain here.
1363 * In time closedown will finish, we empty the write queue and all
1366 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1371 /* Do MTU probing. */
1372 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1374 } else if (result
> 0) {
1378 while ((skb
= sk
->sk_send_head
)) {
1381 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1384 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1388 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1391 if (tso_segs
== 1) {
1392 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1393 (tcp_skb_is_last(sk
, skb
) ?
1394 nonagle
: TCP_NAGLE_PUSH
))))
1397 if (tcp_tso_should_defer(sk
, tp
, skb
))
1403 limit
= tcp_window_allows(tp
, skb
,
1404 mss_now
, cwnd_quota
);
1406 if (skb
->len
< limit
) {
1407 unsigned int trim
= skb
->len
% mss_now
;
1410 limit
= skb
->len
- trim
;
1414 if (skb
->len
> limit
&&
1415 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1418 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1420 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1423 /* Advance the send_head. This one is sent out.
1424 * This call will increment packets_out.
1426 update_send_head(sk
, tp
, skb
);
1428 tcp_minshall_update(tp
, mss_now
, skb
);
1432 if (likely(sent_pkts
)) {
1433 tcp_cwnd_validate(sk
, tp
);
1436 return !tp
->packets_out
&& sk
->sk_send_head
;
1439 /* Push out any pending frames which were held back due to
1440 * TCP_CORK or attempt at coalescing tiny packets.
1441 * The socket must be locked by the caller.
1443 void __tcp_push_pending_frames(struct sock
*sk
, struct tcp_sock
*tp
,
1444 unsigned int cur_mss
, int nonagle
)
1446 struct sk_buff
*skb
= sk
->sk_send_head
;
1449 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1450 tcp_check_probe_timer(sk
, tp
);
1454 /* Send _single_ skb sitting at the send head. This function requires
1455 * true push pending frames to setup probe timer etc.
1457 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1459 struct tcp_sock
*tp
= tcp_sk(sk
);
1460 struct sk_buff
*skb
= sk
->sk_send_head
;
1461 unsigned int tso_segs
, cwnd_quota
;
1463 BUG_ON(!skb
|| skb
->len
< mss_now
);
1465 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1466 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1468 if (likely(cwnd_quota
)) {
1475 limit
= tcp_window_allows(tp
, skb
,
1476 mss_now
, cwnd_quota
);
1478 if (skb
->len
< limit
) {
1479 unsigned int trim
= skb
->len
% mss_now
;
1482 limit
= skb
->len
- trim
;
1486 if (skb
->len
> limit
&&
1487 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1490 /* Send it out now. */
1491 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1493 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1494 update_send_head(sk
, tp
, skb
);
1495 tcp_cwnd_validate(sk
, tp
);
1501 /* This function returns the amount that we can raise the
1502 * usable window based on the following constraints
1504 * 1. The window can never be shrunk once it is offered (RFC 793)
1505 * 2. We limit memory per socket
1508 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1509 * RECV.NEXT + RCV.WIN fixed until:
1510 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1512 * i.e. don't raise the right edge of the window until you can raise
1513 * it at least MSS bytes.
1515 * Unfortunately, the recommended algorithm breaks header prediction,
1516 * since header prediction assumes th->window stays fixed.
1518 * Strictly speaking, keeping th->window fixed violates the receiver
1519 * side SWS prevention criteria. The problem is that under this rule
1520 * a stream of single byte packets will cause the right side of the
1521 * window to always advance by a single byte.
1523 * Of course, if the sender implements sender side SWS prevention
1524 * then this will not be a problem.
1526 * BSD seems to make the following compromise:
1528 * If the free space is less than the 1/4 of the maximum
1529 * space available and the free space is less than 1/2 mss,
1530 * then set the window to 0.
1531 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1532 * Otherwise, just prevent the window from shrinking
1533 * and from being larger than the largest representable value.
1535 * This prevents incremental opening of the window in the regime
1536 * where TCP is limited by the speed of the reader side taking
1537 * data out of the TCP receive queue. It does nothing about
1538 * those cases where the window is constrained on the sender side
1539 * because the pipeline is full.
1541 * BSD also seems to "accidentally" limit itself to windows that are a
1542 * multiple of MSS, at least until the free space gets quite small.
1543 * This would appear to be a side effect of the mbuf implementation.
1544 * Combining these two algorithms results in the observed behavior
1545 * of having a fixed window size at almost all times.
1547 * Below we obtain similar behavior by forcing the offered window to
1548 * a multiple of the mss when it is feasible to do so.
1550 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1551 * Regular options like TIMESTAMP are taken into account.
1553 u32
__tcp_select_window(struct sock
*sk
)
1555 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1556 struct tcp_sock
*tp
= tcp_sk(sk
);
1557 /* MSS for the peer's data. Previous versions used mss_clamp
1558 * here. I don't know if the value based on our guesses
1559 * of peer's MSS is better for the performance. It's more correct
1560 * but may be worse for the performance because of rcv_mss
1561 * fluctuations. --SAW 1998/11/1
1563 int mss
= icsk
->icsk_ack
.rcv_mss
;
1564 int free_space
= tcp_space(sk
);
1565 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1568 if (mss
> full_space
)
1571 if (free_space
< full_space
/2) {
1572 icsk
->icsk_ack
.quick
= 0;
1574 if (tcp_memory_pressure
)
1575 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1577 if (free_space
< mss
)
1581 if (free_space
> tp
->rcv_ssthresh
)
1582 free_space
= tp
->rcv_ssthresh
;
1584 /* Don't do rounding if we are using window scaling, since the
1585 * scaled window will not line up with the MSS boundary anyway.
1587 window
= tp
->rcv_wnd
;
1588 if (tp
->rx_opt
.rcv_wscale
) {
1589 window
= free_space
;
1591 /* Advertise enough space so that it won't get scaled away.
1592 * Import case: prevent zero window announcement if
1593 * 1<<rcv_wscale > mss.
1595 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1596 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1597 << tp
->rx_opt
.rcv_wscale
);
1599 /* Get the largest window that is a nice multiple of mss.
1600 * Window clamp already applied above.
1601 * If our current window offering is within 1 mss of the
1602 * free space we just keep it. This prevents the divide
1603 * and multiply from happening most of the time.
1604 * We also don't do any window rounding when the free space
1607 if (window
<= free_space
- mss
|| window
> free_space
)
1608 window
= (free_space
/mss
)*mss
;
1614 /* Attempt to collapse two adjacent SKB's during retransmission. */
1615 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1617 struct tcp_sock
*tp
= tcp_sk(sk
);
1618 struct sk_buff
*next_skb
= skb
->next
;
1620 /* The first test we must make is that neither of these two
1621 * SKB's are still referenced by someone else.
1623 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1624 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1625 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1627 /* Also punt if next skb has been SACK'd. */
1628 if(TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1631 /* Next skb is out of window. */
1632 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1635 /* Punt if not enough space exists in the first SKB for
1636 * the data in the second, or the total combined payload
1637 * would exceed the MSS.
1639 if ((next_skb_size
> skb_tailroom(skb
)) ||
1640 ((skb_size
+ next_skb_size
) > mss_now
))
1643 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1644 tcp_skb_pcount(next_skb
) != 1);
1646 /* changing transmit queue under us so clear hints */
1647 clear_all_retrans_hints(tp
);
1649 /* Ok. We will be able to collapse the packet. */
1650 __skb_unlink(next_skb
, &sk
->sk_write_queue
);
1652 memcpy(skb_put(skb
, next_skb_size
), next_skb
->data
, next_skb_size
);
1654 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1655 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1657 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1658 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1660 /* Update sequence range on original skb. */
1661 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1663 /* Merge over control information. */
1664 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1665 TCP_SKB_CB(skb
)->flags
= flags
;
1667 /* All done, get rid of second SKB and account for it so
1668 * packet counting does not break.
1670 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1671 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1672 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1673 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
) {
1674 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1675 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1677 /* Reno case is special. Sigh... */
1678 if (!tp
->rx_opt
.sack_ok
&& tp
->sacked_out
) {
1679 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1680 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1683 /* Not quite right: it can be > snd.fack, but
1684 * it is better to underestimate fackets.
1686 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1687 tcp_packets_out_dec(tp
, next_skb
);
1688 sk_stream_free_skb(sk
, next_skb
);
1692 /* Do a simple retransmit without using the backoff mechanisms in
1693 * tcp_timer. This is used for path mtu discovery.
1694 * The socket is already locked here.
1696 void tcp_simple_retransmit(struct sock
*sk
)
1698 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1699 struct tcp_sock
*tp
= tcp_sk(sk
);
1700 struct sk_buff
*skb
;
1701 unsigned int mss
= tcp_current_mss(sk
, 0);
1704 sk_stream_for_retrans_queue(skb
, sk
) {
1705 if (skb
->len
> mss
&&
1706 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1707 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1708 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1709 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1711 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1712 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1713 tp
->lost_out
+= tcp_skb_pcount(skb
);
1719 clear_all_retrans_hints(tp
);
1724 tcp_sync_left_out(tp
);
1726 /* Don't muck with the congestion window here.
1727 * Reason is that we do not increase amount of _data_
1728 * in network, but units changed and effective
1729 * cwnd/ssthresh really reduced now.
1731 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1732 tp
->high_seq
= tp
->snd_nxt
;
1733 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1734 tp
->prior_ssthresh
= 0;
1735 tp
->undo_marker
= 0;
1736 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1738 tcp_xmit_retransmit_queue(sk
);
1741 /* This retransmits one SKB. Policy decisions and retransmit queue
1742 * state updates are done by the caller. Returns non-zero if an
1743 * error occurred which prevented the send.
1745 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1747 struct tcp_sock
*tp
= tcp_sk(sk
);
1748 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1749 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1752 /* Inconslusive MTU probe */
1753 if (icsk
->icsk_mtup
.probe_size
) {
1754 icsk
->icsk_mtup
.probe_size
= 0;
1757 /* Do not sent more than we queued. 1/4 is reserved for possible
1758 * copying overhead: fragmentation, tunneling, mangling etc.
1760 if (atomic_read(&sk
->sk_wmem_alloc
) >
1761 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1764 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1765 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1767 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1771 /* If receiver has shrunk his window, and skb is out of
1772 * new window, do not retransmit it. The exception is the
1773 * case, when window is shrunk to zero. In this case
1774 * our retransmit serves as a zero window probe.
1776 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1777 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1780 if (skb
->len
> cur_mss
) {
1781 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1782 return -ENOMEM
; /* We'll try again later. */
1785 /* Collapse two adjacent packets if worthwhile and we can. */
1786 if(!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1787 (skb
->len
< (cur_mss
>> 1)) &&
1788 (skb
->next
!= sk
->sk_send_head
) &&
1789 (skb
->next
!= (struct sk_buff
*)&sk
->sk_write_queue
) &&
1790 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(skb
->next
)->nr_frags
== 0) &&
1791 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(skb
->next
) == 1) &&
1792 (sysctl_tcp_retrans_collapse
!= 0))
1793 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1795 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1796 return -EHOSTUNREACH
; /* Routing failure or similar. */
1798 /* Some Solaris stacks overoptimize and ignore the FIN on a
1799 * retransmit when old data is attached. So strip it off
1800 * since it is cheap to do so and saves bytes on the network.
1803 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1804 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1805 if (!pskb_trim(skb
, 0)) {
1806 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1807 skb_shinfo(skb
)->gso_segs
= 1;
1808 skb_shinfo(skb
)->gso_size
= 0;
1809 skb_shinfo(skb
)->gso_type
= 0;
1810 skb
->ip_summed
= CHECKSUM_NONE
;
1815 /* Make a copy, if the first transmission SKB clone we made
1816 * is still in somebody's hands, else make a clone.
1818 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1820 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1823 /* Update global TCP statistics. */
1824 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1826 tp
->total_retrans
++;
1828 #if FASTRETRANS_DEBUG > 0
1829 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1830 if (net_ratelimit())
1831 printk(KERN_DEBUG
"retrans_out leaked.\n");
1834 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1835 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1837 /* Save stamp of the first retransmit. */
1838 if (!tp
->retrans_stamp
)
1839 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1843 /* snd_nxt is stored to detect loss of retransmitted segment,
1844 * see tcp_input.c tcp_sacktag_write_queue().
1846 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1851 /* This gets called after a retransmit timeout, and the initially
1852 * retransmitted data is acknowledged. It tries to continue
1853 * resending the rest of the retransmit queue, until either
1854 * we've sent it all or the congestion window limit is reached.
1855 * If doing SACK, the first ACK which comes back for a timeout
1856 * based retransmit packet might feed us FACK information again.
1857 * If so, we use it to avoid unnecessarily retransmissions.
1859 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1861 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1862 struct tcp_sock
*tp
= tcp_sk(sk
);
1863 struct sk_buff
*skb
;
1866 if (tp
->retransmit_skb_hint
) {
1867 skb
= tp
->retransmit_skb_hint
;
1868 packet_cnt
= tp
->retransmit_cnt_hint
;
1870 skb
= sk
->sk_write_queue
.next
;
1874 /* First pass: retransmit lost packets. */
1876 sk_stream_for_retrans_queue_from(skb
, sk
) {
1877 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1879 /* we could do better than to assign each time */
1880 tp
->retransmit_skb_hint
= skb
;
1881 tp
->retransmit_cnt_hint
= packet_cnt
;
1883 /* Assume this retransmit will generate
1884 * only one packet for congestion window
1885 * calculation purposes. This works because
1886 * tcp_retransmit_skb() will chop up the
1887 * packet to be MSS sized and all the
1888 * packet counting works out.
1890 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1893 if (sacked
& TCPCB_LOST
) {
1894 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1895 if (tcp_retransmit_skb(sk
, skb
)) {
1896 tp
->retransmit_skb_hint
= NULL
;
1899 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1900 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1902 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1905 skb_peek(&sk
->sk_write_queue
))
1906 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1907 inet_csk(sk
)->icsk_rto
,
1911 packet_cnt
+= tcp_skb_pcount(skb
);
1912 if (packet_cnt
>= tp
->lost_out
)
1918 /* OK, demanded retransmission is finished. */
1920 /* Forward retransmissions are possible only during Recovery. */
1921 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1924 /* No forward retransmissions in Reno are possible. */
1925 if (!tp
->rx_opt
.sack_ok
)
1928 /* Yeah, we have to make difficult choice between forward transmission
1929 * and retransmission... Both ways have their merits...
1931 * For now we do not retransmit anything, while we have some new
1935 if (tcp_may_send_now(sk
, tp
))
1938 if (tp
->forward_skb_hint
) {
1939 skb
= tp
->forward_skb_hint
;
1940 packet_cnt
= tp
->forward_cnt_hint
;
1942 skb
= sk
->sk_write_queue
.next
;
1946 sk_stream_for_retrans_queue_from(skb
, sk
) {
1947 tp
->forward_cnt_hint
= packet_cnt
;
1948 tp
->forward_skb_hint
= skb
;
1950 /* Similar to the retransmit loop above we
1951 * can pretend that the retransmitted SKB
1952 * we send out here will be composed of one
1953 * real MSS sized packet because tcp_retransmit_skb()
1954 * will fragment it if necessary.
1956 if (++packet_cnt
> tp
->fackets_out
)
1959 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1962 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
1965 /* Ok, retransmit it. */
1966 if (tcp_retransmit_skb(sk
, skb
)) {
1967 tp
->forward_skb_hint
= NULL
;
1971 if (skb
== skb_peek(&sk
->sk_write_queue
))
1972 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1973 inet_csk(sk
)->icsk_rto
,
1976 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
1981 /* Send a fin. The caller locks the socket for us. This cannot be
1982 * allowed to fail queueing a FIN frame under any circumstances.
1984 void tcp_send_fin(struct sock
*sk
)
1986 struct tcp_sock
*tp
= tcp_sk(sk
);
1987 struct sk_buff
*skb
= skb_peek_tail(&sk
->sk_write_queue
);
1990 /* Optimization, tack on the FIN if we have a queue of
1991 * unsent frames. But be careful about outgoing SACKS
1994 mss_now
= tcp_current_mss(sk
, 1);
1996 if (sk
->sk_send_head
!= NULL
) {
1997 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
1998 TCP_SKB_CB(skb
)->end_seq
++;
2001 /* Socket is locked, keep trying until memory is available. */
2003 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
2009 /* Reserve space for headers and prepare control bits. */
2010 skb_reserve(skb
, MAX_TCP_HEADER
);
2012 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2013 TCP_SKB_CB(skb
)->sacked
= 0;
2014 skb_shinfo(skb
)->gso_segs
= 1;
2015 skb_shinfo(skb
)->gso_size
= 0;
2016 skb_shinfo(skb
)->gso_type
= 0;
2018 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2019 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
2020 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2021 tcp_queue_skb(sk
, skb
);
2023 __tcp_push_pending_frames(sk
, tp
, mss_now
, TCP_NAGLE_OFF
);
2026 /* We get here when a process closes a file descriptor (either due to
2027 * an explicit close() or as a byproduct of exit()'ing) and there
2028 * was unread data in the receive queue. This behavior is recommended
2029 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
2031 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2033 struct tcp_sock
*tp
= tcp_sk(sk
);
2034 struct sk_buff
*skb
;
2036 /* NOTE: No TCP options attached and we never retransmit this. */
2037 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2039 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2043 /* Reserve space for headers and prepare control bits. */
2044 skb_reserve(skb
, MAX_TCP_HEADER
);
2046 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2047 TCP_SKB_CB(skb
)->sacked
= 0;
2048 skb_shinfo(skb
)->gso_segs
= 1;
2049 skb_shinfo(skb
)->gso_size
= 0;
2050 skb_shinfo(skb
)->gso_type
= 0;
2053 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
, tp
);
2054 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2055 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2056 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2057 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2060 /* WARNING: This routine must only be called when we have already sent
2061 * a SYN packet that crossed the incoming SYN that caused this routine
2062 * to get called. If this assumption fails then the initial rcv_wnd
2063 * and rcv_wscale values will not be correct.
2065 int tcp_send_synack(struct sock
*sk
)
2067 struct sk_buff
* skb
;
2069 skb
= skb_peek(&sk
->sk_write_queue
);
2070 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
2071 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2074 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
2075 if (skb_cloned(skb
)) {
2076 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2079 __skb_unlink(skb
, &sk
->sk_write_queue
);
2080 skb_header_release(nskb
);
2081 __skb_queue_head(&sk
->sk_write_queue
, nskb
);
2082 sk_stream_free_skb(sk
, skb
);
2083 sk_charge_skb(sk
, nskb
);
2087 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2088 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2090 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2091 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2095 * Prepare a SYN-ACK.
2097 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2098 struct request_sock
*req
)
2100 struct inet_request_sock
*ireq
= inet_rsk(req
);
2101 struct tcp_sock
*tp
= tcp_sk(sk
);
2103 int tcp_header_size
;
2104 struct sk_buff
*skb
;
2105 #ifdef CONFIG_TCP_MD5SIG
2106 struct tcp_md5sig_key
*md5
;
2107 __u8
*md5_hash_location
;
2110 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2114 /* Reserve space for headers. */
2115 skb_reserve(skb
, MAX_TCP_HEADER
);
2117 skb
->dst
= dst_clone(dst
);
2119 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2120 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2121 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2122 /* SACK_PERM is in the place of NOP NOP of TS */
2123 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2125 #ifdef CONFIG_TCP_MD5SIG
2126 /* Are we doing MD5 on this segment? If so - make room for it */
2127 md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
2129 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
2131 skb
->h
.th
= th
= (struct tcphdr
*) skb_push(skb
, tcp_header_size
);
2133 memset(th
, 0, sizeof(struct tcphdr
));
2136 TCP_ECN_make_synack(req
, th
);
2137 th
->source
= inet_sk(sk
)->sport
;
2138 th
->dest
= ireq
->rmt_port
;
2139 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
2140 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2141 TCP_SKB_CB(skb
)->sacked
= 0;
2142 skb_shinfo(skb
)->gso_segs
= 1;
2143 skb_shinfo(skb
)->gso_size
= 0;
2144 skb_shinfo(skb
)->gso_type
= 0;
2145 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2146 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2147 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2149 /* Set this up on the first call only */
2150 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2151 /* tcp_full_space because it is guaranteed to be the first packet */
2152 tcp_select_initial_window(tcp_full_space(sk
),
2153 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2158 ireq
->rcv_wscale
= rcv_wscale
;
2161 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2162 th
->window
= htons(req
->rcv_wnd
);
2164 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2165 tcp_syn_build_options((__be32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2166 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2167 TCP_SKB_CB(skb
)->when
,
2170 #ifdef CONFIG_TCP_MD5SIG
2171 md5
? &md5_hash_location
:
2177 th
->doff
= (tcp_header_size
>> 2);
2178 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
2180 #ifdef CONFIG_TCP_MD5SIG
2181 /* Okay, we have all we need - do the md5 hash if needed */
2183 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
2186 skb
->h
.th
, sk
->sk_protocol
,
2195 * Do all connect socket setups that can be done AF independent.
2197 static void tcp_connect_init(struct sock
*sk
)
2199 struct dst_entry
*dst
= __sk_dst_get(sk
);
2200 struct tcp_sock
*tp
= tcp_sk(sk
);
2203 /* We'll fix this up when we get a response from the other end.
2204 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2206 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2207 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2209 #ifdef CONFIG_TCP_MD5SIG
2210 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2211 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2214 /* If user gave his TCP_MAXSEG, record it to clamp */
2215 if (tp
->rx_opt
.user_mss
)
2216 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2219 tcp_sync_mss(sk
, dst_mtu(dst
));
2221 if (!tp
->window_clamp
)
2222 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2223 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2224 tcp_initialize_rcv_mss(sk
);
2226 tcp_select_initial_window(tcp_full_space(sk
),
2227 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2230 sysctl_tcp_window_scaling
,
2233 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2234 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2237 sock_reset_flag(sk
, SOCK_DONE
);
2239 tcp_init_wl(tp
, tp
->write_seq
, 0);
2240 tp
->snd_una
= tp
->write_seq
;
2241 tp
->snd_sml
= tp
->write_seq
;
2246 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2247 inet_csk(sk
)->icsk_retransmits
= 0;
2248 tcp_clear_retrans(tp
);
2252 * Build a SYN and send it off.
2254 int tcp_connect(struct sock
*sk
)
2256 struct tcp_sock
*tp
= tcp_sk(sk
);
2257 struct sk_buff
*buff
;
2259 tcp_connect_init(sk
);
2261 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2262 if (unlikely(buff
== NULL
))
2265 /* Reserve space for headers. */
2266 skb_reserve(buff
, MAX_TCP_HEADER
);
2268 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
2269 TCP_ECN_send_syn(sk
, tp
, buff
);
2270 TCP_SKB_CB(buff
)->sacked
= 0;
2271 skb_shinfo(buff
)->gso_segs
= 1;
2272 skb_shinfo(buff
)->gso_size
= 0;
2273 skb_shinfo(buff
)->gso_type
= 0;
2275 tp
->snd_nxt
= tp
->write_seq
;
2276 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
2277 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
2280 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2281 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2282 skb_header_release(buff
);
2283 __skb_queue_tail(&sk
->sk_write_queue
, buff
);
2284 sk_charge_skb(sk
, buff
);
2285 tp
->packets_out
+= tcp_skb_pcount(buff
);
2286 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2288 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2289 * in order to make this packet get counted in tcpOutSegs.
2291 tp
->snd_nxt
= tp
->write_seq
;
2292 tp
->pushed_seq
= tp
->write_seq
;
2293 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
2295 /* Timer for repeating the SYN until an answer. */
2296 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2297 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2301 /* Send out a delayed ack, the caller does the policy checking
2302 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2305 void tcp_send_delayed_ack(struct sock
*sk
)
2307 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2308 int ato
= icsk
->icsk_ack
.ato
;
2309 unsigned long timeout
;
2311 if (ato
> TCP_DELACK_MIN
) {
2312 const struct tcp_sock
*tp
= tcp_sk(sk
);
2315 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2316 max_ato
= TCP_DELACK_MAX
;
2318 /* Slow path, intersegment interval is "high". */
2320 /* If some rtt estimate is known, use it to bound delayed ack.
2321 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2325 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
2331 ato
= min(ato
, max_ato
);
2334 /* Stay within the limit we were given */
2335 timeout
= jiffies
+ ato
;
2337 /* Use new timeout only if there wasn't a older one earlier. */
2338 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2339 /* If delack timer was blocked or is about to expire,
2342 if (icsk
->icsk_ack
.blocked
||
2343 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2348 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2349 timeout
= icsk
->icsk_ack
.timeout
;
2351 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2352 icsk
->icsk_ack
.timeout
= timeout
;
2353 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2356 /* This routine sends an ack and also updates the window. */
2357 void tcp_send_ack(struct sock
*sk
)
2359 /* If we have been reset, we may not send again. */
2360 if (sk
->sk_state
!= TCP_CLOSE
) {
2361 struct tcp_sock
*tp
= tcp_sk(sk
);
2362 struct sk_buff
*buff
;
2364 /* We are not putting this on the write queue, so
2365 * tcp_transmit_skb() will set the ownership to this
2368 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2370 inet_csk_schedule_ack(sk
);
2371 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2372 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2373 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2377 /* Reserve space for headers and prepare control bits. */
2378 skb_reserve(buff
, MAX_TCP_HEADER
);
2380 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
2381 TCP_SKB_CB(buff
)->sacked
= 0;
2382 skb_shinfo(buff
)->gso_segs
= 1;
2383 skb_shinfo(buff
)->gso_size
= 0;
2384 skb_shinfo(buff
)->gso_type
= 0;
2386 /* Send it off, this clears delayed acks for us. */
2387 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
, tp
);
2388 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2389 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2393 /* This routine sends a packet with an out of date sequence
2394 * number. It assumes the other end will try to ack it.
2396 * Question: what should we make while urgent mode?
2397 * 4.4BSD forces sending single byte of data. We cannot send
2398 * out of window data, because we have SND.NXT==SND.MAX...
2400 * Current solution: to send TWO zero-length segments in urgent mode:
2401 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2402 * out-of-date with SND.UNA-1 to probe window.
2404 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2406 struct tcp_sock
*tp
= tcp_sk(sk
);
2407 struct sk_buff
*skb
;
2409 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2410 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2414 /* Reserve space for headers and set control bits. */
2415 skb_reserve(skb
, MAX_TCP_HEADER
);
2417 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2418 TCP_SKB_CB(skb
)->sacked
= urgent
;
2419 skb_shinfo(skb
)->gso_segs
= 1;
2420 skb_shinfo(skb
)->gso_size
= 0;
2421 skb_shinfo(skb
)->gso_type
= 0;
2423 /* Use a previous sequence. This should cause the other
2424 * end to send an ack. Don't queue or clone SKB, just
2427 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2428 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2429 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2430 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2433 int tcp_write_wakeup(struct sock
*sk
)
2435 if (sk
->sk_state
!= TCP_CLOSE
) {
2436 struct tcp_sock
*tp
= tcp_sk(sk
);
2437 struct sk_buff
*skb
;
2439 if ((skb
= sk
->sk_send_head
) != NULL
&&
2440 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
2442 unsigned int mss
= tcp_current_mss(sk
, 0);
2443 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
2445 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2446 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2448 /* We are probing the opening of a window
2449 * but the window size is != 0
2450 * must have been a result SWS avoidance ( sender )
2452 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2454 seg_size
= min(seg_size
, mss
);
2455 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2456 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2458 } else if (!tcp_skb_pcount(skb
))
2459 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2461 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2462 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2463 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2465 update_send_head(sk
, tp
, skb
);
2470 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2471 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2472 return tcp_xmit_probe_skb(sk
, 0);
2478 /* A window probe timeout has occurred. If window is not closed send
2479 * a partial packet else a zero probe.
2481 void tcp_send_probe0(struct sock
*sk
)
2483 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2484 struct tcp_sock
*tp
= tcp_sk(sk
);
2487 err
= tcp_write_wakeup(sk
);
2489 if (tp
->packets_out
|| !sk
->sk_send_head
) {
2490 /* Cancel probe timer, if it is not required. */
2491 icsk
->icsk_probes_out
= 0;
2492 icsk
->icsk_backoff
= 0;
2497 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2498 icsk
->icsk_backoff
++;
2499 icsk
->icsk_probes_out
++;
2500 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2501 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2504 /* If packet was not sent due to local congestion,
2505 * do not backoff and do not remember icsk_probes_out.
2506 * Let local senders to fight for local resources.
2508 * Use accumulated backoff yet.
2510 if (!icsk
->icsk_probes_out
)
2511 icsk
->icsk_probes_out
= 1;
2512 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2513 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2514 TCP_RESOURCE_PROBE_INTERVAL
),
2519 EXPORT_SYMBOL(tcp_connect
);
2520 EXPORT_SYMBOL(tcp_make_synack
);
2521 EXPORT_SYMBOL(tcp_simple_retransmit
);
2522 EXPORT_SYMBOL(tcp_sync_mss
);
2523 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);
2524 EXPORT_SYMBOL(tcp_mtup_init
);