2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
8 * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Changes: Pedro Roque : Retransmit queue handled by TCP.
25 * : Fragmentation on mtu decrease
26 * : Segment collapse on retransmit
29 * Linus Torvalds : send_delayed_ack
30 * David S. Miller : Charge memory using the right skb
31 * during syn/ack processing.
32 * David S. Miller : Output engine completely rewritten.
33 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
34 * Cacophonix Gaul : draft-minshall-nagle-01
35 * J Hadi Salim : ECN support
41 #include <linux/compiler.h>
42 #include <linux/module.h>
44 /* People can turn this off for buggy TCP's found in printers etc. */
45 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
47 /* People can turn this on to work with those rare, broken TCPs that
48 * interpret the window field as a signed quantity.
50 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
52 /* This limits the percentage of the congestion window which we
53 * will allow a single TSO frame to consume. Building TSO frames
54 * which are too large can cause TCP streams to be bursty.
56 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
58 int sysctl_tcp_mtu_probing __read_mostly
= 0;
59 int sysctl_tcp_base_mss __read_mostly
= 512;
61 /* By default, RFC2861 behavior. */
62 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
64 static void update_send_head(struct sock
*sk
, struct sk_buff
*skb
)
66 struct tcp_sock
*tp
= tcp_sk(sk
);
68 tcp_advance_send_head(sk
, skb
);
69 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
70 tcp_packets_out_inc(sk
, skb
);
73 /* SND.NXT, if window was not shrunk.
74 * If window has been shrunk, what should we make? It is not clear at all.
75 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
76 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
77 * invalid. OK, let's make this for now:
79 static inline __u32
tcp_acceptable_seq(struct sock
*sk
)
81 struct tcp_sock
*tp
= tcp_sk(sk
);
83 if (!before(tp
->snd_una
+tp
->snd_wnd
, tp
->snd_nxt
))
86 return tp
->snd_una
+tp
->snd_wnd
;
89 /* Calculate mss to advertise in SYN segment.
90 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
92 * 1. It is independent of path mtu.
93 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
94 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
95 * attached devices, because some buggy hosts are confused by
97 * 4. We do not make 3, we advertise MSS, calculated from first
98 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
99 * This may be overridden via information stored in routing table.
100 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
101 * probably even Jumbo".
103 static __u16
tcp_advertise_mss(struct sock
*sk
)
105 struct tcp_sock
*tp
= tcp_sk(sk
);
106 struct dst_entry
*dst
= __sk_dst_get(sk
);
107 int mss
= tp
->advmss
;
109 if (dst
&& dst_metric(dst
, RTAX_ADVMSS
) < mss
) {
110 mss
= dst_metric(dst
, RTAX_ADVMSS
);
117 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
118 * This is the first part of cwnd validation mechanism. */
119 static void tcp_cwnd_restart(struct sock
*sk
, struct dst_entry
*dst
)
121 struct tcp_sock
*tp
= tcp_sk(sk
);
122 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
123 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
124 u32 cwnd
= tp
->snd_cwnd
;
126 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
128 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
129 restart_cwnd
= min(restart_cwnd
, cwnd
);
131 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
133 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
134 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
135 tp
->snd_cwnd_used
= 0;
138 static void tcp_event_data_sent(struct tcp_sock
*tp
,
139 struct sk_buff
*skb
, struct sock
*sk
)
141 struct inet_connection_sock
*icsk
= inet_csk(sk
);
142 const u32 now
= tcp_time_stamp
;
144 if (sysctl_tcp_slow_start_after_idle
&&
145 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
146 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
150 /* If it is a reply for ato after last received
151 * packet, enter pingpong mode.
153 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
154 icsk
->icsk_ack
.pingpong
= 1;
157 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
159 tcp_dec_quickack_mode(sk
, pkts
);
160 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
163 /* Determine a window scaling and initial window to offer.
164 * Based on the assumption that the given amount of space
165 * will be offered. Store the results in the tp structure.
166 * NOTE: for smooth operation initial space offering should
167 * be a multiple of mss if possible. We assume here that mss >= 1.
168 * This MUST be enforced by all callers.
170 void tcp_select_initial_window(int __space
, __u32 mss
,
171 __u32
*rcv_wnd
, __u32
*window_clamp
,
172 int wscale_ok
, __u8
*rcv_wscale
)
174 unsigned int space
= (__space
< 0 ? 0 : __space
);
176 /* If no clamp set the clamp to the max possible scaled window */
177 if (*window_clamp
== 0)
178 (*window_clamp
) = (65535 << 14);
179 space
= min(*window_clamp
, space
);
181 /* Quantize space offering to a multiple of mss if possible. */
183 space
= (space
/ mss
) * mss
;
185 /* NOTE: offering an initial window larger than 32767
186 * will break some buggy TCP stacks. If the admin tells us
187 * it is likely we could be speaking with such a buggy stack
188 * we will truncate our initial window offering to 32K-1
189 * unless the remote has sent us a window scaling option,
190 * which we interpret as a sign the remote TCP is not
191 * misinterpreting the window field as a signed quantity.
193 if (sysctl_tcp_workaround_signed_windows
)
194 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
200 /* Set window scaling on max possible window
201 * See RFC1323 for an explanation of the limit to 14
203 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
204 space
= min_t(u32
, space
, *window_clamp
);
205 while (space
> 65535 && (*rcv_wscale
) < 14) {
211 /* Set initial window to value enough for senders,
212 * following RFC2414. Senders, not following this RFC,
213 * will be satisfied with 2.
215 if (mss
> (1<<*rcv_wscale
)) {
221 if (*rcv_wnd
> init_cwnd
*mss
)
222 *rcv_wnd
= init_cwnd
*mss
;
225 /* Set the clamp no higher than max representable value */
226 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
229 /* Chose a new window to advertise, update state in tcp_sock for the
230 * socket, and return result with RFC1323 scaling applied. The return
231 * value can be stuffed directly into th->window for an outgoing
234 static u16
tcp_select_window(struct sock
*sk
)
236 struct tcp_sock
*tp
= tcp_sk(sk
);
237 u32 cur_win
= tcp_receive_window(tp
);
238 u32 new_win
= __tcp_select_window(sk
);
240 /* Never shrink the offered window */
241 if (new_win
< cur_win
) {
242 /* Danger Will Robinson!
243 * Don't update rcv_wup/rcv_wnd here or else
244 * we will not be able to advertise a zero
245 * window in time. --DaveM
247 * Relax Will Robinson.
251 tp
->rcv_wnd
= new_win
;
252 tp
->rcv_wup
= tp
->rcv_nxt
;
254 /* Make sure we do not exceed the maximum possible
257 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
258 new_win
= min(new_win
, MAX_TCP_WINDOW
);
260 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
262 /* RFC1323 scaling applied */
263 new_win
>>= tp
->rx_opt
.rcv_wscale
;
265 /* If we advertise zero window, disable fast path. */
272 static 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
)));
293 for (this_sack
= 0; this_sack
< tp
->rx_opt
.eff_sacks
; this_sack
++) {
294 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
295 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
298 if (tp
->rx_opt
.dsack
) {
299 tp
->rx_opt
.dsack
= 0;
300 tp
->rx_opt
.eff_sacks
--;
303 #ifdef CONFIG_TCP_MD5SIG
305 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
307 (TCPOPT_MD5SIG
<< 8) |
309 *md5_hash
= (__u8
*)ptr
;
314 /* Construct a tcp options header for a SYN or SYN_ACK packet.
315 * If this is every changed make sure to change the definition of
316 * MAX_SYN_SIZE to match the new maximum number of options that you
319 * Note - that with the RFC2385 TCP option, we make room for the
320 * 16 byte MD5 hash. This will be filled in later, so the pointer for the
321 * location to be filled is passed back up.
323 static void tcp_syn_build_options(__be32
*ptr
, int mss
, int ts
, int sack
,
324 int offer_wscale
, int wscale
, __u32 tstamp
,
325 __u32 ts_recent
, __u8
**md5_hash
)
327 /* We always get an MSS option.
328 * The option bytes which will be seen in normal data
329 * packets should timestamps be used, must be in the MSS
330 * advertised. But we subtract them from tp->mss_cache so
331 * that calculations in tcp_sendmsg are simpler etc.
332 * So account for this fact here if necessary. If we
333 * don't do this correctly, as a receiver we won't
334 * recognize data packets as being full sized when we
335 * should, and thus we won't abide by the delayed ACK
337 * SACKs don't matter, we never delay an ACK when we
338 * have any of those going out.
340 *ptr
++ = htonl((TCPOPT_MSS
<< 24) | (TCPOLEN_MSS
<< 16) | mss
);
343 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
344 (TCPOLEN_SACK_PERM
<< 16) |
345 (TCPOPT_TIMESTAMP
<< 8) |
348 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
350 (TCPOPT_TIMESTAMP
<< 8) |
352 *ptr
++ = htonl(tstamp
); /* TSVAL */
353 *ptr
++ = htonl(ts_recent
); /* TSECR */
355 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
357 (TCPOPT_SACK_PERM
<< 8) |
360 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
361 (TCPOPT_WINDOW
<< 16) |
362 (TCPOLEN_WINDOW
<< 8) |
364 #ifdef CONFIG_TCP_MD5SIG
366 * If MD5 is enabled, then we set the option, and include the size
367 * (always 18). The actual MD5 hash is added just before the
371 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
373 (TCPOPT_MD5SIG
<< 8) |
375 *md5_hash
= (__u8
*) ptr
;
380 /* This routine actually transmits TCP packets queued in by
381 * tcp_do_sendmsg(). This is used by both the initial
382 * transmission and possible later retransmissions.
383 * All SKB's seen here are completely headerless. It is our
384 * job to build the TCP header, and pass the packet down to
385 * IP so it can do the same plus pass the packet off to the
388 * We are working here with either a clone of the original
389 * SKB, or a fresh unique copy made by the retransmit engine.
391 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
, gfp_t gfp_mask
)
393 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
394 struct inet_sock
*inet
;
396 struct tcp_skb_cb
*tcb
;
398 #ifdef CONFIG_TCP_MD5SIG
399 struct tcp_md5sig_key
*md5
;
400 __u8
*md5_hash_location
;
406 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
408 /* If congestion control is doing timestamping, we must
409 * take such a timestamp before we potentially clone/copy.
411 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
412 __net_timestamp(skb
);
414 if (likely(clone_it
)) {
415 if (unlikely(skb_cloned(skb
)))
416 skb
= pskb_copy(skb
, gfp_mask
);
418 skb
= skb_clone(skb
, gfp_mask
);
425 tcb
= TCP_SKB_CB(skb
);
426 tcp_header_size
= tp
->tcp_header_len
;
428 #define SYSCTL_FLAG_TSTAMPS 0x1
429 #define SYSCTL_FLAG_WSCALE 0x2
430 #define SYSCTL_FLAG_SACK 0x4
433 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
434 tcp_header_size
= sizeof(struct tcphdr
) + TCPOLEN_MSS
;
435 if (sysctl_tcp_timestamps
) {
436 tcp_header_size
+= TCPOLEN_TSTAMP_ALIGNED
;
437 sysctl_flags
|= SYSCTL_FLAG_TSTAMPS
;
439 if (sysctl_tcp_window_scaling
) {
440 tcp_header_size
+= TCPOLEN_WSCALE_ALIGNED
;
441 sysctl_flags
|= SYSCTL_FLAG_WSCALE
;
443 if (sysctl_tcp_sack
) {
444 sysctl_flags
|= SYSCTL_FLAG_SACK
;
445 if (!(sysctl_flags
& SYSCTL_FLAG_TSTAMPS
))
446 tcp_header_size
+= TCPOLEN_SACKPERM_ALIGNED
;
448 } else if (unlikely(tp
->rx_opt
.eff_sacks
)) {
449 /* A SACK is 2 pad bytes, a 2 byte header, plus
450 * 2 32-bit sequence numbers for each SACK block.
452 tcp_header_size
+= (TCPOLEN_SACK_BASE_ALIGNED
+
453 (tp
->rx_opt
.eff_sacks
*
454 TCPOLEN_SACK_PERBLOCK
));
457 if (tcp_packets_in_flight(tp
) == 0)
458 tcp_ca_event(sk
, CA_EVENT_TX_START
);
460 #ifdef CONFIG_TCP_MD5SIG
462 * Are we doing MD5 on this segment? If so - make
465 md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
467 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
470 skb_push(skb
, tcp_header_size
);
471 skb_reset_transport_header(skb
);
472 skb_set_owner_w(skb
, sk
);
474 /* Build TCP header and checksum it. */
476 th
->source
= inet
->sport
;
477 th
->dest
= inet
->dport
;
478 th
->seq
= htonl(tcb
->seq
);
479 th
->ack_seq
= htonl(tp
->rcv_nxt
);
480 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
483 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
484 /* RFC1323: The window in SYN & SYN/ACK segments
487 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
489 th
->window
= htons(tcp_select_window(sk
));
494 if (unlikely(tp
->urg_mode
&&
495 between(tp
->snd_up
, tcb
->seq
+1, tcb
->seq
+0xFFFF))) {
496 th
->urg_ptr
= htons(tp
->snd_up
-tcb
->seq
);
500 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
501 tcp_syn_build_options((__be32
*)(th
+ 1),
502 tcp_advertise_mss(sk
),
503 (sysctl_flags
& SYSCTL_FLAG_TSTAMPS
),
504 (sysctl_flags
& SYSCTL_FLAG_SACK
),
505 (sysctl_flags
& SYSCTL_FLAG_WSCALE
),
506 tp
->rx_opt
.rcv_wscale
,
508 tp
->rx_opt
.ts_recent
,
510 #ifdef CONFIG_TCP_MD5SIG
511 md5
? &md5_hash_location
:
515 tcp_build_and_update_options((__be32
*)(th
+ 1),
517 #ifdef CONFIG_TCP_MD5SIG
518 md5
? &md5_hash_location
:
521 TCP_ECN_send(sk
, skb
, tcp_header_size
);
524 #ifdef CONFIG_TCP_MD5SIG
525 /* Calculate the MD5 hash, as we have all we need now */
527 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
536 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
538 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
539 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
541 if (skb
->len
!= tcp_header_size
)
542 tcp_event_data_sent(tp
, skb
, sk
);
544 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
545 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
547 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
548 if (likely(err
<= 0))
551 tcp_enter_cwr(sk
, 1);
553 return net_xmit_eval(err
);
555 #undef SYSCTL_FLAG_TSTAMPS
556 #undef SYSCTL_FLAG_WSCALE
557 #undef SYSCTL_FLAG_SACK
561 /* This routine just queue's the buffer
563 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
564 * otherwise socket can stall.
566 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
568 struct tcp_sock
*tp
= tcp_sk(sk
);
570 /* Advance write_seq and place onto the write_queue. */
571 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
572 skb_header_release(skb
);
573 tcp_add_write_queue_tail(sk
, skb
);
574 sk_charge_skb(sk
, skb
);
577 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
579 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
)) {
580 /* Avoid the costly divide in the normal
583 skb_shinfo(skb
)->gso_segs
= 1;
584 skb_shinfo(skb
)->gso_size
= 0;
585 skb_shinfo(skb
)->gso_type
= 0;
589 factor
= skb
->len
+ (mss_now
- 1);
591 skb_shinfo(skb
)->gso_segs
= factor
;
592 skb_shinfo(skb
)->gso_size
= mss_now
;
593 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
597 /* Function to create two new TCP segments. Shrinks the given segment
598 * to the specified size and appends a new segment with the rest of the
599 * packet to the list. This won't be called frequently, I hope.
600 * Remember, these are still headerless SKBs at this point.
602 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
, unsigned int mss_now
)
604 struct tcp_sock
*tp
= tcp_sk(sk
);
605 struct sk_buff
*buff
;
606 int nsize
, old_factor
;
610 BUG_ON(len
> skb
->len
);
612 clear_all_retrans_hints(tp
);
613 nsize
= skb_headlen(skb
) - len
;
617 if (skb_cloned(skb
) &&
618 skb_is_nonlinear(skb
) &&
619 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
622 /* Get a new skb... force flag on. */
623 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
625 return -ENOMEM
; /* We'll just try again later. */
627 sk_charge_skb(sk
, buff
);
628 nlen
= skb
->len
- len
- nsize
;
629 buff
->truesize
+= nlen
;
630 skb
->truesize
-= nlen
;
632 /* Correct the sequence numbers. */
633 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
634 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
635 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
637 /* PSH and FIN should only be set in the second packet. */
638 flags
= TCP_SKB_CB(skb
)->flags
;
639 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
640 TCP_SKB_CB(buff
)->flags
= flags
;
641 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
642 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_AT_TAIL
;
644 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
645 /* Copy and checksum data tail into the new buffer. */
646 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
, skb_put(buff
, nsize
),
651 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
653 skb
->ip_summed
= CHECKSUM_PARTIAL
;
654 skb_split(skb
, buff
, len
);
657 buff
->ip_summed
= skb
->ip_summed
;
659 /* Looks stupid, but our code really uses when of
660 * skbs, which it never sent before. --ANK
662 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
663 buff
->tstamp
= skb
->tstamp
;
665 old_factor
= tcp_skb_pcount(skb
);
667 /* Fix up tso_factor for both original and new SKB. */
668 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
669 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
671 /* If this packet has been sent out already, we must
672 * adjust the various packet counters.
674 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
675 int diff
= old_factor
- tcp_skb_pcount(skb
) -
676 tcp_skb_pcount(buff
);
678 tp
->packets_out
-= diff
;
680 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
681 tp
->sacked_out
-= diff
;
682 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
683 tp
->retrans_out
-= diff
;
685 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
) {
686 tp
->lost_out
-= diff
;
687 tp
->left_out
-= diff
;
691 /* Adjust Reno SACK estimate. */
692 if (!tp
->rx_opt
.sack_ok
) {
693 tp
->sacked_out
-= diff
;
694 if ((int)tp
->sacked_out
< 0)
696 tcp_sync_left_out(tp
);
699 tp
->fackets_out
-= diff
;
700 if ((int)tp
->fackets_out
< 0)
702 /* SACK fastpath might overwrite it unless dealt with */
703 if (tp
->fastpath_skb_hint
!= NULL
&&
704 after(TCP_SKB_CB(tp
->fastpath_skb_hint
)->seq
,
705 TCP_SKB_CB(skb
)->seq
)) {
706 tp
->fastpath_cnt_hint
-= diff
;
707 if ((int)tp
->fastpath_cnt_hint
< 0)
708 tp
->fastpath_cnt_hint
= 0;
713 /* Link BUFF into the send queue. */
714 skb_header_release(buff
);
715 tcp_insert_write_queue_after(skb
, buff
, sk
);
720 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
721 * eventually). The difference is that pulled data not copied, but
722 * immediately discarded.
724 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
730 for (i
=0; i
<skb_shinfo(skb
)->nr_frags
; i
++) {
731 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
732 put_page(skb_shinfo(skb
)->frags
[i
].page
);
733 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
735 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
737 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
738 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
744 skb_shinfo(skb
)->nr_frags
= k
;
746 skb_reset_tail_pointer(skb
);
747 skb
->data_len
-= len
;
748 skb
->len
= skb
->data_len
;
751 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
753 if (skb_cloned(skb
) &&
754 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
757 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
758 if (unlikely(len
< skb_headlen(skb
)))
759 __skb_pull(skb
, len
);
761 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
763 TCP_SKB_CB(skb
)->seq
+= len
;
764 skb
->ip_summed
= CHECKSUM_PARTIAL
;
766 skb
->truesize
-= len
;
767 sk
->sk_wmem_queued
-= len
;
768 sk
->sk_forward_alloc
+= len
;
769 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
771 /* Any change of skb->len requires recalculation of tso
774 if (tcp_skb_pcount(skb
) > 1)
775 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
, 1));
780 /* Not accounting for SACKs here. */
781 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
783 struct tcp_sock
*tp
= tcp_sk(sk
);
784 struct inet_connection_sock
*icsk
= inet_csk(sk
);
787 /* Calculate base mss without TCP options:
788 It is MMS_S - sizeof(tcphdr) of rfc1122
790 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
792 /* Clamp it (mss_clamp does not include tcp options) */
793 if (mss_now
> tp
->rx_opt
.mss_clamp
)
794 mss_now
= tp
->rx_opt
.mss_clamp
;
796 /* Now subtract optional transport overhead */
797 mss_now
-= icsk
->icsk_ext_hdr_len
;
799 /* Then reserve room for full set of TCP options and 8 bytes of data */
803 /* Now subtract TCP options size, not including SACKs */
804 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
809 /* Inverse of above */
810 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
812 struct tcp_sock
*tp
= tcp_sk(sk
);
813 struct inet_connection_sock
*icsk
= inet_csk(sk
);
818 icsk
->icsk_ext_hdr_len
+
819 icsk
->icsk_af_ops
->net_header_len
;
824 void tcp_mtup_init(struct sock
*sk
)
826 struct tcp_sock
*tp
= tcp_sk(sk
);
827 struct inet_connection_sock
*icsk
= inet_csk(sk
);
829 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
830 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
831 icsk
->icsk_af_ops
->net_header_len
;
832 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
833 icsk
->icsk_mtup
.probe_size
= 0;
836 /* This function synchronize snd mss to current pmtu/exthdr set.
838 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
839 for TCP options, but includes only bare TCP header.
841 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
842 It is minimum of user_mss and mss received with SYN.
843 It also does not include TCP options.
845 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
847 tp->mss_cache is current effective sending mss, including
848 all tcp options except for SACKs. It is evaluated,
849 taking into account current pmtu, but never exceeds
850 tp->rx_opt.mss_clamp.
852 NOTE1. rfc1122 clearly states that advertised MSS
853 DOES NOT include either tcp or ip options.
855 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
856 are READ ONLY outside this function. --ANK (980731)
859 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
861 struct tcp_sock
*tp
= tcp_sk(sk
);
862 struct inet_connection_sock
*icsk
= inet_csk(sk
);
865 if (icsk
->icsk_mtup
.search_high
> pmtu
)
866 icsk
->icsk_mtup
.search_high
= pmtu
;
868 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
870 /* Bound mss with half of window */
871 if (tp
->max_window
&& mss_now
> (tp
->max_window
>>1))
872 mss_now
= max((tp
->max_window
>>1), 68U - tp
->tcp_header_len
);
874 /* And store cached results */
875 icsk
->icsk_pmtu_cookie
= pmtu
;
876 if (icsk
->icsk_mtup
.enabled
)
877 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
878 tp
->mss_cache
= mss_now
;
883 /* Compute the current effective MSS, taking SACKs and IP options,
884 * and even PMTU discovery events into account.
886 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
887 * cannot be large. However, taking into account rare use of URG, this
890 unsigned int tcp_current_mss(struct sock
*sk
, int large_allowed
)
892 struct tcp_sock
*tp
= tcp_sk(sk
);
893 struct dst_entry
*dst
= __sk_dst_get(sk
);
898 mss_now
= tp
->mss_cache
;
900 if (large_allowed
&& sk_can_gso(sk
) && !tp
->urg_mode
)
904 u32 mtu
= dst_mtu(dst
);
905 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
906 mss_now
= tcp_sync_mss(sk
, mtu
);
909 if (tp
->rx_opt
.eff_sacks
)
910 mss_now
-= (TCPOLEN_SACK_BASE_ALIGNED
+
911 (tp
->rx_opt
.eff_sacks
* TCPOLEN_SACK_PERBLOCK
));
913 #ifdef CONFIG_TCP_MD5SIG
914 if (tp
->af_specific
->md5_lookup(sk
, sk
))
915 mss_now
-= TCPOLEN_MD5SIG_ALIGNED
;
918 xmit_size_goal
= mss_now
;
921 xmit_size_goal
= (65535 -
922 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
923 inet_csk(sk
)->icsk_ext_hdr_len
-
926 if (tp
->max_window
&&
927 (xmit_size_goal
> (tp
->max_window
>> 1)))
928 xmit_size_goal
= max((tp
->max_window
>> 1),
929 68U - tp
->tcp_header_len
);
931 xmit_size_goal
-= (xmit_size_goal
% mss_now
);
933 tp
->xmit_size_goal
= xmit_size_goal
;
938 /* Congestion window validation. (RFC2861) */
940 static void tcp_cwnd_validate(struct sock
*sk
)
942 struct tcp_sock
*tp
= tcp_sk(sk
);
943 __u32 packets_out
= tp
->packets_out
;
945 if (packets_out
>= tp
->snd_cwnd
) {
946 /* Network is feed fully. */
947 tp
->snd_cwnd_used
= 0;
948 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
950 /* Network starves. */
951 if (tp
->packets_out
> tp
->snd_cwnd_used
)
952 tp
->snd_cwnd_used
= tp
->packets_out
;
954 if (sysctl_tcp_slow_start_after_idle
&&
955 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
956 tcp_cwnd_application_limited(sk
);
960 static unsigned int tcp_window_allows(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int mss_now
, unsigned int cwnd
)
962 u32 window
, cwnd_len
;
964 window
= (tp
->snd_una
+ tp
->snd_wnd
- TCP_SKB_CB(skb
)->seq
);
965 cwnd_len
= mss_now
* cwnd
;
966 return min(window
, cwnd_len
);
969 /* Can at least one segment of SKB be sent right now, according to the
970 * congestion window rules? If so, return how many segments are allowed.
972 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
)
976 /* Don't be strict about the congestion window for the final FIN. */
977 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
978 tcp_skb_pcount(skb
) == 1)
981 in_flight
= tcp_packets_in_flight(tp
);
983 if (in_flight
< cwnd
)
984 return (cwnd
- in_flight
);
989 /* This must be invoked the first time we consider transmitting
992 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
, unsigned int mss_now
)
994 int tso_segs
= tcp_skb_pcount(skb
);
998 tcp_skb_mss(skb
) != mss_now
)) {
999 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1000 tso_segs
= tcp_skb_pcount(skb
);
1005 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1007 return after(tp
->snd_sml
,tp
->snd_una
) &&
1008 !after(tp
->snd_sml
, tp
->snd_nxt
);
1011 /* Return 0, if packet can be sent now without violation Nagle's rules:
1012 * 1. It is full sized.
1013 * 2. Or it contains FIN. (already checked by caller)
1014 * 3. Or TCP_NODELAY was set.
1015 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1016 * With Minshall's modification: all sent small packets are ACKed.
1019 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1020 const struct sk_buff
*skb
,
1021 unsigned mss_now
, int nonagle
)
1023 return (skb
->len
< mss_now
&&
1024 ((nonagle
&TCP_NAGLE_CORK
) ||
1027 tcp_minshall_check(tp
))));
1030 /* Return non-zero if the Nagle test allows this packet to be
1033 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1034 unsigned int cur_mss
, int nonagle
)
1036 /* Nagle rule does not apply to frames, which sit in the middle of the
1037 * write_queue (they have no chances to get new data).
1039 * This is implemented in the callers, where they modify the 'nonagle'
1040 * argument based upon the location of SKB in the send queue.
1042 if (nonagle
& TCP_NAGLE_PUSH
)
1045 /* Don't use the nagle rule for urgent data (or for the final FIN).
1046 * Nagle can be ignored during F-RTO too (see RFC4138).
1048 if (tp
->urg_mode
|| (tp
->frto_counter
== 2) ||
1049 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1052 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1058 /* Does at least the first segment of SKB fit into the send window? */
1059 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
, unsigned int cur_mss
)
1061 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1063 if (skb
->len
> cur_mss
)
1064 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1066 return !after(end_seq
, tp
->snd_una
+ tp
->snd_wnd
);
1069 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1070 * should be put on the wire right now. If so, it returns the number of
1071 * packets allowed by the congestion window.
1073 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1074 unsigned int cur_mss
, int nonagle
)
1076 struct tcp_sock
*tp
= tcp_sk(sk
);
1077 unsigned int cwnd_quota
;
1079 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1081 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1084 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1086 !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1092 int tcp_may_send_now(struct sock
*sk
)
1094 struct tcp_sock
*tp
= tcp_sk(sk
);
1095 struct sk_buff
*skb
= tcp_send_head(sk
);
1098 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
, 1),
1099 (tcp_skb_is_last(sk
, skb
) ?
1104 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1105 * which is put after SKB on the list. It is very much like
1106 * tcp_fragment() except that it may make several kinds of assumptions
1107 * in order to speed up the splitting operation. In particular, we
1108 * know that all the data is in scatter-gather pages, and that the
1109 * packet has never been sent out before (and thus is not cloned).
1111 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
, unsigned int mss_now
)
1113 struct sk_buff
*buff
;
1114 int nlen
= skb
->len
- len
;
1117 /* All of a TSO frame must be composed of paged data. */
1118 if (skb
->len
!= skb
->data_len
)
1119 return tcp_fragment(sk
, skb
, len
, mss_now
);
1121 buff
= sk_stream_alloc_pskb(sk
, 0, 0, GFP_ATOMIC
);
1122 if (unlikely(buff
== NULL
))
1125 sk_charge_skb(sk
, buff
);
1126 buff
->truesize
+= nlen
;
1127 skb
->truesize
-= nlen
;
1129 /* Correct the sequence numbers. */
1130 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1131 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1132 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1134 /* PSH and FIN should only be set in the second packet. */
1135 flags
= TCP_SKB_CB(skb
)->flags
;
1136 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1137 TCP_SKB_CB(buff
)->flags
= flags
;
1139 /* This packet was never sent out yet, so no SACK bits. */
1140 TCP_SKB_CB(buff
)->sacked
= 0;
1142 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1143 skb_split(skb
, buff
, len
);
1145 /* Fix up tso_factor for both original and new SKB. */
1146 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1147 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1149 /* Link BUFF into the send queue. */
1150 skb_header_release(buff
);
1151 tcp_insert_write_queue_after(skb
, buff
, sk
);
1156 /* Try to defer sending, if possible, in order to minimize the amount
1157 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1159 * This algorithm is from John Heffner.
1161 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1163 struct tcp_sock
*tp
= tcp_sk(sk
);
1164 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1165 u32 send_win
, cong_win
, limit
, in_flight
;
1167 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1170 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1173 /* Defer for less than two clock ticks. */
1174 if (!tp
->tso_deferred
&& ((jiffies
<<1)>>1) - (tp
->tso_deferred
>>1) > 1)
1177 in_flight
= tcp_packets_in_flight(tp
);
1179 BUG_ON(tcp_skb_pcount(skb
) <= 1 ||
1180 (tp
->snd_cwnd
<= in_flight
));
1182 send_win
= (tp
->snd_una
+ tp
->snd_wnd
) - TCP_SKB_CB(skb
)->seq
;
1184 /* From in_flight test above, we know that cwnd > in_flight. */
1185 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1187 limit
= min(send_win
, cong_win
);
1189 /* If a full-sized TSO skb can be sent, do it. */
1193 if (sysctl_tcp_tso_win_divisor
) {
1194 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1196 /* If at least some fraction of a window is available,
1199 chunk
/= sysctl_tcp_tso_win_divisor
;
1203 /* Different approach, try not to defer past a single
1204 * ACK. Receiver should ACK every other full sized
1205 * frame, so if we have space for more than 3 frames
1208 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1212 /* Ok, it looks like it is advisable to defer. */
1213 tp
->tso_deferred
= 1 | (jiffies
<<1);
1218 tp
->tso_deferred
= 0;
1222 /* Create a new MTU probe if we are ready.
1223 * Returns 0 if we should wait to probe (no cwnd available),
1224 * 1 if a probe was sent,
1226 static int tcp_mtu_probe(struct sock
*sk
)
1228 struct tcp_sock
*tp
= tcp_sk(sk
);
1229 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1230 struct sk_buff
*skb
, *nskb
, *next
;
1237 /* Not currently probing/verifying,
1239 * have enough cwnd, and
1240 * not SACKing (the variable headers throw things off) */
1241 if (!icsk
->icsk_mtup
.enabled
||
1242 icsk
->icsk_mtup
.probe_size
||
1243 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1244 tp
->snd_cwnd
< 11 ||
1245 tp
->rx_opt
.eff_sacks
)
1248 /* Very simple search strategy: just double the MSS. */
1249 mss_now
= tcp_current_mss(sk
, 0);
1250 probe_size
= 2*tp
->mss_cache
;
1251 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1252 /* TODO: set timer for probe_converge_event */
1256 /* Have enough data in the send queue to probe? */
1258 if ((skb
= tcp_send_head(sk
)) == NULL
)
1260 while ((len
+= skb
->len
) < probe_size
&& !tcp_skb_is_last(sk
, skb
))
1261 skb
= tcp_write_queue_next(sk
, skb
);
1262 if (len
< probe_size
)
1265 /* Receive window check. */
1266 if (after(TCP_SKB_CB(skb
)->seq
+ probe_size
, tp
->snd_una
+ tp
->snd_wnd
)) {
1267 if (tp
->snd_wnd
< probe_size
)
1273 /* Do we need to wait to drain cwnd? */
1274 pif
= tcp_packets_in_flight(tp
);
1275 if (pif
+ 2 > tp
->snd_cwnd
) {
1276 /* With no packets in flight, don't stall. */
1283 /* We're allowed to probe. Build it now. */
1284 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1286 sk_charge_skb(sk
, nskb
);
1288 skb
= tcp_send_head(sk
);
1289 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1290 tcp_advance_send_head(sk
, skb
);
1292 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1293 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1294 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1295 TCP_SKB_CB(nskb
)->sacked
= 0;
1297 nskb
->ip_summed
= skb
->ip_summed
;
1300 while (len
< probe_size
) {
1301 next
= tcp_write_queue_next(sk
, skb
);
1303 copy
= min_t(int, skb
->len
, probe_size
- len
);
1304 if (nskb
->ip_summed
)
1305 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1307 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1308 skb_put(nskb
, copy
), copy
, nskb
->csum
);
1310 if (skb
->len
<= copy
) {
1311 /* We've eaten all the data from this skb.
1313 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1314 tcp_unlink_write_queue(skb
, sk
);
1315 sk_stream_free_skb(sk
, skb
);
1317 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1318 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1319 if (!skb_shinfo(skb
)->nr_frags
) {
1320 skb_pull(skb
, copy
);
1321 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1322 skb
->csum
= csum_partial(skb
->data
, skb
->len
, 0);
1324 __pskb_trim_head(skb
, copy
);
1325 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1327 TCP_SKB_CB(skb
)->seq
+= copy
;
1333 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1335 /* We're ready to send. If this fails, the probe will
1336 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1337 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1338 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1339 /* Decrement cwnd here because we are sending
1340 * effectively two packets. */
1342 update_send_head(sk
, nskb
);
1344 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1345 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1346 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1355 /* This routine writes packets to the network. It advances the
1356 * send_head. This happens as incoming acks open up the remote
1359 * Returns 1, if no segments are in flight and we have queued segments, but
1360 * cannot send anything now because of SWS or another problem.
1362 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
)
1364 struct tcp_sock
*tp
= tcp_sk(sk
);
1365 struct sk_buff
*skb
;
1366 unsigned int tso_segs
, sent_pkts
;
1370 /* If we are closed, the bytes will have to remain here.
1371 * In time closedown will finish, we empty the write queue and all
1374 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1379 /* Do MTU probing. */
1380 if ((result
= tcp_mtu_probe(sk
)) == 0) {
1382 } else if (result
> 0) {
1386 while ((skb
= tcp_send_head(sk
))) {
1389 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1392 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1396 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1399 if (tso_segs
== 1) {
1400 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1401 (tcp_skb_is_last(sk
, skb
) ?
1402 nonagle
: TCP_NAGLE_PUSH
))))
1405 if (tcp_tso_should_defer(sk
, skb
))
1411 limit
= tcp_window_allows(tp
, skb
,
1412 mss_now
, cwnd_quota
);
1414 if (skb
->len
< limit
) {
1415 unsigned int trim
= skb
->len
% mss_now
;
1418 limit
= skb
->len
- trim
;
1422 if (skb
->len
> limit
&&
1423 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1426 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1428 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
)))
1431 /* Advance the send_head. This one is sent out.
1432 * This call will increment packets_out.
1434 update_send_head(sk
, skb
);
1436 tcp_minshall_update(tp
, mss_now
, skb
);
1440 if (likely(sent_pkts
)) {
1441 tcp_cwnd_validate(sk
);
1444 return !tp
->packets_out
&& tcp_send_head(sk
);
1447 /* Push out any pending frames which were held back due to
1448 * TCP_CORK or attempt at coalescing tiny packets.
1449 * The socket must be locked by the caller.
1451 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1454 struct sk_buff
*skb
= tcp_send_head(sk
);
1457 if (tcp_write_xmit(sk
, cur_mss
, nonagle
))
1458 tcp_check_probe_timer(sk
);
1462 /* Send _single_ skb sitting at the send head. This function requires
1463 * true push pending frames to setup probe timer etc.
1465 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1467 struct tcp_sock
*tp
= tcp_sk(sk
);
1468 struct sk_buff
*skb
= tcp_send_head(sk
);
1469 unsigned int tso_segs
, cwnd_quota
;
1471 BUG_ON(!skb
|| skb
->len
< mss_now
);
1473 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1474 cwnd_quota
= tcp_snd_test(sk
, skb
, mss_now
, TCP_NAGLE_PUSH
);
1476 if (likely(cwnd_quota
)) {
1483 limit
= tcp_window_allows(tp
, skb
,
1484 mss_now
, cwnd_quota
);
1486 if (skb
->len
< limit
) {
1487 unsigned int trim
= skb
->len
% mss_now
;
1490 limit
= skb
->len
- trim
;
1494 if (skb
->len
> limit
&&
1495 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1498 /* Send it out now. */
1499 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1501 if (likely(!tcp_transmit_skb(sk
, skb
, 1, sk
->sk_allocation
))) {
1502 update_send_head(sk
, skb
);
1503 tcp_cwnd_validate(sk
);
1509 /* This function returns the amount that we can raise the
1510 * usable window based on the following constraints
1512 * 1. The window can never be shrunk once it is offered (RFC 793)
1513 * 2. We limit memory per socket
1516 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1517 * RECV.NEXT + RCV.WIN fixed until:
1518 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1520 * i.e. don't raise the right edge of the window until you can raise
1521 * it at least MSS bytes.
1523 * Unfortunately, the recommended algorithm breaks header prediction,
1524 * since header prediction assumes th->window stays fixed.
1526 * Strictly speaking, keeping th->window fixed violates the receiver
1527 * side SWS prevention criteria. The problem is that under this rule
1528 * a stream of single byte packets will cause the right side of the
1529 * window to always advance by a single byte.
1531 * Of course, if the sender implements sender side SWS prevention
1532 * then this will not be a problem.
1534 * BSD seems to make the following compromise:
1536 * If the free space is less than the 1/4 of the maximum
1537 * space available and the free space is less than 1/2 mss,
1538 * then set the window to 0.
1539 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1540 * Otherwise, just prevent the window from shrinking
1541 * and from being larger than the largest representable value.
1543 * This prevents incremental opening of the window in the regime
1544 * where TCP is limited by the speed of the reader side taking
1545 * data out of the TCP receive queue. It does nothing about
1546 * those cases where the window is constrained on the sender side
1547 * because the pipeline is full.
1549 * BSD also seems to "accidentally" limit itself to windows that are a
1550 * multiple of MSS, at least until the free space gets quite small.
1551 * This would appear to be a side effect of the mbuf implementation.
1552 * Combining these two algorithms results in the observed behavior
1553 * of having a fixed window size at almost all times.
1555 * Below we obtain similar behavior by forcing the offered window to
1556 * a multiple of the mss when it is feasible to do so.
1558 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1559 * Regular options like TIMESTAMP are taken into account.
1561 u32
__tcp_select_window(struct sock
*sk
)
1563 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1564 struct tcp_sock
*tp
= tcp_sk(sk
);
1565 /* MSS for the peer's data. Previous versions used mss_clamp
1566 * here. I don't know if the value based on our guesses
1567 * of peer's MSS is better for the performance. It's more correct
1568 * but may be worse for the performance because of rcv_mss
1569 * fluctuations. --SAW 1998/11/1
1571 int mss
= icsk
->icsk_ack
.rcv_mss
;
1572 int free_space
= tcp_space(sk
);
1573 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1576 if (mss
> full_space
)
1579 if (free_space
< full_space
/2) {
1580 icsk
->icsk_ack
.quick
= 0;
1582 if (tcp_memory_pressure
)
1583 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U*tp
->advmss
);
1585 if (free_space
< mss
)
1589 if (free_space
> tp
->rcv_ssthresh
)
1590 free_space
= tp
->rcv_ssthresh
;
1592 /* Don't do rounding if we are using window scaling, since the
1593 * scaled window will not line up with the MSS boundary anyway.
1595 window
= tp
->rcv_wnd
;
1596 if (tp
->rx_opt
.rcv_wscale
) {
1597 window
= free_space
;
1599 /* Advertise enough space so that it won't get scaled away.
1600 * Import case: prevent zero window announcement if
1601 * 1<<rcv_wscale > mss.
1603 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1604 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1605 << tp
->rx_opt
.rcv_wscale
);
1607 /* Get the largest window that is a nice multiple of mss.
1608 * Window clamp already applied above.
1609 * If our current window offering is within 1 mss of the
1610 * free space we just keep it. This prevents the divide
1611 * and multiply from happening most of the time.
1612 * We also don't do any window rounding when the free space
1615 if (window
<= free_space
- mss
|| window
> free_space
)
1616 window
= (free_space
/mss
)*mss
;
1617 else if (mss
== full_space
&&
1618 free_space
> window
+ full_space
/2)
1619 window
= free_space
;
1625 /* Attempt to collapse two adjacent SKB's during retransmission. */
1626 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*skb
, int mss_now
)
1628 struct tcp_sock
*tp
= tcp_sk(sk
);
1629 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1631 /* The first test we must make is that neither of these two
1632 * SKB's are still referenced by someone else.
1634 if (!skb_cloned(skb
) && !skb_cloned(next_skb
)) {
1635 int skb_size
= skb
->len
, next_skb_size
= next_skb
->len
;
1636 u16 flags
= TCP_SKB_CB(skb
)->flags
;
1638 /* Also punt if next skb has been SACK'd. */
1639 if (TCP_SKB_CB(next_skb
)->sacked
& TCPCB_SACKED_ACKED
)
1642 /* Next skb is out of window. */
1643 if (after(TCP_SKB_CB(next_skb
)->end_seq
, tp
->snd_una
+tp
->snd_wnd
))
1646 /* Punt if not enough space exists in the first SKB for
1647 * the data in the second, or the total combined payload
1648 * would exceed the MSS.
1650 if ((next_skb_size
> skb_tailroom(skb
)) ||
1651 ((skb_size
+ next_skb_size
) > mss_now
))
1654 BUG_ON(tcp_skb_pcount(skb
) != 1 ||
1655 tcp_skb_pcount(next_skb
) != 1);
1657 /* changing transmit queue under us so clear hints */
1658 clear_all_retrans_hints(tp
);
1660 /* Ok. We will be able to collapse the packet. */
1661 tcp_unlink_write_queue(next_skb
, sk
);
1663 skb_copy_from_linear_data(next_skb
,
1664 skb_put(skb
, next_skb_size
),
1667 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1668 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1670 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1671 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1673 /* Update sequence range on original skb. */
1674 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1676 /* Merge over control information. */
1677 flags
|= TCP_SKB_CB(next_skb
)->flags
; /* This moves PSH/FIN etc. over */
1678 TCP_SKB_CB(skb
)->flags
= flags
;
1680 /* All done, get rid of second SKB and account for it so
1681 * packet counting does not break.
1683 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
&(TCPCB_EVER_RETRANS
|TCPCB_AT_TAIL
);
1684 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_SACKED_RETRANS
)
1685 tp
->retrans_out
-= tcp_skb_pcount(next_skb
);
1686 if (TCP_SKB_CB(next_skb
)->sacked
&TCPCB_LOST
) {
1687 tp
->lost_out
-= tcp_skb_pcount(next_skb
);
1688 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1690 /* Reno case is special. Sigh... */
1691 if (!tp
->rx_opt
.sack_ok
&& tp
->sacked_out
) {
1692 tcp_dec_pcount_approx(&tp
->sacked_out
, next_skb
);
1693 tp
->left_out
-= tcp_skb_pcount(next_skb
);
1696 /* Not quite right: it can be > snd.fack, but
1697 * it is better to underestimate fackets.
1699 tcp_dec_pcount_approx(&tp
->fackets_out
, next_skb
);
1700 tcp_packets_out_dec(tp
, next_skb
);
1701 sk_stream_free_skb(sk
, next_skb
);
1705 /* Do a simple retransmit without using the backoff mechanisms in
1706 * tcp_timer. This is used for path mtu discovery.
1707 * The socket is already locked here.
1709 void tcp_simple_retransmit(struct sock
*sk
)
1711 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1712 struct tcp_sock
*tp
= tcp_sk(sk
);
1713 struct sk_buff
*skb
;
1714 unsigned int mss
= tcp_current_mss(sk
, 0);
1717 tcp_for_write_queue(skb
, sk
) {
1718 if (skb
== tcp_send_head(sk
))
1720 if (skb
->len
> mss
&&
1721 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1722 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1723 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1724 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1726 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_LOST
)) {
1727 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1728 tp
->lost_out
+= tcp_skb_pcount(skb
);
1734 clear_all_retrans_hints(tp
);
1739 tcp_sync_left_out(tp
);
1741 /* Don't muck with the congestion window here.
1742 * Reason is that we do not increase amount of _data_
1743 * in network, but units changed and effective
1744 * cwnd/ssthresh really reduced now.
1746 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
1747 tp
->high_seq
= tp
->snd_nxt
;
1748 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1749 tp
->prior_ssthresh
= 0;
1750 tp
->undo_marker
= 0;
1751 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1753 tcp_xmit_retransmit_queue(sk
);
1756 /* This retransmits one SKB. Policy decisions and retransmit queue
1757 * state updates are done by the caller. Returns non-zero if an
1758 * error occurred which prevented the send.
1760 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
1762 struct tcp_sock
*tp
= tcp_sk(sk
);
1763 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1764 unsigned int cur_mss
= tcp_current_mss(sk
, 0);
1767 /* Inconslusive MTU probe */
1768 if (icsk
->icsk_mtup
.probe_size
) {
1769 icsk
->icsk_mtup
.probe_size
= 0;
1772 /* Do not sent more than we queued. 1/4 is reserved for possible
1773 * copying overhead: fragmentation, tunneling, mangling etc.
1775 if (atomic_read(&sk
->sk_wmem_alloc
) >
1776 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
1779 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
1780 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1782 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
1786 /* If receiver has shrunk his window, and skb is out of
1787 * new window, do not retransmit it. The exception is the
1788 * case, when window is shrunk to zero. In this case
1789 * our retransmit serves as a zero window probe.
1791 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)
1792 && TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
1795 if (skb
->len
> cur_mss
) {
1796 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
1797 return -ENOMEM
; /* We'll try again later. */
1800 /* Collapse two adjacent packets if worthwhile and we can. */
1801 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
) &&
1802 (skb
->len
< (cur_mss
>> 1)) &&
1803 (tcp_write_queue_next(sk
, skb
) != tcp_send_head(sk
)) &&
1804 (!tcp_skb_is_last(sk
, skb
)) &&
1805 (skb_shinfo(skb
)->nr_frags
== 0 && skb_shinfo(tcp_write_queue_next(sk
, skb
))->nr_frags
== 0) &&
1806 (tcp_skb_pcount(skb
) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk
, skb
)) == 1) &&
1807 (sysctl_tcp_retrans_collapse
!= 0))
1808 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
1810 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
1811 return -EHOSTUNREACH
; /* Routing failure or similar. */
1813 /* Some Solaris stacks overoptimize and ignore the FIN on a
1814 * retransmit when old data is attached. So strip it off
1815 * since it is cheap to do so and saves bytes on the network.
1818 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1819 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
1820 if (!pskb_trim(skb
, 0)) {
1821 TCP_SKB_CB(skb
)->seq
= TCP_SKB_CB(skb
)->end_seq
- 1;
1822 skb_shinfo(skb
)->gso_segs
= 1;
1823 skb_shinfo(skb
)->gso_size
= 0;
1824 skb_shinfo(skb
)->gso_type
= 0;
1825 skb
->ip_summed
= CHECKSUM_NONE
;
1830 /* Make a copy, if the first transmission SKB clone we made
1831 * is still in somebody's hands, else make a clone.
1833 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1835 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
1838 /* Update global TCP statistics. */
1839 TCP_INC_STATS(TCP_MIB_RETRANSSEGS
);
1841 tp
->total_retrans
++;
1843 #if FASTRETRANS_DEBUG > 0
1844 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) {
1845 if (net_ratelimit())
1846 printk(KERN_DEBUG
"retrans_out leaked.\n");
1849 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
1850 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1852 /* Save stamp of the first retransmit. */
1853 if (!tp
->retrans_stamp
)
1854 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
1858 /* snd_nxt is stored to detect loss of retransmitted segment,
1859 * see tcp_input.c tcp_sacktag_write_queue().
1861 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
1866 /* This gets called after a retransmit timeout, and the initially
1867 * retransmitted data is acknowledged. It tries to continue
1868 * resending the rest of the retransmit queue, until either
1869 * we've sent it all or the congestion window limit is reached.
1870 * If doing SACK, the first ACK which comes back for a timeout
1871 * based retransmit packet might feed us FACK information again.
1872 * If so, we use it to avoid unnecessarily retransmissions.
1874 void tcp_xmit_retransmit_queue(struct sock
*sk
)
1876 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1877 struct tcp_sock
*tp
= tcp_sk(sk
);
1878 struct sk_buff
*skb
;
1881 if (tp
->retransmit_skb_hint
) {
1882 skb
= tp
->retransmit_skb_hint
;
1883 packet_cnt
= tp
->retransmit_cnt_hint
;
1885 skb
= tcp_write_queue_head(sk
);
1889 /* First pass: retransmit lost packets. */
1891 tcp_for_write_queue_from(skb
, sk
) {
1892 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1894 if (skb
== tcp_send_head(sk
))
1896 /* we could do better than to assign each time */
1897 tp
->retransmit_skb_hint
= skb
;
1898 tp
->retransmit_cnt_hint
= packet_cnt
;
1900 /* Assume this retransmit will generate
1901 * only one packet for congestion window
1902 * calculation purposes. This works because
1903 * tcp_retransmit_skb() will chop up the
1904 * packet to be MSS sized and all the
1905 * packet counting works out.
1907 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1910 if (sacked
& TCPCB_LOST
) {
1911 if (!(sacked
&(TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))) {
1912 if (tcp_retransmit_skb(sk
, skb
)) {
1913 tp
->retransmit_skb_hint
= NULL
;
1916 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1917 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS
);
1919 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS
);
1921 if (skb
== tcp_write_queue_head(sk
))
1922 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1923 inet_csk(sk
)->icsk_rto
,
1927 packet_cnt
+= tcp_skb_pcount(skb
);
1928 if (packet_cnt
>= tp
->lost_out
)
1934 /* OK, demanded retransmission is finished. */
1936 /* Forward retransmissions are possible only during Recovery. */
1937 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1940 /* No forward retransmissions in Reno are possible. */
1941 if (!tp
->rx_opt
.sack_ok
)
1944 /* Yeah, we have to make difficult choice between forward transmission
1945 * and retransmission... Both ways have their merits...
1947 * For now we do not retransmit anything, while we have some new
1951 if (tcp_may_send_now(sk
))
1954 if (tp
->forward_skb_hint
) {
1955 skb
= tp
->forward_skb_hint
;
1956 packet_cnt
= tp
->forward_cnt_hint
;
1958 skb
= tcp_write_queue_head(sk
);
1962 tcp_for_write_queue_from(skb
, sk
) {
1963 if (skb
== tcp_send_head(sk
))
1965 tp
->forward_cnt_hint
= packet_cnt
;
1966 tp
->forward_skb_hint
= skb
;
1968 /* Similar to the retransmit loop above we
1969 * can pretend that the retransmitted SKB
1970 * we send out here will be composed of one
1971 * real MSS sized packet because tcp_retransmit_skb()
1972 * will fragment it if necessary.
1974 if (++packet_cnt
> tp
->fackets_out
)
1977 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
1980 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
)
1983 /* Ok, retransmit it. */
1984 if (tcp_retransmit_skb(sk
, skb
)) {
1985 tp
->forward_skb_hint
= NULL
;
1989 if (skb
== tcp_write_queue_head(sk
))
1990 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1991 inet_csk(sk
)->icsk_rto
,
1994 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS
);
1999 /* Send a fin. The caller locks the socket for us. This cannot be
2000 * allowed to fail queueing a FIN frame under any circumstances.
2002 void tcp_send_fin(struct sock
*sk
)
2004 struct tcp_sock
*tp
= tcp_sk(sk
);
2005 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2008 /* Optimization, tack on the FIN if we have a queue of
2009 * unsent frames. But be careful about outgoing SACKS
2012 mss_now
= tcp_current_mss(sk
, 1);
2014 if (tcp_send_head(sk
) != NULL
) {
2015 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2016 TCP_SKB_CB(skb
)->end_seq
++;
2019 /* Socket is locked, keep trying until memory is available. */
2021 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, GFP_KERNEL
);
2027 /* Reserve space for headers and prepare control bits. */
2028 skb_reserve(skb
, MAX_TCP_HEADER
);
2030 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2031 TCP_SKB_CB(skb
)->sacked
= 0;
2032 skb_shinfo(skb
)->gso_segs
= 1;
2033 skb_shinfo(skb
)->gso_size
= 0;
2034 skb_shinfo(skb
)->gso_type
= 0;
2036 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2037 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
2038 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2039 tcp_queue_skb(sk
, skb
);
2041 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2044 /* We get here when a process closes a file descriptor (either due to
2045 * an explicit close() or as a byproduct of exit()'ing) and there
2046 * was unread data in the receive queue. This behavior is recommended
2047 * by RFC 2525, section 2.17. -DaveM
2049 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2051 struct sk_buff
*skb
;
2053 /* NOTE: No TCP options attached and we never retransmit this. */
2054 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2056 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2060 /* Reserve space for headers and prepare control bits. */
2061 skb_reserve(skb
, MAX_TCP_HEADER
);
2063 TCP_SKB_CB(skb
)->flags
= (TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2064 TCP_SKB_CB(skb
)->sacked
= 0;
2065 skb_shinfo(skb
)->gso_segs
= 1;
2066 skb_shinfo(skb
)->gso_size
= 0;
2067 skb_shinfo(skb
)->gso_type
= 0;
2070 TCP_SKB_CB(skb
)->seq
= tcp_acceptable_seq(sk
);
2071 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2072 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2073 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2074 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED
);
2077 /* WARNING: This routine must only be called when we have already sent
2078 * a SYN packet that crossed the incoming SYN that caused this routine
2079 * to get called. If this assumption fails then the initial rcv_wnd
2080 * and rcv_wscale values will not be correct.
2082 int tcp_send_synack(struct sock
*sk
)
2084 struct sk_buff
* skb
;
2086 skb
= tcp_write_queue_head(sk
);
2087 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_SYN
)) {
2088 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2091 if (!(TCP_SKB_CB(skb
)->flags
&TCPCB_FLAG_ACK
)) {
2092 if (skb_cloned(skb
)) {
2093 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2096 tcp_unlink_write_queue(skb
, sk
);
2097 skb_header_release(nskb
);
2098 __tcp_add_write_queue_head(sk
, nskb
);
2099 sk_stream_free_skb(sk
, skb
);
2100 sk_charge_skb(sk
, nskb
);
2104 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2105 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2107 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2108 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2112 * Prepare a SYN-ACK.
2114 struct sk_buff
* tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2115 struct request_sock
*req
)
2117 struct inet_request_sock
*ireq
= inet_rsk(req
);
2118 struct tcp_sock
*tp
= tcp_sk(sk
);
2120 int tcp_header_size
;
2121 struct sk_buff
*skb
;
2122 #ifdef CONFIG_TCP_MD5SIG
2123 struct tcp_md5sig_key
*md5
;
2124 __u8
*md5_hash_location
;
2127 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2131 /* Reserve space for headers. */
2132 skb_reserve(skb
, MAX_TCP_HEADER
);
2134 skb
->dst
= dst_clone(dst
);
2136 tcp_header_size
= (sizeof(struct tcphdr
) + TCPOLEN_MSS
+
2137 (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0) +
2138 (ireq
->wscale_ok
? TCPOLEN_WSCALE_ALIGNED
: 0) +
2139 /* SACK_PERM is in the place of NOP NOP of TS */
2140 ((ireq
->sack_ok
&& !ireq
->tstamp_ok
) ? TCPOLEN_SACKPERM_ALIGNED
: 0));
2142 #ifdef CONFIG_TCP_MD5SIG
2143 /* Are we doing MD5 on this segment? If so - make room for it */
2144 md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
2146 tcp_header_size
+= TCPOLEN_MD5SIG_ALIGNED
;
2148 skb_push(skb
, tcp_header_size
);
2149 skb_reset_transport_header(skb
);
2152 memset(th
, 0, sizeof(struct tcphdr
));
2155 TCP_ECN_make_synack(req
, th
);
2156 th
->source
= inet_sk(sk
)->sport
;
2157 th
->dest
= ireq
->rmt_port
;
2158 TCP_SKB_CB(skb
)->seq
= tcp_rsk(req
)->snt_isn
;
2159 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ 1;
2160 TCP_SKB_CB(skb
)->sacked
= 0;
2161 skb_shinfo(skb
)->gso_segs
= 1;
2162 skb_shinfo(skb
)->gso_size
= 0;
2163 skb_shinfo(skb
)->gso_type
= 0;
2164 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2165 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2166 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2168 /* Set this up on the first call only */
2169 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2170 /* tcp_full_space because it is guaranteed to be the first packet */
2171 tcp_select_initial_window(tcp_full_space(sk
),
2172 dst_metric(dst
, RTAX_ADVMSS
) - (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2177 ireq
->rcv_wscale
= rcv_wscale
;
2180 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2181 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2183 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2184 tcp_syn_build_options((__be32
*)(th
+ 1), dst_metric(dst
, RTAX_ADVMSS
), ireq
->tstamp_ok
,
2185 ireq
->sack_ok
, ireq
->wscale_ok
, ireq
->rcv_wscale
,
2186 TCP_SKB_CB(skb
)->when
,
2189 #ifdef CONFIG_TCP_MD5SIG
2190 md5
? &md5_hash_location
:
2196 th
->doff
= (tcp_header_size
>> 2);
2197 TCP_INC_STATS(TCP_MIB_OUTSEGS
);
2199 #ifdef CONFIG_TCP_MD5SIG
2200 /* Okay, we have all we need - do the md5 hash if needed */
2202 tp
->af_specific
->calc_md5_hash(md5_hash_location
,
2205 tcp_hdr(skb
), sk
->sk_protocol
,
2214 * Do all connect socket setups that can be done AF independent.
2216 static void tcp_connect_init(struct sock
*sk
)
2218 struct dst_entry
*dst
= __sk_dst_get(sk
);
2219 struct tcp_sock
*tp
= tcp_sk(sk
);
2222 /* We'll fix this up when we get a response from the other end.
2223 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2225 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2226 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2228 #ifdef CONFIG_TCP_MD5SIG
2229 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2230 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2233 /* If user gave his TCP_MAXSEG, record it to clamp */
2234 if (tp
->rx_opt
.user_mss
)
2235 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2238 tcp_sync_mss(sk
, dst_mtu(dst
));
2240 if (!tp
->window_clamp
)
2241 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2242 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2243 tcp_initialize_rcv_mss(sk
);
2245 tcp_select_initial_window(tcp_full_space(sk
),
2246 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2249 sysctl_tcp_window_scaling
,
2252 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2253 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2256 sock_reset_flag(sk
, SOCK_DONE
);
2258 tcp_init_wl(tp
, tp
->write_seq
, 0);
2259 tp
->snd_una
= tp
->write_seq
;
2260 tp
->snd_sml
= tp
->write_seq
;
2265 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2266 inet_csk(sk
)->icsk_retransmits
= 0;
2267 tcp_clear_retrans(tp
);
2271 * Build a SYN and send it off.
2273 int tcp_connect(struct sock
*sk
)
2275 struct tcp_sock
*tp
= tcp_sk(sk
);
2276 struct sk_buff
*buff
;
2278 tcp_connect_init(sk
);
2280 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2281 if (unlikely(buff
== NULL
))
2284 /* Reserve space for headers. */
2285 skb_reserve(buff
, MAX_TCP_HEADER
);
2287 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_SYN
;
2288 TCP_ECN_send_syn(sk
, buff
);
2289 TCP_SKB_CB(buff
)->sacked
= 0;
2290 skb_shinfo(buff
)->gso_segs
= 1;
2291 skb_shinfo(buff
)->gso_size
= 0;
2292 skb_shinfo(buff
)->gso_type
= 0;
2294 tp
->snd_nxt
= tp
->write_seq
;
2295 TCP_SKB_CB(buff
)->seq
= tp
->write_seq
++;
2296 TCP_SKB_CB(buff
)->end_seq
= tp
->write_seq
;
2299 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2300 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2301 skb_header_release(buff
);
2302 __tcp_add_write_queue_tail(sk
, buff
);
2303 sk_charge_skb(sk
, buff
);
2304 tp
->packets_out
+= tcp_skb_pcount(buff
);
2305 tcp_transmit_skb(sk
, buff
, 1, GFP_KERNEL
);
2307 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2308 * in order to make this packet get counted in tcpOutSegs.
2310 tp
->snd_nxt
= tp
->write_seq
;
2311 tp
->pushed_seq
= tp
->write_seq
;
2312 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS
);
2314 /* Timer for repeating the SYN until an answer. */
2315 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2316 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2320 /* Send out a delayed ack, the caller does the policy checking
2321 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2324 void tcp_send_delayed_ack(struct sock
*sk
)
2326 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2327 int ato
= icsk
->icsk_ack
.ato
;
2328 unsigned long timeout
;
2330 if (ato
> TCP_DELACK_MIN
) {
2331 const struct tcp_sock
*tp
= tcp_sk(sk
);
2334 if (icsk
->icsk_ack
.pingpong
|| (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2335 max_ato
= TCP_DELACK_MAX
;
2337 /* Slow path, intersegment interval is "high". */
2339 /* If some rtt estimate is known, use it to bound delayed ack.
2340 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2344 int rtt
= max(tp
->srtt
>>3, TCP_DELACK_MIN
);
2350 ato
= min(ato
, max_ato
);
2353 /* Stay within the limit we were given */
2354 timeout
= jiffies
+ ato
;
2356 /* Use new timeout only if there wasn't a older one earlier. */
2357 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2358 /* If delack timer was blocked or is about to expire,
2361 if (icsk
->icsk_ack
.blocked
||
2362 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2367 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2368 timeout
= icsk
->icsk_ack
.timeout
;
2370 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2371 icsk
->icsk_ack
.timeout
= timeout
;
2372 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2375 /* This routine sends an ack and also updates the window. */
2376 void tcp_send_ack(struct sock
*sk
)
2378 /* If we have been reset, we may not send again. */
2379 if (sk
->sk_state
!= TCP_CLOSE
) {
2380 struct sk_buff
*buff
;
2382 /* We are not putting this on the write queue, so
2383 * tcp_transmit_skb() will set the ownership to this
2386 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2388 inet_csk_schedule_ack(sk
);
2389 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2390 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2391 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2395 /* Reserve space for headers and prepare control bits. */
2396 skb_reserve(buff
, MAX_TCP_HEADER
);
2398 TCP_SKB_CB(buff
)->flags
= TCPCB_FLAG_ACK
;
2399 TCP_SKB_CB(buff
)->sacked
= 0;
2400 skb_shinfo(buff
)->gso_segs
= 1;
2401 skb_shinfo(buff
)->gso_size
= 0;
2402 skb_shinfo(buff
)->gso_type
= 0;
2404 /* Send it off, this clears delayed acks for us. */
2405 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(buff
)->end_seq
= tcp_acceptable_seq(sk
);
2406 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2407 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2411 /* This routine sends a packet with an out of date sequence
2412 * number. It assumes the other end will try to ack it.
2414 * Question: what should we make while urgent mode?
2415 * 4.4BSD forces sending single byte of data. We cannot send
2416 * out of window data, because we have SND.NXT==SND.MAX...
2418 * Current solution: to send TWO zero-length segments in urgent mode:
2419 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2420 * out-of-date with SND.UNA-1 to probe window.
2422 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2424 struct tcp_sock
*tp
= tcp_sk(sk
);
2425 struct sk_buff
*skb
;
2427 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2428 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2432 /* Reserve space for headers and set control bits. */
2433 skb_reserve(skb
, MAX_TCP_HEADER
);
2435 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
2436 TCP_SKB_CB(skb
)->sacked
= urgent
;
2437 skb_shinfo(skb
)->gso_segs
= 1;
2438 skb_shinfo(skb
)->gso_size
= 0;
2439 skb_shinfo(skb
)->gso_type
= 0;
2441 /* Use a previous sequence. This should cause the other
2442 * end to send an ack. Don't queue or clone SKB, just
2445 TCP_SKB_CB(skb
)->seq
= urgent
? tp
->snd_una
: tp
->snd_una
- 1;
2446 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
;
2447 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2448 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2451 int tcp_write_wakeup(struct sock
*sk
)
2453 if (sk
->sk_state
!= TCP_CLOSE
) {
2454 struct tcp_sock
*tp
= tcp_sk(sk
);
2455 struct sk_buff
*skb
;
2457 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2458 before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
+tp
->snd_wnd
)) {
2460 unsigned int mss
= tcp_current_mss(sk
, 0);
2461 unsigned int seg_size
= tp
->snd_una
+tp
->snd_wnd
-TCP_SKB_CB(skb
)->seq
;
2463 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2464 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2466 /* We are probing the opening of a window
2467 * but the window size is != 0
2468 * must have been a result SWS avoidance ( sender )
2470 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2472 seg_size
= min(seg_size
, mss
);
2473 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2474 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2476 } else if (!tcp_skb_pcount(skb
))
2477 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2479 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2480 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2481 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2483 update_send_head(sk
, skb
);
2488 between(tp
->snd_up
, tp
->snd_una
+1, tp
->snd_una
+0xFFFF))
2489 tcp_xmit_probe_skb(sk
, TCPCB_URG
);
2490 return tcp_xmit_probe_skb(sk
, 0);
2496 /* A window probe timeout has occurred. If window is not closed send
2497 * a partial packet else a zero probe.
2499 void tcp_send_probe0(struct sock
*sk
)
2501 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2502 struct tcp_sock
*tp
= tcp_sk(sk
);
2505 err
= tcp_write_wakeup(sk
);
2507 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2508 /* Cancel probe timer, if it is not required. */
2509 icsk
->icsk_probes_out
= 0;
2510 icsk
->icsk_backoff
= 0;
2515 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2516 icsk
->icsk_backoff
++;
2517 icsk
->icsk_probes_out
++;
2518 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2519 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2522 /* If packet was not sent due to local congestion,
2523 * do not backoff and do not remember icsk_probes_out.
2524 * Let local senders to fight for local resources.
2526 * Use accumulated backoff yet.
2528 if (!icsk
->icsk_probes_out
)
2529 icsk
->icsk_probes_out
= 1;
2530 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2531 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2532 TCP_RESOURCE_PROBE_INTERVAL
),
2537 EXPORT_SYMBOL(tcp_connect
);
2538 EXPORT_SYMBOL(tcp_make_synack
);
2539 EXPORT_SYMBOL(tcp_simple_retransmit
);
2540 EXPORT_SYMBOL(tcp_sync_mss
);
2541 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor
);
2542 EXPORT_SYMBOL(tcp_mtup_init
);