2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #include <linux/module.h>
66 #include <linux/sysctl.h>
69 #include <net/inet_common.h>
70 #include <linux/ipsec.h>
71 #include <asm/unaligned.h>
72 #include <net/netdma.h>
74 int sysctl_tcp_timestamps __read_mostly
= 1;
75 int sysctl_tcp_window_scaling __read_mostly
= 1;
76 int sysctl_tcp_sack __read_mostly
= 1;
77 int sysctl_tcp_fack __read_mostly
= 1;
78 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
79 int sysctl_tcp_ecn __read_mostly
;
80 int sysctl_tcp_dsack __read_mostly
= 1;
81 int sysctl_tcp_app_win __read_mostly
= 31;
82 int sysctl_tcp_adv_win_scale __read_mostly
= 2;
84 int sysctl_tcp_stdurg __read_mostly
;
85 int sysctl_tcp_rfc1337 __read_mostly
;
86 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
87 int sysctl_tcp_frto __read_mostly
= 2;
88 int sysctl_tcp_frto_response __read_mostly
;
89 int sysctl_tcp_nometrics_save __read_mostly
;
91 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
92 int sysctl_tcp_abc __read_mostly
;
94 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
95 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
96 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
97 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
98 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
99 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
100 #define FLAG_ECE 0x40 /* ECE in this ACK */
101 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
102 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
103 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
104 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
105 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
106 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
107 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
109 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
110 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
111 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
112 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
113 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
115 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
116 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
118 /* Adapt the MSS value used to make delayed ack decision to the
121 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
123 struct inet_connection_sock
*icsk
= inet_csk(sk
);
124 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
127 icsk
->icsk_ack
.last_seg_size
= 0;
129 /* skb->len may jitter because of SACKs, even if peer
130 * sends good full-sized frames.
132 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
133 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
134 icsk
->icsk_ack
.rcv_mss
= len
;
136 /* Otherwise, we make more careful check taking into account,
137 * that SACKs block is variable.
139 * "len" is invariant segment length, including TCP header.
141 len
+= skb
->data
- skb_transport_header(skb
);
142 if (len
>= TCP_MIN_RCVMSS
+ sizeof(struct tcphdr
) ||
143 /* If PSH is not set, packet should be
144 * full sized, provided peer TCP is not badly broken.
145 * This observation (if it is correct 8)) allows
146 * to handle super-low mtu links fairly.
148 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
149 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
150 /* Subtract also invariant (if peer is RFC compliant),
151 * tcp header plus fixed timestamp option length.
152 * Resulting "len" is MSS free of SACK jitter.
154 len
-= tcp_sk(sk
)->tcp_header_len
;
155 icsk
->icsk_ack
.last_seg_size
= len
;
157 icsk
->icsk_ack
.rcv_mss
= len
;
161 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
162 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
163 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
167 static void tcp_incr_quickack(struct sock
*sk
)
169 struct inet_connection_sock
*icsk
= inet_csk(sk
);
170 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
174 if (quickacks
> icsk
->icsk_ack
.quick
)
175 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
178 void tcp_enter_quickack_mode(struct sock
*sk
)
180 struct inet_connection_sock
*icsk
= inet_csk(sk
);
181 tcp_incr_quickack(sk
);
182 icsk
->icsk_ack
.pingpong
= 0;
183 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
186 /* Send ACKs quickly, if "quick" count is not exhausted
187 * and the session is not interactive.
190 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
192 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
193 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
196 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
198 if (tp
->ecn_flags
& TCP_ECN_OK
)
199 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
202 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, struct sk_buff
*skb
)
204 if (tcp_hdr(skb
)->cwr
)
205 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
208 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
210 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
213 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, struct sk_buff
*skb
)
215 if (tp
->ecn_flags
& TCP_ECN_OK
) {
216 if (INET_ECN_is_ce(TCP_SKB_CB(skb
)->flags
))
217 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
218 /* Funny extension: if ECT is not set on a segment,
219 * it is surely retransmit. It is not in ECN RFC,
220 * but Linux follows this rule. */
221 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb
)->flags
)))
222 tcp_enter_quickack_mode((struct sock
*)tp
);
226 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, struct tcphdr
*th
)
228 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
229 tp
->ecn_flags
&= ~TCP_ECN_OK
;
232 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, struct tcphdr
*th
)
234 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
235 tp
->ecn_flags
&= ~TCP_ECN_OK
;
238 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock
*tp
, struct tcphdr
*th
)
240 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
245 /* Buffer size and advertised window tuning.
247 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
250 static void tcp_fixup_sndbuf(struct sock
*sk
)
252 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
253 sizeof(struct sk_buff
);
255 if (sk
->sk_sndbuf
< 3 * sndmem
)
256 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
259 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
261 * All tcp_full_space() is split to two parts: "network" buffer, allocated
262 * forward and advertised in receiver window (tp->rcv_wnd) and
263 * "application buffer", required to isolate scheduling/application
264 * latencies from network.
265 * window_clamp is maximal advertised window. It can be less than
266 * tcp_full_space(), in this case tcp_full_space() - window_clamp
267 * is reserved for "application" buffer. The less window_clamp is
268 * the smoother our behaviour from viewpoint of network, but the lower
269 * throughput and the higher sensitivity of the connection to losses. 8)
271 * rcv_ssthresh is more strict window_clamp used at "slow start"
272 * phase to predict further behaviour of this connection.
273 * It is used for two goals:
274 * - to enforce header prediction at sender, even when application
275 * requires some significant "application buffer". It is check #1.
276 * - to prevent pruning of receive queue because of misprediction
277 * of receiver window. Check #2.
279 * The scheme does not work when sender sends good segments opening
280 * window and then starts to feed us spaghetti. But it should work
281 * in common situations. Otherwise, we have to rely on queue collapsing.
284 /* Slow part of check#2. */
285 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
287 struct tcp_sock
*tp
= tcp_sk(sk
);
289 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
290 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
292 while (tp
->rcv_ssthresh
<= window
) {
293 if (truesize
<= skb
->len
)
294 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
302 static void tcp_grow_window(struct sock
*sk
, struct sk_buff
*skb
)
304 struct tcp_sock
*tp
= tcp_sk(sk
);
307 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
308 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
309 !tcp_memory_pressure
) {
312 /* Check #2. Increase window, if skb with such overhead
313 * will fit to rcvbuf in future.
315 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
316 incr
= 2 * tp
->advmss
;
318 incr
= __tcp_grow_window(sk
, skb
);
321 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
323 inet_csk(sk
)->icsk_ack
.quick
|= 1;
328 /* 3. Tuning rcvbuf, when connection enters established state. */
330 static void tcp_fixup_rcvbuf(struct sock
*sk
)
332 struct tcp_sock
*tp
= tcp_sk(sk
);
333 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
335 /* Try to select rcvbuf so that 4 mss-sized segments
336 * will fit to window and corresponding skbs will fit to our rcvbuf.
337 * (was 3; 4 is minimum to allow fast retransmit to work.)
339 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
341 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
342 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
345 /* 4. Try to fixup all. It is made immediately after connection enters
348 static void tcp_init_buffer_space(struct sock
*sk
)
350 struct tcp_sock
*tp
= tcp_sk(sk
);
353 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
354 tcp_fixup_rcvbuf(sk
);
355 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
356 tcp_fixup_sndbuf(sk
);
358 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
360 maxwin
= tcp_full_space(sk
);
362 if (tp
->window_clamp
>= maxwin
) {
363 tp
->window_clamp
= maxwin
;
365 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
366 tp
->window_clamp
= max(maxwin
-
367 (maxwin
>> sysctl_tcp_app_win
),
371 /* Force reservation of one segment. */
372 if (sysctl_tcp_app_win
&&
373 tp
->window_clamp
> 2 * tp
->advmss
&&
374 tp
->window_clamp
+ tp
->advmss
> maxwin
)
375 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
377 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
378 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
381 /* 5. Recalculate window clamp after socket hit its memory bounds. */
382 static void tcp_clamp_window(struct sock
*sk
)
384 struct tcp_sock
*tp
= tcp_sk(sk
);
385 struct inet_connection_sock
*icsk
= inet_csk(sk
);
387 icsk
->icsk_ack
.quick
= 0;
389 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
390 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
391 !tcp_memory_pressure
&&
392 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
393 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
396 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
397 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
400 /* Initialize RCV_MSS value.
401 * RCV_MSS is an our guess about MSS used by the peer.
402 * We haven't any direct information about the MSS.
403 * It's better to underestimate the RCV_MSS rather than overestimate.
404 * Overestimations make us ACKing less frequently than needed.
405 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
407 void tcp_initialize_rcv_mss(struct sock
*sk
)
409 struct tcp_sock
*tp
= tcp_sk(sk
);
410 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
412 hint
= min(hint
, tp
->rcv_wnd
/ 2);
413 hint
= min(hint
, TCP_MIN_RCVMSS
);
414 hint
= max(hint
, TCP_MIN_MSS
);
416 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
419 /* Receiver "autotuning" code.
421 * The algorithm for RTT estimation w/o timestamps is based on
422 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
423 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
425 * More detail on this code can be found at
426 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
427 * though this reference is out of date. A new paper
430 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
432 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
438 if (new_sample
!= 0) {
439 /* If we sample in larger samples in the non-timestamp
440 * case, we could grossly overestimate the RTT especially
441 * with chatty applications or bulk transfer apps which
442 * are stalled on filesystem I/O.
444 * Also, since we are only going for a minimum in the
445 * non-timestamp case, we do not smooth things out
446 * else with timestamps disabled convergence takes too
450 m
-= (new_sample
>> 3);
452 } else if (m
< new_sample
)
455 /* No previous measure. */
459 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
460 tp
->rcv_rtt_est
.rtt
= new_sample
;
463 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
465 if (tp
->rcv_rtt_est
.time
== 0)
467 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
469 tcp_rcv_rtt_update(tp
, jiffies
- tp
->rcv_rtt_est
.time
, 1);
472 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
473 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
476 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
477 const struct sk_buff
*skb
)
479 struct tcp_sock
*tp
= tcp_sk(sk
);
480 if (tp
->rx_opt
.rcv_tsecr
&&
481 (TCP_SKB_CB(skb
)->end_seq
-
482 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
483 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
487 * This function should be called every time data is copied to user space.
488 * It calculates the appropriate TCP receive buffer space.
490 void tcp_rcv_space_adjust(struct sock
*sk
)
492 struct tcp_sock
*tp
= tcp_sk(sk
);
496 if (tp
->rcvq_space
.time
== 0)
499 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
500 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
503 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
505 space
= max(tp
->rcvq_space
.space
, space
);
507 if (tp
->rcvq_space
.space
!= space
) {
510 tp
->rcvq_space
.space
= space
;
512 if (sysctl_tcp_moderate_rcvbuf
&&
513 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
514 int new_clamp
= space
;
516 /* Receive space grows, normalize in order to
517 * take into account packet headers and sk_buff
518 * structure overhead.
523 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
524 16 + sizeof(struct sk_buff
));
525 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
528 space
= min(space
, sysctl_tcp_rmem
[2]);
529 if (space
> sk
->sk_rcvbuf
) {
530 sk
->sk_rcvbuf
= space
;
532 /* Make the window clamp follow along. */
533 tp
->window_clamp
= new_clamp
;
539 tp
->rcvq_space
.seq
= tp
->copied_seq
;
540 tp
->rcvq_space
.time
= tcp_time_stamp
;
543 /* There is something which you must keep in mind when you analyze the
544 * behavior of the tp->ato delayed ack timeout interval. When a
545 * connection starts up, we want to ack as quickly as possible. The
546 * problem is that "good" TCP's do slow start at the beginning of data
547 * transmission. The means that until we send the first few ACK's the
548 * sender will sit on his end and only queue most of his data, because
549 * he can only send snd_cwnd unacked packets at any given time. For
550 * each ACK we send, he increments snd_cwnd and transmits more of his
553 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
555 struct tcp_sock
*tp
= tcp_sk(sk
);
556 struct inet_connection_sock
*icsk
= inet_csk(sk
);
559 inet_csk_schedule_ack(sk
);
561 tcp_measure_rcv_mss(sk
, skb
);
563 tcp_rcv_rtt_measure(tp
);
565 now
= tcp_time_stamp
;
567 if (!icsk
->icsk_ack
.ato
) {
568 /* The _first_ data packet received, initialize
569 * delayed ACK engine.
571 tcp_incr_quickack(sk
);
572 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
574 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
576 if (m
<= TCP_ATO_MIN
/ 2) {
577 /* The fastest case is the first. */
578 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
579 } else if (m
< icsk
->icsk_ack
.ato
) {
580 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
581 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
582 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
583 } else if (m
> icsk
->icsk_rto
) {
584 /* Too long gap. Apparently sender failed to
585 * restart window, so that we send ACKs quickly.
587 tcp_incr_quickack(sk
);
591 icsk
->icsk_ack
.lrcvtime
= now
;
593 TCP_ECN_check_ce(tp
, skb
);
596 tcp_grow_window(sk
, skb
);
599 static u32
tcp_rto_min(struct sock
*sk
)
601 struct dst_entry
*dst
= __sk_dst_get(sk
);
602 u32 rto_min
= TCP_RTO_MIN
;
604 if (dst
&& dst_metric_locked(dst
, RTAX_RTO_MIN
))
605 rto_min
= dst_metric_rtt(dst
, RTAX_RTO_MIN
);
609 /* Called to compute a smoothed rtt estimate. The data fed to this
610 * routine either comes from timestamps, or from segments that were
611 * known _not_ to have been retransmitted [see Karn/Partridge
612 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
613 * piece by Van Jacobson.
614 * NOTE: the next three routines used to be one big routine.
615 * To save cycles in the RFC 1323 implementation it was better to break
616 * it up into three procedures. -- erics
618 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
620 struct tcp_sock
*tp
= tcp_sk(sk
);
621 long m
= mrtt
; /* RTT */
623 /* The following amusing code comes from Jacobson's
624 * article in SIGCOMM '88. Note that rtt and mdev
625 * are scaled versions of rtt and mean deviation.
626 * This is designed to be as fast as possible
627 * m stands for "measurement".
629 * On a 1990 paper the rto value is changed to:
630 * RTO = rtt + 4 * mdev
632 * Funny. This algorithm seems to be very broken.
633 * These formulae increase RTO, when it should be decreased, increase
634 * too slowly, when it should be increased quickly, decrease too quickly
635 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
636 * does not matter how to _calculate_ it. Seems, it was trap
637 * that VJ failed to avoid. 8)
642 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
643 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
645 m
= -m
; /* m is now abs(error) */
646 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
647 /* This is similar to one of Eifel findings.
648 * Eifel blocks mdev updates when rtt decreases.
649 * This solution is a bit different: we use finer gain
650 * for mdev in this case (alpha*beta).
651 * Like Eifel it also prevents growth of rto,
652 * but also it limits too fast rto decreases,
653 * happening in pure Eifel.
658 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
660 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
661 if (tp
->mdev
> tp
->mdev_max
) {
662 tp
->mdev_max
= tp
->mdev
;
663 if (tp
->mdev_max
> tp
->rttvar
)
664 tp
->rttvar
= tp
->mdev_max
;
666 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
667 if (tp
->mdev_max
< tp
->rttvar
)
668 tp
->rttvar
-= (tp
->rttvar
- tp
->mdev_max
) >> 2;
669 tp
->rtt_seq
= tp
->snd_nxt
;
670 tp
->mdev_max
= tcp_rto_min(sk
);
673 /* no previous measure. */
674 tp
->srtt
= m
<< 3; /* take the measured time to be rtt */
675 tp
->mdev
= m
<< 1; /* make sure rto = 3*rtt */
676 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
677 tp
->rtt_seq
= tp
->snd_nxt
;
681 /* Calculate rto without backoff. This is the second half of Van Jacobson's
682 * routine referred to above.
684 static inline void tcp_set_rto(struct sock
*sk
)
686 const struct tcp_sock
*tp
= tcp_sk(sk
);
687 /* Old crap is replaced with new one. 8)
690 * 1. If rtt variance happened to be less 50msec, it is hallucination.
691 * It cannot be less due to utterly erratic ACK generation made
692 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
693 * to do with delayed acks, because at cwnd>2 true delack timeout
694 * is invisible. Actually, Linux-2.4 also generates erratic
695 * ACKs in some circumstances.
697 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
699 /* 2. Fixups made earlier cannot be right.
700 * If we do not estimate RTO correctly without them,
701 * all the algo is pure shit and should be replaced
702 * with correct one. It is exactly, which we pretend to do.
705 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
706 * guarantees that rto is higher.
708 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
709 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
712 /* Save metrics learned by this TCP session.
713 This function is called only, when TCP finishes successfully
714 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
716 void tcp_update_metrics(struct sock
*sk
)
718 struct tcp_sock
*tp
= tcp_sk(sk
);
719 struct dst_entry
*dst
= __sk_dst_get(sk
);
721 if (sysctl_tcp_nometrics_save
)
726 if (dst
&& (dst
->flags
& DST_HOST
)) {
727 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
731 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
732 /* This session failed to estimate rtt. Why?
733 * Probably, no packets returned in time.
736 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
737 dst
->metrics
[RTAX_RTT
- 1] = 0;
741 rtt
= dst_metric_rtt(dst
, RTAX_RTT
);
744 /* If newly calculated rtt larger than stored one,
745 * store new one. Otherwise, use EWMA. Remember,
746 * rtt overestimation is always better than underestimation.
748 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
750 set_dst_metric_rtt(dst
, RTAX_RTT
, tp
->srtt
);
752 set_dst_metric_rtt(dst
, RTAX_RTT
, rtt
- (m
>> 3));
755 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
760 /* Scale deviation to rttvar fixed point */
765 var
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
769 var
-= (var
- m
) >> 2;
771 set_dst_metric_rtt(dst
, RTAX_RTTVAR
, var
);
774 if (tp
->snd_ssthresh
>= 0xFFFF) {
775 /* Slow start still did not finish. */
776 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
777 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
778 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
779 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
780 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
781 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
782 dst
->metrics
[RTAX_CWND
- 1] = tp
->snd_cwnd
;
783 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
784 icsk
->icsk_ca_state
== TCP_CA_Open
) {
785 /* Cong. avoidance phase, cwnd is reliable. */
786 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
787 dst
->metrics
[RTAX_SSTHRESH
-1] =
788 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
789 if (!dst_metric_locked(dst
, RTAX_CWND
))
790 dst
->metrics
[RTAX_CWND
-1] = (dst_metric(dst
, RTAX_CWND
) + tp
->snd_cwnd
) >> 1;
792 /* Else slow start did not finish, cwnd is non-sense,
793 ssthresh may be also invalid.
795 if (!dst_metric_locked(dst
, RTAX_CWND
))
796 dst
->metrics
[RTAX_CWND
-1] = (dst_metric(dst
, RTAX_CWND
) + tp
->snd_ssthresh
) >> 1;
797 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
798 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
799 tp
->snd_ssthresh
> dst_metric(dst
, RTAX_SSTHRESH
))
800 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
803 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
804 if (dst_metric(dst
, RTAX_REORDERING
) < tp
->reordering
&&
805 tp
->reordering
!= sysctl_tcp_reordering
)
806 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
811 /* Numbers are taken from RFC3390.
813 * John Heffner states:
815 * The RFC specifies a window of no more than 4380 bytes
816 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
817 * is a bit misleading because they use a clamp at 4380 bytes
818 * rather than use a multiplier in the relevant range.
820 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
822 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
825 if (tp
->mss_cache
> 1460)
828 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
830 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
833 /* Set slow start threshold and cwnd not falling to slow start */
834 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
836 struct tcp_sock
*tp
= tcp_sk(sk
);
837 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
839 tp
->prior_ssthresh
= 0;
841 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
844 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
845 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
846 tcp_packets_in_flight(tp
) + 1U);
847 tp
->snd_cwnd_cnt
= 0;
848 tp
->high_seq
= tp
->snd_nxt
;
849 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
850 TCP_ECN_queue_cwr(tp
);
852 tcp_set_ca_state(sk
, TCP_CA_CWR
);
857 * Packet counting of FACK is based on in-order assumptions, therefore TCP
858 * disables it when reordering is detected
860 static void tcp_disable_fack(struct tcp_sock
*tp
)
862 /* RFC3517 uses different metric in lost marker => reset on change */
864 tp
->lost_skb_hint
= NULL
;
865 tp
->rx_opt
.sack_ok
&= ~2;
868 /* Take a notice that peer is sending D-SACKs */
869 static void tcp_dsack_seen(struct tcp_sock
*tp
)
871 tp
->rx_opt
.sack_ok
|= 4;
874 /* Initialize metrics on socket. */
876 static void tcp_init_metrics(struct sock
*sk
)
878 struct tcp_sock
*tp
= tcp_sk(sk
);
879 struct dst_entry
*dst
= __sk_dst_get(sk
);
886 if (dst_metric_locked(dst
, RTAX_CWND
))
887 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
888 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
889 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
890 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
891 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
893 if (dst_metric(dst
, RTAX_REORDERING
) &&
894 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
895 tcp_disable_fack(tp
);
896 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
899 if (dst_metric(dst
, RTAX_RTT
) == 0)
902 if (!tp
->srtt
&& dst_metric_rtt(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
905 /* Initial rtt is determined from SYN,SYN-ACK.
906 * The segment is small and rtt may appear much
907 * less than real one. Use per-dst memory
908 * to make it more realistic.
910 * A bit of theory. RTT is time passed after "normal" sized packet
911 * is sent until it is ACKed. In normal circumstances sending small
912 * packets force peer to delay ACKs and calculation is correct too.
913 * The algorithm is adaptive and, provided we follow specs, it
914 * NEVER underestimate RTT. BUT! If peer tries to make some clever
915 * tricks sort of "quick acks" for time long enough to decrease RTT
916 * to low value, and then abruptly stops to do it and starts to delay
917 * ACKs, wait for troubles.
919 if (dst_metric_rtt(dst
, RTAX_RTT
) > tp
->srtt
) {
920 tp
->srtt
= dst_metric_rtt(dst
, RTAX_RTT
);
921 tp
->rtt_seq
= tp
->snd_nxt
;
923 if (dst_metric_rtt(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
924 tp
->mdev
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
925 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
928 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
930 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
931 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
935 /* Play conservative. If timestamps are not
936 * supported, TCP will fail to recalculate correct
937 * rtt, if initial rto is too small. FORGET ALL AND RESET!
939 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
941 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
942 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
946 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
949 struct tcp_sock
*tp
= tcp_sk(sk
);
950 if (metric
> tp
->reordering
) {
953 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
955 /* This exciting event is worth to be remembered. 8) */
957 mib_idx
= LINUX_MIB_TCPTSREORDER
;
958 else if (tcp_is_reno(tp
))
959 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
960 else if (tcp_is_fack(tp
))
961 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
963 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
965 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
966 #if FASTRETRANS_DEBUG > 1
967 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
968 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
972 tp
->undo_marker
? tp
->undo_retrans
: 0);
974 tcp_disable_fack(tp
);
978 /* This must be called before lost_out is incremented */
979 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
981 if ((tp
->retransmit_skb_hint
== NULL
) ||
982 before(TCP_SKB_CB(skb
)->seq
,
983 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
984 tp
->retransmit_skb_hint
= skb
;
987 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
988 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
991 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
993 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
994 tcp_verify_retransmit_hint(tp
, skb
);
996 tp
->lost_out
+= tcp_skb_pcount(skb
);
997 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1001 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
1002 struct sk_buff
*skb
)
1004 tcp_verify_retransmit_hint(tp
, skb
);
1006 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1007 tp
->lost_out
+= tcp_skb_pcount(skb
);
1008 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1012 /* This procedure tags the retransmission queue when SACKs arrive.
1014 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1015 * Packets in queue with these bits set are counted in variables
1016 * sacked_out, retrans_out and lost_out, correspondingly.
1018 * Valid combinations are:
1019 * Tag InFlight Description
1020 * 0 1 - orig segment is in flight.
1021 * S 0 - nothing flies, orig reached receiver.
1022 * L 0 - nothing flies, orig lost by net.
1023 * R 2 - both orig and retransmit are in flight.
1024 * L|R 1 - orig is lost, retransmit is in flight.
1025 * S|R 1 - orig reached receiver, retrans is still in flight.
1026 * (L|S|R is logically valid, it could occur when L|R is sacked,
1027 * but it is equivalent to plain S and code short-curcuits it to S.
1028 * L|S is logically invalid, it would mean -1 packet in flight 8))
1030 * These 6 states form finite state machine, controlled by the following events:
1031 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1032 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1033 * 3. Loss detection event of one of three flavors:
1034 * A. Scoreboard estimator decided the packet is lost.
1035 * A'. Reno "three dupacks" marks head of queue lost.
1036 * A''. Its FACK modfication, head until snd.fack is lost.
1037 * B. SACK arrives sacking data transmitted after never retransmitted
1038 * hole was sent out.
1039 * C. SACK arrives sacking SND.NXT at the moment, when the
1040 * segment was retransmitted.
1041 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1043 * It is pleasant to note, that state diagram turns out to be commutative,
1044 * so that we are allowed not to be bothered by order of our actions,
1045 * when multiple events arrive simultaneously. (see the function below).
1047 * Reordering detection.
1048 * --------------------
1049 * Reordering metric is maximal distance, which a packet can be displaced
1050 * in packet stream. With SACKs we can estimate it:
1052 * 1. SACK fills old hole and the corresponding segment was not
1053 * ever retransmitted -> reordering. Alas, we cannot use it
1054 * when segment was retransmitted.
1055 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1056 * for retransmitted and already SACKed segment -> reordering..
1057 * Both of these heuristics are not used in Loss state, when we cannot
1058 * account for retransmits accurately.
1060 * SACK block validation.
1061 * ----------------------
1063 * SACK block range validation checks that the received SACK block fits to
1064 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1065 * Note that SND.UNA is not included to the range though being valid because
1066 * it means that the receiver is rather inconsistent with itself reporting
1067 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1068 * perfectly valid, however, in light of RFC2018 which explicitly states
1069 * that "SACK block MUST reflect the newest segment. Even if the newest
1070 * segment is going to be discarded ...", not that it looks very clever
1071 * in case of head skb. Due to potentional receiver driven attacks, we
1072 * choose to avoid immediate execution of a walk in write queue due to
1073 * reneging and defer head skb's loss recovery to standard loss recovery
1074 * procedure that will eventually trigger (nothing forbids us doing this).
1076 * Implements also blockage to start_seq wrap-around. Problem lies in the
1077 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1078 * there's no guarantee that it will be before snd_nxt (n). The problem
1079 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1082 * <- outs wnd -> <- wrapzone ->
1083 * u e n u_w e_w s n_w
1085 * |<------------+------+----- TCP seqno space --------------+---------->|
1086 * ...-- <2^31 ->| |<--------...
1087 * ...---- >2^31 ------>| |<--------...
1089 * Current code wouldn't be vulnerable but it's better still to discard such
1090 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1091 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1092 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1093 * equal to the ideal case (infinite seqno space without wrap caused issues).
1095 * With D-SACK the lower bound is extended to cover sequence space below
1096 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1097 * again, D-SACK block must not to go across snd_una (for the same reason as
1098 * for the normal SACK blocks, explained above). But there all simplicity
1099 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1100 * fully below undo_marker they do not affect behavior in anyway and can
1101 * therefore be safely ignored. In rare cases (which are more or less
1102 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1103 * fragmentation and packet reordering past skb's retransmission. To consider
1104 * them correctly, the acceptable range must be extended even more though
1105 * the exact amount is rather hard to quantify. However, tp->max_window can
1106 * be used as an exaggerated estimate.
1108 static int tcp_is_sackblock_valid(struct tcp_sock
*tp
, int is_dsack
,
1109 u32 start_seq
, u32 end_seq
)
1111 /* Too far in future, or reversed (interpretation is ambiguous) */
1112 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1115 /* Nasty start_seq wrap-around check (see comments above) */
1116 if (!before(start_seq
, tp
->snd_nxt
))
1119 /* In outstanding window? ...This is valid exit for D-SACKs too.
1120 * start_seq == snd_una is non-sensical (see comments above)
1122 if (after(start_seq
, tp
->snd_una
))
1125 if (!is_dsack
|| !tp
->undo_marker
)
1128 /* ...Then it's D-SACK, and must reside below snd_una completely */
1129 if (!after(end_seq
, tp
->snd_una
))
1132 if (!before(start_seq
, tp
->undo_marker
))
1136 if (!after(end_seq
, tp
->undo_marker
))
1139 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1140 * start_seq < undo_marker and end_seq >= undo_marker.
1142 return !before(start_seq
, end_seq
- tp
->max_window
);
1145 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1146 * Event "C". Later note: FACK people cheated me again 8), we have to account
1147 * for reordering! Ugly, but should help.
1149 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1150 * less than what is now known to be received by the other end (derived from
1151 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1152 * retransmitted skbs to avoid some costly processing per ACKs.
1154 static void tcp_mark_lost_retrans(struct sock
*sk
)
1156 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1157 struct tcp_sock
*tp
= tcp_sk(sk
);
1158 struct sk_buff
*skb
;
1160 u32 new_low_seq
= tp
->snd_nxt
;
1161 u32 received_upto
= tcp_highest_sack_seq(tp
);
1163 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1164 !after(received_upto
, tp
->lost_retrans_low
) ||
1165 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1168 tcp_for_write_queue(skb
, sk
) {
1169 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1171 if (skb
== tcp_send_head(sk
))
1173 if (cnt
== tp
->retrans_out
)
1175 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1178 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1181 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1182 * constraint here (see above) but figuring out that at
1183 * least tp->reordering SACK blocks reside between ack_seq
1184 * and received_upto is not easy task to do cheaply with
1185 * the available datastructures.
1187 * Whether FACK should check here for tp->reordering segs
1188 * in-between one could argue for either way (it would be
1189 * rather simple to implement as we could count fack_count
1190 * during the walk and do tp->fackets_out - fack_count).
1192 if (after(received_upto
, ack_seq
)) {
1193 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1194 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1196 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1197 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1199 if (before(ack_seq
, new_low_seq
))
1200 new_low_seq
= ack_seq
;
1201 cnt
+= tcp_skb_pcount(skb
);
1205 if (tp
->retrans_out
)
1206 tp
->lost_retrans_low
= new_low_seq
;
1209 static int tcp_check_dsack(struct sock
*sk
, struct sk_buff
*ack_skb
,
1210 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1213 struct tcp_sock
*tp
= tcp_sk(sk
);
1214 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1215 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1218 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1221 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1222 } else if (num_sacks
> 1) {
1223 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1224 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1226 if (!after(end_seq_0
, end_seq_1
) &&
1227 !before(start_seq_0
, start_seq_1
)) {
1230 NET_INC_STATS_BH(sock_net(sk
),
1231 LINUX_MIB_TCPDSACKOFORECV
);
1235 /* D-SACK for already forgotten data... Do dumb counting. */
1237 !after(end_seq_0
, prior_snd_una
) &&
1238 after(end_seq_0
, tp
->undo_marker
))
1244 struct tcp_sacktag_state
{
1250 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1251 * the incoming SACK may not exactly match but we can find smaller MSS
1252 * aligned portion of it that matches. Therefore we might need to fragment
1253 * which may fail and creates some hassle (caller must handle error case
1256 * FIXME: this could be merged to shift decision code
1258 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1259 u32 start_seq
, u32 end_seq
)
1262 unsigned int pkt_len
;
1265 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1266 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1268 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1269 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1270 mss
= tcp_skb_mss(skb
);
1271 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1274 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1278 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1283 /* Round if necessary so that SACKs cover only full MSSes
1284 * and/or the remaining small portion (if present)
1286 if (pkt_len
> mss
) {
1287 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1288 if (!in_sack
&& new_len
< pkt_len
) {
1290 if (new_len
> skb
->len
)
1295 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1303 static u8
tcp_sacktag_one(struct sk_buff
*skb
, struct sock
*sk
,
1304 struct tcp_sacktag_state
*state
,
1305 int dup_sack
, int pcount
)
1307 struct tcp_sock
*tp
= tcp_sk(sk
);
1308 u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1309 int fack_count
= state
->fack_count
;
1311 /* Account D-SACK for retransmitted packet. */
1312 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1313 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1315 if (sacked
& TCPCB_SACKED_ACKED
)
1316 state
->reord
= min(fack_count
, state
->reord
);
1319 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1320 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1323 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1324 if (sacked
& TCPCB_SACKED_RETRANS
) {
1325 /* If the segment is not tagged as lost,
1326 * we do not clear RETRANS, believing
1327 * that retransmission is still in flight.
1329 if (sacked
& TCPCB_LOST
) {
1330 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1331 tp
->lost_out
-= pcount
;
1332 tp
->retrans_out
-= pcount
;
1335 if (!(sacked
& TCPCB_RETRANS
)) {
1336 /* New sack for not retransmitted frame,
1337 * which was in hole. It is reordering.
1339 if (before(TCP_SKB_CB(skb
)->seq
,
1340 tcp_highest_sack_seq(tp
)))
1341 state
->reord
= min(fack_count
,
1344 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1345 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
))
1346 state
->flag
|= FLAG_ONLY_ORIG_SACKED
;
1349 if (sacked
& TCPCB_LOST
) {
1350 sacked
&= ~TCPCB_LOST
;
1351 tp
->lost_out
-= pcount
;
1355 sacked
|= TCPCB_SACKED_ACKED
;
1356 state
->flag
|= FLAG_DATA_SACKED
;
1357 tp
->sacked_out
+= pcount
;
1359 fack_count
+= pcount
;
1361 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1362 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1363 before(TCP_SKB_CB(skb
)->seq
,
1364 TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1365 tp
->lost_cnt_hint
+= pcount
;
1367 if (fack_count
> tp
->fackets_out
)
1368 tp
->fackets_out
= fack_count
;
1371 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1372 * frames and clear it. undo_retrans is decreased above, L|R frames
1373 * are accounted above as well.
1375 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1376 sacked
&= ~TCPCB_SACKED_RETRANS
;
1377 tp
->retrans_out
-= pcount
;
1383 static int tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1384 struct tcp_sacktag_state
*state
,
1385 unsigned int pcount
, int shifted
, int mss
,
1388 struct tcp_sock
*tp
= tcp_sk(sk
);
1389 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1393 /* Tweak before seqno plays */
1394 if (!tcp_is_fack(tp
) && tcp_is_sack(tp
) && tp
->lost_skb_hint
&&
1395 !before(TCP_SKB_CB(tp
->lost_skb_hint
)->seq
, TCP_SKB_CB(skb
)->seq
))
1396 tp
->lost_cnt_hint
+= pcount
;
1398 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1399 TCP_SKB_CB(skb
)->seq
+= shifted
;
1401 skb_shinfo(prev
)->gso_segs
+= pcount
;
1402 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1403 skb_shinfo(skb
)->gso_segs
-= pcount
;
1405 /* When we're adding to gso_segs == 1, gso_size will be zero,
1406 * in theory this shouldn't be necessary but as long as DSACK
1407 * code can come after this skb later on it's better to keep
1408 * setting gso_size to something.
1410 if (!skb_shinfo(prev
)->gso_size
) {
1411 skb_shinfo(prev
)->gso_size
= mss
;
1412 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1415 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1416 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1417 skb_shinfo(skb
)->gso_size
= 0;
1418 skb_shinfo(skb
)->gso_type
= 0;
1421 /* We discard results */
1422 tcp_sacktag_one(skb
, sk
, state
, dup_sack
, pcount
);
1424 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1425 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1428 BUG_ON(!tcp_skb_pcount(skb
));
1429 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1433 /* Whole SKB was eaten :-) */
1435 if (skb
== tp
->retransmit_skb_hint
)
1436 tp
->retransmit_skb_hint
= prev
;
1437 if (skb
== tp
->scoreboard_skb_hint
)
1438 tp
->scoreboard_skb_hint
= prev
;
1439 if (skb
== tp
->lost_skb_hint
) {
1440 tp
->lost_skb_hint
= prev
;
1441 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1444 TCP_SKB_CB(skb
)->flags
|= TCP_SKB_CB(prev
)->flags
;
1445 if (skb
== tcp_highest_sack(sk
))
1446 tcp_advance_highest_sack(sk
, skb
);
1448 tcp_unlink_write_queue(skb
, sk
);
1449 sk_wmem_free_skb(sk
, skb
);
1451 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1456 /* I wish gso_size would have a bit more sane initialization than
1457 * something-or-zero which complicates things
1459 static int tcp_skb_seglen(struct sk_buff
*skb
)
1461 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1464 /* Shifting pages past head area doesn't work */
1465 static int skb_can_shift(struct sk_buff
*skb
)
1467 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1470 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1473 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1474 struct tcp_sacktag_state
*state
,
1475 u32 start_seq
, u32 end_seq
,
1478 struct tcp_sock
*tp
= tcp_sk(sk
);
1479 struct sk_buff
*prev
;
1485 if (!sk_can_gso(sk
))
1488 /* Normally R but no L won't result in plain S */
1490 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1492 if (!skb_can_shift(skb
))
1494 /* This frame is about to be dropped (was ACKed). */
1495 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1498 /* Can only happen with delayed DSACK + discard craziness */
1499 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1501 prev
= tcp_write_queue_prev(sk
, skb
);
1503 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1506 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1507 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1511 pcount
= tcp_skb_pcount(skb
);
1512 mss
= tcp_skb_seglen(skb
);
1514 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1515 * drop this restriction as unnecessary
1517 if (mss
!= tcp_skb_seglen(prev
))
1520 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1522 /* CHECKME: This is non-MSS split case only?, this will
1523 * cause skipped skbs due to advancing loop btw, original
1524 * has that feature too
1526 if (tcp_skb_pcount(skb
) <= 1)
1529 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1531 /* TODO: head merge to next could be attempted here
1532 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1533 * though it might not be worth of the additional hassle
1535 * ...we can probably just fallback to what was done
1536 * previously. We could try merging non-SACKed ones
1537 * as well but it probably isn't going to buy off
1538 * because later SACKs might again split them, and
1539 * it would make skb timestamp tracking considerably
1545 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1547 BUG_ON(len
> skb
->len
);
1549 /* MSS boundaries should be honoured or else pcount will
1550 * severely break even though it makes things bit trickier.
1551 * Optimize common case to avoid most of the divides
1553 mss
= tcp_skb_mss(skb
);
1555 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1556 * drop this restriction as unnecessary
1558 if (mss
!= tcp_skb_seglen(prev
))
1563 } else if (len
< mss
) {
1571 if (!skb_shift(prev
, skb
, len
))
1573 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1576 /* Hole filled allows collapsing with the next as well, this is very
1577 * useful when hole on every nth skb pattern happens
1579 if (prev
== tcp_write_queue_tail(sk
))
1581 skb
= tcp_write_queue_next(sk
, prev
);
1583 if (!skb_can_shift(skb
) ||
1584 (skb
== tcp_send_head(sk
)) ||
1585 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1586 (mss
!= tcp_skb_seglen(skb
)))
1590 if (skb_shift(prev
, skb
, len
)) {
1591 pcount
+= tcp_skb_pcount(skb
);
1592 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1596 state
->fack_count
+= pcount
;
1603 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1607 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1608 struct tcp_sack_block
*next_dup
,
1609 struct tcp_sacktag_state
*state
,
1610 u32 start_seq
, u32 end_seq
,
1613 struct tcp_sock
*tp
= tcp_sk(sk
);
1614 struct sk_buff
*tmp
;
1616 tcp_for_write_queue_from(skb
, sk
) {
1618 int dup_sack
= dup_sack_in
;
1620 if (skb
== tcp_send_head(sk
))
1623 /* queue is in-order => we can short-circuit the walk early */
1624 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1627 if ((next_dup
!= NULL
) &&
1628 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1629 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1630 next_dup
->start_seq
,
1636 /* skb reference here is a bit tricky to get right, since
1637 * shifting can eat and free both this skb and the next,
1638 * so not even _safe variant of the loop is enough.
1641 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1642 start_seq
, end_seq
, dup_sack
);
1651 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1657 if (unlikely(in_sack
< 0))
1661 TCP_SKB_CB(skb
)->sacked
= tcp_sacktag_one(skb
, sk
,
1664 tcp_skb_pcount(skb
));
1666 if (!before(TCP_SKB_CB(skb
)->seq
,
1667 tcp_highest_sack_seq(tp
)))
1668 tcp_advance_highest_sack(sk
, skb
);
1671 state
->fack_count
+= tcp_skb_pcount(skb
);
1676 /* Avoid all extra work that is being done by sacktag while walking in
1679 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1680 struct tcp_sacktag_state
*state
,
1683 tcp_for_write_queue_from(skb
, sk
) {
1684 if (skb
== tcp_send_head(sk
))
1687 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1690 state
->fack_count
+= tcp_skb_pcount(skb
);
1695 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1697 struct tcp_sack_block
*next_dup
,
1698 struct tcp_sacktag_state
*state
,
1701 if (next_dup
== NULL
)
1704 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1705 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1706 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1707 next_dup
->start_seq
, next_dup
->end_seq
,
1714 static int tcp_sack_cache_ok(struct tcp_sock
*tp
, struct tcp_sack_block
*cache
)
1716 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1720 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
,
1723 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1724 struct tcp_sock
*tp
= tcp_sk(sk
);
1725 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1726 TCP_SKB_CB(ack_skb
)->sacked
);
1727 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1728 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1729 struct tcp_sack_block
*cache
;
1730 struct tcp_sacktag_state state
;
1731 struct sk_buff
*skb
;
1732 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1734 int found_dup_sack
= 0;
1736 int first_sack_index
;
1739 state
.reord
= tp
->packets_out
;
1741 if (!tp
->sacked_out
) {
1742 if (WARN_ON(tp
->fackets_out
))
1743 tp
->fackets_out
= 0;
1744 tcp_highest_sack_reset(sk
);
1747 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1748 num_sacks
, prior_snd_una
);
1750 state
.flag
|= FLAG_DSACKING_ACK
;
1752 /* Eliminate too old ACKs, but take into
1753 * account more or less fresh ones, they can
1754 * contain valid SACK info.
1756 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1759 if (!tp
->packets_out
)
1763 first_sack_index
= 0;
1764 for (i
= 0; i
< num_sacks
; i
++) {
1765 int dup_sack
= !i
&& found_dup_sack
;
1767 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1768 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1770 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1771 sp
[used_sacks
].start_seq
,
1772 sp
[used_sacks
].end_seq
)) {
1776 if (!tp
->undo_marker
)
1777 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1779 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1781 /* Don't count olds caused by ACK reordering */
1782 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1783 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1785 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1788 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1790 first_sack_index
= -1;
1794 /* Ignore very old stuff early */
1795 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1801 /* order SACK blocks to allow in order walk of the retrans queue */
1802 for (i
= used_sacks
- 1; i
> 0; i
--) {
1803 for (j
= 0; j
< i
; j
++) {
1804 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1805 struct tcp_sack_block tmp
;
1811 /* Track where the first SACK block goes to */
1812 if (j
== first_sack_index
)
1813 first_sack_index
= j
+ 1;
1818 skb
= tcp_write_queue_head(sk
);
1819 state
.fack_count
= 0;
1822 if (!tp
->sacked_out
) {
1823 /* It's already past, so skip checking against it */
1824 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1826 cache
= tp
->recv_sack_cache
;
1827 /* Skip empty blocks in at head of the cache */
1828 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1833 while (i
< used_sacks
) {
1834 u32 start_seq
= sp
[i
].start_seq
;
1835 u32 end_seq
= sp
[i
].end_seq
;
1836 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1837 struct tcp_sack_block
*next_dup
= NULL
;
1839 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1840 next_dup
= &sp
[i
+ 1];
1842 /* Event "B" in the comment above. */
1843 if (after(end_seq
, tp
->high_seq
))
1844 state
.flag
|= FLAG_DATA_LOST
;
1846 /* Skip too early cached blocks */
1847 while (tcp_sack_cache_ok(tp
, cache
) &&
1848 !before(start_seq
, cache
->end_seq
))
1851 /* Can skip some work by looking recv_sack_cache? */
1852 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1853 after(end_seq
, cache
->start_seq
)) {
1856 if (before(start_seq
, cache
->start_seq
)) {
1857 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1859 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1866 /* Rest of the block already fully processed? */
1867 if (!after(end_seq
, cache
->end_seq
))
1870 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1874 /* ...tail remains todo... */
1875 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1876 /* ...but better entrypoint exists! */
1877 skb
= tcp_highest_sack(sk
);
1880 state
.fack_count
= tp
->fackets_out
;
1885 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1886 /* Check overlap against next cached too (past this one already) */
1891 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1892 skb
= tcp_highest_sack(sk
);
1895 state
.fack_count
= tp
->fackets_out
;
1897 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1900 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1901 start_seq
, end_seq
, dup_sack
);
1904 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1905 * due to in-order walk
1907 if (after(end_seq
, tp
->frto_highmark
))
1908 state
.flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1913 /* Clear the head of the cache sack blocks so we can skip it next time */
1914 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1915 tp
->recv_sack_cache
[i
].start_seq
= 0;
1916 tp
->recv_sack_cache
[i
].end_seq
= 0;
1918 for (j
= 0; j
< used_sacks
; j
++)
1919 tp
->recv_sack_cache
[i
++] = sp
[j
];
1921 tcp_mark_lost_retrans(sk
);
1923 tcp_verify_left_out(tp
);
1925 if ((state
.reord
< tp
->fackets_out
) &&
1926 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1927 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1928 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1932 #if FASTRETRANS_DEBUG > 0
1933 WARN_ON((int)tp
->sacked_out
< 0);
1934 WARN_ON((int)tp
->lost_out
< 0);
1935 WARN_ON((int)tp
->retrans_out
< 0);
1936 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1941 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1942 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1944 static int tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1948 holes
= max(tp
->lost_out
, 1U);
1949 holes
= min(holes
, tp
->packets_out
);
1951 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1952 tp
->sacked_out
= tp
->packets_out
- holes
;
1958 /* If we receive more dupacks than we expected counting segments
1959 * in assumption of absent reordering, interpret this as reordering.
1960 * The only another reason could be bug in receiver TCP.
1962 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1964 struct tcp_sock
*tp
= tcp_sk(sk
);
1965 if (tcp_limit_reno_sacked(tp
))
1966 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1969 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1971 static void tcp_add_reno_sack(struct sock
*sk
)
1973 struct tcp_sock
*tp
= tcp_sk(sk
);
1975 tcp_check_reno_reordering(sk
, 0);
1976 tcp_verify_left_out(tp
);
1979 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1981 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1983 struct tcp_sock
*tp
= tcp_sk(sk
);
1986 /* One ACK acked hole. The rest eat duplicate ACKs. */
1987 if (acked
- 1 >= tp
->sacked_out
)
1990 tp
->sacked_out
-= acked
- 1;
1992 tcp_check_reno_reordering(sk
, acked
);
1993 tcp_verify_left_out(tp
);
1996 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
2001 static int tcp_is_sackfrto(const struct tcp_sock
*tp
)
2003 return (sysctl_tcp_frto
== 0x2) && !tcp_is_reno(tp
);
2006 /* F-RTO can only be used if TCP has never retransmitted anything other than
2007 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2009 int tcp_use_frto(struct sock
*sk
)
2011 const struct tcp_sock
*tp
= tcp_sk(sk
);
2012 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2013 struct sk_buff
*skb
;
2015 if (!sysctl_tcp_frto
)
2018 /* MTU probe and F-RTO won't really play nicely along currently */
2019 if (icsk
->icsk_mtup
.probe_size
)
2022 if (tcp_is_sackfrto(tp
))
2025 /* Avoid expensive walking of rexmit queue if possible */
2026 if (tp
->retrans_out
> 1)
2029 skb
= tcp_write_queue_head(sk
);
2030 if (tcp_skb_is_last(sk
, skb
))
2032 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
2033 tcp_for_write_queue_from(skb
, sk
) {
2034 if (skb
== tcp_send_head(sk
))
2036 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2038 /* Short-circuit when first non-SACKed skb has been checked */
2039 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2045 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2046 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2047 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2048 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2049 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2050 * bits are handled if the Loss state is really to be entered (in
2051 * tcp_enter_frto_loss).
2053 * Do like tcp_enter_loss() would; when RTO expires the second time it
2055 * "Reduce ssthresh if it has not yet been made inside this window."
2057 void tcp_enter_frto(struct sock
*sk
)
2059 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2060 struct tcp_sock
*tp
= tcp_sk(sk
);
2061 struct sk_buff
*skb
;
2063 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
2064 tp
->snd_una
== tp
->high_seq
||
2065 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
2066 !icsk
->icsk_retransmits
)) {
2067 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2068 /* Our state is too optimistic in ssthresh() call because cwnd
2069 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2070 * recovery has not yet completed. Pattern would be this: RTO,
2071 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2073 * RFC4138 should be more specific on what to do, even though
2074 * RTO is quite unlikely to occur after the first Cumulative ACK
2075 * due to back-off and complexity of triggering events ...
2077 if (tp
->frto_counter
) {
2079 stored_cwnd
= tp
->snd_cwnd
;
2081 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2082 tp
->snd_cwnd
= stored_cwnd
;
2084 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2086 /* ... in theory, cong.control module could do "any tricks" in
2087 * ssthresh(), which means that ca_state, lost bits and lost_out
2088 * counter would have to be faked before the call occurs. We
2089 * consider that too expensive, unlikely and hacky, so modules
2090 * using these in ssthresh() must deal these incompatibility
2091 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2093 tcp_ca_event(sk
, CA_EVENT_FRTO
);
2096 tp
->undo_marker
= tp
->snd_una
;
2097 tp
->undo_retrans
= 0;
2099 skb
= tcp_write_queue_head(sk
);
2100 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2101 tp
->undo_marker
= 0;
2102 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2103 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2104 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2106 tcp_verify_left_out(tp
);
2108 /* Too bad if TCP was application limited */
2109 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2111 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2112 * The last condition is necessary at least in tp->frto_counter case.
2114 if (tcp_is_sackfrto(tp
) && (tp
->frto_counter
||
2115 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
2116 after(tp
->high_seq
, tp
->snd_una
)) {
2117 tp
->frto_highmark
= tp
->high_seq
;
2119 tp
->frto_highmark
= tp
->snd_nxt
;
2121 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
2122 tp
->high_seq
= tp
->snd_nxt
;
2123 tp
->frto_counter
= 1;
2126 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2127 * which indicates that we should follow the traditional RTO recovery,
2128 * i.e. mark everything lost and do go-back-N retransmission.
2130 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
2132 struct tcp_sock
*tp
= tcp_sk(sk
);
2133 struct sk_buff
*skb
;
2136 tp
->retrans_out
= 0;
2137 if (tcp_is_reno(tp
))
2138 tcp_reset_reno_sack(tp
);
2140 tcp_for_write_queue(skb
, sk
) {
2141 if (skb
== tcp_send_head(sk
))
2144 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2146 * Count the retransmission made on RTO correctly (only when
2147 * waiting for the first ACK and did not get it)...
2149 if ((tp
->frto_counter
== 1) && !(flag
& FLAG_DATA_ACKED
)) {
2150 /* For some reason this R-bit might get cleared? */
2151 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
2152 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2153 /* ...enter this if branch just for the first segment */
2154 flag
|= FLAG_DATA_ACKED
;
2156 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2157 tp
->undo_marker
= 0;
2158 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2161 /* Marking forward transmissions that were made after RTO lost
2162 * can cause unnecessary retransmissions in some scenarios,
2163 * SACK blocks will mitigate that in some but not in all cases.
2164 * We used to not mark them but it was causing break-ups with
2165 * receivers that do only in-order receival.
2167 * TODO: we could detect presence of such receiver and select
2168 * different behavior per flow.
2170 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2171 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2172 tp
->lost_out
+= tcp_skb_pcount(skb
);
2173 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2176 tcp_verify_left_out(tp
);
2178 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
2179 tp
->snd_cwnd_cnt
= 0;
2180 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2181 tp
->frto_counter
= 0;
2182 tp
->bytes_acked
= 0;
2184 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2185 sysctl_tcp_reordering
);
2186 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2187 tp
->high_seq
= tp
->snd_nxt
;
2188 TCP_ECN_queue_cwr(tp
);
2190 tcp_clear_all_retrans_hints(tp
);
2193 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
2195 tp
->retrans_out
= 0;
2198 tp
->undo_marker
= 0;
2199 tp
->undo_retrans
= 0;
2202 void tcp_clear_retrans(struct tcp_sock
*tp
)
2204 tcp_clear_retrans_partial(tp
);
2206 tp
->fackets_out
= 0;
2210 /* Enter Loss state. If "how" is not zero, forget all SACK information
2211 * and reset tags completely, otherwise preserve SACKs. If receiver
2212 * dropped its ofo queue, we will know this due to reneging detection.
2214 void tcp_enter_loss(struct sock
*sk
, int how
)
2216 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2217 struct tcp_sock
*tp
= tcp_sk(sk
);
2218 struct sk_buff
*skb
;
2220 /* Reduce ssthresh if it has not yet been made inside this window. */
2221 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
2222 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
2223 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2224 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2225 tcp_ca_event(sk
, CA_EVENT_LOSS
);
2228 tp
->snd_cwnd_cnt
= 0;
2229 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2231 tp
->bytes_acked
= 0;
2232 tcp_clear_retrans_partial(tp
);
2234 if (tcp_is_reno(tp
))
2235 tcp_reset_reno_sack(tp
);
2238 /* Push undo marker, if it was plain RTO and nothing
2239 * was retransmitted. */
2240 tp
->undo_marker
= tp
->snd_una
;
2243 tp
->fackets_out
= 0;
2245 tcp_clear_all_retrans_hints(tp
);
2247 tcp_for_write_queue(skb
, sk
) {
2248 if (skb
== tcp_send_head(sk
))
2251 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2252 tp
->undo_marker
= 0;
2253 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
2254 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
2255 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
2256 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2257 tp
->lost_out
+= tcp_skb_pcount(skb
);
2258 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2261 tcp_verify_left_out(tp
);
2263 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2264 sysctl_tcp_reordering
);
2265 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2266 tp
->high_seq
= tp
->snd_nxt
;
2267 TCP_ECN_queue_cwr(tp
);
2268 /* Abort F-RTO algorithm if one is in progress */
2269 tp
->frto_counter
= 0;
2272 /* If ACK arrived pointing to a remembered SACK, it means that our
2273 * remembered SACKs do not reflect real state of receiver i.e.
2274 * receiver _host_ is heavily congested (or buggy).
2276 * Do processing similar to RTO timeout.
2278 static int tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2280 if (flag
& FLAG_SACK_RENEGING
) {
2281 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2282 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
2284 tcp_enter_loss(sk
, 1);
2285 icsk
->icsk_retransmits
++;
2286 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
2287 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2288 icsk
->icsk_rto
, TCP_RTO_MAX
);
2294 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
2296 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2299 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2300 * counter when SACK is enabled (without SACK, sacked_out is used for
2303 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2304 * segments up to the highest received SACK block so far and holes in
2307 * With reordering, holes may still be in flight, so RFC3517 recovery
2308 * uses pure sacked_out (total number of SACKed segments) even though
2309 * it violates the RFC that uses duplicate ACKs, often these are equal
2310 * but when e.g. out-of-window ACKs or packet duplication occurs,
2311 * they differ. Since neither occurs due to loss, TCP should really
2314 static inline int tcp_dupack_heurestics(struct tcp_sock
*tp
)
2316 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2319 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
2321 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
2324 static inline int tcp_head_timedout(struct sock
*sk
)
2326 struct tcp_sock
*tp
= tcp_sk(sk
);
2328 return tp
->packets_out
&&
2329 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
2332 /* Linux NewReno/SACK/FACK/ECN state machine.
2333 * --------------------------------------
2335 * "Open" Normal state, no dubious events, fast path.
2336 * "Disorder" In all the respects it is "Open",
2337 * but requires a bit more attention. It is entered when
2338 * we see some SACKs or dupacks. It is split of "Open"
2339 * mainly to move some processing from fast path to slow one.
2340 * "CWR" CWND was reduced due to some Congestion Notification event.
2341 * It can be ECN, ICMP source quench, local device congestion.
2342 * "Recovery" CWND was reduced, we are fast-retransmitting.
2343 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2345 * tcp_fastretrans_alert() is entered:
2346 * - each incoming ACK, if state is not "Open"
2347 * - when arrived ACK is unusual, namely:
2352 * Counting packets in flight is pretty simple.
2354 * in_flight = packets_out - left_out + retrans_out
2356 * packets_out is SND.NXT-SND.UNA counted in packets.
2358 * retrans_out is number of retransmitted segments.
2360 * left_out is number of segments left network, but not ACKed yet.
2362 * left_out = sacked_out + lost_out
2364 * sacked_out: Packets, which arrived to receiver out of order
2365 * and hence not ACKed. With SACKs this number is simply
2366 * amount of SACKed data. Even without SACKs
2367 * it is easy to give pretty reliable estimate of this number,
2368 * counting duplicate ACKs.
2370 * lost_out: Packets lost by network. TCP has no explicit
2371 * "loss notification" feedback from network (for now).
2372 * It means that this number can be only _guessed_.
2373 * Actually, it is the heuristics to predict lossage that
2374 * distinguishes different algorithms.
2376 * F.e. after RTO, when all the queue is considered as lost,
2377 * lost_out = packets_out and in_flight = retrans_out.
2379 * Essentially, we have now two algorithms counting
2382 * FACK: It is the simplest heuristics. As soon as we decided
2383 * that something is lost, we decide that _all_ not SACKed
2384 * packets until the most forward SACK are lost. I.e.
2385 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2386 * It is absolutely correct estimate, if network does not reorder
2387 * packets. And it loses any connection to reality when reordering
2388 * takes place. We use FACK by default until reordering
2389 * is suspected on the path to this destination.
2391 * NewReno: when Recovery is entered, we assume that one segment
2392 * is lost (classic Reno). While we are in Recovery and
2393 * a partial ACK arrives, we assume that one more packet
2394 * is lost (NewReno). This heuristics are the same in NewReno
2397 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2398 * deflation etc. CWND is real congestion window, never inflated, changes
2399 * only according to classic VJ rules.
2401 * Really tricky (and requiring careful tuning) part of algorithm
2402 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2403 * The first determines the moment _when_ we should reduce CWND and,
2404 * hence, slow down forward transmission. In fact, it determines the moment
2405 * when we decide that hole is caused by loss, rather than by a reorder.
2407 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2408 * holes, caused by lost packets.
2410 * And the most logically complicated part of algorithm is undo
2411 * heuristics. We detect false retransmits due to both too early
2412 * fast retransmit (reordering) and underestimated RTO, analyzing
2413 * timestamps and D-SACKs. When we detect that some segments were
2414 * retransmitted by mistake and CWND reduction was wrong, we undo
2415 * window reduction and abort recovery phase. This logic is hidden
2416 * inside several functions named tcp_try_undo_<something>.
2419 /* This function decides, when we should leave Disordered state
2420 * and enter Recovery phase, reducing congestion window.
2422 * Main question: may we further continue forward transmission
2423 * with the same cwnd?
2425 static int tcp_time_to_recover(struct sock
*sk
)
2427 struct tcp_sock
*tp
= tcp_sk(sk
);
2430 /* Do not perform any recovery during F-RTO algorithm */
2431 if (tp
->frto_counter
)
2434 /* Trick#1: The loss is proven. */
2438 /* Not-A-Trick#2 : Classic rule... */
2439 if (tcp_dupack_heurestics(tp
) > tp
->reordering
)
2442 /* Trick#3 : when we use RFC2988 timer restart, fast
2443 * retransmit can be triggered by timeout of queue head.
2445 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2448 /* Trick#4: It is still not OK... But will it be useful to delay
2451 packets_out
= tp
->packets_out
;
2452 if (packets_out
<= tp
->reordering
&&
2453 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2454 !tcp_may_send_now(sk
)) {
2455 /* We have nothing to send. This connection is limited
2456 * either by receiver window or by application.
2464 /* New heuristics: it is possible only after we switched to restart timer
2465 * each time when something is ACKed. Hence, we can detect timed out packets
2466 * during fast retransmit without falling to slow start.
2468 * Usefulness of this as is very questionable, since we should know which of
2469 * the segments is the next to timeout which is relatively expensive to find
2470 * in general case unless we add some data structure just for that. The
2471 * current approach certainly won't find the right one too often and when it
2472 * finally does find _something_ it usually marks large part of the window
2473 * right away (because a retransmission with a larger timestamp blocks the
2474 * loop from advancing). -ij
2476 static void tcp_timeout_skbs(struct sock
*sk
)
2478 struct tcp_sock
*tp
= tcp_sk(sk
);
2479 struct sk_buff
*skb
;
2481 if (!tcp_is_fack(tp
) || !tcp_head_timedout(sk
))
2484 skb
= tp
->scoreboard_skb_hint
;
2485 if (tp
->scoreboard_skb_hint
== NULL
)
2486 skb
= tcp_write_queue_head(sk
);
2488 tcp_for_write_queue_from(skb
, sk
) {
2489 if (skb
== tcp_send_head(sk
))
2491 if (!tcp_skb_timedout(sk
, skb
))
2494 tcp_skb_mark_lost(tp
, skb
);
2497 tp
->scoreboard_skb_hint
= skb
;
2499 tcp_verify_left_out(tp
);
2502 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2503 * is against sacked "cnt", otherwise it's against facked "cnt"
2505 static void tcp_mark_head_lost(struct sock
*sk
, int packets
)
2507 struct tcp_sock
*tp
= tcp_sk(sk
);
2508 struct sk_buff
*skb
;
2513 WARN_ON(packets
> tp
->packets_out
);
2514 if (tp
->lost_skb_hint
) {
2515 skb
= tp
->lost_skb_hint
;
2516 cnt
= tp
->lost_cnt_hint
;
2518 skb
= tcp_write_queue_head(sk
);
2522 tcp_for_write_queue_from(skb
, sk
) {
2523 if (skb
== tcp_send_head(sk
))
2525 /* TODO: do this better */
2526 /* this is not the most efficient way to do this... */
2527 tp
->lost_skb_hint
= skb
;
2528 tp
->lost_cnt_hint
= cnt
;
2530 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
2534 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2535 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2536 cnt
+= tcp_skb_pcount(skb
);
2538 if (cnt
> packets
) {
2539 if (tcp_is_sack(tp
) || (oldcnt
>= packets
))
2542 mss
= skb_shinfo(skb
)->gso_size
;
2543 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2549 tcp_skb_mark_lost(tp
, skb
);
2551 tcp_verify_left_out(tp
);
2554 /* Account newly detected lost packet(s) */
2556 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2558 struct tcp_sock
*tp
= tcp_sk(sk
);
2560 if (tcp_is_reno(tp
)) {
2561 tcp_mark_head_lost(sk
, 1);
2562 } else if (tcp_is_fack(tp
)) {
2563 int lost
= tp
->fackets_out
- tp
->reordering
;
2566 tcp_mark_head_lost(sk
, lost
);
2568 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2569 if (sacked_upto
< fast_rexmit
)
2570 sacked_upto
= fast_rexmit
;
2571 tcp_mark_head_lost(sk
, sacked_upto
);
2574 tcp_timeout_skbs(sk
);
2577 /* CWND moderation, preventing bursts due to too big ACKs
2578 * in dubious situations.
2580 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2582 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2583 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2584 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2587 /* Lower bound on congestion window is slow start threshold
2588 * unless congestion avoidance choice decides to overide it.
2590 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2592 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2594 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2597 /* Decrease cwnd each second ack. */
2598 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2600 struct tcp_sock
*tp
= tcp_sk(sk
);
2601 int decr
= tp
->snd_cwnd_cnt
+ 1;
2603 if ((flag
& (FLAG_ANY_PROGRESS
| FLAG_DSACKING_ACK
)) ||
2604 (tcp_is_reno(tp
) && !(flag
& FLAG_NOT_DUP
))) {
2605 tp
->snd_cwnd_cnt
= decr
& 1;
2608 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2609 tp
->snd_cwnd
-= decr
;
2611 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2612 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2616 /* Nothing was retransmitted or returned timestamp is less
2617 * than timestamp of the first retransmission.
2619 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2621 return !tp
->retrans_stamp
||
2622 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2623 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2626 /* Undo procedures. */
2628 #if FASTRETRANS_DEBUG > 1
2629 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2631 struct tcp_sock
*tp
= tcp_sk(sk
);
2632 struct inet_sock
*inet
= inet_sk(sk
);
2634 if (sk
->sk_family
== AF_INET
) {
2635 printk(KERN_DEBUG
"Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2637 &inet
->daddr
, ntohs(inet
->dport
),
2638 tp
->snd_cwnd
, tcp_left_out(tp
),
2639 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2642 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2643 else if (sk
->sk_family
== AF_INET6
) {
2644 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2645 printk(KERN_DEBUG
"Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2647 &np
->daddr
, ntohs(inet
->dport
),
2648 tp
->snd_cwnd
, tcp_left_out(tp
),
2649 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2655 #define DBGUNDO(x...) do { } while (0)
2658 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2660 struct tcp_sock
*tp
= tcp_sk(sk
);
2662 if (tp
->prior_ssthresh
) {
2663 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2665 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2666 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2668 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2670 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2671 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2672 TCP_ECN_withdraw_cwr(tp
);
2675 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2677 tcp_moderate_cwnd(tp
);
2678 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2681 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2683 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2686 /* People celebrate: "We love our President!" */
2687 static int tcp_try_undo_recovery(struct sock
*sk
)
2689 struct tcp_sock
*tp
= tcp_sk(sk
);
2691 if (tcp_may_undo(tp
)) {
2694 /* Happy end! We did not retransmit anything
2695 * or our original transmission succeeded.
2697 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2698 tcp_undo_cwr(sk
, 1);
2699 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2700 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2702 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2704 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2705 tp
->undo_marker
= 0;
2707 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2708 /* Hold old state until something *above* high_seq
2709 * is ACKed. For Reno it is MUST to prevent false
2710 * fast retransmits (RFC2582). SACK TCP is safe. */
2711 tcp_moderate_cwnd(tp
);
2714 tcp_set_ca_state(sk
, TCP_CA_Open
);
2718 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2719 static void tcp_try_undo_dsack(struct sock
*sk
)
2721 struct tcp_sock
*tp
= tcp_sk(sk
);
2723 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2724 DBGUNDO(sk
, "D-SACK");
2725 tcp_undo_cwr(sk
, 1);
2726 tp
->undo_marker
= 0;
2727 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2731 /* Undo during fast recovery after partial ACK. */
2733 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2735 struct tcp_sock
*tp
= tcp_sk(sk
);
2736 /* Partial ACK arrived. Force Hoe's retransmit. */
2737 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2739 if (tcp_may_undo(tp
)) {
2740 /* Plain luck! Hole if filled with delayed
2741 * packet, rather than with a retransmit.
2743 if (tp
->retrans_out
== 0)
2744 tp
->retrans_stamp
= 0;
2746 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2749 tcp_undo_cwr(sk
, 0);
2750 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2752 /* So... Do not make Hoe's retransmit yet.
2753 * If the first packet was delayed, the rest
2754 * ones are most probably delayed as well.
2761 /* Undo during loss recovery after partial ACK. */
2762 static int tcp_try_undo_loss(struct sock
*sk
)
2764 struct tcp_sock
*tp
= tcp_sk(sk
);
2766 if (tcp_may_undo(tp
)) {
2767 struct sk_buff
*skb
;
2768 tcp_for_write_queue(skb
, sk
) {
2769 if (skb
== tcp_send_head(sk
))
2771 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2774 tcp_clear_all_retrans_hints(tp
);
2776 DBGUNDO(sk
, "partial loss");
2778 tcp_undo_cwr(sk
, 1);
2779 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2780 inet_csk(sk
)->icsk_retransmits
= 0;
2781 tp
->undo_marker
= 0;
2782 if (tcp_is_sack(tp
))
2783 tcp_set_ca_state(sk
, TCP_CA_Open
);
2789 static inline void tcp_complete_cwr(struct sock
*sk
)
2791 struct tcp_sock
*tp
= tcp_sk(sk
);
2792 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2793 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2794 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2797 static void tcp_try_keep_open(struct sock
*sk
)
2799 struct tcp_sock
*tp
= tcp_sk(sk
);
2800 int state
= TCP_CA_Open
;
2802 if (tcp_left_out(tp
) || tp
->retrans_out
|| tp
->undo_marker
)
2803 state
= TCP_CA_Disorder
;
2805 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2806 tcp_set_ca_state(sk
, state
);
2807 tp
->high_seq
= tp
->snd_nxt
;
2811 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2813 struct tcp_sock
*tp
= tcp_sk(sk
);
2815 tcp_verify_left_out(tp
);
2817 if (!tp
->frto_counter
&& tp
->retrans_out
== 0)
2818 tp
->retrans_stamp
= 0;
2820 if (flag
& FLAG_ECE
)
2821 tcp_enter_cwr(sk
, 1);
2823 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2824 tcp_try_keep_open(sk
);
2825 tcp_moderate_cwnd(tp
);
2827 tcp_cwnd_down(sk
, flag
);
2831 static void tcp_mtup_probe_failed(struct sock
*sk
)
2833 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2835 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2836 icsk
->icsk_mtup
.probe_size
= 0;
2839 static void tcp_mtup_probe_success(struct sock
*sk
)
2841 struct tcp_sock
*tp
= tcp_sk(sk
);
2842 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2844 /* FIXME: breaks with very large cwnd */
2845 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2846 tp
->snd_cwnd
= tp
->snd_cwnd
*
2847 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2848 icsk
->icsk_mtup
.probe_size
;
2849 tp
->snd_cwnd_cnt
= 0;
2850 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2851 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2853 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2854 icsk
->icsk_mtup
.probe_size
= 0;
2855 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2858 /* Do a simple retransmit without using the backoff mechanisms in
2859 * tcp_timer. This is used for path mtu discovery.
2860 * The socket is already locked here.
2862 void tcp_simple_retransmit(struct sock
*sk
)
2864 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2865 struct tcp_sock
*tp
= tcp_sk(sk
);
2866 struct sk_buff
*skb
;
2867 unsigned int mss
= tcp_current_mss(sk
, 0);
2868 u32 prior_lost
= tp
->lost_out
;
2870 tcp_for_write_queue(skb
, sk
) {
2871 if (skb
== tcp_send_head(sk
))
2873 if (tcp_skb_seglen(skb
) > mss
&&
2874 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2875 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2876 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2877 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2879 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2883 tcp_clear_retrans_hints_partial(tp
);
2885 if (prior_lost
== tp
->lost_out
)
2888 if (tcp_is_reno(tp
))
2889 tcp_limit_reno_sacked(tp
);
2891 tcp_verify_left_out(tp
);
2893 /* Don't muck with the congestion window here.
2894 * Reason is that we do not increase amount of _data_
2895 * in network, but units changed and effective
2896 * cwnd/ssthresh really reduced now.
2898 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2899 tp
->high_seq
= tp
->snd_nxt
;
2900 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2901 tp
->prior_ssthresh
= 0;
2902 tp
->undo_marker
= 0;
2903 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2905 tcp_xmit_retransmit_queue(sk
);
2908 /* Process an event, which can update packets-in-flight not trivially.
2909 * Main goal of this function is to calculate new estimate for left_out,
2910 * taking into account both packets sitting in receiver's buffer and
2911 * packets lost by network.
2913 * Besides that it does CWND reduction, when packet loss is detected
2914 * and changes state of machine.
2916 * It does _not_ decide what to send, it is made in function
2917 * tcp_xmit_retransmit_queue().
2919 static void tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2921 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2922 struct tcp_sock
*tp
= tcp_sk(sk
);
2923 int is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
2924 int do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2925 (tcp_fackets_out(tp
) > tp
->reordering
));
2926 int fast_rexmit
= 0, mib_idx
;
2928 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2930 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2931 tp
->fackets_out
= 0;
2933 /* Now state machine starts.
2934 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2935 if (flag
& FLAG_ECE
)
2936 tp
->prior_ssthresh
= 0;
2938 /* B. In all the states check for reneging SACKs. */
2939 if (tcp_check_sack_reneging(sk
, flag
))
2942 /* C. Process data loss notification, provided it is valid. */
2943 if (tcp_is_fack(tp
) && (flag
& FLAG_DATA_LOST
) &&
2944 before(tp
->snd_una
, tp
->high_seq
) &&
2945 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2946 tp
->fackets_out
> tp
->reordering
) {
2947 tcp_mark_head_lost(sk
, tp
->fackets_out
- tp
->reordering
);
2948 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSS
);
2951 /* D. Check consistency of the current state. */
2952 tcp_verify_left_out(tp
);
2954 /* E. Check state exit conditions. State can be terminated
2955 * when high_seq is ACKed. */
2956 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2957 WARN_ON(tp
->retrans_out
!= 0);
2958 tp
->retrans_stamp
= 0;
2959 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2960 switch (icsk
->icsk_ca_state
) {
2962 icsk
->icsk_retransmits
= 0;
2963 if (tcp_try_undo_recovery(sk
))
2968 /* CWR is to be held something *above* high_seq
2969 * is ACKed for CWR bit to reach receiver. */
2970 if (tp
->snd_una
!= tp
->high_seq
) {
2971 tcp_complete_cwr(sk
);
2972 tcp_set_ca_state(sk
, TCP_CA_Open
);
2976 case TCP_CA_Disorder
:
2977 tcp_try_undo_dsack(sk
);
2978 if (!tp
->undo_marker
||
2979 /* For SACK case do not Open to allow to undo
2980 * catching for all duplicate ACKs. */
2981 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2982 tp
->undo_marker
= 0;
2983 tcp_set_ca_state(sk
, TCP_CA_Open
);
2987 case TCP_CA_Recovery
:
2988 if (tcp_is_reno(tp
))
2989 tcp_reset_reno_sack(tp
);
2990 if (tcp_try_undo_recovery(sk
))
2992 tcp_complete_cwr(sk
);
2997 /* F. Process state. */
2998 switch (icsk
->icsk_ca_state
) {
2999 case TCP_CA_Recovery
:
3000 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
3001 if (tcp_is_reno(tp
) && is_dupack
)
3002 tcp_add_reno_sack(sk
);
3004 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
3007 if (flag
& FLAG_DATA_ACKED
)
3008 icsk
->icsk_retransmits
= 0;
3009 if (tcp_is_reno(tp
) && flag
& FLAG_SND_UNA_ADVANCED
)
3010 tcp_reset_reno_sack(tp
);
3011 if (!tcp_try_undo_loss(sk
)) {
3012 tcp_moderate_cwnd(tp
);
3013 tcp_xmit_retransmit_queue(sk
);
3016 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
3018 /* Loss is undone; fall through to processing in Open state. */
3020 if (tcp_is_reno(tp
)) {
3021 if (flag
& FLAG_SND_UNA_ADVANCED
)
3022 tcp_reset_reno_sack(tp
);
3024 tcp_add_reno_sack(sk
);
3027 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
3028 tcp_try_undo_dsack(sk
);
3030 if (!tcp_time_to_recover(sk
)) {
3031 tcp_try_to_open(sk
, flag
);
3035 /* MTU probe failure: don't reduce cwnd */
3036 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
3037 icsk
->icsk_mtup
.probe_size
&&
3038 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
3039 tcp_mtup_probe_failed(sk
);
3040 /* Restores the reduction we did in tcp_mtup_probe() */
3042 tcp_simple_retransmit(sk
);
3046 /* Otherwise enter Recovery state */
3048 if (tcp_is_reno(tp
))
3049 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
3051 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
3053 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3055 tp
->high_seq
= tp
->snd_nxt
;
3056 tp
->prior_ssthresh
= 0;
3057 tp
->undo_marker
= tp
->snd_una
;
3058 tp
->undo_retrans
= tp
->retrans_out
;
3060 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
3061 if (!(flag
& FLAG_ECE
))
3062 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
3063 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
3064 TCP_ECN_queue_cwr(tp
);
3067 tp
->bytes_acked
= 0;
3068 tp
->snd_cwnd_cnt
= 0;
3069 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
3073 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
3074 tcp_update_scoreboard(sk
, fast_rexmit
);
3075 tcp_cwnd_down(sk
, flag
);
3076 tcp_xmit_retransmit_queue(sk
);
3079 static void tcp_valid_rtt_meas(struct sock
*sk
, u32 seq_rtt
)
3081 tcp_rtt_estimator(sk
, seq_rtt
);
3083 inet_csk(sk
)->icsk_backoff
= 0;
3086 /* Read draft-ietf-tcplw-high-performance before mucking
3087 * with this code. (Supersedes RFC1323)
3089 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
3091 /* RTTM Rule: A TSecr value received in a segment is used to
3092 * update the averaged RTT measurement only if the segment
3093 * acknowledges some new data, i.e., only if it advances the
3094 * left edge of the send window.
3096 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3097 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3099 * Changed: reset backoff as soon as we see the first valid sample.
3100 * If we do not, we get strongly overestimated rto. With timestamps
3101 * samples are accepted even from very old segments: f.e., when rtt=1
3102 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3103 * answer arrives rto becomes 120 seconds! If at least one of segments
3104 * in window is lost... Voila. --ANK (010210)
3106 struct tcp_sock
*tp
= tcp_sk(sk
);
3108 tcp_valid_rtt_meas(sk
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
3111 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
3113 /* We don't have a timestamp. Can only use
3114 * packets that are not retransmitted to determine
3115 * rtt estimates. Also, we must not reset the
3116 * backoff for rto until we get a non-retransmitted
3117 * packet. This allows us to deal with a situation
3118 * where the network delay has increased suddenly.
3119 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3122 if (flag
& FLAG_RETRANS_DATA_ACKED
)
3125 tcp_valid_rtt_meas(sk
, seq_rtt
);
3128 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
3131 const struct tcp_sock
*tp
= tcp_sk(sk
);
3132 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3133 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3134 tcp_ack_saw_tstamp(sk
, flag
);
3135 else if (seq_rtt
>= 0)
3136 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
3139 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
3141 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3142 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
3143 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
3146 /* Restart timer after forward progress on connection.
3147 * RFC2988 recommends to restart timer to now+rto.
3149 static void tcp_rearm_rto(struct sock
*sk
)
3151 struct tcp_sock
*tp
= tcp_sk(sk
);
3153 if (!tp
->packets_out
) {
3154 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3156 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3157 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3161 /* If we get here, the whole TSO packet has not been acked. */
3162 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3164 struct tcp_sock
*tp
= tcp_sk(sk
);
3167 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3169 packets_acked
= tcp_skb_pcount(skb
);
3170 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3172 packets_acked
-= tcp_skb_pcount(skb
);
3174 if (packets_acked
) {
3175 BUG_ON(tcp_skb_pcount(skb
) == 0);
3176 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3179 return packets_acked
;
3182 /* Remove acknowledged frames from the retransmission queue. If our packet
3183 * is before the ack sequence we can discard it as it's confirmed to have
3184 * arrived at the other end.
3186 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3189 struct tcp_sock
*tp
= tcp_sk(sk
);
3190 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3191 struct sk_buff
*skb
;
3192 u32 now
= tcp_time_stamp
;
3193 int fully_acked
= 1;
3196 u32 reord
= tp
->packets_out
;
3197 u32 prior_sacked
= tp
->sacked_out
;
3199 s32 ca_seq_rtt
= -1;
3200 ktime_t last_ackt
= net_invalid_timestamp();
3202 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3203 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3205 u8 sacked
= scb
->sacked
;
3207 /* Determine how many packets and what bytes were acked, tso and else */
3208 if (after(scb
->end_seq
, tp
->snd_una
)) {
3209 if (tcp_skb_pcount(skb
) == 1 ||
3210 !after(tp
->snd_una
, scb
->seq
))
3213 acked_pcount
= tcp_tso_acked(sk
, skb
);
3219 acked_pcount
= tcp_skb_pcount(skb
);
3222 if (sacked
& TCPCB_RETRANS
) {
3223 if (sacked
& TCPCB_SACKED_RETRANS
)
3224 tp
->retrans_out
-= acked_pcount
;
3225 flag
|= FLAG_RETRANS_DATA_ACKED
;
3228 if ((flag
& FLAG_DATA_ACKED
) || (acked_pcount
> 1))
3229 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
3231 ca_seq_rtt
= now
- scb
->when
;
3232 last_ackt
= skb
->tstamp
;
3234 seq_rtt
= ca_seq_rtt
;
3236 if (!(sacked
& TCPCB_SACKED_ACKED
))
3237 reord
= min(pkts_acked
, reord
);
3240 if (sacked
& TCPCB_SACKED_ACKED
)
3241 tp
->sacked_out
-= acked_pcount
;
3242 if (sacked
& TCPCB_LOST
)
3243 tp
->lost_out
-= acked_pcount
;
3245 tp
->packets_out
-= acked_pcount
;
3246 pkts_acked
+= acked_pcount
;
3248 /* Initial outgoing SYN's get put onto the write_queue
3249 * just like anything else we transmit. It is not
3250 * true data, and if we misinform our callers that
3251 * this ACK acks real data, we will erroneously exit
3252 * connection startup slow start one packet too
3253 * quickly. This is severely frowned upon behavior.
3255 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
3256 flag
|= FLAG_DATA_ACKED
;
3258 flag
|= FLAG_SYN_ACKED
;
3259 tp
->retrans_stamp
= 0;
3265 tcp_unlink_write_queue(skb
, sk
);
3266 sk_wmem_free_skb(sk
, skb
);
3267 tp
->scoreboard_skb_hint
= NULL
;
3268 if (skb
== tp
->retransmit_skb_hint
)
3269 tp
->retransmit_skb_hint
= NULL
;
3270 if (skb
== tp
->lost_skb_hint
)
3271 tp
->lost_skb_hint
= NULL
;
3274 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3275 tp
->snd_up
= tp
->snd_una
;
3277 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3278 flag
|= FLAG_SACK_RENEGING
;
3280 if (flag
& FLAG_ACKED
) {
3281 const struct tcp_congestion_ops
*ca_ops
3282 = inet_csk(sk
)->icsk_ca_ops
;
3284 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3285 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3286 tcp_mtup_probe_success(sk
);
3289 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
3292 if (tcp_is_reno(tp
)) {
3293 tcp_remove_reno_sacks(sk
, pkts_acked
);
3297 /* Non-retransmitted hole got filled? That's reordering */
3298 if (reord
< prior_fackets
)
3299 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3301 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3302 prior_sacked
- tp
->sacked_out
;
3303 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3306 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3308 if (ca_ops
->pkts_acked
) {
3311 /* Is the ACK triggering packet unambiguous? */
3312 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3313 /* High resolution needed and available? */
3314 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3315 !ktime_equal(last_ackt
,
3316 net_invalid_timestamp()))
3317 rtt_us
= ktime_us_delta(ktime_get_real(),
3319 else if (ca_seq_rtt
> 0)
3320 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3323 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3327 #if FASTRETRANS_DEBUG > 0
3328 WARN_ON((int)tp
->sacked_out
< 0);
3329 WARN_ON((int)tp
->lost_out
< 0);
3330 WARN_ON((int)tp
->retrans_out
< 0);
3331 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3332 icsk
= inet_csk(sk
);
3334 printk(KERN_DEBUG
"Leak l=%u %d\n",
3335 tp
->lost_out
, icsk
->icsk_ca_state
);
3338 if (tp
->sacked_out
) {
3339 printk(KERN_DEBUG
"Leak s=%u %d\n",
3340 tp
->sacked_out
, icsk
->icsk_ca_state
);
3343 if (tp
->retrans_out
) {
3344 printk(KERN_DEBUG
"Leak r=%u %d\n",
3345 tp
->retrans_out
, icsk
->icsk_ca_state
);
3346 tp
->retrans_out
= 0;
3353 static void tcp_ack_probe(struct sock
*sk
)
3355 const struct tcp_sock
*tp
= tcp_sk(sk
);
3356 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3358 /* Was it a usable window open? */
3360 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3361 icsk
->icsk_backoff
= 0;
3362 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3363 /* Socket must be waked up by subsequent tcp_data_snd_check().
3364 * This function is not for random using!
3367 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3368 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3373 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3375 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3376 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
3379 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3381 const struct tcp_sock
*tp
= tcp_sk(sk
);
3382 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3383 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
3386 /* Check that window update is acceptable.
3387 * The function assumes that snd_una<=ack<=snd_next.
3389 static inline int tcp_may_update_window(const struct tcp_sock
*tp
,
3390 const u32 ack
, const u32 ack_seq
,
3393 return (after(ack
, tp
->snd_una
) ||
3394 after(ack_seq
, tp
->snd_wl1
) ||
3395 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
3398 /* Update our send window.
3400 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3401 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3403 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
3406 struct tcp_sock
*tp
= tcp_sk(sk
);
3408 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3410 if (likely(!tcp_hdr(skb
)->syn
))
3411 nwin
<<= tp
->rx_opt
.snd_wscale
;
3413 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3414 flag
|= FLAG_WIN_UPDATE
;
3415 tcp_update_wl(tp
, ack_seq
);
3417 if (tp
->snd_wnd
!= nwin
) {
3420 /* Note, it is the only place, where
3421 * fast path is recovered for sending TCP.
3424 tcp_fast_path_check(sk
);
3426 if (nwin
> tp
->max_window
) {
3427 tp
->max_window
= nwin
;
3428 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3438 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3439 * continue in congestion avoidance.
3441 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
3443 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
3444 tp
->snd_cwnd_cnt
= 0;
3445 tp
->bytes_acked
= 0;
3446 TCP_ECN_queue_cwr(tp
);
3447 tcp_moderate_cwnd(tp
);
3450 /* A conservative spurious RTO response algorithm: reduce cwnd using
3451 * rate halving and continue in congestion avoidance.
3453 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
3455 tcp_enter_cwr(sk
, 0);
3458 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
3460 if (flag
& FLAG_ECE
)
3461 tcp_ratehalving_spur_to_response(sk
);
3463 tcp_undo_cwr(sk
, 1);
3466 /* F-RTO spurious RTO detection algorithm (RFC4138)
3468 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3469 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3470 * window (but not to or beyond highest sequence sent before RTO):
3471 * On First ACK, send two new segments out.
3472 * On Second ACK, RTO was likely spurious. Do spurious response (response
3473 * algorithm is not part of the F-RTO detection algorithm
3474 * given in RFC4138 but can be selected separately).
3475 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3476 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3477 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3478 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3480 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3481 * original window even after we transmit two new data segments.
3484 * on first step, wait until first cumulative ACK arrives, then move to
3485 * the second step. In second step, the next ACK decides.
3487 * F-RTO is implemented (mainly) in four functions:
3488 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3489 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3490 * called when tcp_use_frto() showed green light
3491 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3492 * - tcp_enter_frto_loss() is called if there is not enough evidence
3493 * to prove that the RTO is indeed spurious. It transfers the control
3494 * from F-RTO to the conventional RTO recovery
3496 static int tcp_process_frto(struct sock
*sk
, int flag
)
3498 struct tcp_sock
*tp
= tcp_sk(sk
);
3500 tcp_verify_left_out(tp
);
3502 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3503 if (flag
& FLAG_DATA_ACKED
)
3504 inet_csk(sk
)->icsk_retransmits
= 0;
3506 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
3507 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
3508 tp
->undo_marker
= 0;
3510 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
3511 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
3515 if (!tcp_is_sackfrto(tp
)) {
3516 /* RFC4138 shortcoming in step 2; should also have case c):
3517 * ACK isn't duplicate nor advances window, e.g., opposite dir
3520 if (!(flag
& FLAG_ANY_PROGRESS
) && (flag
& FLAG_NOT_DUP
))
3523 if (!(flag
& FLAG_DATA_ACKED
)) {
3524 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
3529 if (!(flag
& FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
3530 /* Prevent sending of new data. */
3531 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
3532 tcp_packets_in_flight(tp
));
3536 if ((tp
->frto_counter
>= 2) &&
3537 (!(flag
& FLAG_FORWARD_PROGRESS
) ||
3538 ((flag
& FLAG_DATA_SACKED
) &&
3539 !(flag
& FLAG_ONLY_ORIG_SACKED
)))) {
3540 /* RFC4138 shortcoming (see comment above) */
3541 if (!(flag
& FLAG_FORWARD_PROGRESS
) &&
3542 (flag
& FLAG_NOT_DUP
))
3545 tcp_enter_frto_loss(sk
, 3, flag
);
3550 if (tp
->frto_counter
== 1) {
3551 /* tcp_may_send_now needs to see updated state */
3552 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3553 tp
->frto_counter
= 2;
3555 if (!tcp_may_send_now(sk
))
3556 tcp_enter_frto_loss(sk
, 2, flag
);
3560 switch (sysctl_tcp_frto_response
) {
3562 tcp_undo_spur_to_response(sk
, flag
);
3565 tcp_conservative_spur_to_response(tp
);
3568 tcp_ratehalving_spur_to_response(sk
);
3571 tp
->frto_counter
= 0;
3572 tp
->undo_marker
= 0;
3573 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSPURIOUSRTOS
);
3578 /* This routine deals with incoming acks, but not outgoing ones. */
3579 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
3581 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3582 struct tcp_sock
*tp
= tcp_sk(sk
);
3583 u32 prior_snd_una
= tp
->snd_una
;
3584 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3585 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3586 u32 prior_in_flight
;
3591 /* If the ack is newer than sent or older than previous acks
3592 * then we can probably ignore it.
3594 if (after(ack
, tp
->snd_nxt
))
3595 goto uninteresting_ack
;
3597 if (before(ack
, prior_snd_una
))
3600 if (after(ack
, prior_snd_una
))
3601 flag
|= FLAG_SND_UNA_ADVANCED
;
3603 if (sysctl_tcp_abc
) {
3604 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3605 tp
->bytes_acked
+= ack
- prior_snd_una
;
3606 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3607 /* we assume just one segment left network */
3608 tp
->bytes_acked
+= min(ack
- prior_snd_una
,
3612 prior_fackets
= tp
->fackets_out
;
3613 prior_in_flight
= tcp_packets_in_flight(tp
);
3615 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3616 /* Window is constant, pure forward advance.
3617 * No more checks are required.
3618 * Note, we use the fact that SND.UNA>=SND.WL2.
3620 tcp_update_wl(tp
, ack_seq
);
3622 flag
|= FLAG_WIN_UPDATE
;
3624 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3626 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3628 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3631 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3633 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3635 if (TCP_SKB_CB(skb
)->sacked
)
3636 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3638 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3641 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3644 /* We passed data and got it acked, remove any soft error
3645 * log. Something worked...
3647 sk
->sk_err_soft
= 0;
3648 icsk
->icsk_probes_out
= 0;
3649 tp
->rcv_tstamp
= tcp_time_stamp
;
3650 prior_packets
= tp
->packets_out
;
3654 /* See if we can take anything off of the retransmit queue. */
3655 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
);
3657 if (tp
->frto_counter
)
3658 frto_cwnd
= tcp_process_frto(sk
, flag
);
3659 /* Guarantee sacktag reordering detection against wrap-arounds */
3660 if (before(tp
->frto_highmark
, tp
->snd_una
))
3661 tp
->frto_highmark
= 0;
3663 if (tcp_ack_is_dubious(sk
, flag
)) {
3664 /* Advance CWND, if state allows this. */
3665 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3666 tcp_may_raise_cwnd(sk
, flag
))
3667 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3668 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
,
3671 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3672 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3675 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3676 dst_confirm(sk
->sk_dst_cache
);
3681 /* If this ack opens up a zero window, clear backoff. It was
3682 * being used to time the probes, and is probably far higher than
3683 * it needs to be for normal retransmission.
3685 if (tcp_send_head(sk
))
3690 if (TCP_SKB_CB(skb
)->sacked
) {
3691 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3692 if (icsk
->icsk_ca_state
== TCP_CA_Open
)
3693 tcp_try_keep_open(sk
);
3697 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3701 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3702 * But, this can also be called on packets in the established flow when
3703 * the fast version below fails.
3705 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
,
3709 struct tcphdr
*th
= tcp_hdr(skb
);
3710 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3712 ptr
= (unsigned char *)(th
+ 1);
3713 opt_rx
->saw_tstamp
= 0;
3715 while (length
> 0) {
3716 int opcode
= *ptr
++;
3722 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3727 if (opsize
< 2) /* "silly options" */
3729 if (opsize
> length
)
3730 return; /* don't parse partial options */
3733 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3734 u16 in_mss
= get_unaligned_be16(ptr
);
3736 if (opt_rx
->user_mss
&&
3737 opt_rx
->user_mss
< in_mss
)
3738 in_mss
= opt_rx
->user_mss
;
3739 opt_rx
->mss_clamp
= in_mss
;
3744 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3745 !estab
&& sysctl_tcp_window_scaling
) {
3746 __u8 snd_wscale
= *(__u8
*)ptr
;
3747 opt_rx
->wscale_ok
= 1;
3748 if (snd_wscale
> 14) {
3749 if (net_ratelimit())
3750 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3751 "scaling value %d >14 received.\n",
3755 opt_rx
->snd_wscale
= snd_wscale
;
3758 case TCPOPT_TIMESTAMP
:
3759 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3760 ((estab
&& opt_rx
->tstamp_ok
) ||
3761 (!estab
&& sysctl_tcp_timestamps
))) {
3762 opt_rx
->saw_tstamp
= 1;
3763 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3764 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3767 case TCPOPT_SACK_PERM
:
3768 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3769 !estab
&& sysctl_tcp_sack
) {
3770 opt_rx
->sack_ok
= 1;
3771 tcp_sack_reset(opt_rx
);
3776 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3777 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3779 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3782 #ifdef CONFIG_TCP_MD5SIG
3785 * The MD5 Hash has already been
3786 * checked (see tcp_v{4,6}_do_rcv()).
3798 static int tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, struct tcphdr
*th
)
3800 __be32
*ptr
= (__be32
*)(th
+ 1);
3802 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3803 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3804 tp
->rx_opt
.saw_tstamp
= 1;
3806 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3808 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3814 /* Fast parse options. This hopes to only see timestamps.
3815 * If it is wrong it falls back on tcp_parse_options().
3817 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3818 struct tcp_sock
*tp
)
3820 if (th
->doff
== sizeof(struct tcphdr
) >> 2) {
3821 tp
->rx_opt
.saw_tstamp
= 0;
3823 } else if (tp
->rx_opt
.tstamp_ok
&&
3824 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3825 if (tcp_parse_aligned_timestamp(tp
, th
))
3828 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3832 #ifdef CONFIG_TCP_MD5SIG
3834 * Parse MD5 Signature option
3836 u8
*tcp_parse_md5sig_option(struct tcphdr
*th
)
3838 int length
= (th
->doff
<< 2) - sizeof (*th
);
3839 u8
*ptr
= (u8
*)(th
+ 1);
3841 /* If the TCP option is too short, we can short cut */
3842 if (length
< TCPOLEN_MD5SIG
)
3845 while (length
> 0) {
3846 int opcode
= *ptr
++;
3857 if (opsize
< 2 || opsize
> length
)
3859 if (opcode
== TCPOPT_MD5SIG
)
3869 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3871 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3872 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3875 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3877 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3878 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3879 * extra check below makes sure this can only happen
3880 * for pure ACK frames. -DaveM
3882 * Not only, also it occurs for expired timestamps.
3885 if (tcp_paws_check(&tp
->rx_opt
, 0))
3886 tcp_store_ts_recent(tp
);
3890 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3892 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3893 * it can pass through stack. So, the following predicate verifies that
3894 * this segment is not used for anything but congestion avoidance or
3895 * fast retransmit. Moreover, we even are able to eliminate most of such
3896 * second order effects, if we apply some small "replay" window (~RTO)
3897 * to timestamp space.
3899 * All these measures still do not guarantee that we reject wrapped ACKs
3900 * on networks with high bandwidth, when sequence space is recycled fastly,
3901 * but it guarantees that such events will be very rare and do not affect
3902 * connection seriously. This doesn't look nice, but alas, PAWS is really
3905 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3906 * states that events when retransmit arrives after original data are rare.
3907 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3908 * the biggest problem on large power networks even with minor reordering.
3909 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3910 * up to bandwidth of 18Gigabit/sec. 8) ]
3913 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3915 struct tcp_sock
*tp
= tcp_sk(sk
);
3916 struct tcphdr
*th
= tcp_hdr(skb
);
3917 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3918 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3920 return (/* 1. Pure ACK with correct sequence number. */
3921 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3923 /* 2. ... and duplicate ACK. */
3924 ack
== tp
->snd_una
&&
3926 /* 3. ... and does not update window. */
3927 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3929 /* 4. ... and sits in replay window. */
3930 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3933 static inline int tcp_paws_discard(const struct sock
*sk
,
3934 const struct sk_buff
*skb
)
3936 const struct tcp_sock
*tp
= tcp_sk(sk
);
3938 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3939 !tcp_disordered_ack(sk
, skb
);
3942 /* Check segment sequence number for validity.
3944 * Segment controls are considered valid, if the segment
3945 * fits to the window after truncation to the window. Acceptability
3946 * of data (and SYN, FIN, of course) is checked separately.
3947 * See tcp_data_queue(), for example.
3949 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3950 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3951 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3952 * (borrowed from freebsd)
3955 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3957 return !before(end_seq
, tp
->rcv_wup
) &&
3958 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3961 /* When we get a reset we do this. */
3962 static void tcp_reset(struct sock
*sk
)
3964 /* We want the right error as BSD sees it (and indeed as we do). */
3965 switch (sk
->sk_state
) {
3967 sk
->sk_err
= ECONNREFUSED
;
3969 case TCP_CLOSE_WAIT
:
3975 sk
->sk_err
= ECONNRESET
;
3978 if (!sock_flag(sk
, SOCK_DEAD
))
3979 sk
->sk_error_report(sk
);
3985 * Process the FIN bit. This now behaves as it is supposed to work
3986 * and the FIN takes effect when it is validly part of sequence
3987 * space. Not before when we get holes.
3989 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3990 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3993 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3994 * close and we go into CLOSING (and later onto TIME-WAIT)
3996 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3998 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
4000 struct tcp_sock
*tp
= tcp_sk(sk
);
4002 inet_csk_schedule_ack(sk
);
4004 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4005 sock_set_flag(sk
, SOCK_DONE
);
4007 switch (sk
->sk_state
) {
4009 case TCP_ESTABLISHED
:
4010 /* Move to CLOSE_WAIT */
4011 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4012 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4015 case TCP_CLOSE_WAIT
:
4017 /* Received a retransmission of the FIN, do
4022 /* RFC793: Remain in the LAST-ACK state. */
4026 /* This case occurs when a simultaneous close
4027 * happens, we must ack the received FIN and
4028 * enter the CLOSING state.
4031 tcp_set_state(sk
, TCP_CLOSING
);
4034 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4036 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4039 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4040 * cases we should never reach this piece of code.
4042 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
4043 __func__
, sk
->sk_state
);
4047 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4048 * Probably, we should reset in this case. For now drop them.
4050 __skb_queue_purge(&tp
->out_of_order_queue
);
4051 if (tcp_is_sack(tp
))
4052 tcp_sack_reset(&tp
->rx_opt
);
4055 if (!sock_flag(sk
, SOCK_DEAD
)) {
4056 sk
->sk_state_change(sk
);
4058 /* Do not send POLL_HUP for half duplex close. */
4059 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4060 sk
->sk_state
== TCP_CLOSE
)
4061 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4063 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4067 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4070 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4071 if (before(seq
, sp
->start_seq
))
4072 sp
->start_seq
= seq
;
4073 if (after(end_seq
, sp
->end_seq
))
4074 sp
->end_seq
= end_seq
;
4080 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4082 struct tcp_sock
*tp
= tcp_sk(sk
);
4084 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4087 if (before(seq
, tp
->rcv_nxt
))
4088 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4090 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4092 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
4094 tp
->rx_opt
.dsack
= 1;
4095 tp
->duplicate_sack
[0].start_seq
= seq
;
4096 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4100 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4102 struct tcp_sock
*tp
= tcp_sk(sk
);
4104 if (!tp
->rx_opt
.dsack
)
4105 tcp_dsack_set(sk
, seq
, end_seq
);
4107 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4110 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
4112 struct tcp_sock
*tp
= tcp_sk(sk
);
4114 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4115 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4116 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4117 tcp_enter_quickack_mode(sk
);
4119 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4120 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4122 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4123 end_seq
= tp
->rcv_nxt
;
4124 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4131 /* These routines update the SACK block as out-of-order packets arrive or
4132 * in-order packets close up the sequence space.
4134 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4137 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4138 struct tcp_sack_block
*swalk
= sp
+ 1;
4140 /* See if the recent change to the first SACK eats into
4141 * or hits the sequence space of other SACK blocks, if so coalesce.
4143 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4144 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4147 /* Zap SWALK, by moving every further SACK up by one slot.
4148 * Decrease num_sacks.
4150 tp
->rx_opt
.num_sacks
--;
4151 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4155 this_sack
++, swalk
++;
4159 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
,
4160 struct tcp_sack_block
*sack2
)
4164 tmp
= sack1
->start_seq
;
4165 sack1
->start_seq
= sack2
->start_seq
;
4166 sack2
->start_seq
= tmp
;
4168 tmp
= sack1
->end_seq
;
4169 sack1
->end_seq
= sack2
->end_seq
;
4170 sack2
->end_seq
= tmp
;
4173 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4175 struct tcp_sock
*tp
= tcp_sk(sk
);
4176 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4177 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4183 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4184 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4185 /* Rotate this_sack to the first one. */
4186 for (; this_sack
> 0; this_sack
--, sp
--)
4187 tcp_sack_swap(sp
, sp
- 1);
4189 tcp_sack_maybe_coalesce(tp
);
4194 /* Could not find an adjacent existing SACK, build a new one,
4195 * put it at the front, and shift everyone else down. We
4196 * always know there is at least one SACK present already here.
4198 * If the sack array is full, forget about the last one.
4200 if (this_sack
>= TCP_NUM_SACKS
) {
4202 tp
->rx_opt
.num_sacks
--;
4205 for (; this_sack
> 0; this_sack
--, sp
--)
4209 /* Build the new head SACK, and we're done. */
4210 sp
->start_seq
= seq
;
4211 sp
->end_seq
= end_seq
;
4212 tp
->rx_opt
.num_sacks
++;
4215 /* RCV.NXT advances, some SACKs should be eaten. */
4217 static void tcp_sack_remove(struct tcp_sock
*tp
)
4219 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4220 int num_sacks
= tp
->rx_opt
.num_sacks
;
4223 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4224 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4225 tp
->rx_opt
.num_sacks
= 0;
4229 for (this_sack
= 0; this_sack
< num_sacks
;) {
4230 /* Check if the start of the sack is covered by RCV.NXT. */
4231 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4234 /* RCV.NXT must cover all the block! */
4235 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4237 /* Zap this SACK, by moving forward any other SACKS. */
4238 for (i
=this_sack
+1; i
< num_sacks
; i
++)
4239 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4246 tp
->rx_opt
.num_sacks
= num_sacks
;
4249 /* This one checks to see if we can put data from the
4250 * out_of_order queue into the receive_queue.
4252 static void tcp_ofo_queue(struct sock
*sk
)
4254 struct tcp_sock
*tp
= tcp_sk(sk
);
4255 __u32 dsack_high
= tp
->rcv_nxt
;
4256 struct sk_buff
*skb
;
4258 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4259 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4262 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4263 __u32 dsack
= dsack_high
;
4264 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4265 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4266 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4269 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4270 SOCK_DEBUG(sk
, "ofo packet was already received \n");
4271 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4275 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4276 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4277 TCP_SKB_CB(skb
)->end_seq
);
4279 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4280 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4281 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4282 if (tcp_hdr(skb
)->fin
)
4283 tcp_fin(skb
, sk
, tcp_hdr(skb
));
4287 static int tcp_prune_ofo_queue(struct sock
*sk
);
4288 static int tcp_prune_queue(struct sock
*sk
);
4290 static inline int tcp_try_rmem_schedule(struct sock
*sk
, unsigned int size
)
4292 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4293 !sk_rmem_schedule(sk
, size
)) {
4295 if (tcp_prune_queue(sk
) < 0)
4298 if (!sk_rmem_schedule(sk
, size
)) {
4299 if (!tcp_prune_ofo_queue(sk
))
4302 if (!sk_rmem_schedule(sk
, size
))
4309 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4311 struct tcphdr
*th
= tcp_hdr(skb
);
4312 struct tcp_sock
*tp
= tcp_sk(sk
);
4315 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4318 __skb_pull(skb
, th
->doff
* 4);
4320 TCP_ECN_accept_cwr(tp
, skb
);
4322 tp
->rx_opt
.dsack
= 0;
4324 /* Queue data for delivery to the user.
4325 * Packets in sequence go to the receive queue.
4326 * Out of sequence packets to the out_of_order_queue.
4328 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4329 if (tcp_receive_window(tp
) == 0)
4332 /* Ok. In sequence. In window. */
4333 if (tp
->ucopy
.task
== current
&&
4334 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4335 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4336 int chunk
= min_t(unsigned int, skb
->len
,
4339 __set_current_state(TASK_RUNNING
);
4342 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4343 tp
->ucopy
.len
-= chunk
;
4344 tp
->copied_seq
+= chunk
;
4345 eaten
= (chunk
== skb
->len
&& !th
->fin
);
4346 tcp_rcv_space_adjust(sk
);
4354 tcp_try_rmem_schedule(sk
, skb
->truesize
))
4357 skb_set_owner_r(skb
, sk
);
4358 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4360 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4362 tcp_event_data_recv(sk
, skb
);
4364 tcp_fin(skb
, sk
, th
);
4366 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4369 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4370 * gap in queue is filled.
4372 if (skb_queue_empty(&tp
->out_of_order_queue
))
4373 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4376 if (tp
->rx_opt
.num_sacks
)
4377 tcp_sack_remove(tp
);
4379 tcp_fast_path_check(sk
);
4383 else if (!sock_flag(sk
, SOCK_DEAD
))
4384 sk
->sk_data_ready(sk
, 0);
4388 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4389 /* A retransmit, 2nd most common case. Force an immediate ack. */
4390 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4391 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4394 tcp_enter_quickack_mode(sk
);
4395 inet_csk_schedule_ack(sk
);
4401 /* Out of window. F.e. zero window probe. */
4402 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4405 tcp_enter_quickack_mode(sk
);
4407 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4408 /* Partial packet, seq < rcv_next < end_seq */
4409 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4410 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4411 TCP_SKB_CB(skb
)->end_seq
);
4413 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4415 /* If window is closed, drop tail of packet. But after
4416 * remembering D-SACK for its head made in previous line.
4418 if (!tcp_receive_window(tp
))
4423 TCP_ECN_check_ce(tp
, skb
);
4425 if (tcp_try_rmem_schedule(sk
, skb
->truesize
))
4428 /* Disable header prediction. */
4430 inet_csk_schedule_ack(sk
);
4432 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4433 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4435 skb_set_owner_r(skb
, sk
);
4437 if (!skb_peek(&tp
->out_of_order_queue
)) {
4438 /* Initial out of order segment, build 1 SACK. */
4439 if (tcp_is_sack(tp
)) {
4440 tp
->rx_opt
.num_sacks
= 1;
4441 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4442 tp
->selective_acks
[0].end_seq
=
4443 TCP_SKB_CB(skb
)->end_seq
;
4445 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4447 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
4448 u32 seq
= TCP_SKB_CB(skb
)->seq
;
4449 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4451 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4452 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4454 if (!tp
->rx_opt
.num_sacks
||
4455 tp
->selective_acks
[0].end_seq
!= seq
)
4458 /* Common case: data arrive in order after hole. */
4459 tp
->selective_acks
[0].end_seq
= end_seq
;
4463 /* Find place to insert this segment. */
4465 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4467 } while ((skb1
= skb1
->prev
) !=
4468 (struct sk_buff
*)&tp
->out_of_order_queue
);
4470 /* Do skb overlap to previous one? */
4471 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
4472 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4473 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4474 /* All the bits are present. Drop. */
4476 tcp_dsack_set(sk
, seq
, end_seq
);
4479 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4480 /* Partial overlap. */
4481 tcp_dsack_set(sk
, seq
,
4482 TCP_SKB_CB(skb1
)->end_seq
);
4487 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4489 /* And clean segments covered by new one as whole. */
4490 while ((skb1
= skb
->next
) !=
4491 (struct sk_buff
*)&tp
->out_of_order_queue
&&
4492 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
4493 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4494 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4498 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4499 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4500 TCP_SKB_CB(skb1
)->end_seq
);
4505 if (tcp_is_sack(tp
))
4506 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4510 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4511 struct sk_buff_head
*list
)
4513 struct sk_buff
*next
= skb
->next
;
4515 __skb_unlink(skb
, list
);
4517 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4522 /* Collapse contiguous sequence of skbs head..tail with
4523 * sequence numbers start..end.
4524 * Segments with FIN/SYN are not collapsed (only because this
4528 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4529 struct sk_buff
*head
, struct sk_buff
*tail
,
4532 struct sk_buff
*skb
;
4534 /* First, check that queue is collapsible and find
4535 * the point where collapsing can be useful. */
4536 for (skb
= head
; skb
!= tail
;) {
4537 /* No new bits? It is possible on ofo queue. */
4538 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4539 skb
= tcp_collapse_one(sk
, skb
, list
);
4543 /* The first skb to collapse is:
4545 * - bloated or contains data before "start" or
4546 * overlaps to the next one.
4548 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4549 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4550 before(TCP_SKB_CB(skb
)->seq
, start
) ||
4551 (skb
->next
!= tail
&&
4552 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
4555 /* Decided to skip this, advance start seq. */
4556 start
= TCP_SKB_CB(skb
)->end_seq
;
4559 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4562 while (before(start
, end
)) {
4563 struct sk_buff
*nskb
;
4564 unsigned int header
= skb_headroom(skb
);
4565 int copy
= SKB_MAX_ORDER(header
, 0);
4567 /* Too big header? This can happen with IPv6. */
4570 if (end
- start
< copy
)
4572 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4576 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4577 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4579 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4581 skb_reserve(nskb
, header
);
4582 memcpy(nskb
->head
, skb
->head
, header
);
4583 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4584 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4585 __skb_queue_before(list
, skb
, nskb
);
4586 skb_set_owner_r(nskb
, sk
);
4588 /* Copy data, releasing collapsed skbs. */
4590 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4591 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4595 size
= min(copy
, size
);
4596 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4598 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4602 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4603 skb
= tcp_collapse_one(sk
, skb
, list
);
4605 tcp_hdr(skb
)->syn
||
4613 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4614 * and tcp_collapse() them until all the queue is collapsed.
4616 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4618 struct tcp_sock
*tp
= tcp_sk(sk
);
4619 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4620 struct sk_buff
*head
;
4626 start
= TCP_SKB_CB(skb
)->seq
;
4627 end
= TCP_SKB_CB(skb
)->end_seq
;
4633 /* Segment is terminated when we see gap or when
4634 * we are at the end of all the queue. */
4635 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
4636 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4637 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4638 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4639 head
, skb
, start
, end
);
4641 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
4643 /* Start new segment */
4644 start
= TCP_SKB_CB(skb
)->seq
;
4645 end
= TCP_SKB_CB(skb
)->end_seq
;
4647 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4648 start
= TCP_SKB_CB(skb
)->seq
;
4649 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4650 end
= TCP_SKB_CB(skb
)->end_seq
;
4656 * Purge the out-of-order queue.
4657 * Return true if queue was pruned.
4659 static int tcp_prune_ofo_queue(struct sock
*sk
)
4661 struct tcp_sock
*tp
= tcp_sk(sk
);
4664 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4665 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4666 __skb_queue_purge(&tp
->out_of_order_queue
);
4668 /* Reset SACK state. A conforming SACK implementation will
4669 * do the same at a timeout based retransmit. When a connection
4670 * is in a sad state like this, we care only about integrity
4671 * of the connection not performance.
4673 if (tp
->rx_opt
.sack_ok
)
4674 tcp_sack_reset(&tp
->rx_opt
);
4681 /* Reduce allocated memory if we can, trying to get
4682 * the socket within its memory limits again.
4684 * Return less than zero if we should start dropping frames
4685 * until the socket owning process reads some of the data
4686 * to stabilize the situation.
4688 static int tcp_prune_queue(struct sock
*sk
)
4690 struct tcp_sock
*tp
= tcp_sk(sk
);
4692 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4694 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4696 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4697 tcp_clamp_window(sk
);
4698 else if (tcp_memory_pressure
)
4699 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4701 tcp_collapse_ofo_queue(sk
);
4702 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4703 sk
->sk_receive_queue
.next
,
4704 (struct sk_buff
*)&sk
->sk_receive_queue
,
4705 tp
->copied_seq
, tp
->rcv_nxt
);
4708 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4711 /* Collapsing did not help, destructive actions follow.
4712 * This must not ever occur. */
4714 tcp_prune_ofo_queue(sk
);
4716 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4719 /* If we are really being abused, tell the caller to silently
4720 * drop receive data on the floor. It will get retransmitted
4721 * and hopefully then we'll have sufficient space.
4723 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4725 /* Massive buffer overcommit. */
4730 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4731 * As additional protections, we do not touch cwnd in retransmission phases,
4732 * and if application hit its sndbuf limit recently.
4734 void tcp_cwnd_application_limited(struct sock
*sk
)
4736 struct tcp_sock
*tp
= tcp_sk(sk
);
4738 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4739 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4740 /* Limited by application or receiver window. */
4741 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4742 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4743 if (win_used
< tp
->snd_cwnd
) {
4744 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4745 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4747 tp
->snd_cwnd_used
= 0;
4749 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4752 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4754 struct tcp_sock
*tp
= tcp_sk(sk
);
4756 /* If the user specified a specific send buffer setting, do
4759 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4762 /* If we are under global TCP memory pressure, do not expand. */
4763 if (tcp_memory_pressure
)
4766 /* If we are under soft global TCP memory pressure, do not expand. */
4767 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4770 /* If we filled the congestion window, do not expand. */
4771 if (tp
->packets_out
>= tp
->snd_cwnd
)
4777 /* When incoming ACK allowed to free some skb from write_queue,
4778 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4779 * on the exit from tcp input handler.
4781 * PROBLEM: sndbuf expansion does not work well with largesend.
4783 static void tcp_new_space(struct sock
*sk
)
4785 struct tcp_sock
*tp
= tcp_sk(sk
);
4787 if (tcp_should_expand_sndbuf(sk
)) {
4788 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4789 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
4790 int demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4791 tp
->reordering
+ 1);
4792 sndmem
*= 2 * demanded
;
4793 if (sndmem
> sk
->sk_sndbuf
)
4794 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4795 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4798 sk
->sk_write_space(sk
);
4801 static void tcp_check_space(struct sock
*sk
)
4803 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4804 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4805 if (sk
->sk_socket
&&
4806 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4811 static inline void tcp_data_snd_check(struct sock
*sk
)
4813 tcp_push_pending_frames(sk
);
4814 tcp_check_space(sk
);
4818 * Check if sending an ack is needed.
4820 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4822 struct tcp_sock
*tp
= tcp_sk(sk
);
4824 /* More than one full frame received... */
4825 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
4826 /* ... and right edge of window advances far enough.
4827 * (tcp_recvmsg() will send ACK otherwise). Or...
4829 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4830 /* We ACK each frame or... */
4831 tcp_in_quickack_mode(sk
) ||
4832 /* We have out of order data. */
4833 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4834 /* Then ack it now */
4837 /* Else, send delayed ack. */
4838 tcp_send_delayed_ack(sk
);
4842 static inline void tcp_ack_snd_check(struct sock
*sk
)
4844 if (!inet_csk_ack_scheduled(sk
)) {
4845 /* We sent a data segment already. */
4848 __tcp_ack_snd_check(sk
, 1);
4852 * This routine is only called when we have urgent data
4853 * signaled. Its the 'slow' part of tcp_urg. It could be
4854 * moved inline now as tcp_urg is only called from one
4855 * place. We handle URGent data wrong. We have to - as
4856 * BSD still doesn't use the correction from RFC961.
4857 * For 1003.1g we should support a new option TCP_STDURG to permit
4858 * either form (or just set the sysctl tcp_stdurg).
4861 static void tcp_check_urg(struct sock
*sk
, struct tcphdr
*th
)
4863 struct tcp_sock
*tp
= tcp_sk(sk
);
4864 u32 ptr
= ntohs(th
->urg_ptr
);
4866 if (ptr
&& !sysctl_tcp_stdurg
)
4868 ptr
+= ntohl(th
->seq
);
4870 /* Ignore urgent data that we've already seen and read. */
4871 if (after(tp
->copied_seq
, ptr
))
4874 /* Do not replay urg ptr.
4876 * NOTE: interesting situation not covered by specs.
4877 * Misbehaving sender may send urg ptr, pointing to segment,
4878 * which we already have in ofo queue. We are not able to fetch
4879 * such data and will stay in TCP_URG_NOTYET until will be eaten
4880 * by recvmsg(). Seems, we are not obliged to handle such wicked
4881 * situations. But it is worth to think about possibility of some
4882 * DoSes using some hypothetical application level deadlock.
4884 if (before(ptr
, tp
->rcv_nxt
))
4887 /* Do we already have a newer (or duplicate) urgent pointer? */
4888 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4891 /* Tell the world about our new urgent pointer. */
4894 /* We may be adding urgent data when the last byte read was
4895 * urgent. To do this requires some care. We cannot just ignore
4896 * tp->copied_seq since we would read the last urgent byte again
4897 * as data, nor can we alter copied_seq until this data arrives
4898 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4900 * NOTE. Double Dutch. Rendering to plain English: author of comment
4901 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4902 * and expect that both A and B disappear from stream. This is _wrong_.
4903 * Though this happens in BSD with high probability, this is occasional.
4904 * Any application relying on this is buggy. Note also, that fix "works"
4905 * only in this artificial test. Insert some normal data between A and B and we will
4906 * decline of BSD again. Verdict: it is better to remove to trap
4909 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4910 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4911 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4913 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4914 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4919 tp
->urg_data
= TCP_URG_NOTYET
;
4922 /* Disable header prediction. */
4926 /* This is the 'fast' part of urgent handling. */
4927 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4929 struct tcp_sock
*tp
= tcp_sk(sk
);
4931 /* Check if we get a new urgent pointer - normally not. */
4933 tcp_check_urg(sk
, th
);
4935 /* Do we wait for any urgent data? - normally not... */
4936 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4937 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4940 /* Is the urgent pointer pointing into this packet? */
4941 if (ptr
< skb
->len
) {
4943 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4945 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4946 if (!sock_flag(sk
, SOCK_DEAD
))
4947 sk
->sk_data_ready(sk
, 0);
4952 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4954 struct tcp_sock
*tp
= tcp_sk(sk
);
4955 int chunk
= skb
->len
- hlen
;
4959 if (skb_csum_unnecessary(skb
))
4960 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4962 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4966 tp
->ucopy
.len
-= chunk
;
4967 tp
->copied_seq
+= chunk
;
4968 tcp_rcv_space_adjust(sk
);
4975 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4976 struct sk_buff
*skb
)
4980 if (sock_owned_by_user(sk
)) {
4982 result
= __tcp_checksum_complete(skb
);
4985 result
= __tcp_checksum_complete(skb
);
4990 static inline int tcp_checksum_complete_user(struct sock
*sk
,
4991 struct sk_buff
*skb
)
4993 return !skb_csum_unnecessary(skb
) &&
4994 __tcp_checksum_complete_user(sk
, skb
);
4997 #ifdef CONFIG_NET_DMA
4998 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
5001 struct tcp_sock
*tp
= tcp_sk(sk
);
5002 int chunk
= skb
->len
- hlen
;
5004 int copied_early
= 0;
5006 if (tp
->ucopy
.wakeup
)
5009 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
5010 tp
->ucopy
.dma_chan
= dma_find_channel(DMA_MEMCPY
);
5012 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
5014 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
5016 tp
->ucopy
.iov
, chunk
,
5017 tp
->ucopy
.pinned_list
);
5022 tp
->ucopy
.dma_cookie
= dma_cookie
;
5025 tp
->ucopy
.len
-= chunk
;
5026 tp
->copied_seq
+= chunk
;
5027 tcp_rcv_space_adjust(sk
);
5029 if ((tp
->ucopy
.len
== 0) ||
5030 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
5031 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
5032 tp
->ucopy
.wakeup
= 1;
5033 sk
->sk_data_ready(sk
, 0);
5035 } else if (chunk
> 0) {
5036 tp
->ucopy
.wakeup
= 1;
5037 sk
->sk_data_ready(sk
, 0);
5040 return copied_early
;
5042 #endif /* CONFIG_NET_DMA */
5044 /* Does PAWS and seqno based validation of an incoming segment, flags will
5045 * play significant role here.
5047 static int tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5048 struct tcphdr
*th
, int syn_inerr
)
5050 struct tcp_sock
*tp
= tcp_sk(sk
);
5052 /* RFC1323: H1. Apply PAWS check first. */
5053 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5054 tcp_paws_discard(sk
, skb
)) {
5056 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5057 tcp_send_dupack(sk
, skb
);
5060 /* Reset is accepted even if it did not pass PAWS. */
5063 /* Step 1: check sequence number */
5064 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5065 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5066 * (RST) segments are validated by checking their SEQ-fields."
5067 * And page 69: "If an incoming segment is not acceptable,
5068 * an acknowledgment should be sent in reply (unless the RST
5069 * bit is set, if so drop the segment and return)".
5072 tcp_send_dupack(sk
, skb
);
5076 /* Step 2: check RST bit */
5082 /* ts_recent update must be made after we are sure that the packet
5085 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
5087 /* step 3: check security and precedence [ignored] */
5089 /* step 4: Check for a SYN in window. */
5090 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
5092 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5093 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONSYN
);
5106 * TCP receive function for the ESTABLISHED state.
5108 * It is split into a fast path and a slow path. The fast path is
5110 * - A zero window was announced from us - zero window probing
5111 * is only handled properly in the slow path.
5112 * - Out of order segments arrived.
5113 * - Urgent data is expected.
5114 * - There is no buffer space left
5115 * - Unexpected TCP flags/window values/header lengths are received
5116 * (detected by checking the TCP header against pred_flags)
5117 * - Data is sent in both directions. Fast path only supports pure senders
5118 * or pure receivers (this means either the sequence number or the ack
5119 * value must stay constant)
5120 * - Unexpected TCP option.
5122 * When these conditions are not satisfied it drops into a standard
5123 * receive procedure patterned after RFC793 to handle all cases.
5124 * The first three cases are guaranteed by proper pred_flags setting,
5125 * the rest is checked inline. Fast processing is turned on in
5126 * tcp_data_queue when everything is OK.
5128 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5129 struct tcphdr
*th
, unsigned len
)
5131 struct tcp_sock
*tp
= tcp_sk(sk
);
5135 * Header prediction.
5136 * The code loosely follows the one in the famous
5137 * "30 instruction TCP receive" Van Jacobson mail.
5139 * Van's trick is to deposit buffers into socket queue
5140 * on a device interrupt, to call tcp_recv function
5141 * on the receive process context and checksum and copy
5142 * the buffer to user space. smart...
5144 * Our current scheme is not silly either but we take the
5145 * extra cost of the net_bh soft interrupt processing...
5146 * We do checksum and copy also but from device to kernel.
5149 tp
->rx_opt
.saw_tstamp
= 0;
5151 /* pred_flags is 0xS?10 << 16 + snd_wnd
5152 * if header_prediction is to be made
5153 * 'S' will always be tp->tcp_header_len >> 2
5154 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5155 * turn it off (when there are holes in the receive
5156 * space for instance)
5157 * PSH flag is ignored.
5160 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5161 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
5162 int tcp_header_len
= tp
->tcp_header_len
;
5164 /* Timestamp header prediction: tcp_header_len
5165 * is automatically equal to th->doff*4 due to pred_flags
5169 /* Check timestamp */
5170 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5171 /* No? Slow path! */
5172 if (!tcp_parse_aligned_timestamp(tp
, th
))
5175 /* If PAWS failed, check it more carefully in slow path */
5176 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5179 /* DO NOT update ts_recent here, if checksum fails
5180 * and timestamp was corrupted part, it will result
5181 * in a hung connection since we will drop all
5182 * future packets due to the PAWS test.
5186 if (len
<= tcp_header_len
) {
5187 /* Bulk data transfer: sender */
5188 if (len
== tcp_header_len
) {
5189 /* Predicted packet is in window by definition.
5190 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5191 * Hence, check seq<=rcv_wup reduces to:
5193 if (tcp_header_len
==
5194 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5195 tp
->rcv_nxt
== tp
->rcv_wup
)
5196 tcp_store_ts_recent(tp
);
5198 /* We know that such packets are checksummed
5201 tcp_ack(sk
, skb
, 0);
5203 tcp_data_snd_check(sk
);
5205 } else { /* Header too small */
5206 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5211 int copied_early
= 0;
5213 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5214 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5215 #ifdef CONFIG_NET_DMA
5216 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5221 if (tp
->ucopy
.task
== current
&&
5222 sock_owned_by_user(sk
) && !copied_early
) {
5223 __set_current_state(TASK_RUNNING
);
5225 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5229 /* Predicted packet is in window by definition.
5230 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5231 * Hence, check seq<=rcv_wup reduces to:
5233 if (tcp_header_len
==
5234 (sizeof(struct tcphdr
) +
5235 TCPOLEN_TSTAMP_ALIGNED
) &&
5236 tp
->rcv_nxt
== tp
->rcv_wup
)
5237 tcp_store_ts_recent(tp
);
5239 tcp_rcv_rtt_measure_ts(sk
, skb
);
5241 __skb_pull(skb
, tcp_header_len
);
5242 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5243 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5246 tcp_cleanup_rbuf(sk
, skb
->len
);
5249 if (tcp_checksum_complete_user(sk
, skb
))
5252 /* Predicted packet is in window by definition.
5253 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5254 * Hence, check seq<=rcv_wup reduces to:
5256 if (tcp_header_len
==
5257 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5258 tp
->rcv_nxt
== tp
->rcv_wup
)
5259 tcp_store_ts_recent(tp
);
5261 tcp_rcv_rtt_measure_ts(sk
, skb
);
5263 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5266 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5268 /* Bulk data transfer: receiver */
5269 __skb_pull(skb
, tcp_header_len
);
5270 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
5271 skb_set_owner_r(skb
, sk
);
5272 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5275 tcp_event_data_recv(sk
, skb
);
5277 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5278 /* Well, only one small jumplet in fast path... */
5279 tcp_ack(sk
, skb
, FLAG_DATA
);
5280 tcp_data_snd_check(sk
);
5281 if (!inet_csk_ack_scheduled(sk
))
5285 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5286 __tcp_ack_snd_check(sk
, 0);
5288 #ifdef CONFIG_NET_DMA
5290 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5296 sk
->sk_data_ready(sk
, 0);
5302 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5306 * Standard slow path.
5309 res
= tcp_validate_incoming(sk
, skb
, th
, 1);
5315 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5317 tcp_rcv_rtt_measure_ts(sk
, skb
);
5319 /* Process urgent data. */
5320 tcp_urg(sk
, skb
, th
);
5322 /* step 7: process the segment text */
5323 tcp_data_queue(sk
, skb
);
5325 tcp_data_snd_check(sk
);
5326 tcp_ack_snd_check(sk
);
5330 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5337 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5338 struct tcphdr
*th
, unsigned len
)
5340 struct tcp_sock
*tp
= tcp_sk(sk
);
5341 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5342 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5344 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
5348 * "If the state is SYN-SENT then
5349 * first check the ACK bit
5350 * If the ACK bit is set
5351 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5352 * a reset (unless the RST bit is set, if so drop
5353 * the segment and return)"
5355 * We do not send data with SYN, so that RFC-correct
5358 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
5359 goto reset_and_undo
;
5361 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5362 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5364 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5365 goto reset_and_undo
;
5368 /* Now ACK is acceptable.
5370 * "If the RST bit is set
5371 * If the ACK was acceptable then signal the user "error:
5372 * connection reset", drop the segment, enter CLOSED state,
5373 * delete TCB, and return."
5382 * "fifth, if neither of the SYN or RST bits is set then
5383 * drop the segment and return."
5389 goto discard_and_undo
;
5392 * "If the SYN bit is on ...
5393 * are acceptable then ...
5394 * (our SYN has been ACKed), change the connection
5395 * state to ESTABLISHED..."
5398 TCP_ECN_rcv_synack(tp
, th
);
5400 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5401 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5403 /* Ok.. it's good. Set up sequence numbers and
5404 * move to established.
5406 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5407 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5409 /* RFC1323: The window in SYN & SYN/ACK segments is
5412 tp
->snd_wnd
= ntohs(th
->window
);
5413 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5415 if (!tp
->rx_opt
.wscale_ok
) {
5416 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5417 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5420 if (tp
->rx_opt
.saw_tstamp
) {
5421 tp
->rx_opt
.tstamp_ok
= 1;
5422 tp
->tcp_header_len
=
5423 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5424 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5425 tcp_store_ts_recent(tp
);
5427 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5430 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5431 tcp_enable_fack(tp
);
5434 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5435 tcp_initialize_rcv_mss(sk
);
5437 /* Remember, tcp_poll() does not lock socket!
5438 * Change state from SYN-SENT only after copied_seq
5439 * is initialized. */
5440 tp
->copied_seq
= tp
->rcv_nxt
;
5442 tcp_set_state(sk
, TCP_ESTABLISHED
);
5444 security_inet_conn_established(sk
, skb
);
5446 /* Make sure socket is routed, for correct metrics. */
5447 icsk
->icsk_af_ops
->rebuild_header(sk
);
5449 tcp_init_metrics(sk
);
5451 tcp_init_congestion_control(sk
);
5453 /* Prevent spurious tcp_cwnd_restart() on first data
5456 tp
->lsndtime
= tcp_time_stamp
;
5458 tcp_init_buffer_space(sk
);
5460 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5461 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5463 if (!tp
->rx_opt
.snd_wscale
)
5464 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5468 if (!sock_flag(sk
, SOCK_DEAD
)) {
5469 sk
->sk_state_change(sk
);
5470 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5473 if (sk
->sk_write_pending
||
5474 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5475 icsk
->icsk_ack
.pingpong
) {
5476 /* Save one ACK. Data will be ready after
5477 * several ticks, if write_pending is set.
5479 * It may be deleted, but with this feature tcpdumps
5480 * look so _wonderfully_ clever, that I was not able
5481 * to stand against the temptation 8) --ANK
5483 inet_csk_schedule_ack(sk
);
5484 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5485 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
5486 tcp_incr_quickack(sk
);
5487 tcp_enter_quickack_mode(sk
);
5488 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5489 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5500 /* No ACK in the segment */
5504 * "If the RST bit is set
5506 * Otherwise (no ACK) drop the segment and return."
5509 goto discard_and_undo
;
5513 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5514 tcp_paws_reject(&tp
->rx_opt
, 0))
5515 goto discard_and_undo
;
5518 /* We see SYN without ACK. It is attempt of
5519 * simultaneous connect with crossed SYNs.
5520 * Particularly, it can be connect to self.
5522 tcp_set_state(sk
, TCP_SYN_RECV
);
5524 if (tp
->rx_opt
.saw_tstamp
) {
5525 tp
->rx_opt
.tstamp_ok
= 1;
5526 tcp_store_ts_recent(tp
);
5527 tp
->tcp_header_len
=
5528 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5530 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5533 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5534 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5536 /* RFC1323: The window in SYN & SYN/ACK segments is
5539 tp
->snd_wnd
= ntohs(th
->window
);
5540 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5541 tp
->max_window
= tp
->snd_wnd
;
5543 TCP_ECN_rcv_syn(tp
, th
);
5546 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5547 tcp_initialize_rcv_mss(sk
);
5549 tcp_send_synack(sk
);
5551 /* Note, we could accept data and URG from this segment.
5552 * There are no obstacles to make this.
5554 * However, if we ignore data in ACKless segments sometimes,
5555 * we have no reasons to accept it sometimes.
5556 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5557 * is not flawless. So, discard packet for sanity.
5558 * Uncomment this return to process the data.
5565 /* "fifth, if neither of the SYN or RST bits is set then
5566 * drop the segment and return."
5570 tcp_clear_options(&tp
->rx_opt
);
5571 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5575 tcp_clear_options(&tp
->rx_opt
);
5576 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5581 * This function implements the receiving procedure of RFC 793 for
5582 * all states except ESTABLISHED and TIME_WAIT.
5583 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5584 * address independent.
5587 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5588 struct tcphdr
*th
, unsigned len
)
5590 struct tcp_sock
*tp
= tcp_sk(sk
);
5591 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5595 tp
->rx_opt
.saw_tstamp
= 0;
5597 switch (sk
->sk_state
) {
5609 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5612 /* Now we have several options: In theory there is
5613 * nothing else in the frame. KA9Q has an option to
5614 * send data with the syn, BSD accepts data with the
5615 * syn up to the [to be] advertised window and
5616 * Solaris 2.1 gives you a protocol error. For now
5617 * we just ignore it, that fits the spec precisely
5618 * and avoids incompatibilities. It would be nice in
5619 * future to drop through and process the data.
5621 * Now that TTCP is starting to be used we ought to
5623 * But, this leaves one open to an easy denial of
5624 * service attack, and SYN cookies can't defend
5625 * against this problem. So, we drop the data
5626 * in the interest of security over speed unless
5627 * it's still in use.
5635 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5639 /* Do step6 onward by hand. */
5640 tcp_urg(sk
, skb
, th
);
5642 tcp_data_snd_check(sk
);
5646 res
= tcp_validate_incoming(sk
, skb
, th
, 0);
5650 /* step 5: check the ACK field */
5652 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5654 switch (sk
->sk_state
) {
5657 tp
->copied_seq
= tp
->rcv_nxt
;
5659 tcp_set_state(sk
, TCP_ESTABLISHED
);
5660 sk
->sk_state_change(sk
);
5662 /* Note, that this wakeup is only for marginal
5663 * crossed SYN case. Passively open sockets
5664 * are not waked up, because sk->sk_sleep ==
5665 * NULL and sk->sk_socket == NULL.
5669 SOCK_WAKE_IO
, POLL_OUT
);
5671 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5672 tp
->snd_wnd
= ntohs(th
->window
) <<
5673 tp
->rx_opt
.snd_wscale
;
5674 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5676 /* tcp_ack considers this ACK as duplicate
5677 * and does not calculate rtt.
5678 * Fix it at least with timestamps.
5680 if (tp
->rx_opt
.saw_tstamp
&&
5681 tp
->rx_opt
.rcv_tsecr
&& !tp
->srtt
)
5682 tcp_ack_saw_tstamp(sk
, 0);
5684 if (tp
->rx_opt
.tstamp_ok
)
5685 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5687 /* Make sure socket is routed, for
5690 icsk
->icsk_af_ops
->rebuild_header(sk
);
5692 tcp_init_metrics(sk
);
5694 tcp_init_congestion_control(sk
);
5696 /* Prevent spurious tcp_cwnd_restart() on
5697 * first data packet.
5699 tp
->lsndtime
= tcp_time_stamp
;
5702 tcp_initialize_rcv_mss(sk
);
5703 tcp_init_buffer_space(sk
);
5704 tcp_fast_path_on(tp
);
5711 if (tp
->snd_una
== tp
->write_seq
) {
5712 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5713 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5714 dst_confirm(sk
->sk_dst_cache
);
5716 if (!sock_flag(sk
, SOCK_DEAD
))
5717 /* Wake up lingering close() */
5718 sk
->sk_state_change(sk
);
5722 if (tp
->linger2
< 0 ||
5723 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5724 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5726 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5730 tmo
= tcp_fin_time(sk
);
5731 if (tmo
> TCP_TIMEWAIT_LEN
) {
5732 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5733 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5734 /* Bad case. We could lose such FIN otherwise.
5735 * It is not a big problem, but it looks confusing
5736 * and not so rare event. We still can lose it now,
5737 * if it spins in bh_lock_sock(), but it is really
5740 inet_csk_reset_keepalive_timer(sk
, tmo
);
5742 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5750 if (tp
->snd_una
== tp
->write_seq
) {
5751 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5757 if (tp
->snd_una
== tp
->write_seq
) {
5758 tcp_update_metrics(sk
);
5767 /* step 6: check the URG bit */
5768 tcp_urg(sk
, skb
, th
);
5770 /* step 7: process the segment text */
5771 switch (sk
->sk_state
) {
5772 case TCP_CLOSE_WAIT
:
5775 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5779 /* RFC 793 says to queue data in these states,
5780 * RFC 1122 says we MUST send a reset.
5781 * BSD 4.4 also does reset.
5783 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5784 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5785 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5786 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5792 case TCP_ESTABLISHED
:
5793 tcp_data_queue(sk
, skb
);
5798 /* tcp_data could move socket to TIME-WAIT */
5799 if (sk
->sk_state
!= TCP_CLOSE
) {
5800 tcp_data_snd_check(sk
);
5801 tcp_ack_snd_check(sk
);
5811 EXPORT_SYMBOL(sysctl_tcp_ecn
);
5812 EXPORT_SYMBOL(sysctl_tcp_reordering
);
5813 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
5814 EXPORT_SYMBOL(tcp_parse_options
);
5815 #ifdef CONFIG_TCP_MD5SIG
5816 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
5818 EXPORT_SYMBOL(tcp_rcv_established
);
5819 EXPORT_SYMBOL(tcp_rcv_state_process
);
5820 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);