Version:  2.0.40 2.2.26 2.4.37 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19

Linux/net/ipv4/tcp_input.c

  1 /*
  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.
  5  *
  6  *              Implementation of the Transmission Control Protocol(TCP).
  7  *
  8  * Authors:     Ross Biro
  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>
 19  */
 20 
 21 /*
 22  * Changes:
 23  *              Pedro Roque     :       Fast Retransmit/Recovery.
 24  *                                      Two receive queues.
 25  *                                      Retransmit queue handled by TCP.
 26  *                                      Better retransmit timer handling.
 27  *                                      New congestion avoidance.
 28  *                                      Header prediction.
 29  *                                      Variable renaming.
 30  *
 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
 44  *                                      timestamps.
 45  *              Andrey Savochkin:       Check sequence numbers correctly when
 46  *                                      removing SACKs due to in sequence incoming
 47  *                                      data segments.
 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
 55  *                                      fast path.
 56  *              J Hadi Salim:           ECN support
 57  *              Andrei Gurtov,
 58  *              Pasi Sarolahti,
 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
 62  */
 63 
 64 #define pr_fmt(fmt) "TCP: " fmt
 65 
 66 #include <linux/mm.h>
 67 #include <linux/slab.h>
 68 #include <linux/module.h>
 69 #include <linux/sysctl.h>
 70 #include <linux/kernel.h>
 71 #include <linux/prefetch.h>
 72 #include <net/dst.h>
 73 #include <net/tcp.h>
 74 #include <net/inet_common.h>
 75 #include <linux/ipsec.h>
 76 #include <asm/unaligned.h>
 77 #include <linux/errqueue.h>
 78 
 79 int sysctl_tcp_timestamps __read_mostly = 1;
 80 int sysctl_tcp_window_scaling __read_mostly = 1;
 81 int sysctl_tcp_sack __read_mostly = 1;
 82 int sysctl_tcp_fack __read_mostly = 1;
 83 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
 84 int sysctl_tcp_max_reordering __read_mostly = 300;
 85 EXPORT_SYMBOL(sysctl_tcp_reordering);
 86 int sysctl_tcp_dsack __read_mostly = 1;
 87 int sysctl_tcp_app_win __read_mostly = 31;
 88 int sysctl_tcp_adv_win_scale __read_mostly = 1;
 89 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
 90 
 91 /* rfc5961 challenge ack rate limiting */
 92 int sysctl_tcp_challenge_ack_limit = 100;
 93 
 94 int sysctl_tcp_stdurg __read_mostly;
 95 int sysctl_tcp_rfc1337 __read_mostly;
 96 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
 97 int sysctl_tcp_frto __read_mostly = 2;
 98 
 99 int sysctl_tcp_thin_dupack __read_mostly;
100 
101 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
102 int sysctl_tcp_early_retrans __read_mostly = 3;
103 
104 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
105 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
106 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
107 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
108 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
109 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
110 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
111 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
112 #define FLAG_ORIG_SACK_ACKED    0x200 /* Never retransmitted data are (s)acked  */
113 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
114 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
115 #define FLAG_SACK_RENEGING      0x2000 /* snd_una advanced to a sacked seq */
116 #define FLAG_UPDATE_TS_RECENT   0x4000 /* tcp_replace_ts_recent() */
117 
118 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
119 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
120 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
121 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
122 
123 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
124 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
125 
126 /* Adapt the MSS value used to make delayed ack decision to the
127  * real world.
128  */
129 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
130 {
131         struct inet_connection_sock *icsk = inet_csk(sk);
132         const unsigned int lss = icsk->icsk_ack.last_seg_size;
133         unsigned int len;
134 
135         icsk->icsk_ack.last_seg_size = 0;
136 
137         /* skb->len may jitter because of SACKs, even if peer
138          * sends good full-sized frames.
139          */
140         len = skb_shinfo(skb)->gso_size ? : skb->len;
141         if (len >= icsk->icsk_ack.rcv_mss) {
142                 icsk->icsk_ack.rcv_mss = len;
143         } else {
144                 /* Otherwise, we make more careful check taking into account,
145                  * that SACKs block is variable.
146                  *
147                  * "len" is invariant segment length, including TCP header.
148                  */
149                 len += skb->data - skb_transport_header(skb);
150                 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
151                     /* If PSH is not set, packet should be
152                      * full sized, provided peer TCP is not badly broken.
153                      * This observation (if it is correct 8)) allows
154                      * to handle super-low mtu links fairly.
155                      */
156                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
157                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
158                         /* Subtract also invariant (if peer is RFC compliant),
159                          * tcp header plus fixed timestamp option length.
160                          * Resulting "len" is MSS free of SACK jitter.
161                          */
162                         len -= tcp_sk(sk)->tcp_header_len;
163                         icsk->icsk_ack.last_seg_size = len;
164                         if (len == lss) {
165                                 icsk->icsk_ack.rcv_mss = len;
166                                 return;
167                         }
168                 }
169                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
170                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
171                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
172         }
173 }
174 
175 static void tcp_incr_quickack(struct sock *sk)
176 {
177         struct inet_connection_sock *icsk = inet_csk(sk);
178         unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
179 
180         if (quickacks == 0)
181                 quickacks = 2;
182         if (quickacks > icsk->icsk_ack.quick)
183                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
184 }
185 
186 static void tcp_enter_quickack_mode(struct sock *sk)
187 {
188         struct inet_connection_sock *icsk = inet_csk(sk);
189         tcp_incr_quickack(sk);
190         icsk->icsk_ack.pingpong = 0;
191         icsk->icsk_ack.ato = TCP_ATO_MIN;
192 }
193 
194 /* Send ACKs quickly, if "quick" count is not exhausted
195  * and the session is not interactive.
196  */
197 
198 static inline bool tcp_in_quickack_mode(const struct sock *sk)
199 {
200         const struct inet_connection_sock *icsk = inet_csk(sk);
201 
202         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
203 }
204 
205 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
206 {
207         if (tp->ecn_flags & TCP_ECN_OK)
208                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
209 }
210 
211 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
212 {
213         if (tcp_hdr(skb)->cwr)
214                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
215 }
216 
217 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
218 {
219         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
220 }
221 
222 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
223 {
224         switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
225         case INET_ECN_NOT_ECT:
226                 /* Funny extension: if ECT is not set on a segment,
227                  * and we already seen ECT on a previous segment,
228                  * it is probably a retransmit.
229                  */
230                 if (tp->ecn_flags & TCP_ECN_SEEN)
231                         tcp_enter_quickack_mode((struct sock *)tp);
232                 break;
233         case INET_ECN_CE:
234                 if (tcp_ca_needs_ecn((struct sock *)tp))
235                         tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
236 
237                 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
238                         /* Better not delay acks, sender can have a very low cwnd */
239                         tcp_enter_quickack_mode((struct sock *)tp);
240                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
241                 }
242                 tp->ecn_flags |= TCP_ECN_SEEN;
243                 break;
244         default:
245                 if (tcp_ca_needs_ecn((struct sock *)tp))
246                         tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
247                 tp->ecn_flags |= TCP_ECN_SEEN;
248                 break;
249         }
250 }
251 
252 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
253 {
254         if (tp->ecn_flags & TCP_ECN_OK)
255                 __tcp_ecn_check_ce(tp, skb);
256 }
257 
258 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
259 {
260         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
261                 tp->ecn_flags &= ~TCP_ECN_OK;
262 }
263 
264 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
265 {
266         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
267                 tp->ecn_flags &= ~TCP_ECN_OK;
268 }
269 
270 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
271 {
272         if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
273                 return true;
274         return false;
275 }
276 
277 /* Buffer size and advertised window tuning.
278  *
279  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
280  */
281 
282 static void tcp_sndbuf_expand(struct sock *sk)
283 {
284         const struct tcp_sock *tp = tcp_sk(sk);
285         int sndmem, per_mss;
286         u32 nr_segs;
287 
288         /* Worst case is non GSO/TSO : each frame consumes one skb
289          * and skb->head is kmalloced using power of two area of memory
290          */
291         per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
292                   MAX_TCP_HEADER +
293                   SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
294 
295         per_mss = roundup_pow_of_two(per_mss) +
296                   SKB_DATA_ALIGN(sizeof(struct sk_buff));
297 
298         nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
299         nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
300 
301         /* Fast Recovery (RFC 5681 3.2) :
302          * Cubic needs 1.7 factor, rounded to 2 to include
303          * extra cushion (application might react slowly to POLLOUT)
304          */
305         sndmem = 2 * nr_segs * per_mss;
306 
307         if (sk->sk_sndbuf < sndmem)
308                 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
309 }
310 
311 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
312  *
313  * All tcp_full_space() is split to two parts: "network" buffer, allocated
314  * forward and advertised in receiver window (tp->rcv_wnd) and
315  * "application buffer", required to isolate scheduling/application
316  * latencies from network.
317  * window_clamp is maximal advertised window. It can be less than
318  * tcp_full_space(), in this case tcp_full_space() - window_clamp
319  * is reserved for "application" buffer. The less window_clamp is
320  * the smoother our behaviour from viewpoint of network, but the lower
321  * throughput and the higher sensitivity of the connection to losses. 8)
322  *
323  * rcv_ssthresh is more strict window_clamp used at "slow start"
324  * phase to predict further behaviour of this connection.
325  * It is used for two goals:
326  * - to enforce header prediction at sender, even when application
327  *   requires some significant "application buffer". It is check #1.
328  * - to prevent pruning of receive queue because of misprediction
329  *   of receiver window. Check #2.
330  *
331  * The scheme does not work when sender sends good segments opening
332  * window and then starts to feed us spaghetti. But it should work
333  * in common situations. Otherwise, we have to rely on queue collapsing.
334  */
335 
336 /* Slow part of check#2. */
337 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
338 {
339         struct tcp_sock *tp = tcp_sk(sk);
340         /* Optimize this! */
341         int truesize = tcp_win_from_space(skb->truesize) >> 1;
342         int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
343 
344         while (tp->rcv_ssthresh <= window) {
345                 if (truesize <= skb->len)
346                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
347 
348                 truesize >>= 1;
349                 window >>= 1;
350         }
351         return 0;
352 }
353 
354 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
355 {
356         struct tcp_sock *tp = tcp_sk(sk);
357 
358         /* Check #1 */
359         if (tp->rcv_ssthresh < tp->window_clamp &&
360             (int)tp->rcv_ssthresh < tcp_space(sk) &&
361             !sk_under_memory_pressure(sk)) {
362                 int incr;
363 
364                 /* Check #2. Increase window, if skb with such overhead
365                  * will fit to rcvbuf in future.
366                  */
367                 if (tcp_win_from_space(skb->truesize) <= skb->len)
368                         incr = 2 * tp->advmss;
369                 else
370                         incr = __tcp_grow_window(sk, skb);
371 
372                 if (incr) {
373                         incr = max_t(int, incr, 2 * skb->len);
374                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
375                                                tp->window_clamp);
376                         inet_csk(sk)->icsk_ack.quick |= 1;
377                 }
378         }
379 }
380 
381 /* 3. Tuning rcvbuf, when connection enters established state. */
382 static void tcp_fixup_rcvbuf(struct sock *sk)
383 {
384         u32 mss = tcp_sk(sk)->advmss;
385         int rcvmem;
386 
387         rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
388                  tcp_default_init_rwnd(mss);
389 
390         /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
391          * Allow enough cushion so that sender is not limited by our window
392          */
393         if (sysctl_tcp_moderate_rcvbuf)
394                 rcvmem <<= 2;
395 
396         if (sk->sk_rcvbuf < rcvmem)
397                 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
398 }
399 
400 /* 4. Try to fixup all. It is made immediately after connection enters
401  *    established state.
402  */
403 void tcp_init_buffer_space(struct sock *sk)
404 {
405         struct tcp_sock *tp = tcp_sk(sk);
406         int maxwin;
407 
408         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
409                 tcp_fixup_rcvbuf(sk);
410         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
411                 tcp_sndbuf_expand(sk);
412 
413         tp->rcvq_space.space = tp->rcv_wnd;
414         tp->rcvq_space.time = tcp_time_stamp;
415         tp->rcvq_space.seq = tp->copied_seq;
416 
417         maxwin = tcp_full_space(sk);
418 
419         if (tp->window_clamp >= maxwin) {
420                 tp->window_clamp = maxwin;
421 
422                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
423                         tp->window_clamp = max(maxwin -
424                                                (maxwin >> sysctl_tcp_app_win),
425                                                4 * tp->advmss);
426         }
427 
428         /* Force reservation of one segment. */
429         if (sysctl_tcp_app_win &&
430             tp->window_clamp > 2 * tp->advmss &&
431             tp->window_clamp + tp->advmss > maxwin)
432                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
433 
434         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
435         tp->snd_cwnd_stamp = tcp_time_stamp;
436 }
437 
438 /* 5. Recalculate window clamp after socket hit its memory bounds. */
439 static void tcp_clamp_window(struct sock *sk)
440 {
441         struct tcp_sock *tp = tcp_sk(sk);
442         struct inet_connection_sock *icsk = inet_csk(sk);
443 
444         icsk->icsk_ack.quick = 0;
445 
446         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
447             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
448             !sk_under_memory_pressure(sk) &&
449             sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
450                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
451                                     sysctl_tcp_rmem[2]);
452         }
453         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
454                 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
455 }
456 
457 /* Initialize RCV_MSS value.
458  * RCV_MSS is an our guess about MSS used by the peer.
459  * We haven't any direct information about the MSS.
460  * It's better to underestimate the RCV_MSS rather than overestimate.
461  * Overestimations make us ACKing less frequently than needed.
462  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
463  */
464 void tcp_initialize_rcv_mss(struct sock *sk)
465 {
466         const struct tcp_sock *tp = tcp_sk(sk);
467         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
468 
469         hint = min(hint, tp->rcv_wnd / 2);
470         hint = min(hint, TCP_MSS_DEFAULT);
471         hint = max(hint, TCP_MIN_MSS);
472 
473         inet_csk(sk)->icsk_ack.rcv_mss = hint;
474 }
475 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
476 
477 /* Receiver "autotuning" code.
478  *
479  * The algorithm for RTT estimation w/o timestamps is based on
480  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
481  * <http://public.lanl.gov/radiant/pubs.html#DRS>
482  *
483  * More detail on this code can be found at
484  * <http://staff.psc.edu/jheffner/>,
485  * though this reference is out of date.  A new paper
486  * is pending.
487  */
488 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
489 {
490         u32 new_sample = tp->rcv_rtt_est.rtt;
491         long m = sample;
492 
493         if (m == 0)
494                 m = 1;
495 
496         if (new_sample != 0) {
497                 /* If we sample in larger samples in the non-timestamp
498                  * case, we could grossly overestimate the RTT especially
499                  * with chatty applications or bulk transfer apps which
500                  * are stalled on filesystem I/O.
501                  *
502                  * Also, since we are only going for a minimum in the
503                  * non-timestamp case, we do not smooth things out
504                  * else with timestamps disabled convergence takes too
505                  * long.
506                  */
507                 if (!win_dep) {
508                         m -= (new_sample >> 3);
509                         new_sample += m;
510                 } else {
511                         m <<= 3;
512                         if (m < new_sample)
513                                 new_sample = m;
514                 }
515         } else {
516                 /* No previous measure. */
517                 new_sample = m << 3;
518         }
519 
520         if (tp->rcv_rtt_est.rtt != new_sample)
521                 tp->rcv_rtt_est.rtt = new_sample;
522 }
523 
524 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
525 {
526         if (tp->rcv_rtt_est.time == 0)
527                 goto new_measure;
528         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
529                 return;
530         tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
531 
532 new_measure:
533         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
534         tp->rcv_rtt_est.time = tcp_time_stamp;
535 }
536 
537 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
538                                           const struct sk_buff *skb)
539 {
540         struct tcp_sock *tp = tcp_sk(sk);
541         if (tp->rx_opt.rcv_tsecr &&
542             (TCP_SKB_CB(skb)->end_seq -
543              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
544                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
545 }
546 
547 /*
548  * This function should be called every time data is copied to user space.
549  * It calculates the appropriate TCP receive buffer space.
550  */
551 void tcp_rcv_space_adjust(struct sock *sk)
552 {
553         struct tcp_sock *tp = tcp_sk(sk);
554         int time;
555         int copied;
556 
557         time = tcp_time_stamp - tp->rcvq_space.time;
558         if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
559                 return;
560 
561         /* Number of bytes copied to user in last RTT */
562         copied = tp->copied_seq - tp->rcvq_space.seq;
563         if (copied <= tp->rcvq_space.space)
564                 goto new_measure;
565 
566         /* A bit of theory :
567          * copied = bytes received in previous RTT, our base window
568          * To cope with packet losses, we need a 2x factor
569          * To cope with slow start, and sender growing its cwin by 100 %
570          * every RTT, we need a 4x factor, because the ACK we are sending
571          * now is for the next RTT, not the current one :
572          * <prev RTT . ><current RTT .. ><next RTT .... >
573          */
574 
575         if (sysctl_tcp_moderate_rcvbuf &&
576             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
577                 int rcvwin, rcvmem, rcvbuf;
578 
579                 /* minimal window to cope with packet losses, assuming
580                  * steady state. Add some cushion because of small variations.
581                  */
582                 rcvwin = (copied << 1) + 16 * tp->advmss;
583 
584                 /* If rate increased by 25%,
585                  *      assume slow start, rcvwin = 3 * copied
586                  * If rate increased by 50%,
587                  *      assume sender can use 2x growth, rcvwin = 4 * copied
588                  */
589                 if (copied >=
590                     tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
591                         if (copied >=
592                             tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
593                                 rcvwin <<= 1;
594                         else
595                                 rcvwin += (rcvwin >> 1);
596                 }
597 
598                 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
599                 while (tcp_win_from_space(rcvmem) < tp->advmss)
600                         rcvmem += 128;
601 
602                 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
603                 if (rcvbuf > sk->sk_rcvbuf) {
604                         sk->sk_rcvbuf = rcvbuf;
605 
606                         /* Make the window clamp follow along.  */
607                         tp->window_clamp = rcvwin;
608                 }
609         }
610         tp->rcvq_space.space = copied;
611 
612 new_measure:
613         tp->rcvq_space.seq = tp->copied_seq;
614         tp->rcvq_space.time = tcp_time_stamp;
615 }
616 
617 /* There is something which you must keep in mind when you analyze the
618  * behavior of the tp->ato delayed ack timeout interval.  When a
619  * connection starts up, we want to ack as quickly as possible.  The
620  * problem is that "good" TCP's do slow start at the beginning of data
621  * transmission.  The means that until we send the first few ACK's the
622  * sender will sit on his end and only queue most of his data, because
623  * he can only send snd_cwnd unacked packets at any given time.  For
624  * each ACK we send, he increments snd_cwnd and transmits more of his
625  * queue.  -DaveM
626  */
627 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
628 {
629         struct tcp_sock *tp = tcp_sk(sk);
630         struct inet_connection_sock *icsk = inet_csk(sk);
631         u32 now;
632 
633         inet_csk_schedule_ack(sk);
634 
635         tcp_measure_rcv_mss(sk, skb);
636 
637         tcp_rcv_rtt_measure(tp);
638 
639         now = tcp_time_stamp;
640 
641         if (!icsk->icsk_ack.ato) {
642                 /* The _first_ data packet received, initialize
643                  * delayed ACK engine.
644                  */
645                 tcp_incr_quickack(sk);
646                 icsk->icsk_ack.ato = TCP_ATO_MIN;
647         } else {
648                 int m = now - icsk->icsk_ack.lrcvtime;
649 
650                 if (m <= TCP_ATO_MIN / 2) {
651                         /* The fastest case is the first. */
652                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
653                 } else if (m < icsk->icsk_ack.ato) {
654                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
655                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
656                                 icsk->icsk_ack.ato = icsk->icsk_rto;
657                 } else if (m > icsk->icsk_rto) {
658                         /* Too long gap. Apparently sender failed to
659                          * restart window, so that we send ACKs quickly.
660                          */
661                         tcp_incr_quickack(sk);
662                         sk_mem_reclaim(sk);
663                 }
664         }
665         icsk->icsk_ack.lrcvtime = now;
666 
667         tcp_ecn_check_ce(tp, skb);
668 
669         if (skb->len >= 128)
670                 tcp_grow_window(sk, skb);
671 }
672 
673 /* Called to compute a smoothed rtt estimate. The data fed to this
674  * routine either comes from timestamps, or from segments that were
675  * known _not_ to have been retransmitted [see Karn/Partridge
676  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
677  * piece by Van Jacobson.
678  * NOTE: the next three routines used to be one big routine.
679  * To save cycles in the RFC 1323 implementation it was better to break
680  * it up into three procedures. -- erics
681  */
682 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
683 {
684         struct tcp_sock *tp = tcp_sk(sk);
685         long m = mrtt_us; /* RTT */
686         u32 srtt = tp->srtt_us;
687 
688         /*      The following amusing code comes from Jacobson's
689          *      article in SIGCOMM '88.  Note that rtt and mdev
690          *      are scaled versions of rtt and mean deviation.
691          *      This is designed to be as fast as possible
692          *      m stands for "measurement".
693          *
694          *      On a 1990 paper the rto value is changed to:
695          *      RTO = rtt + 4 * mdev
696          *
697          * Funny. This algorithm seems to be very broken.
698          * These formulae increase RTO, when it should be decreased, increase
699          * too slowly, when it should be increased quickly, decrease too quickly
700          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
701          * does not matter how to _calculate_ it. Seems, it was trap
702          * that VJ failed to avoid. 8)
703          */
704         if (srtt != 0) {
705                 m -= (srtt >> 3);       /* m is now error in rtt est */
706                 srtt += m;              /* rtt = 7/8 rtt + 1/8 new */
707                 if (m < 0) {
708                         m = -m;         /* m is now abs(error) */
709                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
710                         /* This is similar to one of Eifel findings.
711                          * Eifel blocks mdev updates when rtt decreases.
712                          * This solution is a bit different: we use finer gain
713                          * for mdev in this case (alpha*beta).
714                          * Like Eifel it also prevents growth of rto,
715                          * but also it limits too fast rto decreases,
716                          * happening in pure Eifel.
717                          */
718                         if (m > 0)
719                                 m >>= 3;
720                 } else {
721                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
722                 }
723                 tp->mdev_us += m;               /* mdev = 3/4 mdev + 1/4 new */
724                 if (tp->mdev_us > tp->mdev_max_us) {
725                         tp->mdev_max_us = tp->mdev_us;
726                         if (tp->mdev_max_us > tp->rttvar_us)
727                                 tp->rttvar_us = tp->mdev_max_us;
728                 }
729                 if (after(tp->snd_una, tp->rtt_seq)) {
730                         if (tp->mdev_max_us < tp->rttvar_us)
731                                 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
732                         tp->rtt_seq = tp->snd_nxt;
733                         tp->mdev_max_us = tcp_rto_min_us(sk);
734                 }
735         } else {
736                 /* no previous measure. */
737                 srtt = m << 3;          /* take the measured time to be rtt */
738                 tp->mdev_us = m << 1;   /* make sure rto = 3*rtt */
739                 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
740                 tp->mdev_max_us = tp->rttvar_us;
741                 tp->rtt_seq = tp->snd_nxt;
742         }
743         tp->srtt_us = max(1U, srtt);
744 }
745 
746 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
747  * Note: TCP stack does not yet implement pacing.
748  * FQ packet scheduler can be used to implement cheap but effective
749  * TCP pacing, to smooth the burst on large writes when packets
750  * in flight is significantly lower than cwnd (or rwin)
751  */
752 static void tcp_update_pacing_rate(struct sock *sk)
753 {
754         const struct tcp_sock *tp = tcp_sk(sk);
755         u64 rate;
756 
757         /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
758         rate = (u64)tp->mss_cache * 2 * (USEC_PER_SEC << 3);
759 
760         rate *= max(tp->snd_cwnd, tp->packets_out);
761 
762         if (likely(tp->srtt_us))
763                 do_div(rate, tp->srtt_us);
764 
765         /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
766          * without any lock. We want to make sure compiler wont store
767          * intermediate values in this location.
768          */
769         ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
770                                                 sk->sk_max_pacing_rate);
771 }
772 
773 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
774  * routine referred to above.
775  */
776 static void tcp_set_rto(struct sock *sk)
777 {
778         const struct tcp_sock *tp = tcp_sk(sk);
779         /* Old crap is replaced with new one. 8)
780          *
781          * More seriously:
782          * 1. If rtt variance happened to be less 50msec, it is hallucination.
783          *    It cannot be less due to utterly erratic ACK generation made
784          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
785          *    to do with delayed acks, because at cwnd>2 true delack timeout
786          *    is invisible. Actually, Linux-2.4 also generates erratic
787          *    ACKs in some circumstances.
788          */
789         inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
790 
791         /* 2. Fixups made earlier cannot be right.
792          *    If we do not estimate RTO correctly without them,
793          *    all the algo is pure shit and should be replaced
794          *    with correct one. It is exactly, which we pretend to do.
795          */
796 
797         /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
798          * guarantees that rto is higher.
799          */
800         tcp_bound_rto(sk);
801 }
802 
803 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
804 {
805         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
806 
807         if (!cwnd)
808                 cwnd = TCP_INIT_CWND;
809         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
810 }
811 
812 /*
813  * Packet counting of FACK is based on in-order assumptions, therefore TCP
814  * disables it when reordering is detected
815  */
816 void tcp_disable_fack(struct tcp_sock *tp)
817 {
818         /* RFC3517 uses different metric in lost marker => reset on change */
819         if (tcp_is_fack(tp))
820                 tp->lost_skb_hint = NULL;
821         tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
822 }
823 
824 /* Take a notice that peer is sending D-SACKs */
825 static void tcp_dsack_seen(struct tcp_sock *tp)
826 {
827         tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
828 }
829 
830 static void tcp_update_reordering(struct sock *sk, const int metric,
831                                   const int ts)
832 {
833         struct tcp_sock *tp = tcp_sk(sk);
834         if (metric > tp->reordering) {
835                 int mib_idx;
836 
837                 tp->reordering = min(sysctl_tcp_max_reordering, metric);
838 
839                 /* This exciting event is worth to be remembered. 8) */
840                 if (ts)
841                         mib_idx = LINUX_MIB_TCPTSREORDER;
842                 else if (tcp_is_reno(tp))
843                         mib_idx = LINUX_MIB_TCPRENOREORDER;
844                 else if (tcp_is_fack(tp))
845                         mib_idx = LINUX_MIB_TCPFACKREORDER;
846                 else
847                         mib_idx = LINUX_MIB_TCPSACKREORDER;
848 
849                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
850 #if FASTRETRANS_DEBUG > 1
851                 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
852                          tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
853                          tp->reordering,
854                          tp->fackets_out,
855                          tp->sacked_out,
856                          tp->undo_marker ? tp->undo_retrans : 0);
857 #endif
858                 tcp_disable_fack(tp);
859         }
860 
861         if (metric > 0)
862                 tcp_disable_early_retrans(tp);
863 }
864 
865 /* This must be called before lost_out is incremented */
866 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
867 {
868         if ((tp->retransmit_skb_hint == NULL) ||
869             before(TCP_SKB_CB(skb)->seq,
870                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
871                 tp->retransmit_skb_hint = skb;
872 
873         if (!tp->lost_out ||
874             after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
875                 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
876 }
877 
878 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
879 {
880         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
881                 tcp_verify_retransmit_hint(tp, skb);
882 
883                 tp->lost_out += tcp_skb_pcount(skb);
884                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
885         }
886 }
887 
888 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
889                                             struct sk_buff *skb)
890 {
891         tcp_verify_retransmit_hint(tp, skb);
892 
893         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
894                 tp->lost_out += tcp_skb_pcount(skb);
895                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
896         }
897 }
898 
899 /* This procedure tags the retransmission queue when SACKs arrive.
900  *
901  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
902  * Packets in queue with these bits set are counted in variables
903  * sacked_out, retrans_out and lost_out, correspondingly.
904  *
905  * Valid combinations are:
906  * Tag  InFlight        Description
907  * 0    1               - orig segment is in flight.
908  * S    0               - nothing flies, orig reached receiver.
909  * L    0               - nothing flies, orig lost by net.
910  * R    2               - both orig and retransmit are in flight.
911  * L|R  1               - orig is lost, retransmit is in flight.
912  * S|R  1               - orig reached receiver, retrans is still in flight.
913  * (L|S|R is logically valid, it could occur when L|R is sacked,
914  *  but it is equivalent to plain S and code short-curcuits it to S.
915  *  L|S is logically invalid, it would mean -1 packet in flight 8))
916  *
917  * These 6 states form finite state machine, controlled by the following events:
918  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
919  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
920  * 3. Loss detection event of two flavors:
921  *      A. Scoreboard estimator decided the packet is lost.
922  *         A'. Reno "three dupacks" marks head of queue lost.
923  *         A''. Its FACK modification, head until snd.fack is lost.
924  *      B. SACK arrives sacking SND.NXT at the moment, when the
925  *         segment was retransmitted.
926  * 4. D-SACK added new rule: D-SACK changes any tag to S.
927  *
928  * It is pleasant to note, that state diagram turns out to be commutative,
929  * so that we are allowed not to be bothered by order of our actions,
930  * when multiple events arrive simultaneously. (see the function below).
931  *
932  * Reordering detection.
933  * --------------------
934  * Reordering metric is maximal distance, which a packet can be displaced
935  * in packet stream. With SACKs we can estimate it:
936  *
937  * 1. SACK fills old hole and the corresponding segment was not
938  *    ever retransmitted -> reordering. Alas, we cannot use it
939  *    when segment was retransmitted.
940  * 2. The last flaw is solved with D-SACK. D-SACK arrives
941  *    for retransmitted and already SACKed segment -> reordering..
942  * Both of these heuristics are not used in Loss state, when we cannot
943  * account for retransmits accurately.
944  *
945  * SACK block validation.
946  * ----------------------
947  *
948  * SACK block range validation checks that the received SACK block fits to
949  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
950  * Note that SND.UNA is not included to the range though being valid because
951  * it means that the receiver is rather inconsistent with itself reporting
952  * SACK reneging when it should advance SND.UNA. Such SACK block this is
953  * perfectly valid, however, in light of RFC2018 which explicitly states
954  * that "SACK block MUST reflect the newest segment.  Even if the newest
955  * segment is going to be discarded ...", not that it looks very clever
956  * in case of head skb. Due to potentional receiver driven attacks, we
957  * choose to avoid immediate execution of a walk in write queue due to
958  * reneging and defer head skb's loss recovery to standard loss recovery
959  * procedure that will eventually trigger (nothing forbids us doing this).
960  *
961  * Implements also blockage to start_seq wrap-around. Problem lies in the
962  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
963  * there's no guarantee that it will be before snd_nxt (n). The problem
964  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
965  * wrap (s_w):
966  *
967  *         <- outs wnd ->                          <- wrapzone ->
968  *         u     e      n                         u_w   e_w  s n_w
969  *         |     |      |                          |     |   |  |
970  * |<------------+------+----- TCP seqno space --------------+---------->|
971  * ...-- <2^31 ->|                                           |<--------...
972  * ...---- >2^31 ------>|                                    |<--------...
973  *
974  * Current code wouldn't be vulnerable but it's better still to discard such
975  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
976  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
977  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
978  * equal to the ideal case (infinite seqno space without wrap caused issues).
979  *
980  * With D-SACK the lower bound is extended to cover sequence space below
981  * SND.UNA down to undo_marker, which is the last point of interest. Yet
982  * again, D-SACK block must not to go across snd_una (for the same reason as
983  * for the normal SACK blocks, explained above). But there all simplicity
984  * ends, TCP might receive valid D-SACKs below that. As long as they reside
985  * fully below undo_marker they do not affect behavior in anyway and can
986  * therefore be safely ignored. In rare cases (which are more or less
987  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
988  * fragmentation and packet reordering past skb's retransmission. To consider
989  * them correctly, the acceptable range must be extended even more though
990  * the exact amount is rather hard to quantify. However, tp->max_window can
991  * be used as an exaggerated estimate.
992  */
993 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
994                                    u32 start_seq, u32 end_seq)
995 {
996         /* Too far in future, or reversed (interpretation is ambiguous) */
997         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
998                 return false;
999 
1000         /* Nasty start_seq wrap-around check (see comments above) */
1001         if (!before(start_seq, tp->snd_nxt))
1002                 return false;
1003 
1004         /* In outstanding window? ...This is valid exit for D-SACKs too.
1005          * start_seq == snd_una is non-sensical (see comments above)
1006          */
1007         if (after(start_seq, tp->snd_una))
1008                 return true;
1009 
1010         if (!is_dsack || !tp->undo_marker)
1011                 return false;
1012 
1013         /* ...Then it's D-SACK, and must reside below snd_una completely */
1014         if (after(end_seq, tp->snd_una))
1015                 return false;
1016 
1017         if (!before(start_seq, tp->undo_marker))
1018                 return true;
1019 
1020         /* Too old */
1021         if (!after(end_seq, tp->undo_marker))
1022                 return false;
1023 
1024         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1025          *   start_seq < undo_marker and end_seq >= undo_marker.
1026          */
1027         return !before(start_seq, end_seq - tp->max_window);
1028 }
1029 
1030 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1031  * Event "B". Later note: FACK people cheated me again 8), we have to account
1032  * for reordering! Ugly, but should help.
1033  *
1034  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1035  * less than what is now known to be received by the other end (derived from
1036  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1037  * retransmitted skbs to avoid some costly processing per ACKs.
1038  */
1039 static void tcp_mark_lost_retrans(struct sock *sk)
1040 {
1041         const struct inet_connection_sock *icsk = inet_csk(sk);
1042         struct tcp_sock *tp = tcp_sk(sk);
1043         struct sk_buff *skb;
1044         int cnt = 0;
1045         u32 new_low_seq = tp->snd_nxt;
1046         u32 received_upto = tcp_highest_sack_seq(tp);
1047 
1048         if (!tcp_is_fack(tp) || !tp->retrans_out ||
1049             !after(received_upto, tp->lost_retrans_low) ||
1050             icsk->icsk_ca_state != TCP_CA_Recovery)
1051                 return;
1052 
1053         tcp_for_write_queue(skb, sk) {
1054                 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1055 
1056                 if (skb == tcp_send_head(sk))
1057                         break;
1058                 if (cnt == tp->retrans_out)
1059                         break;
1060                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1061                         continue;
1062 
1063                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1064                         continue;
1065 
1066                 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1067                  * constraint here (see above) but figuring out that at
1068                  * least tp->reordering SACK blocks reside between ack_seq
1069                  * and received_upto is not easy task to do cheaply with
1070                  * the available datastructures.
1071                  *
1072                  * Whether FACK should check here for tp->reordering segs
1073                  * in-between one could argue for either way (it would be
1074                  * rather simple to implement as we could count fack_count
1075                  * during the walk and do tp->fackets_out - fack_count).
1076                  */
1077                 if (after(received_upto, ack_seq)) {
1078                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1079                         tp->retrans_out -= tcp_skb_pcount(skb);
1080 
1081                         tcp_skb_mark_lost_uncond_verify(tp, skb);
1082                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1083                 } else {
1084                         if (before(ack_seq, new_low_seq))
1085                                 new_low_seq = ack_seq;
1086                         cnt += tcp_skb_pcount(skb);
1087                 }
1088         }
1089 
1090         if (tp->retrans_out)
1091                 tp->lost_retrans_low = new_low_seq;
1092 }
1093 
1094 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1095                             struct tcp_sack_block_wire *sp, int num_sacks,
1096                             u32 prior_snd_una)
1097 {
1098         struct tcp_sock *tp = tcp_sk(sk);
1099         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1100         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1101         bool dup_sack = false;
1102 
1103         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1104                 dup_sack = true;
1105                 tcp_dsack_seen(tp);
1106                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1107         } else if (num_sacks > 1) {
1108                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1109                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1110 
1111                 if (!after(end_seq_0, end_seq_1) &&
1112                     !before(start_seq_0, start_seq_1)) {
1113                         dup_sack = true;
1114                         tcp_dsack_seen(tp);
1115                         NET_INC_STATS_BH(sock_net(sk),
1116                                         LINUX_MIB_TCPDSACKOFORECV);
1117                 }
1118         }
1119 
1120         /* D-SACK for already forgotten data... Do dumb counting. */
1121         if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1122             !after(end_seq_0, prior_snd_una) &&
1123             after(end_seq_0, tp->undo_marker))
1124                 tp->undo_retrans--;
1125 
1126         return dup_sack;
1127 }
1128 
1129 struct tcp_sacktag_state {
1130         int     reord;
1131         int     fack_count;
1132         long    rtt_us; /* RTT measured by SACKing never-retransmitted data */
1133         int     flag;
1134 };
1135 
1136 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1137  * the incoming SACK may not exactly match but we can find smaller MSS
1138  * aligned portion of it that matches. Therefore we might need to fragment
1139  * which may fail and creates some hassle (caller must handle error case
1140  * returns).
1141  *
1142  * FIXME: this could be merged to shift decision code
1143  */
1144 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1145                                   u32 start_seq, u32 end_seq)
1146 {
1147         int err;
1148         bool in_sack;
1149         unsigned int pkt_len;
1150         unsigned int mss;
1151 
1152         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1153                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1154 
1155         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1156             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1157                 mss = tcp_skb_mss(skb);
1158                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1159 
1160                 if (!in_sack) {
1161                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1162                         if (pkt_len < mss)
1163                                 pkt_len = mss;
1164                 } else {
1165                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1166                         if (pkt_len < mss)
1167                                 return -EINVAL;
1168                 }
1169 
1170                 /* Round if necessary so that SACKs cover only full MSSes
1171                  * and/or the remaining small portion (if present)
1172                  */
1173                 if (pkt_len > mss) {
1174                         unsigned int new_len = (pkt_len / mss) * mss;
1175                         if (!in_sack && new_len < pkt_len) {
1176                                 new_len += mss;
1177                                 if (new_len >= skb->len)
1178                                         return 0;
1179                         }
1180                         pkt_len = new_len;
1181                 }
1182                 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1183                 if (err < 0)
1184                         return err;
1185         }
1186 
1187         return in_sack;
1188 }
1189 
1190 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1191 static u8 tcp_sacktag_one(struct sock *sk,
1192                           struct tcp_sacktag_state *state, u8 sacked,
1193                           u32 start_seq, u32 end_seq,
1194                           int dup_sack, int pcount,
1195                           const struct skb_mstamp *xmit_time)
1196 {
1197         struct tcp_sock *tp = tcp_sk(sk);
1198         int fack_count = state->fack_count;
1199 
1200         /* Account D-SACK for retransmitted packet. */
1201         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1202                 if (tp->undo_marker && tp->undo_retrans > 0 &&
1203                     after(end_seq, tp->undo_marker))
1204                         tp->undo_retrans--;
1205                 if (sacked & TCPCB_SACKED_ACKED)
1206                         state->reord = min(fack_count, state->reord);
1207         }
1208 
1209         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1210         if (!after(end_seq, tp->snd_una))
1211                 return sacked;
1212 
1213         if (!(sacked & TCPCB_SACKED_ACKED)) {
1214                 if (sacked & TCPCB_SACKED_RETRANS) {
1215                         /* If the segment is not tagged as lost,
1216                          * we do not clear RETRANS, believing
1217                          * that retransmission is still in flight.
1218                          */
1219                         if (sacked & TCPCB_LOST) {
1220                                 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1221                                 tp->lost_out -= pcount;
1222                                 tp->retrans_out -= pcount;
1223                         }
1224                 } else {
1225                         if (!(sacked & TCPCB_RETRANS)) {
1226                                 /* New sack for not retransmitted frame,
1227                                  * which was in hole. It is reordering.
1228                                  */
1229                                 if (before(start_seq,
1230                                            tcp_highest_sack_seq(tp)))
1231                                         state->reord = min(fack_count,
1232                                                            state->reord);
1233                                 if (!after(end_seq, tp->high_seq))
1234                                         state->flag |= FLAG_ORIG_SACK_ACKED;
1235                                 /* Pick the earliest sequence sacked for RTT */
1236                                 if (state->rtt_us < 0) {
1237                                         struct skb_mstamp now;
1238 
1239                                         skb_mstamp_get(&now);
1240                                         state->rtt_us = skb_mstamp_us_delta(&now,
1241                                                                 xmit_time);
1242                                 }
1243                         }
1244 
1245                         if (sacked & TCPCB_LOST) {
1246                                 sacked &= ~TCPCB_LOST;
1247                                 tp->lost_out -= pcount;
1248                         }
1249                 }
1250 
1251                 sacked |= TCPCB_SACKED_ACKED;
1252                 state->flag |= FLAG_DATA_SACKED;
1253                 tp->sacked_out += pcount;
1254 
1255                 fack_count += pcount;
1256 
1257                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1258                 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1259                     before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1260                         tp->lost_cnt_hint += pcount;
1261 
1262                 if (fack_count > tp->fackets_out)
1263                         tp->fackets_out = fack_count;
1264         }
1265 
1266         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1267          * frames and clear it. undo_retrans is decreased above, L|R frames
1268          * are accounted above as well.
1269          */
1270         if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1271                 sacked &= ~TCPCB_SACKED_RETRANS;
1272                 tp->retrans_out -= pcount;
1273         }
1274 
1275         return sacked;
1276 }
1277 
1278 /* Shift newly-SACKed bytes from this skb to the immediately previous
1279  * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1280  */
1281 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1282                             struct tcp_sacktag_state *state,
1283                             unsigned int pcount, int shifted, int mss,
1284                             bool dup_sack)
1285 {
1286         struct tcp_sock *tp = tcp_sk(sk);
1287         struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1288         u32 start_seq = TCP_SKB_CB(skb)->seq;   /* start of newly-SACKed */
1289         u32 end_seq = start_seq + shifted;      /* end of newly-SACKed */
1290 
1291         BUG_ON(!pcount);
1292 
1293         /* Adjust counters and hints for the newly sacked sequence
1294          * range but discard the return value since prev is already
1295          * marked. We must tag the range first because the seq
1296          * advancement below implicitly advances
1297          * tcp_highest_sack_seq() when skb is highest_sack.
1298          */
1299         tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1300                         start_seq, end_seq, dup_sack, pcount,
1301                         &skb->skb_mstamp);
1302 
1303         if (skb == tp->lost_skb_hint)
1304                 tp->lost_cnt_hint += pcount;
1305 
1306         TCP_SKB_CB(prev)->end_seq += shifted;
1307         TCP_SKB_CB(skb)->seq += shifted;
1308 
1309         tcp_skb_pcount_add(prev, pcount);
1310         BUG_ON(tcp_skb_pcount(skb) < pcount);
1311         tcp_skb_pcount_add(skb, -pcount);
1312 
1313         /* When we're adding to gso_segs == 1, gso_size will be zero,
1314          * in theory this shouldn't be necessary but as long as DSACK
1315          * code can come after this skb later on it's better to keep
1316          * setting gso_size to something.
1317          */
1318         if (!skb_shinfo(prev)->gso_size) {
1319                 skb_shinfo(prev)->gso_size = mss;
1320                 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1321         }
1322 
1323         /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1324         if (tcp_skb_pcount(skb) <= 1) {
1325                 skb_shinfo(skb)->gso_size = 0;
1326                 skb_shinfo(skb)->gso_type = 0;
1327         }
1328 
1329         /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1330         TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1331 
1332         if (skb->len > 0) {
1333                 BUG_ON(!tcp_skb_pcount(skb));
1334                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1335                 return false;
1336         }
1337 
1338         /* Whole SKB was eaten :-) */
1339 
1340         if (skb == tp->retransmit_skb_hint)
1341                 tp->retransmit_skb_hint = prev;
1342         if (skb == tp->lost_skb_hint) {
1343                 tp->lost_skb_hint = prev;
1344                 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1345         }
1346 
1347         TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1348         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1349                 TCP_SKB_CB(prev)->end_seq++;
1350 
1351         if (skb == tcp_highest_sack(sk))
1352                 tcp_advance_highest_sack(sk, skb);
1353 
1354         tcp_unlink_write_queue(skb, sk);
1355         sk_wmem_free_skb(sk, skb);
1356 
1357         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1358 
1359         return true;
1360 }
1361 
1362 /* I wish gso_size would have a bit more sane initialization than
1363  * something-or-zero which complicates things
1364  */
1365 static int tcp_skb_seglen(const struct sk_buff *skb)
1366 {
1367         return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1368 }
1369 
1370 /* Shifting pages past head area doesn't work */
1371 static int skb_can_shift(const struct sk_buff *skb)
1372 {
1373         return !skb_headlen(skb) && skb_is_nonlinear(skb);
1374 }
1375 
1376 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1377  * skb.
1378  */
1379 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1380                                           struct tcp_sacktag_state *state,
1381                                           u32 start_seq, u32 end_seq,
1382                                           bool dup_sack)
1383 {
1384         struct tcp_sock *tp = tcp_sk(sk);
1385         struct sk_buff *prev;
1386         int mss;
1387         int pcount = 0;
1388         int len;
1389         int in_sack;
1390 
1391         if (!sk_can_gso(sk))
1392                 goto fallback;
1393 
1394         /* Normally R but no L won't result in plain S */
1395         if (!dup_sack &&
1396             (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1397                 goto fallback;
1398         if (!skb_can_shift(skb))
1399                 goto fallback;
1400         /* This frame is about to be dropped (was ACKed). */
1401         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1402                 goto fallback;
1403 
1404         /* Can only happen with delayed DSACK + discard craziness */
1405         if (unlikely(skb == tcp_write_queue_head(sk)))
1406                 goto fallback;
1407         prev = tcp_write_queue_prev(sk, skb);
1408 
1409         if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1410                 goto fallback;
1411 
1412         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1413                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1414 
1415         if (in_sack) {
1416                 len = skb->len;
1417                 pcount = tcp_skb_pcount(skb);
1418                 mss = tcp_skb_seglen(skb);
1419 
1420                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1421                  * drop this restriction as unnecessary
1422                  */
1423                 if (mss != tcp_skb_seglen(prev))
1424                         goto fallback;
1425         } else {
1426                 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1427                         goto noop;
1428                 /* CHECKME: This is non-MSS split case only?, this will
1429                  * cause skipped skbs due to advancing loop btw, original
1430                  * has that feature too
1431                  */
1432                 if (tcp_skb_pcount(skb) <= 1)
1433                         goto noop;
1434 
1435                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1436                 if (!in_sack) {
1437                         /* TODO: head merge to next could be attempted here
1438                          * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1439                          * though it might not be worth of the additional hassle
1440                          *
1441                          * ...we can probably just fallback to what was done
1442                          * previously. We could try merging non-SACKed ones
1443                          * as well but it probably isn't going to buy off
1444                          * because later SACKs might again split them, and
1445                          * it would make skb timestamp tracking considerably
1446                          * harder problem.
1447                          */
1448                         goto fallback;
1449                 }
1450 
1451                 len = end_seq - TCP_SKB_CB(skb)->seq;
1452                 BUG_ON(len < 0);
1453                 BUG_ON(len > skb->len);
1454 
1455                 /* MSS boundaries should be honoured or else pcount will
1456                  * severely break even though it makes things bit trickier.
1457                  * Optimize common case to avoid most of the divides
1458                  */
1459                 mss = tcp_skb_mss(skb);
1460 
1461                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1462                  * drop this restriction as unnecessary
1463                  */
1464                 if (mss != tcp_skb_seglen(prev))
1465                         goto fallback;
1466 
1467                 if (len == mss) {
1468                         pcount = 1;
1469                 } else if (len < mss) {
1470                         goto noop;
1471                 } else {
1472                         pcount = len / mss;
1473                         len = pcount * mss;
1474                 }
1475         }
1476 
1477         /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1478         if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1479                 goto fallback;
1480 
1481         if (!skb_shift(prev, skb, len))
1482                 goto fallback;
1483         if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1484                 goto out;
1485 
1486         /* Hole filled allows collapsing with the next as well, this is very
1487          * useful when hole on every nth skb pattern happens
1488          */
1489         if (prev == tcp_write_queue_tail(sk))
1490                 goto out;
1491         skb = tcp_write_queue_next(sk, prev);
1492 
1493         if (!skb_can_shift(skb) ||
1494             (skb == tcp_send_head(sk)) ||
1495             ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1496             (mss != tcp_skb_seglen(skb)))
1497                 goto out;
1498 
1499         len = skb->len;
1500         if (skb_shift(prev, skb, len)) {
1501                 pcount += tcp_skb_pcount(skb);
1502                 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1503         }
1504 
1505 out:
1506         state->fack_count += pcount;
1507         return prev;
1508 
1509 noop:
1510         return skb;
1511 
1512 fallback:
1513         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1514         return NULL;
1515 }
1516 
1517 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1518                                         struct tcp_sack_block *next_dup,
1519                                         struct tcp_sacktag_state *state,
1520                                         u32 start_seq, u32 end_seq,
1521                                         bool dup_sack_in)
1522 {
1523         struct tcp_sock *tp = tcp_sk(sk);
1524         struct sk_buff *tmp;
1525 
1526         tcp_for_write_queue_from(skb, sk) {
1527                 int in_sack = 0;
1528                 bool dup_sack = dup_sack_in;
1529 
1530                 if (skb == tcp_send_head(sk))
1531                         break;
1532 
1533                 /* queue is in-order => we can short-circuit the walk early */
1534                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1535                         break;
1536 
1537                 if ((next_dup != NULL) &&
1538                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1539                         in_sack = tcp_match_skb_to_sack(sk, skb,
1540                                                         next_dup->start_seq,
1541                                                         next_dup->end_seq);
1542                         if (in_sack > 0)
1543                                 dup_sack = true;
1544                 }
1545 
1546                 /* skb reference here is a bit tricky to get right, since
1547                  * shifting can eat and free both this skb and the next,
1548                  * so not even _safe variant of the loop is enough.
1549                  */
1550                 if (in_sack <= 0) {
1551                         tmp = tcp_shift_skb_data(sk, skb, state,
1552                                                  start_seq, end_seq, dup_sack);
1553                         if (tmp != NULL) {
1554                                 if (tmp != skb) {
1555                                         skb = tmp;
1556                                         continue;
1557                                 }
1558 
1559                                 in_sack = 0;
1560                         } else {
1561                                 in_sack = tcp_match_skb_to_sack(sk, skb,
1562                                                                 start_seq,
1563                                                                 end_seq);
1564                         }
1565                 }
1566 
1567                 if (unlikely(in_sack < 0))
1568                         break;
1569 
1570                 if (in_sack) {
1571                         TCP_SKB_CB(skb)->sacked =
1572                                 tcp_sacktag_one(sk,
1573                                                 state,
1574                                                 TCP_SKB_CB(skb)->sacked,
1575                                                 TCP_SKB_CB(skb)->seq,
1576                                                 TCP_SKB_CB(skb)->end_seq,
1577                                                 dup_sack,
1578                                                 tcp_skb_pcount(skb),
1579                                                 &skb->skb_mstamp);
1580 
1581                         if (!before(TCP_SKB_CB(skb)->seq,
1582                                     tcp_highest_sack_seq(tp)))
1583                                 tcp_advance_highest_sack(sk, skb);
1584                 }
1585 
1586                 state->fack_count += tcp_skb_pcount(skb);
1587         }
1588         return skb;
1589 }
1590 
1591 /* Avoid all extra work that is being done by sacktag while walking in
1592  * a normal way
1593  */
1594 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1595                                         struct tcp_sacktag_state *state,
1596                                         u32 skip_to_seq)
1597 {
1598         tcp_for_write_queue_from(skb, sk) {
1599                 if (skb == tcp_send_head(sk))
1600                         break;
1601 
1602                 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1603                         break;
1604 
1605                 state->fack_count += tcp_skb_pcount(skb);
1606         }
1607         return skb;
1608 }
1609 
1610 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1611                                                 struct sock *sk,
1612                                                 struct tcp_sack_block *next_dup,
1613                                                 struct tcp_sacktag_state *state,
1614                                                 u32 skip_to_seq)
1615 {
1616         if (next_dup == NULL)
1617                 return skb;
1618 
1619         if (before(next_dup->start_seq, skip_to_seq)) {
1620                 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1621                 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1622                                        next_dup->start_seq, next_dup->end_seq,
1623                                        1);
1624         }
1625 
1626         return skb;
1627 }
1628 
1629 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1630 {
1631         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1632 }
1633 
1634 static int
1635 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1636                         u32 prior_snd_una, long *sack_rtt_us)
1637 {
1638         struct tcp_sock *tp = tcp_sk(sk);
1639         const unsigned char *ptr = (skb_transport_header(ack_skb) +
1640                                     TCP_SKB_CB(ack_skb)->sacked);
1641         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1642         struct tcp_sack_block sp[TCP_NUM_SACKS];
1643         struct tcp_sack_block *cache;
1644         struct tcp_sacktag_state state;
1645         struct sk_buff *skb;
1646         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1647         int used_sacks;
1648         bool found_dup_sack = false;
1649         int i, j;
1650         int first_sack_index;
1651 
1652         state.flag = 0;
1653         state.reord = tp->packets_out;
1654         state.rtt_us = -1L;
1655 
1656         if (!tp->sacked_out) {
1657                 if (WARN_ON(tp->fackets_out))
1658                         tp->fackets_out = 0;
1659                 tcp_highest_sack_reset(sk);
1660         }
1661 
1662         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1663                                          num_sacks, prior_snd_una);
1664         if (found_dup_sack)
1665                 state.flag |= FLAG_DSACKING_ACK;
1666 
1667         /* Eliminate too old ACKs, but take into
1668          * account more or less fresh ones, they can
1669          * contain valid SACK info.
1670          */
1671         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1672                 return 0;
1673 
1674         if (!tp->packets_out)
1675                 goto out;
1676 
1677         used_sacks = 0;
1678         first_sack_index = 0;
1679         for (i = 0; i < num_sacks; i++) {
1680                 bool dup_sack = !i && found_dup_sack;
1681 
1682                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1683                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1684 
1685                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1686                                             sp[used_sacks].start_seq,
1687                                             sp[used_sacks].end_seq)) {
1688                         int mib_idx;
1689 
1690                         if (dup_sack) {
1691                                 if (!tp->undo_marker)
1692                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1693                                 else
1694                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1695                         } else {
1696                                 /* Don't count olds caused by ACK reordering */
1697                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1698                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1699                                         continue;
1700                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1701                         }
1702 
1703                         NET_INC_STATS_BH(sock_net(sk), mib_idx);
1704                         if (i == 0)
1705                                 first_sack_index = -1;
1706                         continue;
1707                 }
1708 
1709                 /* Ignore very old stuff early */
1710                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1711                         continue;
1712 
1713                 used_sacks++;
1714         }
1715 
1716         /* order SACK blocks to allow in order walk of the retrans queue */
1717         for (i = used_sacks - 1; i > 0; i--) {
1718                 for (j = 0; j < i; j++) {
1719                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1720                                 swap(sp[j], sp[j + 1]);
1721 
1722                                 /* Track where the first SACK block goes to */
1723                                 if (j == first_sack_index)
1724                                         first_sack_index = j + 1;
1725                         }
1726                 }
1727         }
1728 
1729         skb = tcp_write_queue_head(sk);
1730         state.fack_count = 0;
1731         i = 0;
1732 
1733         if (!tp->sacked_out) {
1734                 /* It's already past, so skip checking against it */
1735                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1736         } else {
1737                 cache = tp->recv_sack_cache;
1738                 /* Skip empty blocks in at head of the cache */
1739                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1740                        !cache->end_seq)
1741                         cache++;
1742         }
1743 
1744         while (i < used_sacks) {
1745                 u32 start_seq = sp[i].start_seq;
1746                 u32 end_seq = sp[i].end_seq;
1747                 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1748                 struct tcp_sack_block *next_dup = NULL;
1749 
1750                 if (found_dup_sack && ((i + 1) == first_sack_index))
1751                         next_dup = &sp[i + 1];
1752 
1753                 /* Skip too early cached blocks */
1754                 while (tcp_sack_cache_ok(tp, cache) &&
1755                        !before(start_seq, cache->end_seq))
1756                         cache++;
1757 
1758                 /* Can skip some work by looking recv_sack_cache? */
1759                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1760                     after(end_seq, cache->start_seq)) {
1761 
1762                         /* Head todo? */
1763                         if (before(start_seq, cache->start_seq)) {
1764                                 skb = tcp_sacktag_skip(skb, sk, &state,
1765                                                        start_seq);
1766                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1767                                                        &state,
1768                                                        start_seq,
1769                                                        cache->start_seq,
1770                                                        dup_sack);
1771                         }
1772 
1773                         /* Rest of the block already fully processed? */
1774                         if (!after(end_seq, cache->end_seq))
1775                                 goto advance_sp;
1776 
1777                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1778                                                        &state,
1779                                                        cache->end_seq);
1780 
1781                         /* ...tail remains todo... */
1782                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1783                                 /* ...but better entrypoint exists! */
1784                                 skb = tcp_highest_sack(sk);
1785                                 if (skb == NULL)
1786                                         break;
1787                                 state.fack_count = tp->fackets_out;
1788                                 cache++;
1789                                 goto walk;
1790                         }
1791 
1792                         skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1793                         /* Check overlap against next cached too (past this one already) */
1794                         cache++;
1795                         continue;
1796                 }
1797 
1798                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1799                         skb = tcp_highest_sack(sk);
1800                         if (skb == NULL)
1801                                 break;
1802                         state.fack_count = tp->fackets_out;
1803                 }
1804                 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1805 
1806 walk:
1807                 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1808                                        start_seq, end_seq, dup_sack);
1809 
1810 advance_sp:
1811                 i++;
1812         }
1813 
1814         /* Clear the head of the cache sack blocks so we can skip it next time */
1815         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1816                 tp->recv_sack_cache[i].start_seq = 0;
1817                 tp->recv_sack_cache[i].end_seq = 0;
1818         }
1819         for (j = 0; j < used_sacks; j++)
1820                 tp->recv_sack_cache[i++] = sp[j];
1821 
1822         tcp_mark_lost_retrans(sk);
1823 
1824         tcp_verify_left_out(tp);
1825 
1826         if ((state.reord < tp->fackets_out) &&
1827             ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1828                 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1829 
1830 out:
1831 
1832 #if FASTRETRANS_DEBUG > 0
1833         WARN_ON((int)tp->sacked_out < 0);
1834         WARN_ON((int)tp->lost_out < 0);
1835         WARN_ON((int)tp->retrans_out < 0);
1836         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1837 #endif
1838         *sack_rtt_us = state.rtt_us;
1839         return state.flag;
1840 }
1841 
1842 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1843  * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1844  */
1845 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1846 {
1847         u32 holes;
1848 
1849         holes = max(tp->lost_out, 1U);
1850         holes = min(holes, tp->packets_out);
1851 
1852         if ((tp->sacked_out + holes) > tp->packets_out) {
1853                 tp->sacked_out = tp->packets_out - holes;
1854                 return true;
1855         }
1856         return false;
1857 }
1858 
1859 /* If we receive more dupacks than we expected counting segments
1860  * in assumption of absent reordering, interpret this as reordering.
1861  * The only another reason could be bug in receiver TCP.
1862  */
1863 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1864 {
1865         struct tcp_sock *tp = tcp_sk(sk);
1866         if (tcp_limit_reno_sacked(tp))
1867                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1868 }
1869 
1870 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1871 
1872 static void tcp_add_reno_sack(struct sock *sk)
1873 {
1874         struct tcp_sock *tp = tcp_sk(sk);
1875         tp->sacked_out++;
1876         tcp_check_reno_reordering(sk, 0);
1877         tcp_verify_left_out(tp);
1878 }
1879 
1880 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1881 
1882 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1883 {
1884         struct tcp_sock *tp = tcp_sk(sk);
1885 
1886         if (acked > 0) {
1887                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1888                 if (acked - 1 >= tp->sacked_out)
1889                         tp->sacked_out = 0;
1890                 else
1891                         tp->sacked_out -= acked - 1;
1892         }
1893         tcp_check_reno_reordering(sk, acked);
1894         tcp_verify_left_out(tp);
1895 }
1896 
1897 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1898 {
1899         tp->sacked_out = 0;
1900 }
1901 
1902 void tcp_clear_retrans(struct tcp_sock *tp)
1903 {
1904         tp->retrans_out = 0;
1905         tp->lost_out = 0;
1906         tp->undo_marker = 0;
1907         tp->undo_retrans = -1;
1908         tp->fackets_out = 0;
1909         tp->sacked_out = 0;
1910 }
1911 
1912 static inline void tcp_init_undo(struct tcp_sock *tp)
1913 {
1914         tp->undo_marker = tp->snd_una;
1915         /* Retransmission still in flight may cause DSACKs later. */
1916         tp->undo_retrans = tp->retrans_out ? : -1;
1917 }
1918 
1919 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1920  * and reset tags completely, otherwise preserve SACKs. If receiver
1921  * dropped its ofo queue, we will know this due to reneging detection.
1922  */
1923 void tcp_enter_loss(struct sock *sk)
1924 {
1925         const struct inet_connection_sock *icsk = inet_csk(sk);
1926         struct tcp_sock *tp = tcp_sk(sk);
1927         struct sk_buff *skb;
1928         bool new_recovery = false;
1929         bool is_reneg;                  /* is receiver reneging on SACKs? */
1930 
1931         /* Reduce ssthresh if it has not yet been made inside this window. */
1932         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1933             !after(tp->high_seq, tp->snd_una) ||
1934             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1935                 new_recovery = true;
1936                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1937                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1938                 tcp_ca_event(sk, CA_EVENT_LOSS);
1939                 tcp_init_undo(tp);
1940         }
1941         tp->snd_cwnd       = 1;
1942         tp->snd_cwnd_cnt   = 0;
1943         tp->snd_cwnd_stamp = tcp_time_stamp;
1944 
1945         tp->retrans_out = 0;
1946         tp->lost_out = 0;
1947 
1948         if (tcp_is_reno(tp))
1949                 tcp_reset_reno_sack(tp);
1950 
1951         skb = tcp_write_queue_head(sk);
1952         is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1953         if (is_reneg) {
1954                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1955                 tp->sacked_out = 0;
1956                 tp->fackets_out = 0;
1957         }
1958         tcp_clear_all_retrans_hints(tp);
1959 
1960         tcp_for_write_queue(skb, sk) {
1961                 if (skb == tcp_send_head(sk))
1962                         break;
1963 
1964                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1965                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
1966                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1967                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1968                         tp->lost_out += tcp_skb_pcount(skb);
1969                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1970                 }
1971         }
1972         tcp_verify_left_out(tp);
1973 
1974         /* Timeout in disordered state after receiving substantial DUPACKs
1975          * suggests that the degree of reordering is over-estimated.
1976          */
1977         if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1978             tp->sacked_out >= sysctl_tcp_reordering)
1979                 tp->reordering = min_t(unsigned int, tp->reordering,
1980                                        sysctl_tcp_reordering);
1981         tcp_set_ca_state(sk, TCP_CA_Loss);
1982         tp->high_seq = tp->snd_nxt;
1983         tcp_ecn_queue_cwr(tp);
1984 
1985         /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1986          * loss recovery is underway except recurring timeout(s) on
1987          * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1988          */
1989         tp->frto = sysctl_tcp_frto &&
1990                    (new_recovery || icsk->icsk_retransmits) &&
1991                    !inet_csk(sk)->icsk_mtup.probe_size;
1992 }
1993 
1994 /* If ACK arrived pointing to a remembered SACK, it means that our
1995  * remembered SACKs do not reflect real state of receiver i.e.
1996  * receiver _host_ is heavily congested (or buggy).
1997  *
1998  * To avoid big spurious retransmission bursts due to transient SACK
1999  * scoreboard oddities that look like reneging, we give the receiver a
2000  * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2001  * restore sanity to the SACK scoreboard. If the apparent reneging
2002  * persists until this RTO then we'll clear the SACK scoreboard.
2003  */
2004 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2005 {
2006         if (flag & FLAG_SACK_RENEGING) {
2007                 struct tcp_sock *tp = tcp_sk(sk);
2008                 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2009                                           msecs_to_jiffies(10));
2010 
2011                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2012                                           delay, TCP_RTO_MAX);
2013                 return true;
2014         }
2015         return false;
2016 }
2017 
2018 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2019 {
2020         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2021 }
2022 
2023 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2024  * counter when SACK is enabled (without SACK, sacked_out is used for
2025  * that purpose).
2026  *
2027  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2028  * segments up to the highest received SACK block so far and holes in
2029  * between them.
2030  *
2031  * With reordering, holes may still be in flight, so RFC3517 recovery
2032  * uses pure sacked_out (total number of SACKed segments) even though
2033  * it violates the RFC that uses duplicate ACKs, often these are equal
2034  * but when e.g. out-of-window ACKs or packet duplication occurs,
2035  * they differ. Since neither occurs due to loss, TCP should really
2036  * ignore them.
2037  */
2038 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2039 {
2040         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2041 }
2042 
2043 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2044 {
2045         struct tcp_sock *tp = tcp_sk(sk);
2046         unsigned long delay;
2047 
2048         /* Delay early retransmit and entering fast recovery for
2049          * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2050          * available, or RTO is scheduled to fire first.
2051          */
2052         if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2053             (flag & FLAG_ECE) || !tp->srtt_us)
2054                 return false;
2055 
2056         delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2057                     msecs_to_jiffies(2));
2058 
2059         if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2060                 return false;
2061 
2062         inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2063                                   TCP_RTO_MAX);
2064         return true;
2065 }
2066 
2067 /* Linux NewReno/SACK/FACK/ECN state machine.
2068  * --------------------------------------
2069  *
2070  * "Open"       Normal state, no dubious events, fast path.
2071  * "Disorder"   In all the respects it is "Open",
2072  *              but requires a bit more attention. It is entered when
2073  *              we see some SACKs or dupacks. It is split of "Open"
2074  *              mainly to move some processing from fast path to slow one.
2075  * "CWR"        CWND was reduced due to some Congestion Notification event.
2076  *              It can be ECN, ICMP source quench, local device congestion.
2077  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2078  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2079  *
2080  * tcp_fastretrans_alert() is entered:
2081  * - each incoming ACK, if state is not "Open"
2082  * - when arrived ACK is unusual, namely:
2083  *      * SACK
2084  *      * Duplicate ACK.
2085  *      * ECN ECE.
2086  *
2087  * Counting packets in flight is pretty simple.
2088  *
2089  *      in_flight = packets_out - left_out + retrans_out
2090  *
2091  *      packets_out is SND.NXT-SND.UNA counted in packets.
2092  *
2093  *      retrans_out is number of retransmitted segments.
2094  *
2095  *      left_out is number of segments left network, but not ACKed yet.
2096  *
2097  *              left_out = sacked_out + lost_out
2098  *
2099  *     sacked_out: Packets, which arrived to receiver out of order
2100  *                 and hence not ACKed. With SACKs this number is simply
2101  *                 amount of SACKed data. Even without SACKs
2102  *                 it is easy to give pretty reliable estimate of this number,
2103  *                 counting duplicate ACKs.
2104  *
2105  *       lost_out: Packets lost by network. TCP has no explicit
2106  *                 "loss notification" feedback from network (for now).
2107  *                 It means that this number can be only _guessed_.
2108  *                 Actually, it is the heuristics to predict lossage that
2109  *                 distinguishes different algorithms.
2110  *
2111  *      F.e. after RTO, when all the queue is considered as lost,
2112  *      lost_out = packets_out and in_flight = retrans_out.
2113  *
2114  *              Essentially, we have now two algorithms counting
2115  *              lost packets.
2116  *
2117  *              FACK: It is the simplest heuristics. As soon as we decided
2118  *              that something is lost, we decide that _all_ not SACKed
2119  *              packets until the most forward SACK are lost. I.e.
2120  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2121  *              It is absolutely correct estimate, if network does not reorder
2122  *              packets. And it loses any connection to reality when reordering
2123  *              takes place. We use FACK by default until reordering
2124  *              is suspected on the path to this destination.
2125  *
2126  *              NewReno: when Recovery is entered, we assume that one segment
2127  *              is lost (classic Reno). While we are in Recovery and
2128  *              a partial ACK arrives, we assume that one more packet
2129  *              is lost (NewReno). This heuristics are the same in NewReno
2130  *              and SACK.
2131  *
2132  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2133  *  deflation etc. CWND is real congestion window, never inflated, changes
2134  *  only according to classic VJ rules.
2135  *
2136  * Really tricky (and requiring careful tuning) part of algorithm
2137  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2138  * The first determines the moment _when_ we should reduce CWND and,
2139  * hence, slow down forward transmission. In fact, it determines the moment
2140  * when we decide that hole is caused by loss, rather than by a reorder.
2141  *
2142  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2143  * holes, caused by lost packets.
2144  *
2145  * And the most logically complicated part of algorithm is undo
2146  * heuristics. We detect false retransmits due to both too early
2147  * fast retransmit (reordering) and underestimated RTO, analyzing
2148  * timestamps and D-SACKs. When we detect that some segments were
2149  * retransmitted by mistake and CWND reduction was wrong, we undo
2150  * window reduction and abort recovery phase. This logic is hidden
2151  * inside several functions named tcp_try_undo_<something>.
2152  */
2153 
2154 /* This function decides, when we should leave Disordered state
2155  * and enter Recovery phase, reducing congestion window.
2156  *
2157  * Main question: may we further continue forward transmission
2158  * with the same cwnd?
2159  */
2160 static bool tcp_time_to_recover(struct sock *sk, int flag)
2161 {
2162         struct tcp_sock *tp = tcp_sk(sk);
2163         __u32 packets_out;
2164 
2165         /* Trick#1: The loss is proven. */
2166         if (tp->lost_out)
2167                 return true;
2168 
2169         /* Not-A-Trick#2 : Classic rule... */
2170         if (tcp_dupack_heuristics(tp) > tp->reordering)
2171                 return true;
2172 
2173         /* Trick#4: It is still not OK... But will it be useful to delay
2174          * recovery more?
2175          */
2176         packets_out = tp->packets_out;
2177         if (packets_out <= tp->reordering &&
2178             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2179             !tcp_may_send_now(sk)) {
2180                 /* We have nothing to send. This connection is limited
2181                  * either by receiver window or by application.
2182                  */
2183                 return true;
2184         }
2185 
2186         /* If a thin stream is detected, retransmit after first
2187          * received dupack. Employ only if SACK is supported in order
2188          * to avoid possible corner-case series of spurious retransmissions
2189          * Use only if there are no unsent data.
2190          */
2191         if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2192             tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2193             tcp_is_sack(tp) && !tcp_send_head(sk))
2194                 return true;
2195 
2196         /* Trick#6: TCP early retransmit, per RFC5827.  To avoid spurious
2197          * retransmissions due to small network reorderings, we implement
2198          * Mitigation A.3 in the RFC and delay the retransmission for a short
2199          * interval if appropriate.
2200          */
2201         if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2202             (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2203             !tcp_may_send_now(sk))
2204                 return !tcp_pause_early_retransmit(sk, flag);
2205 
2206         return false;
2207 }
2208 
2209 /* Detect loss in event "A" above by marking head of queue up as lost.
2210  * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2211  * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2212  * has at least tp->reordering SACKed seqments above it; "packets" refers to
2213  * the maximum SACKed segments to pass before reaching this limit.
2214  */
2215 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2216 {
2217         struct tcp_sock *tp = tcp_sk(sk);
2218         struct sk_buff *skb;
2219         int cnt, oldcnt;
2220         int err;
2221         unsigned int mss;
2222         /* Use SACK to deduce losses of new sequences sent during recovery */
2223         const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2224 
2225         WARN_ON(packets > tp->packets_out);
2226         if (tp->lost_skb_hint) {
2227                 skb = tp->lost_skb_hint;
2228                 cnt = tp->lost_cnt_hint;
2229                 /* Head already handled? */
2230                 if (mark_head && skb != tcp_write_queue_head(sk))
2231                         return;
2232         } else {
2233                 skb = tcp_write_queue_head(sk);
2234                 cnt = 0;
2235         }
2236 
2237         tcp_for_write_queue_from(skb, sk) {
2238                 if (skb == tcp_send_head(sk))
2239                         break;
2240                 /* TODO: do this better */
2241                 /* this is not the most efficient way to do this... */
2242                 tp->lost_skb_hint = skb;
2243                 tp->lost_cnt_hint = cnt;
2244 
2245                 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2246                         break;
2247 
2248                 oldcnt = cnt;
2249                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2250                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2251                         cnt += tcp_skb_pcount(skb);
2252 
2253                 if (cnt > packets) {
2254                         if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2255                             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2256                             (oldcnt >= packets))
2257                                 break;
2258 
2259                         mss = skb_shinfo(skb)->gso_size;
2260                         err = tcp_fragment(sk, skb, (packets - oldcnt) * mss,
2261                                            mss, GFP_ATOMIC);
2262                         if (err < 0)
2263                                 break;
2264                         cnt = packets;
2265                 }
2266 
2267                 tcp_skb_mark_lost(tp, skb);
2268 
2269                 if (mark_head)
2270                         break;
2271         }
2272         tcp_verify_left_out(tp);
2273 }
2274 
2275 /* Account newly detected lost packet(s) */
2276 
2277 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2278 {
2279         struct tcp_sock *tp = tcp_sk(sk);
2280 
2281         if (tcp_is_reno(tp)) {
2282                 tcp_mark_head_lost(sk, 1, 1);
2283         } else if (tcp_is_fack(tp)) {
2284                 int lost = tp->fackets_out - tp->reordering;
2285                 if (lost <= 0)
2286                         lost = 1;
2287                 tcp_mark_head_lost(sk, lost, 0);
2288         } else {
2289                 int sacked_upto = tp->sacked_out - tp->reordering;
2290                 if (sacked_upto >= 0)
2291                         tcp_mark_head_lost(sk, sacked_upto, 0);
2292                 else if (fast_rexmit)
2293                         tcp_mark_head_lost(sk, 1, 1);
2294         }
2295 }
2296 
2297 /* CWND moderation, preventing bursts due to too big ACKs
2298  * in dubious situations.
2299  */
2300 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2301 {
2302         tp->snd_cwnd = min(tp->snd_cwnd,
2303                            tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2304         tp->snd_cwnd_stamp = tcp_time_stamp;
2305 }
2306 
2307 /* Nothing was retransmitted or returned timestamp is less
2308  * than timestamp of the first retransmission.
2309  */
2310 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2311 {
2312         return !tp->retrans_stamp ||
2313                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2314                  before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2315 }
2316 
2317 /* Undo procedures. */
2318 
2319 /* We can clear retrans_stamp when there are no retransmissions in the
2320  * window. It would seem that it is trivially available for us in
2321  * tp->retrans_out, however, that kind of assumptions doesn't consider
2322  * what will happen if errors occur when sending retransmission for the
2323  * second time. ...It could the that such segment has only
2324  * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2325  * the head skb is enough except for some reneging corner cases that
2326  * are not worth the effort.
2327  *
2328  * Main reason for all this complexity is the fact that connection dying
2329  * time now depends on the validity of the retrans_stamp, in particular,
2330  * that successive retransmissions of a segment must not advance
2331  * retrans_stamp under any conditions.
2332  */
2333 static bool tcp_any_retrans_done(const struct sock *sk)
2334 {
2335         const struct tcp_sock *tp = tcp_sk(sk);
2336         struct sk_buff *skb;
2337 
2338         if (tp->retrans_out)
2339                 return true;
2340 
2341         skb = tcp_write_queue_head(sk);
2342         if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2343                 return true;
2344 
2345         return false;
2346 }
2347 
2348 #if FASTRETRANS_DEBUG > 1
2349 static void DBGUNDO(struct sock *sk, const char *msg)
2350 {
2351         struct tcp_sock *tp = tcp_sk(sk);
2352         struct inet_sock *inet = inet_sk(sk);
2353 
2354         if (sk->sk_family == AF_INET) {
2355                 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2356                          msg,
2357                          &inet->inet_daddr, ntohs(inet->inet_dport),
2358                          tp->snd_cwnd, tcp_left_out(tp),
2359                          tp->snd_ssthresh, tp->prior_ssthresh,
2360                          tp->packets_out);
2361         }
2362 #if IS_ENABLED(CONFIG_IPV6)
2363         else if (sk->sk_family == AF_INET6) {
2364                 struct ipv6_pinfo *np = inet6_sk(sk);
2365                 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2366                          msg,
2367                          &np->daddr, ntohs(inet->inet_dport),
2368                          tp->snd_cwnd, tcp_left_out(tp),
2369                          tp->snd_ssthresh, tp->prior_ssthresh,
2370                          tp->packets_out);
2371         }
2372 #endif
2373 }
2374 #else
2375 #define DBGUNDO(x...) do { } while (0)
2376 #endif
2377 
2378 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2379 {
2380         struct tcp_sock *tp = tcp_sk(sk);
2381 
2382         if (unmark_loss) {
2383                 struct sk_buff *skb;
2384 
2385                 tcp_for_write_queue(skb, sk) {
2386                         if (skb == tcp_send_head(sk))
2387                                 break;
2388                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2389                 }
2390                 tp->lost_out = 0;
2391                 tcp_clear_all_retrans_hints(tp);
2392         }
2393 
2394         if (tp->prior_ssthresh) {
2395                 const struct inet_connection_sock *icsk = inet_csk(sk);
2396 
2397                 if (icsk->icsk_ca_ops->undo_cwnd)
2398                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2399                 else
2400                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2401 
2402                 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2403                         tp->snd_ssthresh = tp->prior_ssthresh;
2404                         tcp_ecn_withdraw_cwr(tp);
2405                 }
2406         } else {
2407                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2408         }
2409         tp->snd_cwnd_stamp = tcp_time_stamp;
2410         tp->undo_marker = 0;
2411 }
2412 
2413 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2414 {
2415         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2416 }
2417 
2418 /* People celebrate: "We love our President!" */
2419 static bool tcp_try_undo_recovery(struct sock *sk)
2420 {
2421         struct tcp_sock *tp = tcp_sk(sk);
2422 
2423         if (tcp_may_undo(tp)) {
2424                 int mib_idx;
2425 
2426                 /* Happy end! We did not retransmit anything
2427                  * or our original transmission succeeded.
2428                  */
2429                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2430                 tcp_undo_cwnd_reduction(sk, false);
2431                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2432                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2433                 else
2434                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2435 
2436                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2437         }
2438         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2439                 /* Hold old state until something *above* high_seq
2440                  * is ACKed. For Reno it is MUST to prevent false
2441                  * fast retransmits (RFC2582). SACK TCP is safe. */
2442                 tcp_moderate_cwnd(tp);
2443                 if (!tcp_any_retrans_done(sk))
2444                         tp->retrans_stamp = 0;
2445                 return true;
2446         }
2447         tcp_set_ca_state(sk, TCP_CA_Open);
2448         return false;
2449 }
2450 
2451 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2452 static bool tcp_try_undo_dsack(struct sock *sk)
2453 {
2454         struct tcp_sock *tp = tcp_sk(sk);
2455 
2456         if (tp->undo_marker && !tp->undo_retrans) {
2457                 DBGUNDO(sk, "D-SACK");
2458                 tcp_undo_cwnd_reduction(sk, false);
2459                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2460                 return true;
2461         }
2462         return false;
2463 }
2464 
2465 /* Undo during loss recovery after partial ACK or using F-RTO. */
2466 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2467 {
2468         struct tcp_sock *tp = tcp_sk(sk);
2469 
2470         if (frto_undo || tcp_may_undo(tp)) {
2471                 tcp_undo_cwnd_reduction(sk, true);
2472 
2473                 DBGUNDO(sk, "partial loss");
2474                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2475                 if (frto_undo)
2476                         NET_INC_STATS_BH(sock_net(sk),
2477                                          LINUX_MIB_TCPSPURIOUSRTOS);
2478                 inet_csk(sk)->icsk_retransmits = 0;
2479                 if (frto_undo || tcp_is_sack(tp))
2480                         tcp_set_ca_state(sk, TCP_CA_Open);
2481                 return true;
2482         }
2483         return false;
2484 }
2485 
2486 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2487  * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2488  * It computes the number of packets to send (sndcnt) based on packets newly
2489  * delivered:
2490  *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2491  *      cwnd reductions across a full RTT.
2492  *   2) If packets in flight is lower than ssthresh (such as due to excess
2493  *      losses and/or application stalls), do not perform any further cwnd
2494  *      reductions, but instead slow start up to ssthresh.
2495  */
2496 static void tcp_init_cwnd_reduction(struct sock *sk)
2497 {
2498         struct tcp_sock *tp = tcp_sk(sk);
2499 
2500         tp->high_seq = tp->snd_nxt;
2501         tp->tlp_high_seq = 0;
2502         tp->snd_cwnd_cnt = 0;
2503         tp->prior_cwnd = tp->snd_cwnd;
2504         tp->prr_delivered = 0;
2505         tp->prr_out = 0;
2506         tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2507         tcp_ecn_queue_cwr(tp);
2508 }
2509 
2510 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2511                                int fast_rexmit)
2512 {
2513         struct tcp_sock *tp = tcp_sk(sk);
2514         int sndcnt = 0;
2515         int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2516         int newly_acked_sacked = prior_unsacked -
2517                                  (tp->packets_out - tp->sacked_out);
2518 
2519         tp->prr_delivered += newly_acked_sacked;
2520         if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2521                 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2522                                tp->prior_cwnd - 1;
2523                 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2524         } else {
2525                 sndcnt = min_t(int, delta,
2526                                max_t(int, tp->prr_delivered - tp->prr_out,
2527                                      newly_acked_sacked) + 1);
2528         }
2529 
2530         sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2531         tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2532 }
2533 
2534 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2535 {
2536         struct tcp_sock *tp = tcp_sk(sk);
2537 
2538         /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2539         if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2540             (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2541                 tp->snd_cwnd = tp->snd_ssthresh;
2542                 tp->snd_cwnd_stamp = tcp_time_stamp;
2543         }
2544         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2545 }
2546 
2547 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2548 void tcp_enter_cwr(struct sock *sk)
2549 {
2550         struct tcp_sock *tp = tcp_sk(sk);
2551 
2552         tp->prior_ssthresh = 0;
2553         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2554                 tp->undo_marker = 0;
2555                 tcp_init_cwnd_reduction(sk);
2556                 tcp_set_ca_state(sk, TCP_CA_CWR);
2557         }
2558 }
2559 
2560 static void tcp_try_keep_open(struct sock *sk)
2561 {
2562         struct tcp_sock *tp = tcp_sk(sk);
2563         int state = TCP_CA_Open;
2564 
2565         if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2566                 state = TCP_CA_Disorder;
2567 
2568         if (inet_csk(sk)->icsk_ca_state != state) {
2569                 tcp_set_ca_state(sk, state);
2570                 tp->high_seq = tp->snd_nxt;
2571         }
2572 }
2573 
2574 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2575 {
2576         struct tcp_sock *tp = tcp_sk(sk);
2577 
2578         tcp_verify_left_out(tp);
2579 
2580         if (!tcp_any_retrans_done(sk))
2581                 tp->retrans_stamp = 0;
2582 
2583         if (flag & FLAG_ECE)
2584                 tcp_enter_cwr(sk);
2585 
2586         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2587                 tcp_try_keep_open(sk);
2588         } else {
2589                 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2590         }
2591 }
2592 
2593 static void tcp_mtup_probe_failed(struct sock *sk)
2594 {
2595         struct inet_connection_sock *icsk = inet_csk(sk);
2596 
2597         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2598         icsk->icsk_mtup.probe_size = 0;
2599 }
2600 
2601 static void tcp_mtup_probe_success(struct sock *sk)
2602 {
2603         struct tcp_sock *tp = tcp_sk(sk);
2604         struct inet_connection_sock *icsk = inet_csk(sk);
2605 
2606         /* FIXME: breaks with very large cwnd */
2607         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2608         tp->snd_cwnd = tp->snd_cwnd *
2609                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2610                        icsk->icsk_mtup.probe_size;
2611         tp->snd_cwnd_cnt = 0;
2612         tp->snd_cwnd_stamp = tcp_time_stamp;
2613         tp->snd_ssthresh = tcp_current_ssthresh(sk);
2614 
2615         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2616         icsk->icsk_mtup.probe_size = 0;
2617         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2618 }
2619 
2620 /* Do a simple retransmit without using the backoff mechanisms in
2621  * tcp_timer. This is used for path mtu discovery.
2622  * The socket is already locked here.
2623  */
2624 void tcp_simple_retransmit(struct sock *sk)
2625 {
2626         const struct inet_connection_sock *icsk = inet_csk(sk);
2627         struct tcp_sock *tp = tcp_sk(sk);
2628         struct sk_buff *skb;
2629         unsigned int mss = tcp_current_mss(sk);
2630         u32 prior_lost = tp->lost_out;
2631 
2632         tcp_for_write_queue(skb, sk) {
2633                 if (skb == tcp_send_head(sk))
2634                         break;
2635                 if (tcp_skb_seglen(skb) > mss &&
2636                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2637                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2638                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2639                                 tp->retrans_out -= tcp_skb_pcount(skb);
2640                         }
2641                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2642                 }
2643         }
2644 
2645         tcp_clear_retrans_hints_partial(tp);
2646 
2647         if (prior_lost == tp->lost_out)
2648                 return;
2649 
2650         if (tcp_is_reno(tp))
2651                 tcp_limit_reno_sacked(tp);
2652 
2653         tcp_verify_left_out(tp);
2654 
2655         /* Don't muck with the congestion window here.
2656          * Reason is that we do not increase amount of _data_
2657          * in network, but units changed and effective
2658          * cwnd/ssthresh really reduced now.
2659          */
2660         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2661                 tp->high_seq = tp->snd_nxt;
2662                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2663                 tp->prior_ssthresh = 0;
2664                 tp->undo_marker = 0;
2665                 tcp_set_ca_state(sk, TCP_CA_Loss);
2666         }
2667         tcp_xmit_retransmit_queue(sk);
2668 }
2669 EXPORT_SYMBOL(tcp_simple_retransmit);
2670 
2671 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2672 {
2673         struct tcp_sock *tp = tcp_sk(sk);
2674         int mib_idx;
2675 
2676         if (tcp_is_reno(tp))
2677                 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2678         else
2679                 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2680 
2681         NET_INC_STATS_BH(sock_net(sk), mib_idx);
2682 
2683         tp->prior_ssthresh = 0;
2684         tcp_init_undo(tp);
2685 
2686         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2687                 if (!ece_ack)
2688                         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2689                 tcp_init_cwnd_reduction(sk);
2690         }
2691         tcp_set_ca_state(sk, TCP_CA_Recovery);
2692 }
2693 
2694 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2695  * recovered or spurious. Otherwise retransmits more on partial ACKs.
2696  */
2697 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2698 {
2699         struct tcp_sock *tp = tcp_sk(sk);
2700         bool recovered = !before(tp->snd_una, tp->high_seq);
2701 
2702         if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2703                 /* Step 3.b. A timeout is spurious if not all data are
2704                  * lost, i.e., never-retransmitted data are (s)acked.
2705                  */
2706                 if (tcp_try_undo_loss(sk, flag & FLAG_ORIG_SACK_ACKED))
2707                         return;
2708 
2709                 if (after(tp->snd_nxt, tp->high_seq) &&
2710                     (flag & FLAG_DATA_SACKED || is_dupack)) {
2711                         tp->frto = 0; /* Loss was real: 2nd part of step 3.a */
2712                 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2713                         tp->high_seq = tp->snd_nxt;
2714                         __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2715                                                   TCP_NAGLE_OFF);
2716                         if (after(tp->snd_nxt, tp->high_seq))
2717                                 return; /* Step 2.b */
2718                         tp->frto = 0;
2719                 }
2720         }
2721 
2722         if (recovered) {
2723                 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2724                 tcp_try_undo_recovery(sk);
2725                 return;
2726         }
2727         if (tcp_is_reno(tp)) {
2728                 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2729                  * delivered. Lower inflight to clock out (re)tranmissions.
2730                  */
2731                 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2732                         tcp_add_reno_sack(sk);
2733                 else if (flag & FLAG_SND_UNA_ADVANCED)
2734                         tcp_reset_reno_sack(tp);
2735         }
2736         if (tcp_try_undo_loss(sk, false))
2737                 return;
2738         tcp_xmit_retransmit_queue(sk);
2739 }
2740 
2741 /* Undo during fast recovery after partial ACK. */
2742 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2743                                  const int prior_unsacked)
2744 {
2745         struct tcp_sock *tp = tcp_sk(sk);
2746 
2747         if (tp->undo_marker && tcp_packet_delayed(tp)) {
2748                 /* Plain luck! Hole if filled with delayed
2749                  * packet, rather than with a retransmit.
2750                  */
2751                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2752 
2753                 /* We are getting evidence that the reordering degree is higher
2754                  * than we realized. If there are no retransmits out then we
2755                  * can undo. Otherwise we clock out new packets but do not
2756                  * mark more packets lost or retransmit more.
2757                  */
2758                 if (tp->retrans_out) {
2759                         tcp_cwnd_reduction(sk, prior_unsacked, 0);
2760                         return true;
2761                 }
2762 
2763                 if (!tcp_any_retrans_done(sk))
2764                         tp->retrans_stamp = 0;
2765 
2766                 DBGUNDO(sk, "partial recovery");
2767                 tcp_undo_cwnd_reduction(sk, true);
2768                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2769                 tcp_try_keep_open(sk);
2770                 return true;
2771         }
2772         return false;
2773 }
2774 
2775 /* Process an event, which can update packets-in-flight not trivially.
2776  * Main goal of this function is to calculate new estimate for left_out,
2777  * taking into account both packets sitting in receiver's buffer and
2778  * packets lost by network.
2779  *
2780  * Besides that it does CWND reduction, when packet loss is detected
2781  * and changes state of machine.
2782  *
2783  * It does _not_ decide what to send, it is made in function
2784  * tcp_xmit_retransmit_queue().
2785  */
2786 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2787                                   const int prior_unsacked,
2788                                   bool is_dupack, int flag)
2789 {
2790         struct inet_connection_sock *icsk = inet_csk(sk);
2791         struct tcp_sock *tp = tcp_sk(sk);
2792         bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2793                                     (tcp_fackets_out(tp) > tp->reordering));
2794         int fast_rexmit = 0;
2795 
2796         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2797                 tp->sacked_out = 0;
2798         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2799                 tp->fackets_out = 0;
2800 
2801         /* Now state machine starts.
2802          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2803         if (flag & FLAG_ECE)
2804                 tp->prior_ssthresh = 0;
2805 
2806         /* B. In all the states check for reneging SACKs. */
2807         if (tcp_check_sack_reneging(sk, flag))
2808                 return;
2809 
2810         /* C. Check consistency of the current state. */
2811         tcp_verify_left_out(tp);
2812 
2813         /* D. Check state exit conditions. State can be terminated
2814          *    when high_seq is ACKed. */
2815         if (icsk->icsk_ca_state == TCP_CA_Open) {
2816                 WARN_ON(tp->retrans_out != 0);
2817                 tp->retrans_stamp = 0;
2818         } else if (!before(tp->snd_una, tp->high_seq)) {
2819                 switch (icsk->icsk_ca_state) {
2820                 case TCP_CA_CWR:
2821                         /* CWR is to be held something *above* high_seq
2822                          * is ACKed for CWR bit to reach receiver. */
2823                         if (tp->snd_una != tp->high_seq) {
2824                                 tcp_end_cwnd_reduction(sk);
2825                                 tcp_set_ca_state(sk, TCP_CA_Open);
2826                         }
2827                         break;
2828 
2829                 case TCP_CA_Recovery:
2830                         if (tcp_is_reno(tp))
2831                                 tcp_reset_reno_sack(tp);
2832                         if (tcp_try_undo_recovery(sk))
2833                                 return;
2834                         tcp_end_cwnd_reduction(sk);
2835                         break;
2836                 }
2837         }
2838 
2839         /* E. Process state. */
2840         switch (icsk->icsk_ca_state) {
2841         case TCP_CA_Recovery:
2842                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2843                         if (tcp_is_reno(tp) && is_dupack)
2844                                 tcp_add_reno_sack(sk);
2845                 } else {
2846                         if (tcp_try_undo_partial(sk, acked, prior_unsacked))
2847                                 return;
2848                         /* Partial ACK arrived. Force fast retransmit. */
2849                         do_lost = tcp_is_reno(tp) ||
2850                                   tcp_fackets_out(tp) > tp->reordering;
2851                 }
2852                 if (tcp_try_undo_dsack(sk)) {
2853                         tcp_try_keep_open(sk);
2854                         return;
2855                 }
2856                 break;
2857         case TCP_CA_Loss:
2858                 tcp_process_loss(sk, flag, is_dupack);
2859                 if (icsk->icsk_ca_state != TCP_CA_Open)
2860                         return;
2861                 /* Fall through to processing in Open state. */
2862         default:
2863                 if (tcp_is_reno(tp)) {
2864                         if (flag & FLAG_SND_UNA_ADVANCED)
2865                                 tcp_reset_reno_sack(tp);
2866                         if (is_dupack)
2867                                 tcp_add_reno_sack(sk);
2868                 }
2869 
2870                 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2871                         tcp_try_undo_dsack(sk);
2872 
2873                 if (!tcp_time_to_recover(sk, flag)) {
2874                         tcp_try_to_open(sk, flag, prior_unsacked);
2875                         return;
2876                 }
2877 
2878                 /* MTU probe failure: don't reduce cwnd */
2879                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2880                     icsk->icsk_mtup.probe_size &&
2881                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2882                         tcp_mtup_probe_failed(sk);
2883                         /* Restores the reduction we did in tcp_mtup_probe() */
2884                         tp->snd_cwnd++;
2885                         tcp_simple_retransmit(sk);
2886                         return;
2887                 }
2888 
2889                 /* Otherwise enter Recovery state */
2890                 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2891                 fast_rexmit = 1;
2892         }
2893 
2894         if (do_lost)
2895                 tcp_update_scoreboard(sk, fast_rexmit);
2896         tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
2897         tcp_xmit_retransmit_queue(sk);
2898 }
2899 
2900 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2901                                       long seq_rtt_us, long sack_rtt_us)
2902 {
2903         const struct tcp_sock *tp = tcp_sk(sk);
2904 
2905         /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2906          * broken middle-boxes or peers may corrupt TS-ECR fields. But
2907          * Karn's algorithm forbids taking RTT if some retransmitted data
2908          * is acked (RFC6298).
2909          */
2910         if (flag & FLAG_RETRANS_DATA_ACKED)
2911                 seq_rtt_us = -1L;
2912 
2913         if (seq_rtt_us < 0)
2914                 seq_rtt_us = sack_rtt_us;
2915 
2916         /* RTTM Rule: A TSecr value received in a segment is used to
2917          * update the averaged RTT measurement only if the segment
2918          * acknowledges some new data, i.e., only if it advances the
2919          * left edge of the send window.
2920          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2921          */
2922         if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2923             flag & FLAG_ACKED)
2924                 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2925 
2926         if (seq_rtt_us < 0)
2927                 return false;
2928 
2929         tcp_rtt_estimator(sk, seq_rtt_us);
2930         tcp_set_rto(sk);
2931 
2932         /* RFC6298: only reset backoff on valid RTT measurement. */
2933         inet_csk(sk)->icsk_backoff = 0;
2934         return true;
2935 }
2936 
2937 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2938 static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
2939 {
2940         struct tcp_sock *tp = tcp_sk(sk);
2941         long seq_rtt_us = -1L;
2942 
2943         if (synack_stamp && !tp->total_retrans)
2944                 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - synack_stamp);
2945 
2946         /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2947          * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2948          */
2949         if (!tp->srtt_us)
2950                 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt_us, -1L);
2951 }
2952 
2953 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2954 {
2955         const struct inet_connection_sock *icsk = inet_csk(sk);
2956 
2957         icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2958         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2959 }
2960 
2961 /* Restart timer after forward progress on connection.
2962  * RFC2988 recommends to restart timer to now+rto.
2963  */
2964 void tcp_rearm_rto(struct sock *sk)
2965 {
2966         const struct inet_connection_sock *icsk = inet_csk(sk);
2967         struct tcp_sock *tp = tcp_sk(sk);
2968 
2969         /* If the retrans timer is currently being used by Fast Open
2970          * for SYN-ACK retrans purpose, stay put.
2971          */
2972         if (tp->fastopen_rsk)
2973                 return;
2974 
2975         if (!tp->packets_out) {
2976                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2977         } else {
2978                 u32 rto = inet_csk(sk)->icsk_rto;
2979                 /* Offset the time elapsed after installing regular RTO */
2980                 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2981                     icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2982                         struct sk_buff *skb = tcp_write_queue_head(sk);
2983                         const u32 rto_time_stamp =
2984                                 tcp_skb_timestamp(skb) + rto;
2985                         s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2986                         /* delta may not be positive if the socket is locked
2987                          * when the retrans timer fires and is rescheduled.
2988                          */
2989                         if (delta > 0)
2990                                 rto = delta;
2991                 }
2992                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2993                                           TCP_RTO_MAX);
2994         }
2995 }
2996 
2997 /* This function is called when the delayed ER timer fires. TCP enters
2998  * fast recovery and performs fast-retransmit.
2999  */
3000 void tcp_resume_early_retransmit(struct sock *sk)
3001 {
3002         struct tcp_sock *tp = tcp_sk(sk);
3003 
3004         tcp_rearm_rto(sk);
3005 
3006         /* Stop if ER is disabled after the delayed ER timer is scheduled */
3007         if (!tp->do_early_retrans)
3008                 return;
3009 
3010         tcp_enter_recovery(sk, false);
3011         tcp_update_scoreboard(sk, 1);
3012         tcp_xmit_retransmit_queue(sk);
3013 }
3014 
3015 /* If we get here, the whole TSO packet has not been acked. */
3016 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3017 {
3018         struct tcp_sock *tp = tcp_sk(sk);
3019         u32 packets_acked;
3020 
3021         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3022 
3023         packets_acked = tcp_skb_pcount(skb);
3024         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3025                 return 0;
3026         packets_acked -= tcp_skb_pcount(skb);
3027 
3028         if (packets_acked) {
3029                 BUG_ON(tcp_skb_pcount(skb) == 0);
3030                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3031         }
3032 
3033         return packets_acked;
3034 }
3035 
3036 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3037                            u32 prior_snd_una)
3038 {
3039         const struct skb_shared_info *shinfo;
3040 
3041         /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3042         if (likely(!(sk->sk_tsflags & SOF_TIMESTAMPING_TX_ACK)))
3043                 return;
3044 
3045         shinfo = skb_shinfo(skb);
3046         if ((shinfo->tx_flags & SKBTX_ACK_TSTAMP) &&
3047             between(shinfo->tskey, prior_snd_una, tcp_sk(sk)->snd_una - 1))
3048                 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3049 }
3050 
3051 /* Remove acknowledged frames from the retransmission queue. If our packet
3052  * is before the ack sequence we can discard it as it's confirmed to have
3053  * arrived at the other end.
3054  */
3055 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3056                                u32 prior_snd_una, long sack_rtt_us)
3057 {
3058         const struct inet_connection_sock *icsk = inet_csk(sk);
3059         struct skb_mstamp first_ackt, last_ackt, now;
3060         struct tcp_sock *tp = tcp_sk(sk);
3061         u32 prior_sacked = tp->sacked_out;
3062         u32 reord = tp->packets_out;
3063         bool fully_acked = true;
3064         long ca_seq_rtt_us = -1L;
3065         long seq_rtt_us = -1L;
3066         struct sk_buff *skb;
3067         u32 pkts_acked = 0;
3068         bool rtt_update;
3069         int flag = 0;
3070 
3071         first_ackt.v64 = 0;
3072 
3073         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3074                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3075                 u8 sacked = scb->sacked;
3076                 u32 acked_pcount;
3077 
3078                 tcp_ack_tstamp(sk, skb, prior_snd_una);
3079 
3080                 /* Determine how many packets and what bytes were acked, tso and else */
3081                 if (after(scb->end_seq, tp->snd_una)) {
3082                         if (tcp_skb_pcount(skb) == 1 ||
3083                             !after(tp->snd_una, scb->seq))
3084                                 break;
3085 
3086                         acked_pcount = tcp_tso_acked(sk, skb);
3087                         if (!acked_pcount)
3088                                 break;
3089 
3090                         fully_acked = false;
3091                 } else {
3092                         /* Speedup tcp_unlink_write_queue() and next loop */
3093                         prefetchw(skb->next);
3094                         acked_pcount = tcp_skb_pcount(skb);
3095                 }
3096 
3097                 if (unlikely(sacked & TCPCB_RETRANS)) {
3098                         if (sacked & TCPCB_SACKED_RETRANS)
3099                                 tp->retrans_out -= acked_pcount;
3100                         flag |= FLAG_RETRANS_DATA_ACKED;
3101                 } else {
3102                         last_ackt = skb->skb_mstamp;
3103                         WARN_ON_ONCE(last_ackt.v64 == 0);
3104                         if (!first_ackt.v64)
3105                                 first_ackt = last_ackt;
3106 
3107                         if (!(sacked & TCPCB_SACKED_ACKED))
3108                                 reord = min(pkts_acked, reord);
3109                         if (!after(scb->end_seq, tp->high_seq))
3110                                 flag |= FLAG_ORIG_SACK_ACKED;
3111                 }
3112 
3113                 if (sacked & TCPCB_SACKED_ACKED)
3114                         tp->sacked_out -= acked_pcount;
3115                 if (sacked & TCPCB_LOST)
3116                         tp->lost_out -= acked_pcount;
3117 
3118                 tp->packets_out -= acked_pcount;
3119                 pkts_acked += acked_pcount;
3120 
3121                 /* Initial outgoing SYN's get put onto the write_queue
3122                  * just like anything else we transmit.  It is not
3123                  * true data, and if we misinform our callers that
3124                  * this ACK acks real data, we will erroneously exit
3125                  * connection startup slow start one packet too
3126                  * quickly.  This is severely frowned upon behavior.
3127                  */
3128                 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3129                         flag |= FLAG_DATA_ACKED;
3130                 } else {
3131                         flag |= FLAG_SYN_ACKED;
3132                         tp->retrans_stamp = 0;
3133                 }
3134 
3135                 if (!fully_acked)
3136                         break;
3137 
3138                 tcp_unlink_write_queue(skb, sk);
3139                 sk_wmem_free_skb(sk, skb);
3140                 if (unlikely(skb == tp->retransmit_skb_hint))
3141                         tp->retransmit_skb_hint = NULL;
3142                 if (unlikely(skb == tp->lost_skb_hint))
3143                         tp->lost_skb_hint = NULL;
3144         }
3145 
3146         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3147                 tp->snd_up = tp->snd_una;
3148 
3149         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3150                 flag |= FLAG_SACK_RENEGING;
3151 
3152         skb_mstamp_get(&now);
3153         if (likely(first_ackt.v64)) {
3154                 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3155                 ca_seq_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3156         }
3157 
3158         rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us);
3159 
3160         if (flag & FLAG_ACKED) {
3161                 const struct tcp_congestion_ops *ca_ops
3162                         = inet_csk(sk)->icsk_ca_ops;
3163 
3164                 tcp_rearm_rto(sk);
3165                 if (unlikely(icsk->icsk_mtup.probe_size &&
3166                              !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3167                         tcp_mtup_probe_success(sk);
3168                 }
3169 
3170                 if (tcp_is_reno(tp)) {
3171                         tcp_remove_reno_sacks(sk, pkts_acked);
3172                 } else {
3173                         int delta;
3174 
3175                         /* Non-retransmitted hole got filled? That's reordering */
3176                         if (reord < prior_fackets)
3177                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3178 
3179                         delta = tcp_is_fack(tp) ? pkts_acked :
3180                                                   prior_sacked - tp->sacked_out;
3181                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3182                 }
3183 
3184                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3185 
3186                 if (ca_ops->pkts_acked)
3187                         ca_ops->pkts_acked(sk, pkts_acked, ca_seq_rtt_us);
3188 
3189         } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3190                    sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3191                 /* Do not re-arm RTO if the sack RTT is measured from data sent
3192                  * after when the head was last (re)transmitted. Otherwise the
3193                  * timeout may continue to extend in loss recovery.
3194                  */
3195                 tcp_rearm_rto(sk);
3196         }
3197 
3198 #if FASTRETRANS_DEBUG > 0
3199         WARN_ON((int)tp->sacked_out < 0);
3200         WARN_ON((int)tp->lost_out < 0);
3201         WARN_ON((int)tp->retrans_out < 0);
3202         if (!tp->packets_out && tcp_is_sack(tp)) {
3203                 icsk = inet_csk(sk);
3204                 if (tp->lost_out) {
3205                         pr_debug("Leak l=%u %d\n",
3206                                  tp->lost_out, icsk->icsk_ca_state);
3207                         tp->lost_out = 0;
3208                 }
3209                 if (tp->sacked_out) {
3210                         pr_debug("Leak s=%u %d\n",
3211                                  tp->sacked_out, icsk->icsk_ca_state);
3212                         tp->sacked_out = 0;
3213                 }
3214                 if (tp->retrans_out) {
3215                         pr_debug("Leak r=%u %d\n",
3216                                  tp->retrans_out, icsk->icsk_ca_state);
3217                         tp->retrans_out = 0;
3218                 }
3219         }
3220 #endif
3221         return flag;
3222 }
3223 
3224 static void tcp_ack_probe(struct sock *sk)
3225 {
3226         const struct tcp_sock *tp = tcp_sk(sk);
3227         struct inet_connection_sock *icsk = inet_csk(sk);
3228 
3229         /* Was it a usable window open? */
3230 
3231         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3232                 icsk->icsk_backoff = 0;
3233                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3234                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3235                  * This function is not for random using!
3236                  */
3237         } else {
3238                 unsigned long when = inet_csk_rto_backoff(icsk, TCP_RTO_MAX);
3239 
3240                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3241                                           when, TCP_RTO_MAX);
3242         }
3243 }
3244 
3245 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3246 {
3247         return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3248                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3249 }
3250 
3251 /* Decide wheather to run the increase function of congestion control. */
3252 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3253 {
3254         if (tcp_in_cwnd_reduction(sk))
3255                 return false;
3256 
3257         /* If reordering is high then always grow cwnd whenever data is
3258          * delivered regardless of its ordering. Otherwise stay conservative
3259          * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3260          * new SACK or ECE mark may first advance cwnd here and later reduce
3261          * cwnd in tcp_fastretrans_alert() based on more states.
3262          */
3263         if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3264                 return flag & FLAG_FORWARD_PROGRESS;
3265 
3266         return flag & FLAG_DATA_ACKED;
3267 }
3268 
3269 /* Check that window update is acceptable.
3270  * The function assumes that snd_una<=ack<=snd_next.
3271  */
3272 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3273                                         const u32 ack, const u32 ack_seq,
3274                                         const u32 nwin)
3275 {
3276         return  after(ack, tp->snd_una) ||
3277                 after(ack_seq, tp->snd_wl1) ||
3278                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3279 }
3280 
3281 /* Update our send window.
3282  *
3283  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3284  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3285  */
3286 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3287                                  u32 ack_seq)
3288 {
3289         struct tcp_sock *tp = tcp_sk(sk);
3290         int flag = 0;
3291         u32 nwin = ntohs(tcp_hdr(skb)->window);
3292 
3293         if (likely(!tcp_hdr(skb)->syn))
3294                 nwin <<= tp->rx_opt.snd_wscale;
3295 
3296         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3297                 flag |= FLAG_WIN_UPDATE;
3298                 tcp_update_wl(tp, ack_seq);
3299 
3300                 if (tp->snd_wnd != nwin) {
3301                         tp->snd_wnd = nwin;
3302 
3303                         /* Note, it is the only place, where
3304                          * fast path is recovered for sending TCP.
3305                          */
3306                         tp->pred_flags = 0;
3307                         tcp_fast_path_check(sk);
3308 
3309                         if (nwin > tp->max_window) {
3310                                 tp->max_window = nwin;
3311                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3312                         }
3313                 }
3314         }
3315 
3316         tp->snd_una = ack;
3317 
3318         return flag;
3319 }
3320 
3321 /* RFC 5961 7 [ACK Throttling] */
3322 static void tcp_send_challenge_ack(struct sock *sk)
3323 {
3324         /* unprotected vars, we dont care of overwrites */
3325         static u32 challenge_timestamp;
3326         static unsigned int challenge_count;
3327         u32 now = jiffies / HZ;
3328 
3329         if (now != challenge_timestamp) {
3330                 challenge_timestamp = now;
3331                 challenge_count = 0;
3332         }
3333         if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3334                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3335                 tcp_send_ack(sk);
3336         }
3337 }
3338 
3339 static void tcp_store_ts_recent(struct tcp_sock *tp)
3340 {
3341         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3342         tp->rx_opt.ts_recent_stamp = get_seconds();
3343 }
3344 
3345 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3346 {
3347         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3348                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3349                  * extra check below makes sure this can only happen
3350                  * for pure ACK frames.  -DaveM
3351                  *
3352                  * Not only, also it occurs for expired timestamps.
3353                  */
3354 
3355                 if (tcp_paws_check(&tp->rx_opt, 0))
3356                         tcp_store_ts_recent(tp);
3357         }
3358 }
3359 
3360 /* This routine deals with acks during a TLP episode.
3361  * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3362  */
3363 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3364 {
3365         struct tcp_sock *tp = tcp_sk(sk);
3366         bool is_tlp_dupack = (ack == tp->tlp_high_seq) &&
3367                              !(flag & (FLAG_SND_UNA_ADVANCED |
3368                                        FLAG_NOT_DUP | FLAG_DATA_SACKED));
3369 
3370         /* Mark the end of TLP episode on receiving TLP dupack or when
3371          * ack is after tlp_high_seq.
3372          */
3373         if (is_tlp_dupack) {
3374                 tp->tlp_high_seq = 0;
3375                 return;
3376         }
3377 
3378         if (after(ack, tp->tlp_high_seq)) {
3379                 tp->tlp_high_seq = 0;
3380                 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3381                 if (!(flag & FLAG_DSACKING_ACK)) {
3382                         tcp_init_cwnd_reduction(sk);
3383                         tcp_set_ca_state(sk, TCP_CA_CWR);
3384                         tcp_end_cwnd_reduction(sk);
3385                         tcp_try_keep_open(sk);
3386                         NET_INC_STATS_BH(sock_net(sk),
3387                                          LINUX_MIB_TCPLOSSPROBERECOVERY);
3388                 }
3389         }
3390 }
3391 
3392 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3393 {
3394         const struct inet_connection_sock *icsk = inet_csk(sk);
3395 
3396         if (icsk->icsk_ca_ops->in_ack_event)
3397                 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3398 }
3399 
3400 /* This routine deals with incoming acks, but not outgoing ones. */
3401 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3402 {
3403         struct inet_connection_sock *icsk = inet_csk(sk);
3404         struct tcp_sock *tp = tcp_sk(sk);
3405         u32 prior_snd_una = tp->snd_una;
3406         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3407         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3408         bool is_dupack = false;
3409         u32 prior_fackets;
3410         int prior_packets = tp->packets_out;
3411         const int prior_unsacked = tp->packets_out - tp->sacked_out;
3412         int acked = 0; /* Number of packets newly acked */
3413         long sack_rtt_us = -1L;
3414 
3415         /* We very likely will need to access write queue head. */
3416         prefetchw(sk->sk_write_queue.next);
3417 
3418         /* If the ack is older than previous acks
3419          * then we can probably ignore it.
3420          */
3421         if (before(ack, prior_snd_una)) {
3422                 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3423                 if (before(ack, prior_snd_una - tp->max_window)) {
3424                         tcp_send_challenge_ack(sk);
3425                         return -1;
3426                 }
3427                 goto old_ack;
3428         }
3429 
3430         /* If the ack includes data we haven't sent yet, discard
3431          * this segment (RFC793 Section 3.9).
3432          */
3433         if (after(ack, tp->snd_nxt))
3434                 goto invalid_ack;
3435 
3436         if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3437             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3438                 tcp_rearm_rto(sk);
3439 
3440         if (after(ack, prior_snd_una)) {
3441                 flag |= FLAG_SND_UNA_ADVANCED;
3442                 icsk->icsk_retransmits = 0;
3443         }
3444 
3445         prior_fackets = tp->fackets_out;
3446 
3447         /* ts_recent update must be made after we are sure that the packet
3448          * is in window.
3449          */
3450         if (flag & FLAG_UPDATE_TS_RECENT)
3451                 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3452 
3453         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3454                 /* Window is constant, pure forward advance.
3455                  * No more checks are required.
3456                  * Note, we use the fact that SND.UNA>=SND.WL2.
3457                  */
3458                 tcp_update_wl(tp, ack_seq);
3459                 tp->snd_una = ack;
3460                 flag |= FLAG_WIN_UPDATE;
3461 
3462                 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3463 
3464                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3465         } else {
3466                 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3467 
3468                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3469                         flag |= FLAG_DATA;
3470                 else
3471                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3472 
3473                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3474 
3475                 if (TCP_SKB_CB(skb)->sacked)
3476                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3477                                                         &sack_rtt_us);
3478 
3479                 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3480                         flag |= FLAG_ECE;
3481                         ack_ev_flags |= CA_ACK_ECE;
3482                 }
3483 
3484                 if (flag & FLAG_WIN_UPDATE)
3485                         ack_ev_flags |= CA_ACK_WIN_UPDATE;
3486 
3487                 tcp_in_ack_event(sk, ack_ev_flags);
3488         }
3489 
3490         /* We passed data and got it acked, remove any soft error
3491          * log. Something worked...
3492          */
3493         sk->sk_err_soft = 0;
3494         icsk->icsk_probes_out = 0;
3495         tp->rcv_tstamp = tcp_time_stamp;
3496         if (!prior_packets)
3497                 goto no_queue;
3498 
3499         /* See if we can take anything off of the retransmit queue. */
3500         acked = tp->packets_out;
3501         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
3502                                     sack_rtt_us);
3503         acked -= tp->packets_out;
3504 
3505         /* Advance cwnd if state allows */
3506         if (tcp_may_raise_cwnd(sk, flag))
3507                 tcp_cong_avoid(sk, ack, acked);
3508 
3509         if (tcp_ack_is_dubious(sk, flag)) {
3510                 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3511                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3512                                       is_dupack, flag);
3513         }
3514         if (tp->tlp_high_seq)
3515                 tcp_process_tlp_ack(sk, ack, flag);
3516 
3517         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3518                 struct dst_entry *dst = __sk_dst_get(sk);
3519                 if (dst)
3520                         dst_confirm(dst);
3521         }
3522 
3523         if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3524                 tcp_schedule_loss_probe(sk);
3525         tcp_update_pacing_rate(sk);
3526         return 1;
3527 
3528 no_queue:
3529         /* If data was DSACKed, see if we can undo a cwnd reduction. */
3530         if (flag & FLAG_DSACKING_ACK)
3531                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3532                                       is_dupack, flag);
3533         /* If this ack opens up a zero window, clear backoff.  It was
3534          * being used to time the probes, and is probably far higher than
3535          * it needs to be for normal retransmission.
3536          */
3537         if (tcp_send_head(sk))
3538                 tcp_ack_probe(sk);
3539 
3540         if (tp->tlp_high_seq)
3541                 tcp_process_tlp_ack(sk, ack, flag);
3542         return 1;
3543 
3544 invalid_ack:
3545         SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3546         return -1;
3547 
3548 old_ack:
3549         /* If data was SACKed, tag it and see if we should send more data.
3550          * If data was DSACKed, see if we can undo a cwnd reduction.
3551          */
3552         if (TCP_SKB_CB(skb)->sacked) {
3553                 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3554                                                 &sack_rtt_us);
3555                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3556                                       is_dupack, flag);
3557         }
3558 
3559         SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3560         return 0;
3561 }
3562 
3563 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3564  * But, this can also be called on packets in the established flow when
3565  * the fast version below fails.
3566  */
3567 void tcp_parse_options(const struct sk_buff *skb,
3568                        struct tcp_options_received *opt_rx, int estab,
3569                        struct tcp_fastopen_cookie *foc)
3570 {
3571         const unsigned char *ptr;
3572         const struct tcphdr *th = tcp_hdr(skb);
3573         int length = (th->doff * 4) - sizeof(struct tcphdr);
3574 
3575         ptr = (const unsigned char *)(th + 1);
3576         opt_rx->saw_tstamp = 0;
3577 
3578         while (length > 0) {
3579                 int opcode = *ptr++;
3580                 int opsize;
3581 
3582                 switch (opcode) {
3583                 case TCPOPT_EOL:
3584                         return;
3585                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3586                         length--;
3587                         continue;
3588                 default:
3589                         opsize = *ptr++;
3590                         if (opsize < 2) /* "silly options" */
3591                                 return;
3592                         if (opsize > length)
3593                                 return; /* don't parse partial options */
3594                         switch (opcode) {
3595                         case TCPOPT_MSS:
3596                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3597                                         u16 in_mss = get_unaligned_be16(ptr);
3598                                         if (in_mss) {
3599                                                 if (opt_rx->user_mss &&
3600                                                     opt_rx->user_mss < in_mss)
3601                                                         in_mss = opt_rx->user_mss;
3602                                                 opt_rx->mss_clamp = in_mss;
3603                                         }
3604                                 }
3605                                 break;
3606                         case TCPOPT_WINDOW:
3607                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3608                                     !estab && sysctl_tcp_window_scaling) {
3609                                         __u8 snd_wscale = *(__u8 *)ptr;
3610                                         opt_rx->wscale_ok = 1;
3611                                         if (snd_wscale > 14) {
3612                                                 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3613                                                                      __func__,
3614                                                                      snd_wscale);
3615                                                 snd_wscale = 14;
3616                                         }
3617                                         opt_rx->snd_wscale = snd_wscale;
3618                                 }
3619                                 break;
3620                         case TCPOPT_TIMESTAMP:
3621                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3622                                     ((estab && opt_rx->tstamp_ok) ||
3623                                      (!estab && sysctl_tcp_timestamps))) {
3624                                         opt_rx->saw_tstamp = 1;
3625                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3626                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3627                                 }
3628                                 break;
3629                         case TCPOPT_SACK_PERM:
3630                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3631                                     !estab && sysctl_tcp_sack) {
3632                                         opt_rx->sack_ok = TCP_SACK_SEEN;
3633                                         tcp_sack_reset(opt_rx);
3634                                 }
3635                                 break;
3636 
3637                         case TCPOPT_SACK:
3638                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3639                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3640                                    opt_rx->sack_ok) {
3641                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3642                                 }
3643                                 break;
3644 #ifdef CONFIG_TCP_MD5SIG
3645                         case TCPOPT_MD5SIG:
3646                                 /*
3647                                  * The MD5 Hash has already been
3648                                  * checked (see tcp_v{4,6}_do_rcv()).
3649                                  */
3650                                 break;
3651 #endif
3652                         case TCPOPT_EXP:
3653                                 /* Fast Open option shares code 254 using a
3654                                  * 16 bits magic number. It's valid only in
3655                                  * SYN or SYN-ACK with an even size.
3656                                  */
3657                                 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3658                                     get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3659                                     foc == NULL || !th->syn || (opsize & 1))
3660                                         break;
3661                                 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3662                                 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3663                                     foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3664                                         memcpy(foc->val, ptr + 2, foc->len);
3665                                 else if (foc->len != 0)
3666                                         foc->len = -1;
3667                                 break;
3668 
3669                         }
3670                         ptr += opsize-2;
3671                         length -= opsize;
3672                 }
3673         }
3674 }
3675 EXPORT_SYMBOL(tcp_parse_options);
3676 
3677 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3678 {
3679         const __be32 *ptr = (const __be32 *)(th + 1);
3680 
3681         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3682                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3683                 tp->rx_opt.saw_tstamp = 1;
3684                 ++ptr;
3685                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3686                 ++ptr;
3687                 if (*ptr)
3688                         tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3689                 else
3690                         tp->rx_opt.rcv_tsecr = 0;
3691                 return true;
3692         }
3693         return false;
3694 }
3695 
3696 /* Fast parse options. This hopes to only see timestamps.
3697  * If it is wrong it falls back on tcp_parse_options().
3698  */
3699 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3700                                    const struct tcphdr *th, struct tcp_sock *tp)
3701 {
3702         /* In the spirit of fast parsing, compare doff directly to constant
3703          * values.  Because equality is used, short doff can be ignored here.
3704          */
3705         if (th->doff == (sizeof(*th) / 4)) {
3706                 tp->rx_opt.saw_tstamp = 0;
3707                 return false;
3708         } else if (tp->rx_opt.tstamp_ok &&
3709                    th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3710                 if (tcp_parse_aligned_timestamp(tp, th))
3711                         return true;
3712         }
3713 
3714         tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3715         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3716                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3717 
3718         return true;
3719 }
3720 
3721 #ifdef CONFIG_TCP_MD5SIG
3722 /*
3723  * Parse MD5 Signature option
3724  */
3725 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3726 {
3727         int length = (th->doff << 2) - sizeof(*th);
3728         const u8 *ptr = (const u8 *)(th + 1);
3729 
3730         /* If the TCP option is too short, we can short cut */
3731         if (length < TCPOLEN_MD5SIG)
3732                 return NULL;
3733 
3734         while (length > 0) {
3735                 int opcode = *ptr++;
3736                 int opsize;
3737 
3738                 switch (opcode) {
3739                 case TCPOPT_EOL:
3740                         return NULL;
3741                 case TCPOPT_NOP:
3742                         length--;
3743                         continue;
3744                 default:
3745                         opsize = *ptr++;
3746                         if (opsize < 2 || opsize > length)
3747                                 return NULL;
3748                         if (opcode == TCPOPT_MD5SIG)
3749                                 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3750                 }
3751                 ptr += opsize - 2;
3752                 length -= opsize;
3753         }
3754         return NULL;
3755 }
3756 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3757 #endif
3758 
3759 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3760  *
3761  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3762  * it can pass through stack. So, the following predicate verifies that
3763  * this segment is not used for anything but congestion avoidance or
3764  * fast retransmit. Moreover, we even are able to eliminate most of such
3765  * second order effects, if we apply some small "replay" window (~RTO)
3766  * to timestamp space.
3767  *
3768  * All these measures still do not guarantee that we reject wrapped ACKs
3769  * on networks with high bandwidth, when sequence space is recycled fastly,
3770  * but it guarantees that such events will be very rare and do not affect
3771  * connection seriously. This doesn't look nice, but alas, PAWS is really
3772  * buggy extension.
3773  *
3774  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3775  * states that events when retransmit arrives after original data are rare.
3776  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3777  * the biggest problem on large power networks even with minor reordering.
3778  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3779  * up to bandwidth of 18Gigabit/sec. 8) ]
3780  */
3781 
3782 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3783 {
3784         const struct tcp_sock *tp = tcp_sk(sk);
3785         const struct tcphdr *th = tcp_hdr(skb);
3786         u32 seq = TCP_SKB_CB(skb)->seq;
3787         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3788 
3789         return (/* 1. Pure ACK with correct sequence number. */
3790                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3791 
3792                 /* 2. ... and duplicate ACK. */
3793                 ack == tp->snd_una &&
3794 
3795                 /* 3. ... and does not update window. */
3796                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3797 
3798                 /* 4. ... and sits in replay window. */
3799                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3800 }
3801 
3802 static inline bool tcp_paws_discard(const struct sock *sk,
3803                                    const struct sk_buff *skb)
3804 {
3805         const struct tcp_sock *tp = tcp_sk(sk);
3806 
3807         return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3808                !tcp_disordered_ack(sk, skb);
3809 }
3810 
3811 /* Check segment sequence number for validity.
3812  *
3813  * Segment controls are considered valid, if the segment
3814  * fits to the window after truncation to the window. Acceptability
3815  * of data (and SYN, FIN, of course) is checked separately.
3816  * See tcp_data_queue(), for example.
3817  *
3818  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3819  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3820  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3821  * (borrowed from freebsd)
3822  */
3823 
3824 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3825 {
3826         return  !before(end_seq, tp->rcv_wup) &&
3827                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3828 }
3829 
3830 /* When we get a reset we do this. */
3831 void tcp_reset(struct sock *sk)
3832 {
3833         /* We want the right error as BSD sees it (and indeed as we do). */
3834         switch (sk->sk_state) {
3835         case TCP_SYN_SENT:
3836                 sk->sk_err = ECONNREFUSED;
3837                 break;
3838         case TCP_CLOSE_WAIT:
3839                 sk->sk_err = EPIPE;
3840                 break;
3841         case TCP_CLOSE:
3842                 return;
3843         default:
3844                 sk->sk_err = ECONNRESET;
3845         }
3846         /* This barrier is coupled with smp_rmb() in tcp_poll() */
3847         smp_wmb();
3848 
3849         if (!sock_flag(sk, SOCK_DEAD))
3850                 sk->sk_error_report(sk);
3851 
3852         tcp_done(sk);
3853 }
3854 
3855 /*
3856  *      Process the FIN bit. This now behaves as it is supposed to work
3857  *      and the FIN takes effect when it is validly part of sequence
3858  *      space. Not before when we get holes.
3859  *
3860  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3861  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3862  *      TIME-WAIT)
3863  *
3864  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3865  *      close and we go into CLOSING (and later onto TIME-WAIT)
3866  *
3867  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3868  */
3869 static void tcp_fin(struct sock *sk)
3870 {
3871         struct tcp_sock *tp = tcp_sk(sk);
3872         const struct dst_entry *dst;
3873 
3874         inet_csk_schedule_ack(sk);
3875 
3876         sk->sk_shutdown |= RCV_SHUTDOWN;
3877         sock_set_flag(sk, SOCK_DONE);
3878 
3879         switch (sk->sk_state) {
3880         case TCP_SYN_RECV:
3881         case TCP_ESTABLISHED:
3882                 /* Move to CLOSE_WAIT */
3883                 tcp_set_state(sk, TCP_CLOSE_WAIT);
3884                 dst = __sk_dst_get(sk);
3885                 if (!dst || !dst_metric(dst, RTAX_QUICKACK))
3886                         inet_csk(sk)->icsk_ack.pingpong = 1;
3887                 break;
3888 
3889         case TCP_CLOSE_WAIT:
3890         case TCP_CLOSING:
3891                 /* Received a retransmission of the FIN, do
3892                  * nothing.
3893                  */
3894                 break;
3895         case TCP_LAST_ACK:
3896                 /* RFC793: Remain in the LAST-ACK state. */
3897                 break;
3898 
3899         case TCP_FIN_WAIT1:
3900                 /* This case occurs when a simultaneous close
3901                  * happens, we must ack the received FIN and
3902                  * enter the CLOSING state.
3903                  */
3904                 tcp_send_ack(sk);
3905                 tcp_set_state(sk, TCP_CLOSING);
3906                 break;
3907         case TCP_FIN_WAIT2:
3908                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3909                 tcp_send_ack(sk);
3910                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3911                 break;
3912         default:
3913                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3914                  * cases we should never reach this piece of code.
3915                  */
3916                 pr_err("%s: Impossible, sk->sk_state=%d\n",
3917                        __func__, sk->sk_state);
3918                 break;
3919         }
3920 
3921         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3922          * Probably, we should reset in this case. For now drop them.
3923          */
3924         __skb_queue_purge(&tp->out_of_order_queue);
3925         if (tcp_is_sack(tp))
3926                 tcp_sack_reset(&tp->rx_opt);
3927         sk_mem_reclaim(sk);
3928 
3929         if (!sock_flag(sk, SOCK_DEAD)) {
3930                 sk->sk_state_change(sk);
3931 
3932                 /* Do not send POLL_HUP for half duplex close. */
3933                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3934                     sk->sk_state == TCP_CLOSE)
3935                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3936                 else
3937                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3938         }
3939 }
3940 
3941 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3942                                   u32 end_seq)
3943 {
3944         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3945                 if (before(seq, sp->start_seq))
3946                         sp->start_seq = seq;
3947                 if (after(end_seq, sp->end_seq))
3948                         sp->end_seq = end_seq;
3949                 return true;
3950         }
3951         return false;
3952 }
3953 
3954 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3955 {
3956         struct tcp_sock *tp = tcp_sk(sk);
3957 
3958         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3959                 int mib_idx;
3960 
3961                 if (before(seq, tp->rcv_nxt))
3962                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3963                 else
3964                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3965 
3966                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3967 
3968                 tp->rx_opt.dsack = 1;
3969                 tp->duplicate_sack[0].start_seq = seq;
3970                 tp->duplicate_sack[0].end_seq = end_seq;
3971         }
3972 }
3973 
3974 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3975 {
3976         struct tcp_sock *tp = tcp_sk(sk);
3977 
3978         if (!tp->rx_opt.dsack)
3979                 tcp_dsack_set(sk, seq, end_seq);
3980         else
3981                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3982 }
3983 
3984 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
3985 {
3986         struct tcp_sock *tp = tcp_sk(sk);
3987 
3988         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3989             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3990                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3991                 tcp_enter_quickack_mode(sk);
3992 
3993                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3994                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3995 
3996                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3997                                 end_seq = tp->rcv_nxt;
3998                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
3999                 }
4000         }
4001 
4002         tcp_send_ack(sk);
4003 }
4004 
4005 /* These routines update the SACK block as out-of-order packets arrive or
4006  * in-order packets close up the sequence space.
4007  */
4008 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4009 {
4010         int this_sack;
4011         struct tcp_sack_block *sp = &tp->selective_acks[0];
4012         struct tcp_sack_block *swalk = sp + 1;
4013 
4014         /* See if the recent change to the first SACK eats into
4015          * or hits the sequence space of other SACK blocks, if so coalesce.
4016          */
4017         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4018                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4019                         int i;
4020 
4021                         /* Zap SWALK, by moving every further SACK up by one slot.
4022                          * Decrease num_sacks.
4023                          */
4024                         tp->rx_opt.num_sacks--;
4025                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4026                                 sp[i] = sp[i + 1];
4027                         continue;
4028                 }
4029                 this_sack++, swalk++;
4030         }
4031 }
4032 
4033 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4034 {
4035         struct tcp_sock *tp = tcp_sk(sk);
4036         struct tcp_sack_block *sp = &tp->selective_acks[0];
4037         int cur_sacks = tp->rx_opt.num_sacks;
4038         int this_sack;
4039 
4040         if (!cur_sacks)
4041                 goto new_sack;
4042 
4043         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4044                 if (tcp_sack_extend(sp, seq, end_seq)) {
4045                         /* Rotate this_sack to the first one. */
4046                         for (; this_sack > 0; this_sack--, sp--)
4047                                 swap(*sp, *(sp - 1));
4048                         if (cur_sacks > 1)
4049                                 tcp_sack_maybe_coalesce(tp);
4050                         return;
4051                 }
4052         }
4053 
4054         /* Could not find an adjacent existing SACK, build a new one,
4055          * put it at the front, and shift everyone else down.  We
4056          * always know there is at least one SACK present already here.
4057          *
4058          * If the sack array is full, forget about the last one.
4059          */
4060         if (this_sack >= TCP_NUM_SACKS) {
4061                 this_sack--;
4062                 tp->rx_opt.num_sacks--;
4063                 sp--;
4064         }
4065         for (; this_sack > 0; this_sack--, sp--)
4066                 *sp = *(sp - 1);
4067 
4068 new_sack:
4069         /* Build the new head SACK, and we're done. */
4070         sp->start_seq = seq;
4071         sp->end_seq = end_seq;
4072         tp->rx_opt.num_sacks++;
4073 }
4074 
4075 /* RCV.NXT advances, some SACKs should be eaten. */
4076 
4077 static void tcp_sack_remove(struct tcp_sock *tp)
4078 {
4079         struct tcp_sack_block *sp = &tp->selective_acks[0];
4080         int num_sacks = tp->rx_opt.num_sacks;
4081         int this_sack;
4082 
4083         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4084         if (skb_queue_empty(&tp->out_of_order_queue)) {
4085                 tp->rx_opt.num_sacks = 0;
4086                 return;
4087         }
4088 
4089         for (this_sack = 0; this_sack < num_sacks;) {
4090                 /* Check if the start of the sack is covered by RCV.NXT. */
4091                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4092                         int i;
4093 
4094                         /* RCV.NXT must cover all the block! */
4095                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4096 
4097                         /* Zap this SACK, by moving forward any other SACKS. */
4098                         for (i = this_sack+1; i < num_sacks; i++)
4099                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4100                         num_sacks--;
4101                         continue;
4102                 }
4103                 this_sack++;
4104                 sp++;
4105         }
4106         tp->rx_opt.num_sacks = num_sacks;
4107 }
4108 
4109 /**
4110  * tcp_try_coalesce - try to merge skb to prior one
4111  * @sk: socket
4112  * @to: prior buffer
4113  * @from: buffer to add in queue
4114  * @fragstolen: pointer to boolean
4115  *
4116  * Before queueing skb @from after @to, try to merge them
4117  * to reduce overall memory use and queue lengths, if cost is small.
4118  * Packets in ofo or receive queues can stay a long time.
4119  * Better try to coalesce them right now to avoid future collapses.
4120  * Returns true if caller should free @from instead of queueing it
4121  */
4122 static bool tcp_try_coalesce(struct sock *sk,
4123                              struct sk_buff *to,
4124                              struct sk_buff *from,
4125                              bool *fragstolen)
4126 {
4127         int delta;
4128 
4129         *fragstolen = false;
4130 
4131         /* Its possible this segment overlaps with prior segment in queue */
4132         if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4133                 return false;
4134 
4135         if (!skb_try_coalesce(to, from, fragstolen, &delta))
4136                 return false;
4137 
4138         atomic_add(delta, &sk->sk_rmem_alloc);
4139         sk_mem_charge(sk, delta);
4140         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4141         TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4142         TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4143         TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4144         return true;
4145 }
4146 
4147 /* This one checks to see if we can put data from the
4148  * out_of_order queue into the receive_queue.
4149  */
4150 static void tcp_ofo_queue(struct sock *sk)
4151 {
4152         struct tcp_sock *tp = tcp_sk(sk);
4153         __u32 dsack_high = tp->rcv_nxt;
4154         struct sk_buff *skb, *tail;
4155         bool fragstolen, eaten;
4156 
4157         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4158                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4159                         break;
4160 
4161                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4162                         __u32 dsack = dsack_high;
4163                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4164                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4165                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4166                 }
4167 
4168                 __skb_unlink(skb, &tp->out_of_order_queue);
4169                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4170                         SOCK_DEBUG(sk, "ofo packet was already received\n");
4171                         __kfree_skb(skb);
4172                         continue;
4173                 }
4174                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4175                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4176                            TCP_SKB_CB(skb)->end_seq);
4177 
4178                 tail = skb_peek_tail(&sk->sk_receive_queue);
4179                 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4180                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4181                 if (!eaten)
4182                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4183                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4184                         tcp_fin(sk);
4185                 if (eaten)
4186                         kfree_skb_partial(skb, fragstolen);
4187         }
4188 }
4189 
4190 static bool tcp_prune_ofo_queue(struct sock *sk);
4191 static int tcp_prune_queue(struct sock *sk);
4192 
4193 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4194                                  unsigned int size)
4195 {
4196         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4197             !sk_rmem_schedule(sk, skb, size)) {
4198 
4199                 if (tcp_prune_queue(sk) < 0)
4200                         return -1;
4201 
4202                 if (!sk_rmem_schedule(sk, skb, size)) {
4203                         if (!tcp_prune_ofo_queue(sk))
4204                                 return -1;
4205 
4206                         if (!sk_rmem_schedule(sk, skb, size))
4207                                 return -1;
4208                 }
4209         }
4210         return 0;
4211 }
4212 
4213 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4214 {
4215         struct tcp_sock *tp = tcp_sk(sk);
4216         struct sk_buff *skb1;
4217         u32 seq, end_seq;
4218 
4219         tcp_ecn_check_ce(tp, skb);
4220 
4221         if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4222                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4223                 __kfree_skb(skb);
4224                 return;
4225         }
4226 
4227         /* Disable header prediction. */
4228         tp->pred_flags = 0;
4229         inet_csk_schedule_ack(sk);
4230 
4231         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4232         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4233                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4234 
4235         skb1 = skb_peek_tail(&tp->out_of_order_queue);
4236         if (!skb1) {
4237                 /* Initial out of order segment, build 1 SACK. */
4238                 if (tcp_is_sack(tp)) {
4239                         tp->rx_opt.num_sacks = 1;
4240                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4241                         tp->selective_acks[0].end_seq =
4242                                                 TCP_SKB_CB(skb)->end_seq;
4243                 }
4244                 __skb_queue_head(&tp->out_of_order_queue, skb);
4245                 goto end;
4246         }
4247 
4248         seq = TCP_SKB_CB(skb)->seq;
4249         end_seq = TCP_SKB_CB(skb)->end_seq;
4250 
4251         if (seq == TCP_SKB_CB(skb1)->end_seq) {
4252                 bool fragstolen;
4253 
4254                 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4255                         __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4256                 } else {
4257                         tcp_grow_window(sk, skb);
4258                         kfree_skb_partial(skb, fragstolen);
4259                         skb = NULL;
4260                 }
4261 
4262                 if (!tp->rx_opt.num_sacks ||
4263                     tp->selective_acks[0].end_seq != seq)
4264                         goto add_sack;
4265 
4266                 /* Common case: data arrive in order after hole. */
4267                 tp->selective_acks[0].end_seq = end_seq;
4268                 goto end;
4269         }
4270 
4271         /* Find place to insert this segment. */
4272         while (1) {
4273                 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4274                         break;
4275                 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4276                         skb1 = NULL;
4277                         break;
4278                 }
4279                 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4280         }
4281 
4282         /* Do skb overlap to previous one? */
4283         if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4284                 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4285                         /* All the bits are present. Drop. */
4286                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4287                         __kfree_skb(skb);
4288                         skb = NULL;
4289                         tcp_dsack_set(sk, seq, end_seq);
4290                         goto add_sack;
4291                 }
4292                 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4293                         /* Partial overlap. */
4294                         tcp_dsack_set(sk, seq,
4295                                       TCP_SKB_CB(skb1)->end_seq);
4296                 } else {
4297                         if (skb_queue_is_first(&tp->out_of_order_queue,
4298                                                skb1))
4299                                 skb1 = NULL;
4300                         else
4301                                 skb1 = skb_queue_prev(
4302                                         &tp->out_of_order_queue,
4303                                         skb1);
4304                 }
4305         }
4306         if (!skb1)
4307                 __skb_queue_head(&tp->out_of_order_queue, skb);
4308         else
4309                 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4310 
4311         /* And clean segments covered by new one as whole. */
4312         while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4313                 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4314 
4315                 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4316                         break;
4317                 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4318                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4319                                          end_seq);
4320                         break;
4321                 }
4322                 __skb_unlink(skb1, &tp->out_of_order_queue);
4323                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4324                                  TCP_SKB_CB(skb1)->end_seq);
4325                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4326                 __kfree_skb(skb1);
4327         }
4328 
4329 add_sack:
4330         if (tcp_is_sack(tp))
4331                 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4332 end:
4333         if (skb) {
4334                 tcp_grow_window(sk, skb);
4335                 skb_set_owner_r(skb, sk);
4336         }
4337 }
4338 
4339 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4340                   bool *fragstolen)
4341 {
4342         int eaten;
4343         struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4344 
4345         __skb_pull(skb, hdrlen);
4346         eaten = (tail &&
4347                  tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4348         tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4349         if (!eaten) {
4350                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4351                 skb_set_owner_r(skb, sk);
4352         }
4353         return eaten;
4354 }
4355 
4356 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4357 {
4358         struct sk_buff *skb;
4359         bool fragstolen;
4360 
4361         if (size == 0)
4362                 return 0;
4363 
4364         skb = alloc_skb(size, sk->sk_allocation);
4365         if (!skb)
4366                 goto err;
4367 
4368         if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4369                 goto err_free;
4370 
4371         if (memcpy_from_msg(skb_put(skb, size), msg, size))
4372                 goto err_free;
4373 
4374         TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4375         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4376         TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4377 
4378         if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4379                 WARN_ON_ONCE(fragstolen); /* should not happen */
4380                 __kfree_skb(skb);
4381         }
4382         return size;
4383 
4384 err_free:
4385         kfree_skb(skb);
4386 err:
4387         return -ENOMEM;
4388 }
4389 
4390 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4391 {
4392         struct tcp_sock *tp = tcp_sk(sk);
4393         int eaten = -1;
4394         bool fragstolen = false;
4395 
4396         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4397                 goto drop;
4398 
4399         skb_dst_drop(skb);
4400         __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4401 
4402         tcp_ecn_accept_cwr(tp, skb);
4403 
4404         tp->rx_opt.dsack = 0;
4405 
4406         /*  Queue data for delivery to the user.
4407          *  Packets in sequence go to the receive queue.
4408          *  Out of sequence packets to the out_of_order_queue.
4409          */
4410         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4411                 if (tcp_receive_window(tp) == 0)
4412                         goto out_of_window;
4413 
4414                 /* Ok. In sequence. In window. */
4415                 if (tp->ucopy.task == current &&
4416                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4417                     sock_owned_by_user(sk) && !tp->urg_data) {
4418                         int chunk = min_t(unsigned int, skb->len,
4419                                           tp->ucopy.len);
4420 
4421                         __set_current_state(TASK_RUNNING);
4422 
4423                         local_bh_enable();
4424                         if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4425                                 tp->ucopy.len -= chunk;
4426                                 tp->copied_seq += chunk;
4427                                 eaten = (chunk == skb->len);
4428                                 tcp_rcv_space_adjust(sk);
4429                         }
4430                         local_bh_disable();
4431                 }
4432 
4433                 if (eaten <= 0) {
4434 queue_and_out:
4435                         if (eaten < 0 &&
4436                             tcp_try_rmem_schedule(sk, skb, skb->truesize))
4437                                 goto drop;
4438 
4439                         eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4440                 }
4441                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4442                 if (skb->len)
4443                         tcp_event_data_recv(sk, skb);
4444                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4445                         tcp_fin(sk);
4446 
4447                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4448                         tcp_ofo_queue(sk);
4449 
4450                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4451                          * gap in queue is filled.
4452                          */
4453                         if (skb_queue_empty(&tp->out_of_order_queue))
4454                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4455                 }
4456 
4457                 if (tp->rx_opt.num_sacks)
4458                         tcp_sack_remove(tp);
4459 
4460                 tcp_fast_path_check(sk);
4461 
4462                 if (eaten > 0)
4463                         kfree_skb_partial(skb, fragstolen);
4464                 if (!sock_flag(sk, SOCK_DEAD))
4465                         sk->sk_data_ready(sk);
4466                 return;
4467         }
4468 
4469         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4470                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4471                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4472                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4473 
4474 out_of_window:
4475                 tcp_enter_quickack_mode(sk);
4476                 inet_csk_schedule_ack(sk);
4477 drop:
4478                 __kfree_skb(skb);
4479                 return;
4480         }
4481 
4482         /* Out of window. F.e. zero window probe. */
4483         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4484                 goto out_of_window;
4485 
4486         tcp_enter_quickack_mode(sk);
4487 
4488         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4489                 /* Partial packet, seq < rcv_next < end_seq */
4490                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4491                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4492                            TCP_SKB_CB(skb)->end_seq);
4493 
4494                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4495 
4496                 /* If window is closed, drop tail of packet. But after
4497                  * remembering D-SACK for its head made in previous line.
4498                  */
4499                 if (!tcp_receive_window(tp))
4500                         goto out_of_window;
4501                 goto queue_and_out;
4502         }
4503 
4504         tcp_data_queue_ofo(sk, skb);
4505 }
4506 
4507 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4508                                         struct sk_buff_head *list)
4509 {
4510         struct sk_buff *next = NULL;
4511 
4512         if (!skb_queue_is_last(list, skb))
4513                 next = skb_queue_next(list, skb);
4514 
4515         __skb_unlink(skb, list);
4516         __kfree_skb(skb);
4517         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4518 
4519         return next;
4520 }
4521 
4522 /* Collapse contiguous sequence of skbs head..tail with
4523  * sequence numbers start..end.
4524  *
4525  * If tail is NULL, this means until the end of the list.
4526  *
4527  * Segments with FIN/SYN are not collapsed (only because this
4528  * simplifies code)
4529  */
4530 static void
4531 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4532              struct sk_buff *head, struct sk_buff *tail,
4533              u32 start, u32 end)
4534 {
4535         struct sk_buff *skb, *n;
4536         bool end_of_skbs;
4537 
4538         /* First, check that queue is collapsible and find
4539          * the point where collapsing can be useful. */
4540         skb = head;
4541 restart:
4542         end_of_skbs = true;
4543         skb_queue_walk_from_safe(list, skb, n) {
4544                 if (skb == tail)
4545                         break;
4546                 /* No new bits? It is possible on ofo queue. */
4547                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4548                         skb = tcp_collapse_one(sk, skb, list);
4549                         if (!skb)
4550                                 break;
4551                         goto restart;
4552                 }
4553 
4554                 /* The first skb to collapse is:
4555                  * - not SYN/FIN and
4556                  * - bloated or contains data before "start" or
4557                  *   overlaps to the next one.
4558                  */
4559                 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4560                     (tcp_win_from_space(skb->truesize) > skb->len ||
4561                      before(TCP_SKB_CB(skb)->seq, start))) {
4562                         end_of_skbs = false;
4563                         break;
4564                 }
4565 
4566                 if (!skb_queue_is_last(list, skb)) {
4567                         struct sk_buff *next = skb_queue_next(list, skb);
4568                         if (next != tail &&
4569                             TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4570                                 end_of_skbs = false;
4571                                 break;
4572                         }
4573                 }
4574 
4575                 /* Decided to skip this, advance start seq. */
4576                 start = TCP_SKB_CB(skb)->end_seq;
4577         }
4578         if (end_of_skbs ||
4579             (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4580                 return;
4581 
4582         while (before(start, end)) {
4583                 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4584                 struct sk_buff *nskb;
4585 
4586                 nskb = alloc_skb(copy, GFP_ATOMIC);
4587                 if (!nskb)
4588                         return;
4589 
4590                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4591                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4592                 __skb_queue_before(list, skb, nskb);
4593                 skb_set_owner_r(nskb, sk);
4594 
4595                 /* Copy data, releasing collapsed skbs. */
4596                 while (copy > 0) {
4597                         int offset = start - TCP_SKB_CB(skb)->seq;
4598                         int size = TCP_SKB_CB(skb)->end_seq - start;
4599 
4600                         BUG_ON(offset < 0);
4601                         if (size > 0) {
4602                                 size = min(copy, size);
4603                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4604                                         BUG();
4605                                 TCP_SKB_CB(nskb)->end_seq += size;
4606                                 copy -= size;
4607                                 start += size;
4608                         }
4609                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4610                                 skb = tcp_collapse_one(sk, skb, list);
4611                                 if (!skb ||
4612                                     skb == tail ||
4613                                     (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4614                                         return;
4615                         }
4616                 }
4617         }
4618 }
4619 
4620 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4621  * and tcp_collapse() them until all the queue is collapsed.
4622  */
4623 static void tcp_collapse_ofo_queue(struct sock *sk)
4624 {
4625         struct tcp_sock *tp = tcp_sk(sk);
4626         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4627         struct sk_buff *head;
4628         u32 start, end;
4629 
4630         if (skb == NULL)
4631                 return;
4632 
4633         start = TCP_SKB_CB(skb)->seq;
4634         end = TCP_SKB_CB(skb)->end_seq;
4635         head = skb;
4636 
4637         for (;;) {
4638                 struct sk_buff *next = NULL;
4639 
4640                 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4641                         next = skb_queue_next(&tp->out_of_order_queue, skb);
4642                 skb = next;
4643 
4644                 /* Segment is terminated when we see gap or when
4645                  * we are at the end of all the queue. */
4646                 if (!skb ||
4647                     after(TCP_SKB_CB(skb)->seq, end) ||
4648                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4649                         tcp_collapse(sk, &tp->out_of_order_queue,
4650                                      head, skb, start, end);
4651                         head = skb;
4652                         if (!skb)
4653                                 break;
4654                         /* Start new segment */
4655                         start = TCP_SKB_CB(skb)->seq;
4656                         end = TCP_SKB_CB(skb)->end_seq;
4657                 } else {
4658                         if (before(TCP_SKB_CB(skb)->seq, start))
4659                                 start = TCP_SKB_CB(skb)->seq;
4660                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4661                                 end = TCP_SKB_CB(skb)->end_seq;
4662                 }
4663         }
4664 }
4665 
4666 /*
4667  * Purge the out-of-order queue.
4668  * Return true if queue was pruned.
4669  */
4670 static bool tcp_prune_ofo_queue(struct sock *sk)
4671 {
4672         struct tcp_sock *tp = tcp_sk(sk);
4673         bool res = false;
4674 
4675         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4676                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4677                 __skb_queue_purge(&tp->out_of_order_queue);
4678 
4679                 /* Reset SACK state.  A conforming SACK implementation will
4680                  * do the same at a timeout based retransmit.  When a connection
4681                  * is in a sad state like this, we care only about integrity
4682                  * of the connection not performance.
4683                  */
4684                 if (tp->rx_opt.sack_ok)
4685                         tcp_sack_reset(&tp->rx_opt);
4686                 sk_mem_reclaim(sk);
4687                 res = true;
4688         }
4689         return res;
4690 }
4691 
4692 /* Reduce allocated memory if we can, trying to get
4693  * the socket within its memory limits again.
4694  *
4695  * Return less than zero if we should start dropping frames
4696  * until the socket owning process reads some of the data
4697  * to stabilize the situation.
4698  */
4699 static int tcp_prune_queue(struct sock *sk)
4700 {
4701         struct tcp_sock *tp = tcp_sk(sk);
4702 
4703         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4704 
4705         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4706 
4707         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4708                 tcp_clamp_window(sk);
4709         else if (sk_under_memory_pressure(sk))
4710                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4711 
4712         tcp_collapse_ofo_queue(sk);
4713         if (!skb_queue_empty(&sk->sk_receive_queue))
4714                 tcp_collapse(sk, &sk->sk_receive_queue,
4715                              skb_peek(&sk->sk_receive_queue),
4716                              NULL,
4717                              tp->copied_seq, tp->rcv_nxt);
4718         sk_mem_reclaim(sk);
4719 
4720         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4721                 return 0;
4722 
4723         /* Collapsing did not help, destructive actions follow.
4724          * This must not ever occur. */
4725 
4726         tcp_prune_ofo_queue(sk);
4727 
4728         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4729                 return 0;
4730 
4731         /* If we are really being abused, tell the caller to silently
4732          * drop receive data on the floor.  It will get retransmitted
4733          * and hopefully then we'll have sufficient space.
4734          */
4735         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4736 
4737         /* Massive buffer overcommit. */
4738         tp->pred_flags = 0;
4739         return -1;
4740 }
4741 
4742 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4743 {
4744         const struct tcp_sock *tp = tcp_sk(sk);
4745 
4746         /* If the user specified a specific send buffer setting, do
4747          * not modify it.
4748          */
4749         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4750                 return false;
4751 
4752         /* If we are under global TCP memory pressure, do not expand.  */
4753         if (sk_under_memory_pressure(sk))
4754                 return false;
4755 
4756         /* If we are under soft global TCP memory pressure, do not expand.  */
4757         if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4758                 return false;
4759 
4760         /* If we filled the congestion window, do not expand.  */
4761         if (tp->packets_out >= tp->snd_cwnd)
4762                 return false;
4763 
4764         return true;
4765 }
4766 
4767 /* When incoming ACK allowed to free some skb from write_queue,
4768  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4769  * on the exit from tcp input handler.
4770  *
4771  * PROBLEM: sndbuf expansion does not work well with largesend.
4772  */
4773 static void tcp_new_space(struct sock *sk)
4774 {
4775         struct tcp_sock *tp = tcp_sk(sk);
4776 
4777         if (tcp_should_expand_sndbuf(sk)) {
4778                 tcp_sndbuf_expand(sk);
4779                 tp->snd_cwnd_stamp = tcp_time_stamp;
4780         }
4781 
4782         sk->sk_write_space(sk);
4783 }
4784 
4785 static void tcp_check_space(struct sock *sk)
4786 {
4787         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4788                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4789                 if (sk->sk_socket &&
4790                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4791                         tcp_new_space(sk);
4792         }
4793 }
4794 
4795 static inline void tcp_data_snd_check(struct sock *sk)
4796 {
4797         tcp_push_pending_frames(sk);
4798         tcp_check_space(sk);
4799 }
4800 
4801 /*
4802  * Check if sending an ack is needed.
4803  */
4804 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4805 {
4806         struct tcp_sock *tp = tcp_sk(sk);
4807 
4808             /* More than one full frame received... */
4809         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4810              /* ... and right edge of window advances far enough.
4811               * (tcp_recvmsg() will send ACK otherwise). Or...
4812               */
4813              __tcp_select_window(sk) >= tp->rcv_wnd) ||
4814             /* We ACK each frame or... */
4815             tcp_in_quickack_mode(sk) ||
4816             /* We have out of order data. */
4817             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4818                 /* Then ack it now */
4819                 tcp_send_ack(sk);
4820         } else {
4821                 /* Else, send delayed ack. */
4822                 tcp_send_delayed_ack(sk);
4823         }
4824 }
4825 
4826 static inline void tcp_ack_snd_check(struct sock *sk)
4827 {
4828         if (!inet_csk_ack_scheduled(sk)) {
4829                 /* We sent a data segment already. */
4830                 return;
4831         }
4832         __tcp_ack_snd_check(sk, 1);
4833 }
4834 
4835 /*
4836  *      This routine is only called when we have urgent data
4837  *      signaled. Its the 'slow' part of tcp_urg. It could be
4838  *      moved inline now as tcp_urg is only called from one
4839  *      place. We handle URGent data wrong. We have to - as
4840  *      BSD still doesn't use the correction from RFC961.
4841  *      For 1003.1g we should support a new option TCP_STDURG to permit
4842  *      either form (or just set the sysctl tcp_stdurg).
4843  */
4844 
4845 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4846 {
4847         struct tcp_sock *tp = tcp_sk(sk);
4848         u32 ptr = ntohs(th->urg_ptr);
4849 
4850         if (ptr && !sysctl_tcp_stdurg)
4851                 ptr--;
4852         ptr += ntohl(th->seq);
4853 
4854         /* Ignore urgent data that we've already seen and read. */
4855         if (after(tp->copied_seq, ptr))
4856                 return;
4857 
4858         /* Do not replay urg ptr.
4859          *
4860          * NOTE: interesting situation not covered by specs.
4861          * Misbehaving sender may send urg ptr, pointing to segment,
4862          * which we already have in ofo queue. We are not able to fetch
4863          * such data and will stay in TCP_URG_NOTYET until will be eaten
4864          * by recvmsg(). Seems, we are not obliged to handle such wicked
4865          * situations. But it is worth to think about possibility of some
4866          * DoSes using some hypothetical application level deadlock.
4867          */
4868         if (before(ptr, tp->rcv_nxt))
4869                 return;
4870 
4871         /* Do we already have a newer (or duplicate) urgent pointer? */
4872         if (tp->urg_data && !after(ptr, tp->urg_seq))
4873                 return;
4874 
4875         /* Tell the world about our new urgent pointer. */
4876         sk_send_sigurg(sk);
4877 
4878         /* We may be adding urgent data when the last byte read was
4879          * urgent. To do this requires some care. We cannot just ignore
4880          * tp->copied_seq since we would read the last urgent byte again
4881          * as data, nor can we alter copied_seq until this data arrives
4882          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4883          *
4884          * NOTE. Double Dutch. Rendering to plain English: author of comment
4885          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4886          * and expect that both A and B disappear from stream. This is _wrong_.
4887          * Though this happens in BSD with high probability, this is occasional.
4888          * Any application relying on this is buggy. Note also, that fix "works"
4889          * only in this artificial test. Insert some normal data between A and B and we will
4890          * decline of BSD again. Verdict: it is better to remove to trap
4891          * buggy users.
4892          */
4893         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4894             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4895                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4896                 tp->copied_seq++;
4897                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4898                         __skb_unlink(skb, &sk->sk_receive_queue);
4899                         __kfree_skb(skb);
4900                 }
4901         }
4902 
4903         tp->urg_data = TCP_URG_NOTYET;
4904         tp->urg_seq = ptr;
4905 
4906         /* Disable header prediction. */
4907         tp->pred_flags = 0;
4908 }
4909 
4910 /* This is the 'fast' part of urgent handling. */
4911 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4912 {
4913         struct tcp_sock *tp = tcp_sk(sk);
4914 
4915         /* Check if we get a new urgent pointer - normally not. */
4916         if (th->urg)
4917                 tcp_check_urg(sk, th);
4918 
4919         /* Do we wait for any urgent data? - normally not... */
4920         if (tp->urg_data == TCP_URG_NOTYET) {
4921                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4922                           th->syn;
4923 
4924                 /* Is the urgent pointer pointing into this packet? */
4925                 if (ptr < skb->len) {
4926                         u8 tmp;
4927                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4928                                 BUG();
4929                         tp->urg_data = TCP_URG_VALID | tmp;
4930                         if (!sock_flag(sk, SOCK_DEAD))
4931                                 sk->sk_data_ready(sk);
4932                 }
4933         }
4934 }
4935 
4936 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4937 {
4938         struct tcp_sock *tp = tcp_sk(sk);
4939         int chunk = skb->len - hlen;
4940         int err;
4941 
4942         local_bh_enable();
4943         if (skb_csum_unnecessary(skb))
4944                 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
4945         else
4946                 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
4947 
4948         if (!err) {
4949                 tp->ucopy.len -= chunk;
4950                 tp->copied_seq += chunk;
4951                 tcp_rcv_space_adjust(sk);
4952         }
4953 
4954         local_bh_disable();
4955         return err;
4956 }
4957 
4958 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4959                                             struct sk_buff *skb)
4960 {
4961         __sum16 result;
4962 
4963         if (sock_owned_by_user(sk)) {
4964                 local_bh_enable();
4965                 result = __tcp_checksum_complete(skb);
4966                 local_bh_disable();
4967         } else {
4968                 result = __tcp_checksum_complete(skb);
4969         }
4970         return result;
4971 }
4972 
4973 static inline bool tcp_checksum_complete_user(struct sock *sk,
4974                                              struct sk_buff *skb)
4975 {
4976         return !skb_csum_unnecessary(skb) &&
4977                __tcp_checksum_complete_user(sk, skb);
4978 }
4979 
4980 /* Does PAWS and seqno based validation of an incoming segment, flags will
4981  * play significant role here.
4982  */
4983 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
4984                                   const struct tcphdr *th, int syn_inerr)
4985 {
4986         struct tcp_sock *tp = tcp_sk(sk);
4987 
4988         /* RFC1323: H1. Apply PAWS check first. */
4989         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4990             tcp_paws_discard(sk, skb)) {
4991                 if (!th->rst) {
4992                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
4993                         tcp_send_dupack(sk, skb);
4994                         goto discard;
4995                 }
4996                 /* Reset is accepted even if it did not pass PAWS. */
4997         }
4998 
4999         /* Step 1: check sequence number */
5000         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5001                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5002                  * (RST) segments are validated by checking their SEQ-fields."
5003                  * And page 69: "If an incoming segment is not acceptable,
5004                  * an acknowledgment should be sent in reply (unless the RST
5005                  * bit is set, if so drop the segment and return)".
5006                  */
5007                 if (!th->rst) {
5008                         if (th->syn)
5009                                 goto syn_challenge;
5010                         tcp_send_dupack(sk, skb);
5011                 }
5012                 goto discard;
5013         }
5014 
5015         /* Step 2: check RST bit */
5016         if (th->rst) {
5017                 /* RFC 5961 3.2 :
5018                  * If sequence number exactly matches RCV.NXT, then
5019                  *     RESET the connection
5020                  * else
5021                  *     Send a challenge ACK
5022                  */
5023                 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5024                         tcp_reset(sk);
5025                 else
5026                         tcp_send_challenge_ack(sk);
5027                 goto discard;
5028         }
5029 
5030         /* step 3: check security and precedence [ignored] */
5031 
5032         /* step 4: Check for a SYN
5033          * RFC 5961 4.2 : Send a challenge ack
5034          */
5035         if (th->syn) {
5036 syn_challenge:
5037                 if (syn_inerr)
5038                         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5039                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5040                 tcp_send_challenge_ack(sk);
5041                 goto discard;
5042         }
5043 
5044         return true;
5045 
5046 discard:
5047         __kfree_skb(skb);
5048         return false;
5049 }
5050 
5051 /*
5052  *      TCP receive function for the ESTABLISHED state.
5053  *
5054  *      It is split into a fast path and a slow path. The fast path is
5055  *      disabled when:
5056  *      - A zero window was announced from us - zero window probing
5057  *        is only handled properly in the slow path.
5058  *      - Out of order segments arrived.
5059  *      - Urgent data is expected.
5060  *      - There is no buffer space left
5061  *      - Unexpected TCP flags/window values/header lengths are received
5062  *        (detected by checking the TCP header against pred_flags)
5063  *      - Data is sent in both directions. Fast path only supports pure senders
5064  *        or pure receivers (this means either the sequence number or the ack
5065  *        value must stay constant)
5066  *      - Unexpected TCP option.
5067  *
5068  *      When these conditions are not satisfied it drops into a standard
5069  *      receive procedure patterned after RFC793 to handle all cases.
5070  *      The first three cases are guaranteed by proper pred_flags setting,
5071  *      the rest is checked inline. Fast processing is turned on in
5072  *      tcp_data_queue when everything is OK.
5073  */
5074 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5075                          const struct tcphdr *th, unsigned int len)
5076 {
5077         struct tcp_sock *tp = tcp_sk(sk);
5078 
5079         if (unlikely(sk->sk_rx_dst == NULL))
5080                 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5081         /*
5082          *      Header prediction.
5083          *      The code loosely follows the one in the famous
5084          *      "30 instruction TCP receive" Van Jacobson mail.
5085          *
5086          *      Van's trick is to deposit buffers into socket queue
5087          *      on a device interrupt, to call tcp_recv function
5088          *      on the receive process context and checksum and copy
5089          *      the buffer to user space. smart...
5090          *
5091          *      Our current scheme is not silly either but we take the
5092          *      extra cost of the net_bh soft interrupt processing...
5093          *      We do checksum and copy also but from device to kernel.
5094          */
5095 
5096         tp->rx_opt.saw_tstamp = 0;
5097 
5098         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5099          *      if header_prediction is to be made
5100          *      'S' will always be tp->tcp_header_len >> 2
5101          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5102          *  turn it off (when there are holes in the receive
5103          *       space for instance)
5104          *      PSH flag is ignored.
5105          */
5106 
5107         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5108             TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5109             !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5110                 int tcp_header_len = tp->tcp_header_len;
5111 
5112                 /* Timestamp header prediction: tcp_header_len
5113                  * is automatically equal to th->doff*4 due to pred_flags
5114                  * match.
5115                  */
5116 
5117                 /* Check timestamp */
5118                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5119                         /* No? Slow path! */
5120                         if (!tcp_parse_aligned_timestamp(tp, th))
5121                                 goto slow_path;
5122 
5123                         /* If PAWS failed, check it more carefully in slow path */
5124                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5125                                 goto slow_path;
5126 
5127                         /* DO NOT update ts_recent here, if checksum fails
5128                          * and timestamp was corrupted part, it will result
5129                          * in a hung connection since we will drop all
5130                          * future packets due to the PAWS test.
5131                          */
5132                 }
5133 
5134                 if (len <= tcp_header_len) {
5135                         /* Bulk data transfer: sender */
5136                         if (len == tcp_header_len) {
5137                                 /* Predicted packet is in window by definition.
5138                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5139                                  * Hence, check seq<=rcv_wup reduces to:
5140                                  */
5141                                 if (tcp_header_len ==
5142                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5143                                     tp->rcv_nxt == tp->rcv_wup)
5144                                         tcp_store_ts_recent(tp);
5145 
5146                                 /* We know that such packets are checksummed
5147                                  * on entry.
5148                                  */
5149                                 tcp_ack(sk, skb, 0);
5150                                 __kfree_skb(skb);
5151                                 tcp_data_snd_check(sk);
5152                                 return;
5153                         } else { /* Header too small */
5154                                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5155                                 goto discard;
5156                         }
5157                 } else {
5158                         int eaten = 0;
5159                         bool fragstolen = false;
5160 
5161                         if (tp->ucopy.task == current &&
5162                             tp->copied_seq == tp->rcv_nxt &&
5163                             len - tcp_header_len <= tp->ucopy.len &&
5164                             sock_owned_by_user(sk)) {
5165                                 __set_current_state(TASK_RUNNING);
5166 
5167                                 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5168                                         /* Predicted packet is in window by definition.
5169                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5170                                          * Hence, check seq<=rcv_wup reduces to:
5171                                          */
5172                                         if (tcp_header_len ==
5173                                             (sizeof(struct tcphdr) +
5174                                              TCPOLEN_TSTAMP_ALIGNED) &&
5175                                             tp->rcv_nxt == tp->rcv_wup)
5176                                                 tcp_store_ts_recent(tp);
5177 
5178                                         tcp_rcv_rtt_measure_ts(sk, skb);
5179 
5180                                         __skb_pull(skb, tcp_header_len);
5181                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5182                                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5183                                         eaten = 1;
5184                                 }
5185                         }
5186                         if (!eaten) {
5187                                 if (tcp_checksum_complete_user(sk, skb))
5188                                         goto csum_error;
5189 
5190                                 if ((int)skb->truesize > sk->sk_forward_alloc)
5191                                         goto step5;
5192 
5193                                 /* Predicted packet is in window by definition.
5194                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5195                                  * Hence, check seq<=rcv_wup reduces to:
5196                                  */
5197                                 if (tcp_header_len ==
5198                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5199                                     tp->rcv_nxt == tp->rcv_wup)
5200                                         tcp_store_ts_recent(tp);
5201 
5202                                 tcp_rcv_rtt_measure_ts(sk, skb);
5203 
5204                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5205 
5206                                 /* Bulk data transfer: receiver */
5207                                 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5208                                                       &fragstolen);
5209                         }
5210 
5211                         tcp_event_data_recv(sk, skb);
5212 
5213                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5214                                 /* Well, only one small jumplet in fast path... */
5215                                 tcp_ack(sk, skb, FLAG_DATA);
5216                                 tcp_data_snd_check(sk);
5217                                 if (!inet_csk_ack_scheduled(sk))
5218                                         goto no_ack;
5219                         }
5220 
5221                         __tcp_ack_snd_check(sk, 0);
5222 no_ack:
5223                         if (eaten)
5224                                 kfree_skb_partial(skb, fragstolen);
5225                         sk->sk_data_ready(sk);
5226                         return;
5227                 }
5228         }
5229 
5230 slow_path:
5231         if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5232                 goto csum_error;
5233 
5234         if (!th->ack && !th->rst && !th->syn)
5235                 goto discard;
5236 
5237         /*
5238          *      Standard slow path.
5239          */
5240 
5241         if (!tcp_validate_incoming(sk, skb, th, 1))
5242                 return;
5243 
5244 step5:
5245         if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5246                 goto discard;
5247 
5248         tcp_rcv_rtt_measure_ts(sk, skb);
5249 
5250         /* Process urgent data. */
5251         tcp_urg(sk, skb, th);
5252 
5253         /* step 7: process the segment text */
5254         tcp_data_queue(sk, skb);
5255 
5256         tcp_data_snd_check(sk);
5257         tcp_ack_snd_check(sk);
5258         return;
5259 
5260 csum_error:
5261         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5262         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5263 
5264 discard:
5265         __kfree_skb(skb);
5266 }
5267 EXPORT_SYMBOL(tcp_rcv_established);
5268 
5269 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5270 {
5271         struct tcp_sock *tp = tcp_sk(sk);
5272         struct inet_connection_sock *icsk = inet_csk(sk);
5273 
5274         tcp_set_state(sk, TCP_ESTABLISHED);
5275 
5276         if (skb != NULL) {
5277                 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5278                 security_inet_conn_established(sk, skb);
5279         }
5280 
5281         /* Make sure socket is routed, for correct metrics.  */
5282         icsk->icsk_af_ops->rebuild_header(sk);
5283 
5284         tcp_init_metrics(sk);
5285 
5286         tcp_init_congestion_control(sk);
5287 
5288         /* Prevent spurious tcp_cwnd_restart() on first data
5289          * packet.
5290          */
5291         tp->lsndtime = tcp_time_stamp;
5292 
5293         tcp_init_buffer_space(sk);
5294 
5295         if (sock_flag(sk, SOCK_KEEPOPEN))
5296                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5297 
5298         if (!tp->rx_opt.snd_wscale)
5299                 __tcp_fast_path_on(tp, tp->snd_wnd);
5300         else
5301                 tp->pred_flags = 0;
5302 
5303         if (!sock_flag(sk, SOCK_DEAD)) {
5304                 sk->sk_state_change(sk);
5305                 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5306         }
5307 }
5308 
5309 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5310                                     struct tcp_fastopen_cookie *cookie)
5311 {
5312         struct tcp_sock *tp = tcp_sk(sk);
5313         struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5314         u16 mss = tp->rx_opt.mss_clamp;
5315         bool syn_drop;
5316 
5317         if (mss == tp->rx_opt.user_mss) {
5318                 struct tcp_options_received opt;
5319 
5320                 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5321                 tcp_clear_options(&opt);
5322                 opt.user_mss = opt.mss_clamp = 0;
5323                 tcp_parse_options(synack, &opt, 0, NULL);
5324                 mss = opt.mss_clamp;
5325         }
5326 
5327         if (!tp->syn_fastopen)  /* Ignore an unsolicited cookie */
5328                 cookie->len = -1;
5329 
5330         /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5331          * the remote receives only the retransmitted (regular) SYNs: either
5332          * the original SYN-data or the corresponding SYN-ACK is lost.
5333          */
5334         syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
5335 
5336         tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
5337 
5338         if (data) { /* Retransmit unacked data in SYN */
5339                 tcp_for_write_queue_from(data, sk) {
5340                         if (data == tcp_send_head(sk) ||
5341                             __tcp_retransmit_skb(sk, data))
5342                                 break;
5343                 }
5344                 tcp_rearm_rto(sk);
5345                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5346                 return true;
5347         }
5348         tp->syn_data_acked = tp->syn_data;
5349         if (tp->syn_data_acked)
5350                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5351         return false;
5352 }
5353 
5354 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5355                                          const struct tcphdr *th, unsigned int len)
5356 {
5357         struct inet_connection_sock *icsk = inet_csk(sk);
5358         struct tcp_sock *tp = tcp_sk(sk);
5359         struct tcp_fastopen_cookie foc = { .len = -1 };
5360         int saved_clamp = tp->rx_opt.mss_clamp;
5361 
5362         tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5363         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5364                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5365 
5366         if (th->ack) {
5367                 /* rfc793:
5368                  * "If the state is SYN-SENT then
5369                  *    first check the ACK bit
5370                  *      If the ACK bit is set
5371                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5372                  *        a reset (unless the RST bit is set, if so drop
5373                  *        the segment and return)"
5374                  */
5375                 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5376                     after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5377                         goto reset_and_undo;
5378 
5379                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5380                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5381                              tcp_time_stamp)) {
5382                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5383                         goto reset_and_undo;
5384                 }
5385 
5386                 /* Now ACK is acceptable.
5387                  *
5388                  * "If the RST bit is set
5389                  *    If the ACK was acceptable then signal the user "error:
5390                  *    connection reset", drop the segment, enter CLOSED state,
5391                  *    delete TCB, and return."
5392                  */
5393 
5394                 if (th->rst) {
5395                         tcp_reset(sk);
5396                         goto discard;
5397                 }
5398 
5399                 /* rfc793:
5400                  *   "fifth, if neither of the SYN or RST bits is set then
5401                  *    drop the segment and return."
5402                  *
5403                  *    See note below!
5404                  *                                        --ANK(990513)
5405                  */
5406                 if (!th->syn)
5407                         goto discard_and_undo;
5408 
5409                 /* rfc793:
5410                  *   "If the SYN bit is on ...
5411                  *    are acceptable then ...
5412                  *    (our SYN has been ACKed), change the connection
5413                  *    state to ESTABLISHED..."
5414                  */
5415 
5416                 tcp_ecn_rcv_synack(tp, th);
5417 
5418                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5419                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5420 
5421                 /* Ok.. it's good. Set up sequence numbers and
5422                  * move to established.
5423                  */
5424                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5425                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5426 
5427                 /* RFC1323: The window in SYN & SYN/ACK segments is
5428                  * never scaled.
5429                  */
5430                 tp->snd_wnd = ntohs(th->window);
5431 
5432                 if (!tp->rx_opt.wscale_ok) {
5433                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5434                         tp->window_clamp = min(tp->window_clamp, 65535U);
5435                 }
5436 
5437                 if (tp->rx_opt.saw_tstamp) {
5438                         tp->rx_opt.tstamp_ok       = 1;
5439                         tp->tcp_header_len =
5440                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5441                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5442                         tcp_store_ts_recent(tp);
5443                 } else {
5444                         tp->tcp_header_len = sizeof(struct tcphdr);
5445                 }
5446 
5447                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5448                         tcp_enable_fack(tp);
5449 
5450                 tcp_mtup_init(sk);
5451                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5452                 tcp_initialize_rcv_mss(sk);
5453 
5454                 /* Remember, tcp_poll() does not lock socket!
5455                  * Change state from SYN-SENT only after copied_seq
5456                  * is initialized. */
5457                 tp->copied_seq = tp->rcv_nxt;
5458 
5459                 smp_mb();
5460 
5461                 tcp_finish_connect(sk, skb);
5462 
5463                 if ((tp->syn_fastopen || tp->syn_data) &&
5464                     tcp_rcv_fastopen_synack(sk, skb, &foc))
5465                         return -1;
5466 
5467                 if (sk->sk_write_pending ||
5468                     icsk->icsk_accept_queue.rskq_defer_accept ||
5469                     icsk->icsk_ack.pingpong) {
5470                         /* Save one ACK. Data will be ready after
5471                          * several ticks, if write_pending is set.
5472                          *
5473                          * It may be deleted, but with this feature tcpdumps
5474                          * look so _wonderfully_ clever, that I was not able
5475                          * to stand against the temptation 8)     --ANK
5476                          */
5477                         inet_csk_schedule_ack(sk);
5478                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5479                         tcp_enter_quickack_mode(sk);
5480                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5481                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5482 
5483 discard:
5484                         __kfree_skb(skb);
5485                         return 0;
5486                 } else {
5487                         tcp_send_ack(sk);
5488                 }
5489                 return -1;
5490         }
5491 
5492         /* No ACK in the segment */
5493 
5494         if (th->rst) {
5495                 /* rfc793:
5496                  * "If the RST bit is set
5497                  *
5498                  *      Otherwise (no ACK) drop the segment and return."
5499                  */
5500 
5501                 goto discard_and_undo;
5502         }
5503 
5504         /* PAWS check. */
5505         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5506             tcp_paws_reject(&tp->rx_opt, 0))
5507                 goto discard_and_undo;
5508 
5509         if (th->syn) {
5510                 /* We see SYN without ACK. It is attempt of
5511                  * simultaneous connect with crossed SYNs.
5512                  * Particularly, it can be connect to self.
5513                  */
5514                 tcp_set_state(sk, TCP_SYN_RECV);
5515 
5516                 if (tp->rx_opt.saw_tstamp) {
5517                         tp->rx_opt.tstamp_ok = 1;
5518                         tcp_store_ts_recent(tp);
5519                         tp->tcp_header_len =
5520                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5521                 } else {
5522                         tp->tcp_header_len = sizeof(struct tcphdr);
5523                 }
5524 
5525                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5526                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5527 
5528                 /* RFC1323: The window in SYN & SYN/ACK segments is
5529                  * never scaled.
5530                  */
5531                 tp->snd_wnd    = ntohs(th->window);
5532                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5533                 tp->max_window = tp->snd_wnd;
5534 
5535                 tcp_ecn_rcv_syn(tp, th);
5536 
5537                 tcp_mtup_init(sk);
5538                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5539                 tcp_initialize_rcv_mss(sk);
5540 
5541                 tcp_send_synack(sk);
5542 #if 0
5543                 /* Note, we could accept data and URG from this segment.
5544                  * There are no obstacles to make this (except that we must
5545                  * either change tcp_recvmsg() to prevent it from returning data
5546                  * before 3WHS completes per RFC793, or employ TCP Fast Open).
5547                  *
5548                  * However, if we ignore data in ACKless segments sometimes,
5549                  * we have no reasons to accept it sometimes.
5550                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5551                  * is not flawless. So, discard packet for sanity.
5552                  * Uncomment this return to process the data.
5553                  */
5554                 return -1;
5555 #else
5556                 goto discard;
5557 #endif
5558         }
5559         /* "fifth, if neither of the SYN or RST bits is set then
5560          * drop the segment and return."
5561          */
5562 
5563 discard_and_undo:
5564         tcp_clear_options(&tp->rx_opt);
5565         tp->rx_opt.mss_clamp = saved_clamp;
5566         goto discard;
5567 
5568 reset_and_undo:
5569         tcp_clear_options(&tp->rx_opt);
5570         tp->rx_opt.mss_clamp = saved_clamp;
5571         return 1;
5572 }
5573 
5574 /*
5575  *      This function implements the receiving procedure of RFC 793 for
5576  *      all states except ESTABLISHED and TIME_WAIT.
5577  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5578  *      address independent.
5579  */
5580 
5581 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5582                           const struct tcphdr *th, unsigned int len)
5583 {
5584         struct tcp_sock *tp = tcp_sk(sk);
5585         struct inet_connection_sock *icsk = inet_csk(sk);
5586         struct request_sock *req;
5587         int queued = 0;
5588         bool acceptable;
5589         u32 synack_stamp;
5590 
5591         tp->rx_opt.saw_tstamp = 0;
5592 
5593         switch (sk->sk_state) {
5594         case TCP_CLOSE:
5595                 goto discard;
5596 
5597         case TCP_LISTEN:
5598                 if (th->ack)
5599                         return 1;
5600 
5601                 if (th->rst)
5602                         goto discard;
5603 
5604                 if (th->syn) {
5605                         if (th->fin)
5606                                 goto discard;
5607                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5608                                 return 1;
5609 
5610                         /* Now we have several options: In theory there is
5611                          * nothing else in the frame. KA9Q has an option to
5612                          * send data with the syn, BSD accepts data with the
5613                          * syn up to the [to be] advertised window and
5614                          * Solaris 2.1 gives you a protocol error. For now
5615                          * we just ignore it, that fits the spec precisely
5616                          * and avoids incompatibilities. It would be nice in
5617                          * future to drop through and process the data.
5618                          *
5619                          * Now that TTCP is starting to be used we ought to
5620                          * queue this data.
5621                          * But, this leaves one open to an easy denial of
5622                          * service attack, and SYN cookies can't defend
5623                          * against this problem. So, we drop the data
5624                          * in the interest of security over speed unless
5625                          * it's still in use.
5626                          */
5627                         kfree_skb(skb);
5628                         return 0;
5629                 }
5630                 goto discard;
5631 
5632         case TCP_SYN_SENT:
5633                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5634                 if (queued >= 0)
5635                         return queued;
5636 
5637                 /* Do step6 onward by hand. */
5638                 tcp_urg(sk, skb, th);
5639                 __kfree_skb(skb);
5640                 tcp_data_snd_check(sk);
5641                 return 0;
5642         }
5643 
5644         req = tp->fastopen_rsk;
5645         if (req != NULL) {
5646                 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5647                     sk->sk_state != TCP_FIN_WAIT1);
5648 
5649                 if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
5650                         goto discard;
5651         }
5652 
5653         if (!th->ack && !th->rst && !th->syn)
5654                 goto discard;
5655 
5656         if (!tcp_validate_incoming(sk, skb, th, 0))
5657                 return 0;
5658 
5659         /* step 5: check the ACK field */
5660         acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5661                                       FLAG_UPDATE_TS_RECENT) > 0;
5662 
5663         switch (sk->sk_state) {
5664         case TCP_SYN_RECV:
5665                 if (!acceptable)
5666                         return 1;
5667 
5668                 /* Once we leave TCP_SYN_RECV, we no longer need req
5669                  * so release it.
5670                  */
5671                 if (req) {
5672                         synack_stamp = tcp_rsk(req)->snt_synack;
5673                         tp->total_retrans = req->num_retrans;
5674                         reqsk_fastopen_remove(sk, req, false);
5675                 } else {
5676                         synack_stamp = tp->lsndtime;
5677                         /* Make sure socket is routed, for correct metrics. */
5678                         icsk->icsk_af_ops->rebuild_header(sk);
5679                         tcp_init_congestion_control(sk);
5680 
5681                         tcp_mtup_init(sk);
5682                         tp->copied_seq = tp->rcv_nxt;
5683                         tcp_init_buffer_space(sk);
5684                 }
5685                 smp_mb();
5686                 tcp_set_state(sk, TCP_ESTABLISHED);
5687                 sk->sk_state_change(sk);
5688 
5689                 /* Note, that this wakeup is only for marginal crossed SYN case.
5690                  * Passively open sockets are not waked up, because
5691                  * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5692                  */
5693                 if (sk->sk_socket)
5694                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5695 
5696                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5697                 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5698                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5699                 tcp_synack_rtt_meas(sk, synack_stamp);
5700 
5701                 if (tp->rx_opt.tstamp_ok)
5702                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5703 
5704                 if (req) {
5705                         /* Re-arm the timer because data may have been sent out.
5706                          * This is similar to the regular data transmission case
5707                          * when new data has just been ack'ed.
5708                          *
5709                          * (TFO) - we could try to be more aggressive and
5710                          * retransmitting any data sooner based on when they
5711                          * are sent out.
5712                          */
5713                         tcp_rearm_rto(sk);
5714                 } else
5715                         tcp_init_metrics(sk);
5716 
5717                 tcp_update_pacing_rate(sk);
5718 
5719                 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5720                 tp->lsndtime = tcp_time_stamp;
5721 
5722                 tcp_initialize_rcv_mss(sk);
5723                 tcp_fast_path_on(tp);
5724                 break;
5725 
5726         case TCP_FIN_WAIT1: {
5727                 struct dst_entry *dst;
5728                 int tmo;
5729 
5730                 /* If we enter the TCP_FIN_WAIT1 state and we are a
5731                  * Fast Open socket and this is the first acceptable
5732                  * ACK we have received, this would have acknowledged
5733                  * our SYNACK so stop the SYNACK timer.
5734                  */
5735                 if (req != NULL) {
5736                         /* Return RST if ack_seq is invalid.
5737                          * Note that RFC793 only says to generate a
5738                          * DUPACK for it but for TCP Fast Open it seems
5739                          * better to treat this case like TCP_SYN_RECV
5740                          * above.
5741                          */
5742                         if (!acceptable)
5743                                 return 1;
5744                         /* We no longer need the request sock. */
5745                         reqsk_fastopen_remove(sk, req, false);
5746                         tcp_rearm_rto(sk);
5747                 }
5748                 if (tp->snd_una != tp->write_seq)
5749                         break;
5750 
5751                 tcp_set_state(sk, TCP_FIN_WAIT2);
5752                 sk->sk_shutdown |= SEND_SHUTDOWN;
5753 
5754                 dst = __sk_dst_get(sk);
5755                 if (dst)
5756                         dst_confirm(dst);
5757 
5758                 if (!sock_flag(sk, SOCK_DEAD)) {
5759                         /* Wake up lingering close() */
5760                         sk->sk_state_change(sk);
5761                         break;
5762                 }
5763 
5764                 if (tp->linger2 < 0 ||
5765                     (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5766                      after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5767                         tcp_done(sk);
5768                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5769                         return 1;
5770                 }
5771 
5772                 tmo = tcp_fin_time(sk);
5773                 if (tmo > TCP_TIMEWAIT_LEN) {
5774                         inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5775                 } else if (th->fin || sock_owned_by_user(sk)) {
5776                         /* Bad case. We could lose such FIN otherwise.
5777                          * It is not a big problem, but it looks confusing
5778                          * and not so rare event. We still can lose it now,
5779                          * if it spins in bh_lock_sock(), but it is really
5780                          * marginal case.
5781                          */
5782                         inet_csk_reset_keepalive_timer(sk, tmo);
5783                 } else {
5784                         tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5785                         goto discard;
5786                 }
5787                 break;
5788         }
5789 
5790         case TCP_CLOSING:
5791                 if (tp->snd_una == tp->write_seq) {
5792                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5793                         goto discard;
5794                 }
5795                 break;
5796 
5797         case TCP_LAST_ACK:
5798                 if (tp->snd_una == tp->write_seq) {
5799                         tcp_update_metrics(sk);
5800                         tcp_done(sk);
5801                         goto discard;
5802                 }
5803                 break;
5804         }
5805 
5806         /* step 6: check the URG bit */
5807         tcp_urg(sk, skb, th);
5808 
5809         /* step 7: process the segment text */
5810         switch (sk->sk_state) {
5811         case TCP_CLOSE_WAIT:
5812         case TCP_CLOSING:
5813         case TCP_LAST_ACK:
5814                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5815                         break;
5816         case TCP_FIN_WAIT1:
5817         case TCP_FIN_WAIT2:
5818                 /* RFC 793 says to queue data in these states,
5819                  * RFC 1122 says we MUST send a reset.
5820                  * BSD 4.4 also does reset.
5821                  */
5822                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5823                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5824                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5825                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5826                                 tcp_reset(sk);
5827                                 return 1;
5828                         }
5829                 }
5830                 /* Fall through */
5831         case TCP_ESTABLISHED:
5832                 tcp_data_queue(sk, skb);
5833                 queued = 1;
5834                 break;
5835         }
5836 
5837         /* tcp_data could move socket to TIME-WAIT */
5838         if (sk->sk_state != TCP_CLOSE) {
5839                 tcp_data_snd_check(sk);
5840                 tcp_ack_snd_check(sk);
5841         }
5842 
5843         if (!queued) {
5844 discard:
5845                 __kfree_skb(skb);
5846         }
5847         return 0;
5848 }
5849 EXPORT_SYMBOL(tcp_rcv_state_process);
5850 
5851 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
5852 {
5853         struct inet_request_sock *ireq = inet_rsk(req);
5854 
5855         if (family == AF_INET)
5856                 net_dbg_ratelimited("drop open request from %pI4/%u\n",
5857                                     &ireq->ir_rmt_addr, port);
5858 #if IS_ENABLED(CONFIG_IPV6)
5859         else if (family == AF_INET6)
5860                 net_dbg_ratelimited("drop open request from %pI6/%u\n",
5861                                     &ireq->ir_v6_rmt_addr, port);
5862 #endif
5863 }
5864 
5865 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
5866  *
5867  * If we receive a SYN packet with these bits set, it means a
5868  * network is playing bad games with TOS bits. In order to
5869  * avoid possible false congestion notifications, we disable
5870  * TCP ECN negotiation.
5871  *
5872  * Exception: tcp_ca wants ECN. This is required for DCTCP
5873  * congestion control; it requires setting ECT on all packets,
5874  * including SYN. We inverse the test in this case: If our
5875  * local socket wants ECN, but peer only set ece/cwr (but not
5876  * ECT in IP header) its probably a non-DCTCP aware sender.
5877  */
5878 static void tcp_ecn_create_request(struct request_sock *req,
5879                                    const struct sk_buff *skb,
5880                                    const struct sock *listen_sk,
5881                                    const struct dst_entry *dst)
5882 {
5883         const struct tcphdr *th = tcp_hdr(skb);
5884         const struct net *net = sock_net(listen_sk);
5885         bool th_ecn = th->ece && th->cwr;
5886         bool ect, need_ecn, ecn_ok;
5887 
5888         if (!th_ecn)
5889                 return;
5890 
5891         ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
5892         need_ecn = tcp_ca_needs_ecn(listen_sk);
5893         ecn_ok = net->ipv4.sysctl_tcp_ecn || dst_feature(dst, RTAX_FEATURE_ECN);
5894 
5895         if (!ect && !need_ecn && ecn_ok)
5896                 inet_rsk(req)->ecn_ok = 1;
5897         else if (ect && need_ecn)
5898                 inet_rsk(req)->ecn_ok = 1;
5899 }
5900 
5901 int tcp_conn_request(struct request_sock_ops *rsk_ops,
5902                      const struct tcp_request_sock_ops *af_ops,
5903                      struct sock *sk, struct sk_buff *skb)
5904 {
5905         struct tcp_options_received tmp_opt;
5906         struct request_sock *req;
5907         struct tcp_sock *tp = tcp_sk(sk);
5908         struct dst_entry *dst = NULL;
5909         __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
5910         bool want_cookie = false, fastopen;
5911         struct flowi fl;
5912         struct tcp_fastopen_cookie foc = { .len = -1 };
5913         int err;
5914 
5915 
5916         /* TW buckets are converted to open requests without
5917          * limitations, they conserve resources and peer is
5918          * evidently real one.
5919          */
5920         if ((sysctl_tcp_syncookies == 2 ||
5921              inet_csk_reqsk_queue_is_full(sk)) && !isn) {
5922                 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
5923                 if (!want_cookie)
5924                         goto drop;
5925         }
5926 
5927 
5928         /* Accept backlog is full. If we have already queued enough
5929          * of warm entries in syn queue, drop request. It is better than
5930          * clogging syn queue with openreqs with exponentially increasing
5931          * timeout.
5932          */
5933         if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
5934                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
5935                 goto drop;
5936         }
5937 
5938         req = inet_reqsk_alloc(rsk_ops);
5939         if (!req)
5940                 goto drop;
5941 
5942         tcp_rsk(req)->af_specific = af_ops;
5943 
5944         tcp_clear_options(&tmp_opt);
5945         tmp_opt.mss_clamp = af_ops->mss_clamp;
5946         tmp_opt.user_mss  = tp->rx_opt.user_mss;
5947         tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
5948 
5949         if (want_cookie && !tmp_opt.saw_tstamp)
5950                 tcp_clear_options(&tmp_opt);
5951 
5952         tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
5953         tcp_openreq_init(req, &tmp_opt, skb, sk);
5954 
5955         af_ops->init_req(req, sk, skb);
5956 
5957         if (security_inet_conn_request(sk, skb, req))
5958                 goto drop_and_free;
5959 
5960         if (!want_cookie && !isn) {
5961                 /* VJ's idea. We save last timestamp seen
5962                  * from the destination in peer table, when entering
5963                  * state TIME-WAIT, and check against it before
5964                  * accepting new connection request.
5965                  *
5966                  * If "isn" is not zero, this request hit alive
5967                  * timewait bucket, so that all the necessary checks
5968                  * are made in the function processing timewait state.
5969                  */
5970                 if (tcp_death_row.sysctl_tw_recycle) {
5971                         bool strict;
5972 
5973                         dst = af_ops->route_req(sk, &fl, req, &strict);
5974 
5975                         if (dst && strict &&
5976                             !tcp_peer_is_proven(req, dst, true,
5977                                                 tmp_opt.saw_tstamp)) {
5978                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
5979                                 goto drop_and_release;
5980                         }
5981                 }
5982                 /* Kill the following clause, if you dislike this way. */
5983                 else if (!sysctl_tcp_syncookies &&
5984                          (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
5985                           (sysctl_max_syn_backlog >> 2)) &&
5986                          !tcp_peer_is_proven(req, dst, false,
5987                                              tmp_opt.saw_tstamp)) {
5988                         /* Without syncookies last quarter of
5989                          * backlog is filled with destinations,
5990                          * proven to be alive.
5991                          * It means that we continue to communicate
5992                          * to destinations, already remembered
5993                          * to the moment of synflood.
5994                          */
5995                         pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
5996                                     rsk_ops->family);
5997                         goto drop_and_release;
5998                 }
5999 
6000                 isn = af_ops->init_seq(skb);
6001         }
6002         if (!dst) {
6003                 dst = af_ops->route_req(sk, &fl, req, NULL);
6004                 if (!dst)
6005                         goto drop_and_free;
6006         }
6007 
6008         tcp_ecn_create_request(req, skb, sk, dst);
6009 
6010         if (want_cookie) {
6011                 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6012                 req->cookie_ts = tmp_opt.tstamp_ok;
6013                 if (!tmp_opt.tstamp_ok)
6014                         inet_rsk(req)->ecn_ok = 0;
6015         }
6016 
6017         tcp_rsk(req)->snt_isn = isn;
6018         tcp_openreq_init_rwin(req, sk, dst);
6019         fastopen = !want_cookie &&
6020                    tcp_try_fastopen(sk, skb, req, &foc, dst);
6021         err = af_ops->send_synack(sk, dst, &fl, req,
6022                                   skb_get_queue_mapping(skb), &foc);
6023         if (!fastopen) {
6024                 if (err || want_cookie)
6025                         goto drop_and_free;
6026 
6027                 tcp_rsk(req)->listener = NULL;
6028                 af_ops->queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6029         }
6030 
6031         return 0;
6032 
6033 drop_and_release:
6034         dst_release(dst);
6035 drop_and_free:
6036         reqsk_free(req);
6037 drop:
6038         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS);
6039         return 0;
6040 }
6041 EXPORT_SYMBOL(tcp_conn_request);
6042 

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us