Version:  2.0.40 2.2.26 2.4.37 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 4.0 4.1

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 int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
104 
105 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
106 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
107 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
108 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
109 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
110 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
111 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
112 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
113 #define FLAG_ORIG_SACK_ACKED    0x200 /* Never retransmitted data are (s)acked  */
114 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
115 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
116 #define FLAG_SACK_RENEGING      0x2000 /* snd_una advanced to a sacked seq */
117 #define FLAG_UPDATE_TS_RECENT   0x4000 /* tcp_replace_ts_recent() */
118 
119 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
120 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
121 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
122 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
123 
124 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
125 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
126 
127 /* Adapt the MSS value used to make delayed ack decision to the
128  * real world.
129  */
130 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
131 {
132         struct inet_connection_sock *icsk = inet_csk(sk);
133         const unsigned int lss = icsk->icsk_ack.last_seg_size;
134         unsigned int len;
135 
136         icsk->icsk_ack.last_seg_size = 0;
137 
138         /* skb->len may jitter because of SACKs, even if peer
139          * sends good full-sized frames.
140          */
141         len = skb_shinfo(skb)->gso_size ? : skb->len;
142         if (len >= icsk->icsk_ack.rcv_mss) {
143                 icsk->icsk_ack.rcv_mss = len;
144         } else {
145                 /* Otherwise, we make more careful check taking into account,
146                  * that SACKs block is variable.
147                  *
148                  * "len" is invariant segment length, including TCP header.
149                  */
150                 len += skb->data - skb_transport_header(skb);
151                 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
152                     /* If PSH is not set, packet should be
153                      * full sized, provided peer TCP is not badly broken.
154                      * This observation (if it is correct 8)) allows
155                      * to handle super-low mtu links fairly.
156                      */
157                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
158                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
159                         /* Subtract also invariant (if peer is RFC compliant),
160                          * tcp header plus fixed timestamp option length.
161                          * Resulting "len" is MSS free of SACK jitter.
162                          */
163                         len -= tcp_sk(sk)->tcp_header_len;
164                         icsk->icsk_ack.last_seg_size = len;
165                         if (len == lss) {
166                                 icsk->icsk_ack.rcv_mss = len;
167                                 return;
168                         }
169                 }
170                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
171                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
172                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
173         }
174 }
175 
176 static void tcp_incr_quickack(struct sock *sk)
177 {
178         struct inet_connection_sock *icsk = inet_csk(sk);
179         unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
180 
181         if (quickacks == 0)
182                 quickacks = 2;
183         if (quickacks > icsk->icsk_ack.quick)
184                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
185 }
186 
187 static void tcp_enter_quickack_mode(struct sock *sk)
188 {
189         struct inet_connection_sock *icsk = inet_csk(sk);
190         tcp_incr_quickack(sk);
191         icsk->icsk_ack.pingpong = 0;
192         icsk->icsk_ack.ato = TCP_ATO_MIN;
193 }
194 
195 /* Send ACKs quickly, if "quick" count is not exhausted
196  * and the session is not interactive.
197  */
198 
199 static inline bool tcp_in_quickack_mode(const struct sock *sk)
200 {
201         const struct inet_connection_sock *icsk = inet_csk(sk);
202 
203         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
204 }
205 
206 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
207 {
208         if (tp->ecn_flags & TCP_ECN_OK)
209                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
210 }
211 
212 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
213 {
214         if (tcp_hdr(skb)->cwr)
215                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
216 }
217 
218 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
219 {
220         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
221 }
222 
223 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
224 {
225         switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
226         case INET_ECN_NOT_ECT:
227                 /* Funny extension: if ECT is not set on a segment,
228                  * and we already seen ECT on a previous segment,
229                  * it is probably a retransmit.
230                  */
231                 if (tp->ecn_flags & TCP_ECN_SEEN)
232                         tcp_enter_quickack_mode((struct sock *)tp);
233                 break;
234         case INET_ECN_CE:
235                 if (tcp_ca_needs_ecn((struct sock *)tp))
236                         tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
237 
238                 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
239                         /* Better not delay acks, sender can have a very low cwnd */
240                         tcp_enter_quickack_mode((struct sock *)tp);
241                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
242                 }
243                 tp->ecn_flags |= TCP_ECN_SEEN;
244                 break;
245         default:
246                 if (tcp_ca_needs_ecn((struct sock *)tp))
247                         tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
248                 tp->ecn_flags |= TCP_ECN_SEEN;
249                 break;
250         }
251 }
252 
253 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
254 {
255         if (tp->ecn_flags & TCP_ECN_OK)
256                 __tcp_ecn_check_ce(tp, skb);
257 }
258 
259 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
260 {
261         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
262                 tp->ecn_flags &= ~TCP_ECN_OK;
263 }
264 
265 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
266 {
267         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
268                 tp->ecn_flags &= ~TCP_ECN_OK;
269 }
270 
271 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
272 {
273         if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
274                 return true;
275         return false;
276 }
277 
278 /* Buffer size and advertised window tuning.
279  *
280  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
281  */
282 
283 static void tcp_sndbuf_expand(struct sock *sk)
284 {
285         const struct tcp_sock *tp = tcp_sk(sk);
286         int sndmem, per_mss;
287         u32 nr_segs;
288 
289         /* Worst case is non GSO/TSO : each frame consumes one skb
290          * and skb->head is kmalloced using power of two area of memory
291          */
292         per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
293                   MAX_TCP_HEADER +
294                   SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
295 
296         per_mss = roundup_pow_of_two(per_mss) +
297                   SKB_DATA_ALIGN(sizeof(struct sk_buff));
298 
299         nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
300         nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
301 
302         /* Fast Recovery (RFC 5681 3.2) :
303          * Cubic needs 1.7 factor, rounded to 2 to include
304          * extra cushion (application might react slowly to POLLOUT)
305          */
306         sndmem = 2 * nr_segs * per_mss;
307 
308         if (sk->sk_sndbuf < sndmem)
309                 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
310 }
311 
312 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
313  *
314  * All tcp_full_space() is split to two parts: "network" buffer, allocated
315  * forward and advertised in receiver window (tp->rcv_wnd) and
316  * "application buffer", required to isolate scheduling/application
317  * latencies from network.
318  * window_clamp is maximal advertised window. It can be less than
319  * tcp_full_space(), in this case tcp_full_space() - window_clamp
320  * is reserved for "application" buffer. The less window_clamp is
321  * the smoother our behaviour from viewpoint of network, but the lower
322  * throughput and the higher sensitivity of the connection to losses. 8)
323  *
324  * rcv_ssthresh is more strict window_clamp used at "slow start"
325  * phase to predict further behaviour of this connection.
326  * It is used for two goals:
327  * - to enforce header prediction at sender, even when application
328  *   requires some significant "application buffer". It is check #1.
329  * - to prevent pruning of receive queue because of misprediction
330  *   of receiver window. Check #2.
331  *
332  * The scheme does not work when sender sends good segments opening
333  * window and then starts to feed us spaghetti. But it should work
334  * in common situations. Otherwise, we have to rely on queue collapsing.
335  */
336 
337 /* Slow part of check#2. */
338 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
339 {
340         struct tcp_sock *tp = tcp_sk(sk);
341         /* Optimize this! */
342         int truesize = tcp_win_from_space(skb->truesize) >> 1;
343         int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
344 
345         while (tp->rcv_ssthresh <= window) {
346                 if (truesize <= skb->len)
347                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
348 
349                 truesize >>= 1;
350                 window >>= 1;
351         }
352         return 0;
353 }
354 
355 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
356 {
357         struct tcp_sock *tp = tcp_sk(sk);
358 
359         /* Check #1 */
360         if (tp->rcv_ssthresh < tp->window_clamp &&
361             (int)tp->rcv_ssthresh < tcp_space(sk) &&
362             !sk_under_memory_pressure(sk)) {
363                 int incr;
364 
365                 /* Check #2. Increase window, if skb with such overhead
366                  * will fit to rcvbuf in future.
367                  */
368                 if (tcp_win_from_space(skb->truesize) <= skb->len)
369                         incr = 2 * tp->advmss;
370                 else
371                         incr = __tcp_grow_window(sk, skb);
372 
373                 if (incr) {
374                         incr = max_t(int, incr, 2 * skb->len);
375                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
376                                                tp->window_clamp);
377                         inet_csk(sk)->icsk_ack.quick |= 1;
378                 }
379         }
380 }
381 
382 /* 3. Tuning rcvbuf, when connection enters established state. */
383 static void tcp_fixup_rcvbuf(struct sock *sk)
384 {
385         u32 mss = tcp_sk(sk)->advmss;
386         int rcvmem;
387 
388         rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
389                  tcp_default_init_rwnd(mss);
390 
391         /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
392          * Allow enough cushion so that sender is not limited by our window
393          */
394         if (sysctl_tcp_moderate_rcvbuf)
395                 rcvmem <<= 2;
396 
397         if (sk->sk_rcvbuf < rcvmem)
398                 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
399 }
400 
401 /* 4. Try to fixup all. It is made immediately after connection enters
402  *    established state.
403  */
404 void tcp_init_buffer_space(struct sock *sk)
405 {
406         struct tcp_sock *tp = tcp_sk(sk);
407         int maxwin;
408 
409         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
410                 tcp_fixup_rcvbuf(sk);
411         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
412                 tcp_sndbuf_expand(sk);
413 
414         tp->rcvq_space.space = tp->rcv_wnd;
415         tp->rcvq_space.time = tcp_time_stamp;
416         tp->rcvq_space.seq = tp->copied_seq;
417 
418         maxwin = tcp_full_space(sk);
419 
420         if (tp->window_clamp >= maxwin) {
421                 tp->window_clamp = maxwin;
422 
423                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
424                         tp->window_clamp = max(maxwin -
425                                                (maxwin >> sysctl_tcp_app_win),
426                                                4 * tp->advmss);
427         }
428 
429         /* Force reservation of one segment. */
430         if (sysctl_tcp_app_win &&
431             tp->window_clamp > 2 * tp->advmss &&
432             tp->window_clamp + tp->advmss > maxwin)
433                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
434 
435         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
436         tp->snd_cwnd_stamp = tcp_time_stamp;
437 }
438 
439 /* 5. Recalculate window clamp after socket hit its memory bounds. */
440 static void tcp_clamp_window(struct sock *sk)
441 {
442         struct tcp_sock *tp = tcp_sk(sk);
443         struct inet_connection_sock *icsk = inet_csk(sk);
444 
445         icsk->icsk_ack.quick = 0;
446 
447         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
448             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
449             !sk_under_memory_pressure(sk) &&
450             sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
451                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
452                                     sysctl_tcp_rmem[2]);
453         }
454         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
455                 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
456 }
457 
458 /* Initialize RCV_MSS value.
459  * RCV_MSS is an our guess about MSS used by the peer.
460  * We haven't any direct information about the MSS.
461  * It's better to underestimate the RCV_MSS rather than overestimate.
462  * Overestimations make us ACKing less frequently than needed.
463  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
464  */
465 void tcp_initialize_rcv_mss(struct sock *sk)
466 {
467         const struct tcp_sock *tp = tcp_sk(sk);
468         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
469 
470         hint = min(hint, tp->rcv_wnd / 2);
471         hint = min(hint, TCP_MSS_DEFAULT);
472         hint = max(hint, TCP_MIN_MSS);
473 
474         inet_csk(sk)->icsk_ack.rcv_mss = hint;
475 }
476 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
477 
478 /* Receiver "autotuning" code.
479  *
480  * The algorithm for RTT estimation w/o timestamps is based on
481  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
482  * <http://public.lanl.gov/radiant/pubs.html#DRS>
483  *
484  * More detail on this code can be found at
485  * <http://staff.psc.edu/jheffner/>,
486  * though this reference is out of date.  A new paper
487  * is pending.
488  */
489 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
490 {
491         u32 new_sample = tp->rcv_rtt_est.rtt;
492         long m = sample;
493 
494         if (m == 0)
495                 m = 1;
496 
497         if (new_sample != 0) {
498                 /* If we sample in larger samples in the non-timestamp
499                  * case, we could grossly overestimate the RTT especially
500                  * with chatty applications or bulk transfer apps which
501                  * are stalled on filesystem I/O.
502                  *
503                  * Also, since we are only going for a minimum in the
504                  * non-timestamp case, we do not smooth things out
505                  * else with timestamps disabled convergence takes too
506                  * long.
507                  */
508                 if (!win_dep) {
509                         m -= (new_sample >> 3);
510                         new_sample += m;
511                 } else {
512                         m <<= 3;
513                         if (m < new_sample)
514                                 new_sample = m;
515                 }
516         } else {
517                 /* No previous measure. */
518                 new_sample = m << 3;
519         }
520 
521         if (tp->rcv_rtt_est.rtt != new_sample)
522                 tp->rcv_rtt_est.rtt = new_sample;
523 }
524 
525 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
526 {
527         if (tp->rcv_rtt_est.time == 0)
528                 goto new_measure;
529         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
530                 return;
531         tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
532 
533 new_measure:
534         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
535         tp->rcv_rtt_est.time = tcp_time_stamp;
536 }
537 
538 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
539                                           const struct sk_buff *skb)
540 {
541         struct tcp_sock *tp = tcp_sk(sk);
542         if (tp->rx_opt.rcv_tsecr &&
543             (TCP_SKB_CB(skb)->end_seq -
544              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
545                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
546 }
547 
548 /*
549  * This function should be called every time data is copied to user space.
550  * It calculates the appropriate TCP receive buffer space.
551  */
552 void tcp_rcv_space_adjust(struct sock *sk)
553 {
554         struct tcp_sock *tp = tcp_sk(sk);
555         int time;
556         int copied;
557 
558         time = tcp_time_stamp - tp->rcvq_space.time;
559         if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
560                 return;
561 
562         /* Number of bytes copied to user in last RTT */
563         copied = tp->copied_seq - tp->rcvq_space.seq;
564         if (copied <= tp->rcvq_space.space)
565                 goto new_measure;
566 
567         /* A bit of theory :
568          * copied = bytes received in previous RTT, our base window
569          * To cope with packet losses, we need a 2x factor
570          * To cope with slow start, and sender growing its cwin by 100 %
571          * every RTT, we need a 4x factor, because the ACK we are sending
572          * now is for the next RTT, not the current one :
573          * <prev RTT . ><current RTT .. ><next RTT .... >
574          */
575 
576         if (sysctl_tcp_moderate_rcvbuf &&
577             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
578                 int rcvwin, rcvmem, rcvbuf;
579 
580                 /* minimal window to cope with packet losses, assuming
581                  * steady state. Add some cushion because of small variations.
582                  */
583                 rcvwin = (copied << 1) + 16 * tp->advmss;
584 
585                 /* If rate increased by 25%,
586                  *      assume slow start, rcvwin = 3 * copied
587                  * If rate increased by 50%,
588                  *      assume sender can use 2x growth, rcvwin = 4 * copied
589                  */
590                 if (copied >=
591                     tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
592                         if (copied >=
593                             tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
594                                 rcvwin <<= 1;
595                         else
596                                 rcvwin += (rcvwin >> 1);
597                 }
598 
599                 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
600                 while (tcp_win_from_space(rcvmem) < tp->advmss)
601                         rcvmem += 128;
602 
603                 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
604                 if (rcvbuf > sk->sk_rcvbuf) {
605                         sk->sk_rcvbuf = rcvbuf;
606 
607                         /* Make the window clamp follow along.  */
608                         tp->window_clamp = rcvwin;
609                 }
610         }
611         tp->rcvq_space.space = copied;
612 
613 new_measure:
614         tp->rcvq_space.seq = tp->copied_seq;
615         tp->rcvq_space.time = tcp_time_stamp;
616 }
617 
618 /* There is something which you must keep in mind when you analyze the
619  * behavior of the tp->ato delayed ack timeout interval.  When a
620  * connection starts up, we want to ack as quickly as possible.  The
621  * problem is that "good" TCP's do slow start at the beginning of data
622  * transmission.  The means that until we send the first few ACK's the
623  * sender will sit on his end and only queue most of his data, because
624  * he can only send snd_cwnd unacked packets at any given time.  For
625  * each ACK we send, he increments snd_cwnd and transmits more of his
626  * queue.  -DaveM
627  */
628 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
629 {
630         struct tcp_sock *tp = tcp_sk(sk);
631         struct inet_connection_sock *icsk = inet_csk(sk);
632         u32 now;
633 
634         inet_csk_schedule_ack(sk);
635 
636         tcp_measure_rcv_mss(sk, skb);
637 
638         tcp_rcv_rtt_measure(tp);
639 
640         now = tcp_time_stamp;
641 
642         if (!icsk->icsk_ack.ato) {
643                 /* The _first_ data packet received, initialize
644                  * delayed ACK engine.
645                  */
646                 tcp_incr_quickack(sk);
647                 icsk->icsk_ack.ato = TCP_ATO_MIN;
648         } else {
649                 int m = now - icsk->icsk_ack.lrcvtime;
650 
651                 if (m <= TCP_ATO_MIN / 2) {
652                         /* The fastest case is the first. */
653                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
654                 } else if (m < icsk->icsk_ack.ato) {
655                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
656                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
657                                 icsk->icsk_ack.ato = icsk->icsk_rto;
658                 } else if (m > icsk->icsk_rto) {
659                         /* Too long gap. Apparently sender failed to
660                          * restart window, so that we send ACKs quickly.
661                          */
662                         tcp_incr_quickack(sk);
663                         sk_mem_reclaim(sk);
664                 }
665         }
666         icsk->icsk_ack.lrcvtime = now;
667 
668         tcp_ecn_check_ce(tp, skb);
669 
670         if (skb->len >= 128)
671                 tcp_grow_window(sk, skb);
672 }
673 
674 /* Called to compute a smoothed rtt estimate. The data fed to this
675  * routine either comes from timestamps, or from segments that were
676  * known _not_ to have been retransmitted [see Karn/Partridge
677  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
678  * piece by Van Jacobson.
679  * NOTE: the next three routines used to be one big routine.
680  * To save cycles in the RFC 1323 implementation it was better to break
681  * it up into three procedures. -- erics
682  */
683 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
684 {
685         struct tcp_sock *tp = tcp_sk(sk);
686         long m = mrtt_us; /* RTT */
687         u32 srtt = tp->srtt_us;
688 
689         /*      The following amusing code comes from Jacobson's
690          *      article in SIGCOMM '88.  Note that rtt and mdev
691          *      are scaled versions of rtt and mean deviation.
692          *      This is designed to be as fast as possible
693          *      m stands for "measurement".
694          *
695          *      On a 1990 paper the rto value is changed to:
696          *      RTO = rtt + 4 * mdev
697          *
698          * Funny. This algorithm seems to be very broken.
699          * These formulae increase RTO, when it should be decreased, increase
700          * too slowly, when it should be increased quickly, decrease too quickly
701          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
702          * does not matter how to _calculate_ it. Seems, it was trap
703          * that VJ failed to avoid. 8)
704          */
705         if (srtt != 0) {
706                 m -= (srtt >> 3);       /* m is now error in rtt est */
707                 srtt += m;              /* rtt = 7/8 rtt + 1/8 new */
708                 if (m < 0) {
709                         m = -m;         /* m is now abs(error) */
710                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
711                         /* This is similar to one of Eifel findings.
712                          * Eifel blocks mdev updates when rtt decreases.
713                          * This solution is a bit different: we use finer gain
714                          * for mdev in this case (alpha*beta).
715                          * Like Eifel it also prevents growth of rto,
716                          * but also it limits too fast rto decreases,
717                          * happening in pure Eifel.
718                          */
719                         if (m > 0)
720                                 m >>= 3;
721                 } else {
722                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
723                 }
724                 tp->mdev_us += m;               /* mdev = 3/4 mdev + 1/4 new */
725                 if (tp->mdev_us > tp->mdev_max_us) {
726                         tp->mdev_max_us = tp->mdev_us;
727                         if (tp->mdev_max_us > tp->rttvar_us)
728                                 tp->rttvar_us = tp->mdev_max_us;
729                 }
730                 if (after(tp->snd_una, tp->rtt_seq)) {
731                         if (tp->mdev_max_us < tp->rttvar_us)
732                                 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
733                         tp->rtt_seq = tp->snd_nxt;
734                         tp->mdev_max_us = tcp_rto_min_us(sk);
735                 }
736         } else {
737                 /* no previous measure. */
738                 srtt = m << 3;          /* take the measured time to be rtt */
739                 tp->mdev_us = m << 1;   /* make sure rto = 3*rtt */
740                 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
741                 tp->mdev_max_us = tp->rttvar_us;
742                 tp->rtt_seq = tp->snd_nxt;
743         }
744         tp->srtt_us = max(1U, srtt);
745 }
746 
747 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
748  * Note: TCP stack does not yet implement pacing.
749  * FQ packet scheduler can be used to implement cheap but effective
750  * TCP pacing, to smooth the burst on large writes when packets
751  * in flight is significantly lower than cwnd (or rwin)
752  */
753 static void tcp_update_pacing_rate(struct sock *sk)
754 {
755         const struct tcp_sock *tp = tcp_sk(sk);
756         u64 rate;
757 
758         /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
759         rate = (u64)tp->mss_cache * 2 * (USEC_PER_SEC << 3);
760 
761         rate *= max(tp->snd_cwnd, tp->packets_out);
762 
763         if (likely(tp->srtt_us))
764                 do_div(rate, tp->srtt_us);
765 
766         /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
767          * without any lock. We want to make sure compiler wont store
768          * intermediate values in this location.
769          */
770         ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
771                                                 sk->sk_max_pacing_rate);
772 }
773 
774 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
775  * routine referred to above.
776  */
777 static void tcp_set_rto(struct sock *sk)
778 {
779         const struct tcp_sock *tp = tcp_sk(sk);
780         /* Old crap is replaced with new one. 8)
781          *
782          * More seriously:
783          * 1. If rtt variance happened to be less 50msec, it is hallucination.
784          *    It cannot be less due to utterly erratic ACK generation made
785          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
786          *    to do with delayed acks, because at cwnd>2 true delack timeout
787          *    is invisible. Actually, Linux-2.4 also generates erratic
788          *    ACKs in some circumstances.
789          */
790         inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
791 
792         /* 2. Fixups made earlier cannot be right.
793          *    If we do not estimate RTO correctly without them,
794          *    all the algo is pure shit and should be replaced
795          *    with correct one. It is exactly, which we pretend to do.
796          */
797 
798         /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
799          * guarantees that rto is higher.
800          */
801         tcp_bound_rto(sk);
802 }
803 
804 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
805 {
806         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
807 
808         if (!cwnd)
809                 cwnd = TCP_INIT_CWND;
810         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
811 }
812 
813 /*
814  * Packet counting of FACK is based on in-order assumptions, therefore TCP
815  * disables it when reordering is detected
816  */
817 void tcp_disable_fack(struct tcp_sock *tp)
818 {
819         /* RFC3517 uses different metric in lost marker => reset on change */
820         if (tcp_is_fack(tp))
821                 tp->lost_skb_hint = NULL;
822         tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
823 }
824 
825 /* Take a notice that peer is sending D-SACKs */
826 static void tcp_dsack_seen(struct tcp_sock *tp)
827 {
828         tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
829 }
830 
831 static void tcp_update_reordering(struct sock *sk, const int metric,
832                                   const int ts)
833 {
834         struct tcp_sock *tp = tcp_sk(sk);
835         if (metric > tp->reordering) {
836                 int mib_idx;
837 
838                 tp->reordering = min(sysctl_tcp_max_reordering, metric);
839 
840                 /* This exciting event is worth to be remembered. 8) */
841                 if (ts)
842                         mib_idx = LINUX_MIB_TCPTSREORDER;
843                 else if (tcp_is_reno(tp))
844                         mib_idx = LINUX_MIB_TCPRENOREORDER;
845                 else if (tcp_is_fack(tp))
846                         mib_idx = LINUX_MIB_TCPFACKREORDER;
847                 else
848                         mib_idx = LINUX_MIB_TCPSACKREORDER;
849 
850                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
851 #if FASTRETRANS_DEBUG > 1
852                 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
853                          tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
854                          tp->reordering,
855                          tp->fackets_out,
856                          tp->sacked_out,
857                          tp->undo_marker ? tp->undo_retrans : 0);
858 #endif
859                 tcp_disable_fack(tp);
860         }
861 
862         if (metric > 0)
863                 tcp_disable_early_retrans(tp);
864 }
865 
866 /* This must be called before lost_out is incremented */
867 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
868 {
869         if (!tp->retransmit_skb_hint ||
870             before(TCP_SKB_CB(skb)->seq,
871                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
872                 tp->retransmit_skb_hint = skb;
873 
874         if (!tp->lost_out ||
875             after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
876                 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
877 }
878 
879 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
880 {
881         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
882                 tcp_verify_retransmit_hint(tp, skb);
883 
884                 tp->lost_out += tcp_skb_pcount(skb);
885                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
886         }
887 }
888 
889 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
890                                             struct sk_buff *skb)
891 {
892         tcp_verify_retransmit_hint(tp, skb);
893 
894         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
895                 tp->lost_out += tcp_skb_pcount(skb);
896                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
897         }
898 }
899 
900 /* This procedure tags the retransmission queue when SACKs arrive.
901  *
902  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
903  * Packets in queue with these bits set are counted in variables
904  * sacked_out, retrans_out and lost_out, correspondingly.
905  *
906  * Valid combinations are:
907  * Tag  InFlight        Description
908  * 0    1               - orig segment is in flight.
909  * S    0               - nothing flies, orig reached receiver.
910  * L    0               - nothing flies, orig lost by net.
911  * R    2               - both orig and retransmit are in flight.
912  * L|R  1               - orig is lost, retransmit is in flight.
913  * S|R  1               - orig reached receiver, retrans is still in flight.
914  * (L|S|R is logically valid, it could occur when L|R is sacked,
915  *  but it is equivalent to plain S and code short-curcuits it to S.
916  *  L|S is logically invalid, it would mean -1 packet in flight 8))
917  *
918  * These 6 states form finite state machine, controlled by the following events:
919  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
920  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
921  * 3. Loss detection event of two flavors:
922  *      A. Scoreboard estimator decided the packet is lost.
923  *         A'. Reno "three dupacks" marks head of queue lost.
924  *         A''. Its FACK modification, head until snd.fack is lost.
925  *      B. SACK arrives sacking SND.NXT at the moment, when the
926  *         segment was retransmitted.
927  * 4. D-SACK added new rule: D-SACK changes any tag to S.
928  *
929  * It is pleasant to note, that state diagram turns out to be commutative,
930  * so that we are allowed not to be bothered by order of our actions,
931  * when multiple events arrive simultaneously. (see the function below).
932  *
933  * Reordering detection.
934  * --------------------
935  * Reordering metric is maximal distance, which a packet can be displaced
936  * in packet stream. With SACKs we can estimate it:
937  *
938  * 1. SACK fills old hole and the corresponding segment was not
939  *    ever retransmitted -> reordering. Alas, we cannot use it
940  *    when segment was retransmitted.
941  * 2. The last flaw is solved with D-SACK. D-SACK arrives
942  *    for retransmitted and already SACKed segment -> reordering..
943  * Both of these heuristics are not used in Loss state, when we cannot
944  * account for retransmits accurately.
945  *
946  * SACK block validation.
947  * ----------------------
948  *
949  * SACK block range validation checks that the received SACK block fits to
950  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
951  * Note that SND.UNA is not included to the range though being valid because
952  * it means that the receiver is rather inconsistent with itself reporting
953  * SACK reneging when it should advance SND.UNA. Such SACK block this is
954  * perfectly valid, however, in light of RFC2018 which explicitly states
955  * that "SACK block MUST reflect the newest segment.  Even if the newest
956  * segment is going to be discarded ...", not that it looks very clever
957  * in case of head skb. Due to potentional receiver driven attacks, we
958  * choose to avoid immediate execution of a walk in write queue due to
959  * reneging and defer head skb's loss recovery to standard loss recovery
960  * procedure that will eventually trigger (nothing forbids us doing this).
961  *
962  * Implements also blockage to start_seq wrap-around. Problem lies in the
963  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
964  * there's no guarantee that it will be before snd_nxt (n). The problem
965  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
966  * wrap (s_w):
967  *
968  *         <- outs wnd ->                          <- wrapzone ->
969  *         u     e      n                         u_w   e_w  s n_w
970  *         |     |      |                          |     |   |  |
971  * |<------------+------+----- TCP seqno space --------------+---------->|
972  * ...-- <2^31 ->|                                           |<--------...
973  * ...---- >2^31 ------>|                                    |<--------...
974  *
975  * Current code wouldn't be vulnerable but it's better still to discard such
976  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
977  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
978  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
979  * equal to the ideal case (infinite seqno space without wrap caused issues).
980  *
981  * With D-SACK the lower bound is extended to cover sequence space below
982  * SND.UNA down to undo_marker, which is the last point of interest. Yet
983  * again, D-SACK block must not to go across snd_una (for the same reason as
984  * for the normal SACK blocks, explained above). But there all simplicity
985  * ends, TCP might receive valid D-SACKs below that. As long as they reside
986  * fully below undo_marker they do not affect behavior in anyway and can
987  * therefore be safely ignored. In rare cases (which are more or less
988  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
989  * fragmentation and packet reordering past skb's retransmission. To consider
990  * them correctly, the acceptable range must be extended even more though
991  * the exact amount is rather hard to quantify. However, tp->max_window can
992  * be used as an exaggerated estimate.
993  */
994 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
995                                    u32 start_seq, u32 end_seq)
996 {
997         /* Too far in future, or reversed (interpretation is ambiguous) */
998         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
999                 return false;
1000 
1001         /* Nasty start_seq wrap-around check (see comments above) */
1002         if (!before(start_seq, tp->snd_nxt))
1003                 return false;
1004 
1005         /* In outstanding window? ...This is valid exit for D-SACKs too.
1006          * start_seq == snd_una is non-sensical (see comments above)
1007          */
1008         if (after(start_seq, tp->snd_una))
1009                 return true;
1010 
1011         if (!is_dsack || !tp->undo_marker)
1012                 return false;
1013 
1014         /* ...Then it's D-SACK, and must reside below snd_una completely */
1015         if (after(end_seq, tp->snd_una))
1016                 return false;
1017 
1018         if (!before(start_seq, tp->undo_marker))
1019                 return true;
1020 
1021         /* Too old */
1022         if (!after(end_seq, tp->undo_marker))
1023                 return false;
1024 
1025         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1026          *   start_seq < undo_marker and end_seq >= undo_marker.
1027          */
1028         return !before(start_seq, end_seq - tp->max_window);
1029 }
1030 
1031 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1032  * Event "B". Later note: FACK people cheated me again 8), we have to account
1033  * for reordering! Ugly, but should help.
1034  *
1035  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1036  * less than what is now known to be received by the other end (derived from
1037  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1038  * retransmitted skbs to avoid some costly processing per ACKs.
1039  */
1040 static void tcp_mark_lost_retrans(struct sock *sk)
1041 {
1042         const struct inet_connection_sock *icsk = inet_csk(sk);
1043         struct tcp_sock *tp = tcp_sk(sk);
1044         struct sk_buff *skb;
1045         int cnt = 0;
1046         u32 new_low_seq = tp->snd_nxt;
1047         u32 received_upto = tcp_highest_sack_seq(tp);
1048 
1049         if (!tcp_is_fack(tp) || !tp->retrans_out ||
1050             !after(received_upto, tp->lost_retrans_low) ||
1051             icsk->icsk_ca_state != TCP_CA_Recovery)
1052                 return;
1053 
1054         tcp_for_write_queue(skb, sk) {
1055                 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1056 
1057                 if (skb == tcp_send_head(sk))
1058                         break;
1059                 if (cnt == tp->retrans_out)
1060                         break;
1061                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1062                         continue;
1063 
1064                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1065                         continue;
1066 
1067                 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1068                  * constraint here (see above) but figuring out that at
1069                  * least tp->reordering SACK blocks reside between ack_seq
1070                  * and received_upto is not easy task to do cheaply with
1071                  * the available datastructures.
1072                  *
1073                  * Whether FACK should check here for tp->reordering segs
1074                  * in-between one could argue for either way (it would be
1075                  * rather simple to implement as we could count fack_count
1076                  * during the walk and do tp->fackets_out - fack_count).
1077                  */
1078                 if (after(received_upto, ack_seq)) {
1079                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1080                         tp->retrans_out -= tcp_skb_pcount(skb);
1081 
1082                         tcp_skb_mark_lost_uncond_verify(tp, skb);
1083                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1084                 } else {
1085                         if (before(ack_seq, new_low_seq))
1086                                 new_low_seq = ack_seq;
1087                         cnt += tcp_skb_pcount(skb);
1088                 }
1089         }
1090 
1091         if (tp->retrans_out)
1092                 tp->lost_retrans_low = new_low_seq;
1093 }
1094 
1095 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1096                             struct tcp_sack_block_wire *sp, int num_sacks,
1097                             u32 prior_snd_una)
1098 {
1099         struct tcp_sock *tp = tcp_sk(sk);
1100         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1101         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1102         bool dup_sack = false;
1103 
1104         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1105                 dup_sack = true;
1106                 tcp_dsack_seen(tp);
1107                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1108         } else if (num_sacks > 1) {
1109                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1110                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1111 
1112                 if (!after(end_seq_0, end_seq_1) &&
1113                     !before(start_seq_0, start_seq_1)) {
1114                         dup_sack = true;
1115                         tcp_dsack_seen(tp);
1116                         NET_INC_STATS_BH(sock_net(sk),
1117                                         LINUX_MIB_TCPDSACKOFORECV);
1118                 }
1119         }
1120 
1121         /* D-SACK for already forgotten data... Do dumb counting. */
1122         if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1123             !after(end_seq_0, prior_snd_una) &&
1124             after(end_seq_0, tp->undo_marker))
1125                 tp->undo_retrans--;
1126 
1127         return dup_sack;
1128 }
1129 
1130 struct tcp_sacktag_state {
1131         int     reord;
1132         int     fack_count;
1133         long    rtt_us; /* RTT measured by SACKing never-retransmitted data */
1134         int     flag;
1135 };
1136 
1137 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1138  * the incoming SACK may not exactly match but we can find smaller MSS
1139  * aligned portion of it that matches. Therefore we might need to fragment
1140  * which may fail and creates some hassle (caller must handle error case
1141  * returns).
1142  *
1143  * FIXME: this could be merged to shift decision code
1144  */
1145 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1146                                   u32 start_seq, u32 end_seq)
1147 {
1148         int err;
1149         bool in_sack;
1150         unsigned int pkt_len;
1151         unsigned int mss;
1152 
1153         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1154                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1155 
1156         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1157             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1158                 mss = tcp_skb_mss(skb);
1159                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1160 
1161                 if (!in_sack) {
1162                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1163                         if (pkt_len < mss)
1164                                 pkt_len = mss;
1165                 } else {
1166                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1167                         if (pkt_len < mss)
1168                                 return -EINVAL;
1169                 }
1170 
1171                 /* Round if necessary so that SACKs cover only full MSSes
1172                  * and/or the remaining small portion (if present)
1173                  */
1174                 if (pkt_len > mss) {
1175                         unsigned int new_len = (pkt_len / mss) * mss;
1176                         if (!in_sack && new_len < pkt_len) {
1177                                 new_len += mss;
1178                                 if (new_len >= skb->len)
1179                                         return 0;
1180                         }
1181                         pkt_len = new_len;
1182                 }
1183                 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1184                 if (err < 0)
1185                         return err;
1186         }
1187 
1188         return in_sack;
1189 }
1190 
1191 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1192 static u8 tcp_sacktag_one(struct sock *sk,
1193                           struct tcp_sacktag_state *state, u8 sacked,
1194                           u32 start_seq, u32 end_seq,
1195                           int dup_sack, int pcount,
1196                           const struct skb_mstamp *xmit_time)
1197 {
1198         struct tcp_sock *tp = tcp_sk(sk);
1199         int fack_count = state->fack_count;
1200 
1201         /* Account D-SACK for retransmitted packet. */
1202         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1203                 if (tp->undo_marker && tp->undo_retrans > 0 &&
1204                     after(end_seq, tp->undo_marker))
1205                         tp->undo_retrans--;
1206                 if (sacked & TCPCB_SACKED_ACKED)
1207                         state->reord = min(fack_count, state->reord);
1208         }
1209 
1210         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1211         if (!after(end_seq, tp->snd_una))
1212                 return sacked;
1213 
1214         if (!(sacked & TCPCB_SACKED_ACKED)) {
1215                 if (sacked & TCPCB_SACKED_RETRANS) {
1216                         /* If the segment is not tagged as lost,
1217                          * we do not clear RETRANS, believing
1218                          * that retransmission is still in flight.
1219                          */
1220                         if (sacked & TCPCB_LOST) {
1221                                 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1222                                 tp->lost_out -= pcount;
1223                                 tp->retrans_out -= pcount;
1224                         }
1225                 } else {
1226                         if (!(sacked & TCPCB_RETRANS)) {
1227                                 /* New sack for not retransmitted frame,
1228                                  * which was in hole. It is reordering.
1229                                  */
1230                                 if (before(start_seq,
1231                                            tcp_highest_sack_seq(tp)))
1232                                         state->reord = min(fack_count,
1233                                                            state->reord);
1234                                 if (!after(end_seq, tp->high_seq))
1235                                         state->flag |= FLAG_ORIG_SACK_ACKED;
1236                                 /* Pick the earliest sequence sacked for RTT */
1237                                 if (state->rtt_us < 0) {
1238                                         struct skb_mstamp now;
1239 
1240                                         skb_mstamp_get(&now);
1241                                         state->rtt_us = skb_mstamp_us_delta(&now,
1242                                                                 xmit_time);
1243                                 }
1244                         }
1245 
1246                         if (sacked & TCPCB_LOST) {
1247                                 sacked &= ~TCPCB_LOST;
1248                                 tp->lost_out -= pcount;
1249                         }
1250                 }
1251 
1252                 sacked |= TCPCB_SACKED_ACKED;
1253                 state->flag |= FLAG_DATA_SACKED;
1254                 tp->sacked_out += pcount;
1255 
1256                 fack_count += pcount;
1257 
1258                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1259                 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1260                     before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1261                         tp->lost_cnt_hint += pcount;
1262 
1263                 if (fack_count > tp->fackets_out)
1264                         tp->fackets_out = fack_count;
1265         }
1266 
1267         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1268          * frames and clear it. undo_retrans is decreased above, L|R frames
1269          * are accounted above as well.
1270          */
1271         if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1272                 sacked &= ~TCPCB_SACKED_RETRANS;
1273                 tp->retrans_out -= pcount;
1274         }
1275 
1276         return sacked;
1277 }
1278 
1279 /* Shift newly-SACKed bytes from this skb to the immediately previous
1280  * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1281  */
1282 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1283                             struct tcp_sacktag_state *state,
1284                             unsigned int pcount, int shifted, int mss,
1285                             bool dup_sack)
1286 {
1287         struct tcp_sock *tp = tcp_sk(sk);
1288         struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1289         u32 start_seq = TCP_SKB_CB(skb)->seq;   /* start of newly-SACKed */
1290         u32 end_seq = start_seq + shifted;      /* end of newly-SACKed */
1291 
1292         BUG_ON(!pcount);
1293 
1294         /* Adjust counters and hints for the newly sacked sequence
1295          * range but discard the return value since prev is already
1296          * marked. We must tag the range first because the seq
1297          * advancement below implicitly advances
1298          * tcp_highest_sack_seq() when skb is highest_sack.
1299          */
1300         tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1301                         start_seq, end_seq, dup_sack, pcount,
1302                         &skb->skb_mstamp);
1303 
1304         if (skb == tp->lost_skb_hint)
1305                 tp->lost_cnt_hint += pcount;
1306 
1307         TCP_SKB_CB(prev)->end_seq += shifted;
1308         TCP_SKB_CB(skb)->seq += shifted;
1309 
1310         tcp_skb_pcount_add(prev, pcount);
1311         BUG_ON(tcp_skb_pcount(skb) < pcount);
1312         tcp_skb_pcount_add(skb, -pcount);
1313 
1314         /* When we're adding to gso_segs == 1, gso_size will be zero,
1315          * in theory this shouldn't be necessary but as long as DSACK
1316          * code can come after this skb later on it's better to keep
1317          * setting gso_size to something.
1318          */
1319         if (!skb_shinfo(prev)->gso_size) {
1320                 skb_shinfo(prev)->gso_size = mss;
1321                 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1322         }
1323 
1324         /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1325         if (tcp_skb_pcount(skb) <= 1) {
1326                 skb_shinfo(skb)->gso_size = 0;
1327                 skb_shinfo(skb)->gso_type = 0;
1328         }
1329 
1330         /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1331         TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1332 
1333         if (skb->len > 0) {
1334                 BUG_ON(!tcp_skb_pcount(skb));
1335                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1336                 return false;
1337         }
1338 
1339         /* Whole SKB was eaten :-) */
1340 
1341         if (skb == tp->retransmit_skb_hint)
1342                 tp->retransmit_skb_hint = prev;
1343         if (skb == tp->lost_skb_hint) {
1344                 tp->lost_skb_hint = prev;
1345                 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1346         }
1347 
1348         TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1349         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1350                 TCP_SKB_CB(prev)->end_seq++;
1351 
1352         if (skb == tcp_highest_sack(sk))
1353                 tcp_advance_highest_sack(sk, skb);
1354 
1355         tcp_unlink_write_queue(skb, sk);
1356         sk_wmem_free_skb(sk, skb);
1357 
1358         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1359 
1360         return true;
1361 }
1362 
1363 /* I wish gso_size would have a bit more sane initialization than
1364  * something-or-zero which complicates things
1365  */
1366 static int tcp_skb_seglen(const struct sk_buff *skb)
1367 {
1368         return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1369 }
1370 
1371 /* Shifting pages past head area doesn't work */
1372 static int skb_can_shift(const struct sk_buff *skb)
1373 {
1374         return !skb_headlen(skb) && skb_is_nonlinear(skb);
1375 }
1376 
1377 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1378  * skb.
1379  */
1380 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1381                                           struct tcp_sacktag_state *state,
1382                                           u32 start_seq, u32 end_seq,
1383                                           bool dup_sack)
1384 {
1385         struct tcp_sock *tp = tcp_sk(sk);
1386         struct sk_buff *prev;
1387         int mss;
1388         int pcount = 0;
1389         int len;
1390         int in_sack;
1391 
1392         if (!sk_can_gso(sk))
1393                 goto fallback;
1394 
1395         /* Normally R but no L won't result in plain S */
1396         if (!dup_sack &&
1397             (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1398                 goto fallback;
1399         if (!skb_can_shift(skb))
1400                 goto fallback;
1401         /* This frame is about to be dropped (was ACKed). */
1402         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1403                 goto fallback;
1404 
1405         /* Can only happen with delayed DSACK + discard craziness */
1406         if (unlikely(skb == tcp_write_queue_head(sk)))
1407                 goto fallback;
1408         prev = tcp_write_queue_prev(sk, skb);
1409 
1410         if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1411                 goto fallback;
1412 
1413         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1414                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1415 
1416         if (in_sack) {
1417                 len = skb->len;
1418                 pcount = tcp_skb_pcount(skb);
1419                 mss = tcp_skb_seglen(skb);
1420 
1421                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1422                  * drop this restriction as unnecessary
1423                  */
1424                 if (mss != tcp_skb_seglen(prev))
1425                         goto fallback;
1426         } else {
1427                 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1428                         goto noop;
1429                 /* CHECKME: This is non-MSS split case only?, this will
1430                  * cause skipped skbs due to advancing loop btw, original
1431                  * has that feature too
1432                  */
1433                 if (tcp_skb_pcount(skb) <= 1)
1434                         goto noop;
1435 
1436                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1437                 if (!in_sack) {
1438                         /* TODO: head merge to next could be attempted here
1439                          * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1440                          * though it might not be worth of the additional hassle
1441                          *
1442                          * ...we can probably just fallback to what was done
1443                          * previously. We could try merging non-SACKed ones
1444                          * as well but it probably isn't going to buy off
1445                          * because later SACKs might again split them, and
1446                          * it would make skb timestamp tracking considerably
1447                          * harder problem.
1448                          */
1449                         goto fallback;
1450                 }
1451 
1452                 len = end_seq - TCP_SKB_CB(skb)->seq;
1453                 BUG_ON(len < 0);
1454                 BUG_ON(len > skb->len);
1455 
1456                 /* MSS boundaries should be honoured or else pcount will
1457                  * severely break even though it makes things bit trickier.
1458                  * Optimize common case to avoid most of the divides
1459                  */
1460                 mss = tcp_skb_mss(skb);
1461 
1462                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1463                  * drop this restriction as unnecessary
1464                  */
1465                 if (mss != tcp_skb_seglen(prev))
1466                         goto fallback;
1467 
1468                 if (len == mss) {
1469                         pcount = 1;
1470                 } else if (len < mss) {
1471                         goto noop;
1472                 } else {
1473                         pcount = len / mss;
1474                         len = pcount * mss;
1475                 }
1476         }
1477 
1478         /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1479         if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1480                 goto fallback;
1481 
1482         if (!skb_shift(prev, skb, len))
1483                 goto fallback;
1484         if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1485                 goto out;
1486 
1487         /* Hole filled allows collapsing with the next as well, this is very
1488          * useful when hole on every nth skb pattern happens
1489          */
1490         if (prev == tcp_write_queue_tail(sk))
1491                 goto out;
1492         skb = tcp_write_queue_next(sk, prev);
1493 
1494         if (!skb_can_shift(skb) ||
1495             (skb == tcp_send_head(sk)) ||
1496             ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1497             (mss != tcp_skb_seglen(skb)))
1498                 goto out;
1499 
1500         len = skb->len;
1501         if (skb_shift(prev, skb, len)) {
1502                 pcount += tcp_skb_pcount(skb);
1503                 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1504         }
1505 
1506 out:
1507         state->fack_count += pcount;
1508         return prev;
1509 
1510 noop:
1511         return skb;
1512 
1513 fallback:
1514         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1515         return NULL;
1516 }
1517 
1518 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1519                                         struct tcp_sack_block *next_dup,
1520                                         struct tcp_sacktag_state *state,
1521                                         u32 start_seq, u32 end_seq,
1522                                         bool dup_sack_in)
1523 {
1524         struct tcp_sock *tp = tcp_sk(sk);
1525         struct sk_buff *tmp;
1526 
1527         tcp_for_write_queue_from(skb, sk) {
1528                 int in_sack = 0;
1529                 bool dup_sack = dup_sack_in;
1530 
1531                 if (skb == tcp_send_head(sk))
1532                         break;
1533 
1534                 /* queue is in-order => we can short-circuit the walk early */
1535                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1536                         break;
1537 
1538                 if (next_dup  &&
1539                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1540                         in_sack = tcp_match_skb_to_sack(sk, skb,
1541                                                         next_dup->start_seq,
1542                                                         next_dup->end_seq);
1543                         if (in_sack > 0)
1544                                 dup_sack = true;
1545                 }
1546 
1547                 /* skb reference here is a bit tricky to get right, since
1548                  * shifting can eat and free both this skb and the next,
1549                  * so not even _safe variant of the loop is enough.
1550                  */
1551                 if (in_sack <= 0) {
1552                         tmp = tcp_shift_skb_data(sk, skb, state,
1553                                                  start_seq, end_seq, dup_sack);
1554                         if (tmp) {
1555                                 if (tmp != skb) {
1556                                         skb = tmp;
1557                                         continue;
1558                                 }
1559 
1560                                 in_sack = 0;
1561                         } else {
1562                                 in_sack = tcp_match_skb_to_sack(sk, skb,
1563                                                                 start_seq,
1564                                                                 end_seq);
1565                         }
1566                 }
1567 
1568                 if (unlikely(in_sack < 0))
1569                         break;
1570 
1571                 if (in_sack) {
1572                         TCP_SKB_CB(skb)->sacked =
1573                                 tcp_sacktag_one(sk,
1574                                                 state,
1575                                                 TCP_SKB_CB(skb)->sacked,
1576                                                 TCP_SKB_CB(skb)->seq,
1577                                                 TCP_SKB_CB(skb)->end_seq,
1578                                                 dup_sack,
1579                                                 tcp_skb_pcount(skb),
1580                                                 &skb->skb_mstamp);
1581 
1582                         if (!before(TCP_SKB_CB(skb)->seq,
1583                                     tcp_highest_sack_seq(tp)))
1584                                 tcp_advance_highest_sack(sk, skb);
1585                 }
1586 
1587                 state->fack_count += tcp_skb_pcount(skb);
1588         }
1589         return skb;
1590 }
1591 
1592 /* Avoid all extra work that is being done by sacktag while walking in
1593  * a normal way
1594  */
1595 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1596                                         struct tcp_sacktag_state *state,
1597                                         u32 skip_to_seq)
1598 {
1599         tcp_for_write_queue_from(skb, sk) {
1600                 if (skb == tcp_send_head(sk))
1601                         break;
1602 
1603                 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1604                         break;
1605 
1606                 state->fack_count += tcp_skb_pcount(skb);
1607         }
1608         return skb;
1609 }
1610 
1611 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1612                                                 struct sock *sk,
1613                                                 struct tcp_sack_block *next_dup,
1614                                                 struct tcp_sacktag_state *state,
1615                                                 u32 skip_to_seq)
1616 {
1617         if (!next_dup)
1618                 return skb;
1619 
1620         if (before(next_dup->start_seq, skip_to_seq)) {
1621                 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1622                 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1623                                        next_dup->start_seq, next_dup->end_seq,
1624                                        1);
1625         }
1626 
1627         return skb;
1628 }
1629 
1630 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1631 {
1632         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1633 }
1634 
1635 static int
1636 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1637                         u32 prior_snd_una, long *sack_rtt_us)
1638 {
1639         struct tcp_sock *tp = tcp_sk(sk);
1640         const unsigned char *ptr = (skb_transport_header(ack_skb) +
1641                                     TCP_SKB_CB(ack_skb)->sacked);
1642         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1643         struct tcp_sack_block sp[TCP_NUM_SACKS];
1644         struct tcp_sack_block *cache;
1645         struct tcp_sacktag_state state;
1646         struct sk_buff *skb;
1647         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1648         int used_sacks;
1649         bool found_dup_sack = false;
1650         int i, j;
1651         int first_sack_index;
1652 
1653         state.flag = 0;
1654         state.reord = tp->packets_out;
1655         state.rtt_us = -1L;
1656 
1657         if (!tp->sacked_out) {
1658                 if (WARN_ON(tp->fackets_out))
1659                         tp->fackets_out = 0;
1660                 tcp_highest_sack_reset(sk);
1661         }
1662 
1663         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1664                                          num_sacks, prior_snd_una);
1665         if (found_dup_sack)
1666                 state.flag |= FLAG_DSACKING_ACK;
1667 
1668         /* Eliminate too old ACKs, but take into
1669          * account more or less fresh ones, they can
1670          * contain valid SACK info.
1671          */
1672         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1673                 return 0;
1674 
1675         if (!tp->packets_out)
1676                 goto out;
1677 
1678         used_sacks = 0;
1679         first_sack_index = 0;
1680         for (i = 0; i < num_sacks; i++) {
1681                 bool dup_sack = !i && found_dup_sack;
1682 
1683                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1684                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1685 
1686                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1687                                             sp[used_sacks].start_seq,
1688                                             sp[used_sacks].end_seq)) {
1689                         int mib_idx;
1690 
1691                         if (dup_sack) {
1692                                 if (!tp->undo_marker)
1693                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1694                                 else
1695                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1696                         } else {
1697                                 /* Don't count olds caused by ACK reordering */
1698                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1699                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1700                                         continue;
1701                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1702                         }
1703 
1704                         NET_INC_STATS_BH(sock_net(sk), mib_idx);
1705                         if (i == 0)
1706                                 first_sack_index = -1;
1707                         continue;
1708                 }
1709 
1710                 /* Ignore very old stuff early */
1711                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1712                         continue;
1713 
1714                 used_sacks++;
1715         }
1716 
1717         /* order SACK blocks to allow in order walk of the retrans queue */
1718         for (i = used_sacks - 1; i > 0; i--) {
1719                 for (j = 0; j < i; j++) {
1720                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1721                                 swap(sp[j], sp[j + 1]);
1722 
1723                                 /* Track where the first SACK block goes to */
1724                                 if (j == first_sack_index)
1725                                         first_sack_index = j + 1;
1726                         }
1727                 }
1728         }
1729 
1730         skb = tcp_write_queue_head(sk);
1731         state.fack_count = 0;
1732         i = 0;
1733 
1734         if (!tp->sacked_out) {
1735                 /* It's already past, so skip checking against it */
1736                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1737         } else {
1738                 cache = tp->recv_sack_cache;
1739                 /* Skip empty blocks in at head of the cache */
1740                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1741                        !cache->end_seq)
1742                         cache++;
1743         }
1744 
1745         while (i < used_sacks) {
1746                 u32 start_seq = sp[i].start_seq;
1747                 u32 end_seq = sp[i].end_seq;
1748                 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1749                 struct tcp_sack_block *next_dup = NULL;
1750 
1751                 if (found_dup_sack && ((i + 1) == first_sack_index))
1752                         next_dup = &sp[i + 1];
1753 
1754                 /* Skip too early cached blocks */
1755                 while (tcp_sack_cache_ok(tp, cache) &&
1756                        !before(start_seq, cache->end_seq))
1757                         cache++;
1758 
1759                 /* Can skip some work by looking recv_sack_cache? */
1760                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1761                     after(end_seq, cache->start_seq)) {
1762 
1763                         /* Head todo? */
1764                         if (before(start_seq, cache->start_seq)) {
1765                                 skb = tcp_sacktag_skip(skb, sk, &state,
1766                                                        start_seq);
1767                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1768                                                        &state,
1769                                                        start_seq,
1770                                                        cache->start_seq,
1771                                                        dup_sack);
1772                         }
1773 
1774                         /* Rest of the block already fully processed? */
1775                         if (!after(end_seq, cache->end_seq))
1776                                 goto advance_sp;
1777 
1778                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1779                                                        &state,
1780                                                        cache->end_seq);
1781 
1782                         /* ...tail remains todo... */
1783                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1784                                 /* ...but better entrypoint exists! */
1785                                 skb = tcp_highest_sack(sk);
1786                                 if (!skb)
1787                                         break;
1788                                 state.fack_count = tp->fackets_out;
1789                                 cache++;
1790                                 goto walk;
1791                         }
1792 
1793                         skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1794                         /* Check overlap against next cached too (past this one already) */
1795                         cache++;
1796                         continue;
1797                 }
1798 
1799                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1800                         skb = tcp_highest_sack(sk);
1801                         if (!skb)
1802                                 break;
1803                         state.fack_count = tp->fackets_out;
1804                 }
1805                 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1806 
1807 walk:
1808                 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1809                                        start_seq, end_seq, dup_sack);
1810 
1811 advance_sp:
1812                 i++;
1813         }
1814 
1815         /* Clear the head of the cache sack blocks so we can skip it next time */
1816         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1817                 tp->recv_sack_cache[i].start_seq = 0;
1818                 tp->recv_sack_cache[i].end_seq = 0;
1819         }
1820         for (j = 0; j < used_sacks; j++)
1821                 tp->recv_sack_cache[i++] = sp[j];
1822 
1823         if ((state.reord < tp->fackets_out) &&
1824             ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1825                 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1826 
1827         tcp_mark_lost_retrans(sk);
1828         tcp_verify_left_out(tp);
1829 out:
1830 
1831 #if FASTRETRANS_DEBUG > 0
1832         WARN_ON((int)tp->sacked_out < 0);
1833         WARN_ON((int)tp->lost_out < 0);
1834         WARN_ON((int)tp->retrans_out < 0);
1835         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1836 #endif
1837         *sack_rtt_us = state.rtt_us;
1838         return state.flag;
1839 }
1840 
1841 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1842  * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1843  */
1844 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1845 {
1846         u32 holes;
1847 
1848         holes = max(tp->lost_out, 1U);
1849         holes = min(holes, tp->packets_out);
1850 
1851         if ((tp->sacked_out + holes) > tp->packets_out) {
1852                 tp->sacked_out = tp->packets_out - holes;
1853                 return true;
1854         }
1855         return false;
1856 }
1857 
1858 /* If we receive more dupacks than we expected counting segments
1859  * in assumption of absent reordering, interpret this as reordering.
1860  * The only another reason could be bug in receiver TCP.
1861  */
1862 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1863 {
1864         struct tcp_sock *tp = tcp_sk(sk);
1865         if (tcp_limit_reno_sacked(tp))
1866                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1867 }
1868 
1869 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1870 
1871 static void tcp_add_reno_sack(struct sock *sk)
1872 {
1873         struct tcp_sock *tp = tcp_sk(sk);
1874         tp->sacked_out++;
1875         tcp_check_reno_reordering(sk, 0);
1876         tcp_verify_left_out(tp);
1877 }
1878 
1879 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1880 
1881 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1882 {
1883         struct tcp_sock *tp = tcp_sk(sk);
1884 
1885         if (acked > 0) {
1886                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1887                 if (acked - 1 >= tp->sacked_out)
1888                         tp->sacked_out = 0;
1889                 else
1890                         tp->sacked_out -= acked - 1;
1891         }
1892         tcp_check_reno_reordering(sk, acked);
1893         tcp_verify_left_out(tp);
1894 }
1895 
1896 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1897 {
1898         tp->sacked_out = 0;
1899 }
1900 
1901 void tcp_clear_retrans(struct tcp_sock *tp)
1902 {
1903         tp->retrans_out = 0;
1904         tp->lost_out = 0;
1905         tp->undo_marker = 0;
1906         tp->undo_retrans = -1;
1907         tp->fackets_out = 0;
1908         tp->sacked_out = 0;
1909 }
1910 
1911 static inline void tcp_init_undo(struct tcp_sock *tp)
1912 {
1913         tp->undo_marker = tp->snd_una;
1914         /* Retransmission still in flight may cause DSACKs later. */
1915         tp->undo_retrans = tp->retrans_out ? : -1;
1916 }
1917 
1918 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1919  * and reset tags completely, otherwise preserve SACKs. If receiver
1920  * dropped its ofo queue, we will know this due to reneging detection.
1921  */
1922 void tcp_enter_loss(struct sock *sk)
1923 {
1924         const struct inet_connection_sock *icsk = inet_csk(sk);
1925         struct tcp_sock *tp = tcp_sk(sk);
1926         struct sk_buff *skb;
1927         bool new_recovery = false;
1928         bool is_reneg;                  /* is receiver reneging on SACKs? */
1929 
1930         /* Reduce ssthresh if it has not yet been made inside this window. */
1931         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1932             !after(tp->high_seq, tp->snd_una) ||
1933             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1934                 new_recovery = true;
1935                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1936                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1937                 tcp_ca_event(sk, CA_EVENT_LOSS);
1938                 tcp_init_undo(tp);
1939         }
1940         tp->snd_cwnd       = 1;
1941         tp->snd_cwnd_cnt   = 0;
1942         tp->snd_cwnd_stamp = tcp_time_stamp;
1943 
1944         tp->retrans_out = 0;
1945         tp->lost_out = 0;
1946 
1947         if (tcp_is_reno(tp))
1948                 tcp_reset_reno_sack(tp);
1949 
1950         skb = tcp_write_queue_head(sk);
1951         is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1952         if (is_reneg) {
1953                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1954                 tp->sacked_out = 0;
1955                 tp->fackets_out = 0;
1956         }
1957         tcp_clear_all_retrans_hints(tp);
1958 
1959         tcp_for_write_queue(skb, sk) {
1960                 if (skb == tcp_send_head(sk))
1961                         break;
1962 
1963                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1964                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
1965                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1966                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1967                         tp->lost_out += tcp_skb_pcount(skb);
1968                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1969                 }
1970         }
1971         tcp_verify_left_out(tp);
1972 
1973         /* Timeout in disordered state after receiving substantial DUPACKs
1974          * suggests that the degree of reordering is over-estimated.
1975          */
1976         if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1977             tp->sacked_out >= sysctl_tcp_reordering)
1978                 tp->reordering = min_t(unsigned int, tp->reordering,
1979                                        sysctl_tcp_reordering);
1980         tcp_set_ca_state(sk, TCP_CA_Loss);
1981         tp->high_seq = tp->snd_nxt;
1982         tcp_ecn_queue_cwr(tp);
1983 
1984         /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1985          * loss recovery is underway except recurring timeout(s) on
1986          * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1987          */
1988         tp->frto = sysctl_tcp_frto &&
1989                    (new_recovery || icsk->icsk_retransmits) &&
1990                    !inet_csk(sk)->icsk_mtup.probe_size;
1991 }
1992 
1993 /* If ACK arrived pointing to a remembered SACK, it means that our
1994  * remembered SACKs do not reflect real state of receiver i.e.
1995  * receiver _host_ is heavily congested (or buggy).
1996  *
1997  * To avoid big spurious retransmission bursts due to transient SACK
1998  * scoreboard oddities that look like reneging, we give the receiver a
1999  * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2000  * restore sanity to the SACK scoreboard. If the apparent reneging
2001  * persists until this RTO then we'll clear the SACK scoreboard.
2002  */
2003 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2004 {
2005         if (flag & FLAG_SACK_RENEGING) {
2006                 struct tcp_sock *tp = tcp_sk(sk);
2007                 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2008                                           msecs_to_jiffies(10));
2009 
2010                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2011                                           delay, TCP_RTO_MAX);
2012                 return true;
2013         }
2014         return false;
2015 }
2016 
2017 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2018 {
2019         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2020 }
2021 
2022 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2023  * counter when SACK is enabled (without SACK, sacked_out is used for
2024  * that purpose).
2025  *
2026  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2027  * segments up to the highest received SACK block so far and holes in
2028  * between them.
2029  *
2030  * With reordering, holes may still be in flight, so RFC3517 recovery
2031  * uses pure sacked_out (total number of SACKed segments) even though
2032  * it violates the RFC that uses duplicate ACKs, often these are equal
2033  * but when e.g. out-of-window ACKs or packet duplication occurs,
2034  * they differ. Since neither occurs due to loss, TCP should really
2035  * ignore them.
2036  */
2037 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2038 {
2039         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2040 }
2041 
2042 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2043 {
2044         struct tcp_sock *tp = tcp_sk(sk);
2045         unsigned long delay;
2046 
2047         /* Delay early retransmit and entering fast recovery for
2048          * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2049          * available, or RTO is scheduled to fire first.
2050          */
2051         if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2052             (flag & FLAG_ECE) || !tp->srtt_us)
2053                 return false;
2054 
2055         delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2056                     msecs_to_jiffies(2));
2057 
2058         if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2059                 return false;
2060 
2061         inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2062                                   TCP_RTO_MAX);
2063         return true;
2064 }
2065 
2066 /* Linux NewReno/SACK/FACK/ECN state machine.
2067  * --------------------------------------
2068  *
2069  * "Open"       Normal state, no dubious events, fast path.
2070  * "Disorder"   In all the respects it is "Open",
2071  *              but requires a bit more attention. It is entered when
2072  *              we see some SACKs or dupacks. It is split of "Open"
2073  *              mainly to move some processing from fast path to slow one.
2074  * "CWR"        CWND was reduced due to some Congestion Notification event.
2075  *              It can be ECN, ICMP source quench, local device congestion.
2076  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2077  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2078  *
2079  * tcp_fastretrans_alert() is entered:
2080  * - each incoming ACK, if state is not "Open"
2081  * - when arrived ACK is unusual, namely:
2082  *      * SACK
2083  *      * Duplicate ACK.
2084  *      * ECN ECE.
2085  *
2086  * Counting packets in flight is pretty simple.
2087  *
2088  *      in_flight = packets_out - left_out + retrans_out
2089  *
2090  *      packets_out is SND.NXT-SND.UNA counted in packets.
2091  *
2092  *      retrans_out is number of retransmitted segments.
2093  *
2094  *      left_out is number of segments left network, but not ACKed yet.
2095  *
2096  *              left_out = sacked_out + lost_out
2097  *
2098  *     sacked_out: Packets, which arrived to receiver out of order
2099  *                 and hence not ACKed. With SACKs this number is simply
2100  *                 amount of SACKed data. Even without SACKs
2101  *                 it is easy to give pretty reliable estimate of this number,
2102  *                 counting duplicate ACKs.
2103  *
2104  *       lost_out: Packets lost by network. TCP has no explicit
2105  *                 "loss notification" feedback from network (for now).
2106  *                 It means that this number can be only _guessed_.
2107  *                 Actually, it is the heuristics to predict lossage that
2108  *                 distinguishes different algorithms.
2109  *
2110  *      F.e. after RTO, when all the queue is considered as lost,
2111  *      lost_out = packets_out and in_flight = retrans_out.
2112  *
2113  *              Essentially, we have now two algorithms counting
2114  *              lost packets.
2115  *
2116  *              FACK: It is the simplest heuristics. As soon as we decided
2117  *              that something is lost, we decide that _all_ not SACKed
2118  *              packets until the most forward SACK are lost. I.e.
2119  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2120  *              It is absolutely correct estimate, if network does not reorder
2121  *              packets. And it loses any connection to reality when reordering
2122  *              takes place. We use FACK by default until reordering
2123  *              is suspected on the path to this destination.
2124  *
2125  *              NewReno: when Recovery is entered, we assume that one segment
2126  *              is lost (classic Reno). While we are in Recovery and
2127  *              a partial ACK arrives, we assume that one more packet
2128  *              is lost (NewReno). This heuristics are the same in NewReno
2129  *              and SACK.
2130  *
2131  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2132  *  deflation etc. CWND is real congestion window, never inflated, changes
2133  *  only according to classic VJ rules.
2134  *
2135  * Really tricky (and requiring careful tuning) part of algorithm
2136  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2137  * The first determines the moment _when_ we should reduce CWND and,
2138  * hence, slow down forward transmission. In fact, it determines the moment
2139  * when we decide that hole is caused by loss, rather than by a reorder.
2140  *
2141  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2142  * holes, caused by lost packets.
2143  *
2144  * And the most logically complicated part of algorithm is undo
2145  * heuristics. We detect false retransmits due to both too early
2146  * fast retransmit (reordering) and underestimated RTO, analyzing
2147  * timestamps and D-SACKs. When we detect that some segments were
2148  * retransmitted by mistake and CWND reduction was wrong, we undo
2149  * window reduction and abort recovery phase. This logic is hidden
2150  * inside several functions named tcp_try_undo_<something>.
2151  */
2152 
2153 /* This function decides, when we should leave Disordered state
2154  * and enter Recovery phase, reducing congestion window.
2155  *
2156  * Main question: may we further continue forward transmission
2157  * with the same cwnd?
2158  */
2159 static bool tcp_time_to_recover(struct sock *sk, int flag)
2160 {
2161         struct tcp_sock *tp = tcp_sk(sk);
2162         __u32 packets_out;
2163 
2164         /* Trick#1: The loss is proven. */
2165         if (tp->lost_out)
2166                 return true;
2167 
2168         /* Not-A-Trick#2 : Classic rule... */
2169         if (tcp_dupack_heuristics(tp) > tp->reordering)
2170                 return true;
2171 
2172         /* Trick#4: It is still not OK... But will it be useful to delay
2173          * recovery more?
2174          */
2175         packets_out = tp->packets_out;
2176         if (packets_out <= tp->reordering &&
2177             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2178             !tcp_may_send_now(sk)) {
2179                 /* We have nothing to send. This connection is limited
2180                  * either by receiver window or by application.
2181                  */
2182                 return true;
2183         }
2184 
2185         /* If a thin stream is detected, retransmit after first
2186          * received dupack. Employ only if SACK is supported in order
2187          * to avoid possible corner-case series of spurious retransmissions
2188          * Use only if there are no unsent data.
2189          */
2190         if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2191             tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2192             tcp_is_sack(tp) && !tcp_send_head(sk))
2193                 return true;
2194 
2195         /* Trick#6: TCP early retransmit, per RFC5827.  To avoid spurious
2196          * retransmissions due to small network reorderings, we implement
2197          * Mitigation A.3 in the RFC and delay the retransmission for a short
2198          * interval if appropriate.
2199          */
2200         if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2201             (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2202             !tcp_may_send_now(sk))
2203                 return !tcp_pause_early_retransmit(sk, flag);
2204 
2205         return false;
2206 }
2207 
2208 /* Detect loss in event "A" above by marking head of queue up as lost.
2209  * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2210  * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2211  * has at least tp->reordering SACKed seqments above it; "packets" refers to
2212  * the maximum SACKed segments to pass before reaching this limit.
2213  */
2214 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2215 {
2216         struct tcp_sock *tp = tcp_sk(sk);
2217         struct sk_buff *skb;
2218         int cnt, oldcnt;
2219         int err;
2220         unsigned int mss;
2221         /* Use SACK to deduce losses of new sequences sent during recovery */
2222         const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2223 
2224         WARN_ON(packets > tp->packets_out);
2225         if (tp->lost_skb_hint) {
2226                 skb = tp->lost_skb_hint;
2227                 cnt = tp->lost_cnt_hint;
2228                 /* Head already handled? */
2229                 if (mark_head && skb != tcp_write_queue_head(sk))
2230                         return;
2231         } else {
2232                 skb = tcp_write_queue_head(sk);
2233                 cnt = 0;
2234         }
2235 
2236         tcp_for_write_queue_from(skb, sk) {
2237                 if (skb == tcp_send_head(sk))
2238                         break;
2239                 /* TODO: do this better */
2240                 /* this is not the most efficient way to do this... */
2241                 tp->lost_skb_hint = skb;
2242                 tp->lost_cnt_hint = cnt;
2243 
2244                 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2245                         break;
2246 
2247                 oldcnt = cnt;
2248                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2249                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2250                         cnt += tcp_skb_pcount(skb);
2251 
2252                 if (cnt > packets) {
2253                         if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2254                             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2255                             (oldcnt >= packets))
2256                                 break;
2257 
2258                         mss = skb_shinfo(skb)->gso_size;
2259                         err = tcp_fragment(sk, skb, (packets - oldcnt) * mss,
2260                                            mss, GFP_ATOMIC);
2261                         if (err < 0)
2262                                 break;
2263                         cnt = packets;
2264                 }
2265 
2266                 tcp_skb_mark_lost(tp, skb);
2267 
2268                 if (mark_head)
2269                         break;
2270         }
2271         tcp_verify_left_out(tp);
2272 }
2273 
2274 /* Account newly detected lost packet(s) */
2275 
2276 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2277 {
2278         struct tcp_sock *tp = tcp_sk(sk);
2279 
2280         if (tcp_is_reno(tp)) {
2281                 tcp_mark_head_lost(sk, 1, 1);
2282         } else if (tcp_is_fack(tp)) {
2283                 int lost = tp->fackets_out - tp->reordering;
2284                 if (lost <= 0)
2285                         lost = 1;
2286                 tcp_mark_head_lost(sk, lost, 0);
2287         } else {
2288                 int sacked_upto = tp->sacked_out - tp->reordering;
2289                 if (sacked_upto >= 0)
2290                         tcp_mark_head_lost(sk, sacked_upto, 0);
2291                 else if (fast_rexmit)
2292                         tcp_mark_head_lost(sk, 1, 1);
2293         }
2294 }
2295 
2296 /* CWND moderation, preventing bursts due to too big ACKs
2297  * in dubious situations.
2298  */
2299 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2300 {
2301         tp->snd_cwnd = min(tp->snd_cwnd,
2302                            tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2303         tp->snd_cwnd_stamp = tcp_time_stamp;
2304 }
2305 
2306 /* Nothing was retransmitted or returned timestamp is less
2307  * than timestamp of the first retransmission.
2308  */
2309 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2310 {
2311         return !tp->retrans_stamp ||
2312                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2313                  before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2314 }
2315 
2316 /* Undo procedures. */
2317 
2318 /* We can clear retrans_stamp when there are no retransmissions in the
2319  * window. It would seem that it is trivially available for us in
2320  * tp->retrans_out, however, that kind of assumptions doesn't consider
2321  * what will happen if errors occur when sending retransmission for the
2322  * second time. ...It could the that such segment has only
2323  * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2324  * the head skb is enough except for some reneging corner cases that
2325  * are not worth the effort.
2326  *
2327  * Main reason for all this complexity is the fact that connection dying
2328  * time now depends on the validity of the retrans_stamp, in particular,
2329  * that successive retransmissions of a segment must not advance
2330  * retrans_stamp under any conditions.
2331  */
2332 static bool tcp_any_retrans_done(const struct sock *sk)
2333 {
2334         const struct tcp_sock *tp = tcp_sk(sk);
2335         struct sk_buff *skb;
2336 
2337         if (tp->retrans_out)
2338                 return true;
2339 
2340         skb = tcp_write_queue_head(sk);
2341         if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2342                 return true;
2343 
2344         return false;
2345 }
2346 
2347 #if FASTRETRANS_DEBUG > 1
2348 static void DBGUNDO(struct sock *sk, const char *msg)
2349 {
2350         struct tcp_sock *tp = tcp_sk(sk);
2351         struct inet_sock *inet = inet_sk(sk);
2352 
2353         if (sk->sk_family == AF_INET) {
2354                 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2355                          msg,
2356                          &inet->inet_daddr, ntohs(inet->inet_dport),
2357                          tp->snd_cwnd, tcp_left_out(tp),
2358                          tp->snd_ssthresh, tp->prior_ssthresh,
2359                          tp->packets_out);
2360         }
2361 #if IS_ENABLED(CONFIG_IPV6)
2362         else if (sk->sk_family == AF_INET6) {
2363                 struct ipv6_pinfo *np = inet6_sk(sk);
2364                 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2365                          msg,
2366                          &np->daddr, ntohs(inet->inet_dport),
2367                          tp->snd_cwnd, tcp_left_out(tp),
2368                          tp->snd_ssthresh, tp->prior_ssthresh,
2369                          tp->packets_out);
2370         }
2371 #endif
2372 }
2373 #else
2374 #define DBGUNDO(x...) do { } while (0)
2375 #endif
2376 
2377 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2378 {
2379         struct tcp_sock *tp = tcp_sk(sk);
2380 
2381         if (unmark_loss) {
2382                 struct sk_buff *skb;
2383 
2384                 tcp_for_write_queue(skb, sk) {
2385                         if (skb == tcp_send_head(sk))
2386                                 break;
2387                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2388                 }
2389                 tp->lost_out = 0;
2390                 tcp_clear_all_retrans_hints(tp);
2391         }
2392 
2393         if (tp->prior_ssthresh) {
2394                 const struct inet_connection_sock *icsk = inet_csk(sk);
2395 
2396                 if (icsk->icsk_ca_ops->undo_cwnd)
2397                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2398                 else
2399                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2400 
2401                 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2402                         tp->snd_ssthresh = tp->prior_ssthresh;
2403                         tcp_ecn_withdraw_cwr(tp);
2404                 }
2405         } else {
2406                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2407         }
2408         tp->snd_cwnd_stamp = tcp_time_stamp;
2409         tp->undo_marker = 0;
2410 }
2411 
2412 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2413 {
2414         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2415 }
2416 
2417 /* People celebrate: "We love our President!" */
2418 static bool tcp_try_undo_recovery(struct sock *sk)
2419 {
2420         struct tcp_sock *tp = tcp_sk(sk);
2421 
2422         if (tcp_may_undo(tp)) {
2423                 int mib_idx;
2424 
2425                 /* Happy end! We did not retransmit anything
2426                  * or our original transmission succeeded.
2427                  */
2428                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2429                 tcp_undo_cwnd_reduction(sk, false);
2430                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2431                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2432                 else
2433                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2434 
2435                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2436         }
2437         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2438                 /* Hold old state until something *above* high_seq
2439                  * is ACKed. For Reno it is MUST to prevent false
2440                  * fast retransmits (RFC2582). SACK TCP is safe. */
2441                 tcp_moderate_cwnd(tp);
2442                 if (!tcp_any_retrans_done(sk))
2443                         tp->retrans_stamp = 0;
2444                 return true;
2445         }
2446         tcp_set_ca_state(sk, TCP_CA_Open);
2447         return false;
2448 }
2449 
2450 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2451 static bool tcp_try_undo_dsack(struct sock *sk)
2452 {
2453         struct tcp_sock *tp = tcp_sk(sk);
2454 
2455         if (tp->undo_marker && !tp->undo_retrans) {
2456                 DBGUNDO(sk, "D-SACK");
2457                 tcp_undo_cwnd_reduction(sk, false);
2458                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2459                 return true;
2460         }
2461         return false;
2462 }
2463 
2464 /* Undo during loss recovery after partial ACK or using F-RTO. */
2465 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2466 {
2467         struct tcp_sock *tp = tcp_sk(sk);
2468 
2469         if (frto_undo || tcp_may_undo(tp)) {
2470                 tcp_undo_cwnd_reduction(sk, true);
2471 
2472                 DBGUNDO(sk, "partial loss");
2473                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2474                 if (frto_undo)
2475                         NET_INC_STATS_BH(sock_net(sk),
2476                                          LINUX_MIB_TCPSPURIOUSRTOS);
2477                 inet_csk(sk)->icsk_retransmits = 0;
2478                 if (frto_undo || tcp_is_sack(tp))
2479                         tcp_set_ca_state(sk, TCP_CA_Open);
2480                 return true;
2481         }
2482         return false;
2483 }
2484 
2485 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2486  * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2487  * It computes the number of packets to send (sndcnt) based on packets newly
2488  * delivered:
2489  *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2490  *      cwnd reductions across a full RTT.
2491  *   2) If packets in flight is lower than ssthresh (such as due to excess
2492  *      losses and/or application stalls), do not perform any further cwnd
2493  *      reductions, but instead slow start up to ssthresh.
2494  */
2495 static void tcp_init_cwnd_reduction(struct sock *sk)
2496 {
2497         struct tcp_sock *tp = tcp_sk(sk);
2498 
2499         tp->high_seq = tp->snd_nxt;
2500         tp->tlp_high_seq = 0;
2501         tp->snd_cwnd_cnt = 0;
2502         tp->prior_cwnd = tp->snd_cwnd;
2503         tp->prr_delivered = 0;
2504         tp->prr_out = 0;
2505         tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2506         tcp_ecn_queue_cwr(tp);
2507 }
2508 
2509 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2510                                int fast_rexmit)
2511 {
2512         struct tcp_sock *tp = tcp_sk(sk);
2513         int sndcnt = 0;
2514         int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2515         int newly_acked_sacked = prior_unsacked -
2516                                  (tp->packets_out - tp->sacked_out);
2517 
2518         tp->prr_delivered += newly_acked_sacked;
2519         if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2520                 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2521                                tp->prior_cwnd - 1;
2522                 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2523         } else {
2524                 sndcnt = min_t(int, delta,
2525                                max_t(int, tp->prr_delivered - tp->prr_out,
2526                                      newly_acked_sacked) + 1);
2527         }
2528 
2529         sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2530         tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2531 }
2532 
2533 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2534 {
2535         struct tcp_sock *tp = tcp_sk(sk);
2536 
2537         /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2538         if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2539             (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2540                 tp->snd_cwnd = tp->snd_ssthresh;
2541                 tp->snd_cwnd_stamp = tcp_time_stamp;
2542         }
2543         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2544 }
2545 
2546 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2547 void tcp_enter_cwr(struct sock *sk)
2548 {
2549         struct tcp_sock *tp = tcp_sk(sk);
2550 
2551         tp->prior_ssthresh = 0;
2552         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2553                 tp->undo_marker = 0;
2554                 tcp_init_cwnd_reduction(sk);
2555                 tcp_set_ca_state(sk, TCP_CA_CWR);
2556         }
2557 }
2558 
2559 static void tcp_try_keep_open(struct sock *sk)
2560 {
2561         struct tcp_sock *tp = tcp_sk(sk);
2562         int state = TCP_CA_Open;
2563 
2564         if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2565                 state = TCP_CA_Disorder;
2566 
2567         if (inet_csk(sk)->icsk_ca_state != state) {
2568                 tcp_set_ca_state(sk, state);
2569                 tp->high_seq = tp->snd_nxt;
2570         }
2571 }
2572 
2573 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2574 {
2575         struct tcp_sock *tp = tcp_sk(sk);
2576 
2577         tcp_verify_left_out(tp);
2578 
2579         if (!tcp_any_retrans_done(sk))
2580                 tp->retrans_stamp = 0;
2581 
2582         if (flag & FLAG_ECE)
2583                 tcp_enter_cwr(sk);
2584 
2585         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2586                 tcp_try_keep_open(sk);
2587         } else {
2588                 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2589         }
2590 }
2591 
2592 static void tcp_mtup_probe_failed(struct sock *sk)
2593 {
2594         struct inet_connection_sock *icsk = inet_csk(sk);
2595 
2596         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2597         icsk->icsk_mtup.probe_size = 0;
2598 }
2599 
2600 static void tcp_mtup_probe_success(struct sock *sk)
2601 {
2602         struct tcp_sock *tp = tcp_sk(sk);
2603         struct inet_connection_sock *icsk = inet_csk(sk);
2604 
2605         /* FIXME: breaks with very large cwnd */
2606         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2607         tp->snd_cwnd = tp->snd_cwnd *
2608                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2609                        icsk->icsk_mtup.probe_size;
2610         tp->snd_cwnd_cnt = 0;
2611         tp->snd_cwnd_stamp = tcp_time_stamp;
2612         tp->snd_ssthresh = tcp_current_ssthresh(sk);
2613 
2614         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2615         icsk->icsk_mtup.probe_size = 0;
2616         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2617 }
2618 
2619 /* Do a simple retransmit without using the backoff mechanisms in
2620  * tcp_timer. This is used for path mtu discovery.
2621  * The socket is already locked here.
2622  */
2623 void tcp_simple_retransmit(struct sock *sk)
2624 {
2625         const struct inet_connection_sock *icsk = inet_csk(sk);
2626         struct tcp_sock *tp = tcp_sk(sk);
2627         struct sk_buff *skb;
2628         unsigned int mss = tcp_current_mss(sk);
2629         u32 prior_lost = tp->lost_out;
2630 
2631         tcp_for_write_queue(skb, sk) {
2632                 if (skb == tcp_send_head(sk))
2633                         break;
2634                 if (tcp_skb_seglen(skb) > mss &&
2635                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2636                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2637                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2638                                 tp->retrans_out -= tcp_skb_pcount(skb);
2639                         }
2640                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2641                 }
2642         }
2643 
2644         tcp_clear_retrans_hints_partial(tp);
2645 
2646         if (prior_lost == tp->lost_out)
2647                 return;
2648 
2649         if (tcp_is_reno(tp))
2650                 tcp_limit_reno_sacked(tp);
2651 
2652         tcp_verify_left_out(tp);
2653 
2654         /* Don't muck with the congestion window here.
2655          * Reason is that we do not increase amount of _data_
2656          * in network, but units changed and effective
2657          * cwnd/ssthresh really reduced now.
2658          */
2659         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2660                 tp->high_seq = tp->snd_nxt;
2661                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2662                 tp->prior_ssthresh = 0;
2663                 tp->undo_marker = 0;
2664                 tcp_set_ca_state(sk, TCP_CA_Loss);
2665         }
2666         tcp_xmit_retransmit_queue(sk);
2667 }
2668 EXPORT_SYMBOL(tcp_simple_retransmit);
2669 
2670 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2671 {
2672         struct tcp_sock *tp = tcp_sk(sk);
2673         int mib_idx;
2674 
2675         if (tcp_is_reno(tp))
2676                 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2677         else
2678                 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2679 
2680         NET_INC_STATS_BH(sock_net(sk), mib_idx);
2681 
2682         tp->prior_ssthresh = 0;
2683         tcp_init_undo(tp);
2684 
2685         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2686                 if (!ece_ack)
2687                         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2688                 tcp_init_cwnd_reduction(sk);
2689         }
2690         tcp_set_ca_state(sk, TCP_CA_Recovery);
2691 }
2692 
2693 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2694  * recovered or spurious. Otherwise retransmits more on partial ACKs.
2695  */
2696 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2697 {
2698         struct tcp_sock *tp = tcp_sk(sk);
2699         bool recovered = !before(tp->snd_una, tp->high_seq);
2700 
2701         if ((flag & FLAG_SND_UNA_ADVANCED) &&
2702             tcp_try_undo_loss(sk, false))
2703                 return;
2704 
2705         if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2706                 /* Step 3.b. A timeout is spurious if not all data are
2707                  * lost, i.e., never-retransmitted data are (s)acked.
2708                  */
2709                 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2710                     tcp_try_undo_loss(sk, true))
2711                         return;
2712 
2713                 if (after(tp->snd_nxt, tp->high_seq)) {
2714                         if (flag & FLAG_DATA_SACKED || is_dupack)
2715                                 tp->frto = 0; /* Step 3.a. loss was real */
2716                 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2717                         tp->high_seq = tp->snd_nxt;
2718                         __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2719                                                   TCP_NAGLE_OFF);
2720                         if (after(tp->snd_nxt, tp->high_seq))
2721                                 return; /* Step 2.b */
2722                         tp->frto = 0;
2723                 }
2724         }
2725 
2726         if (recovered) {
2727                 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2728                 tcp_try_undo_recovery(sk);
2729                 return;
2730         }
2731         if (tcp_is_reno(tp)) {
2732                 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2733                  * delivered. Lower inflight to clock out (re)tranmissions.
2734                  */
2735                 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2736                         tcp_add_reno_sack(sk);
2737                 else if (flag & FLAG_SND_UNA_ADVANCED)
2738                         tcp_reset_reno_sack(tp);
2739         }
2740         tcp_xmit_retransmit_queue(sk);
2741 }
2742 
2743 /* Undo during fast recovery after partial ACK. */
2744 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2745                                  const int prior_unsacked)
2746 {
2747         struct tcp_sock *tp = tcp_sk(sk);
2748 
2749         if (tp->undo_marker && tcp_packet_delayed(tp)) {
2750                 /* Plain luck! Hole if filled with delayed
2751                  * packet, rather than with a retransmit.
2752                  */
2753                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2754 
2755                 /* We are getting evidence that the reordering degree is higher
2756                  * than we realized. If there are no retransmits out then we
2757                  * can undo. Otherwise we clock out new packets but do not
2758                  * mark more packets lost or retransmit more.
2759                  */
2760                 if (tp->retrans_out) {
2761                         tcp_cwnd_reduction(sk, prior_unsacked, 0);
2762                         return true;
2763                 }
2764 
2765                 if (!tcp_any_retrans_done(sk))
2766                         tp->retrans_stamp = 0;
2767 
2768                 DBGUNDO(sk, "partial recovery");
2769                 tcp_undo_cwnd_reduction(sk, true);
2770                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2771                 tcp_try_keep_open(sk);
2772                 return true;
2773         }
2774         return false;
2775 }
2776 
2777 /* Process an event, which can update packets-in-flight not trivially.
2778  * Main goal of this function is to calculate new estimate for left_out,
2779  * taking into account both packets sitting in receiver's buffer and
2780  * packets lost by network.
2781  *
2782  * Besides that it does CWND reduction, when packet loss is detected
2783  * and changes state of machine.
2784  *
2785  * It does _not_ decide what to send, it is made in function
2786  * tcp_xmit_retransmit_queue().
2787  */
2788 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2789                                   const int prior_unsacked,
2790                                   bool is_dupack, int flag)
2791 {
2792         struct inet_connection_sock *icsk = inet_csk(sk);
2793         struct tcp_sock *tp = tcp_sk(sk);
2794         bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2795                                     (tcp_fackets_out(tp) > tp->reordering));
2796         int fast_rexmit = 0;
2797 
2798         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2799                 tp->sacked_out = 0;
2800         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2801                 tp->fackets_out = 0;
2802 
2803         /* Now state machine starts.
2804          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2805         if (flag & FLAG_ECE)
2806                 tp->prior_ssthresh = 0;
2807 
2808         /* B. In all the states check for reneging SACKs. */
2809         if (tcp_check_sack_reneging(sk, flag))
2810                 return;
2811 
2812         /* C. Check consistency of the current state. */
2813         tcp_verify_left_out(tp);
2814 
2815         /* D. Check state exit conditions. State can be terminated
2816          *    when high_seq is ACKed. */
2817         if (icsk->icsk_ca_state == TCP_CA_Open) {
2818                 WARN_ON(tp->retrans_out != 0);
2819                 tp->retrans_stamp = 0;
2820         } else if (!before(tp->snd_una, tp->high_seq)) {
2821                 switch (icsk->icsk_ca_state) {
2822                 case TCP_CA_CWR:
2823                         /* CWR is to be held something *above* high_seq
2824                          * is ACKed for CWR bit to reach receiver. */
2825                         if (tp->snd_una != tp->high_seq) {
2826                                 tcp_end_cwnd_reduction(sk);
2827                                 tcp_set_ca_state(sk, TCP_CA_Open);
2828                         }
2829                         break;
2830 
2831                 case TCP_CA_Recovery:
2832                         if (tcp_is_reno(tp))
2833                                 tcp_reset_reno_sack(tp);
2834                         if (tcp_try_undo_recovery(sk))
2835                                 return;
2836                         tcp_end_cwnd_reduction(sk);
2837                         break;
2838                 }
2839         }
2840 
2841         /* E. Process state. */
2842         switch (icsk->icsk_ca_state) {
2843         case TCP_CA_Recovery:
2844                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2845                         if (tcp_is_reno(tp) && is_dupack)
2846                                 tcp_add_reno_sack(sk);
2847                 } else {
2848                         if (tcp_try_undo_partial(sk, acked, prior_unsacked))
2849                                 return;
2850                         /* Partial ACK arrived. Force fast retransmit. */
2851                         do_lost = tcp_is_reno(tp) ||
2852                                   tcp_fackets_out(tp) > tp->reordering;
2853                 }
2854                 if (tcp_try_undo_dsack(sk)) {
2855                         tcp_try_keep_open(sk);
2856                         return;
2857                 }
2858                 break;
2859         case TCP_CA_Loss:
2860                 tcp_process_loss(sk, flag, is_dupack);
2861                 if (icsk->icsk_ca_state != TCP_CA_Open)
2862                         return;
2863                 /* Fall through to processing in Open state. */
2864         default:
2865                 if (tcp_is_reno(tp)) {
2866                         if (flag & FLAG_SND_UNA_ADVANCED)
2867                                 tcp_reset_reno_sack(tp);
2868                         if (is_dupack)
2869                                 tcp_add_reno_sack(sk);
2870                 }
2871 
2872                 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2873                         tcp_try_undo_dsack(sk);
2874 
2875                 if (!tcp_time_to_recover(sk, flag)) {
2876                         tcp_try_to_open(sk, flag, prior_unsacked);
2877                         return;
2878                 }
2879 
2880                 /* MTU probe failure: don't reduce cwnd */
2881                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2882                     icsk->icsk_mtup.probe_size &&
2883                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2884                         tcp_mtup_probe_failed(sk);
2885                         /* Restores the reduction we did in tcp_mtup_probe() */
2886                         tp->snd_cwnd++;
2887                         tcp_simple_retransmit(sk);
2888                         return;
2889                 }
2890 
2891                 /* Otherwise enter Recovery state */
2892                 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2893                 fast_rexmit = 1;
2894         }
2895 
2896         if (do_lost)
2897                 tcp_update_scoreboard(sk, fast_rexmit);
2898         tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
2899         tcp_xmit_retransmit_queue(sk);
2900 }
2901 
2902 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2903                                       long seq_rtt_us, long sack_rtt_us)
2904 {
2905         const struct tcp_sock *tp = tcp_sk(sk);
2906 
2907         /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2908          * broken middle-boxes or peers may corrupt TS-ECR fields. But
2909          * Karn's algorithm forbids taking RTT if some retransmitted data
2910          * is acked (RFC6298).
2911          */
2912         if (flag & FLAG_RETRANS_DATA_ACKED)
2913                 seq_rtt_us = -1L;
2914 
2915         if (seq_rtt_us < 0)
2916                 seq_rtt_us = sack_rtt_us;
2917 
2918         /* RTTM Rule: A TSecr value received in a segment is used to
2919          * update the averaged RTT measurement only if the segment
2920          * acknowledges some new data, i.e., only if it advances the
2921          * left edge of the send window.
2922          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2923          */
2924         if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2925             flag & FLAG_ACKED)
2926                 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2927 
2928         if (seq_rtt_us < 0)
2929                 return false;
2930 
2931         tcp_rtt_estimator(sk, seq_rtt_us);
2932         tcp_set_rto(sk);
2933 
2934         /* RFC6298: only reset backoff on valid RTT measurement. */
2935         inet_csk(sk)->icsk_backoff = 0;
2936         return true;
2937 }
2938 
2939 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2940 static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
2941 {
2942         struct tcp_sock *tp = tcp_sk(sk);
2943         long seq_rtt_us = -1L;
2944 
2945         if (synack_stamp && !tp->total_retrans)
2946                 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - synack_stamp);
2947 
2948         /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2949          * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2950          */
2951         if (!tp->srtt_us)
2952                 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt_us, -1L);
2953 }
2954 
2955 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2956 {
2957         const struct inet_connection_sock *icsk = inet_csk(sk);
2958 
2959         icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2960         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2961 }
2962 
2963 /* Restart timer after forward progress on connection.
2964  * RFC2988 recommends to restart timer to now+rto.
2965  */
2966 void tcp_rearm_rto(struct sock *sk)
2967 {
2968         const struct inet_connection_sock *icsk = inet_csk(sk);
2969         struct tcp_sock *tp = tcp_sk(sk);
2970 
2971         /* If the retrans timer is currently being used by Fast Open
2972          * for SYN-ACK retrans purpose, stay put.
2973          */
2974         if (tp->fastopen_rsk)
2975                 return;
2976 
2977         if (!tp->packets_out) {
2978                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2979         } else {
2980                 u32 rto = inet_csk(sk)->icsk_rto;
2981                 /* Offset the time elapsed after installing regular RTO */
2982                 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2983                     icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2984                         struct sk_buff *skb = tcp_write_queue_head(sk);
2985                         const u32 rto_time_stamp =
2986                                 tcp_skb_timestamp(skb) + rto;
2987                         s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2988                         /* delta may not be positive if the socket is locked
2989                          * when the retrans timer fires and is rescheduled.
2990                          */
2991                         if (delta > 0)
2992                                 rto = delta;
2993                 }
2994                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2995                                           TCP_RTO_MAX);
2996         }
2997 }
2998 
2999 /* This function is called when the delayed ER timer fires. TCP enters
3000  * fast recovery and performs fast-retransmit.
3001  */
3002 void tcp_resume_early_retransmit(struct sock *sk)
3003 {
3004         struct tcp_sock *tp = tcp_sk(sk);
3005 
3006         tcp_rearm_rto(sk);
3007 
3008         /* Stop if ER is disabled after the delayed ER timer is scheduled */
3009         if (!tp->do_early_retrans)
3010                 return;
3011 
3012         tcp_enter_recovery(sk, false);
3013         tcp_update_scoreboard(sk, 1);
3014         tcp_xmit_retransmit_queue(sk);
3015 }
3016 
3017 /* If we get here, the whole TSO packet has not been acked. */
3018 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3019 {
3020         struct tcp_sock *tp = tcp_sk(sk);
3021         u32 packets_acked;
3022 
3023         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3024 
3025         packets_acked = tcp_skb_pcount(skb);
3026         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3027                 return 0;
3028         packets_acked -= tcp_skb_pcount(skb);
3029 
3030         if (packets_acked) {
3031                 BUG_ON(tcp_skb_pcount(skb) == 0);
3032                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3033         }
3034 
3035         return packets_acked;
3036 }
3037 
3038 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3039                            u32 prior_snd_una)
3040 {
3041         const struct skb_shared_info *shinfo;
3042 
3043         /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3044         if (likely(!(sk->sk_tsflags & SOF_TIMESTAMPING_TX_ACK)))
3045                 return;
3046 
3047         shinfo = skb_shinfo(skb);
3048         if ((shinfo->tx_flags & SKBTX_ACK_TSTAMP) &&
3049             between(shinfo->tskey, prior_snd_una, tcp_sk(sk)->snd_una - 1))
3050                 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3051 }
3052 
3053 /* Remove acknowledged frames from the retransmission queue. If our packet
3054  * is before the ack sequence we can discard it as it's confirmed to have
3055  * arrived at the other end.
3056  */
3057 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3058                                u32 prior_snd_una, long sack_rtt_us)
3059 {
3060         const struct inet_connection_sock *icsk = inet_csk(sk);
3061         struct skb_mstamp first_ackt, last_ackt, now;
3062         struct tcp_sock *tp = tcp_sk(sk);
3063         u32 prior_sacked = tp->sacked_out;
3064         u32 reord = tp->packets_out;
3065         bool fully_acked = true;
3066         long ca_seq_rtt_us = -1L;
3067         long seq_rtt_us = -1L;
3068         struct sk_buff *skb;
3069         u32 pkts_acked = 0;
3070         bool rtt_update;
3071         int flag = 0;
3072 
3073         first_ackt.v64 = 0;
3074 
3075         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3076                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3077                 u8 sacked = scb->sacked;
3078                 u32 acked_pcount;
3079 
3080                 tcp_ack_tstamp(sk, skb, prior_snd_una);
3081 
3082                 /* Determine how many packets and what bytes were acked, tso and else */
3083                 if (after(scb->end_seq, tp->snd_una)) {
3084                         if (tcp_skb_pcount(skb) == 1 ||
3085                             !after(tp->snd_una, scb->seq))
3086                                 break;
3087 
3088                         acked_pcount = tcp_tso_acked(sk, skb);
3089                         if (!acked_pcount)
3090                                 break;
3091 
3092                         fully_acked = false;
3093                 } else {
3094                         /* Speedup tcp_unlink_write_queue() and next loop */
3095                         prefetchw(skb->next);
3096                         acked_pcount = tcp_skb_pcount(skb);
3097                 }
3098 
3099                 if (unlikely(sacked & TCPCB_RETRANS)) {
3100                         if (sacked & TCPCB_SACKED_RETRANS)
3101                                 tp->retrans_out -= acked_pcount;
3102                         flag |= FLAG_RETRANS_DATA_ACKED;
3103                 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3104                         last_ackt = skb->skb_mstamp;
3105                         WARN_ON_ONCE(last_ackt.v64 == 0);
3106                         if (!first_ackt.v64)
3107                                 first_ackt = last_ackt;
3108 
3109                         reord = min(pkts_acked, reord);
3110                         if (!after(scb->end_seq, tp->high_seq))
3111                                 flag |= FLAG_ORIG_SACK_ACKED;
3112                 }
3113 
3114                 if (sacked & TCPCB_SACKED_ACKED)
3115                         tp->sacked_out -= acked_pcount;
3116                 if (sacked & TCPCB_LOST)
3117                         tp->lost_out -= acked_pcount;
3118 
3119                 tp->packets_out -= acked_pcount;
3120                 pkts_acked += acked_pcount;
3121 
3122                 /* Initial outgoing SYN's get put onto the write_queue
3123                  * just like anything else we transmit.  It is not
3124                  * true data, and if we misinform our callers that
3125                  * this ACK acks real data, we will erroneously exit
3126                  * connection startup slow start one packet too
3127                  * quickly.  This is severely frowned upon behavior.
3128                  */
3129                 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3130                         flag |= FLAG_DATA_ACKED;
3131                 } else {
3132                         flag |= FLAG_SYN_ACKED;
3133                         tp->retrans_stamp = 0;
3134                 }
3135 
3136                 if (!fully_acked)
3137                         break;
3138 
3139                 tcp_unlink_write_queue(skb, sk);
3140                 sk_wmem_free_skb(sk, skb);
3141                 if (unlikely(skb == tp->retransmit_skb_hint))
3142                         tp->retransmit_skb_hint = NULL;
3143                 if (unlikely(skb == tp->lost_skb_hint))
3144                         tp->lost_skb_hint = NULL;
3145         }
3146 
3147         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3148                 tp->snd_up = tp->snd_una;
3149 
3150         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3151                 flag |= FLAG_SACK_RENEGING;
3152 
3153         skb_mstamp_get(&now);
3154         if (likely(first_ackt.v64)) {
3155                 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3156                 ca_seq_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3157         }
3158 
3159         rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us);
3160 
3161         if (flag & FLAG_ACKED) {
3162                 const struct tcp_congestion_ops *ca_ops
3163                         = inet_csk(sk)->icsk_ca_ops;
3164 
3165                 tcp_rearm_rto(sk);
3166                 if (unlikely(icsk->icsk_mtup.probe_size &&
3167                              !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3168                         tcp_mtup_probe_success(sk);
3169                 }
3170 
3171                 if (tcp_is_reno(tp)) {
3172                         tcp_remove_reno_sacks(sk, pkts_acked);
3173                 } else {
3174                         int delta;
3175 
3176                         /* Non-retransmitted hole got filled? That's reordering */
3177                         if (reord < prior_fackets)
3178                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3179 
3180                         delta = tcp_is_fack(tp) ? pkts_acked :
3181                                                   prior_sacked - tp->sacked_out;
3182                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3183                 }
3184 
3185                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3186 
3187                 if (ca_ops->pkts_acked) {
3188                         long rtt_us = min_t(ulong, ca_seq_rtt_us, sack_rtt_us);
3189                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3190                 }
3191 
3192         } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3193                    sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3194                 /* Do not re-arm RTO if the sack RTT is measured from data sent
3195                  * after when the head was last (re)transmitted. Otherwise the
3196                  * timeout may continue to extend in loss recovery.
3197                  */
3198                 tcp_rearm_rto(sk);
3199         }
3200 
3201 #if FASTRETRANS_DEBUG > 0
3202         WARN_ON((int)tp->sacked_out < 0);
3203         WARN_ON((int)tp->lost_out < 0);
3204         WARN_ON((int)tp->retrans_out < 0);
3205         if (!tp->packets_out && tcp_is_sack(tp)) {
3206                 icsk = inet_csk(sk);
3207                 if (tp->lost_out) {
3208                         pr_debug("Leak l=%u %d\n",
3209                                  tp->lost_out, icsk->icsk_ca_state);
3210                         tp->lost_out = 0;
3211                 }
3212                 if (tp->sacked_out) {
3213                         pr_debug("Leak s=%u %d\n",
3214                                  tp->sacked_out, icsk->icsk_ca_state);
3215                         tp->sacked_out = 0;
3216                 }
3217                 if (tp->retrans_out) {
3218                         pr_debug("Leak r=%u %d\n",
3219                                  tp->retrans_out, icsk->icsk_ca_state);
3220                         tp->retrans_out = 0;
3221                 }
3222         }
3223 #endif
3224         return flag;
3225 }
3226 
3227 static void tcp_ack_probe(struct sock *sk)
3228 {
3229         const struct tcp_sock *tp = tcp_sk(sk);
3230         struct inet_connection_sock *icsk = inet_csk(sk);
3231 
3232         /* Was it a usable window open? */
3233 
3234         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3235                 icsk->icsk_backoff = 0;
3236                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3237                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3238                  * This function is not for random using!
3239                  */
3240         } else {
3241                 unsigned long when = inet_csk_rto_backoff(icsk, TCP_RTO_MAX);
3242 
3243                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3244                                           when, TCP_RTO_MAX);
3245         }
3246 }
3247 
3248 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3249 {
3250         return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3251                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3252 }
3253 
3254 /* Decide wheather to run the increase function of congestion control. */
3255 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3256 {
3257         if (tcp_in_cwnd_reduction(sk))
3258                 return false;
3259 
3260         /* If reordering is high then always grow cwnd whenever data is
3261          * delivered regardless of its ordering. Otherwise stay conservative
3262          * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3263          * new SACK or ECE mark may first advance cwnd here and later reduce
3264          * cwnd in tcp_fastretrans_alert() based on more states.
3265          */
3266         if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3267                 return flag & FLAG_FORWARD_PROGRESS;
3268 
3269         return flag & FLAG_DATA_ACKED;
3270 }
3271 
3272 /* Check that window update is acceptable.
3273  * The function assumes that snd_una<=ack<=snd_next.
3274  */
3275 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3276                                         const u32 ack, const u32 ack_seq,
3277                                         const u32 nwin)
3278 {
3279         return  after(ack, tp->snd_una) ||
3280                 after(ack_seq, tp->snd_wl1) ||
3281                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3282 }
3283 
3284 /* If we update tp->snd_una, also update tp->bytes_acked */
3285 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3286 {
3287         u32 delta = ack - tp->snd_una;
3288 
3289         u64_stats_update_begin(&tp->syncp);
3290         tp->bytes_acked += delta;
3291         u64_stats_update_end(&tp->syncp);
3292         tp->snd_una = ack;
3293 }
3294 
3295 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3296 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3297 {
3298         u32 delta = seq - tp->rcv_nxt;
3299 
3300         u64_stats_update_begin(&tp->syncp);
3301         tp->bytes_received += delta;
3302         u64_stats_update_end(&tp->syncp);
3303         tp->rcv_nxt = seq;
3304 }
3305 
3306 /* Update our send window.
3307  *
3308  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3309  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3310  */
3311 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3312                                  u32 ack_seq)
3313 {
3314         struct tcp_sock *tp = tcp_sk(sk);
3315         int flag = 0;
3316         u32 nwin = ntohs(tcp_hdr(skb)->window);
3317 
3318         if (likely(!tcp_hdr(skb)->syn))
3319                 nwin <<= tp->rx_opt.snd_wscale;
3320 
3321         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3322                 flag |= FLAG_WIN_UPDATE;
3323                 tcp_update_wl(tp, ack_seq);
3324 
3325                 if (tp->snd_wnd != nwin) {
3326                         tp->snd_wnd = nwin;
3327 
3328                         /* Note, it is the only place, where
3329                          * fast path is recovered for sending TCP.
3330                          */
3331                         tp->pred_flags = 0;
3332                         tcp_fast_path_check(sk);
3333 
3334                         if (nwin > tp->max_window) {
3335                                 tp->max_window = nwin;
3336                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3337                         }
3338                 }
3339         }
3340 
3341         tcp_snd_una_update(tp, ack);
3342 
3343         return flag;
3344 }
3345 
3346 /* Return true if we're currently rate-limiting out-of-window ACKs and
3347  * thus shouldn't send a dupack right now. We rate-limit dupacks in
3348  * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3349  * attacks that send repeated SYNs or ACKs for the same connection. To
3350  * do this, we do not send a duplicate SYNACK or ACK if the remote
3351  * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3352  */
3353 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3354                           int mib_idx, u32 *last_oow_ack_time)
3355 {
3356         /* Data packets without SYNs are not likely part of an ACK loop. */
3357         if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3358             !tcp_hdr(skb)->syn)
3359                 goto not_rate_limited;
3360 
3361         if (*last_oow_ack_time) {
3362                 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3363 
3364                 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3365                         NET_INC_STATS_BH(net, mib_idx);
3366                         return true;    /* rate-limited: don't send yet! */
3367                 }
3368         }
3369 
3370         *last_oow_ack_time = tcp_time_stamp;
3371 
3372 not_rate_limited:
3373         return false;   /* not rate-limited: go ahead, send dupack now! */
3374 }
3375 
3376 /* RFC 5961 7 [ACK Throttling] */
3377 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3378 {
3379         /* unprotected vars, we dont care of overwrites */
3380         static u32 challenge_timestamp;
3381         static unsigned int challenge_count;
3382         struct tcp_sock *tp = tcp_sk(sk);
3383         u32 now;
3384 
3385         /* First check our per-socket dupack rate limit. */
3386         if (tcp_oow_rate_limited(sock_net(sk), skb,
3387                                  LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3388                                  &tp->last_oow_ack_time))
3389                 return;
3390 
3391         /* Then check the check host-wide RFC 5961 rate limit. */
3392         now = jiffies / HZ;
3393         if (now != challenge_timestamp) {
3394                 challenge_timestamp = now;
3395                 challenge_count = 0;
3396         }
3397         if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3398                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3399                 tcp_send_ack(sk);
3400         }
3401 }
3402 
3403 static void tcp_store_ts_recent(struct tcp_sock *tp)
3404 {
3405         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3406         tp->rx_opt.ts_recent_stamp = get_seconds();
3407 }
3408 
3409 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3410 {
3411         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3412                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3413                  * extra check below makes sure this can only happen
3414                  * for pure ACK frames.  -DaveM
3415                  *
3416                  * Not only, also it occurs for expired timestamps.
3417                  */
3418 
3419                 if (tcp_paws_check(&tp->rx_opt, 0))
3420                         tcp_store_ts_recent(tp);
3421         }
3422 }
3423 
3424 /* This routine deals with acks during a TLP episode.
3425  * We mark the end of a TLP episode on receiving TLP dupack or when
3426  * ack is after tlp_high_seq.
3427  * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3428  */
3429 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3430 {
3431         struct tcp_sock *tp = tcp_sk(sk);
3432 
3433         if (before(ack, tp->tlp_high_seq))
3434                 return;
3435 
3436         if (flag & FLAG_DSACKING_ACK) {
3437                 /* This DSACK means original and TLP probe arrived; no loss */
3438                 tp->tlp_high_seq = 0;
3439         } else if (after(ack, tp->tlp_high_seq)) {
3440                 /* ACK advances: there was a loss, so reduce cwnd. Reset
3441                  * tlp_high_seq in tcp_init_cwnd_reduction()
3442                  */
3443                 tcp_init_cwnd_reduction(sk);
3444                 tcp_set_ca_state(sk, TCP_CA_CWR);
3445                 tcp_end_cwnd_reduction(sk);
3446                 tcp_try_keep_open(sk);
3447                 NET_INC_STATS_BH(sock_net(sk),
3448                                  LINUX_MIB_TCPLOSSPROBERECOVERY);
3449         } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3450                              FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3451                 /* Pure dupack: original and TLP probe arrived; no loss */
3452                 tp->tlp_high_seq = 0;
3453         }
3454 }
3455 
3456 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3457 {
3458         const struct inet_connection_sock *icsk = inet_csk(sk);
3459 
3460         if (icsk->icsk_ca_ops->in_ack_event)
3461                 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3462 }
3463 
3464 /* This routine deals with incoming acks, but not outgoing ones. */
3465 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3466 {
3467         struct inet_connection_sock *icsk = inet_csk(sk);
3468         struct tcp_sock *tp = tcp_sk(sk);
3469         u32 prior_snd_una = tp->snd_una;
3470         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3471         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3472         bool is_dupack = false;
3473         u32 prior_fackets;
3474         int prior_packets = tp->packets_out;
3475         const int prior_unsacked = tp->packets_out - tp->sacked_out;
3476         int acked = 0; /* Number of packets newly acked */
3477         long sack_rtt_us = -1L;
3478 
3479         /* We very likely will need to access write queue head. */
3480         prefetchw(sk->sk_write_queue.next);
3481 
3482         /* If the ack is older than previous acks
3483          * then we can probably ignore it.
3484          */
3485         if (before(ack, prior_snd_una)) {
3486                 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3487                 if (before(ack, prior_snd_una - tp->max_window)) {
3488                         tcp_send_challenge_ack(sk, skb);
3489                         return -1;
3490                 }
3491                 goto old_ack;
3492         }
3493 
3494         /* If the ack includes data we haven't sent yet, discard
3495          * this segment (RFC793 Section 3.9).
3496          */
3497         if (after(ack, tp->snd_nxt))
3498                 goto invalid_ack;
3499 
3500         if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3501             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3502                 tcp_rearm_rto(sk);
3503 
3504         if (after(ack, prior_snd_una)) {
3505                 flag |= FLAG_SND_UNA_ADVANCED;
3506                 icsk->icsk_retransmits = 0;
3507         }
3508 
3509         prior_fackets = tp->fackets_out;
3510 
3511         /* ts_recent update must be made after we are sure that the packet
3512          * is in window.
3513          */
3514         if (flag & FLAG_UPDATE_TS_RECENT)
3515                 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3516 
3517         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3518                 /* Window is constant, pure forward advance.
3519                  * No more checks are required.
3520                  * Note, we use the fact that SND.UNA>=SND.WL2.
3521                  */
3522                 tcp_update_wl(tp, ack_seq);
3523                 tcp_snd_una_update(tp, ack);
3524                 flag |= FLAG_WIN_UPDATE;
3525 
3526                 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3527 
3528                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3529         } else {
3530                 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3531 
3532                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3533                         flag |= FLAG_DATA;
3534                 else
3535                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3536 
3537                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3538 
3539                 if (TCP_SKB_CB(skb)->sacked)
3540                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3541                                                         &sack_rtt_us);
3542 
3543                 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3544                         flag |= FLAG_ECE;
3545                         ack_ev_flags |= CA_ACK_ECE;
3546                 }
3547 
3548                 if (flag & FLAG_WIN_UPDATE)
3549                         ack_ev_flags |= CA_ACK_WIN_UPDATE;
3550 
3551                 tcp_in_ack_event(sk, ack_ev_flags);
3552         }
3553 
3554         /* We passed data and got it acked, remove any soft error
3555          * log. Something worked...
3556          */
3557         sk->sk_err_soft = 0;
3558         icsk->icsk_probes_out = 0;
3559         tp->rcv_tstamp = tcp_time_stamp;
3560         if (!prior_packets)
3561                 goto no_queue;
3562 
3563         /* See if we can take anything off of the retransmit queue. */
3564         acked = tp->packets_out;
3565         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
3566                                     sack_rtt_us);
3567         acked -= tp->packets_out;
3568 
3569         /* Advance cwnd if state allows */
3570         if (tcp_may_raise_cwnd(sk, flag))
3571                 tcp_cong_avoid(sk, ack, acked);
3572 
3573         if (tcp_ack_is_dubious(sk, flag)) {
3574                 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3575                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3576                                       is_dupack, flag);
3577         }
3578         if (tp->tlp_high_seq)
3579                 tcp_process_tlp_ack(sk, ack, flag);
3580 
3581         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3582                 struct dst_entry *dst = __sk_dst_get(sk);
3583                 if (dst)
3584                         dst_confirm(dst);
3585         }
3586 
3587         if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3588                 tcp_schedule_loss_probe(sk);
3589         tcp_update_pacing_rate(sk);
3590         return 1;
3591 
3592 no_queue:
3593         /* If data was DSACKed, see if we can undo a cwnd reduction. */
3594         if (flag & FLAG_DSACKING_ACK)
3595                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3596                                       is_dupack, flag);
3597         /* If this ack opens up a zero window, clear backoff.  It was
3598          * being used to time the probes, and is probably far higher than
3599          * it needs to be for normal retransmission.
3600          */
3601         if (tcp_send_head(sk))
3602                 tcp_ack_probe(sk);
3603 
3604         if (tp->tlp_high_seq)
3605                 tcp_process_tlp_ack(sk, ack, flag);
3606         return 1;
3607 
3608 invalid_ack:
3609         SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3610         return -1;
3611 
3612 old_ack:
3613         /* If data was SACKed, tag it and see if we should send more data.
3614          * If data was DSACKed, see if we can undo a cwnd reduction.
3615          */
3616         if (TCP_SKB_CB(skb)->sacked) {
3617                 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3618                                                 &sack_rtt_us);
3619                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3620                                       is_dupack, flag);
3621         }
3622 
3623         SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3624         return 0;
3625 }
3626 
3627 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3628                                       bool syn, struct tcp_fastopen_cookie *foc,
3629                                       bool exp_opt)
3630 {
3631         /* Valid only in SYN or SYN-ACK with an even length.  */
3632         if (!foc || !syn || len < 0 || (len & 1))
3633                 return;
3634 
3635         if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3636             len <= TCP_FASTOPEN_COOKIE_MAX)
3637                 memcpy(foc->val, cookie, len);
3638         else if (len != 0)
3639                 len = -1;
3640         foc->len = len;
3641         foc->exp = exp_opt;
3642 }
3643 
3644 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3645  * But, this can also be called on packets in the established flow when
3646  * the fast version below fails.
3647  */
3648 void tcp_parse_options(const struct sk_buff *skb,
3649                        struct tcp_options_received *opt_rx, int estab,
3650                        struct tcp_fastopen_cookie *foc)
3651 {
3652         const unsigned char *ptr;
3653         const struct tcphdr *th = tcp_hdr(skb);
3654         int length = (th->doff * 4) - sizeof(struct tcphdr);
3655 
3656         ptr = (const unsigned char *)(th + 1);
3657         opt_rx->saw_tstamp = 0;
3658 
3659         while (length > 0) {
3660                 int opcode = *ptr++;
3661                 int opsize;
3662 
3663                 switch (opcode) {
3664                 case TCPOPT_EOL:
3665                         return;
3666                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3667                         length--;
3668                         continue;
3669                 default:
3670                         opsize = *ptr++;
3671                         if (opsize < 2) /* "silly options" */
3672                                 return;
3673                         if (opsize > length)
3674                                 return; /* don't parse partial options */
3675                         switch (opcode) {
3676                         case TCPOPT_MSS:
3677                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3678                                         u16 in_mss = get_unaligned_be16(ptr);
3679                                         if (in_mss) {
3680                                                 if (opt_rx->user_mss &&
3681                                                     opt_rx->user_mss < in_mss)
3682                                                         in_mss = opt_rx->user_mss;
3683                                                 opt_rx->mss_clamp = in_mss;
3684                                         }
3685                                 }
3686                                 break;
3687                         case TCPOPT_WINDOW:
3688                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3689                                     !estab && sysctl_tcp_window_scaling) {
3690                                         __u8 snd_wscale = *(__u8 *)ptr;
3691                                         opt_rx->wscale_ok = 1;
3692                                         if (snd_wscale > 14) {
3693                                                 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3694                                                                      __func__,
3695                                                                      snd_wscale);
3696                                                 snd_wscale = 14;
3697                                         }
3698                                         opt_rx->snd_wscale = snd_wscale;
3699                                 }
3700                                 break;
3701                         case TCPOPT_TIMESTAMP:
3702                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3703                                     ((estab && opt_rx->tstamp_ok) ||
3704                                      (!estab && sysctl_tcp_timestamps))) {
3705                                         opt_rx->saw_tstamp = 1;
3706                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3707                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3708                                 }
3709                                 break;
3710                         case TCPOPT_SACK_PERM:
3711                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3712                                     !estab && sysctl_tcp_sack) {
3713                                         opt_rx->sack_ok = TCP_SACK_SEEN;
3714                                         tcp_sack_reset(opt_rx);
3715                                 }
3716                                 break;
3717 
3718                         case TCPOPT_SACK:
3719                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3720                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3721                                    opt_rx->sack_ok) {
3722                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3723                                 }
3724                                 break;
3725 #ifdef CONFIG_TCP_MD5SIG
3726                         case TCPOPT_MD5SIG:
3727                                 /*
3728                                  * The MD5 Hash has already been
3729                                  * checked (see tcp_v{4,6}_do_rcv()).
3730                                  */
3731                                 break;
3732 #endif
3733                         case TCPOPT_FASTOPEN:
3734                                 tcp_parse_fastopen_option(
3735                                         opsize - TCPOLEN_FASTOPEN_BASE,
3736                                         ptr, th->syn, foc, false);
3737                                 break;
3738 
3739                         case TCPOPT_EXP:
3740                                 /* Fast Open option shares code 254 using a
3741                                  * 16 bits magic number.
3742                                  */
3743                                 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3744                                     get_unaligned_be16(ptr) ==
3745                                     TCPOPT_FASTOPEN_MAGIC)
3746                                         tcp_parse_fastopen_option(opsize -
3747                                                 TCPOLEN_EXP_FASTOPEN_BASE,
3748                                                 ptr + 2, th->syn, foc, true);
3749                                 break;
3750 
3751                         }
3752                         ptr += opsize-2;
3753                         length -= opsize;
3754                 }
3755         }
3756 }
3757 EXPORT_SYMBOL(tcp_parse_options);
3758 
3759 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3760 {
3761         const __be32 *ptr = (const __be32 *)(th + 1);
3762 
3763         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3764                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3765                 tp->rx_opt.saw_tstamp = 1;
3766                 ++ptr;
3767                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3768                 ++ptr;
3769                 if (*ptr)
3770                         tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3771                 else
3772                         tp->rx_opt.rcv_tsecr = 0;
3773                 return true;
3774         }
3775         return false;
3776 }
3777 
3778 /* Fast parse options. This hopes to only see timestamps.
3779  * If it is wrong it falls back on tcp_parse_options().
3780  */
3781 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3782                                    const struct tcphdr *th, struct tcp_sock *tp)
3783 {
3784         /* In the spirit of fast parsing, compare doff directly to constant
3785          * values.  Because equality is used, short doff can be ignored here.
3786          */
3787         if (th->doff == (sizeof(*th) / 4)) {
3788                 tp->rx_opt.saw_tstamp = 0;
3789                 return false;
3790         } else if (tp->rx_opt.tstamp_ok &&
3791                    th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3792                 if (tcp_parse_aligned_timestamp(tp, th))
3793                         return true;
3794         }
3795 
3796         tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3797         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3798                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3799 
3800         return true;
3801 }
3802 
3803 #ifdef CONFIG_TCP_MD5SIG
3804 /*
3805  * Parse MD5 Signature option
3806  */
3807 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3808 {
3809         int length = (th->doff << 2) - sizeof(*th);
3810         const u8 *ptr = (const u8 *)(th + 1);
3811 
3812         /* If the TCP option is too short, we can short cut */
3813         if (length < TCPOLEN_MD5SIG)
3814                 return NULL;
3815 
3816         while (length > 0) {
3817                 int opcode = *ptr++;
3818                 int opsize;
3819 
3820                 switch (opcode) {
3821                 case TCPOPT_EOL:
3822                         return NULL;
3823                 case TCPOPT_NOP:
3824                         length--;
3825                         continue;
3826                 default:
3827                         opsize = *ptr++;
3828                         if (opsize < 2 || opsize > length)
3829                                 return NULL;
3830                         if (opcode == TCPOPT_MD5SIG)
3831                                 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3832                 }
3833                 ptr += opsize - 2;
3834                 length -= opsize;
3835         }
3836         return NULL;
3837 }
3838 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3839 #endif
3840 
3841 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3842  *
3843  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3844  * it can pass through stack. So, the following predicate verifies that
3845  * this segment is not used for anything but congestion avoidance or
3846  * fast retransmit. Moreover, we even are able to eliminate most of such
3847  * second order effects, if we apply some small "replay" window (~RTO)
3848  * to timestamp space.
3849  *
3850  * All these measures still do not guarantee that we reject wrapped ACKs
3851  * on networks with high bandwidth, when sequence space is recycled fastly,
3852  * but it guarantees that such events will be very rare and do not affect
3853  * connection seriously. This doesn't look nice, but alas, PAWS is really
3854  * buggy extension.
3855  *
3856  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3857  * states that events when retransmit arrives after original data are rare.
3858  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3859  * the biggest problem on large power networks even with minor reordering.
3860  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3861  * up to bandwidth of 18Gigabit/sec. 8) ]
3862  */
3863 
3864 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3865 {
3866         const struct tcp_sock *tp = tcp_sk(sk);
3867         const struct tcphdr *th = tcp_hdr(skb);
3868         u32 seq = TCP_SKB_CB(skb)->seq;
3869         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3870 
3871         return (/* 1. Pure ACK with correct sequence number. */
3872                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3873 
3874                 /* 2. ... and duplicate ACK. */
3875                 ack == tp->snd_una &&
3876 
3877                 /* 3. ... and does not update window. */
3878                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3879 
3880                 /* 4. ... and sits in replay window. */
3881                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3882 }
3883 
3884 static inline bool tcp_paws_discard(const struct sock *sk,
3885                                    const struct sk_buff *skb)
3886 {
3887         const struct tcp_sock *tp = tcp_sk(sk);
3888 
3889         return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3890                !tcp_disordered_ack(sk, skb);
3891 }
3892 
3893 /* Check segment sequence number for validity.
3894  *
3895  * Segment controls are considered valid, if the segment
3896  * fits to the window after truncation to the window. Acceptability
3897  * of data (and SYN, FIN, of course) is checked separately.
3898  * See tcp_data_queue(), for example.
3899  *
3900  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3901  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3902  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3903  * (borrowed from freebsd)
3904  */
3905 
3906 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3907 {
3908         return  !before(end_seq, tp->rcv_wup) &&
3909                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3910 }
3911 
3912 /* When we get a reset we do this. */
3913 void tcp_reset(struct sock *sk)
3914 {
3915         /* We want the right error as BSD sees it (and indeed as we do). */
3916         switch (sk->sk_state) {
3917         case TCP_SYN_SENT:
3918                 sk->sk_err = ECONNREFUSED;
3919                 break;
3920         case TCP_CLOSE_WAIT:
3921                 sk->sk_err = EPIPE;
3922                 break;
3923         case TCP_CLOSE:
3924                 return;
3925         default:
3926                 sk->sk_err = ECONNRESET;
3927         }
3928         /* This barrier is coupled with smp_rmb() in tcp_poll() */
3929         smp_wmb();
3930 
3931         if (!sock_flag(sk, SOCK_DEAD))
3932                 sk->sk_error_report(sk);
3933 
3934         tcp_done(sk);
3935 }
3936 
3937 /*
3938  *      Process the FIN bit. This now behaves as it is supposed to work
3939  *      and the FIN takes effect when it is validly part of sequence
3940  *      space. Not before when we get holes.
3941  *
3942  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3943  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3944  *      TIME-WAIT)
3945  *
3946  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3947  *      close and we go into CLOSING (and later onto TIME-WAIT)
3948  *
3949  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3950  */
3951 static void tcp_fin(struct sock *sk)
3952 {
3953         struct tcp_sock *tp = tcp_sk(sk);
3954         const struct dst_entry *dst;
3955 
3956         inet_csk_schedule_ack(sk);
3957 
3958         sk->sk_shutdown |= RCV_SHUTDOWN;
3959         sock_set_flag(sk, SOCK_DONE);
3960 
3961         switch (sk->sk_state) {
3962         case TCP_SYN_RECV:
3963         case TCP_ESTABLISHED:
3964                 /* Move to CLOSE_WAIT */
3965                 tcp_set_state(sk, TCP_CLOSE_WAIT);
3966                 dst = __sk_dst_get(sk);
3967                 if (!dst || !dst_metric(dst, RTAX_QUICKACK))
3968                         inet_csk(sk)->icsk_ack.pingpong = 1;
3969                 break;
3970 
3971         case TCP_CLOSE_WAIT:
3972         case TCP_CLOSING:
3973                 /* Received a retransmission of the FIN, do
3974                  * nothing.
3975                  */
3976                 break;
3977         case TCP_LAST_ACK:
3978                 /* RFC793: Remain in the LAST-ACK state. */
3979                 break;
3980 
3981         case TCP_FIN_WAIT1:
3982                 /* This case occurs when a simultaneous close
3983                  * happens, we must ack the received FIN and
3984                  * enter the CLOSING state.
3985                  */
3986                 tcp_send_ack(sk);
3987                 tcp_set_state(sk, TCP_CLOSING);
3988                 break;
3989         case TCP_FIN_WAIT2:
3990                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3991                 tcp_send_ack(sk);
3992                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3993                 break;
3994         default:
3995                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3996                  * cases we should never reach this piece of code.
3997                  */
3998                 pr_err("%s: Impossible, sk->sk_state=%d\n",
3999                        __func__, sk->sk_state);
4000                 break;
4001         }
4002 
4003         /* It _is_ possible, that we have something out-of-order _after_ FIN.
4004          * Probably, we should reset in this case. For now drop them.
4005          */
4006         __skb_queue_purge(&tp->out_of_order_queue);
4007         if (tcp_is_sack(tp))
4008                 tcp_sack_reset(&tp->rx_opt);
4009         sk_mem_reclaim(sk);
4010 
4011         if (!sock_flag(sk, SOCK_DEAD)) {
4012                 sk->sk_state_change(sk);
4013 
4014                 /* Do not send POLL_HUP for half duplex close. */
4015                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4016                     sk->sk_state == TCP_CLOSE)
4017                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4018                 else
4019                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4020         }
4021 }
4022 
4023 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4024                                   u32 end_seq)
4025 {
4026         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4027                 if (before(seq, sp->start_seq))
4028                         sp->start_seq = seq;
4029                 if (after(end_seq, sp->end_seq))
4030                         sp->end_seq = end_seq;
4031                 return true;
4032         }
4033         return false;
4034 }
4035 
4036 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4037 {
4038         struct tcp_sock *tp = tcp_sk(sk);
4039 
4040         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4041                 int mib_idx;
4042 
4043                 if (before(seq, tp->rcv_nxt))
4044                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4045                 else
4046                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4047 
4048                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4049 
4050                 tp->rx_opt.dsack = 1;
4051                 tp->duplicate_sack[0].start_seq = seq;
4052                 tp->duplicate_sack[0].end_seq = end_seq;
4053         }
4054 }
4055 
4056 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4057 {
4058         struct tcp_sock *tp = tcp_sk(sk);
4059 
4060         if (!tp->rx_opt.dsack)
4061                 tcp_dsack_set(sk, seq, end_seq);
4062         else
4063                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4064 }
4065 
4066 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4067 {
4068         struct tcp_sock *tp = tcp_sk(sk);
4069 
4070         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4071             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4072                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4073                 tcp_enter_quickack_mode(sk);
4074 
4075                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4076                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4077 
4078                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4079                                 end_seq = tp->rcv_nxt;
4080                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4081                 }
4082         }
4083 
4084         tcp_send_ack(sk);
4085 }
4086 
4087 /* These routines update the SACK block as out-of-order packets arrive or
4088  * in-order packets close up the sequence space.
4089  */
4090 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4091 {
4092         int this_sack;
4093         struct tcp_sack_block *sp = &tp->selective_acks[0];
4094         struct tcp_sack_block *swalk = sp + 1;
4095 
4096         /* See if the recent change to the first SACK eats into
4097          * or hits the sequence space of other SACK blocks, if so coalesce.
4098          */
4099         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4100                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4101                         int i;
4102 
4103                         /* Zap SWALK, by moving every further SACK up by one slot.
4104                          * Decrease num_sacks.
4105                          */
4106                         tp->rx_opt.num_sacks--;
4107                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4108                                 sp[i] = sp[i + 1];
4109                         continue;
4110                 }
4111                 this_sack++, swalk++;
4112         }
4113 }
4114 
4115 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4116 {
4117         struct tcp_sock *tp = tcp_sk(sk);
4118         struct tcp_sack_block *sp = &tp->selective_acks[0];
4119         int cur_sacks = tp->rx_opt.num_sacks;
4120         int this_sack;
4121 
4122         if (!cur_sacks)
4123                 goto new_sack;
4124 
4125         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4126                 if (tcp_sack_extend(sp, seq, end_seq)) {
4127                         /* Rotate this_sack to the first one. */
4128                         for (; this_sack > 0; this_sack--, sp--)
4129                                 swap(*sp, *(sp - 1));
4130                         if (cur_sacks > 1)
4131                                 tcp_sack_maybe_coalesce(tp);
4132                         return;
4133                 }
4134         }
4135 
4136         /* Could not find an adjacent existing SACK, build a new one,
4137          * put it at the front, and shift everyone else down.  We
4138          * always know there is at least one SACK present already here.
4139          *
4140          * If the sack array is full, forget about the last one.
4141          */
4142         if (this_sack >= TCP_NUM_SACKS) {
4143                 this_sack--;
4144                 tp->rx_opt.num_sacks--;
4145                 sp--;
4146         }
4147         for (; this_sack > 0; this_sack--, sp--)
4148                 *sp = *(sp - 1);
4149 
4150 new_sack:
4151         /* Build the new head SACK, and we're done. */
4152         sp->start_seq = seq;
4153         sp->end_seq = end_seq;
4154         tp->rx_opt.num_sacks++;
4155 }
4156 
4157 /* RCV.NXT advances, some SACKs should be eaten. */
4158 
4159 static void tcp_sack_remove(struct tcp_sock *tp)
4160 {
4161         struct tcp_sack_block *sp = &tp->selective_acks[0];
4162         int num_sacks = tp->rx_opt.num_sacks;
4163         int this_sack;
4164 
4165         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4166         if (skb_queue_empty(&tp->out_of_order_queue)) {
4167                 tp->rx_opt.num_sacks = 0;
4168                 return;
4169         }
4170 
4171         for (this_sack = 0; this_sack < num_sacks;) {
4172                 /* Check if the start of the sack is covered by RCV.NXT. */
4173                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4174                         int i;
4175 
4176                         /* RCV.NXT must cover all the block! */
4177                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4178 
4179                         /* Zap this SACK, by moving forward any other SACKS. */
4180                         for (i = this_sack+1; i < num_sacks; i++)
4181                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4182                         num_sacks--;
4183                         continue;
4184                 }
4185                 this_sack++;
4186                 sp++;
4187         }
4188         tp->rx_opt.num_sacks = num_sacks;
4189 }
4190 
4191 /**
4192  * tcp_try_coalesce - try to merge skb to prior one
4193  * @sk: socket
4194  * @to: prior buffer
4195  * @from: buffer to add in queue
4196  * @fragstolen: pointer to boolean
4197  *
4198  * Before queueing skb @from after @to, try to merge them
4199  * to reduce overall memory use and queue lengths, if cost is small.
4200  * Packets in ofo or receive queues can stay a long time.
4201  * Better try to coalesce them right now to avoid future collapses.
4202  * Returns true if caller should free @from instead of queueing it
4203  */
4204 static bool tcp_try_coalesce(struct sock *sk,
4205                              struct sk_buff *to,
4206                              struct sk_buff *from,
4207                              bool *fragstolen)
4208 {
4209         int delta;
4210 
4211         *fragstolen = false;
4212 
4213         /* Its possible this segment overlaps with prior segment in queue */
4214         if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4215                 return false;
4216 
4217         if (!skb_try_coalesce(to, from, fragstolen, &delta))
4218                 return false;
4219 
4220         atomic_add(delta, &sk->sk_rmem_alloc);
4221         sk_mem_charge(sk, delta);
4222         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4223         TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4224         TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4225         TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4226         return true;
4227 }
4228 
4229 /* This one checks to see if we can put data from the
4230  * out_of_order queue into the receive_queue.
4231  */
4232 static void tcp_ofo_queue(struct sock *sk)
4233 {
4234         struct tcp_sock *tp = tcp_sk(sk);
4235         __u32 dsack_high = tp->rcv_nxt;
4236         struct sk_buff *skb, *tail;
4237         bool fragstolen, eaten;
4238 
4239         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4240                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4241                         break;
4242 
4243                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4244                         __u32 dsack = dsack_high;
4245                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4246                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4247                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4248                 }
4249 
4250                 __skb_unlink(skb, &tp->out_of_order_queue);
4251                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4252                         SOCK_DEBUG(sk, "ofo packet was already received\n");
4253                         __kfree_skb(skb);
4254                         continue;
4255                 }
4256                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4257                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4258                            TCP_SKB_CB(skb)->end_seq);
4259 
4260                 tail = skb_peek_tail(&sk->sk_receive_queue);
4261                 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4262                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4263                 if (!eaten)
4264                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4265                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4266                         tcp_fin(sk);
4267                 if (eaten)
4268                         kfree_skb_partial(skb, fragstolen);
4269         }
4270 }
4271 
4272 static bool tcp_prune_ofo_queue(struct sock *sk);
4273 static int tcp_prune_queue(struct sock *sk);
4274 
4275 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4276                                  unsigned int size)
4277 {
4278         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4279             !sk_rmem_schedule(sk, skb, size)) {
4280 
4281                 if (tcp_prune_queue(sk) < 0)
4282                         return -1;
4283 
4284                 if (!sk_rmem_schedule(sk, skb, size)) {
4285                         if (!tcp_prune_ofo_queue(sk))
4286                                 return -1;
4287 
4288                         if (!sk_rmem_schedule(sk, skb, size))
4289                                 return -1;
4290                 }
4291         }
4292         return 0;
4293 }
4294 
4295 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4296 {
4297         struct tcp_sock *tp = tcp_sk(sk);
4298         struct sk_buff *skb1;
4299         u32 seq, end_seq;
4300 
4301         tcp_ecn_check_ce(tp, skb);
4302 
4303         if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4304                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4305                 __kfree_skb(skb);
4306                 return;
4307         }
4308 
4309         /* Disable header prediction. */
4310         tp->pred_flags = 0;
4311         inet_csk_schedule_ack(sk);
4312 
4313         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4314         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4315                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4316 
4317         skb1 = skb_peek_tail(&tp->out_of_order_queue);
4318         if (!skb1) {
4319                 /* Initial out of order segment, build 1 SACK. */
4320                 if (tcp_is_sack(tp)) {
4321                         tp->rx_opt.num_sacks = 1;
4322                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4323                         tp->selective_acks[0].end_seq =
4324                                                 TCP_SKB_CB(skb)->end_seq;
4325                 }
4326                 __skb_queue_head(&tp->out_of_order_queue, skb);
4327                 goto end;
4328         }
4329 
4330         seq = TCP_SKB_CB(skb)->seq;
4331         end_seq = TCP_SKB_CB(skb)->end_seq;
4332 
4333         if (seq == TCP_SKB_CB(skb1)->end_seq) {
4334                 bool fragstolen;
4335 
4336                 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4337                         __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4338                 } else {
4339                         tcp_grow_window(sk, skb);
4340                         kfree_skb_partial(skb, fragstolen);
4341                         skb = NULL;
4342                 }
4343 
4344                 if (!tp->rx_opt.num_sacks ||
4345                     tp->selective_acks[0].end_seq != seq)
4346                         goto add_sack;
4347 
4348                 /* Common case: data arrive in order after hole. */
4349                 tp->selective_acks[0].end_seq = end_seq;
4350                 goto end;
4351         }
4352 
4353         /* Find place to insert this segment. */
4354         while (1) {
4355                 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4356                         break;
4357                 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4358                         skb1 = NULL;
4359                         break;
4360                 }
4361                 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4362         }
4363 
4364         /* Do skb overlap to previous one? */
4365         if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4366                 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4367                         /* All the bits are present. Drop. */
4368                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4369                         __kfree_skb(skb);
4370                         skb = NULL;
4371                         tcp_dsack_set(sk, seq, end_seq);
4372                         goto add_sack;
4373                 }
4374                 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4375                         /* Partial overlap. */
4376                         tcp_dsack_set(sk, seq,
4377                                       TCP_SKB_CB(skb1)->end_seq);
4378                 } else {
4379                         if (skb_queue_is_first(&tp->out_of_order_queue,
4380                                                skb1))
4381                                 skb1 = NULL;
4382                         else
4383                                 skb1 = skb_queue_prev(
4384                                         &tp->out_of_order_queue,
4385                                         skb1);
4386                 }
4387         }
4388         if (!skb1)
4389                 __skb_queue_head(&tp->out_of_order_queue, skb);
4390         else
4391                 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4392 
4393         /* And clean segments covered by new one as whole. */
4394         while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4395                 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4396 
4397                 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4398                         break;
4399                 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4400                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4401                                          end_seq);
4402                         break;
4403                 }
4404                 __skb_unlink(skb1, &tp->out_of_order_queue);
4405                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4406                                  TCP_SKB_CB(skb1)->end_seq);
4407                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4408                 __kfree_skb(skb1);
4409         }
4410 
4411 add_sack:
4412         if (tcp_is_sack(tp))
4413                 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4414 end:
4415         if (skb) {
4416                 tcp_grow_window(sk, skb);
4417                 skb_set_owner_r(skb, sk);
4418         }
4419 }
4420 
4421 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4422                   bool *fragstolen)
4423 {
4424         int eaten;
4425         struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4426 
4427         __skb_pull(skb, hdrlen);
4428         eaten = (tail &&
4429                  tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4430         tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4431         if (!eaten) {
4432                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4433                 skb_set_owner_r(skb, sk);
4434         }
4435         return eaten;
4436 }
4437 
4438 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4439 {
4440         struct sk_buff *skb;
4441         bool fragstolen;
4442 
4443         if (size == 0)
4444                 return 0;
4445 
4446         skb = alloc_skb(size, sk->sk_allocation);
4447         if (!skb)
4448                 goto err;
4449 
4450         if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4451                 goto err_free;
4452 
4453         if (memcpy_from_msg(skb_put(skb, size), msg, size))
4454                 goto err_free;
4455 
4456         TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4457         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4458         TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4459 
4460         if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4461                 WARN_ON_ONCE(fragstolen); /* should not happen */
4462                 __kfree_skb(skb);
4463         }
4464         return size;
4465 
4466 err_free:
4467         kfree_skb(skb);
4468 err:
4469         return -ENOMEM;
4470 }
4471 
4472 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4473 {
4474         struct tcp_sock *tp = tcp_sk(sk);
4475         int eaten = -1;
4476         bool fragstolen = false;
4477 
4478         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4479                 goto drop;
4480 
4481         skb_dst_drop(skb);
4482         __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4483 
4484         tcp_ecn_accept_cwr(tp, skb);
4485 
4486         tp->rx_opt.dsack = 0;
4487 
4488         /*  Queue data for delivery to the user.
4489          *  Packets in sequence go to the receive queue.
4490          *  Out of sequence packets to the out_of_order_queue.
4491          */
4492         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4493                 if (tcp_receive_window(tp) == 0)
4494                         goto out_of_window;
4495 
4496                 /* Ok. In sequence. In window. */
4497                 if (tp->ucopy.task == current &&
4498                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4499                     sock_owned_by_user(sk) && !tp->urg_data) {
4500                         int chunk = min_t(unsigned int, skb->len,
4501                                           tp->ucopy.len);
4502 
4503                         __set_current_state(TASK_RUNNING);
4504 
4505                         local_bh_enable();
4506                         if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4507                                 tp->ucopy.len -= chunk;
4508                                 tp->copied_seq += chunk;
4509                                 eaten = (chunk == skb->len);
4510                                 tcp_rcv_space_adjust(sk);
4511                         }
4512                         local_bh_disable();
4513                 }
4514 
4515                 if (eaten <= 0) {
4516 queue_and_out:
4517                         if (eaten < 0 &&
4518                             tcp_try_rmem_schedule(sk, skb, skb->truesize))
4519                                 goto drop;
4520 
4521                         eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4522                 }
4523                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4524                 if (skb->len)
4525                         tcp_event_data_recv(sk, skb);
4526                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4527                         tcp_fin(sk);
4528 
4529                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4530                         tcp_ofo_queue(sk);
4531 
4532                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4533                          * gap in queue is filled.
4534                          */
4535                         if (skb_queue_empty(&tp->out_of_order_queue))
4536                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4537                 }
4538 
4539                 if (tp->rx_opt.num_sacks)
4540                         tcp_sack_remove(tp);
4541 
4542                 tcp_fast_path_check(sk);
4543 
4544                 if (eaten > 0)
4545                         kfree_skb_partial(skb, fragstolen);
4546                 if (!sock_flag(sk, SOCK_DEAD))
4547                         sk->sk_data_ready(sk);
4548                 return;
4549         }
4550 
4551         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4552                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4553                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4554                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4555 
4556 out_of_window:
4557                 tcp_enter_quickack_mode(sk);
4558                 inet_csk_schedule_ack(sk);
4559 drop:
4560                 __kfree_skb(skb);
4561                 return;
4562         }
4563 
4564         /* Out of window. F.e. zero window probe. */
4565         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4566                 goto out_of_window;
4567 
4568         tcp_enter_quickack_mode(sk);
4569 
4570         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4571                 /* Partial packet, seq < rcv_next < end_seq */
4572                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4573                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4574                            TCP_SKB_CB(skb)->end_seq);
4575 
4576                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4577 
4578                 /* If window is closed, drop tail of packet. But after
4579                  * remembering D-SACK for its head made in previous line.
4580                  */
4581                 if (!tcp_receive_window(tp))
4582                         goto out_of_window;
4583                 goto queue_and_out;
4584         }
4585 
4586         tcp_data_queue_ofo(sk, skb);
4587 }
4588 
4589 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4590                                         struct sk_buff_head *list)
4591 {
4592         struct sk_buff *next = NULL;
4593 
4594         if (!skb_queue_is_last(list, skb))
4595                 next = skb_queue_next(list, skb);
4596 
4597         __skb_unlink(skb, list);
4598         __kfree_skb(skb);
4599         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4600 
4601         return next;
4602 }
4603 
4604 /* Collapse contiguous sequence of skbs head..tail with
4605  * sequence numbers start..end.
4606  *
4607  * If tail is NULL, this means until the end of the list.
4608  *
4609  * Segments with FIN/SYN are not collapsed (only because this
4610  * simplifies code)
4611  */
4612 static void
4613 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4614              struct sk_buff *head, struct sk_buff *tail,
4615              u32 start, u32 end)
4616 {
4617         struct sk_buff *skb, *n;
4618         bool end_of_skbs;
4619 
4620         /* First, check that queue is collapsible and find
4621          * the point where collapsing can be useful. */
4622         skb = head;
4623 restart:
4624         end_of_skbs = true;
4625         skb_queue_walk_from_safe(list, skb, n) {
4626                 if (skb == tail)
4627                         break;
4628                 /* No new bits? It is possible on ofo queue. */
4629                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4630                         skb = tcp_collapse_one(sk, skb, list);
4631                         if (!skb)
4632                                 break;
4633                         goto restart;
4634                 }
4635 
4636                 /* The first skb to collapse is:
4637                  * - not SYN/FIN and
4638                  * - bloated or contains data before "start" or
4639                  *   overlaps to the next one.
4640                  */
4641                 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4642                     (tcp_win_from_space(skb->truesize) > skb->len ||
4643                      before(TCP_SKB_CB(skb)->seq, start))) {
4644                         end_of_skbs = false;
4645                         break;
4646                 }
4647 
4648                 if (!skb_queue_is_last(list, skb)) {
4649                         struct sk_buff *next = skb_queue_next(list, skb);
4650                         if (next != tail &&
4651                             TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4652                                 end_of_skbs = false;
4653                                 break;
4654                         }
4655                 }
4656 
4657                 /* Decided to skip this, advance start seq. */
4658                 start = TCP_SKB_CB(skb)->end_seq;
4659         }
4660         if (end_of_skbs ||
4661             (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4662                 return;
4663 
4664         while (before(start, end)) {
4665                 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4666                 struct sk_buff *nskb;
4667 
4668                 nskb = alloc_skb(copy, GFP_ATOMIC);
4669                 if (!nskb)
4670                         return;
4671 
4672                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4673                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4674                 __skb_queue_before(list, skb, nskb);
4675                 skb_set_owner_r(nskb, sk);
4676 
4677                 /* Copy data, releasing collapsed skbs. */
4678                 while (copy > 0) {
4679                         int offset = start - TCP_SKB_CB(skb)->seq;
4680                         int size = TCP_SKB_CB(skb)->end_seq - start;
4681 
4682                         BUG_ON(offset < 0);
4683                         if (size > 0) {
4684                                 size = min(copy, size);
4685                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4686                                         BUG();
4687                                 TCP_SKB_CB(nskb)->end_seq += size;
4688                                 copy -= size;
4689                                 start += size;
4690                         }
4691                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4692                                 skb = tcp_collapse_one(sk, skb, list);
4693                                 if (!skb ||
4694                                     skb == tail ||
4695                                     (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4696                                         return;
4697                         }
4698                 }
4699         }
4700 }
4701 
4702 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4703  * and tcp_collapse() them until all the queue is collapsed.
4704  */
4705 static void tcp_collapse_ofo_queue(struct sock *sk)
4706 {
4707         struct tcp_sock *tp = tcp_sk(sk);
4708         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4709         struct sk_buff *head;
4710         u32 start, end;
4711 
4712         if (!skb)
4713                 return;
4714 
4715         start = TCP_SKB_CB(skb)->seq;
4716         end = TCP_SKB_CB(skb)->end_seq;
4717         head = skb;
4718 
4719         for (;;) {
4720                 struct sk_buff *next = NULL;
4721 
4722                 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4723                         next = skb_queue_next(&tp->out_of_order_queue, skb);
4724                 skb = next;
4725 
4726                 /* Segment is terminated when we see gap or when
4727                  * we are at the end of all the queue. */
4728                 if (!skb ||
4729                     after(TCP_SKB_CB(skb)->seq, end) ||
4730                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4731                         tcp_collapse(sk, &tp->out_of_order_queue,
4732                                      head, skb, start, end);
4733                         head = skb;
4734                         if (!skb)
4735                                 break;
4736                         /* Start new segment */
4737                         start = TCP_SKB_CB(skb)->seq;
4738                         end = TCP_SKB_CB(skb)->end_seq;
4739                 } else {
4740                         if (before(TCP_SKB_CB(skb)->seq, start))
4741                                 start = TCP_SKB_CB(skb)->seq;
4742                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4743                                 end = TCP_SKB_CB(skb)->end_seq;
4744                 }
4745         }
4746 }
4747 
4748 /*
4749  * Purge the out-of-order queue.
4750  * Return true if queue was pruned.
4751  */
4752 static bool tcp_prune_ofo_queue(struct sock *sk)
4753 {
4754         struct tcp_sock *tp = tcp_sk(sk);
4755         bool res = false;
4756 
4757         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4758                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4759                 __skb_queue_purge(&tp->out_of_order_queue);
4760 
4761                 /* Reset SACK state.  A conforming SACK implementation will
4762                  * do the same at a timeout based retransmit.  When a connection
4763                  * is in a sad state like this, we care only about integrity
4764                  * of the connection not performance.
4765                  */
4766                 if (tp->rx_opt.sack_ok)
4767                         tcp_sack_reset(&tp->rx_opt);
4768                 sk_mem_reclaim(sk);
4769                 res = true;
4770         }
4771         return res;
4772 }
4773 
4774 /* Reduce allocated memory if we can, trying to get
4775  * the socket within its memory limits again.
4776  *
4777  * Return less than zero if we should start dropping frames
4778  * until the socket owning process reads some of the data
4779  * to stabilize the situation.
4780  */
4781 static int tcp_prune_queue(struct sock *sk)
4782 {
4783         struct tcp_sock *tp = tcp_sk(sk);
4784 
4785         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4786 
4787         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4788 
4789         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4790                 tcp_clamp_window(sk);
4791         else if (sk_under_memory_pressure(sk))
4792                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4793 
4794         tcp_collapse_ofo_queue(sk);
4795         if (!skb_queue_empty(&sk->sk_receive_queue))
4796                 tcp_collapse(sk, &sk->sk_receive_queue,
4797                              skb_peek(&sk->sk_receive_queue),
4798                              NULL,
4799                              tp->copied_seq, tp->rcv_nxt);
4800         sk_mem_reclaim(sk);
4801 
4802         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4803                 return 0;
4804 
4805         /* Collapsing did not help, destructive actions follow.
4806          * This must not ever occur. */
4807 
4808         tcp_prune_ofo_queue(sk);
4809 
4810         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4811                 return 0;
4812 
4813         /* If we are really being abused, tell the caller to silently
4814          * drop receive data on the floor.  It will get retransmitted
4815          * and hopefully then we'll have sufficient space.
4816          */
4817         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4818 
4819         /* Massive buffer overcommit. */
4820         tp->pred_flags = 0;
4821         return -1;
4822 }
4823 
4824 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4825 {
4826         const struct tcp_sock *tp = tcp_sk(sk);
4827 
4828         /* If the user specified a specific send buffer setting, do
4829          * not modify it.
4830          */
4831         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4832                 return false;
4833 
4834         /* If we are under global TCP memory pressure, do not expand.  */
4835         if (sk_under_memory_pressure(sk))
4836                 return false;
4837 
4838         /* If we are under soft global TCP memory pressure, do not expand.  */
4839         if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4840                 return false;
4841 
4842         /* If we filled the congestion window, do not expand.  */
4843         if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4844                 return false;
4845 
4846         return true;
4847 }
4848 
4849 /* When incoming ACK allowed to free some skb from write_queue,
4850  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4851  * on the exit from tcp input handler.
4852  *
4853  * PROBLEM: sndbuf expansion does not work well with largesend.
4854  */
4855 static void tcp_new_space(struct sock *sk)
4856 {
4857         struct tcp_sock *tp = tcp_sk(sk);
4858 
4859         if (tcp_should_expand_sndbuf(sk)) {
4860                 tcp_sndbuf_expand(sk);
4861                 tp->snd_cwnd_stamp = tcp_time_stamp;
4862         }
4863 
4864         sk->sk_write_space(sk);
4865 }
4866 
4867 static void tcp_check_space(struct sock *sk)
4868 {
4869         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4870                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4871                 /* pairs with tcp_poll() */
4872                 smp_mb__after_atomic();
4873                 if (sk->sk_socket &&
4874                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4875                         tcp_new_space(sk);
4876         }
4877 }
4878 
4879 static inline void tcp_data_snd_check(struct sock *sk)
4880 {
4881         tcp_push_pending_frames(sk);
4882         tcp_check_space(sk);
4883 }
4884 
4885 /*
4886  * Check if sending an ack is needed.
4887  */
4888 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4889 {
4890         struct tcp_sock *tp = tcp_sk(sk);
4891 
4892             /* More than one full frame received... */
4893         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4894              /* ... and right edge of window advances far enough.
4895               * (tcp_recvmsg() will send ACK otherwise). Or...
4896               */
4897              __tcp_select_window(sk) >= tp->rcv_wnd) ||
4898             /* We ACK each frame or... */
4899             tcp_in_quickack_mode(sk) ||
4900             /* We have out of order data. */
4901             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4902                 /* Then ack it now */
4903                 tcp_send_ack(sk);
4904         } else {
4905                 /* Else, send delayed ack. */
4906                 tcp_send_delayed_ack(sk);
4907         }
4908 }
4909 
4910 static inline void tcp_ack_snd_check(struct sock *sk)
4911 {
4912         if (!inet_csk_ack_scheduled(sk)) {
4913                 /* We sent a data segment already. */
4914                 return;
4915         }
4916         __tcp_ack_snd_check(sk, 1);
4917 }
4918 
4919 /*
4920  *      This routine is only called when we have urgent data
4921  *      signaled. Its the 'slow' part of tcp_urg. It could be
4922  *      moved inline now as tcp_urg is only called from one
4923  *      place. We handle URGent data wrong. We have to - as
4924  *      BSD still doesn't use the correction from RFC961.
4925  *      For 1003.1g we should support a new option TCP_STDURG to permit
4926  *      either form (or just set the sysctl tcp_stdurg).
4927  */
4928 
4929 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4930 {
4931         struct tcp_sock *tp = tcp_sk(sk);
4932         u32 ptr = ntohs(th->urg_ptr);
4933 
4934         if (ptr && !sysctl_tcp_stdurg)
4935                 ptr--;
4936         ptr += ntohl(th->seq);
4937 
4938         /* Ignore urgent data that we've already seen and read. */
4939         if (after(tp->copied_seq, ptr))
4940                 return;
4941 
4942         /* Do not replay urg ptr.
4943          *
4944          * NOTE: interesting situation not covered by specs.
4945          * Misbehaving sender may send urg ptr, pointing to segment,
4946          * which we already have in ofo queue. We are not able to fetch
4947          * such data and will stay in TCP_URG_NOTYET until will be eaten
4948          * by recvmsg(). Seems, we are not obliged to handle such wicked
4949          * situations. But it is worth to think about possibility of some
4950          * DoSes using some hypothetical application level deadlock.
4951          */
4952         if (before(ptr, tp->rcv_nxt))
4953                 return;
4954 
4955         /* Do we already have a newer (or duplicate) urgent pointer? */
4956         if (tp->urg_data && !after(ptr, tp->urg_seq))
4957                 return;
4958 
4959         /* Tell the world about our new urgent pointer. */
4960         sk_send_sigurg(sk);
4961 
4962         /* We may be adding urgent data when the last byte read was
4963          * urgent. To do this requires some care. We cannot just ignore
4964          * tp->copied_seq since we would read the last urgent byte again
4965          * as data, nor can we alter copied_seq until this data arrives
4966          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4967          *
4968          * NOTE. Double Dutch. Rendering to plain English: author of comment
4969          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4970          * and expect that both A and B disappear from stream. This is _wrong_.
4971          * Though this happens in BSD with high probability, this is occasional.
4972          * Any application relying on this is buggy. Note also, that fix "works"
4973          * only in this artificial test. Insert some normal data between A and B and we will
4974          * decline of BSD again. Verdict: it is better to remove to trap
4975          * buggy users.
4976          */
4977         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4978             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4979                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4980                 tp->copied_seq++;
4981                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4982                         __skb_unlink(skb, &sk->sk_receive_queue);
4983                         __kfree_skb(skb);
4984                 }
4985         }
4986 
4987         tp->urg_data = TCP_URG_NOTYET;
4988         tp->urg_seq = ptr;
4989 
4990         /* Disable header prediction. */
4991         tp->pred_flags = 0;
4992 }
4993 
4994 /* This is the 'fast' part of urgent handling. */
4995 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4996 {
4997         struct tcp_sock *tp = tcp_sk(sk);
4998 
4999         /* Check if we get a new urgent pointer - normally not. */
5000         if (th->urg)
5001                 tcp_check_urg(sk, th);
5002 
5003         /* Do we wait for any urgent data? - normally not... */
5004         if (tp->urg_data == TCP_URG_NOTYET) {
5005                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5006                           th->syn;
5007 
5008                 /* Is the urgent pointer pointing into this packet? */
5009                 if (ptr < skb->len) {
5010                         u8 tmp;
5011                         if (skb_copy_bits(skb, ptr, &tmp, 1))
5012                                 BUG();
5013                         tp->urg_data = TCP_URG_VALID | tmp;
5014                         if (!sock_flag(sk, SOCK_DEAD))
5015                                 sk->sk_data_ready(sk);
5016                 }
5017         }
5018 }
5019 
5020 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5021 {
5022         struct tcp_sock *tp = tcp_sk(sk);
5023         int chunk = skb->len - hlen;
5024         int err;
5025 
5026         local_bh_enable();
5027         if (skb_csum_unnecessary(skb))
5028                 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5029         else
5030                 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5031 
5032         if (!err) {
5033                 tp->ucopy.len -= chunk;
5034                 tp->copied_seq += chunk;
5035                 tcp_rcv_space_adjust(sk);
5036         }
5037 
5038         local_bh_disable();
5039         return err;
5040 }
5041 
5042 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5043                                             struct sk_buff *skb)
5044 {
5045         __sum16 result;
5046 
5047         if (sock_owned_by_user(sk)) {
5048                 local_bh_enable();
5049                 result = __tcp_checksum_complete(skb);
5050                 local_bh_disable();
5051         } else {
5052                 result = __tcp_checksum_complete(skb);
5053         }
5054         return result;
5055 }
5056 
5057 static inline bool tcp_checksum_complete_user(struct sock *sk,
5058                                              struct sk_buff *skb)
5059 {
5060         return !skb_csum_unnecessary(skb) &&
5061                __tcp_checksum_complete_user(sk, skb);
5062 }
5063 
5064 /* Does PAWS and seqno based validation of an incoming segment, flags will
5065  * play significant role here.
5066  */
5067 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5068                                   const struct tcphdr *th, int syn_inerr)
5069 {
5070         struct tcp_sock *tp = tcp_sk(sk);
5071 
5072         /* RFC1323: H1. Apply PAWS check first. */
5073         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5074             tcp_paws_discard(sk, skb)) {
5075                 if (!th->rst) {
5076                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5077                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5078                                                   LINUX_MIB_TCPACKSKIPPEDPAWS,
5079                                                   &tp->last_oow_ack_time))
5080                                 tcp_send_dupack(sk, skb);
5081                         goto discard;
5082                 }
5083                 /* Reset is accepted even if it did not pass PAWS. */
5084         }
5085 
5086         /* Step 1: check sequence number */
5087         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5088                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5089                  * (RST) segments are validated by checking their SEQ-fields."
5090                  * And page 69: "If an incoming segment is not acceptable,
5091                  * an acknowledgment should be sent in reply (unless the RST
5092                  * bit is set, if so drop the segment and return)".
5093                  */
5094                 if (!th->rst) {
5095                         if (th->syn)
5096                                 goto syn_challenge;
5097                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5098                                                   LINUX_MIB_TCPACKSKIPPEDSEQ,
5099                                                   &tp->last_oow_ack_time))
5100                                 tcp_send_dupack(sk, skb);
5101                 }
5102                 goto discard;
5103         }
5104 
5105         /* Step 2: check RST bit */
5106         if (th->rst) {
5107                 /* RFC 5961 3.2 :
5108                  * If sequence number exactly matches RCV.NXT, then
5109                  *     RESET the connection
5110                  * else
5111                  *     Send a challenge ACK
5112                  */
5113                 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5114                         tcp_reset(sk);
5115                 else
5116                         tcp_send_challenge_ack(sk, skb);
5117                 goto discard;
5118         }
5119 
5120         /* step 3: check security and precedence [ignored] */
5121 
5122         /* step 4: Check for a SYN
5123          * RFC 5961 4.2 : Send a challenge ack
5124          */
5125         if (th->syn) {
5126 syn_challenge:
5127                 if (syn_inerr)
5128                         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5129                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5130                 tcp_send_challenge_ack(sk, skb);
5131                 goto discard;
5132         }
5133 
5134         return true;
5135 
5136 discard:
5137         __kfree_skb(skb);
5138         return false;
5139 }
5140 
5141 /*
5142  *      TCP receive function for the ESTABLISHED state.
5143  *
5144  *      It is split into a fast path and a slow path. The fast path is
5145  *      disabled when:
5146  *      - A zero window was announced from us - zero window probing
5147  *        is only handled properly in the slow path.
5148  *      - Out of order segments arrived.
5149  *      - Urgent data is expected.
5150  *      - There is no buffer space left
5151  *      - Unexpected TCP flags/window values/header lengths are received
5152  *        (detected by checking the TCP header against pred_flags)
5153  *      - Data is sent in both directions. Fast path only supports pure senders
5154  *        or pure receivers (this means either the sequence number or the ack
5155  *        value must stay constant)
5156  *      - Unexpected TCP option.
5157  *
5158  *      When these conditions are not satisfied it drops into a standard
5159  *      receive procedure patterned after RFC793 to handle all cases.
5160  *      The first three cases are guaranteed by proper pred_flags setting,
5161  *      the rest is checked inline. Fast processing is turned on in
5162  *      tcp_data_queue when everything is OK.
5163  */
5164 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5165                          const struct tcphdr *th, unsigned int len)
5166 {
5167         struct tcp_sock *tp = tcp_sk(sk);
5168 
5169         if (unlikely(!sk->sk_rx_dst))
5170                 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5171         /*
5172          *      Header prediction.
5173          *      The code loosely follows the one in the famous
5174          *      "30 instruction TCP receive" Van Jacobson mail.
5175          *
5176          *      Van's trick is to deposit buffers into socket queue
5177          *      on a device interrupt, to call tcp_recv function
5178          *      on the receive process context and checksum and copy
5179          *      the buffer to user space. smart...
5180          *
5181          *      Our current scheme is not silly either but we take the
5182          *      extra cost of the net_bh soft interrupt processing...
5183          *      We do checksum and copy also but from device to kernel.
5184          */
5185 
5186         tp->rx_opt.saw_tstamp = 0;
5187 
5188         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5189          *      if header_prediction is to be made
5190          *      'S' will always be tp->tcp_header_len >> 2
5191          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5192          *  turn it off (when there are holes in the receive
5193          *       space for instance)
5194          *      PSH flag is ignored.
5195          */
5196 
5197         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5198             TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5199             !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5200                 int tcp_header_len = tp->tcp_header_len;
5201 
5202                 /* Timestamp header prediction: tcp_header_len
5203                  * is automatically equal to th->doff*4 due to pred_flags
5204                  * match.
5205                  */
5206 
5207                 /* Check timestamp */
5208                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5209                         /* No? Slow path! */
5210                         if (!tcp_parse_aligned_timestamp(tp, th))
5211                                 goto slow_path;
5212 
5213                         /* If PAWS failed, check it more carefully in slow path */
5214                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5215                                 goto slow_path;
5216 
5217                         /* DO NOT update ts_recent here, if checksum fails
5218                          * and timestamp was corrupted part, it will result
5219                          * in a hung connection since we will drop all
5220                          * future packets due to the PAWS test.
5221                          */
5222                 }
5223 
5224                 if (len <= tcp_header_len) {
5225                         /* Bulk data transfer: sender */
5226                         if (len == tcp_header_len) {
5227                                 /* Predicted packet is in window by definition.
5228                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5229                                  * Hence, check seq<=rcv_wup reduces to:
5230                                  */
5231                                 if (tcp_header_len ==
5232                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5233                                     tp->rcv_nxt == tp->rcv_wup)
5234                                         tcp_store_ts_recent(tp);
5235 
5236                                 /* We know that such packets are checksummed
5237                                  * on entry.
5238                                  */
5239                                 tcp_ack(sk, skb, 0);
5240                                 __kfree_skb(skb);
5241                                 tcp_data_snd_check(sk);
5242                                 return;
5243                         } else { /* Header too small */
5244                                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5245                                 goto discard;
5246                         }
5247                 } else {
5248                         int eaten = 0;
5249                         bool fragstolen = false;
5250 
5251                         if (tp->ucopy.task == current &&
5252                             tp->copied_seq == tp->rcv_nxt &&
5253                             len - tcp_header_len <= tp->ucopy.len &&
5254                             sock_owned_by_user(sk)) {
5255                                 __set_current_state(TASK_RUNNING);
5256 
5257                                 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5258                                         /* Predicted packet is in window by definition.
5259                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5260                                          * Hence, check seq<=rcv_wup reduces to:
5261                                          */
5262                                         if (tcp_header_len ==
5263                                             (sizeof(struct tcphdr) +
5264                                              TCPOLEN_TSTAMP_ALIGNED) &&
5265                                             tp->rcv_nxt == tp->rcv_wup)
5266                                                 tcp_store_ts_recent(tp);
5267 
5268                                         tcp_rcv_rtt_measure_ts(sk, skb);
5269 
5270                                         __skb_pull(skb, tcp_header_len);
5271                                         tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5272                                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5273                                         eaten = 1;
5274                                 }
5275                         }
5276                         if (!eaten) {
5277                                 if (tcp_checksum_complete_user(sk, skb))
5278                                         goto csum_error;
5279 
5280                                 if ((int)skb->truesize > sk->sk_forward_alloc)
5281                                         goto step5;
5282 
5283                                 /* Predicted packet is in window by definition.
5284                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5285                                  * Hence, check seq<=rcv_wup reduces to:
5286                                  */
5287                                 if (tcp_header_len ==
5288                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5289                                     tp->rcv_nxt == tp->rcv_wup)
5290                                         tcp_store_ts_recent(tp);
5291 
5292                                 tcp_rcv_rtt_measure_ts(sk, skb);
5293 
5294                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5295 
5296                                 /* Bulk data transfer: receiver */
5297                                 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5298                                                       &fragstolen);
5299                         }
5300 
5301                         tcp_event_data_recv(sk, skb);
5302 
5303                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5304                                 /* Well, only one small jumplet in fast path... */
5305                                 tcp_ack(sk, skb, FLAG_DATA);
5306                                 tcp_data_snd_check(sk);
5307                                 if (!inet_csk_ack_scheduled(sk))
5308                                         goto no_ack;
5309                         }
5310 
5311                         __tcp_ack_snd_check(sk, 0);
5312 no_ack:
5313                         if (eaten)
5314                                 kfree_skb_partial(skb, fragstolen);
5315                         sk->sk_data_ready(sk);
5316                         return;
5317                 }
5318         }
5319 
5320 slow_path:
5321         if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5322                 goto csum_error;
5323 
5324         if (!th->ack && !th->rst && !th->syn)
5325                 goto discard;
5326 
5327         /*
5328          *      Standard slow path.
5329          */
5330 
5331         if (!tcp_validate_incoming(sk, skb, th, 1))
5332                 return;
5333 
5334 step5:
5335         if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5336                 goto discard;
5337 
5338         tcp_rcv_rtt_measure_ts(sk, skb);
5339 
5340         /* Process urgent data. */
5341         tcp_urg(sk, skb, th);
5342 
5343         /* step 7: process the segment text */
5344         tcp_data_queue(sk, skb);
5345 
5346         tcp_data_snd_check(sk);
5347         tcp_ack_snd_check(sk);
5348         return;
5349 
5350 csum_error:
5351         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5352         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5353 
5354 discard:
5355         __kfree_skb(skb);
5356 }
5357 EXPORT_SYMBOL(tcp_rcv_established);
5358 
5359 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5360 {
5361         struct tcp_sock *tp = tcp_sk(sk);
5362         struct inet_connection_sock *icsk = inet_csk(sk);
5363 
5364         tcp_set_state(sk, TCP_ESTABLISHED);
5365 
5366         if (skb) {
5367                 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5368                 security_inet_conn_established(sk, skb);
5369         }
5370 
5371         /* Make sure socket is routed, for correct metrics.  */
5372         icsk->icsk_af_ops->rebuild_header(sk);
5373 
5374         tcp_init_metrics(sk);
5375 
5376         tcp_init_congestion_control(sk);
5377 
5378         /* Prevent spurious tcp_cwnd_restart() on first data
5379          * packet.
5380          */
5381         tp->lsndtime = tcp_time_stamp;
5382 
5383         tcp_init_buffer_space(sk);
5384 
5385         if (sock_flag(sk, SOCK_KEEPOPEN))
5386                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5387 
5388         if (!tp->rx_opt.snd_wscale)
5389                 __tcp_fast_path_on(tp, tp->snd_wnd);
5390         else
5391                 tp->pred_flags = 0;
5392 
5393         if (!sock_flag(sk, SOCK_DEAD)) {
5394                 sk->sk_state_change(sk);
5395                 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5396         }
5397 }
5398 
5399 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5400                                     struct tcp_fastopen_cookie *cookie)
5401 {
5402         struct tcp_sock *tp = tcp_sk(sk);
5403         struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5404         u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5405         bool syn_drop = false;
5406 
5407         if (mss == tp->rx_opt.user_mss) {
5408                 struct tcp_options_received opt;
5409 
5410                 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5411                 tcp_clear_options(&opt);
5412                 opt.user_mss = opt.mss_clamp = 0;
5413                 tcp_parse_options(synack, &opt, 0, NULL);
5414                 mss = opt.mss_clamp;
5415         }
5416 
5417         if (!tp->syn_fastopen) {
5418                 /* Ignore an unsolicited cookie */
5419                 cookie->len = -1;
5420         } else if (tp->total_retrans) {
5421                 /* SYN timed out and the SYN-ACK neither has a cookie nor
5422                  * acknowledges data. Presumably the remote received only
5423                  * the retransmitted (regular) SYNs: either the original
5424                  * SYN-data or the corresponding SYN-ACK was dropped.
5425                  */
5426                 syn_drop = (cookie->len < 0 && data);
5427         } else if (cookie->len < 0 && !tp->syn_data) {
5428                 /* We requested a cookie but didn't get it. If we did not use
5429                  * the (old) exp opt format then try so next time (try_exp=1).
5430                  * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5431                  */
5432                 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5433         }
5434 
5435         tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5436 
5437         if (data) { /* Retransmit unacked data in SYN */
5438                 tcp_for_write_queue_from(data, sk) {
5439                         if (data == tcp_send_head(sk) ||
5440                             __tcp_retransmit_skb(sk, data))
5441                                 break;
5442                 }
5443                 tcp_rearm_rto(sk);
5444                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5445                 return true;
5446         }
5447         tp->syn_data_acked = tp->syn_data;
5448         if (tp->syn_data_acked)
5449                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5450         return false;
5451 }
5452 
5453 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5454                                          const struct tcphdr *th, unsigned int len)
5455 {
5456         struct inet_connection_sock *icsk = inet_csk(sk);
5457         struct tcp_sock *tp = tcp_sk(sk);
5458         struct tcp_fastopen_cookie foc = { .len = -1 };
5459         int saved_clamp = tp->rx_opt.mss_clamp;
5460 
5461         tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5462         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5463                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5464 
5465         if (th->ack) {
5466                 /* rfc793:
5467                  * "If the state is SYN-SENT then
5468                  *    first check the ACK bit
5469                  *      If the ACK bit is set
5470                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5471                  *        a reset (unless the RST bit is set, if so drop
5472                  *        the segment and return)"
5473                  */
5474                 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5475                     after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5476                         goto reset_and_undo;
5477 
5478                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5479                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5480                              tcp_time_stamp)) {
5481                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5482                         goto reset_and_undo;
5483                 }
5484 
5485                 /* Now ACK is acceptable.
5486                  *
5487                  * "If the RST bit is set
5488                  *    If the ACK was acceptable then signal the user "error:
5489                  *    connection reset", drop the segment, enter CLOSED state,
5490                  *    delete TCB, and return."
5491                  */
5492 
5493                 if (th->rst) {
5494                         tcp_reset(sk);
5495                         goto discard;
5496                 }
5497 
5498                 /* rfc793:
5499                  *   "fifth, if neither of the SYN or RST bits is set then
5500                  *    drop the segment and return."
5501                  *
5502                  *    See note below!
5503                  *                                        --ANK(990513)
5504                  */
5505                 if (!th->syn)
5506                         goto discard_and_undo;
5507 
5508                 /* rfc793:
5509                  *   "If the SYN bit is on ...
5510                  *    are acceptable then ...
5511                  *    (our SYN has been ACKed), change the connection
5512                  *    state to ESTABLISHED..."
5513                  */
5514 
5515                 tcp_ecn_rcv_synack(tp, th);
5516 
5517                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5518                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5519 
5520                 /* Ok.. it's good. Set up sequence numbers and
5521                  * move to established.
5522                  */
5523                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5524                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5525 
5526                 /* RFC1323: The window in SYN & SYN/ACK segments is
5527                  * never scaled.
5528                  */
5529                 tp->snd_wnd = ntohs(th->window);
5530 
5531                 if (!tp->rx_opt.wscale_ok) {
5532                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5533                         tp->window_clamp = min(tp->window_clamp, 65535U);
5534                 }
5535 
5536                 if (tp->rx_opt.saw_tstamp) {
5537                         tp->rx_opt.tstamp_ok       = 1;
5538                         tp->tcp_header_len =
5539                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5540                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5541                         tcp_store_ts_recent(tp);
5542                 } else {
5543                         tp->tcp_header_len = sizeof(struct tcphdr);
5544                 }
5545 
5546                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5547                         tcp_enable_fack(tp);
5548 
5549                 tcp_mtup_init(sk);
5550                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5551                 tcp_initialize_rcv_mss(sk);
5552 
5553                 /* Remember, tcp_poll() does not lock socket!
5554                  * Change state from SYN-SENT only after copied_seq
5555                  * is initialized. */
5556                 tp->copied_seq = tp->rcv_nxt;
5557 
5558                 smp_mb();
5559 
5560                 tcp_finish_connect(sk, skb);
5561 
5562                 if ((tp->syn_fastopen || tp->syn_data) &&
5563                     tcp_rcv_fastopen_synack(sk, skb, &foc))
5564                         return -1;
5565 
5566                 if (sk->sk_write_pending ||
5567                     icsk->icsk_accept_queue.rskq_defer_accept ||
5568                     icsk->icsk_ack.pingpong) {
5569                         /* Save one ACK. Data will be ready after
5570                          * several ticks, if write_pending is set.
5571                          *
5572                          * It may be deleted, but with this feature tcpdumps
5573                          * look so _wonderfully_ clever, that I was not able
5574                          * to stand against the temptation 8)     --ANK
5575                          */
5576                         inet_csk_schedule_ack(sk);
5577                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5578                         tcp_enter_quickack_mode(sk);
5579                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5580                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5581 
5582 discard:
5583                         __kfree_skb(skb);
5584                         return 0;
5585                 } else {
5586                         tcp_send_ack(sk);
5587                 }
5588                 return -1;
5589         }
5590 
5591         /* No ACK in the segment */
5592 
5593         if (th->rst) {
5594                 /* rfc793:
5595                  * "If the RST bit is set
5596                  *
5597                  *      Otherwise (no ACK) drop the segment and return."
5598                  */
5599 
5600                 goto discard_and_undo;
5601         }
5602 
5603         /* PAWS check. */
5604         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5605             tcp_paws_reject(&tp->rx_opt, 0))
5606                 goto discard_and_undo;
5607 
5608         if (th->syn) {
5609                 /* We see SYN without ACK. It is attempt of
5610                  * simultaneous connect with crossed SYNs.
5611                  * Particularly, it can be connect to self.
5612                  */
5613                 tcp_set_state(sk, TCP_SYN_RECV);
5614 
5615                 if (tp->rx_opt.saw_tstamp) {
5616                         tp->rx_opt.tstamp_ok = 1;
5617                         tcp_store_ts_recent(tp);
5618                         tp->tcp_header_len =
5619                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5620                 } else {
5621                         tp->tcp_header_len = sizeof(struct tcphdr);
5622                 }
5623 
5624                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5625                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5626 
5627                 /* RFC1323: The window in SYN & SYN/ACK segments is
5628                  * never scaled.
5629                  */
5630                 tp->snd_wnd    = ntohs(th->window);
5631                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5632                 tp->max_window = tp->snd_wnd;
5633 
5634                 tcp_ecn_rcv_syn(tp, th);
5635 
5636                 tcp_mtup_init(sk);
5637                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5638                 tcp_initialize_rcv_mss(sk);
5639 
5640                 tcp_send_synack(sk);
5641 #if 0
5642                 /* Note, we could accept data and URG from this segment.
5643                  * There are no obstacles to make this (except that we must
5644                  * either change tcp_recvmsg() to prevent it from returning data
5645                  * before 3WHS completes per RFC793, or employ TCP Fast Open).
5646                  *
5647                  * However, if we ignore data in ACKless segments sometimes,
5648                  * we have no reasons to accept it sometimes.
5649                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5650                  * is not flawless. So, discard packet for sanity.
5651                  * Uncomment this return to process the data.
5652                  */
5653                 return -1;
5654 #else
5655                 goto discard;
5656 #endif
5657         }
5658         /* "fifth, if neither of the SYN or RST bits is set then
5659          * drop the segment and return."
5660          */
5661 
5662 discard_and_undo:
5663         tcp_clear_options(&tp->rx_opt);
5664         tp->rx_opt.mss_clamp = saved_clamp;
5665         goto discard;
5666 
5667 reset_and_undo:
5668         tcp_clear_options(&tp->rx_opt);
5669         tp->rx_opt.mss_clamp = saved_clamp;
5670         return 1;
5671 }
5672 
5673 /*
5674  *      This function implements the receiving procedure of RFC 793 for
5675  *      all states except ESTABLISHED and TIME_WAIT.
5676  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5677  *      address independent.
5678  */
5679 
5680 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5681                           const struct tcphdr *th, unsigned int len)
5682 {
5683         struct tcp_sock *tp = tcp_sk(sk);
5684         struct inet_connection_sock *icsk = inet_csk(sk);
5685         struct request_sock *req;
5686         int queued = 0;
5687         bool acceptable;
5688         u32 synack_stamp;
5689 
5690         tp->rx_opt.saw_tstamp = 0;
5691 
5692         switch (sk->sk_state) {
5693         case TCP_CLOSE:
5694                 goto discard;
5695 
5696         case TCP_LISTEN:
5697                 if (th->ack)
5698                         return 1;
5699 
5700                 if (th->rst)
5701                         goto discard;
5702 
5703                 if (th->syn) {
5704                         if (th->fin)
5705                                 goto discard;
5706                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5707                                 return 1;
5708 
5709                         /* Now we have several options: In theory there is
5710                          * nothing else in the frame. KA9Q has an option to
5711                          * send data with the syn, BSD accepts data with the
5712                          * syn up to the [to be] advertised window and
5713                          * Solaris 2.1 gives you a protocol error. For now
5714                          * we just ignore it, that fits the spec precisely
5715                          * and avoids incompatibilities. It would be nice in
5716                          * future to drop through and process the data.
5717                          *
5718                          * Now that TTCP is starting to be used we ought to
5719                          * queue this data.
5720                          * But, this leaves one open to an easy denial of
5721                          * service attack, and SYN cookies can't defend
5722                          * against this problem. So, we drop the data
5723                          * in the interest of security over speed unless
5724                          * it's still in use.
5725                          */
5726                         kfree_skb(skb);
5727                         return 0;
5728                 }
5729                 goto discard;
5730 
5731         case TCP_SYN_SENT:
5732                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5733                 if (queued >= 0)
5734                         return queued;
5735 
5736                 /* Do step6 onward by hand. */
5737                 tcp_urg(sk, skb, th);
5738                 __kfree_skb(skb);
5739                 tcp_data_snd_check(sk);
5740                 return 0;
5741         }
5742 
5743         req = tp->fastopen_rsk;
5744         if (req) {
5745                 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5746                     sk->sk_state != TCP_FIN_WAIT1);
5747 
5748                 if (!tcp_check_req(sk, skb, req, true))
5749                         goto discard;
5750         }
5751 
5752         if (!th->ack && !th->rst && !th->syn)
5753                 goto discard;
5754 
5755         if (!tcp_validate_incoming(sk, skb, th, 0))
5756                 return 0;
5757 
5758         /* step 5: check the ACK field */
5759         acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5760                                       FLAG_UPDATE_TS_RECENT) > 0;
5761 
5762         switch (sk->sk_state) {
5763         case TCP_SYN_RECV:
5764                 if (!acceptable)
5765                         return 1;
5766 
5767                 /* Once we leave TCP_SYN_RECV, we no longer need req
5768                  * so release it.
5769                  */
5770                 if (req) {
5771                         synack_stamp = tcp_rsk(req)->snt_synack;
5772                         tp->total_retrans = req->num_retrans;
5773                         reqsk_fastopen_remove(sk, req, false);
5774                 } else {
5775                         synack_stamp = tp->lsndtime;
5776                         /* Make sure socket is routed, for correct metrics. */
5777                         icsk->icsk_af_ops->rebuild_header(sk);
5778                         tcp_init_congestion_control(sk);
5779 
5780                         tcp_mtup_init(sk);
5781                         tp->copied_seq = tp->rcv_nxt;
5782                         tcp_init_buffer_space(sk);
5783                 }
5784                 smp_mb();
5785                 tcp_set_state(sk, TCP_ESTABLISHED);
5786                 sk->sk_state_change(sk);
5787 
5788                 /* Note, that this wakeup is only for marginal crossed SYN case.
5789                  * Passively open sockets are not waked up, because
5790                  * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5791                  */
5792                 if (sk->sk_socket)
5793                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5794 
5795                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5796                 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5797                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5798                 tcp_synack_rtt_meas(sk, synack_stamp);
5799 
5800                 if (tp->rx_opt.tstamp_ok)
5801                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5802 
5803                 if (req) {
5804                         /* Re-arm the timer because data may have been sent out.
5805                          * This is similar to the regular data transmission case
5806                          * when new data has just been ack'ed.
5807                          *
5808                          * (TFO) - we could try to be more aggressive and
5809                          * retransmitting any data sooner based on when they
5810                          * are sent out.
5811                          */
5812                         tcp_rearm_rto(sk);
5813                 } else
5814                         tcp_init_metrics(sk);
5815 
5816                 tcp_update_pacing_rate(sk);
5817 
5818                 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5819                 tp->lsndtime = tcp_time_stamp;
5820 
5821                 tcp_initialize_rcv_mss(sk);
5822                 tcp_fast_path_on(tp);
5823                 break;
5824 
5825         case TCP_FIN_WAIT1: {
5826                 struct dst_entry *dst;
5827                 int tmo;
5828 
5829                 /* If we enter the TCP_FIN_WAIT1 state and we are a
5830                  * Fast Open socket and this is the first acceptable
5831                  * ACK we have received, this would have acknowledged
5832                  * our SYNACK so stop the SYNACK timer.
5833                  */
5834                 if (req) {
5835                         /* Return RST if ack_seq is invalid.
5836                          * Note that RFC793 only says to generate a
5837                          * DUPACK for it but for TCP Fast Open it seems
5838                          * better to treat this case like TCP_SYN_RECV
5839                          * above.
5840                          */
5841                         if (!acceptable)
5842                                 return 1;
5843                         /* We no longer need the request sock. */
5844                         reqsk_fastopen_remove(sk, req, false);
5845                         tcp_rearm_rto(sk);
5846                 }
5847                 if (tp->snd_una != tp->write_seq)
5848                         break;
5849 
5850                 tcp_set_state(sk, TCP_FIN_WAIT2);
5851                 sk->sk_shutdown |= SEND_SHUTDOWN;
5852 
5853                 dst = __sk_dst_get(sk);
5854                 if (dst)
5855                         dst_confirm(dst);
5856 
5857                 if (!sock_flag(sk, SOCK_DEAD)) {
5858                         /* Wake up lingering close() */
5859                         sk->sk_state_change(sk);
5860                         break;
5861                 }
5862 
5863                 if (tp->linger2 < 0 ||
5864                     (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5865                      after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5866                         tcp_done(sk);
5867                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5868                         return 1;
5869                 }
5870 
5871                 tmo = tcp_fin_time(sk);
5872                 if (tmo > TCP_TIMEWAIT_LEN) {
5873                         inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5874                 } else if (th->fin || sock_owned_by_user(sk)) {
5875                         /* Bad case. We could lose such FIN otherwise.
5876                          * It is not a big problem, but it looks confusing
5877                          * and not so rare event. We still can lose it now,
5878                          * if it spins in bh_lock_sock(), but it is really
5879                          * marginal case.
5880                          */
5881                         inet_csk_reset_keepalive_timer(sk, tmo);
5882                 } else {
5883                         tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5884                         goto discard;
5885                 }
5886                 break;
5887         }
5888 
5889         case TCP_CLOSING:
5890                 if (tp->snd_una == tp->write_seq) {
5891                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5892                         goto discard;
5893                 }
5894                 break;
5895 
5896         case TCP_LAST_ACK:
5897                 if (tp->snd_una == tp->write_seq) {
5898                         tcp_update_metrics(sk);
5899                         tcp_done(sk);
5900                         goto discard;
5901                 }
5902                 break;
5903         }
5904 
5905         /* step 6: check the URG bit */
5906         tcp_urg(sk, skb, th);
5907 
5908         /* step 7: process the segment text */
5909         switch (sk->sk_state) {
5910         case TCP_CLOSE_WAIT:
5911         case TCP_CLOSING:
5912         case TCP_LAST_ACK:
5913                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5914                         break;
5915         case TCP_FIN_WAIT1:
5916         case TCP_FIN_WAIT2:
5917                 /* RFC 793 says to queue data in these states,
5918                  * RFC 1122 says we MUST send a reset.
5919                  * BSD 4.4 also does reset.
5920                  */
5921                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5922                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5923                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5924                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5925                                 tcp_reset(sk);
5926                                 return 1;
5927                         }
5928                 }
5929                 /* Fall through */
5930         case TCP_ESTABLISHED:
5931                 tcp_data_queue(sk, skb);
5932                 queued = 1;
5933                 break;
5934         }
5935 
5936         /* tcp_data could move socket to TIME-WAIT */
5937         if (sk->sk_state != TCP_CLOSE) {
5938                 tcp_data_snd_check(sk);
5939                 tcp_ack_snd_check(sk);
5940         }
5941 
5942         if (!queued) {
5943 discard:
5944                 __kfree_skb(skb);
5945         }
5946         return 0;
5947 }
5948 EXPORT_SYMBOL(tcp_rcv_state_process);
5949 
5950 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
5951 {
5952         struct inet_request_sock *ireq = inet_rsk(req);
5953 
5954         if (family == AF_INET)
5955                 net_dbg_ratelimited("drop open request from %pI4/%u\n",
5956                                     &ireq->ir_rmt_addr, port);
5957 #if IS_ENABLED(CONFIG_IPV6)
5958         else if (family == AF_INET6)
5959                 net_dbg_ratelimited("drop open request from %pI6/%u\n",
5960                                     &ireq->ir_v6_rmt_addr, port);
5961 #endif
5962 }
5963 
5964 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
5965  *
5966  * If we receive a SYN packet with these bits set, it means a
5967  * network is playing bad games with TOS bits. In order to
5968  * avoid possible false congestion notifications, we disable
5969  * TCP ECN negotiation.
5970  *
5971  * Exception: tcp_ca wants ECN. This is required for DCTCP
5972  * congestion control: Linux DCTCP asserts ECT on all packets,
5973  * including SYN, which is most optimal solution; however,
5974  * others, such as FreeBSD do not.
5975  */
5976 static void tcp_ecn_create_request(struct request_sock *req,
5977                                    const struct sk_buff *skb,
5978                                    const struct sock *listen_sk,
5979                                    const struct dst_entry *dst)
5980 {
5981         const struct tcphdr *th = tcp_hdr(skb);
5982         const struct net *net = sock_net(listen_sk);
5983         bool th_ecn = th->ece && th->cwr;
5984         bool ect, ecn_ok;
5985 
5986         if (!th_ecn)
5987                 return;
5988 
5989         ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
5990         ecn_ok = net->ipv4.sysctl_tcp_ecn || dst_feature(dst, RTAX_FEATURE_ECN);
5991 
5992         if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk))
5993                 inet_rsk(req)->ecn_ok = 1;
5994 }
5995 
5996 static void tcp_openreq_init(struct request_sock *req,
5997                              const struct tcp_options_received *rx_opt,
5998                              struct sk_buff *skb, const struct sock *sk)
5999 {
6000         struct inet_request_sock *ireq = inet_rsk(req);
6001 
6002         req->rcv_wnd = 0;               /* So that tcp_send_synack() knows! */
6003         req->cookie_ts = 0;
6004         tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6005         tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6006         tcp_rsk(req)->snt_synack = tcp_time_stamp;
6007         tcp_rsk(req)->last_oow_ack_time = 0;
6008         req->mss = rx_opt->mss_clamp;
6009         req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6010         ireq->tstamp_ok = rx_opt->tstamp_ok;
6011         ireq->sack_ok = rx_opt->sack_ok;
6012         ireq->snd_wscale = rx_opt->snd_wscale;
6013         ireq->wscale_ok = rx_opt->wscale_ok;
6014         ireq->acked = 0;
6015         ireq->ecn_ok = 0;
6016         ireq->ir_rmt_port = tcp_hdr(skb)->source;
6017         ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6018         ireq->ir_mark = inet_request_mark(sk, skb);
6019 }
6020 
6021 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6022                                       struct sock *sk_listener)
6023 {
6024         struct request_sock *req = reqsk_alloc(ops, sk_listener);
6025 
6026         if (req) {
6027                 struct inet_request_sock *ireq = inet_rsk(req);
6028 
6029                 kmemcheck_annotate_bitfield(ireq, flags);
6030                 ireq->opt = NULL;
6031                 atomic64_set(&ireq->ir_cookie, 0);
6032                 ireq->ireq_state = TCP_NEW_SYN_RECV;
6033                 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6034                 ireq->ireq_family = sk_listener->sk_family;
6035         }
6036 
6037         return req;
6038 }
6039 EXPORT_SYMBOL(inet_reqsk_alloc);
6040 
6041 /*
6042  * Return true if a syncookie should be sent
6043  */
6044 static bool tcp_syn_flood_action(struct sock *sk,
6045                                  const struct sk_buff *skb,
6046                                  const char *proto)
6047 {
6048         const char *msg = "Dropping request";
6049         bool want_cookie = false;
6050         struct listen_sock *lopt;
6051 
6052 #ifdef CONFIG_SYN_COOKIES
6053         if (sysctl_tcp_syncookies) {
6054                 msg = "Sending cookies";
6055                 want_cookie = true;
6056                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6057         } else
6058 #endif
6059                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6060 
6061         lopt = inet_csk(sk)->icsk_accept_queue.listen_opt;
6062         if (!lopt->synflood_warned && sysctl_tcp_syncookies != 2) {
6063                 lopt->synflood_warned = 1;
6064                 pr_info("%s: Possible SYN flooding on port %d. %s.  Check SNMP counters.\n",
6065                         proto, ntohs(tcp_hdr(skb)->dest), msg);
6066         }
6067         return want_cookie;
6068 }
6069 
6070 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6071                      const struct tcp_request_sock_ops *af_ops,
6072                      struct sock *sk, struct sk_buff *skb)
6073 {
6074         struct tcp_options_received tmp_opt;
6075         struct request_sock *req;
6076         struct tcp_sock *tp = tcp_sk(sk);
6077         struct dst_entry *dst = NULL;
6078         __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6079         bool want_cookie = false, fastopen;
6080         struct flowi fl;
6081         struct tcp_fastopen_cookie foc = { .len = -1 };
6082         int err;
6083 
6084 
6085         /* TW buckets are converted to open requests without
6086          * limitations, they conserve resources and peer is
6087          * evidently real one.
6088          */
6089         if ((sysctl_tcp_syncookies == 2 ||
6090              inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6091                 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6092                 if (!want_cookie)
6093                         goto drop;
6094         }
6095 
6096 
6097         /* Accept backlog is full. If we have already queued enough
6098          * of warm entries in syn queue, drop request. It is better than
6099          * clogging syn queue with openreqs with exponentially increasing
6100          * timeout.
6101          */
6102         if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
6103                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6104                 goto drop;
6105         }
6106 
6107         req = inet_reqsk_alloc(rsk_ops, sk);
6108         if (!req)
6109                 goto drop;
6110 
6111         tcp_rsk(req)->af_specific = af_ops;
6112 
6113         tcp_clear_options(&tmp_opt);
6114         tmp_opt.mss_clamp = af_ops->mss_clamp;
6115         tmp_opt.user_mss  = tp->rx_opt.user_mss;
6116         tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6117 
6118         if (want_cookie && !tmp_opt.saw_tstamp)
6119                 tcp_clear_options(&tmp_opt);
6120 
6121         tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6122         tcp_openreq_init(req, &tmp_opt, skb, sk);
6123 
6124         /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6125         inet_rsk(req)->ir_iif = sk->sk_bound_dev_if;
6126 
6127         af_ops->init_req(req, sk, skb);
6128 
6129         if (security_inet_conn_request(sk, skb, req))
6130                 goto drop_and_free;
6131 
6132         if (!want_cookie && !isn) {
6133                 /* VJ's idea. We save last timestamp seen
6134                  * from the destination in peer table, when entering
6135                  * state TIME-WAIT, and check against it before
6136                  * accepting new connection request.
6137                  *
6138                  * If "isn" is not zero, this request hit alive
6139                  * timewait bucket, so that all the necessary checks
6140                  * are made in the function processing timewait state.
6141                  */
6142                 if (tcp_death_row.sysctl_tw_recycle) {
6143                         bool strict;
6144 
6145                         dst = af_ops->route_req(sk, &fl, req, &strict);
6146 
6147                         if (dst && strict &&
6148                             !tcp_peer_is_proven(req, dst, true,
6149                                                 tmp_opt.saw_tstamp)) {
6150                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
6151                                 goto drop_and_release;
6152                         }
6153                 }
6154                 /* Kill the following clause, if you dislike this way. */
6155                 else if (!sysctl_tcp_syncookies &&
6156                          (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6157                           (sysctl_max_syn_backlog >> 2)) &&
6158                          !tcp_peer_is_proven(req, dst, false,
6159                                              tmp_opt.saw_tstamp)) {
6160                         /* Without syncookies last quarter of
6161                          * backlog is filled with destinations,
6162                          * proven to be alive.
6163                          * It means that we continue to communicate
6164                          * to destinations, already remembered
6165                          * to the moment of synflood.
6166                          */
6167                         pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6168                                     rsk_ops->family);
6169                         goto drop_and_release;
6170                 }
6171 
6172                 isn = af_ops->init_seq(skb);
6173         }
6174         if (!dst) {
6175                 dst = af_ops->route_req(sk, &fl, req, NULL);
6176                 if (!dst)
6177                         goto drop_and_free;
6178         }
6179 
6180         tcp_ecn_create_request(req, skb, sk, dst);
6181 
6182         if (want_cookie) {
6183                 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6184                 req->cookie_ts = tmp_opt.tstamp_ok;
6185                 if (!tmp_opt.tstamp_ok)
6186                         inet_rsk(req)->ecn_ok = 0;
6187         }
6188 
6189         tcp_rsk(req)->snt_isn = isn;
6190         tcp_openreq_init_rwin(req, sk, dst);
6191         fastopen = !want_cookie &&
6192                    tcp_try_fastopen(sk, skb, req, &foc, dst);
6193         err = af_ops->send_synack(sk, dst, &fl, req,
6194                                   skb_get_queue_mapping(skb), &foc);
6195         if (!fastopen) {
6196                 if (err || want_cookie)
6197                         goto drop_and_free;
6198 
6199                 tcp_rsk(req)->tfo_listener = false;
6200                 af_ops->queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6201         }
6202 
6203         return 0;
6204 
6205 drop_and_release:
6206         dst_release(dst);
6207 drop_and_free:
6208         reqsk_free(req);
6209 drop:
6210         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS);
6211         return 0;
6212 }
6213 EXPORT_SYMBOL(tcp_conn_request);
6214 

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