Version:  2.0.40 2.2.26 2.4.37 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0

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 == NULL) ||
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 != NULL) &&
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 != NULL) &&
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 != NULL) {
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 == NULL)
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 == NULL)
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 == NULL)
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         tcp_mark_lost_retrans(sk);
1824 
1825         tcp_verify_left_out(tp);
1826 
1827         if ((state.reord < tp->fackets_out) &&
1828             ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1829                 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1830 
1831 out:
1832 
1833 #if FASTRETRANS_DEBUG > 0
1834         WARN_ON((int)tp->sacked_out < 0);
1835         WARN_ON((int)tp->lost_out < 0);
1836         WARN_ON((int)tp->retrans_out < 0);
1837         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1838 #endif
1839         *sack_rtt_us = state.rtt_us;
1840         return state.flag;
1841 }
1842 
1843 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1844  * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1845  */
1846 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1847 {
1848         u32 holes;
1849 
1850         holes = max(tp->lost_out, 1U);
1851         holes = min(holes, tp->packets_out);
1852 
1853         if ((tp->sacked_out + holes) > tp->packets_out) {
1854                 tp->sacked_out = tp->packets_out - holes;
1855                 return true;
1856         }
1857         return false;
1858 }
1859 
1860 /* If we receive more dupacks than we expected counting segments
1861  * in assumption of absent reordering, interpret this as reordering.
1862  * The only another reason could be bug in receiver TCP.
1863  */
1864 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1865 {
1866         struct tcp_sock *tp = tcp_sk(sk);
1867         if (tcp_limit_reno_sacked(tp))
1868                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1869 }
1870 
1871 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1872 
1873 static void tcp_add_reno_sack(struct sock *sk)
1874 {
1875         struct tcp_sock *tp = tcp_sk(sk);
1876         tp->sacked_out++;
1877         tcp_check_reno_reordering(sk, 0);
1878         tcp_verify_left_out(tp);
1879 }
1880 
1881 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1882 
1883 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1884 {
1885         struct tcp_sock *tp = tcp_sk(sk);
1886 
1887         if (acked > 0) {
1888                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1889                 if (acked - 1 >= tp->sacked_out)
1890                         tp->sacked_out = 0;
1891                 else
1892                         tp->sacked_out -= acked - 1;
1893         }
1894         tcp_check_reno_reordering(sk, acked);
1895         tcp_verify_left_out(tp);
1896 }
1897 
1898 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1899 {
1900         tp->sacked_out = 0;
1901 }
1902 
1903 void tcp_clear_retrans(struct tcp_sock *tp)
1904 {
1905         tp->retrans_out = 0;
1906         tp->lost_out = 0;
1907         tp->undo_marker = 0;
1908         tp->undo_retrans = -1;
1909         tp->fackets_out = 0;
1910         tp->sacked_out = 0;
1911 }
1912 
1913 static inline void tcp_init_undo(struct tcp_sock *tp)
1914 {
1915         tp->undo_marker = tp->snd_una;
1916         /* Retransmission still in flight may cause DSACKs later. */
1917         tp->undo_retrans = tp->retrans_out ? : -1;
1918 }
1919 
1920 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1921  * and reset tags completely, otherwise preserve SACKs. If receiver
1922  * dropped its ofo queue, we will know this due to reneging detection.
1923  */
1924 void tcp_enter_loss(struct sock *sk)
1925 {
1926         const struct inet_connection_sock *icsk = inet_csk(sk);
1927         struct tcp_sock *tp = tcp_sk(sk);
1928         struct sk_buff *skb;
1929         bool new_recovery = false;
1930         bool is_reneg;                  /* is receiver reneging on SACKs? */
1931 
1932         /* Reduce ssthresh if it has not yet been made inside this window. */
1933         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1934             !after(tp->high_seq, tp->snd_una) ||
1935             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1936                 new_recovery = true;
1937                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1938                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1939                 tcp_ca_event(sk, CA_EVENT_LOSS);
1940                 tcp_init_undo(tp);
1941         }
1942         tp->snd_cwnd       = 1;
1943         tp->snd_cwnd_cnt   = 0;
1944         tp->snd_cwnd_stamp = tcp_time_stamp;
1945 
1946         tp->retrans_out = 0;
1947         tp->lost_out = 0;
1948 
1949         if (tcp_is_reno(tp))
1950                 tcp_reset_reno_sack(tp);
1951 
1952         skb = tcp_write_queue_head(sk);
1953         is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1954         if (is_reneg) {
1955                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1956                 tp->sacked_out = 0;
1957                 tp->fackets_out = 0;
1958         }
1959         tcp_clear_all_retrans_hints(tp);
1960 
1961         tcp_for_write_queue(skb, sk) {
1962                 if (skb == tcp_send_head(sk))
1963                         break;
1964 
1965                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1966                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
1967                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1968                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1969                         tp->lost_out += tcp_skb_pcount(skb);
1970                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1971                 }
1972         }
1973         tcp_verify_left_out(tp);
1974 
1975         /* Timeout in disordered state after receiving substantial DUPACKs
1976          * suggests that the degree of reordering is over-estimated.
1977          */
1978         if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1979             tp->sacked_out >= sysctl_tcp_reordering)
1980                 tp->reordering = min_t(unsigned int, tp->reordering,
1981                                        sysctl_tcp_reordering);
1982         tcp_set_ca_state(sk, TCP_CA_Loss);
1983         tp->high_seq = tp->snd_nxt;
1984         tcp_ecn_queue_cwr(tp);
1985 
1986         /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1987          * loss recovery is underway except recurring timeout(s) on
1988          * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1989          */
1990         tp->frto = sysctl_tcp_frto &&
1991                    (new_recovery || icsk->icsk_retransmits) &&
1992                    !inet_csk(sk)->icsk_mtup.probe_size;
1993 }
1994 
1995 /* If ACK arrived pointing to a remembered SACK, it means that our
1996  * remembered SACKs do not reflect real state of receiver i.e.
1997  * receiver _host_ is heavily congested (or buggy).
1998  *
1999  * To avoid big spurious retransmission bursts due to transient SACK
2000  * scoreboard oddities that look like reneging, we give the receiver a
2001  * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2002  * restore sanity to the SACK scoreboard. If the apparent reneging
2003  * persists until this RTO then we'll clear the SACK scoreboard.
2004  */
2005 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2006 {
2007         if (flag & FLAG_SACK_RENEGING) {
2008                 struct tcp_sock *tp = tcp_sk(sk);
2009                 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2010                                           msecs_to_jiffies(10));
2011 
2012                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2013                                           delay, TCP_RTO_MAX);
2014                 return true;
2015         }
2016         return false;
2017 }
2018 
2019 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2020 {
2021         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2022 }
2023 
2024 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2025  * counter when SACK is enabled (without SACK, sacked_out is used for
2026  * that purpose).
2027  *
2028  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2029  * segments up to the highest received SACK block so far and holes in
2030  * between them.
2031  *
2032  * With reordering, holes may still be in flight, so RFC3517 recovery
2033  * uses pure sacked_out (total number of SACKed segments) even though
2034  * it violates the RFC that uses duplicate ACKs, often these are equal
2035  * but when e.g. out-of-window ACKs or packet duplication occurs,
2036  * they differ. Since neither occurs due to loss, TCP should really
2037  * ignore them.
2038  */
2039 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2040 {
2041         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2042 }
2043 
2044 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2045 {
2046         struct tcp_sock *tp = tcp_sk(sk);
2047         unsigned long delay;
2048 
2049         /* Delay early retransmit and entering fast recovery for
2050          * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2051          * available, or RTO is scheduled to fire first.
2052          */
2053         if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2054             (flag & FLAG_ECE) || !tp->srtt_us)
2055                 return false;
2056 
2057         delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2058                     msecs_to_jiffies(2));
2059 
2060         if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2061                 return false;
2062 
2063         inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2064                                   TCP_RTO_MAX);
2065         return true;
2066 }
2067 
2068 /* Linux NewReno/SACK/FACK/ECN state machine.
2069  * --------------------------------------
2070  *
2071  * "Open"       Normal state, no dubious events, fast path.
2072  * "Disorder"   In all the respects it is "Open",
2073  *              but requires a bit more attention. It is entered when
2074  *              we see some SACKs or dupacks. It is split of "Open"
2075  *              mainly to move some processing from fast path to slow one.
2076  * "CWR"        CWND was reduced due to some Congestion Notification event.
2077  *              It can be ECN, ICMP source quench, local device congestion.
2078  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2079  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2080  *
2081  * tcp_fastretrans_alert() is entered:
2082  * - each incoming ACK, if state is not "Open"
2083  * - when arrived ACK is unusual, namely:
2084  *      * SACK
2085  *      * Duplicate ACK.
2086  *      * ECN ECE.
2087  *
2088  * Counting packets in flight is pretty simple.
2089  *
2090  *      in_flight = packets_out - left_out + retrans_out
2091  *
2092  *      packets_out is SND.NXT-SND.UNA counted in packets.
2093  *
2094  *      retrans_out is number of retransmitted segments.
2095  *
2096  *      left_out is number of segments left network, but not ACKed yet.
2097  *
2098  *              left_out = sacked_out + lost_out
2099  *
2100  *     sacked_out: Packets, which arrived to receiver out of order
2101  *                 and hence not ACKed. With SACKs this number is simply
2102  *                 amount of SACKed data. Even without SACKs
2103  *                 it is easy to give pretty reliable estimate of this number,
2104  *                 counting duplicate ACKs.
2105  *
2106  *       lost_out: Packets lost by network. TCP has no explicit
2107  *                 "loss notification" feedback from network (for now).
2108  *                 It means that this number can be only _guessed_.
2109  *                 Actually, it is the heuristics to predict lossage that
2110  *                 distinguishes different algorithms.
2111  *
2112  *      F.e. after RTO, when all the queue is considered as lost,
2113  *      lost_out = packets_out and in_flight = retrans_out.
2114  *
2115  *              Essentially, we have now two algorithms counting
2116  *              lost packets.
2117  *
2118  *              FACK: It is the simplest heuristics. As soon as we decided
2119  *              that something is lost, we decide that _all_ not SACKed
2120  *              packets until the most forward SACK are lost. I.e.
2121  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2122  *              It is absolutely correct estimate, if network does not reorder
2123  *              packets. And it loses any connection to reality when reordering
2124  *              takes place. We use FACK by default until reordering
2125  *              is suspected on the path to this destination.
2126  *
2127  *              NewReno: when Recovery is entered, we assume that one segment
2128  *              is lost (classic Reno). While we are in Recovery and
2129  *              a partial ACK arrives, we assume that one more packet
2130  *              is lost (NewReno). This heuristics are the same in NewReno
2131  *              and SACK.
2132  *
2133  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2134  *  deflation etc. CWND is real congestion window, never inflated, changes
2135  *  only according to classic VJ rules.
2136  *
2137  * Really tricky (and requiring careful tuning) part of algorithm
2138  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2139  * The first determines the moment _when_ we should reduce CWND and,
2140  * hence, slow down forward transmission. In fact, it determines the moment
2141  * when we decide that hole is caused by loss, rather than by a reorder.
2142  *
2143  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2144  * holes, caused by lost packets.
2145  *
2146  * And the most logically complicated part of algorithm is undo
2147  * heuristics. We detect false retransmits due to both too early
2148  * fast retransmit (reordering) and underestimated RTO, analyzing
2149  * timestamps and D-SACKs. When we detect that some segments were
2150  * retransmitted by mistake and CWND reduction was wrong, we undo
2151  * window reduction and abort recovery phase. This logic is hidden
2152  * inside several functions named tcp_try_undo_<something>.
2153  */
2154 
2155 /* This function decides, when we should leave Disordered state
2156  * and enter Recovery phase, reducing congestion window.
2157  *
2158  * Main question: may we further continue forward transmission
2159  * with the same cwnd?
2160  */
2161 static bool tcp_time_to_recover(struct sock *sk, int flag)
2162 {
2163         struct tcp_sock *tp = tcp_sk(sk);
2164         __u32 packets_out;
2165 
2166         /* Trick#1: The loss is proven. */
2167         if (tp->lost_out)
2168                 return true;
2169 
2170         /* Not-A-Trick#2 : Classic rule... */
2171         if (tcp_dupack_heuristics(tp) > tp->reordering)
2172                 return true;
2173 
2174         /* Trick#4: It is still not OK... But will it be useful to delay
2175          * recovery more?
2176          */
2177         packets_out = tp->packets_out;
2178         if (packets_out <= tp->reordering &&
2179             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2180             !tcp_may_send_now(sk)) {
2181                 /* We have nothing to send. This connection is limited
2182                  * either by receiver window or by application.
2183                  */
2184                 return true;
2185         }
2186 
2187         /* If a thin stream is detected, retransmit after first
2188          * received dupack. Employ only if SACK is supported in order
2189          * to avoid possible corner-case series of spurious retransmissions
2190          * Use only if there are no unsent data.
2191          */
2192         if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2193             tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2194             tcp_is_sack(tp) && !tcp_send_head(sk))
2195                 return true;
2196 
2197         /* Trick#6: TCP early retransmit, per RFC5827.  To avoid spurious
2198          * retransmissions due to small network reorderings, we implement
2199          * Mitigation A.3 in the RFC and delay the retransmission for a short
2200          * interval if appropriate.
2201          */
2202         if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2203             (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2204             !tcp_may_send_now(sk))
2205                 return !tcp_pause_early_retransmit(sk, flag);
2206 
2207         return false;
2208 }
2209 
2210 /* Detect loss in event "A" above by marking head of queue up as lost.
2211  * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2212  * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2213  * has at least tp->reordering SACKed seqments above it; "packets" refers to
2214  * the maximum SACKed segments to pass before reaching this limit.
2215  */
2216 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2217 {
2218         struct tcp_sock *tp = tcp_sk(sk);
2219         struct sk_buff *skb;
2220         int cnt, oldcnt;
2221         int err;
2222         unsigned int mss;
2223         /* Use SACK to deduce losses of new sequences sent during recovery */
2224         const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2225 
2226         WARN_ON(packets > tp->packets_out);
2227         if (tp->lost_skb_hint) {
2228                 skb = tp->lost_skb_hint;
2229                 cnt = tp->lost_cnt_hint;
2230                 /* Head already handled? */
2231                 if (mark_head && skb != tcp_write_queue_head(sk))
2232                         return;
2233         } else {
2234                 skb = tcp_write_queue_head(sk);
2235                 cnt = 0;
2236         }
2237 
2238         tcp_for_write_queue_from(skb, sk) {
2239                 if (skb == tcp_send_head(sk))
2240                         break;
2241                 /* TODO: do this better */
2242                 /* this is not the most efficient way to do this... */
2243                 tp->lost_skb_hint = skb;
2244                 tp->lost_cnt_hint = cnt;
2245 
2246                 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2247                         break;
2248 
2249                 oldcnt = cnt;
2250                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2251                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2252                         cnt += tcp_skb_pcount(skb);
2253 
2254                 if (cnt > packets) {
2255                         if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2256                             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2257                             (oldcnt >= packets))
2258                                 break;
2259 
2260                         mss = skb_shinfo(skb)->gso_size;
2261                         err = tcp_fragment(sk, skb, (packets - oldcnt) * mss,
2262                                            mss, GFP_ATOMIC);
2263                         if (err < 0)
2264                                 break;
2265                         cnt = packets;
2266                 }
2267 
2268                 tcp_skb_mark_lost(tp, skb);
2269 
2270                 if (mark_head)
2271                         break;
2272         }
2273         tcp_verify_left_out(tp);
2274 }
2275 
2276 /* Account newly detected lost packet(s) */
2277 
2278 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2279 {
2280         struct tcp_sock *tp = tcp_sk(sk);
2281 
2282         if (tcp_is_reno(tp)) {
2283                 tcp_mark_head_lost(sk, 1, 1);
2284         } else if (tcp_is_fack(tp)) {
2285                 int lost = tp->fackets_out - tp->reordering;
2286                 if (lost <= 0)
2287                         lost = 1;
2288                 tcp_mark_head_lost(sk, lost, 0);
2289         } else {
2290                 int sacked_upto = tp->sacked_out - tp->reordering;
2291                 if (sacked_upto >= 0)
2292                         tcp_mark_head_lost(sk, sacked_upto, 0);
2293                 else if (fast_rexmit)
2294                         tcp_mark_head_lost(sk, 1, 1);
2295         }
2296 }
2297 
2298 /* CWND moderation, preventing bursts due to too big ACKs
2299  * in dubious situations.
2300  */
2301 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2302 {
2303         tp->snd_cwnd = min(tp->snd_cwnd,
2304                            tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2305         tp->snd_cwnd_stamp = tcp_time_stamp;
2306 }
2307 
2308 /* Nothing was retransmitted or returned timestamp is less
2309  * than timestamp of the first retransmission.
2310  */
2311 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2312 {
2313         return !tp->retrans_stamp ||
2314                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2315                  before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2316 }
2317 
2318 /* Undo procedures. */
2319 
2320 /* We can clear retrans_stamp when there are no retransmissions in the
2321  * window. It would seem that it is trivially available for us in
2322  * tp->retrans_out, however, that kind of assumptions doesn't consider
2323  * what will happen if errors occur when sending retransmission for the
2324  * second time. ...It could the that such segment has only
2325  * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2326  * the head skb is enough except for some reneging corner cases that
2327  * are not worth the effort.
2328  *
2329  * Main reason for all this complexity is the fact that connection dying
2330  * time now depends on the validity of the retrans_stamp, in particular,
2331  * that successive retransmissions of a segment must not advance
2332  * retrans_stamp under any conditions.
2333  */
2334 static bool tcp_any_retrans_done(const struct sock *sk)
2335 {
2336         const struct tcp_sock *tp = tcp_sk(sk);
2337         struct sk_buff *skb;
2338 
2339         if (tp->retrans_out)
2340                 return true;
2341 
2342         skb = tcp_write_queue_head(sk);
2343         if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2344                 return true;
2345 
2346         return false;
2347 }
2348 
2349 #if FASTRETRANS_DEBUG > 1
2350 static void DBGUNDO(struct sock *sk, const char *msg)
2351 {
2352         struct tcp_sock *tp = tcp_sk(sk);
2353         struct inet_sock *inet = inet_sk(sk);
2354 
2355         if (sk->sk_family == AF_INET) {
2356                 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2357                          msg,
2358                          &inet->inet_daddr, ntohs(inet->inet_dport),
2359                          tp->snd_cwnd, tcp_left_out(tp),
2360                          tp->snd_ssthresh, tp->prior_ssthresh,
2361                          tp->packets_out);
2362         }
2363 #if IS_ENABLED(CONFIG_IPV6)
2364         else if (sk->sk_family == AF_INET6) {
2365                 struct ipv6_pinfo *np = inet6_sk(sk);
2366                 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2367                          msg,
2368                          &np->daddr, ntohs(inet->inet_dport),
2369                          tp->snd_cwnd, tcp_left_out(tp),
2370                          tp->snd_ssthresh, tp->prior_ssthresh,
2371                          tp->packets_out);
2372         }
2373 #endif
2374 }
2375 #else
2376 #define DBGUNDO(x...) do { } while (0)
2377 #endif
2378 
2379 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2380 {
2381         struct tcp_sock *tp = tcp_sk(sk);
2382 
2383         if (unmark_loss) {
2384                 struct sk_buff *skb;
2385 
2386                 tcp_for_write_queue(skb, sk) {
2387                         if (skb == tcp_send_head(sk))
2388                                 break;
2389                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2390                 }
2391                 tp->lost_out = 0;
2392                 tcp_clear_all_retrans_hints(tp);
2393         }
2394 
2395         if (tp->prior_ssthresh) {
2396                 const struct inet_connection_sock *icsk = inet_csk(sk);
2397 
2398                 if (icsk->icsk_ca_ops->undo_cwnd)
2399                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2400                 else
2401                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2402 
2403                 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2404                         tp->snd_ssthresh = tp->prior_ssthresh;
2405                         tcp_ecn_withdraw_cwr(tp);
2406                 }
2407         } else {
2408                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2409         }
2410         tp->snd_cwnd_stamp = tcp_time_stamp;
2411         tp->undo_marker = 0;
2412 }
2413 
2414 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2415 {
2416         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2417 }
2418 
2419 /* People celebrate: "We love our President!" */
2420 static bool tcp_try_undo_recovery(struct sock *sk)
2421 {
2422         struct tcp_sock *tp = tcp_sk(sk);
2423 
2424         if (tcp_may_undo(tp)) {
2425                 int mib_idx;
2426 
2427                 /* Happy end! We did not retransmit anything
2428                  * or our original transmission succeeded.
2429                  */
2430                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2431                 tcp_undo_cwnd_reduction(sk, false);
2432                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2433                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2434                 else
2435                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2436 
2437                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2438         }
2439         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2440                 /* Hold old state until something *above* high_seq
2441                  * is ACKed. For Reno it is MUST to prevent false
2442                  * fast retransmits (RFC2582). SACK TCP is safe. */
2443                 tcp_moderate_cwnd(tp);
2444                 if (!tcp_any_retrans_done(sk))
2445                         tp->retrans_stamp = 0;
2446                 return true;
2447         }
2448         tcp_set_ca_state(sk, TCP_CA_Open);
2449         return false;
2450 }
2451 
2452 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2453 static bool tcp_try_undo_dsack(struct sock *sk)
2454 {
2455         struct tcp_sock *tp = tcp_sk(sk);
2456 
2457         if (tp->undo_marker && !tp->undo_retrans) {
2458                 DBGUNDO(sk, "D-SACK");
2459                 tcp_undo_cwnd_reduction(sk, false);
2460                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2461                 return true;
2462         }
2463         return false;
2464 }
2465 
2466 /* Undo during loss recovery after partial ACK or using F-RTO. */
2467 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2468 {
2469         struct tcp_sock *tp = tcp_sk(sk);
2470 
2471         if (frto_undo || tcp_may_undo(tp)) {
2472                 tcp_undo_cwnd_reduction(sk, true);
2473 
2474                 DBGUNDO(sk, "partial loss");
2475                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2476                 if (frto_undo)
2477                         NET_INC_STATS_BH(sock_net(sk),
2478                                          LINUX_MIB_TCPSPURIOUSRTOS);
2479                 inet_csk(sk)->icsk_retransmits = 0;
2480                 if (frto_undo || tcp_is_sack(tp))
2481                         tcp_set_ca_state(sk, TCP_CA_Open);
2482                 return true;
2483         }
2484         return false;
2485 }
2486 
2487 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2488  * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2489  * It computes the number of packets to send (sndcnt) based on packets newly
2490  * delivered:
2491  *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2492  *      cwnd reductions across a full RTT.
2493  *   2) If packets in flight is lower than ssthresh (such as due to excess
2494  *      losses and/or application stalls), do not perform any further cwnd
2495  *      reductions, but instead slow start up to ssthresh.
2496  */
2497 static void tcp_init_cwnd_reduction(struct sock *sk)
2498 {
2499         struct tcp_sock *tp = tcp_sk(sk);
2500 
2501         tp->high_seq = tp->snd_nxt;
2502         tp->tlp_high_seq = 0;
2503         tp->snd_cwnd_cnt = 0;
2504         tp->prior_cwnd = tp->snd_cwnd;
2505         tp->prr_delivered = 0;
2506         tp->prr_out = 0;
2507         tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2508         tcp_ecn_queue_cwr(tp);
2509 }
2510 
2511 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2512                                int fast_rexmit)
2513 {
2514         struct tcp_sock *tp = tcp_sk(sk);
2515         int sndcnt = 0;
2516         int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2517         int newly_acked_sacked = prior_unsacked -
2518                                  (tp->packets_out - tp->sacked_out);
2519 
2520         tp->prr_delivered += newly_acked_sacked;
2521         if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2522                 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2523                                tp->prior_cwnd - 1;
2524                 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2525         } else {
2526                 sndcnt = min_t(int, delta,
2527                                max_t(int, tp->prr_delivered - tp->prr_out,
2528                                      newly_acked_sacked) + 1);
2529         }
2530 
2531         sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2532         tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2533 }
2534 
2535 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2536 {
2537         struct tcp_sock *tp = tcp_sk(sk);
2538 
2539         /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2540         if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2541             (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2542                 tp->snd_cwnd = tp->snd_ssthresh;
2543                 tp->snd_cwnd_stamp = tcp_time_stamp;
2544         }
2545         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2546 }
2547 
2548 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2549 void tcp_enter_cwr(struct sock *sk)
2550 {
2551         struct tcp_sock *tp = tcp_sk(sk);
2552 
2553         tp->prior_ssthresh = 0;
2554         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2555                 tp->undo_marker = 0;
2556                 tcp_init_cwnd_reduction(sk);
2557                 tcp_set_ca_state(sk, TCP_CA_CWR);
2558         }
2559 }
2560 
2561 static void tcp_try_keep_open(struct sock *sk)
2562 {
2563         struct tcp_sock *tp = tcp_sk(sk);
2564         int state = TCP_CA_Open;
2565 
2566         if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2567                 state = TCP_CA_Disorder;
2568 
2569         if (inet_csk(sk)->icsk_ca_state != state) {
2570                 tcp_set_ca_state(sk, state);
2571                 tp->high_seq = tp->snd_nxt;
2572         }
2573 }
2574 
2575 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2576 {
2577         struct tcp_sock *tp = tcp_sk(sk);
2578 
2579         tcp_verify_left_out(tp);
2580 
2581         if (!tcp_any_retrans_done(sk))
2582                 tp->retrans_stamp = 0;
2583 
2584         if (flag & FLAG_ECE)
2585                 tcp_enter_cwr(sk);
2586 
2587         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2588                 tcp_try_keep_open(sk);
2589         } else {
2590                 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2591         }
2592 }
2593 
2594 static void tcp_mtup_probe_failed(struct sock *sk)
2595 {
2596         struct inet_connection_sock *icsk = inet_csk(sk);
2597 
2598         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2599         icsk->icsk_mtup.probe_size = 0;
2600 }
2601 
2602 static void tcp_mtup_probe_success(struct sock *sk)
2603 {
2604         struct tcp_sock *tp = tcp_sk(sk);
2605         struct inet_connection_sock *icsk = inet_csk(sk);
2606 
2607         /* FIXME: breaks with very large cwnd */
2608         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2609         tp->snd_cwnd = tp->snd_cwnd *
2610                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2611                        icsk->icsk_mtup.probe_size;
2612         tp->snd_cwnd_cnt = 0;
2613         tp->snd_cwnd_stamp = tcp_time_stamp;
2614         tp->snd_ssthresh = tcp_current_ssthresh(sk);
2615 
2616         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2617         icsk->icsk_mtup.probe_size = 0;
2618         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2619 }
2620 
2621 /* Do a simple retransmit without using the backoff mechanisms in
2622  * tcp_timer. This is used for path mtu discovery.
2623  * The socket is already locked here.
2624  */
2625 void tcp_simple_retransmit(struct sock *sk)
2626 {
2627         const struct inet_connection_sock *icsk = inet_csk(sk);
2628         struct tcp_sock *tp = tcp_sk(sk);
2629         struct sk_buff *skb;
2630         unsigned int mss = tcp_current_mss(sk);
2631         u32 prior_lost = tp->lost_out;
2632 
2633         tcp_for_write_queue(skb, sk) {
2634                 if (skb == tcp_send_head(sk))
2635                         break;
2636                 if (tcp_skb_seglen(skb) > mss &&
2637                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2638                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2639                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2640                                 tp->retrans_out -= tcp_skb_pcount(skb);
2641                         }
2642                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2643                 }
2644         }
2645 
2646         tcp_clear_retrans_hints_partial(tp);
2647 
2648         if (prior_lost == tp->lost_out)
2649                 return;
2650 
2651         if (tcp_is_reno(tp))
2652                 tcp_limit_reno_sacked(tp);
2653 
2654         tcp_verify_left_out(tp);
2655 
2656         /* Don't muck with the congestion window here.
2657          * Reason is that we do not increase amount of _data_
2658          * in network, but units changed and effective
2659          * cwnd/ssthresh really reduced now.
2660          */
2661         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2662                 tp->high_seq = tp->snd_nxt;
2663                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2664                 tp->prior_ssthresh = 0;
2665                 tp->undo_marker = 0;
2666                 tcp_set_ca_state(sk, TCP_CA_Loss);
2667         }
2668         tcp_xmit_retransmit_queue(sk);
2669 }
2670 EXPORT_SYMBOL(tcp_simple_retransmit);
2671 
2672 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2673 {
2674         struct tcp_sock *tp = tcp_sk(sk);
2675         int mib_idx;
2676 
2677         if (tcp_is_reno(tp))
2678                 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2679         else
2680                 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2681 
2682         NET_INC_STATS_BH(sock_net(sk), mib_idx);
2683 
2684         tp->prior_ssthresh = 0;
2685         tcp_init_undo(tp);
2686 
2687         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2688                 if (!ece_ack)
2689                         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2690                 tcp_init_cwnd_reduction(sk);
2691         }
2692         tcp_set_ca_state(sk, TCP_CA_Recovery);
2693 }
2694 
2695 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2696  * recovered or spurious. Otherwise retransmits more on partial ACKs.
2697  */
2698 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2699 {
2700         struct tcp_sock *tp = tcp_sk(sk);
2701         bool recovered = !before(tp->snd_una, tp->high_seq);
2702 
2703         if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2704                 /* Step 3.b. A timeout is spurious if not all data are
2705                  * lost, i.e., never-retransmitted data are (s)acked.
2706                  */
2707                 if (tcp_try_undo_loss(sk, flag & FLAG_ORIG_SACK_ACKED))
2708                         return;
2709 
2710                 if (after(tp->snd_nxt, tp->high_seq) &&
2711                     (flag & FLAG_DATA_SACKED || is_dupack)) {
2712                         tp->frto = 0; /* Loss was real: 2nd part of step 3.a */
2713                 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2714                         tp->high_seq = tp->snd_nxt;
2715                         __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2716                                                   TCP_NAGLE_OFF);
2717                         if (after(tp->snd_nxt, tp->high_seq))
2718                                 return; /* Step 2.b */
2719                         tp->frto = 0;
2720                 }
2721         }
2722 
2723         if (recovered) {
2724                 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2725                 tcp_try_undo_recovery(sk);
2726                 return;
2727         }
2728         if (tcp_is_reno(tp)) {
2729                 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2730                  * delivered. Lower inflight to clock out (re)tranmissions.
2731                  */
2732                 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2733                         tcp_add_reno_sack(sk);
2734                 else if (flag & FLAG_SND_UNA_ADVANCED)
2735                         tcp_reset_reno_sack(tp);
2736         }
2737         if (tcp_try_undo_loss(sk, false))
2738                 return;
2739         tcp_xmit_retransmit_queue(sk);
2740 }
2741 
2742 /* Undo during fast recovery after partial ACK. */
2743 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2744                                  const int prior_unsacked)
2745 {
2746         struct tcp_sock *tp = tcp_sk(sk);
2747 
2748         if (tp->undo_marker && tcp_packet_delayed(tp)) {
2749                 /* Plain luck! Hole if filled with delayed
2750                  * packet, rather than with a retransmit.
2751                  */
2752                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2753 
2754                 /* We are getting evidence that the reordering degree is higher
2755                  * than we realized. If there are no retransmits out then we
2756                  * can undo. Otherwise we clock out new packets but do not
2757                  * mark more packets lost or retransmit more.
2758                  */
2759                 if (tp->retrans_out) {
2760                         tcp_cwnd_reduction(sk, prior_unsacked, 0);
2761                         return true;
2762                 }
2763 
2764                 if (!tcp_any_retrans_done(sk))
2765                         tp->retrans_stamp = 0;
2766 
2767                 DBGUNDO(sk, "partial recovery");
2768                 tcp_undo_cwnd_reduction(sk, true);
2769                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2770                 tcp_try_keep_open(sk);
2771                 return true;
2772         }
2773         return false;
2774 }
2775 
2776 /* Process an event, which can update packets-in-flight not trivially.
2777  * Main goal of this function is to calculate new estimate for left_out,
2778  * taking into account both packets sitting in receiver's buffer and
2779  * packets lost by network.
2780  *
2781  * Besides that it does CWND reduction, when packet loss is detected
2782  * and changes state of machine.
2783  *
2784  * It does _not_ decide what to send, it is made in function
2785  * tcp_xmit_retransmit_queue().
2786  */
2787 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2788                                   const int prior_unsacked,
2789                                   bool is_dupack, int flag)
2790 {
2791         struct inet_connection_sock *icsk = inet_csk(sk);
2792         struct tcp_sock *tp = tcp_sk(sk);
2793         bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2794                                     (tcp_fackets_out(tp) > tp->reordering));
2795         int fast_rexmit = 0;
2796 
2797         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2798                 tp->sacked_out = 0;
2799         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2800                 tp->fackets_out = 0;
2801 
2802         /* Now state machine starts.
2803          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2804         if (flag & FLAG_ECE)
2805                 tp->prior_ssthresh = 0;
2806 
2807         /* B. In all the states check for reneging SACKs. */
2808         if (tcp_check_sack_reneging(sk, flag))
2809                 return;
2810 
2811         /* C. Check consistency of the current state. */
2812         tcp_verify_left_out(tp);
2813 
2814         /* D. Check state exit conditions. State can be terminated
2815          *    when high_seq is ACKed. */
2816         if (icsk->icsk_ca_state == TCP_CA_Open) {
2817                 WARN_ON(tp->retrans_out != 0);
2818                 tp->retrans_stamp = 0;
2819         } else if (!before(tp->snd_una, tp->high_seq)) {
2820                 switch (icsk->icsk_ca_state) {
2821                 case TCP_CA_CWR:
2822                         /* CWR is to be held something *above* high_seq
2823                          * is ACKed for CWR bit to reach receiver. */
2824                         if (tp->snd_una != tp->high_seq) {
2825                                 tcp_end_cwnd_reduction(sk);
2826                                 tcp_set_ca_state(sk, TCP_CA_Open);
2827                         }
2828                         break;
2829 
2830                 case TCP_CA_Recovery:
2831                         if (tcp_is_reno(tp))
2832                                 tcp_reset_reno_sack(tp);
2833                         if (tcp_try_undo_recovery(sk))
2834                                 return;
2835                         tcp_end_cwnd_reduction(sk);
2836                         break;
2837                 }
2838         }
2839 
2840         /* E. Process state. */
2841         switch (icsk->icsk_ca_state) {
2842         case TCP_CA_Recovery:
2843                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2844                         if (tcp_is_reno(tp) && is_dupack)
2845                                 tcp_add_reno_sack(sk);
2846                 } else {
2847                         if (tcp_try_undo_partial(sk, acked, prior_unsacked))
2848                                 return;
2849                         /* Partial ACK arrived. Force fast retransmit. */
2850                         do_lost = tcp_is_reno(tp) ||
2851                                   tcp_fackets_out(tp) > tp->reordering;
2852                 }
2853                 if (tcp_try_undo_dsack(sk)) {
2854                         tcp_try_keep_open(sk);
2855                         return;
2856                 }
2857                 break;
2858         case TCP_CA_Loss:
2859                 tcp_process_loss(sk, flag, is_dupack);
2860                 if (icsk->icsk_ca_state != TCP_CA_Open)
2861                         return;
2862                 /* Fall through to processing in Open state. */
2863         default:
2864                 if (tcp_is_reno(tp)) {
2865                         if (flag & FLAG_SND_UNA_ADVANCED)
2866                                 tcp_reset_reno_sack(tp);
2867                         if (is_dupack)
2868                                 tcp_add_reno_sack(sk);
2869                 }
2870 
2871                 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2872                         tcp_try_undo_dsack(sk);
2873 
2874                 if (!tcp_time_to_recover(sk, flag)) {
2875                         tcp_try_to_open(sk, flag, prior_unsacked);
2876                         return;
2877                 }
2878 
2879                 /* MTU probe failure: don't reduce cwnd */
2880                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2881                     icsk->icsk_mtup.probe_size &&
2882                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2883                         tcp_mtup_probe_failed(sk);
2884                         /* Restores the reduction we did in tcp_mtup_probe() */
2885                         tp->snd_cwnd++;
2886                         tcp_simple_retransmit(sk);
2887                         return;
2888                 }
2889 
2890                 /* Otherwise enter Recovery state */
2891                 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2892                 fast_rexmit = 1;
2893         }
2894 
2895         if (do_lost)
2896                 tcp_update_scoreboard(sk, fast_rexmit);
2897         tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
2898         tcp_xmit_retransmit_queue(sk);
2899 }
2900 
2901 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2902                                       long seq_rtt_us, long sack_rtt_us)
2903 {
2904         const struct tcp_sock *tp = tcp_sk(sk);
2905 
2906         /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2907          * broken middle-boxes or peers may corrupt TS-ECR fields. But
2908          * Karn's algorithm forbids taking RTT if some retransmitted data
2909          * is acked (RFC6298).
2910          */
2911         if (flag & FLAG_RETRANS_DATA_ACKED)
2912                 seq_rtt_us = -1L;
2913 
2914         if (seq_rtt_us < 0)
2915                 seq_rtt_us = sack_rtt_us;
2916 
2917         /* RTTM Rule: A TSecr value received in a segment is used to
2918          * update the averaged RTT measurement only if the segment
2919          * acknowledges some new data, i.e., only if it advances the
2920          * left edge of the send window.
2921          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2922          */
2923         if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2924             flag & FLAG_ACKED)
2925                 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2926 
2927         if (seq_rtt_us < 0)
2928                 return false;
2929 
2930         tcp_rtt_estimator(sk, seq_rtt_us);
2931         tcp_set_rto(sk);
2932 
2933         /* RFC6298: only reset backoff on valid RTT measurement. */
2934         inet_csk(sk)->icsk_backoff = 0;
2935         return true;
2936 }
2937 
2938 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2939 static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
2940 {
2941         struct tcp_sock *tp = tcp_sk(sk);
2942         long seq_rtt_us = -1L;
2943 
2944         if (synack_stamp && !tp->total_retrans)
2945                 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - synack_stamp);
2946 
2947         /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2948          * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2949          */
2950         if (!tp->srtt_us)
2951                 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt_us, -1L);
2952 }
2953 
2954 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2955 {
2956         const struct inet_connection_sock *icsk = inet_csk(sk);
2957 
2958         icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2959         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2960 }
2961 
2962 /* Restart timer after forward progress on connection.
2963  * RFC2988 recommends to restart timer to now+rto.
2964  */
2965 void tcp_rearm_rto(struct sock *sk)
2966 {
2967         const struct inet_connection_sock *icsk = inet_csk(sk);
2968         struct tcp_sock *tp = tcp_sk(sk);
2969 
2970         /* If the retrans timer is currently being used by Fast Open
2971          * for SYN-ACK retrans purpose, stay put.
2972          */
2973         if (tp->fastopen_rsk)
2974                 return;
2975 
2976         if (!tp->packets_out) {
2977                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2978         } else {
2979                 u32 rto = inet_csk(sk)->icsk_rto;
2980                 /* Offset the time elapsed after installing regular RTO */
2981                 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2982                     icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2983                         struct sk_buff *skb = tcp_write_queue_head(sk);
2984                         const u32 rto_time_stamp =
2985                                 tcp_skb_timestamp(skb) + rto;
2986                         s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2987                         /* delta may not be positive if the socket is locked
2988                          * when the retrans timer fires and is rescheduled.
2989                          */
2990                         if (delta > 0)
2991                                 rto = delta;
2992                 }
2993                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2994                                           TCP_RTO_MAX);
2995         }
2996 }
2997 
2998 /* This function is called when the delayed ER timer fires. TCP enters
2999  * fast recovery and performs fast-retransmit.
3000  */
3001 void tcp_resume_early_retransmit(struct sock *sk)
3002 {
3003         struct tcp_sock *tp = tcp_sk(sk);
3004 
3005         tcp_rearm_rto(sk);
3006 
3007         /* Stop if ER is disabled after the delayed ER timer is scheduled */
3008         if (!tp->do_early_retrans)
3009                 return;
3010 
3011         tcp_enter_recovery(sk, false);
3012         tcp_update_scoreboard(sk, 1);
3013         tcp_xmit_retransmit_queue(sk);
3014 }
3015 
3016 /* If we get here, the whole TSO packet has not been acked. */
3017 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3018 {
3019         struct tcp_sock *tp = tcp_sk(sk);
3020         u32 packets_acked;
3021 
3022         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3023 
3024         packets_acked = tcp_skb_pcount(skb);
3025         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3026                 return 0;
3027         packets_acked -= tcp_skb_pcount(skb);
3028 
3029         if (packets_acked) {
3030                 BUG_ON(tcp_skb_pcount(skb) == 0);
3031                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3032         }
3033 
3034         return packets_acked;
3035 }
3036 
3037 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3038                            u32 prior_snd_una)
3039 {
3040         const struct skb_shared_info *shinfo;
3041 
3042         /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3043         if (likely(!(sk->sk_tsflags & SOF_TIMESTAMPING_TX_ACK)))
3044                 return;
3045 
3046         shinfo = skb_shinfo(skb);
3047         if ((shinfo->tx_flags & SKBTX_ACK_TSTAMP) &&
3048             between(shinfo->tskey, prior_snd_una, tcp_sk(sk)->snd_una - 1))
3049                 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3050 }
3051 
3052 /* Remove acknowledged frames from the retransmission queue. If our packet
3053  * is before the ack sequence we can discard it as it's confirmed to have
3054  * arrived at the other end.
3055  */
3056 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3057                                u32 prior_snd_una, long sack_rtt_us)
3058 {
3059         const struct inet_connection_sock *icsk = inet_csk(sk);
3060         struct skb_mstamp first_ackt, last_ackt, now;
3061         struct tcp_sock *tp = tcp_sk(sk);
3062         u32 prior_sacked = tp->sacked_out;
3063         u32 reord = tp->packets_out;
3064         bool fully_acked = true;
3065         long ca_seq_rtt_us = -1L;
3066         long seq_rtt_us = -1L;
3067         struct sk_buff *skb;
3068         u32 pkts_acked = 0;
3069         bool rtt_update;
3070         int flag = 0;
3071 
3072         first_ackt.v64 = 0;
3073 
3074         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3075                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3076                 u8 sacked = scb->sacked;
3077                 u32 acked_pcount;
3078 
3079                 tcp_ack_tstamp(sk, skb, prior_snd_una);
3080 
3081                 /* Determine how many packets and what bytes were acked, tso and else */
3082                 if (after(scb->end_seq, tp->snd_una)) {
3083                         if (tcp_skb_pcount(skb) == 1 ||
3084                             !after(tp->snd_una, scb->seq))
3085                                 break;
3086 
3087                         acked_pcount = tcp_tso_acked(sk, skb);
3088                         if (!acked_pcount)
3089                                 break;
3090 
3091                         fully_acked = false;
3092                 } else {
3093                         /* Speedup tcp_unlink_write_queue() and next loop */
3094                         prefetchw(skb->next);
3095                         acked_pcount = tcp_skb_pcount(skb);
3096                 }
3097 
3098                 if (unlikely(sacked & TCPCB_RETRANS)) {
3099                         if (sacked & TCPCB_SACKED_RETRANS)
3100                                 tp->retrans_out -= acked_pcount;
3101                         flag |= FLAG_RETRANS_DATA_ACKED;
3102                 } else {
3103                         last_ackt = skb->skb_mstamp;
3104                         WARN_ON_ONCE(last_ackt.v64 == 0);
3105                         if (!first_ackt.v64)
3106                                 first_ackt = last_ackt;
3107 
3108                         if (!(sacked & TCPCB_SACKED_ACKED)) {
3109                                 reord = min(pkts_acked, reord);
3110                                 if (!after(scb->end_seq, tp->high_seq))
3111                                         flag |= FLAG_ORIG_SACK_ACKED;
3112                         }
3113                 }
3114 
3115                 if (sacked & TCPCB_SACKED_ACKED)
3116                         tp->sacked_out -= acked_pcount;
3117                 if (sacked & TCPCB_LOST)
3118                         tp->lost_out -= acked_pcount;
3119 
3120                 tp->packets_out -= acked_pcount;
3121                 pkts_acked += acked_pcount;
3122 
3123                 /* Initial outgoing SYN's get put onto the write_queue
3124                  * just like anything else we transmit.  It is not
3125                  * true data, and if we misinform our callers that
3126                  * this ACK acks real data, we will erroneously exit
3127                  * connection startup slow start one packet too
3128                  * quickly.  This is severely frowned upon behavior.
3129                  */
3130                 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3131                         flag |= FLAG_DATA_ACKED;
3132                 } else {
3133                         flag |= FLAG_SYN_ACKED;
3134                         tp->retrans_stamp = 0;
3135                 }
3136 
3137                 if (!fully_acked)
3138                         break;
3139 
3140                 tcp_unlink_write_queue(skb, sk);
3141                 sk_wmem_free_skb(sk, skb);
3142                 if (unlikely(skb == tp->retransmit_skb_hint))
3143                         tp->retransmit_skb_hint = NULL;
3144                 if (unlikely(skb == tp->lost_skb_hint))
3145                         tp->lost_skb_hint = NULL;
3146         }
3147 
3148         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3149                 tp->snd_up = tp->snd_una;
3150 
3151         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3152                 flag |= FLAG_SACK_RENEGING;
3153 
3154         skb_mstamp_get(&now);
3155         if (likely(first_ackt.v64)) {
3156                 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3157                 ca_seq_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3158         }
3159 
3160         rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us);
3161 
3162         if (flag & FLAG_ACKED) {
3163                 const struct tcp_congestion_ops *ca_ops
3164                         = inet_csk(sk)->icsk_ca_ops;
3165 
3166                 tcp_rearm_rto(sk);
3167                 if (unlikely(icsk->icsk_mtup.probe_size &&
3168                              !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3169                         tcp_mtup_probe_success(sk);
3170                 }
3171 
3172                 if (tcp_is_reno(tp)) {
3173                         tcp_remove_reno_sacks(sk, pkts_acked);
3174                 } else {
3175                         int delta;
3176 
3177                         /* Non-retransmitted hole got filled? That's reordering */
3178                         if (reord < prior_fackets)
3179                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3180 
3181                         delta = tcp_is_fack(tp) ? pkts_acked :
3182                                                   prior_sacked - tp->sacked_out;
3183                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3184                 }
3185 
3186                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3187 
3188                 if (ca_ops->pkts_acked) {
3189                         long rtt_us = min_t(ulong, ca_seq_rtt_us, sack_rtt_us);
3190                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3191                 }
3192 
3193         } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3194                    sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3195                 /* Do not re-arm RTO if the sack RTT is measured from data sent
3196                  * after when the head was last (re)transmitted. Otherwise the
3197                  * timeout may continue to extend in loss recovery.
3198                  */
3199                 tcp_rearm_rto(sk);
3200         }
3201 
3202 #if FASTRETRANS_DEBUG > 0
3203         WARN_ON((int)tp->sacked_out < 0);
3204         WARN_ON((int)tp->lost_out < 0);
3205         WARN_ON((int)tp->retrans_out < 0);
3206         if (!tp->packets_out && tcp_is_sack(tp)) {
3207                 icsk = inet_csk(sk);
3208                 if (tp->lost_out) {
3209                         pr_debug("Leak l=%u %d\n",
3210                                  tp->lost_out, icsk->icsk_ca_state);
3211                         tp->lost_out = 0;
3212                 }
3213                 if (tp->sacked_out) {
3214                         pr_debug("Leak s=%u %d\n",
3215                                  tp->sacked_out, icsk->icsk_ca_state);
3216                         tp->sacked_out = 0;
3217                 }
3218                 if (tp->retrans_out) {
3219                         pr_debug("Leak r=%u %d\n",
3220                                  tp->retrans_out, icsk->icsk_ca_state);
3221                         tp->retrans_out = 0;
3222                 }
3223         }
3224 #endif
3225         return flag;
3226 }
3227 
3228 static void tcp_ack_probe(struct sock *sk)
3229 {
3230         const struct tcp_sock *tp = tcp_sk(sk);
3231         struct inet_connection_sock *icsk = inet_csk(sk);
3232 
3233         /* Was it a usable window open? */
3234 
3235         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3236                 icsk->icsk_backoff = 0;
3237                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3238                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3239                  * This function is not for random using!
3240                  */
3241         } else {
3242                 unsigned long when = inet_csk_rto_backoff(icsk, TCP_RTO_MAX);
3243 
3244                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3245                                           when, TCP_RTO_MAX);
3246         }
3247 }
3248 
3249 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3250 {
3251         return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3252                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3253 }
3254 
3255 /* Decide wheather to run the increase function of congestion control. */
3256 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3257 {
3258         if (tcp_in_cwnd_reduction(sk))
3259                 return false;
3260 
3261         /* If reordering is high then always grow cwnd whenever data is
3262          * delivered regardless of its ordering. Otherwise stay conservative
3263          * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3264          * new SACK or ECE mark may first advance cwnd here and later reduce
3265          * cwnd in tcp_fastretrans_alert() based on more states.
3266          */
3267         if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3268                 return flag & FLAG_FORWARD_PROGRESS;
3269 
3270         return flag & FLAG_DATA_ACKED;
3271 }
3272 
3273 /* Check that window update is acceptable.
3274  * The function assumes that snd_una<=ack<=snd_next.
3275  */
3276 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3277                                         const u32 ack, const u32 ack_seq,
3278                                         const u32 nwin)
3279 {
3280         return  after(ack, tp->snd_una) ||
3281                 after(ack_seq, tp->snd_wl1) ||
3282                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3283 }
3284 
3285 /* Update our send window.
3286  *
3287  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3288  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3289  */
3290 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3291                                  u32 ack_seq)
3292 {
3293         struct tcp_sock *tp = tcp_sk(sk);
3294         int flag = 0;
3295         u32 nwin = ntohs(tcp_hdr(skb)->window);
3296 
3297         if (likely(!tcp_hdr(skb)->syn))
3298                 nwin <<= tp->rx_opt.snd_wscale;
3299 
3300         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3301                 flag |= FLAG_WIN_UPDATE;
3302                 tcp_update_wl(tp, ack_seq);
3303 
3304                 if (tp->snd_wnd != nwin) {
3305                         tp->snd_wnd = nwin;
3306 
3307                         /* Note, it is the only place, where
3308                          * fast path is recovered for sending TCP.
3309                          */
3310                         tp->pred_flags = 0;
3311                         tcp_fast_path_check(sk);
3312 
3313                         if (nwin > tp->max_window) {
3314                                 tp->max_window = nwin;
3315                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3316                         }
3317                 }
3318         }
3319 
3320         tp->snd_una = ack;
3321 
3322         return flag;
3323 }
3324 
3325 /* RFC 5961 7 [ACK Throttling] */
3326 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3327 {
3328         /* unprotected vars, we dont care of overwrites */
3329         static u32 challenge_timestamp;
3330         static unsigned int challenge_count;
3331         struct tcp_sock *tp = tcp_sk(sk);
3332         u32 now;
3333 
3334         /* First check our per-socket dupack rate limit. */
3335         if (tcp_oow_rate_limited(sock_net(sk), skb,
3336                                  LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3337                                  &tp->last_oow_ack_time))
3338                 return;
3339 
3340         /* Then check the check host-wide RFC 5961 rate limit. */
3341         now = jiffies / HZ;
3342         if (now != challenge_timestamp) {
3343                 challenge_timestamp = now;
3344                 challenge_count = 0;
3345         }
3346         if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3347                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3348                 tcp_send_ack(sk);
3349         }
3350 }
3351 
3352 static void tcp_store_ts_recent(struct tcp_sock *tp)
3353 {
3354         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3355         tp->rx_opt.ts_recent_stamp = get_seconds();
3356 }
3357 
3358 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3359 {
3360         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3361                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3362                  * extra check below makes sure this can only happen
3363                  * for pure ACK frames.  -DaveM
3364                  *
3365                  * Not only, also it occurs for expired timestamps.
3366                  */
3367 
3368                 if (tcp_paws_check(&tp->rx_opt, 0))
3369                         tcp_store_ts_recent(tp);
3370         }
3371 }
3372 
3373 /* This routine deals with acks during a TLP episode.
3374  * We mark the end of a TLP episode on receiving TLP dupack or when
3375  * ack is after tlp_high_seq.
3376  * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3377  */
3378 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3379 {
3380         struct tcp_sock *tp = tcp_sk(sk);
3381 
3382         if (before(ack, tp->tlp_high_seq))
3383                 return;
3384 
3385         if (flag & FLAG_DSACKING_ACK) {
3386                 /* This DSACK means original and TLP probe arrived; no loss */
3387                 tp->tlp_high_seq = 0;
3388         } else if (after(ack, tp->tlp_high_seq)) {
3389                 /* ACK advances: there was a loss, so reduce cwnd. Reset
3390                  * tlp_high_seq in tcp_init_cwnd_reduction()
3391                  */
3392                 tcp_init_cwnd_reduction(sk);
3393                 tcp_set_ca_state(sk, TCP_CA_CWR);
3394                 tcp_end_cwnd_reduction(sk);
3395                 tcp_try_keep_open(sk);
3396                 NET_INC_STATS_BH(sock_net(sk),
3397                                  LINUX_MIB_TCPLOSSPROBERECOVERY);
3398         } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3399                              FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3400                 /* Pure dupack: original and TLP probe arrived; no loss */
3401                 tp->tlp_high_seq = 0;
3402         }
3403 }
3404 
3405 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3406 {
3407         const struct inet_connection_sock *icsk = inet_csk(sk);
3408 
3409         if (icsk->icsk_ca_ops->in_ack_event)
3410                 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3411 }
3412 
3413 /* This routine deals with incoming acks, but not outgoing ones. */
3414 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3415 {
3416         struct inet_connection_sock *icsk = inet_csk(sk);
3417         struct tcp_sock *tp = tcp_sk(sk);
3418         u32 prior_snd_una = tp->snd_una;
3419         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3420         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3421         bool is_dupack = false;
3422         u32 prior_fackets;
3423         int prior_packets = tp->packets_out;
3424         const int prior_unsacked = tp->packets_out - tp->sacked_out;
3425         int acked = 0; /* Number of packets newly acked */
3426         long sack_rtt_us = -1L;
3427 
3428         /* We very likely will need to access write queue head. */
3429         prefetchw(sk->sk_write_queue.next);
3430 
3431         /* If the ack is older than previous acks
3432          * then we can probably ignore it.
3433          */
3434         if (before(ack, prior_snd_una)) {
3435                 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3436                 if (before(ack, prior_snd_una - tp->max_window)) {
3437                         tcp_send_challenge_ack(sk, skb);
3438                         return -1;
3439                 }
3440                 goto old_ack;
3441         }
3442 
3443         /* If the ack includes data we haven't sent yet, discard
3444          * this segment (RFC793 Section 3.9).
3445          */
3446         if (after(ack, tp->snd_nxt))
3447                 goto invalid_ack;
3448 
3449         if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3450             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3451                 tcp_rearm_rto(sk);
3452 
3453         if (after(ack, prior_snd_una)) {
3454                 flag |= FLAG_SND_UNA_ADVANCED;
3455                 icsk->icsk_retransmits = 0;
3456         }
3457 
3458         prior_fackets = tp->fackets_out;
3459 
3460         /* ts_recent update must be made after we are sure that the packet
3461          * is in window.
3462          */
3463         if (flag & FLAG_UPDATE_TS_RECENT)
3464                 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3465 
3466         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3467                 /* Window is constant, pure forward advance.
3468                  * No more checks are required.
3469                  * Note, we use the fact that SND.UNA>=SND.WL2.
3470                  */
3471                 tcp_update_wl(tp, ack_seq);
3472                 tp->snd_una = ack;
3473                 flag |= FLAG_WIN_UPDATE;
3474 
3475                 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3476 
3477                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3478         } else {
3479                 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3480 
3481                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3482                         flag |= FLAG_DATA;
3483                 else
3484                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3485 
3486                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3487 
3488                 if (TCP_SKB_CB(skb)->sacked)
3489                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3490                                                         &sack_rtt_us);
3491 
3492                 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3493                         flag |= FLAG_ECE;
3494                         ack_ev_flags |= CA_ACK_ECE;
3495                 }
3496 
3497                 if (flag & FLAG_WIN_UPDATE)
3498                         ack_ev_flags |= CA_ACK_WIN_UPDATE;
3499 
3500                 tcp_in_ack_event(sk, ack_ev_flags);
3501         }
3502 
3503         /* We passed data and got it acked, remove any soft error
3504          * log. Something worked...
3505          */
3506         sk->sk_err_soft = 0;
3507         icsk->icsk_probes_out = 0;
3508         tp->rcv_tstamp = tcp_time_stamp;
3509         if (!prior_packets)
3510                 goto no_queue;
3511 
3512         /* See if we can take anything off of the retransmit queue. */
3513         acked = tp->packets_out;
3514         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
3515                                     sack_rtt_us);
3516         acked -= tp->packets_out;
3517 
3518         /* Advance cwnd if state allows */
3519         if (tcp_may_raise_cwnd(sk, flag))
3520                 tcp_cong_avoid(sk, ack, acked);
3521 
3522         if (tcp_ack_is_dubious(sk, flag)) {
3523                 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3524                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3525                                       is_dupack, flag);
3526         }
3527         if (tp->tlp_high_seq)
3528                 tcp_process_tlp_ack(sk, ack, flag);
3529 
3530         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3531                 struct dst_entry *dst = __sk_dst_get(sk);
3532                 if (dst)
3533                         dst_confirm(dst);
3534         }
3535 
3536         if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3537                 tcp_schedule_loss_probe(sk);
3538         tcp_update_pacing_rate(sk);
3539         return 1;
3540 
3541 no_queue:
3542         /* If data was DSACKed, see if we can undo a cwnd reduction. */
3543         if (flag & FLAG_DSACKING_ACK)
3544                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3545                                       is_dupack, flag);
3546         /* If this ack opens up a zero window, clear backoff.  It was
3547          * being used to time the probes, and is probably far higher than
3548          * it needs to be for normal retransmission.
3549          */
3550         if (tcp_send_head(sk))
3551                 tcp_ack_probe(sk);
3552 
3553         if (tp->tlp_high_seq)
3554                 tcp_process_tlp_ack(sk, ack, flag);
3555         return 1;
3556 
3557 invalid_ack:
3558         SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3559         return -1;
3560 
3561 old_ack:
3562         /* If data was SACKed, tag it and see if we should send more data.
3563          * If data was DSACKed, see if we can undo a cwnd reduction.
3564          */
3565         if (TCP_SKB_CB(skb)->sacked) {
3566                 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3567                                                 &sack_rtt_us);
3568                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3569                                       is_dupack, flag);
3570         }
3571 
3572         SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3573         return 0;
3574 }
3575 
3576 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3577  * But, this can also be called on packets in the established flow when
3578  * the fast version below fails.
3579  */
3580 void tcp_parse_options(const struct sk_buff *skb,
3581                        struct tcp_options_received *opt_rx, int estab,
3582                        struct tcp_fastopen_cookie *foc)
3583 {
3584         const unsigned char *ptr;
3585         const struct tcphdr *th = tcp_hdr(skb);
3586         int length = (th->doff * 4) - sizeof(struct tcphdr);
3587 
3588         ptr = (const unsigned char *)(th + 1);
3589         opt_rx->saw_tstamp = 0;
3590 
3591         while (length > 0) {
3592                 int opcode = *ptr++;
3593                 int opsize;
3594 
3595                 switch (opcode) {
3596                 case TCPOPT_EOL:
3597                         return;
3598                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3599                         length--;
3600                         continue;
3601                 default:
3602                         opsize = *ptr++;
3603                         if (opsize < 2) /* "silly options" */
3604                                 return;
3605                         if (opsize > length)
3606                                 return; /* don't parse partial options */
3607                         switch (opcode) {
3608                         case TCPOPT_MSS:
3609                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3610                                         u16 in_mss = get_unaligned_be16(ptr);
3611                                         if (in_mss) {
3612                                                 if (opt_rx->user_mss &&
3613                                                     opt_rx->user_mss < in_mss)
3614                                                         in_mss = opt_rx->user_mss;
3615                                                 opt_rx->mss_clamp = in_mss;
3616                                         }
3617                                 }
3618                                 break;
3619                         case TCPOPT_WINDOW:
3620                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3621                                     !estab && sysctl_tcp_window_scaling) {
3622                                         __u8 snd_wscale = *(__u8 *)ptr;
3623                                         opt_rx->wscale_ok = 1;
3624                                         if (snd_wscale > 14) {
3625                                                 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3626                                                                      __func__,
3627                                                                      snd_wscale);
3628                                                 snd_wscale = 14;
3629                                         }
3630                                         opt_rx->snd_wscale = snd_wscale;
3631                                 }
3632                                 break;
3633                         case TCPOPT_TIMESTAMP:
3634                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3635                                     ((estab && opt_rx->tstamp_ok) ||
3636                                      (!estab && sysctl_tcp_timestamps))) {
3637                                         opt_rx->saw_tstamp = 1;
3638                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3639                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3640                                 }
3641                                 break;
3642                         case TCPOPT_SACK_PERM:
3643                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3644                                     !estab && sysctl_tcp_sack) {
3645                                         opt_rx->sack_ok = TCP_SACK_SEEN;
3646                                         tcp_sack_reset(opt_rx);
3647                                 }
3648                                 break;
3649 
3650                         case TCPOPT_SACK:
3651                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3652                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3653                                    opt_rx->sack_ok) {
3654                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3655                                 }
3656                                 break;
3657 #ifdef CONFIG_TCP_MD5SIG
3658                         case TCPOPT_MD5SIG:
3659                                 /*
3660                                  * The MD5 Hash has already been
3661                                  * checked (see tcp_v{4,6}_do_rcv()).
3662                                  */
3663                                 break;
3664 #endif
3665                         case TCPOPT_EXP:
3666                                 /* Fast Open option shares code 254 using a
3667                                  * 16 bits magic number. It's valid only in
3668                                  * SYN or SYN-ACK with an even size.
3669                                  */
3670                                 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3671                                     get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3672                                     foc == NULL || !th->syn || (opsize & 1))
3673                                         break;
3674                                 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3675                                 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3676                                     foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3677                                         memcpy(foc->val, ptr + 2, foc->len);
3678                                 else if (foc->len != 0)
3679                                         foc->len = -1;
3680                                 break;
3681 
3682                         }
3683                         ptr += opsize-2;
3684                         length -= opsize;
3685                 }
3686         }
3687 }
3688 EXPORT_SYMBOL(tcp_parse_options);
3689 
3690 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3691 {
3692         const __be32 *ptr = (const __be32 *)(th + 1);
3693 
3694         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3695                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3696                 tp->rx_opt.saw_tstamp = 1;
3697                 ++ptr;
3698                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3699                 ++ptr;
3700                 if (*ptr)
3701                         tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3702                 else
3703                         tp->rx_opt.rcv_tsecr = 0;
3704                 return true;
3705         }
3706         return false;
3707 }
3708 
3709 /* Fast parse options. This hopes to only see timestamps.
3710  * If it is wrong it falls back on tcp_parse_options().
3711  */
3712 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3713                                    const struct tcphdr *th, struct tcp_sock *tp)
3714 {
3715         /* In the spirit of fast parsing, compare doff directly to constant
3716          * values.  Because equality is used, short doff can be ignored here.
3717          */
3718         if (th->doff == (sizeof(*th) / 4)) {
3719                 tp->rx_opt.saw_tstamp = 0;
3720                 return false;
3721         } else if (tp->rx_opt.tstamp_ok &&
3722                    th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3723                 if (tcp_parse_aligned_timestamp(tp, th))
3724                         return true;
3725         }
3726 
3727         tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3728         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3729                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3730 
3731         return true;
3732 }
3733 
3734 #ifdef CONFIG_TCP_MD5SIG
3735 /*
3736  * Parse MD5 Signature option
3737  */
3738 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3739 {
3740         int length = (th->doff << 2) - sizeof(*th);
3741         const u8 *ptr = (const u8 *)(th + 1);
3742 
3743         /* If the TCP option is too short, we can short cut */
3744         if (length < TCPOLEN_MD5SIG)
3745                 return NULL;
3746 
3747         while (length > 0) {
3748                 int opcode = *ptr++;
3749                 int opsize;
3750 
3751                 switch (opcode) {
3752                 case TCPOPT_EOL:
3753                         return NULL;
3754                 case TCPOPT_NOP:
3755                         length--;
3756                         continue;
3757                 default:
3758                         opsize = *ptr++;
3759                         if (opsize < 2 || opsize > length)
3760                                 return NULL;
3761                         if (opcode == TCPOPT_MD5SIG)
3762                                 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3763                 }
3764                 ptr += opsize - 2;
3765                 length -= opsize;
3766         }
3767         return NULL;
3768 }
3769 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3770 #endif
3771 
3772 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3773  *
3774  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3775  * it can pass through stack. So, the following predicate verifies that
3776  * this segment is not used for anything but congestion avoidance or
3777  * fast retransmit. Moreover, we even are able to eliminate most of such
3778  * second order effects, if we apply some small "replay" window (~RTO)
3779  * to timestamp space.
3780  *
3781  * All these measures still do not guarantee that we reject wrapped ACKs
3782  * on networks with high bandwidth, when sequence space is recycled fastly,
3783  * but it guarantees that such events will be very rare and do not affect
3784  * connection seriously. This doesn't look nice, but alas, PAWS is really
3785  * buggy extension.
3786  *
3787  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3788  * states that events when retransmit arrives after original data are rare.
3789  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3790  * the biggest problem on large power networks even with minor reordering.
3791  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3792  * up to bandwidth of 18Gigabit/sec. 8) ]
3793  */
3794 
3795 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3796 {
3797         const struct tcp_sock *tp = tcp_sk(sk);
3798         const struct tcphdr *th = tcp_hdr(skb);
3799         u32 seq = TCP_SKB_CB(skb)->seq;
3800         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3801 
3802         return (/* 1. Pure ACK with correct sequence number. */
3803                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3804 
3805                 /* 2. ... and duplicate ACK. */
3806                 ack == tp->snd_una &&
3807 
3808                 /* 3. ... and does not update window. */
3809                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3810 
3811                 /* 4. ... and sits in replay window. */
3812                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3813 }
3814 
3815 static inline bool tcp_paws_discard(const struct sock *sk,
3816                                    const struct sk_buff *skb)
3817 {
3818         const struct tcp_sock *tp = tcp_sk(sk);
3819 
3820         return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3821                !tcp_disordered_ack(sk, skb);
3822 }
3823 
3824 /* Check segment sequence number for validity.
3825  *
3826  * Segment controls are considered valid, if the segment
3827  * fits to the window after truncation to the window. Acceptability
3828  * of data (and SYN, FIN, of course) is checked separately.
3829  * See tcp_data_queue(), for example.
3830  *
3831  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3832  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3833  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3834  * (borrowed from freebsd)
3835  */
3836 
3837 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3838 {
3839         return  !before(end_seq, tp->rcv_wup) &&
3840                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3841 }
3842 
3843 /* When we get a reset we do this. */
3844 void tcp_reset(struct sock *sk)
3845 {
3846         /* We want the right error as BSD sees it (and indeed as we do). */
3847         switch (sk->sk_state) {
3848         case TCP_SYN_SENT:
3849                 sk->sk_err = ECONNREFUSED;
3850                 break;
3851         case TCP_CLOSE_WAIT:
3852                 sk->sk_err = EPIPE;
3853                 break;
3854         case TCP_CLOSE:
3855                 return;
3856         default:
3857                 sk->sk_err = ECONNRESET;
3858         }
3859         /* This barrier is coupled with smp_rmb() in tcp_poll() */
3860         smp_wmb();
3861 
3862         if (!sock_flag(sk, SOCK_DEAD))
3863                 sk->sk_error_report(sk);
3864 
3865         tcp_done(sk);
3866 }
3867 
3868 /*
3869  *      Process the FIN bit. This now behaves as it is supposed to work
3870  *      and the FIN takes effect when it is validly part of sequence
3871  *      space. Not before when we get holes.
3872  *
3873  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3874  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3875  *      TIME-WAIT)
3876  *
3877  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3878  *      close and we go into CLOSING (and later onto TIME-WAIT)
3879  *
3880  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3881  */
3882 static void tcp_fin(struct sock *sk)
3883 {
3884         struct tcp_sock *tp = tcp_sk(sk);
3885         const struct dst_entry *dst;
3886 
3887         inet_csk_schedule_ack(sk);
3888 
3889         sk->sk_shutdown |= RCV_SHUTDOWN;
3890         sock_set_flag(sk, SOCK_DONE);
3891 
3892         switch (sk->sk_state) {
3893         case TCP_SYN_RECV:
3894         case TCP_ESTABLISHED:
3895                 /* Move to CLOSE_WAIT */
3896                 tcp_set_state(sk, TCP_CLOSE_WAIT);
3897                 dst = __sk_dst_get(sk);
3898                 if (!dst || !dst_metric(dst, RTAX_QUICKACK))
3899                         inet_csk(sk)->icsk_ack.pingpong = 1;
3900                 break;
3901 
3902         case TCP_CLOSE_WAIT:
3903         case TCP_CLOSING:
3904                 /* Received a retransmission of the FIN, do
3905                  * nothing.
3906                  */
3907                 break;
3908         case TCP_LAST_ACK:
3909                 /* RFC793: Remain in the LAST-ACK state. */
3910                 break;
3911 
3912         case TCP_FIN_WAIT1:
3913                 /* This case occurs when a simultaneous close
3914                  * happens, we must ack the received FIN and
3915                  * enter the CLOSING state.
3916                  */
3917                 tcp_send_ack(sk);
3918                 tcp_set_state(sk, TCP_CLOSING);
3919                 break;
3920         case TCP_FIN_WAIT2:
3921                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3922                 tcp_send_ack(sk);
3923                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3924                 break;
3925         default:
3926                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3927                  * cases we should never reach this piece of code.
3928                  */
3929                 pr_err("%s: Impossible, sk->sk_state=%d\n",
3930                        __func__, sk->sk_state);
3931                 break;
3932         }
3933 
3934         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3935          * Probably, we should reset in this case. For now drop them.
3936          */
3937         __skb_queue_purge(&tp->out_of_order_queue);
3938         if (tcp_is_sack(tp))
3939                 tcp_sack_reset(&tp->rx_opt);
3940         sk_mem_reclaim(sk);
3941 
3942         if (!sock_flag(sk, SOCK_DEAD)) {
3943                 sk->sk_state_change(sk);
3944 
3945                 /* Do not send POLL_HUP for half duplex close. */
3946                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3947                     sk->sk_state == TCP_CLOSE)
3948                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3949                 else
3950                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3951         }
3952 }
3953 
3954 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3955                                   u32 end_seq)
3956 {
3957         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3958                 if (before(seq, sp->start_seq))
3959                         sp->start_seq = seq;
3960                 if (after(end_seq, sp->end_seq))
3961                         sp->end_seq = end_seq;
3962                 return true;
3963         }
3964         return false;
3965 }
3966 
3967 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3968 {
3969         struct tcp_sock *tp = tcp_sk(sk);
3970 
3971         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3972                 int mib_idx;
3973 
3974                 if (before(seq, tp->rcv_nxt))
3975                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3976                 else
3977                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3978 
3979                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3980 
3981                 tp->rx_opt.dsack = 1;
3982                 tp->duplicate_sack[0].start_seq = seq;
3983                 tp->duplicate_sack[0].end_seq = end_seq;
3984         }
3985 }
3986 
3987 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3988 {
3989         struct tcp_sock *tp = tcp_sk(sk);
3990 
3991         if (!tp->rx_opt.dsack)
3992                 tcp_dsack_set(sk, seq, end_seq);
3993         else
3994                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3995 }
3996 
3997 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
3998 {
3999         struct tcp_sock *tp = tcp_sk(sk);
4000 
4001         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4002             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4003                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4004                 tcp_enter_quickack_mode(sk);
4005 
4006                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4007                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4008 
4009                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4010                                 end_seq = tp->rcv_nxt;
4011                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4012                 }
4013         }
4014 
4015         tcp_send_ack(sk);
4016 }
4017 
4018 /* These routines update the SACK block as out-of-order packets arrive or
4019  * in-order packets close up the sequence space.
4020  */
4021 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4022 {
4023         int this_sack;
4024         struct tcp_sack_block *sp = &tp->selective_acks[0];
4025         struct tcp_sack_block *swalk = sp + 1;
4026 
4027         /* See if the recent change to the first SACK eats into
4028          * or hits the sequence space of other SACK blocks, if so coalesce.
4029          */
4030         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4031                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4032                         int i;
4033 
4034                         /* Zap SWALK, by moving every further SACK up by one slot.
4035                          * Decrease num_sacks.
4036                          */
4037                         tp->rx_opt.num_sacks--;
4038                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4039                                 sp[i] = sp[i + 1];
4040                         continue;
4041                 }
4042                 this_sack++, swalk++;
4043         }
4044 }
4045 
4046 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4047 {
4048         struct tcp_sock *tp = tcp_sk(sk);
4049         struct tcp_sack_block *sp = &tp->selective_acks[0];
4050         int cur_sacks = tp->rx_opt.num_sacks;
4051         int this_sack;
4052 
4053         if (!cur_sacks)
4054                 goto new_sack;
4055 
4056         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4057                 if (tcp_sack_extend(sp, seq, end_seq)) {
4058                         /* Rotate this_sack to the first one. */
4059                         for (; this_sack > 0; this_sack--, sp--)
4060                                 swap(*sp, *(sp - 1));
4061                         if (cur_sacks > 1)
4062                                 tcp_sack_maybe_coalesce(tp);
4063                         return;
4064                 }
4065         }
4066 
4067         /* Could not find an adjacent existing SACK, build a new one,
4068          * put it at the front, and shift everyone else down.  We
4069          * always know there is at least one SACK present already here.
4070          *
4071          * If the sack array is full, forget about the last one.
4072          */
4073         if (this_sack >= TCP_NUM_SACKS) {
4074                 this_sack--;
4075                 tp->rx_opt.num_sacks--;
4076                 sp--;
4077         }
4078         for (; this_sack > 0; this_sack--, sp--)
4079                 *sp = *(sp - 1);
4080 
4081 new_sack:
4082         /* Build the new head SACK, and we're done. */
4083         sp->start_seq = seq;
4084         sp->end_seq = end_seq;
4085         tp->rx_opt.num_sacks++;
4086 }
4087 
4088 /* RCV.NXT advances, some SACKs should be eaten. */
4089 
4090 static void tcp_sack_remove(struct tcp_sock *tp)
4091 {
4092         struct tcp_sack_block *sp = &tp->selective_acks[0];
4093         int num_sacks = tp->rx_opt.num_sacks;
4094         int this_sack;
4095 
4096         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4097         if (skb_queue_empty(&tp->out_of_order_queue)) {
4098                 tp->rx_opt.num_sacks = 0;
4099                 return;
4100         }
4101 
4102         for (this_sack = 0; this_sack < num_sacks;) {
4103                 /* Check if the start of the sack is covered by RCV.NXT. */
4104                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4105                         int i;
4106 
4107                         /* RCV.NXT must cover all the block! */
4108                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4109 
4110                         /* Zap this SACK, by moving forward any other SACKS. */
4111                         for (i = this_sack+1; i < num_sacks; i++)
4112                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4113                         num_sacks--;
4114                         continue;
4115                 }
4116                 this_sack++;
4117                 sp++;
4118         }
4119         tp->rx_opt.num_sacks = num_sacks;
4120 }
4121 
4122 /**
4123  * tcp_try_coalesce - try to merge skb to prior one
4124  * @sk: socket
4125  * @to: prior buffer
4126  * @from: buffer to add in queue
4127  * @fragstolen: pointer to boolean
4128  *
4129  * Before queueing skb @from after @to, try to merge them
4130  * to reduce overall memory use and queue lengths, if cost is small.
4131  * Packets in ofo or receive queues can stay a long time.
4132  * Better try to coalesce them right now to avoid future collapses.
4133  * Returns true if caller should free @from instead of queueing it
4134  */
4135 static bool tcp_try_coalesce(struct sock *sk,
4136                              struct sk_buff *to,
4137                              struct sk_buff *from,
4138                              bool *fragstolen)
4139 {
4140         int delta;
4141 
4142         *fragstolen = false;
4143 
4144         /* Its possible this segment overlaps with prior segment in queue */
4145         if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4146                 return false;
4147 
4148         if (!skb_try_coalesce(to, from, fragstolen, &delta))
4149                 return false;
4150 
4151         atomic_add(delta, &sk->sk_rmem_alloc);
4152         sk_mem_charge(sk, delta);
4153         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4154         TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4155         TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4156         TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4157         return true;
4158 }
4159 
4160 /* This one checks to see if we can put data from the
4161  * out_of_order queue into the receive_queue.
4162  */
4163 static void tcp_ofo_queue(struct sock *sk)
4164 {
4165         struct tcp_sock *tp = tcp_sk(sk);
4166         __u32 dsack_high = tp->rcv_nxt;
4167         struct sk_buff *skb, *tail;
4168         bool fragstolen, eaten;
4169 
4170         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4171                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4172                         break;
4173 
4174                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4175                         __u32 dsack = dsack_high;
4176                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4177                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4178                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4179                 }
4180 
4181                 __skb_unlink(skb, &tp->out_of_order_queue);
4182                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4183                         SOCK_DEBUG(sk, "ofo packet was already received\n");
4184                         __kfree_skb(skb);
4185                         continue;
4186                 }
4187                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4188                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4189                            TCP_SKB_CB(skb)->end_seq);
4190 
4191                 tail = skb_peek_tail(&sk->sk_receive_queue);
4192                 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4193                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4194                 if (!eaten)
4195                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4196                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4197                         tcp_fin(sk);
4198                 if (eaten)
4199                         kfree_skb_partial(skb, fragstolen);
4200         }
4201 }
4202 
4203 static bool tcp_prune_ofo_queue(struct sock *sk);
4204 static int tcp_prune_queue(struct sock *sk);
4205 
4206 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4207                                  unsigned int size)
4208 {
4209         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4210             !sk_rmem_schedule(sk, skb, size)) {
4211 
4212                 if (tcp_prune_queue(sk) < 0)
4213                         return -1;
4214 
4215                 if (!sk_rmem_schedule(sk, skb, size)) {
4216                         if (!tcp_prune_ofo_queue(sk))
4217                                 return -1;
4218 
4219                         if (!sk_rmem_schedule(sk, skb, size))
4220                                 return -1;
4221                 }
4222         }
4223         return 0;
4224 }
4225 
4226 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4227 {
4228         struct tcp_sock *tp = tcp_sk(sk);
4229         struct sk_buff *skb1;
4230         u32 seq, end_seq;
4231 
4232         tcp_ecn_check_ce(tp, skb);
4233 
4234         if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4235                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4236                 __kfree_skb(skb);
4237                 return;
4238         }
4239 
4240         /* Disable header prediction. */
4241         tp->pred_flags = 0;
4242         inet_csk_schedule_ack(sk);
4243 
4244         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4245         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4246                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4247 
4248         skb1 = skb_peek_tail(&tp->out_of_order_queue);
4249         if (!skb1) {
4250                 /* Initial out of order segment, build 1 SACK. */
4251                 if (tcp_is_sack(tp)) {
4252                         tp->rx_opt.num_sacks = 1;
4253                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4254                         tp->selective_acks[0].end_seq =
4255                                                 TCP_SKB_CB(skb)->end_seq;
4256                 }
4257                 __skb_queue_head(&tp->out_of_order_queue, skb);
4258                 goto end;
4259         }
4260 
4261         seq = TCP_SKB_CB(skb)->seq;
4262         end_seq = TCP_SKB_CB(skb)->end_seq;
4263 
4264         if (seq == TCP_SKB_CB(skb1)->end_seq) {
4265                 bool fragstolen;
4266 
4267                 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4268                         __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4269                 } else {
4270                         tcp_grow_window(sk, skb);
4271                         kfree_skb_partial(skb, fragstolen);
4272                         skb = NULL;
4273                 }
4274 
4275                 if (!tp->rx_opt.num_sacks ||
4276                     tp->selective_acks[0].end_seq != seq)
4277                         goto add_sack;
4278 
4279                 /* Common case: data arrive in order after hole. */
4280                 tp->selective_acks[0].end_seq = end_seq;
4281                 goto end;
4282         }
4283 
4284         /* Find place to insert this segment. */
4285         while (1) {
4286                 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4287                         break;
4288                 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4289                         skb1 = NULL;
4290                         break;
4291                 }
4292                 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4293         }
4294 
4295         /* Do skb overlap to previous one? */
4296         if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4297                 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4298                         /* All the bits are present. Drop. */
4299                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4300                         __kfree_skb(skb);
4301                         skb = NULL;
4302                         tcp_dsack_set(sk, seq, end_seq);
4303                         goto add_sack;
4304                 }
4305                 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4306                         /* Partial overlap. */
4307                         tcp_dsack_set(sk, seq,
4308                                       TCP_SKB_CB(skb1)->end_seq);
4309                 } else {
4310                         if (skb_queue_is_first(&tp->out_of_order_queue,
4311                                                skb1))
4312                                 skb1 = NULL;
4313                         else
4314                                 skb1 = skb_queue_prev(
4315                                         &tp->out_of_order_queue,
4316                                         skb1);
4317                 }
4318         }
4319         if (!skb1)
4320                 __skb_queue_head(&tp->out_of_order_queue, skb);
4321         else
4322                 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4323 
4324         /* And clean segments covered by new one as whole. */
4325         while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4326                 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4327 
4328                 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4329                         break;
4330                 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4331                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4332                                          end_seq);
4333                         break;
4334                 }
4335                 __skb_unlink(skb1, &tp->out_of_order_queue);
4336                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4337                                  TCP_SKB_CB(skb1)->end_seq);
4338                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4339                 __kfree_skb(skb1);
4340         }
4341 
4342 add_sack:
4343         if (tcp_is_sack(tp))
4344                 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4345 end:
4346         if (skb) {
4347                 tcp_grow_window(sk, skb);
4348                 skb_set_owner_r(skb, sk);
4349         }
4350 }
4351 
4352 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4353                   bool *fragstolen)
4354 {
4355         int eaten;
4356         struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4357 
4358         __skb_pull(skb, hdrlen);
4359         eaten = (tail &&
4360                  tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4361         tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4362         if (!eaten) {
4363                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4364                 skb_set_owner_r(skb, sk);
4365         }
4366         return eaten;
4367 }
4368 
4369 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4370 {
4371         struct sk_buff *skb;
4372         bool fragstolen;
4373 
4374         if (size == 0)
4375                 return 0;
4376 
4377         skb = alloc_skb(size, sk->sk_allocation);
4378         if (!skb)
4379                 goto err;
4380 
4381         if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4382                 goto err_free;
4383 
4384         if (memcpy_from_msg(skb_put(skb, size), msg, size))
4385                 goto err_free;
4386 
4387         TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4388         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4389         TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4390 
4391         if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4392                 WARN_ON_ONCE(fragstolen); /* should not happen */
4393                 __kfree_skb(skb);
4394         }
4395         return size;
4396 
4397 err_free:
4398         kfree_skb(skb);
4399 err:
4400         return -ENOMEM;
4401 }
4402 
4403 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4404 {
4405         struct tcp_sock *tp = tcp_sk(sk);
4406         int eaten = -1;
4407         bool fragstolen = false;
4408 
4409         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4410                 goto drop;
4411 
4412         skb_dst_drop(skb);
4413         __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4414 
4415         tcp_ecn_accept_cwr(tp, skb);
4416 
4417         tp->rx_opt.dsack = 0;
4418 
4419         /*  Queue data for delivery to the user.
4420          *  Packets in sequence go to the receive queue.
4421          *  Out of sequence packets to the out_of_order_queue.
4422          */
4423         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4424                 if (tcp_receive_window(tp) == 0)
4425                         goto out_of_window;
4426 
4427                 /* Ok. In sequence. In window. */
4428                 if (tp->ucopy.task == current &&
4429                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4430                     sock_owned_by_user(sk) && !tp->urg_data) {
4431                         int chunk = min_t(unsigned int, skb->len,
4432                                           tp->ucopy.len);
4433 
4434                         __set_current_state(TASK_RUNNING);
4435 
4436                         local_bh_enable();
4437                         if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4438                                 tp->ucopy.len -= chunk;
4439                                 tp->copied_seq += chunk;
4440                                 eaten = (chunk == skb->len);
4441                                 tcp_rcv_space_adjust(sk);
4442                         }
4443                         local_bh_disable();
4444                 }
4445 
4446                 if (eaten <= 0) {
4447 queue_and_out:
4448                         if (eaten < 0 &&
4449                             tcp_try_rmem_schedule(sk, skb, skb->truesize))
4450                                 goto drop;
4451 
4452                         eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4453                 }
4454                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4455                 if (skb->len)
4456                         tcp_event_data_recv(sk, skb);
4457                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4458                         tcp_fin(sk);
4459 
4460                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4461                         tcp_ofo_queue(sk);
4462 
4463                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4464                          * gap in queue is filled.
4465                          */
4466                         if (skb_queue_empty(&tp->out_of_order_queue))
4467                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4468                 }
4469 
4470                 if (tp->rx_opt.num_sacks)
4471                         tcp_sack_remove(tp);
4472 
4473                 tcp_fast_path_check(sk);
4474 
4475                 if (eaten > 0)
4476                         kfree_skb_partial(skb, fragstolen);
4477                 if (!sock_flag(sk, SOCK_DEAD))
4478                         sk->sk_data_ready(sk);
4479                 return;
4480         }
4481 
4482         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4483                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4484                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4485                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4486 
4487 out_of_window:
4488                 tcp_enter_quickack_mode(sk);
4489                 inet_csk_schedule_ack(sk);
4490 drop:
4491                 __kfree_skb(skb);
4492                 return;
4493         }
4494 
4495         /* Out of window. F.e. zero window probe. */
4496         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4497                 goto out_of_window;
4498 
4499         tcp_enter_quickack_mode(sk);
4500 
4501         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4502                 /* Partial packet, seq < rcv_next < end_seq */
4503                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4504                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4505                            TCP_SKB_CB(skb)->end_seq);
4506 
4507                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4508 
4509                 /* If window is closed, drop tail of packet. But after
4510                  * remembering D-SACK for its head made in previous line.
4511                  */
4512                 if (!tcp_receive_window(tp))
4513                         goto out_of_window;
4514                 goto queue_and_out;
4515         }
4516 
4517         tcp_data_queue_ofo(sk, skb);
4518 }
4519 
4520 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4521                                         struct sk_buff_head *list)
4522 {
4523         struct sk_buff *next = NULL;
4524 
4525         if (!skb_queue_is_last(list, skb))
4526                 next = skb_queue_next(list, skb);
4527 
4528         __skb_unlink(skb, list);
4529         __kfree_skb(skb);
4530         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4531 
4532         return next;
4533 }
4534 
4535 /* Collapse contiguous sequence of skbs head..tail with
4536  * sequence numbers start..end.
4537  *
4538  * If tail is NULL, this means until the end of the list.
4539  *
4540  * Segments with FIN/SYN are not collapsed (only because this
4541  * simplifies code)
4542  */
4543 static void
4544 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4545              struct sk_buff *head, struct sk_buff *tail,
4546              u32 start, u32 end)
4547 {
4548         struct sk_buff *skb, *n;
4549         bool end_of_skbs;
4550 
4551         /* First, check that queue is collapsible and find
4552          * the point where collapsing can be useful. */
4553         skb = head;
4554 restart:
4555         end_of_skbs = true;
4556         skb_queue_walk_from_safe(list, skb, n) {
4557                 if (skb == tail)
4558                         break;
4559                 /* No new bits? It is possible on ofo queue. */
4560                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4561                         skb = tcp_collapse_one(sk, skb, list);
4562                         if (!skb)
4563                                 break;
4564                         goto restart;
4565                 }
4566 
4567                 /* The first skb to collapse is:
4568                  * - not SYN/FIN and
4569                  * - bloated or contains data before "start" or
4570                  *   overlaps to the next one.
4571                  */
4572                 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4573                     (tcp_win_from_space(skb->truesize) > skb->len ||
4574                      before(TCP_SKB_CB(skb)->seq, start))) {
4575                         end_of_skbs = false;
4576                         break;
4577                 }
4578 
4579                 if (!skb_queue_is_last(list, skb)) {
4580                         struct sk_buff *next = skb_queue_next(list, skb);
4581                         if (next != tail &&
4582                             TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4583                                 end_of_skbs = false;
4584                                 break;
4585                         }
4586                 }
4587 
4588                 /* Decided to skip this, advance start seq. */
4589                 start = TCP_SKB_CB(skb)->end_seq;
4590         }
4591         if (end_of_skbs ||
4592             (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4593                 return;
4594 
4595         while (before(start, end)) {
4596                 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4597                 struct sk_buff *nskb;
4598 
4599                 nskb = alloc_skb(copy, GFP_ATOMIC);
4600                 if (!nskb)
4601                         return;
4602 
4603                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4604                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4605                 __skb_queue_before(list, skb, nskb);
4606                 skb_set_owner_r(nskb, sk);
4607 
4608                 /* Copy data, releasing collapsed skbs. */
4609                 while (copy > 0) {
4610                         int offset = start - TCP_SKB_CB(skb)->seq;
4611                         int size = TCP_SKB_CB(skb)->end_seq - start;
4612 
4613                         BUG_ON(offset < 0);
4614                         if (size > 0) {
4615                                 size = min(copy, size);
4616                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4617                                         BUG();
4618                                 TCP_SKB_CB(nskb)->end_seq += size;
4619                                 copy -= size;
4620                                 start += size;
4621                         }
4622                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4623                                 skb = tcp_collapse_one(sk, skb, list);
4624                                 if (!skb ||
4625                                     skb == tail ||
4626                                     (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4627                                         return;
4628                         }
4629                 }
4630         }
4631 }
4632 
4633 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4634  * and tcp_collapse() them until all the queue is collapsed.
4635  */
4636 static void tcp_collapse_ofo_queue(struct sock *sk)
4637 {
4638         struct tcp_sock *tp = tcp_sk(sk);
4639         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4640         struct sk_buff *head;
4641         u32 start, end;
4642 
4643         if (skb == NULL)
4644                 return;
4645 
4646         start = TCP_SKB_CB(skb)->seq;
4647         end = TCP_SKB_CB(skb)->end_seq;
4648         head = skb;
4649 
4650         for (;;) {
4651                 struct sk_buff *next = NULL;
4652 
4653                 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4654                         next = skb_queue_next(&tp->out_of_order_queue, skb);
4655                 skb = next;
4656 
4657                 /* Segment is terminated when we see gap or when
4658                  * we are at the end of all the queue. */
4659                 if (!skb ||
4660                     after(TCP_SKB_CB(skb)->seq, end) ||
4661                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4662                         tcp_collapse(sk, &tp->out_of_order_queue,
4663                                      head, skb, start, end);
4664                         head = skb;
4665                         if (!skb)
4666                                 break;
4667                         /* Start new segment */
4668                         start = TCP_SKB_CB(skb)->seq;
4669                         end = TCP_SKB_CB(skb)->end_seq;
4670                 } else {
4671                         if (before(TCP_SKB_CB(skb)->seq, start))
4672                                 start = TCP_SKB_CB(skb)->seq;
4673                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4674                                 end = TCP_SKB_CB(skb)->end_seq;
4675                 }
4676         }
4677 }
4678 
4679 /*
4680  * Purge the out-of-order queue.
4681  * Return true if queue was pruned.
4682  */
4683 static bool tcp_prune_ofo_queue(struct sock *sk)
4684 {
4685         struct tcp_sock *tp = tcp_sk(sk);
4686         bool res = false;
4687 
4688         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4689                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4690                 __skb_queue_purge(&tp->out_of_order_queue);
4691 
4692                 /* Reset SACK state.  A conforming SACK implementation will
4693                  * do the same at a timeout based retransmit.  When a connection
4694                  * is in a sad state like this, we care only about integrity
4695                  * of the connection not performance.
4696                  */
4697                 if (tp->rx_opt.sack_ok)
4698                         tcp_sack_reset(&tp->rx_opt);
4699                 sk_mem_reclaim(sk);
4700                 res = true;
4701         }
4702         return res;
4703 }
4704 
4705 /* Reduce allocated memory if we can, trying to get
4706  * the socket within its memory limits again.
4707  *
4708  * Return less than zero if we should start dropping frames
4709  * until the socket owning process reads some of the data
4710  * to stabilize the situation.
4711  */
4712 static int tcp_prune_queue(struct sock *sk)
4713 {
4714         struct tcp_sock *tp = tcp_sk(sk);
4715 
4716         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4717 
4718         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4719 
4720         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4721                 tcp_clamp_window(sk);
4722         else if (sk_under_memory_pressure(sk))
4723                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4724 
4725         tcp_collapse_ofo_queue(sk);
4726         if (!skb_queue_empty(&sk->sk_receive_queue))
4727                 tcp_collapse(sk, &sk->sk_receive_queue,
4728                              skb_peek(&sk->sk_receive_queue),
4729                              NULL,
4730                              tp->copied_seq, tp->rcv_nxt);
4731         sk_mem_reclaim(sk);
4732 
4733         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4734                 return 0;
4735 
4736         /* Collapsing did not help, destructive actions follow.
4737          * This must not ever occur. */
4738 
4739         tcp_prune_ofo_queue(sk);
4740 
4741         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4742                 return 0;
4743 
4744         /* If we are really being abused, tell the caller to silently
4745          * drop receive data on the floor.  It will get retransmitted
4746          * and hopefully then we'll have sufficient space.
4747          */
4748         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4749 
4750         /* Massive buffer overcommit. */
4751         tp->pred_flags = 0;
4752         return -1;
4753 }
4754 
4755 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4756 {
4757         const struct tcp_sock *tp = tcp_sk(sk);
4758 
4759         /* If the user specified a specific send buffer setting, do
4760          * not modify it.
4761          */
4762         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4763                 return false;
4764 
4765         /* If we are under global TCP memory pressure, do not expand.  */
4766         if (sk_under_memory_pressure(sk))
4767                 return false;
4768 
4769         /* If we are under soft global TCP memory pressure, do not expand.  */
4770         if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4771                 return false;
4772 
4773         /* If we filled the congestion window, do not expand.  */
4774         if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4775                 return false;
4776 
4777         return true;
4778 }
4779 
4780 /* When incoming ACK allowed to free some skb from write_queue,
4781  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4782  * on the exit from tcp input handler.
4783  *
4784  * PROBLEM: sndbuf expansion does not work well with largesend.
4785  */
4786 static void tcp_new_space(struct sock *sk)
4787 {
4788         struct tcp_sock *tp = tcp_sk(sk);
4789 
4790         if (tcp_should_expand_sndbuf(sk)) {
4791                 tcp_sndbuf_expand(sk);
4792                 tp->snd_cwnd_stamp = tcp_time_stamp;
4793         }
4794 
4795         sk->sk_write_space(sk);
4796 }
4797 
4798 static void tcp_check_space(struct sock *sk)
4799 {
4800         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4801                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4802                 if (sk->sk_socket &&
4803                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4804                         tcp_new_space(sk);
4805         }
4806 }
4807 
4808 static inline void tcp_data_snd_check(struct sock *sk)
4809 {
4810         tcp_push_pending_frames(sk);
4811         tcp_check_space(sk);
4812 }
4813 
4814 /*
4815  * Check if sending an ack is needed.
4816  */
4817 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4818 {
4819         struct tcp_sock *tp = tcp_sk(sk);
4820 
4821             /* More than one full frame received... */
4822         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4823              /* ... and right edge of window advances far enough.
4824               * (tcp_recvmsg() will send ACK otherwise). Or...
4825               */
4826              __tcp_select_window(sk) >= tp->rcv_wnd) ||
4827             /* We ACK each frame or... */
4828             tcp_in_quickack_mode(sk) ||
4829             /* We have out of order data. */
4830             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4831                 /* Then ack it now */
4832                 tcp_send_ack(sk);
4833         } else {
4834                 /* Else, send delayed ack. */
4835                 tcp_send_delayed_ack(sk);
4836         }
4837 }
4838 
4839 static inline void tcp_ack_snd_check(struct sock *sk)
4840 {
4841         if (!inet_csk_ack_scheduled(sk)) {
4842                 /* We sent a data segment already. */
4843                 return;
4844         }
4845         __tcp_ack_snd_check(sk, 1);
4846 }
4847 
4848 /*
4849  *      This routine is only called when we have urgent data
4850  *      signaled. Its the 'slow' part of tcp_urg. It could be
4851  *      moved inline now as tcp_urg is only called from one
4852  *      place. We handle URGent data wrong. We have to - as
4853  *      BSD still doesn't use the correction from RFC961.
4854  *      For 1003.1g we should support a new option TCP_STDURG to permit
4855  *      either form (or just set the sysctl tcp_stdurg).
4856  */
4857 
4858 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4859 {
4860         struct tcp_sock *tp = tcp_sk(sk);
4861         u32 ptr = ntohs(th->urg_ptr);
4862 
4863         if (ptr && !sysctl_tcp_stdurg)
4864                 ptr--;
4865         ptr += ntohl(th->seq);
4866 
4867         /* Ignore urgent data that we've already seen and read. */
4868         if (after(tp->copied_seq, ptr))
4869                 return;
4870 
4871         /* Do not replay urg ptr.
4872          *
4873          * NOTE: interesting situation not covered by specs.
4874          * Misbehaving sender may send urg ptr, pointing to segment,
4875          * which we already have in ofo queue. We are not able to fetch
4876          * such data and will stay in TCP_URG_NOTYET until will be eaten
4877          * by recvmsg(). Seems, we are not obliged to handle such wicked
4878          * situations. But it is worth to think about possibility of some
4879          * DoSes using some hypothetical application level deadlock.
4880          */
4881         if (before(ptr, tp->rcv_nxt))
4882                 return;
4883 
4884         /* Do we already have a newer (or duplicate) urgent pointer? */
4885         if (tp->urg_data && !after(ptr, tp->urg_seq))
4886                 return;
4887 
4888         /* Tell the world about our new urgent pointer. */
4889         sk_send_sigurg(sk);
4890 
4891         /* We may be adding urgent data when the last byte read was
4892          * urgent. To do this requires some care. We cannot just ignore
4893          * tp->copied_seq since we would read the last urgent byte again
4894          * as data, nor can we alter copied_seq until this data arrives
4895          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4896          *
4897          * NOTE. Double Dutch. Rendering to plain English: author of comment
4898          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4899          * and expect that both A and B disappear from stream. This is _wrong_.
4900          * Though this happens in BSD with high probability, this is occasional.
4901          * Any application relying on this is buggy. Note also, that fix "works"
4902          * only in this artificial test. Insert some normal data between A and B and we will
4903          * decline of BSD again. Verdict: it is better to remove to trap
4904          * buggy users.
4905          */
4906         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4907             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4908                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4909                 tp->copied_seq++;
4910                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4911                         __skb_unlink(skb, &sk->sk_receive_queue);
4912                         __kfree_skb(skb);
4913                 }
4914         }
4915 
4916         tp->urg_data = TCP_URG_NOTYET;
4917         tp->urg_seq = ptr;
4918 
4919         /* Disable header prediction. */
4920         tp->pred_flags = 0;
4921 }
4922 
4923 /* This is the 'fast' part of urgent handling. */
4924 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4925 {
4926         struct tcp_sock *tp = tcp_sk(sk);
4927 
4928         /* Check if we get a new urgent pointer - normally not. */
4929         if (th->urg)
4930                 tcp_check_urg(sk, th);
4931 
4932         /* Do we wait for any urgent data? - normally not... */
4933         if (tp->urg_data == TCP_URG_NOTYET) {
4934                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4935                           th->syn;
4936 
4937                 /* Is the urgent pointer pointing into this packet? */
4938                 if (ptr < skb->len) {
4939                         u8 tmp;
4940                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4941                                 BUG();
4942                         tp->urg_data = TCP_URG_VALID | tmp;
4943                         if (!sock_flag(sk, SOCK_DEAD))
4944                                 sk->sk_data_ready(sk);
4945                 }
4946         }
4947 }
4948 
4949 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4950 {
4951         struct tcp_sock *tp = tcp_sk(sk);
4952         int chunk = skb->len - hlen;
4953         int err;
4954 
4955         local_bh_enable();
4956         if (skb_csum_unnecessary(skb))
4957                 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
4958         else
4959                 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
4960 
4961         if (!err) {
4962                 tp->ucopy.len -= chunk;
4963                 tp->copied_seq += chunk;
4964                 tcp_rcv_space_adjust(sk);
4965         }
4966 
4967         local_bh_disable();
4968         return err;
4969 }
4970 
4971 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4972                                             struct sk_buff *skb)
4973 {
4974         __sum16 result;
4975 
4976         if (sock_owned_by_user(sk)) {
4977                 local_bh_enable();
4978                 result = __tcp_checksum_complete(skb);
4979                 local_bh_disable();
4980         } else {
4981                 result = __tcp_checksum_complete(skb);
4982         }
4983         return result;
4984 }
4985 
4986 static inline bool tcp_checksum_complete_user(struct sock *sk,
4987                                              struct sk_buff *skb)
4988 {
4989         return !skb_csum_unnecessary(skb) &&
4990                __tcp_checksum_complete_user(sk, skb);
4991 }
4992 
4993 /* Does PAWS and seqno based validation of an incoming segment, flags will
4994  * play significant role here.
4995  */
4996 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
4997                                   const struct tcphdr *th, int syn_inerr)
4998 {
4999         struct tcp_sock *tp = tcp_sk(sk);
5000 
5001         /* RFC1323: H1. Apply PAWS check first. */
5002         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5003             tcp_paws_discard(sk, skb)) {
5004                 if (!th->rst) {
5005                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5006                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5007                                                   LINUX_MIB_TCPACKSKIPPEDPAWS,
5008                                                   &tp->last_oow_ack_time))
5009                                 tcp_send_dupack(sk, skb);
5010                         goto discard;
5011                 }
5012                 /* Reset is accepted even if it did not pass PAWS. */
5013         }
5014 
5015         /* Step 1: check sequence number */
5016         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5017                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5018                  * (RST) segments are validated by checking their SEQ-fields."
5019                  * And page 69: "If an incoming segment is not acceptable,
5020                  * an acknowledgment should be sent in reply (unless the RST
5021                  * bit is set, if so drop the segment and return)".
5022                  */
5023                 if (!th->rst) {
5024                         if (th->syn)
5025                                 goto syn_challenge;
5026                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5027                                                   LINUX_MIB_TCPACKSKIPPEDSEQ,
5028                                                   &tp->last_oow_ack_time))
5029                                 tcp_send_dupack(sk, skb);
5030                 }
5031                 goto discard;
5032         }
5033 
5034         /* Step 2: check RST bit */
5035         if (th->rst) {
5036                 /* RFC 5961 3.2 :
5037                  * If sequence number exactly matches RCV.NXT, then
5038                  *     RESET the connection
5039                  * else
5040                  *     Send a challenge ACK
5041                  */
5042                 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5043                         tcp_reset(sk);
5044                 else
5045                         tcp_send_challenge_ack(sk, skb);
5046                 goto discard;
5047         }
5048 
5049         /* step 3: check security and precedence [ignored] */
5050 
5051         /* step 4: Check for a SYN
5052          * RFC 5961 4.2 : Send a challenge ack
5053          */
5054         if (th->syn) {
5055 syn_challenge:
5056                 if (syn_inerr)
5057                         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5058                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5059                 tcp_send_challenge_ack(sk, skb);
5060                 goto discard;
5061         }
5062 
5063         return true;
5064 
5065 discard:
5066         __kfree_skb(skb);
5067         return false;
5068 }
5069 
5070 /*
5071  *      TCP receive function for the ESTABLISHED state.
5072  *
5073  *      It is split into a fast path and a slow path. The fast path is
5074  *      disabled when:
5075  *      - A zero window was announced from us - zero window probing
5076  *        is only handled properly in the slow path.
5077  *      - Out of order segments arrived.
5078  *      - Urgent data is expected.
5079  *      - There is no buffer space left
5080  *      - Unexpected TCP flags/window values/header lengths are received
5081  *        (detected by checking the TCP header against pred_flags)
5082  *      - Data is sent in both directions. Fast path only supports pure senders
5083  *        or pure receivers (this means either the sequence number or the ack
5084  *        value must stay constant)
5085  *      - Unexpected TCP option.
5086  *
5087  *      When these conditions are not satisfied it drops into a standard
5088  *      receive procedure patterned after RFC793 to handle all cases.
5089  *      The first three cases are guaranteed by proper pred_flags setting,
5090  *      the rest is checked inline. Fast processing is turned on in
5091  *      tcp_data_queue when everything is OK.
5092  */
5093 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5094                          const struct tcphdr *th, unsigned int len)
5095 {
5096         struct tcp_sock *tp = tcp_sk(sk);
5097 
5098         if (unlikely(sk->sk_rx_dst == NULL))
5099                 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5100         /*
5101          *      Header prediction.
5102          *      The code loosely follows the one in the famous
5103          *      "30 instruction TCP receive" Van Jacobson mail.
5104          *
5105          *      Van's trick is to deposit buffers into socket queue
5106          *      on a device interrupt, to call tcp_recv function
5107          *      on the receive process context and checksum and copy
5108          *      the buffer to user space. smart...
5109          *
5110          *      Our current scheme is not silly either but we take the
5111          *      extra cost of the net_bh soft interrupt processing...
5112          *      We do checksum and copy also but from device to kernel.
5113          */
5114 
5115         tp->rx_opt.saw_tstamp = 0;
5116 
5117         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5118          *      if header_prediction is to be made
5119          *      'S' will always be tp->tcp_header_len >> 2
5120          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5121          *  turn it off (when there are holes in the receive
5122          *       space for instance)
5123          *      PSH flag is ignored.
5124          */
5125 
5126         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5127             TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5128             !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5129                 int tcp_header_len = tp->tcp_header_len;
5130 
5131                 /* Timestamp header prediction: tcp_header_len
5132                  * is automatically equal to th->doff*4 due to pred_flags
5133                  * match.
5134                  */
5135 
5136                 /* Check timestamp */
5137                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5138                         /* No? Slow path! */
5139                         if (!tcp_parse_aligned_timestamp(tp, th))
5140                                 goto slow_path;
5141 
5142                         /* If PAWS failed, check it more carefully in slow path */
5143                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5144                                 goto slow_path;
5145 
5146                         /* DO NOT update ts_recent here, if checksum fails
5147                          * and timestamp was corrupted part, it will result
5148                          * in a hung connection since we will drop all
5149                          * future packets due to the PAWS test.
5150                          */
5151                 }
5152 
5153                 if (len <= tcp_header_len) {
5154                         /* Bulk data transfer: sender */
5155                         if (len == tcp_header_len) {
5156                                 /* Predicted packet is in window by definition.
5157                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5158                                  * Hence, check seq<=rcv_wup reduces to:
5159                                  */
5160                                 if (tcp_header_len ==
5161                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5162                                     tp->rcv_nxt == tp->rcv_wup)
5163                                         tcp_store_ts_recent(tp);
5164 
5165                                 /* We know that such packets are checksummed
5166                                  * on entry.
5167                                  */
5168                                 tcp_ack(sk, skb, 0);
5169                                 __kfree_skb(skb);
5170                                 tcp_data_snd_check(sk);
5171                                 return;
5172                         } else { /* Header too small */
5173                                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5174                                 goto discard;
5175                         }
5176                 } else {
5177                         int eaten = 0;
5178                         bool fragstolen = false;
5179 
5180                         if (tp->ucopy.task == current &&
5181                             tp->copied_seq == tp->rcv_nxt &&
5182                             len - tcp_header_len <= tp->ucopy.len &&
5183                             sock_owned_by_user(sk)) {
5184                                 __set_current_state(TASK_RUNNING);
5185 
5186                                 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5187                                         /* Predicted packet is in window by definition.
5188                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5189                                          * Hence, check seq<=rcv_wup reduces to:
5190                                          */
5191                                         if (tcp_header_len ==
5192                                             (sizeof(struct tcphdr) +
5193                                              TCPOLEN_TSTAMP_ALIGNED) &&
5194                                             tp->rcv_nxt == tp->rcv_wup)
5195                                                 tcp_store_ts_recent(tp);
5196 
5197                                         tcp_rcv_rtt_measure_ts(sk, skb);
5198 
5199                                         __skb_pull(skb, tcp_header_len);
5200                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5201                                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5202                                         eaten = 1;
5203                                 }
5204                         }
5205                         if (!eaten) {
5206                                 if (tcp_checksum_complete_user(sk, skb))
5207                                         goto csum_error;
5208 
5209                                 if ((int)skb->truesize > sk->sk_forward_alloc)
5210                                         goto step5;
5211 
5212                                 /* Predicted packet is in window by definition.
5213                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5214                                  * Hence, check seq<=rcv_wup reduces to:
5215                                  */
5216                                 if (tcp_header_len ==
5217                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5218                                     tp->rcv_nxt == tp->rcv_wup)
5219                                         tcp_store_ts_recent(tp);
5220 
5221                                 tcp_rcv_rtt_measure_ts(sk, skb);
5222 
5223                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5224 
5225                                 /* Bulk data transfer: receiver */
5226                                 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5227                                                       &fragstolen);
5228                         }
5229 
5230                         tcp_event_data_recv(sk, skb);
5231 
5232                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5233                                 /* Well, only one small jumplet in fast path... */
5234                                 tcp_ack(sk, skb, FLAG_DATA);
5235                                 tcp_data_snd_check(sk);
5236                                 if (!inet_csk_ack_scheduled(sk))
5237                                         goto no_ack;
5238                         }
5239 
5240                         __tcp_ack_snd_check(sk, 0);
5241 no_ack:
5242                         if (eaten)
5243                                 kfree_skb_partial(skb, fragstolen);
5244                         sk->sk_data_ready(sk);
5245                         return;
5246                 }
5247         }
5248 
5249 slow_path:
5250         if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5251                 goto csum_error;
5252 
5253         if (!th->ack && !th->rst && !th->syn)
5254                 goto discard;
5255 
5256         /*
5257          *      Standard slow path.
5258          */
5259 
5260         if (!tcp_validate_incoming(sk, skb, th, 1))
5261                 return;
5262 
5263 step5:
5264         if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5265                 goto discard;
5266 
5267         tcp_rcv_rtt_measure_ts(sk, skb);
5268 
5269         /* Process urgent data. */
5270         tcp_urg(sk, skb, th);
5271 
5272         /* step 7: process the segment text */
5273         tcp_data_queue(sk, skb);
5274 
5275         tcp_data_snd_check(sk);
5276         tcp_ack_snd_check(sk);
5277         return;
5278 
5279 csum_error:
5280         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5281         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5282 
5283 discard:
5284         __kfree_skb(skb);
5285 }
5286 EXPORT_SYMBOL(tcp_rcv_established);
5287 
5288 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5289 {
5290         struct tcp_sock *tp = tcp_sk(sk);
5291         struct inet_connection_sock *icsk = inet_csk(sk);
5292 
5293         tcp_set_state(sk, TCP_ESTABLISHED);
5294 
5295         if (skb != NULL) {
5296                 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5297                 security_inet_conn_established(sk, skb);
5298         }
5299 
5300         /* Make sure socket is routed, for correct metrics.  */
5301         icsk->icsk_af_ops->rebuild_header(sk);
5302 
5303         tcp_init_metrics(sk);
5304 
5305         tcp_init_congestion_control(sk);
5306 
5307         /* Prevent spurious tcp_cwnd_restart() on first data
5308          * packet.
5309          */
5310         tp->lsndtime = tcp_time_stamp;
5311 
5312         tcp_init_buffer_space(sk);
5313 
5314         if (sock_flag(sk, SOCK_KEEPOPEN))
5315                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5316 
5317         if (!tp->rx_opt.snd_wscale)
5318                 __tcp_fast_path_on(tp, tp->snd_wnd);
5319         else
5320                 tp->pred_flags = 0;
5321 
5322         if (!sock_flag(sk, SOCK_DEAD)) {
5323                 sk->sk_state_change(sk);
5324                 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5325         }
5326 }
5327 
5328 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5329                                     struct tcp_fastopen_cookie *cookie)
5330 {
5331         struct tcp_sock *tp = tcp_sk(sk);
5332         struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5333         u16 mss = tp->rx_opt.mss_clamp;
5334         bool syn_drop;
5335 
5336         if (mss == tp->rx_opt.user_mss) {
5337                 struct tcp_options_received opt;
5338 
5339                 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5340                 tcp_clear_options(&opt);
5341                 opt.user_mss = opt.mss_clamp = 0;
5342                 tcp_parse_options(synack, &opt, 0, NULL);
5343                 mss = opt.mss_clamp;
5344         }
5345 
5346         if (!tp->syn_fastopen)  /* Ignore an unsolicited cookie */
5347                 cookie->len = -1;
5348 
5349         /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5350          * the remote receives only the retransmitted (regular) SYNs: either
5351          * the original SYN-data or the corresponding SYN-ACK is lost.
5352          */
5353         syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
5354 
5355         tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
5356 
5357         if (data) { /* Retransmit unacked data in SYN */
5358                 tcp_for_write_queue_from(data, sk) {
5359                         if (data == tcp_send_head(sk) ||
5360                             __tcp_retransmit_skb(sk, data))
5361                                 break;
5362                 }
5363                 tcp_rearm_rto(sk);
5364                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5365                 return true;
5366         }
5367         tp->syn_data_acked = tp->syn_data;
5368         if (tp->syn_data_acked)
5369                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5370         return false;
5371 }
5372 
5373 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5374                                          const struct tcphdr *th, unsigned int len)
5375 {
5376         struct inet_connection_sock *icsk = inet_csk(sk);
5377         struct tcp_sock *tp = tcp_sk(sk);
5378         struct tcp_fastopen_cookie foc = { .len = -1 };
5379         int saved_clamp = tp->rx_opt.mss_clamp;
5380 
5381         tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5382         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5383                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5384 
5385         if (th->ack) {
5386                 /* rfc793:
5387                  * "If the state is SYN-SENT then
5388                  *    first check the ACK bit
5389                  *      If the ACK bit is set
5390                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5391                  *        a reset (unless the RST bit is set, if so drop
5392                  *        the segment and return)"
5393                  */
5394                 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5395                     after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5396                         goto reset_and_undo;
5397 
5398                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5399                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5400                              tcp_time_stamp)) {
5401                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5402                         goto reset_and_undo;
5403                 }
5404 
5405                 /* Now ACK is acceptable.
5406                  *
5407                  * "If the RST bit is set
5408                  *    If the ACK was acceptable then signal the user "error:
5409                  *    connection reset", drop the segment, enter CLOSED state,
5410                  *    delete TCB, and return."
5411                  */
5412 
5413                 if (th->rst) {
5414                         tcp_reset(sk);
5415                         goto discard;
5416                 }
5417 
5418                 /* rfc793:
5419                  *   "fifth, if neither of the SYN or RST bits is set then
5420                  *    drop the segment and return."
5421                  *
5422                  *    See note below!
5423                  *                                        --ANK(990513)
5424                  */
5425                 if (!th->syn)
5426                         goto discard_and_undo;
5427 
5428                 /* rfc793:
5429                  *   "If the SYN bit is on ...
5430                  *    are acceptable then ...
5431                  *    (our SYN has been ACKed), change the connection
5432                  *    state to ESTABLISHED..."
5433                  */
5434 
5435                 tcp_ecn_rcv_synack(tp, th);
5436 
5437                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5438                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5439 
5440                 /* Ok.. it's good. Set up sequence numbers and
5441                  * move to established.
5442                  */
5443                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5444                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5445 
5446                 /* RFC1323: The window in SYN & SYN/ACK segments is
5447                  * never scaled.
5448                  */
5449                 tp->snd_wnd = ntohs(th->window);
5450 
5451                 if (!tp->rx_opt.wscale_ok) {
5452                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5453                         tp->window_clamp = min(tp->window_clamp, 65535U);
5454                 }
5455 
5456                 if (tp->rx_opt.saw_tstamp) {
5457                         tp->rx_opt.tstamp_ok       = 1;
5458                         tp->tcp_header_len =
5459                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5460                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5461                         tcp_store_ts_recent(tp);
5462                 } else {
5463                         tp->tcp_header_len = sizeof(struct tcphdr);
5464                 }
5465 
5466                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5467                         tcp_enable_fack(tp);
5468 
5469                 tcp_mtup_init(sk);
5470                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5471                 tcp_initialize_rcv_mss(sk);
5472 
5473                 /* Remember, tcp_poll() does not lock socket!
5474                  * Change state from SYN-SENT only after copied_seq
5475                  * is initialized. */
5476                 tp->copied_seq = tp->rcv_nxt;
5477 
5478                 smp_mb();
5479 
5480                 tcp_finish_connect(sk, skb);
5481 
5482                 if ((tp->syn_fastopen || tp->syn_data) &&
5483                     tcp_rcv_fastopen_synack(sk, skb, &foc))
5484                         return -1;
5485 
5486                 if (sk->sk_write_pending ||
5487                     icsk->icsk_accept_queue.rskq_defer_accept ||
5488                     icsk->icsk_ack.pingpong) {
5489                         /* Save one ACK. Data will be ready after
5490                          * several ticks, if write_pending is set.
5491                          *
5492                          * It may be deleted, but with this feature tcpdumps
5493                          * look so _wonderfully_ clever, that I was not able
5494                          * to stand against the temptation 8)     --ANK
5495                          */
5496                         inet_csk_schedule_ack(sk);
5497                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5498                         tcp_enter_quickack_mode(sk);
5499                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5500                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5501 
5502 discard:
5503                         __kfree_skb(skb);
5504                         return 0;
5505                 } else {
5506                         tcp_send_ack(sk);
5507                 }
5508                 return -1;
5509         }
5510 
5511         /* No ACK in the segment */
5512 
5513         if (th->rst) {
5514                 /* rfc793:
5515                  * "If the RST bit is set
5516                  *
5517                  *      Otherwise (no ACK) drop the segment and return."
5518                  */
5519 
5520                 goto discard_and_undo;
5521         }
5522 
5523         /* PAWS check. */
5524         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5525             tcp_paws_reject(&tp->rx_opt, 0))
5526                 goto discard_and_undo;
5527 
5528         if (th->syn) {
5529                 /* We see SYN without ACK. It is attempt of
5530                  * simultaneous connect with crossed SYNs.
5531                  * Particularly, it can be connect to self.
5532                  */
5533                 tcp_set_state(sk, TCP_SYN_RECV);
5534 
5535                 if (tp->rx_opt.saw_tstamp) {
5536                         tp->rx_opt.tstamp_ok = 1;
5537                         tcp_store_ts_recent(tp);
5538                         tp->tcp_header_len =
5539                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5540                 } else {
5541                         tp->tcp_header_len = sizeof(struct tcphdr);
5542                 }
5543 
5544                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5545                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5546 
5547                 /* RFC1323: The window in SYN & SYN/ACK segments is
5548                  * never scaled.
5549                  */
5550                 tp->snd_wnd    = ntohs(th->window);
5551                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5552                 tp->max_window = tp->snd_wnd;
5553 
5554                 tcp_ecn_rcv_syn(tp, th);
5555 
5556                 tcp_mtup_init(sk);
5557                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5558                 tcp_initialize_rcv_mss(sk);
5559 
5560                 tcp_send_synack(sk);
5561 #if 0
5562                 /* Note, we could accept data and URG from this segment.
5563                  * There are no obstacles to make this (except that we must
5564                  * either change tcp_recvmsg() to prevent it from returning data
5565                  * before 3WHS completes per RFC793, or employ TCP Fast Open).
5566                  *
5567                  * However, if we ignore data in ACKless segments sometimes,
5568                  * we have no reasons to accept it sometimes.
5569                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5570                  * is not flawless. So, discard packet for sanity.
5571                  * Uncomment this return to process the data.
5572                  */
5573                 return -1;
5574 #else
5575                 goto discard;
5576 #endif
5577         }
5578         /* "fifth, if neither of the SYN or RST bits is set then
5579          * drop the segment and return."
5580          */
5581 
5582 discard_and_undo:
5583         tcp_clear_options(&tp->rx_opt);
5584         tp->rx_opt.mss_clamp = saved_clamp;
5585         goto discard;
5586 
5587 reset_and_undo:
5588         tcp_clear_options(&tp->rx_opt);
5589         tp->rx_opt.mss_clamp = saved_clamp;
5590         return 1;
5591 }
5592 
5593 /*
5594  *      This function implements the receiving procedure of RFC 793 for
5595  *      all states except ESTABLISHED and TIME_WAIT.
5596  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5597  *      address independent.
5598  */
5599 
5600 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5601                           const struct tcphdr *th, unsigned int len)
5602 {
5603         struct tcp_sock *tp = tcp_sk(sk);
5604         struct inet_connection_sock *icsk = inet_csk(sk);
5605         struct request_sock *req;
5606         int queued = 0;
5607         bool acceptable;
5608         u32 synack_stamp;
5609 
5610         tp->rx_opt.saw_tstamp = 0;
5611 
5612         switch (sk->sk_state) {
5613         case TCP_CLOSE:
5614                 goto discard;
5615 
5616         case TCP_LISTEN:
5617                 if (th->ack)
5618                         return 1;
5619 
5620                 if (th->rst)
5621                         goto discard;
5622 
5623                 if (th->syn) {
5624                         if (th->fin)
5625                                 goto discard;
5626                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5627                                 return 1;
5628 
5629                         /* Now we have several options: In theory there is
5630                          * nothing else in the frame. KA9Q has an option to
5631                          * send data with the syn, BSD accepts data with the
5632                          * syn up to the [to be] advertised window and
5633                          * Solaris 2.1 gives you a protocol error. For now
5634                          * we just ignore it, that fits the spec precisely
5635                          * and avoids incompatibilities. It would be nice in
5636                          * future to drop through and process the data.
5637                          *
5638                          * Now that TTCP is starting to be used we ought to
5639                          * queue this data.
5640                          * But, this leaves one open to an easy denial of
5641                          * service attack, and SYN cookies can't defend
5642                          * against this problem. So, we drop the data
5643                          * in the interest of security over speed unless
5644                          * it's still in use.
5645                          */
5646                         kfree_skb(skb);
5647                         return 0;
5648                 }
5649                 goto discard;
5650 
5651         case TCP_SYN_SENT:
5652                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5653                 if (queued >= 0)
5654                         return queued;
5655 
5656                 /* Do step6 onward by hand. */
5657                 tcp_urg(sk, skb, th);
5658                 __kfree_skb(skb);
5659                 tcp_data_snd_check(sk);
5660                 return 0;
5661         }
5662 
5663         req = tp->fastopen_rsk;
5664         if (req != NULL) {
5665                 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5666                     sk->sk_state != TCP_FIN_WAIT1);
5667 
5668                 if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
5669                         goto discard;
5670         }
5671 
5672         if (!th->ack && !th->rst && !th->syn)
5673                 goto discard;
5674 
5675         if (!tcp_validate_incoming(sk, skb, th, 0))
5676                 return 0;
5677 
5678         /* step 5: check the ACK field */
5679         acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5680                                       FLAG_UPDATE_TS_RECENT) > 0;
5681 
5682         switch (sk->sk_state) {
5683         case TCP_SYN_RECV:
5684                 if (!acceptable)
5685                         return 1;
5686 
5687                 /* Once we leave TCP_SYN_RECV, we no longer need req
5688                  * so release it.
5689                  */
5690                 if (req) {
5691                         synack_stamp = tcp_rsk(req)->snt_synack;
5692                         tp->total_retrans = req->num_retrans;
5693                         reqsk_fastopen_remove(sk, req, false);
5694                 } else {
5695                         synack_stamp = tp->lsndtime;
5696                         /* Make sure socket is routed, for correct metrics. */
5697                         icsk->icsk_af_ops->rebuild_header(sk);
5698                         tcp_init_congestion_control(sk);
5699 
5700                         tcp_mtup_init(sk);
5701                         tp->copied_seq = tp->rcv_nxt;
5702                         tcp_init_buffer_space(sk);
5703                 }
5704                 smp_mb();
5705                 tcp_set_state(sk, TCP_ESTABLISHED);
5706                 sk->sk_state_change(sk);
5707 
5708                 /* Note, that this wakeup is only for marginal crossed SYN case.
5709                  * Passively open sockets are not waked up, because
5710                  * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5711                  */
5712                 if (sk->sk_socket)
5713                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5714 
5715                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5716                 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5717                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5718                 tcp_synack_rtt_meas(sk, synack_stamp);
5719 
5720                 if (tp->rx_opt.tstamp_ok)
5721                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5722 
5723                 if (req) {
5724                         /* Re-arm the timer because data may have been sent out.
5725                          * This is similar to the regular data transmission case
5726                          * when new data has just been ack'ed.
5727                          *
5728                          * (TFO) - we could try to be more aggressive and
5729                          * retransmitting any data sooner based on when they
5730                          * are sent out.
5731                          */
5732                         tcp_rearm_rto(sk);
5733                 } else
5734                         tcp_init_metrics(sk);
5735 
5736                 tcp_update_pacing_rate(sk);
5737 
5738                 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5739                 tp->lsndtime = tcp_time_stamp;
5740 
5741                 tcp_initialize_rcv_mss(sk);
5742                 tcp_fast_path_on(tp);
5743                 break;
5744 
5745         case TCP_FIN_WAIT1: {
5746                 struct dst_entry *dst;
5747                 int tmo;
5748 
5749                 /* If we enter the TCP_FIN_WAIT1 state and we are a
5750                  * Fast Open socket and this is the first acceptable
5751                  * ACK we have received, this would have acknowledged
5752                  * our SYNACK so stop the SYNACK timer.
5753                  */
5754                 if (req != NULL) {
5755                         /* Return RST if ack_seq is invalid.
5756                          * Note that RFC793 only says to generate a
5757                          * DUPACK for it but for TCP Fast Open it seems
5758                          * better to treat this case like TCP_SYN_RECV
5759                          * above.
5760                          */
5761                         if (!acceptable)
5762                                 return 1;
5763                         /* We no longer need the request sock. */
5764                         reqsk_fastopen_remove(sk, req, false);
5765                         tcp_rearm_rto(sk);
5766                 }
5767                 if (tp->snd_una != tp->write_seq)
5768                         break;
5769 
5770                 tcp_set_state(sk, TCP_FIN_WAIT2);
5771                 sk->sk_shutdown |= SEND_SHUTDOWN;
5772 
5773                 dst = __sk_dst_get(sk);
5774                 if (dst)
5775                         dst_confirm(dst);
5776 
5777                 if (!sock_flag(sk, SOCK_DEAD)) {
5778                         /* Wake up lingering close() */
5779                         sk->sk_state_change(sk);
5780                         break;
5781                 }
5782 
5783                 if (tp->linger2 < 0 ||
5784                     (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5785                      after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5786                         tcp_done(sk);
5787                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5788                         return 1;
5789                 }
5790 
5791                 tmo = tcp_fin_time(sk);
5792                 if (tmo > TCP_TIMEWAIT_LEN) {
5793                         inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5794                 } else if (th->fin || sock_owned_by_user(sk)) {
5795                         /* Bad case. We could lose such FIN otherwise.
5796                          * It is not a big problem, but it looks confusing
5797                          * and not so rare event. We still can lose it now,
5798                          * if it spins in bh_lock_sock(), but it is really
5799                          * marginal case.
5800                          */
5801                         inet_csk_reset_keepalive_timer(sk, tmo);
5802                 } else {
5803                         tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5804                         goto discard;
5805                 }
5806                 break;
5807         }
5808 
5809         case TCP_CLOSING:
5810                 if (tp->snd_una == tp->write_seq) {
5811                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5812                         goto discard;
5813                 }
5814                 break;
5815 
5816         case TCP_LAST_ACK:
5817                 if (tp->snd_una == tp->write_seq) {
5818                         tcp_update_metrics(sk);
5819                         tcp_done(sk);
5820                         goto discard;
5821                 }
5822                 break;
5823         }
5824 
5825         /* step 6: check the URG bit */
5826         tcp_urg(sk, skb, th);
5827 
5828         /* step 7: process the segment text */
5829         switch (sk->sk_state) {
5830         case TCP_CLOSE_WAIT:
5831         case TCP_CLOSING:
5832         case TCP_LAST_ACK:
5833                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5834                         break;
5835         case TCP_FIN_WAIT1:
5836         case TCP_FIN_WAIT2:
5837                 /* RFC 793 says to queue data in these states,
5838                  * RFC 1122 says we MUST send a reset.
5839                  * BSD 4.4 also does reset.
5840                  */
5841                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5842                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5843                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5844                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5845                                 tcp_reset(sk);
5846                                 return 1;
5847                         }
5848                 }
5849                 /* Fall through */
5850         case TCP_ESTABLISHED:
5851                 tcp_data_queue(sk, skb);
5852                 queued = 1;
5853                 break;
5854         }
5855 
5856         /* tcp_data could move socket to TIME-WAIT */
5857         if (sk->sk_state != TCP_CLOSE) {
5858                 tcp_data_snd_check(sk);
5859                 tcp_ack_snd_check(sk);
5860         }
5861 
5862         if (!queued) {
5863 discard:
5864                 __kfree_skb(skb);
5865         }
5866         return 0;
5867 }
5868 EXPORT_SYMBOL(tcp_rcv_state_process);
5869 
5870 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
5871 {
5872         struct inet_request_sock *ireq = inet_rsk(req);
5873 
5874         if (family == AF_INET)
5875                 net_dbg_ratelimited("drop open request from %pI4/%u\n",
5876                                     &ireq->ir_rmt_addr, port);
5877 #if IS_ENABLED(CONFIG_IPV6)
5878         else if (family == AF_INET6)
5879                 net_dbg_ratelimited("drop open request from %pI6/%u\n",
5880                                     &ireq->ir_v6_rmt_addr, port);
5881 #endif
5882 }
5883 
5884 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
5885  *
5886  * If we receive a SYN packet with these bits set, it means a
5887  * network is playing bad games with TOS bits. In order to
5888  * avoid possible false congestion notifications, we disable
5889  * TCP ECN negotiation.
5890  *
5891  * Exception: tcp_ca wants ECN. This is required for DCTCP
5892  * congestion control: Linux DCTCP asserts ECT on all packets,
5893  * including SYN, which is most optimal solution; however,
5894  * others, such as FreeBSD do not.
5895  */
5896 static void tcp_ecn_create_request(struct request_sock *req,
5897                                    const struct sk_buff *skb,
5898                                    const struct sock *listen_sk,
5899                                    const struct dst_entry *dst)
5900 {
5901         const struct tcphdr *th = tcp_hdr(skb);
5902         const struct net *net = sock_net(listen_sk);
5903         bool th_ecn = th->ece && th->cwr;
5904         bool ect, ecn_ok;
5905 
5906         if (!th_ecn)
5907                 return;
5908 
5909         ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
5910         ecn_ok = net->ipv4.sysctl_tcp_ecn || dst_feature(dst, RTAX_FEATURE_ECN);
5911 
5912         if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk))
5913                 inet_rsk(req)->ecn_ok = 1;
5914 }
5915 
5916 int tcp_conn_request(struct request_sock_ops *rsk_ops,
5917                      const struct tcp_request_sock_ops *af_ops,
5918                      struct sock *sk, struct sk_buff *skb)
5919 {
5920         struct tcp_options_received tmp_opt;
5921         struct request_sock *req;
5922         struct tcp_sock *tp = tcp_sk(sk);
5923         struct dst_entry *dst = NULL;
5924         __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
5925         bool want_cookie = false, fastopen;
5926         struct flowi fl;
5927         struct tcp_fastopen_cookie foc = { .len = -1 };
5928         int err;
5929 
5930 
5931         /* TW buckets are converted to open requests without
5932          * limitations, they conserve resources and peer is
5933          * evidently real one.
5934          */
5935         if ((sysctl_tcp_syncookies == 2 ||
5936              inet_csk_reqsk_queue_is_full(sk)) && !isn) {
5937                 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
5938                 if (!want_cookie)
5939                         goto drop;
5940         }
5941 
5942 
5943         /* Accept backlog is full. If we have already queued enough
5944          * of warm entries in syn queue, drop request. It is better than
5945          * clogging syn queue with openreqs with exponentially increasing
5946          * timeout.
5947          */
5948         if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
5949                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
5950                 goto drop;
5951         }
5952 
5953         req = inet_reqsk_alloc(rsk_ops);
5954         if (!req)
5955                 goto drop;
5956 
5957         tcp_rsk(req)->af_specific = af_ops;
5958 
5959         tcp_clear_options(&tmp_opt);
5960         tmp_opt.mss_clamp = af_ops->mss_clamp;
5961         tmp_opt.user_mss  = tp->rx_opt.user_mss;
5962         tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
5963 
5964         if (want_cookie && !tmp_opt.saw_tstamp)
5965                 tcp_clear_options(&tmp_opt);
5966 
5967         tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
5968         tcp_openreq_init(req, &tmp_opt, skb, sk);
5969 
5970         af_ops->init_req(req, sk, skb);
5971 
5972         if (security_inet_conn_request(sk, skb, req))
5973                 goto drop_and_free;
5974 
5975         if (!want_cookie && !isn) {
5976                 /* VJ's idea. We save last timestamp seen
5977                  * from the destination in peer table, when entering
5978                  * state TIME-WAIT, and check against it before
5979                  * accepting new connection request.
5980                  *
5981                  * If "isn" is not zero, this request hit alive
5982                  * timewait bucket, so that all the necessary checks
5983                  * are made in the function processing timewait state.
5984                  */
5985                 if (tcp_death_row.sysctl_tw_recycle) {
5986                         bool strict;
5987 
5988                         dst = af_ops->route_req(sk, &fl, req, &strict);
5989 
5990                         if (dst && strict &&
5991                             !tcp_peer_is_proven(req, dst, true,
5992                                                 tmp_opt.saw_tstamp)) {
5993                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
5994                                 goto drop_and_release;
5995                         }
5996                 }
5997                 /* Kill the following clause, if you dislike this way. */
5998                 else if (!sysctl_tcp_syncookies &&
5999                          (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6000                           (sysctl_max_syn_backlog >> 2)) &&
6001                          !tcp_peer_is_proven(req, dst, false,
6002                                              tmp_opt.saw_tstamp)) {
6003                         /* Without syncookies last quarter of
6004                          * backlog is filled with destinations,
6005                          * proven to be alive.
6006                          * It means that we continue to communicate
6007                          * to destinations, already remembered
6008                          * to the moment of synflood.
6009                          */
6010                         pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6011                                     rsk_ops->family);
6012                         goto drop_and_release;
6013                 }
6014 
6015                 isn = af_ops->init_seq(skb);
6016         }
6017         if (!dst) {
6018                 dst = af_ops->route_req(sk, &fl, req, NULL);
6019                 if (!dst)
6020                         goto drop_and_free;
6021         }
6022 
6023         tcp_ecn_create_request(req, skb, sk, dst);
6024 
6025         if (want_cookie) {
6026                 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6027                 req->cookie_ts = tmp_opt.tstamp_ok;
6028                 if (!tmp_opt.tstamp_ok)
6029                         inet_rsk(req)->ecn_ok = 0;
6030         }
6031 
6032         tcp_rsk(req)->snt_isn = isn;
6033         tcp_openreq_init_rwin(req, sk, dst);
6034         fastopen = !want_cookie &&
6035                    tcp_try_fastopen(sk, skb, req, &foc, dst);
6036         err = af_ops->send_synack(sk, dst, &fl, req,
6037                                   skb_get_queue_mapping(skb), &foc);
6038         if (!fastopen) {
6039                 if (err || want_cookie)
6040                         goto drop_and_free;
6041 
6042                 tcp_rsk(req)->listener = NULL;
6043                 af_ops->queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6044         }
6045 
6046         return 0;
6047 
6048 drop_and_release:
6049         dst_release(dst);
6050 drop_and_free:
6051         reqsk_free(req);
6052 drop:
6053         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS);
6054         return 0;
6055 }
6056 EXPORT_SYMBOL(tcp_conn_request);
6057 

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