Version:  2.0.40 2.2.26 2.4.37 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2

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             !tcp_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             !tcp_under_memory_pressure(sk) &&
450             sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
451                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
452                                     sysctl_tcp_rmem[2]);
453         }
454         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
455                 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
456 }
457 
458 /* Initialize RCV_MSS value.
459  * RCV_MSS is an our guess about MSS used by the peer.
460  * We haven't any direct information about the MSS.
461  * It's better to underestimate the RCV_MSS rather than overestimate.
462  * Overestimations make us ACKing less frequently than needed.
463  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
464  */
465 void tcp_initialize_rcv_mss(struct sock *sk)
466 {
467         const struct tcp_sock *tp = tcp_sk(sk);
468         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
469 
470         hint = min(hint, tp->rcv_wnd / 2);
471         hint = min(hint, TCP_MSS_DEFAULT);
472         hint = max(hint, TCP_MIN_MSS);
473 
474         inet_csk(sk)->icsk_ack.rcv_mss = hint;
475 }
476 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
477 
478 /* Receiver "autotuning" code.
479  *
480  * The algorithm for RTT estimation w/o timestamps is based on
481  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
482  * <http://public.lanl.gov/radiant/pubs.html#DRS>
483  *
484  * More detail on this code can be found at
485  * <http://staff.psc.edu/jheffner/>,
486  * though this reference is out of date.  A new paper
487  * is pending.
488  */
489 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
490 {
491         u32 new_sample = tp->rcv_rtt_est.rtt;
492         long m = sample;
493 
494         if (m == 0)
495                 m = 1;
496 
497         if (new_sample != 0) {
498                 /* If we sample in larger samples in the non-timestamp
499                  * case, we could grossly overestimate the RTT especially
500                  * with chatty applications or bulk transfer apps which
501                  * are stalled on filesystem I/O.
502                  *
503                  * Also, since we are only going for a minimum in the
504                  * non-timestamp case, we do not smooth things out
505                  * else with timestamps disabled convergence takes too
506                  * long.
507                  */
508                 if (!win_dep) {
509                         m -= (new_sample >> 3);
510                         new_sample += m;
511                 } else {
512                         m <<= 3;
513                         if (m < new_sample)
514                                 new_sample = m;
515                 }
516         } else {
517                 /* No previous measure. */
518                 new_sample = m << 3;
519         }
520 
521         if (tp->rcv_rtt_est.rtt != new_sample)
522                 tp->rcv_rtt_est.rtt = new_sample;
523 }
524 
525 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
526 {
527         if (tp->rcv_rtt_est.time == 0)
528                 goto new_measure;
529         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
530                 return;
531         tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
532 
533 new_measure:
534         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
535         tp->rcv_rtt_est.time = tcp_time_stamp;
536 }
537 
538 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
539                                           const struct sk_buff *skb)
540 {
541         struct tcp_sock *tp = tcp_sk(sk);
542         if (tp->rx_opt.rcv_tsecr &&
543             (TCP_SKB_CB(skb)->end_seq -
544              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
545                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
546 }
547 
548 /*
549  * This function should be called every time data is copied to user space.
550  * It calculates the appropriate TCP receive buffer space.
551  */
552 void tcp_rcv_space_adjust(struct sock *sk)
553 {
554         struct tcp_sock *tp = tcp_sk(sk);
555         int time;
556         int copied;
557 
558         time = tcp_time_stamp - tp->rcvq_space.time;
559         if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
560                 return;
561 
562         /* Number of bytes copied to user in last RTT */
563         copied = tp->copied_seq - tp->rcvq_space.seq;
564         if (copied <= tp->rcvq_space.space)
565                 goto new_measure;
566 
567         /* A bit of theory :
568          * copied = bytes received in previous RTT, our base window
569          * To cope with packet losses, we need a 2x factor
570          * To cope with slow start, and sender growing its cwin by 100 %
571          * every RTT, we need a 4x factor, because the ACK we are sending
572          * now is for the next RTT, not the current one :
573          * <prev RTT . ><current RTT .. ><next RTT .... >
574          */
575 
576         if (sysctl_tcp_moderate_rcvbuf &&
577             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
578                 int rcvwin, rcvmem, rcvbuf;
579 
580                 /* minimal window to cope with packet losses, assuming
581                  * steady state. Add some cushion because of small variations.
582                  */
583                 rcvwin = (copied << 1) + 16 * tp->advmss;
584 
585                 /* If rate increased by 25%,
586                  *      assume slow start, rcvwin = 3 * copied
587                  * If rate increased by 50%,
588                  *      assume sender can use 2x growth, rcvwin = 4 * copied
589                  */
590                 if (copied >=
591                     tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
592                         if (copied >=
593                             tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
594                                 rcvwin <<= 1;
595                         else
596                                 rcvwin += (rcvwin >> 1);
597                 }
598 
599                 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
600                 while (tcp_win_from_space(rcvmem) < tp->advmss)
601                         rcvmem += 128;
602 
603                 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
604                 if (rcvbuf > sk->sk_rcvbuf) {
605                         sk->sk_rcvbuf = rcvbuf;
606 
607                         /* Make the window clamp follow along.  */
608                         tp->window_clamp = rcvwin;
609                 }
610         }
611         tp->rcvq_space.space = copied;
612 
613 new_measure:
614         tp->rcvq_space.seq = tp->copied_seq;
615         tp->rcvq_space.time = tcp_time_stamp;
616 }
617 
618 /* There is something which you must keep in mind when you analyze the
619  * behavior of the tp->ato delayed ack timeout interval.  When a
620  * connection starts up, we want to ack as quickly as possible.  The
621  * problem is that "good" TCP's do slow start at the beginning of data
622  * transmission.  The means that until we send the first few ACK's the
623  * sender will sit on his end and only queue most of his data, because
624  * he can only send snd_cwnd unacked packets at any given time.  For
625  * each ACK we send, he increments snd_cwnd and transmits more of his
626  * queue.  -DaveM
627  */
628 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
629 {
630         struct tcp_sock *tp = tcp_sk(sk);
631         struct inet_connection_sock *icsk = inet_csk(sk);
632         u32 now;
633 
634         inet_csk_schedule_ack(sk);
635 
636         tcp_measure_rcv_mss(sk, skb);
637 
638         tcp_rcv_rtt_measure(tp);
639 
640         now = tcp_time_stamp;
641 
642         if (!icsk->icsk_ack.ato) {
643                 /* The _first_ data packet received, initialize
644                  * delayed ACK engine.
645                  */
646                 tcp_incr_quickack(sk);
647                 icsk->icsk_ack.ato = TCP_ATO_MIN;
648         } else {
649                 int m = now - icsk->icsk_ack.lrcvtime;
650 
651                 if (m <= TCP_ATO_MIN / 2) {
652                         /* The fastest case is the first. */
653                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
654                 } else if (m < icsk->icsk_ack.ato) {
655                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
656                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
657                                 icsk->icsk_ack.ato = icsk->icsk_rto;
658                 } else if (m > icsk->icsk_rto) {
659                         /* Too long gap. Apparently sender failed to
660                          * restart window, so that we send ACKs quickly.
661                          */
662                         tcp_incr_quickack(sk);
663                         sk_mem_reclaim(sk);
664                 }
665         }
666         icsk->icsk_ack.lrcvtime = now;
667 
668         tcp_ecn_check_ce(tp, skb);
669 
670         if (skb->len >= 128)
671                 tcp_grow_window(sk, skb);
672 }
673 
674 /* Called to compute a smoothed rtt estimate. The data fed to this
675  * routine either comes from timestamps, or from segments that were
676  * known _not_ to have been retransmitted [see Karn/Partridge
677  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
678  * piece by Van Jacobson.
679  * NOTE: the next three routines used to be one big routine.
680  * To save cycles in the RFC 1323 implementation it was better to break
681  * it up into three procedures. -- erics
682  */
683 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
684 {
685         struct tcp_sock *tp = tcp_sk(sk);
686         long m = mrtt_us; /* RTT */
687         u32 srtt = tp->srtt_us;
688 
689         /*      The following amusing code comes from Jacobson's
690          *      article in SIGCOMM '88.  Note that rtt and mdev
691          *      are scaled versions of rtt and mean deviation.
692          *      This is designed to be as fast as possible
693          *      m stands for "measurement".
694          *
695          *      On a 1990 paper the rto value is changed to:
696          *      RTO = rtt + 4 * mdev
697          *
698          * Funny. This algorithm seems to be very broken.
699          * These formulae increase RTO, when it should be decreased, increase
700          * too slowly, when it should be increased quickly, decrease too quickly
701          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
702          * does not matter how to _calculate_ it. Seems, it was trap
703          * that VJ failed to avoid. 8)
704          */
705         if (srtt != 0) {
706                 m -= (srtt >> 3);       /* m is now error in rtt est */
707                 srtt += m;              /* rtt = 7/8 rtt + 1/8 new */
708                 if (m < 0) {
709                         m = -m;         /* m is now abs(error) */
710                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
711                         /* This is similar to one of Eifel findings.
712                          * Eifel blocks mdev updates when rtt decreases.
713                          * This solution is a bit different: we use finer gain
714                          * for mdev in this case (alpha*beta).
715                          * Like Eifel it also prevents growth of rto,
716                          * but also it limits too fast rto decreases,
717                          * happening in pure Eifel.
718                          */
719                         if (m > 0)
720                                 m >>= 3;
721                 } else {
722                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
723                 }
724                 tp->mdev_us += m;               /* mdev = 3/4 mdev + 1/4 new */
725                 if (tp->mdev_us > tp->mdev_max_us) {
726                         tp->mdev_max_us = tp->mdev_us;
727                         if (tp->mdev_max_us > tp->rttvar_us)
728                                 tp->rttvar_us = tp->mdev_max_us;
729                 }
730                 if (after(tp->snd_una, tp->rtt_seq)) {
731                         if (tp->mdev_max_us < tp->rttvar_us)
732                                 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
733                         tp->rtt_seq = tp->snd_nxt;
734                         tp->mdev_max_us = tcp_rto_min_us(sk);
735                 }
736         } else {
737                 /* no previous measure. */
738                 srtt = m << 3;          /* take the measured time to be rtt */
739                 tp->mdev_us = m << 1;   /* make sure rto = 3*rtt */
740                 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
741                 tp->mdev_max_us = tp->rttvar_us;
742                 tp->rtt_seq = tp->snd_nxt;
743         }
744         tp->srtt_us = max(1U, srtt);
745 }
746 
747 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
748  * Note: TCP stack does not yet implement pacing.
749  * FQ packet scheduler can be used to implement cheap but effective
750  * TCP pacing, to smooth the burst on large writes when packets
751  * in flight is significantly lower than cwnd (or rwin)
752  */
753 static void tcp_update_pacing_rate(struct sock *sk)
754 {
755         const struct tcp_sock *tp = tcp_sk(sk);
756         u64 rate;
757 
758         /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
759         rate = (u64)tp->mss_cache * 2 * (USEC_PER_SEC << 3);
760 
761         rate *= max(tp->snd_cwnd, tp->packets_out);
762 
763         if (likely(tp->srtt_us))
764                 do_div(rate, tp->srtt_us);
765 
766         /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
767          * without any lock. We want to make sure compiler wont store
768          * intermediate values in this location.
769          */
770         ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
771                                                 sk->sk_max_pacing_rate);
772 }
773 
774 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
775  * routine referred to above.
776  */
777 static void tcp_set_rto(struct sock *sk)
778 {
779         const struct tcp_sock *tp = tcp_sk(sk);
780         /* Old crap is replaced with new one. 8)
781          *
782          * More seriously:
783          * 1. If rtt variance happened to be less 50msec, it is hallucination.
784          *    It cannot be less due to utterly erratic ACK generation made
785          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
786          *    to do with delayed acks, because at cwnd>2 true delack timeout
787          *    is invisible. Actually, Linux-2.4 also generates erratic
788          *    ACKs in some circumstances.
789          */
790         inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
791 
792         /* 2. Fixups made earlier cannot be right.
793          *    If we do not estimate RTO correctly without them,
794          *    all the algo is pure shit and should be replaced
795          *    with correct one. It is exactly, which we pretend to do.
796          */
797 
798         /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
799          * guarantees that rto is higher.
800          */
801         tcp_bound_rto(sk);
802 }
803 
804 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
805 {
806         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
807 
808         if (!cwnd)
809                 cwnd = TCP_INIT_CWND;
810         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
811 }
812 
813 /*
814  * Packet counting of FACK is based on in-order assumptions, therefore TCP
815  * disables it when reordering is detected
816  */
817 void tcp_disable_fack(struct tcp_sock *tp)
818 {
819         /* RFC3517 uses different metric in lost marker => reset on change */
820         if (tcp_is_fack(tp))
821                 tp->lost_skb_hint = NULL;
822         tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
823 }
824 
825 /* Take a notice that peer is sending D-SACKs */
826 static void tcp_dsack_seen(struct tcp_sock *tp)
827 {
828         tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
829 }
830 
831 static void tcp_update_reordering(struct sock *sk, const int metric,
832                                   const int ts)
833 {
834         struct tcp_sock *tp = tcp_sk(sk);
835         if (metric > tp->reordering) {
836                 int mib_idx;
837 
838                 tp->reordering = min(sysctl_tcp_max_reordering, metric);
839 
840                 /* This exciting event is worth to be remembered. 8) */
841                 if (ts)
842                         mib_idx = LINUX_MIB_TCPTSREORDER;
843                 else if (tcp_is_reno(tp))
844                         mib_idx = LINUX_MIB_TCPRENOREORDER;
845                 else if (tcp_is_fack(tp))
846                         mib_idx = LINUX_MIB_TCPFACKREORDER;
847                 else
848                         mib_idx = LINUX_MIB_TCPSACKREORDER;
849 
850                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
851 #if FASTRETRANS_DEBUG > 1
852                 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
853                          tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
854                          tp->reordering,
855                          tp->fackets_out,
856                          tp->sacked_out,
857                          tp->undo_marker ? tp->undo_retrans : 0);
858 #endif
859                 tcp_disable_fack(tp);
860         }
861 
862         if (metric > 0)
863                 tcp_disable_early_retrans(tp);
864 }
865 
866 /* This must be called before lost_out is incremented */
867 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
868 {
869         if (!tp->retransmit_skb_hint ||
870             before(TCP_SKB_CB(skb)->seq,
871                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
872                 tp->retransmit_skb_hint = skb;
873 
874         if (!tp->lost_out ||
875             after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
876                 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
877 }
878 
879 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
880 {
881         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
882                 tcp_verify_retransmit_hint(tp, skb);
883 
884                 tp->lost_out += tcp_skb_pcount(skb);
885                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
886         }
887 }
888 
889 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
890                                             struct sk_buff *skb)
891 {
892         tcp_verify_retransmit_hint(tp, skb);
893 
894         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
895                 tp->lost_out += tcp_skb_pcount(skb);
896                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
897         }
898 }
899 
900 /* This procedure tags the retransmission queue when SACKs arrive.
901  *
902  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
903  * Packets in queue with these bits set are counted in variables
904  * sacked_out, retrans_out and lost_out, correspondingly.
905  *
906  * Valid combinations are:
907  * Tag  InFlight        Description
908  * 0    1               - orig segment is in flight.
909  * S    0               - nothing flies, orig reached receiver.
910  * L    0               - nothing flies, orig lost by net.
911  * R    2               - both orig and retransmit are in flight.
912  * L|R  1               - orig is lost, retransmit is in flight.
913  * S|R  1               - orig reached receiver, retrans is still in flight.
914  * (L|S|R is logically valid, it could occur when L|R is sacked,
915  *  but it is equivalent to plain S and code short-curcuits it to S.
916  *  L|S is logically invalid, it would mean -1 packet in flight 8))
917  *
918  * These 6 states form finite state machine, controlled by the following events:
919  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
920  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
921  * 3. Loss detection event of two flavors:
922  *      A. Scoreboard estimator decided the packet is lost.
923  *         A'. Reno "three dupacks" marks head of queue lost.
924  *         A''. Its FACK modification, head until snd.fack is lost.
925  *      B. SACK arrives sacking SND.NXT at the moment, when the
926  *         segment was retransmitted.
927  * 4. D-SACK added new rule: D-SACK changes any tag to S.
928  *
929  * It is pleasant to note, that state diagram turns out to be commutative,
930  * so that we are allowed not to be bothered by order of our actions,
931  * when multiple events arrive simultaneously. (see the function below).
932  *
933  * Reordering detection.
934  * --------------------
935  * Reordering metric is maximal distance, which a packet can be displaced
936  * in packet stream. With SACKs we can estimate it:
937  *
938  * 1. SACK fills old hole and the corresponding segment was not
939  *    ever retransmitted -> reordering. Alas, we cannot use it
940  *    when segment was retransmitted.
941  * 2. The last flaw is solved with D-SACK. D-SACK arrives
942  *    for retransmitted and already SACKed segment -> reordering..
943  * Both of these heuristics are not used in Loss state, when we cannot
944  * account for retransmits accurately.
945  *
946  * SACK block validation.
947  * ----------------------
948  *
949  * SACK block range validation checks that the received SACK block fits to
950  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
951  * Note that SND.UNA is not included to the range though being valid because
952  * it means that the receiver is rather inconsistent with itself reporting
953  * SACK reneging when it should advance SND.UNA. Such SACK block this is
954  * perfectly valid, however, in light of RFC2018 which explicitly states
955  * that "SACK block MUST reflect the newest segment.  Even if the newest
956  * segment is going to be discarded ...", not that it looks very clever
957  * in case of head skb. Due to potentional receiver driven attacks, we
958  * choose to avoid immediate execution of a walk in write queue due to
959  * reneging and defer head skb's loss recovery to standard loss recovery
960  * procedure that will eventually trigger (nothing forbids us doing this).
961  *
962  * Implements also blockage to start_seq wrap-around. Problem lies in the
963  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
964  * there's no guarantee that it will be before snd_nxt (n). The problem
965  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
966  * wrap (s_w):
967  *
968  *         <- outs wnd ->                          <- wrapzone ->
969  *         u     e      n                         u_w   e_w  s n_w
970  *         |     |      |                          |     |   |  |
971  * |<------------+------+----- TCP seqno space --------------+---------->|
972  * ...-- <2^31 ->|                                           |<--------...
973  * ...---- >2^31 ------>|                                    |<--------...
974  *
975  * Current code wouldn't be vulnerable but it's better still to discard such
976  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
977  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
978  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
979  * equal to the ideal case (infinite seqno space without wrap caused issues).
980  *
981  * With D-SACK the lower bound is extended to cover sequence space below
982  * SND.UNA down to undo_marker, which is the last point of interest. Yet
983  * again, D-SACK block must not to go across snd_una (for the same reason as
984  * for the normal SACK blocks, explained above). But there all simplicity
985  * ends, TCP might receive valid D-SACKs below that. As long as they reside
986  * fully below undo_marker they do not affect behavior in anyway and can
987  * therefore be safely ignored. In rare cases (which are more or less
988  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
989  * fragmentation and packet reordering past skb's retransmission. To consider
990  * them correctly, the acceptable range must be extended even more though
991  * the exact amount is rather hard to quantify. However, tp->max_window can
992  * be used as an exaggerated estimate.
993  */
994 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
995                                    u32 start_seq, u32 end_seq)
996 {
997         /* Too far in future, or reversed (interpretation is ambiguous) */
998         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
999                 return false;
1000 
1001         /* Nasty start_seq wrap-around check (see comments above) */
1002         if (!before(start_seq, tp->snd_nxt))
1003                 return false;
1004 
1005         /* In outstanding window? ...This is valid exit for D-SACKs too.
1006          * start_seq == snd_una is non-sensical (see comments above)
1007          */
1008         if (after(start_seq, tp->snd_una))
1009                 return true;
1010 
1011         if (!is_dsack || !tp->undo_marker)
1012                 return false;
1013 
1014         /* ...Then it's D-SACK, and must reside below snd_una completely */
1015         if (after(end_seq, tp->snd_una))
1016                 return false;
1017 
1018         if (!before(start_seq, tp->undo_marker))
1019                 return true;
1020 
1021         /* Too old */
1022         if (!after(end_seq, tp->undo_marker))
1023                 return false;
1024 
1025         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1026          *   start_seq < undo_marker and end_seq >= undo_marker.
1027          */
1028         return !before(start_seq, end_seq - tp->max_window);
1029 }
1030 
1031 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1032  * Event "B". Later note: FACK people cheated me again 8), we have to account
1033  * for reordering! Ugly, but should help.
1034  *
1035  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1036  * less than what is now known to be received by the other end (derived from
1037  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1038  * retransmitted skbs to avoid some costly processing per ACKs.
1039  */
1040 static void tcp_mark_lost_retrans(struct sock *sk)
1041 {
1042         const struct inet_connection_sock *icsk = inet_csk(sk);
1043         struct tcp_sock *tp = tcp_sk(sk);
1044         struct sk_buff *skb;
1045         int cnt = 0;
1046         u32 new_low_seq = tp->snd_nxt;
1047         u32 received_upto = tcp_highest_sack_seq(tp);
1048 
1049         if (!tcp_is_fack(tp) || !tp->retrans_out ||
1050             !after(received_upto, tp->lost_retrans_low) ||
1051             icsk->icsk_ca_state != TCP_CA_Recovery)
1052                 return;
1053 
1054         tcp_for_write_queue(skb, sk) {
1055                 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1056 
1057                 if (skb == tcp_send_head(sk))
1058                         break;
1059                 if (cnt == tp->retrans_out)
1060                         break;
1061                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1062                         continue;
1063 
1064                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1065                         continue;
1066 
1067                 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1068                  * constraint here (see above) but figuring out that at
1069                  * least tp->reordering SACK blocks reside between ack_seq
1070                  * and received_upto is not easy task to do cheaply with
1071                  * the available datastructures.
1072                  *
1073                  * Whether FACK should check here for tp->reordering segs
1074                  * in-between one could argue for either way (it would be
1075                  * rather simple to implement as we could count fack_count
1076                  * during the walk and do tp->fackets_out - fack_count).
1077                  */
1078                 if (after(received_upto, ack_seq)) {
1079                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1080                         tp->retrans_out -= tcp_skb_pcount(skb);
1081 
1082                         tcp_skb_mark_lost_uncond_verify(tp, skb);
1083                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1084                 } else {
1085                         if (before(ack_seq, new_low_seq))
1086                                 new_low_seq = ack_seq;
1087                         cnt += tcp_skb_pcount(skb);
1088                 }
1089         }
1090 
1091         if (tp->retrans_out)
1092                 tp->lost_retrans_low = new_low_seq;
1093 }
1094 
1095 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1096                             struct tcp_sack_block_wire *sp, int num_sacks,
1097                             u32 prior_snd_una)
1098 {
1099         struct tcp_sock *tp = tcp_sk(sk);
1100         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1101         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1102         bool dup_sack = false;
1103 
1104         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1105                 dup_sack = true;
1106                 tcp_dsack_seen(tp);
1107                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1108         } else if (num_sacks > 1) {
1109                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1110                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1111 
1112                 if (!after(end_seq_0, end_seq_1) &&
1113                     !before(start_seq_0, start_seq_1)) {
1114                         dup_sack = true;
1115                         tcp_dsack_seen(tp);
1116                         NET_INC_STATS_BH(sock_net(sk),
1117                                         LINUX_MIB_TCPDSACKOFORECV);
1118                 }
1119         }
1120 
1121         /* D-SACK for already forgotten data... Do dumb counting. */
1122         if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1123             !after(end_seq_0, prior_snd_una) &&
1124             after(end_seq_0, tp->undo_marker))
1125                 tp->undo_retrans--;
1126 
1127         return dup_sack;
1128 }
1129 
1130 struct tcp_sacktag_state {
1131         int     reord;
1132         int     fack_count;
1133         /* Timestamps for earliest and latest never-retransmitted segment
1134          * that was SACKed. RTO needs the earliest RTT to stay conservative,
1135          * but congestion control should still get an accurate delay signal.
1136          */
1137         struct skb_mstamp first_sackt;
1138         struct skb_mstamp last_sackt;
1139         int     flag;
1140 };
1141 
1142 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1143  * the incoming SACK may not exactly match but we can find smaller MSS
1144  * aligned portion of it that matches. Therefore we might need to fragment
1145  * which may fail and creates some hassle (caller must handle error case
1146  * returns).
1147  *
1148  * FIXME: this could be merged to shift decision code
1149  */
1150 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1151                                   u32 start_seq, u32 end_seq)
1152 {
1153         int err;
1154         bool in_sack;
1155         unsigned int pkt_len;
1156         unsigned int mss;
1157 
1158         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1159                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1160 
1161         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1162             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1163                 mss = tcp_skb_mss(skb);
1164                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1165 
1166                 if (!in_sack) {
1167                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1168                         if (pkt_len < mss)
1169                                 pkt_len = mss;
1170                 } else {
1171                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1172                         if (pkt_len < mss)
1173                                 return -EINVAL;
1174                 }
1175 
1176                 /* Round if necessary so that SACKs cover only full MSSes
1177                  * and/or the remaining small portion (if present)
1178                  */
1179                 if (pkt_len > mss) {
1180                         unsigned int new_len = (pkt_len / mss) * mss;
1181                         if (!in_sack && new_len < pkt_len) {
1182                                 new_len += mss;
1183                                 if (new_len >= skb->len)
1184                                         return 0;
1185                         }
1186                         pkt_len = new_len;
1187                 }
1188                 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1189                 if (err < 0)
1190                         return err;
1191         }
1192 
1193         return in_sack;
1194 }
1195 
1196 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1197 static u8 tcp_sacktag_one(struct sock *sk,
1198                           struct tcp_sacktag_state *state, u8 sacked,
1199                           u32 start_seq, u32 end_seq,
1200                           int dup_sack, int pcount,
1201                           const struct skb_mstamp *xmit_time)
1202 {
1203         struct tcp_sock *tp = tcp_sk(sk);
1204         int fack_count = state->fack_count;
1205 
1206         /* Account D-SACK for retransmitted packet. */
1207         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1208                 if (tp->undo_marker && tp->undo_retrans > 0 &&
1209                     after(end_seq, tp->undo_marker))
1210                         tp->undo_retrans--;
1211                 if (sacked & TCPCB_SACKED_ACKED)
1212                         state->reord = min(fack_count, state->reord);
1213         }
1214 
1215         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1216         if (!after(end_seq, tp->snd_una))
1217                 return sacked;
1218 
1219         if (!(sacked & TCPCB_SACKED_ACKED)) {
1220                 if (sacked & TCPCB_SACKED_RETRANS) {
1221                         /* If the segment is not tagged as lost,
1222                          * we do not clear RETRANS, believing
1223                          * that retransmission is still in flight.
1224                          */
1225                         if (sacked & TCPCB_LOST) {
1226                                 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1227                                 tp->lost_out -= pcount;
1228                                 tp->retrans_out -= pcount;
1229                         }
1230                 } else {
1231                         if (!(sacked & TCPCB_RETRANS)) {
1232                                 /* New sack for not retransmitted frame,
1233                                  * which was in hole. It is reordering.
1234                                  */
1235                                 if (before(start_seq,
1236                                            tcp_highest_sack_seq(tp)))
1237                                         state->reord = min(fack_count,
1238                                                            state->reord);
1239                                 if (!after(end_seq, tp->high_seq))
1240                                         state->flag |= FLAG_ORIG_SACK_ACKED;
1241                                 if (state->first_sackt.v64 == 0)
1242                                         state->first_sackt = *xmit_time;
1243                                 state->last_sackt = *xmit_time;
1244                         }
1245 
1246                         if (sacked & TCPCB_LOST) {
1247                                 sacked &= ~TCPCB_LOST;
1248                                 tp->lost_out -= pcount;
1249                         }
1250                 }
1251 
1252                 sacked |= TCPCB_SACKED_ACKED;
1253                 state->flag |= FLAG_DATA_SACKED;
1254                 tp->sacked_out += pcount;
1255 
1256                 fack_count += pcount;
1257 
1258                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1259                 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1260                     before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1261                         tp->lost_cnt_hint += pcount;
1262 
1263                 if (fack_count > tp->fackets_out)
1264                         tp->fackets_out = fack_count;
1265         }
1266 
1267         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1268          * frames and clear it. undo_retrans is decreased above, L|R frames
1269          * are accounted above as well.
1270          */
1271         if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1272                 sacked &= ~TCPCB_SACKED_RETRANS;
1273                 tp->retrans_out -= pcount;
1274         }
1275 
1276         return sacked;
1277 }
1278 
1279 /* Shift newly-SACKed bytes from this skb to the immediately previous
1280  * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1281  */
1282 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1283                             struct tcp_sacktag_state *state,
1284                             unsigned int pcount, int shifted, int mss,
1285                             bool dup_sack)
1286 {
1287         struct tcp_sock *tp = tcp_sk(sk);
1288         struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1289         u32 start_seq = TCP_SKB_CB(skb)->seq;   /* start of newly-SACKed */
1290         u32 end_seq = start_seq + shifted;      /* end of newly-SACKed */
1291 
1292         BUG_ON(!pcount);
1293 
1294         /* Adjust counters and hints for the newly sacked sequence
1295          * range but discard the return value since prev is already
1296          * marked. We must tag the range first because the seq
1297          * advancement below implicitly advances
1298          * tcp_highest_sack_seq() when skb is highest_sack.
1299          */
1300         tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1301                         start_seq, end_seq, dup_sack, pcount,
1302                         &skb->skb_mstamp);
1303 
1304         if (skb == tp->lost_skb_hint)
1305                 tp->lost_cnt_hint += pcount;
1306 
1307         TCP_SKB_CB(prev)->end_seq += shifted;
1308         TCP_SKB_CB(skb)->seq += shifted;
1309 
1310         tcp_skb_pcount_add(prev, pcount);
1311         BUG_ON(tcp_skb_pcount(skb) < pcount);
1312         tcp_skb_pcount_add(skb, -pcount);
1313 
1314         /* When we're adding to gso_segs == 1, gso_size will be zero,
1315          * in theory this shouldn't be necessary but as long as DSACK
1316          * code can come after this skb later on it's better to keep
1317          * setting gso_size to something.
1318          */
1319         if (!TCP_SKB_CB(prev)->tcp_gso_size)
1320                 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1321 
1322         /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1323         if (tcp_skb_pcount(skb) <= 1)
1324                 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1325 
1326         /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1327         TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1328 
1329         if (skb->len > 0) {
1330                 BUG_ON(!tcp_skb_pcount(skb));
1331                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1332                 return false;
1333         }
1334 
1335         /* Whole SKB was eaten :-) */
1336 
1337         if (skb == tp->retransmit_skb_hint)
1338                 tp->retransmit_skb_hint = prev;
1339         if (skb == tp->lost_skb_hint) {
1340                 tp->lost_skb_hint = prev;
1341                 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1342         }
1343 
1344         TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1345         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1346                 TCP_SKB_CB(prev)->end_seq++;
1347 
1348         if (skb == tcp_highest_sack(sk))
1349                 tcp_advance_highest_sack(sk, skb);
1350 
1351         tcp_unlink_write_queue(skb, sk);
1352         sk_wmem_free_skb(sk, skb);
1353 
1354         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1355 
1356         return true;
1357 }
1358 
1359 /* I wish gso_size would have a bit more sane initialization than
1360  * something-or-zero which complicates things
1361  */
1362 static int tcp_skb_seglen(const struct sk_buff *skb)
1363 {
1364         return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1365 }
1366 
1367 /* Shifting pages past head area doesn't work */
1368 static int skb_can_shift(const struct sk_buff *skb)
1369 {
1370         return !skb_headlen(skb) && skb_is_nonlinear(skb);
1371 }
1372 
1373 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1374  * skb.
1375  */
1376 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1377                                           struct tcp_sacktag_state *state,
1378                                           u32 start_seq, u32 end_seq,
1379                                           bool dup_sack)
1380 {
1381         struct tcp_sock *tp = tcp_sk(sk);
1382         struct sk_buff *prev;
1383         int mss;
1384         int pcount = 0;
1385         int len;
1386         int in_sack;
1387 
1388         if (!sk_can_gso(sk))
1389                 goto fallback;
1390 
1391         /* Normally R but no L won't result in plain S */
1392         if (!dup_sack &&
1393             (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1394                 goto fallback;
1395         if (!skb_can_shift(skb))
1396                 goto fallback;
1397         /* This frame is about to be dropped (was ACKed). */
1398         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1399                 goto fallback;
1400 
1401         /* Can only happen with delayed DSACK + discard craziness */
1402         if (unlikely(skb == tcp_write_queue_head(sk)))
1403                 goto fallback;
1404         prev = tcp_write_queue_prev(sk, skb);
1405 
1406         if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1407                 goto fallback;
1408 
1409         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1410                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1411 
1412         if (in_sack) {
1413                 len = skb->len;
1414                 pcount = tcp_skb_pcount(skb);
1415                 mss = tcp_skb_seglen(skb);
1416 
1417                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1418                  * drop this restriction as unnecessary
1419                  */
1420                 if (mss != tcp_skb_seglen(prev))
1421                         goto fallback;
1422         } else {
1423                 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1424                         goto noop;
1425                 /* CHECKME: This is non-MSS split case only?, this will
1426                  * cause skipped skbs due to advancing loop btw, original
1427                  * has that feature too
1428                  */
1429                 if (tcp_skb_pcount(skb) <= 1)
1430                         goto noop;
1431 
1432                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1433                 if (!in_sack) {
1434                         /* TODO: head merge to next could be attempted here
1435                          * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1436                          * though it might not be worth of the additional hassle
1437                          *
1438                          * ...we can probably just fallback to what was done
1439                          * previously. We could try merging non-SACKed ones
1440                          * as well but it probably isn't going to buy off
1441                          * because later SACKs might again split them, and
1442                          * it would make skb timestamp tracking considerably
1443                          * harder problem.
1444                          */
1445                         goto fallback;
1446                 }
1447 
1448                 len = end_seq - TCP_SKB_CB(skb)->seq;
1449                 BUG_ON(len < 0);
1450                 BUG_ON(len > skb->len);
1451 
1452                 /* MSS boundaries should be honoured or else pcount will
1453                  * severely break even though it makes things bit trickier.
1454                  * Optimize common case to avoid most of the divides
1455                  */
1456                 mss = tcp_skb_mss(skb);
1457 
1458                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1459                  * drop this restriction as unnecessary
1460                  */
1461                 if (mss != tcp_skb_seglen(prev))
1462                         goto fallback;
1463 
1464                 if (len == mss) {
1465                         pcount = 1;
1466                 } else if (len < mss) {
1467                         goto noop;
1468                 } else {
1469                         pcount = len / mss;
1470                         len = pcount * mss;
1471                 }
1472         }
1473 
1474         /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1475         if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1476                 goto fallback;
1477 
1478         if (!skb_shift(prev, skb, len))
1479                 goto fallback;
1480         if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1481                 goto out;
1482 
1483         /* Hole filled allows collapsing with the next as well, this is very
1484          * useful when hole on every nth skb pattern happens
1485          */
1486         if (prev == tcp_write_queue_tail(sk))
1487                 goto out;
1488         skb = tcp_write_queue_next(sk, prev);
1489 
1490         if (!skb_can_shift(skb) ||
1491             (skb == tcp_send_head(sk)) ||
1492             ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1493             (mss != tcp_skb_seglen(skb)))
1494                 goto out;
1495 
1496         len = skb->len;
1497         if (skb_shift(prev, skb, len)) {
1498                 pcount += tcp_skb_pcount(skb);
1499                 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1500         }
1501 
1502 out:
1503         state->fack_count += pcount;
1504         return prev;
1505 
1506 noop:
1507         return skb;
1508 
1509 fallback:
1510         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1511         return NULL;
1512 }
1513 
1514 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1515                                         struct tcp_sack_block *next_dup,
1516                                         struct tcp_sacktag_state *state,
1517                                         u32 start_seq, u32 end_seq,
1518                                         bool dup_sack_in)
1519 {
1520         struct tcp_sock *tp = tcp_sk(sk);
1521         struct sk_buff *tmp;
1522 
1523         tcp_for_write_queue_from(skb, sk) {
1524                 int in_sack = 0;
1525                 bool dup_sack = dup_sack_in;
1526 
1527                 if (skb == tcp_send_head(sk))
1528                         break;
1529 
1530                 /* queue is in-order => we can short-circuit the walk early */
1531                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1532                         break;
1533 
1534                 if (next_dup  &&
1535                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1536                         in_sack = tcp_match_skb_to_sack(sk, skb,
1537                                                         next_dup->start_seq,
1538                                                         next_dup->end_seq);
1539                         if (in_sack > 0)
1540                                 dup_sack = true;
1541                 }
1542 
1543                 /* skb reference here is a bit tricky to get right, since
1544                  * shifting can eat and free both this skb and the next,
1545                  * so not even _safe variant of the loop is enough.
1546                  */
1547                 if (in_sack <= 0) {
1548                         tmp = tcp_shift_skb_data(sk, skb, state,
1549                                                  start_seq, end_seq, dup_sack);
1550                         if (tmp) {
1551                                 if (tmp != skb) {
1552                                         skb = tmp;
1553                                         continue;
1554                                 }
1555 
1556                                 in_sack = 0;
1557                         } else {
1558                                 in_sack = tcp_match_skb_to_sack(sk, skb,
1559                                                                 start_seq,
1560                                                                 end_seq);
1561                         }
1562                 }
1563 
1564                 if (unlikely(in_sack < 0))
1565                         break;
1566 
1567                 if (in_sack) {
1568                         TCP_SKB_CB(skb)->sacked =
1569                                 tcp_sacktag_one(sk,
1570                                                 state,
1571                                                 TCP_SKB_CB(skb)->sacked,
1572                                                 TCP_SKB_CB(skb)->seq,
1573                                                 TCP_SKB_CB(skb)->end_seq,
1574                                                 dup_sack,
1575                                                 tcp_skb_pcount(skb),
1576                                                 &skb->skb_mstamp);
1577 
1578                         if (!before(TCP_SKB_CB(skb)->seq,
1579                                     tcp_highest_sack_seq(tp)))
1580                                 tcp_advance_highest_sack(sk, skb);
1581                 }
1582 
1583                 state->fack_count += tcp_skb_pcount(skb);
1584         }
1585         return skb;
1586 }
1587 
1588 /* Avoid all extra work that is being done by sacktag while walking in
1589  * a normal way
1590  */
1591 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1592                                         struct tcp_sacktag_state *state,
1593                                         u32 skip_to_seq)
1594 {
1595         tcp_for_write_queue_from(skb, sk) {
1596                 if (skb == tcp_send_head(sk))
1597                         break;
1598 
1599                 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1600                         break;
1601 
1602                 state->fack_count += tcp_skb_pcount(skb);
1603         }
1604         return skb;
1605 }
1606 
1607 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1608                                                 struct sock *sk,
1609                                                 struct tcp_sack_block *next_dup,
1610                                                 struct tcp_sacktag_state *state,
1611                                                 u32 skip_to_seq)
1612 {
1613         if (!next_dup)
1614                 return skb;
1615 
1616         if (before(next_dup->start_seq, skip_to_seq)) {
1617                 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1618                 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1619                                        next_dup->start_seq, next_dup->end_seq,
1620                                        1);
1621         }
1622 
1623         return skb;
1624 }
1625 
1626 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1627 {
1628         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1629 }
1630 
1631 static int
1632 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1633                         u32 prior_snd_una, struct tcp_sacktag_state *state)
1634 {
1635         struct tcp_sock *tp = tcp_sk(sk);
1636         const unsigned char *ptr = (skb_transport_header(ack_skb) +
1637                                     TCP_SKB_CB(ack_skb)->sacked);
1638         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1639         struct tcp_sack_block sp[TCP_NUM_SACKS];
1640         struct tcp_sack_block *cache;
1641         struct sk_buff *skb;
1642         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1643         int used_sacks;
1644         bool found_dup_sack = false;
1645         int i, j;
1646         int first_sack_index;
1647 
1648         state->flag = 0;
1649         state->reord = tp->packets_out;
1650 
1651         if (!tp->sacked_out) {
1652                 if (WARN_ON(tp->fackets_out))
1653                         tp->fackets_out = 0;
1654                 tcp_highest_sack_reset(sk);
1655         }
1656 
1657         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1658                                          num_sacks, prior_snd_una);
1659         if (found_dup_sack)
1660                 state->flag |= FLAG_DSACKING_ACK;
1661 
1662         /* Eliminate too old ACKs, but take into
1663          * account more or less fresh ones, they can
1664          * contain valid SACK info.
1665          */
1666         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1667                 return 0;
1668 
1669         if (!tp->packets_out)
1670                 goto out;
1671 
1672         used_sacks = 0;
1673         first_sack_index = 0;
1674         for (i = 0; i < num_sacks; i++) {
1675                 bool dup_sack = !i && found_dup_sack;
1676 
1677                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1678                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1679 
1680                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1681                                             sp[used_sacks].start_seq,
1682                                             sp[used_sacks].end_seq)) {
1683                         int mib_idx;
1684 
1685                         if (dup_sack) {
1686                                 if (!tp->undo_marker)
1687                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1688                                 else
1689                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1690                         } else {
1691                                 /* Don't count olds caused by ACK reordering */
1692                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1693                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1694                                         continue;
1695                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1696                         }
1697 
1698                         NET_INC_STATS_BH(sock_net(sk), mib_idx);
1699                         if (i == 0)
1700                                 first_sack_index = -1;
1701                         continue;
1702                 }
1703 
1704                 /* Ignore very old stuff early */
1705                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1706                         continue;
1707 
1708                 used_sacks++;
1709         }
1710 
1711         /* order SACK blocks to allow in order walk of the retrans queue */
1712         for (i = used_sacks - 1; i > 0; i--) {
1713                 for (j = 0; j < i; j++) {
1714                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1715                                 swap(sp[j], sp[j + 1]);
1716 
1717                                 /* Track where the first SACK block goes to */
1718                                 if (j == first_sack_index)
1719                                         first_sack_index = j + 1;
1720                         }
1721                 }
1722         }
1723 
1724         skb = tcp_write_queue_head(sk);
1725         state->fack_count = 0;
1726         i = 0;
1727 
1728         if (!tp->sacked_out) {
1729                 /* It's already past, so skip checking against it */
1730                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1731         } else {
1732                 cache = tp->recv_sack_cache;
1733                 /* Skip empty blocks in at head of the cache */
1734                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1735                        !cache->end_seq)
1736                         cache++;
1737         }
1738 
1739         while (i < used_sacks) {
1740                 u32 start_seq = sp[i].start_seq;
1741                 u32 end_seq = sp[i].end_seq;
1742                 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1743                 struct tcp_sack_block *next_dup = NULL;
1744 
1745                 if (found_dup_sack && ((i + 1) == first_sack_index))
1746                         next_dup = &sp[i + 1];
1747 
1748                 /* Skip too early cached blocks */
1749                 while (tcp_sack_cache_ok(tp, cache) &&
1750                        !before(start_seq, cache->end_seq))
1751                         cache++;
1752 
1753                 /* Can skip some work by looking recv_sack_cache? */
1754                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1755                     after(end_seq, cache->start_seq)) {
1756 
1757                         /* Head todo? */
1758                         if (before(start_seq, cache->start_seq)) {
1759                                 skb = tcp_sacktag_skip(skb, sk, state,
1760                                                        start_seq);
1761                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1762                                                        state,
1763                                                        start_seq,
1764                                                        cache->start_seq,
1765                                                        dup_sack);
1766                         }
1767 
1768                         /* Rest of the block already fully processed? */
1769                         if (!after(end_seq, cache->end_seq))
1770                                 goto advance_sp;
1771 
1772                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1773                                                        state,
1774                                                        cache->end_seq);
1775 
1776                         /* ...tail remains todo... */
1777                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1778                                 /* ...but better entrypoint exists! */
1779                                 skb = tcp_highest_sack(sk);
1780                                 if (!skb)
1781                                         break;
1782                                 state->fack_count = tp->fackets_out;
1783                                 cache++;
1784                                 goto walk;
1785                         }
1786 
1787                         skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1788                         /* Check overlap against next cached too (past this one already) */
1789                         cache++;
1790                         continue;
1791                 }
1792 
1793                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1794                         skb = tcp_highest_sack(sk);
1795                         if (!skb)
1796                                 break;
1797                         state->fack_count = tp->fackets_out;
1798                 }
1799                 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1800 
1801 walk:
1802                 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1803                                        start_seq, end_seq, dup_sack);
1804 
1805 advance_sp:
1806                 i++;
1807         }
1808 
1809         /* Clear the head of the cache sack blocks so we can skip it next time */
1810         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1811                 tp->recv_sack_cache[i].start_seq = 0;
1812                 tp->recv_sack_cache[i].end_seq = 0;
1813         }
1814         for (j = 0; j < used_sacks; j++)
1815                 tp->recv_sack_cache[i++] = sp[j];
1816 
1817         if ((state->reord < tp->fackets_out) &&
1818             ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1819                 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
1820 
1821         tcp_mark_lost_retrans(sk);
1822         tcp_verify_left_out(tp);
1823 out:
1824 
1825 #if FASTRETRANS_DEBUG > 0
1826         WARN_ON((int)tp->sacked_out < 0);
1827         WARN_ON((int)tp->lost_out < 0);
1828         WARN_ON((int)tp->retrans_out < 0);
1829         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1830 #endif
1831         return state->flag;
1832 }
1833 
1834 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1835  * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1836  */
1837 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1838 {
1839         u32 holes;
1840 
1841         holes = max(tp->lost_out, 1U);
1842         holes = min(holes, tp->packets_out);
1843 
1844         if ((tp->sacked_out + holes) > tp->packets_out) {
1845                 tp->sacked_out = tp->packets_out - holes;
1846                 return true;
1847         }
1848         return false;
1849 }
1850 
1851 /* If we receive more dupacks than we expected counting segments
1852  * in assumption of absent reordering, interpret this as reordering.
1853  * The only another reason could be bug in receiver TCP.
1854  */
1855 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1856 {
1857         struct tcp_sock *tp = tcp_sk(sk);
1858         if (tcp_limit_reno_sacked(tp))
1859                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1860 }
1861 
1862 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1863 
1864 static void tcp_add_reno_sack(struct sock *sk)
1865 {
1866         struct tcp_sock *tp = tcp_sk(sk);
1867         tp->sacked_out++;
1868         tcp_check_reno_reordering(sk, 0);
1869         tcp_verify_left_out(tp);
1870 }
1871 
1872 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1873 
1874 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1875 {
1876         struct tcp_sock *tp = tcp_sk(sk);
1877 
1878         if (acked > 0) {
1879                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1880                 if (acked - 1 >= tp->sacked_out)
1881                         tp->sacked_out = 0;
1882                 else
1883                         tp->sacked_out -= acked - 1;
1884         }
1885         tcp_check_reno_reordering(sk, acked);
1886         tcp_verify_left_out(tp);
1887 }
1888 
1889 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1890 {
1891         tp->sacked_out = 0;
1892 }
1893 
1894 void tcp_clear_retrans(struct tcp_sock *tp)
1895 {
1896         tp->retrans_out = 0;
1897         tp->lost_out = 0;
1898         tp->undo_marker = 0;
1899         tp->undo_retrans = -1;
1900         tp->fackets_out = 0;
1901         tp->sacked_out = 0;
1902 }
1903 
1904 static inline void tcp_init_undo(struct tcp_sock *tp)
1905 {
1906         tp->undo_marker = tp->snd_una;
1907         /* Retransmission still in flight may cause DSACKs later. */
1908         tp->undo_retrans = tp->retrans_out ? : -1;
1909 }
1910 
1911 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1912  * and reset tags completely, otherwise preserve SACKs. If receiver
1913  * dropped its ofo queue, we will know this due to reneging detection.
1914  */
1915 void tcp_enter_loss(struct sock *sk)
1916 {
1917         const struct inet_connection_sock *icsk = inet_csk(sk);
1918         struct tcp_sock *tp = tcp_sk(sk);
1919         struct sk_buff *skb;
1920         bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1921         bool is_reneg;                  /* is receiver reneging on SACKs? */
1922 
1923         /* Reduce ssthresh if it has not yet been made inside this window. */
1924         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1925             !after(tp->high_seq, tp->snd_una) ||
1926             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1927                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1928                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1929                 tcp_ca_event(sk, CA_EVENT_LOSS);
1930                 tcp_init_undo(tp);
1931         }
1932         tp->snd_cwnd       = 1;
1933         tp->snd_cwnd_cnt   = 0;
1934         tp->snd_cwnd_stamp = tcp_time_stamp;
1935 
1936         tp->retrans_out = 0;
1937         tp->lost_out = 0;
1938 
1939         if (tcp_is_reno(tp))
1940                 tcp_reset_reno_sack(tp);
1941 
1942         skb = tcp_write_queue_head(sk);
1943         is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1944         if (is_reneg) {
1945                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1946                 tp->sacked_out = 0;
1947                 tp->fackets_out = 0;
1948         }
1949         tcp_clear_all_retrans_hints(tp);
1950 
1951         tcp_for_write_queue(skb, sk) {
1952                 if (skb == tcp_send_head(sk))
1953                         break;
1954 
1955                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1956                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
1957                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1958                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1959                         tp->lost_out += tcp_skb_pcount(skb);
1960                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1961                 }
1962         }
1963         tcp_verify_left_out(tp);
1964 
1965         /* Timeout in disordered state after receiving substantial DUPACKs
1966          * suggests that the degree of reordering is over-estimated.
1967          */
1968         if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1969             tp->sacked_out >= sysctl_tcp_reordering)
1970                 tp->reordering = min_t(unsigned int, tp->reordering,
1971                                        sysctl_tcp_reordering);
1972         tcp_set_ca_state(sk, TCP_CA_Loss);
1973         tp->high_seq = tp->snd_nxt;
1974         tcp_ecn_queue_cwr(tp);
1975 
1976         /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1977          * loss recovery is underway except recurring timeout(s) on
1978          * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1979          */
1980         tp->frto = sysctl_tcp_frto &&
1981                    (new_recovery || icsk->icsk_retransmits) &&
1982                    !inet_csk(sk)->icsk_mtup.probe_size;
1983 }
1984 
1985 /* If ACK arrived pointing to a remembered SACK, it means that our
1986  * remembered SACKs do not reflect real state of receiver i.e.
1987  * receiver _host_ is heavily congested (or buggy).
1988  *
1989  * To avoid big spurious retransmission bursts due to transient SACK
1990  * scoreboard oddities that look like reneging, we give the receiver a
1991  * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1992  * restore sanity to the SACK scoreboard. If the apparent reneging
1993  * persists until this RTO then we'll clear the SACK scoreboard.
1994  */
1995 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1996 {
1997         if (flag & FLAG_SACK_RENEGING) {
1998                 struct tcp_sock *tp = tcp_sk(sk);
1999                 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2000                                           msecs_to_jiffies(10));
2001 
2002                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2003                                           delay, TCP_RTO_MAX);
2004                 return true;
2005         }
2006         return false;
2007 }
2008 
2009 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2010 {
2011         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2012 }
2013 
2014 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2015  * counter when SACK is enabled (without SACK, sacked_out is used for
2016  * that purpose).
2017  *
2018  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2019  * segments up to the highest received SACK block so far and holes in
2020  * between them.
2021  *
2022  * With reordering, holes may still be in flight, so RFC3517 recovery
2023  * uses pure sacked_out (total number of SACKed segments) even though
2024  * it violates the RFC that uses duplicate ACKs, often these are equal
2025  * but when e.g. out-of-window ACKs or packet duplication occurs,
2026  * they differ. Since neither occurs due to loss, TCP should really
2027  * ignore them.
2028  */
2029 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2030 {
2031         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2032 }
2033 
2034 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2035 {
2036         struct tcp_sock *tp = tcp_sk(sk);
2037         unsigned long delay;
2038 
2039         /* Delay early retransmit and entering fast recovery for
2040          * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2041          * available, or RTO is scheduled to fire first.
2042          */
2043         if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2044             (flag & FLAG_ECE) || !tp->srtt_us)
2045                 return false;
2046 
2047         delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2048                     msecs_to_jiffies(2));
2049 
2050         if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2051                 return false;
2052 
2053         inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2054                                   TCP_RTO_MAX);
2055         return true;
2056 }
2057 
2058 /* Linux NewReno/SACK/FACK/ECN state machine.
2059  * --------------------------------------
2060  *
2061  * "Open"       Normal state, no dubious events, fast path.
2062  * "Disorder"   In all the respects it is "Open",
2063  *              but requires a bit more attention. It is entered when
2064  *              we see some SACKs or dupacks. It is split of "Open"
2065  *              mainly to move some processing from fast path to slow one.
2066  * "CWR"        CWND was reduced due to some Congestion Notification event.
2067  *              It can be ECN, ICMP source quench, local device congestion.
2068  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2069  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2070  *
2071  * tcp_fastretrans_alert() is entered:
2072  * - each incoming ACK, if state is not "Open"
2073  * - when arrived ACK is unusual, namely:
2074  *      * SACK
2075  *      * Duplicate ACK.
2076  *      * ECN ECE.
2077  *
2078  * Counting packets in flight is pretty simple.
2079  *
2080  *      in_flight = packets_out - left_out + retrans_out
2081  *
2082  *      packets_out is SND.NXT-SND.UNA counted in packets.
2083  *
2084  *      retrans_out is number of retransmitted segments.
2085  *
2086  *      left_out is number of segments left network, but not ACKed yet.
2087  *
2088  *              left_out = sacked_out + lost_out
2089  *
2090  *     sacked_out: Packets, which arrived to receiver out of order
2091  *                 and hence not ACKed. With SACKs this number is simply
2092  *                 amount of SACKed data. Even without SACKs
2093  *                 it is easy to give pretty reliable estimate of this number,
2094  *                 counting duplicate ACKs.
2095  *
2096  *       lost_out: Packets lost by network. TCP has no explicit
2097  *                 "loss notification" feedback from network (for now).
2098  *                 It means that this number can be only _guessed_.
2099  *                 Actually, it is the heuristics to predict lossage that
2100  *                 distinguishes different algorithms.
2101  *
2102  *      F.e. after RTO, when all the queue is considered as lost,
2103  *      lost_out = packets_out and in_flight = retrans_out.
2104  *
2105  *              Essentially, we have now two algorithms counting
2106  *              lost packets.
2107  *
2108  *              FACK: It is the simplest heuristics. As soon as we decided
2109  *              that something is lost, we decide that _all_ not SACKed
2110  *              packets until the most forward SACK are lost. I.e.
2111  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2112  *              It is absolutely correct estimate, if network does not reorder
2113  *              packets. And it loses any connection to reality when reordering
2114  *              takes place. We use FACK by default until reordering
2115  *              is suspected on the path to this destination.
2116  *
2117  *              NewReno: when Recovery is entered, we assume that one segment
2118  *              is lost (classic Reno). While we are in Recovery and
2119  *              a partial ACK arrives, we assume that one more packet
2120  *              is lost (NewReno). This heuristics are the same in NewReno
2121  *              and SACK.
2122  *
2123  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2124  *  deflation etc. CWND is real congestion window, never inflated, changes
2125  *  only according to classic VJ rules.
2126  *
2127  * Really tricky (and requiring careful tuning) part of algorithm
2128  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2129  * The first determines the moment _when_ we should reduce CWND and,
2130  * hence, slow down forward transmission. In fact, it determines the moment
2131  * when we decide that hole is caused by loss, rather than by a reorder.
2132  *
2133  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2134  * holes, caused by lost packets.
2135  *
2136  * And the most logically complicated part of algorithm is undo
2137  * heuristics. We detect false retransmits due to both too early
2138  * fast retransmit (reordering) and underestimated RTO, analyzing
2139  * timestamps and D-SACKs. When we detect that some segments were
2140  * retransmitted by mistake and CWND reduction was wrong, we undo
2141  * window reduction and abort recovery phase. This logic is hidden
2142  * inside several functions named tcp_try_undo_<something>.
2143  */
2144 
2145 /* This function decides, when we should leave Disordered state
2146  * and enter Recovery phase, reducing congestion window.
2147  *
2148  * Main question: may we further continue forward transmission
2149  * with the same cwnd?
2150  */
2151 static bool tcp_time_to_recover(struct sock *sk, int flag)
2152 {
2153         struct tcp_sock *tp = tcp_sk(sk);
2154         __u32 packets_out;
2155 
2156         /* Trick#1: The loss is proven. */
2157         if (tp->lost_out)
2158                 return true;
2159 
2160         /* Not-A-Trick#2 : Classic rule... */
2161         if (tcp_dupack_heuristics(tp) > tp->reordering)
2162                 return true;
2163 
2164         /* Trick#4: It is still not OK... But will it be useful to delay
2165          * recovery more?
2166          */
2167         packets_out = tp->packets_out;
2168         if (packets_out <= tp->reordering &&
2169             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2170             !tcp_may_send_now(sk)) {
2171                 /* We have nothing to send. This connection is limited
2172                  * either by receiver window or by application.
2173                  */
2174                 return true;
2175         }
2176 
2177         /* If a thin stream is detected, retransmit after first
2178          * received dupack. Employ only if SACK is supported in order
2179          * to avoid possible corner-case series of spurious retransmissions
2180          * Use only if there are no unsent data.
2181          */
2182         if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2183             tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2184             tcp_is_sack(tp) && !tcp_send_head(sk))
2185                 return true;
2186 
2187         /* Trick#6: TCP early retransmit, per RFC5827.  To avoid spurious
2188          * retransmissions due to small network reorderings, we implement
2189          * Mitigation A.3 in the RFC and delay the retransmission for a short
2190          * interval if appropriate.
2191          */
2192         if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2193             (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2194             !tcp_may_send_now(sk))
2195                 return !tcp_pause_early_retransmit(sk, flag);
2196 
2197         return false;
2198 }
2199 
2200 /* Detect loss in event "A" above by marking head of queue up as lost.
2201  * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2202  * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2203  * has at least tp->reordering SACKed seqments above it; "packets" refers to
2204  * the maximum SACKed segments to pass before reaching this limit.
2205  */
2206 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2207 {
2208         struct tcp_sock *tp = tcp_sk(sk);
2209         struct sk_buff *skb;
2210         int cnt, oldcnt;
2211         int err;
2212         unsigned int mss;
2213         /* Use SACK to deduce losses of new sequences sent during recovery */
2214         const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2215 
2216         WARN_ON(packets > tp->packets_out);
2217         if (tp->lost_skb_hint) {
2218                 skb = tp->lost_skb_hint;
2219                 cnt = tp->lost_cnt_hint;
2220                 /* Head already handled? */
2221                 if (mark_head && skb != tcp_write_queue_head(sk))
2222                         return;
2223         } else {
2224                 skb = tcp_write_queue_head(sk);
2225                 cnt = 0;
2226         }
2227 
2228         tcp_for_write_queue_from(skb, sk) {
2229                 if (skb == tcp_send_head(sk))
2230                         break;
2231                 /* TODO: do this better */
2232                 /* this is not the most efficient way to do this... */
2233                 tp->lost_skb_hint = skb;
2234                 tp->lost_cnt_hint = cnt;
2235 
2236                 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2237                         break;
2238 
2239                 oldcnt = cnt;
2240                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2241                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2242                         cnt += tcp_skb_pcount(skb);
2243 
2244                 if (cnt > packets) {
2245                         if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2246                             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2247                             (oldcnt >= packets))
2248                                 break;
2249 
2250                         mss = tcp_skb_mss(skb);
2251                         err = tcp_fragment(sk, skb, (packets - oldcnt) * mss,
2252                                            mss, GFP_ATOMIC);
2253                         if (err < 0)
2254                                 break;
2255                         cnt = packets;
2256                 }
2257 
2258                 tcp_skb_mark_lost(tp, skb);
2259 
2260                 if (mark_head)
2261                         break;
2262         }
2263         tcp_verify_left_out(tp);
2264 }
2265 
2266 /* Account newly detected lost packet(s) */
2267 
2268 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2269 {
2270         struct tcp_sock *tp = tcp_sk(sk);
2271 
2272         if (tcp_is_reno(tp)) {
2273                 tcp_mark_head_lost(sk, 1, 1);
2274         } else if (tcp_is_fack(tp)) {
2275                 int lost = tp->fackets_out - tp->reordering;
2276                 if (lost <= 0)
2277                         lost = 1;
2278                 tcp_mark_head_lost(sk, lost, 0);
2279         } else {
2280                 int sacked_upto = tp->sacked_out - tp->reordering;
2281                 if (sacked_upto >= 0)
2282                         tcp_mark_head_lost(sk, sacked_upto, 0);
2283                 else if (fast_rexmit)
2284                         tcp_mark_head_lost(sk, 1, 1);
2285         }
2286 }
2287 
2288 /* CWND moderation, preventing bursts due to too big ACKs
2289  * in dubious situations.
2290  */
2291 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2292 {
2293         tp->snd_cwnd = min(tp->snd_cwnd,
2294                            tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2295         tp->snd_cwnd_stamp = tcp_time_stamp;
2296 }
2297 
2298 /* Nothing was retransmitted or returned timestamp is less
2299  * than timestamp of the first retransmission.
2300  */
2301 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2302 {
2303         return !tp->retrans_stamp ||
2304                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2305                  before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2306 }
2307 
2308 /* Undo procedures. */
2309 
2310 /* We can clear retrans_stamp when there are no retransmissions in the
2311  * window. It would seem that it is trivially available for us in
2312  * tp->retrans_out, however, that kind of assumptions doesn't consider
2313  * what will happen if errors occur when sending retransmission for the
2314  * second time. ...It could the that such segment has only
2315  * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2316  * the head skb is enough except for some reneging corner cases that
2317  * are not worth the effort.
2318  *
2319  * Main reason for all this complexity is the fact that connection dying
2320  * time now depends on the validity of the retrans_stamp, in particular,
2321  * that successive retransmissions of a segment must not advance
2322  * retrans_stamp under any conditions.
2323  */
2324 static bool tcp_any_retrans_done(const struct sock *sk)
2325 {
2326         const struct tcp_sock *tp = tcp_sk(sk);
2327         struct sk_buff *skb;
2328 
2329         if (tp->retrans_out)
2330                 return true;
2331 
2332         skb = tcp_write_queue_head(sk);
2333         if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2334                 return true;
2335 
2336         return false;
2337 }
2338 
2339 #if FASTRETRANS_DEBUG > 1
2340 static void DBGUNDO(struct sock *sk, const char *msg)
2341 {
2342         struct tcp_sock *tp = tcp_sk(sk);
2343         struct inet_sock *inet = inet_sk(sk);
2344 
2345         if (sk->sk_family == AF_INET) {
2346                 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2347                          msg,
2348                          &inet->inet_daddr, ntohs(inet->inet_dport),
2349                          tp->snd_cwnd, tcp_left_out(tp),
2350                          tp->snd_ssthresh, tp->prior_ssthresh,
2351                          tp->packets_out);
2352         }
2353 #if IS_ENABLED(CONFIG_IPV6)
2354         else if (sk->sk_family == AF_INET6) {
2355                 struct ipv6_pinfo *np = inet6_sk(sk);
2356                 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2357                          msg,
2358                          &np->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 #endif
2364 }
2365 #else
2366 #define DBGUNDO(x...) do { } while (0)
2367 #endif
2368 
2369 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2370 {
2371         struct tcp_sock *tp = tcp_sk(sk);
2372 
2373         if (unmark_loss) {
2374                 struct sk_buff *skb;
2375 
2376                 tcp_for_write_queue(skb, sk) {
2377                         if (skb == tcp_send_head(sk))
2378                                 break;
2379                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2380                 }
2381                 tp->lost_out = 0;
2382                 tcp_clear_all_retrans_hints(tp);
2383         }
2384 
2385         if (tp->prior_ssthresh) {
2386                 const struct inet_connection_sock *icsk = inet_csk(sk);
2387 
2388                 if (icsk->icsk_ca_ops->undo_cwnd)
2389                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2390                 else
2391                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2392 
2393                 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2394                         tp->snd_ssthresh = tp->prior_ssthresh;
2395                         tcp_ecn_withdraw_cwr(tp);
2396                 }
2397         } else {
2398                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2399         }
2400         tp->snd_cwnd_stamp = tcp_time_stamp;
2401         tp->undo_marker = 0;
2402 }
2403 
2404 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2405 {
2406         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2407 }
2408 
2409 /* People celebrate: "We love our President!" */
2410 static bool tcp_try_undo_recovery(struct sock *sk)
2411 {
2412         struct tcp_sock *tp = tcp_sk(sk);
2413 
2414         if (tcp_may_undo(tp)) {
2415                 int mib_idx;
2416 
2417                 /* Happy end! We did not retransmit anything
2418                  * or our original transmission succeeded.
2419                  */
2420                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2421                 tcp_undo_cwnd_reduction(sk, false);
2422                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2423                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2424                 else
2425                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2426 
2427                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2428         }
2429         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2430                 /* Hold old state until something *above* high_seq
2431                  * is ACKed. For Reno it is MUST to prevent false
2432                  * fast retransmits (RFC2582). SACK TCP is safe. */
2433                 tcp_moderate_cwnd(tp);
2434                 if (!tcp_any_retrans_done(sk))
2435                         tp->retrans_stamp = 0;
2436                 return true;
2437         }
2438         tcp_set_ca_state(sk, TCP_CA_Open);
2439         return false;
2440 }
2441 
2442 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2443 static bool tcp_try_undo_dsack(struct sock *sk)
2444 {
2445         struct tcp_sock *tp = tcp_sk(sk);
2446 
2447         if (tp->undo_marker && !tp->undo_retrans) {
2448                 DBGUNDO(sk, "D-SACK");
2449                 tcp_undo_cwnd_reduction(sk, false);
2450                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2451                 return true;
2452         }
2453         return false;
2454 }
2455 
2456 /* Undo during loss recovery after partial ACK or using F-RTO. */
2457 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2458 {
2459         struct tcp_sock *tp = tcp_sk(sk);
2460 
2461         if (frto_undo || tcp_may_undo(tp)) {
2462                 tcp_undo_cwnd_reduction(sk, true);
2463 
2464                 DBGUNDO(sk, "partial loss");
2465                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2466                 if (frto_undo)
2467                         NET_INC_STATS_BH(sock_net(sk),
2468                                          LINUX_MIB_TCPSPURIOUSRTOS);
2469                 inet_csk(sk)->icsk_retransmits = 0;
2470                 if (frto_undo || tcp_is_sack(tp))
2471                         tcp_set_ca_state(sk, TCP_CA_Open);
2472                 return true;
2473         }
2474         return false;
2475 }
2476 
2477 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2478  * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2479  * It computes the number of packets to send (sndcnt) based on packets newly
2480  * delivered:
2481  *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2482  *      cwnd reductions across a full RTT.
2483  *   2) If packets in flight is lower than ssthresh (such as due to excess
2484  *      losses and/or application stalls), do not perform any further cwnd
2485  *      reductions, but instead slow start up to ssthresh.
2486  */
2487 static void tcp_init_cwnd_reduction(struct sock *sk)
2488 {
2489         struct tcp_sock *tp = tcp_sk(sk);
2490 
2491         tp->high_seq = tp->snd_nxt;
2492         tp->tlp_high_seq = 0;
2493         tp->snd_cwnd_cnt = 0;
2494         tp->prior_cwnd = tp->snd_cwnd;
2495         tp->prr_delivered = 0;
2496         tp->prr_out = 0;
2497         tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2498         tcp_ecn_queue_cwr(tp);
2499 }
2500 
2501 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2502                                int fast_rexmit)
2503 {
2504         struct tcp_sock *tp = tcp_sk(sk);
2505         int sndcnt = 0;
2506         int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2507         int newly_acked_sacked = prior_unsacked -
2508                                  (tp->packets_out - tp->sacked_out);
2509 
2510         tp->prr_delivered += newly_acked_sacked;
2511         if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2512                 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2513                                tp->prior_cwnd - 1;
2514                 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2515         } else {
2516                 sndcnt = min_t(int, delta,
2517                                max_t(int, tp->prr_delivered - tp->prr_out,
2518                                      newly_acked_sacked) + 1);
2519         }
2520 
2521         sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2522         tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2523 }
2524 
2525 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2526 {
2527         struct tcp_sock *tp = tcp_sk(sk);
2528 
2529         /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2530         if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2531             (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2532                 tp->snd_cwnd = tp->snd_ssthresh;
2533                 tp->snd_cwnd_stamp = tcp_time_stamp;
2534         }
2535         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2536 }
2537 
2538 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2539 void tcp_enter_cwr(struct sock *sk)
2540 {
2541         struct tcp_sock *tp = tcp_sk(sk);
2542 
2543         tp->prior_ssthresh = 0;
2544         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2545                 tp->undo_marker = 0;
2546                 tcp_init_cwnd_reduction(sk);
2547                 tcp_set_ca_state(sk, TCP_CA_CWR);
2548         }
2549 }
2550 EXPORT_SYMBOL(tcp_enter_cwr);
2551 
2552 static void tcp_try_keep_open(struct sock *sk)
2553 {
2554         struct tcp_sock *tp = tcp_sk(sk);
2555         int state = TCP_CA_Open;
2556 
2557         if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2558                 state = TCP_CA_Disorder;
2559 
2560         if (inet_csk(sk)->icsk_ca_state != state) {
2561                 tcp_set_ca_state(sk, state);
2562                 tp->high_seq = tp->snd_nxt;
2563         }
2564 }
2565 
2566 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2567 {
2568         struct tcp_sock *tp = tcp_sk(sk);
2569 
2570         tcp_verify_left_out(tp);
2571 
2572         if (!tcp_any_retrans_done(sk))
2573                 tp->retrans_stamp = 0;
2574 
2575         if (flag & FLAG_ECE)
2576                 tcp_enter_cwr(sk);
2577 
2578         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2579                 tcp_try_keep_open(sk);
2580         } else {
2581                 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2582         }
2583 }
2584 
2585 static void tcp_mtup_probe_failed(struct sock *sk)
2586 {
2587         struct inet_connection_sock *icsk = inet_csk(sk);
2588 
2589         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2590         icsk->icsk_mtup.probe_size = 0;
2591 }
2592 
2593 static void tcp_mtup_probe_success(struct sock *sk)
2594 {
2595         struct tcp_sock *tp = tcp_sk(sk);
2596         struct inet_connection_sock *icsk = inet_csk(sk);
2597 
2598         /* FIXME: breaks with very large cwnd */
2599         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2600         tp->snd_cwnd = tp->snd_cwnd *
2601                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2602                        icsk->icsk_mtup.probe_size;
2603         tp->snd_cwnd_cnt = 0;
2604         tp->snd_cwnd_stamp = tcp_time_stamp;
2605         tp->snd_ssthresh = tcp_current_ssthresh(sk);
2606 
2607         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2608         icsk->icsk_mtup.probe_size = 0;
2609         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2610 }
2611 
2612 /* Do a simple retransmit without using the backoff mechanisms in
2613  * tcp_timer. This is used for path mtu discovery.
2614  * The socket is already locked here.
2615  */
2616 void tcp_simple_retransmit(struct sock *sk)
2617 {
2618         const struct inet_connection_sock *icsk = inet_csk(sk);
2619         struct tcp_sock *tp = tcp_sk(sk);
2620         struct sk_buff *skb;
2621         unsigned int mss = tcp_current_mss(sk);
2622         u32 prior_lost = tp->lost_out;
2623 
2624         tcp_for_write_queue(skb, sk) {
2625                 if (skb == tcp_send_head(sk))
2626                         break;
2627                 if (tcp_skb_seglen(skb) > mss &&
2628                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2629                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2630                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2631                                 tp->retrans_out -= tcp_skb_pcount(skb);
2632                         }
2633                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2634                 }
2635         }
2636 
2637         tcp_clear_retrans_hints_partial(tp);
2638 
2639         if (prior_lost == tp->lost_out)
2640                 return;
2641 
2642         if (tcp_is_reno(tp))
2643                 tcp_limit_reno_sacked(tp);
2644 
2645         tcp_verify_left_out(tp);
2646 
2647         /* Don't muck with the congestion window here.
2648          * Reason is that we do not increase amount of _data_
2649          * in network, but units changed and effective
2650          * cwnd/ssthresh really reduced now.
2651          */
2652         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2653                 tp->high_seq = tp->snd_nxt;
2654                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2655                 tp->prior_ssthresh = 0;
2656                 tp->undo_marker = 0;
2657                 tcp_set_ca_state(sk, TCP_CA_Loss);
2658         }
2659         tcp_xmit_retransmit_queue(sk);
2660 }
2661 EXPORT_SYMBOL(tcp_simple_retransmit);
2662 
2663 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2664 {
2665         struct tcp_sock *tp = tcp_sk(sk);
2666         int mib_idx;
2667 
2668         if (tcp_is_reno(tp))
2669                 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2670         else
2671                 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2672 
2673         NET_INC_STATS_BH(sock_net(sk), mib_idx);
2674 
2675         tp->prior_ssthresh = 0;
2676         tcp_init_undo(tp);
2677 
2678         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2679                 if (!ece_ack)
2680                         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2681                 tcp_init_cwnd_reduction(sk);
2682         }
2683         tcp_set_ca_state(sk, TCP_CA_Recovery);
2684 }
2685 
2686 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2687  * recovered or spurious. Otherwise retransmits more on partial ACKs.
2688  */
2689 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2690 {
2691         struct tcp_sock *tp = tcp_sk(sk);
2692         bool recovered = !before(tp->snd_una, tp->high_seq);
2693 
2694         if ((flag & FLAG_SND_UNA_ADVANCED) &&
2695             tcp_try_undo_loss(sk, false))
2696                 return;
2697 
2698         if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2699                 /* Step 3.b. A timeout is spurious if not all data are
2700                  * lost, i.e., never-retransmitted data are (s)acked.
2701                  */
2702                 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2703                     tcp_try_undo_loss(sk, true))
2704                         return;
2705 
2706                 if (after(tp->snd_nxt, tp->high_seq)) {
2707                         if (flag & FLAG_DATA_SACKED || is_dupack)
2708                                 tp->frto = 0; /* Step 3.a. loss was real */
2709                 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2710                         tp->high_seq = tp->snd_nxt;
2711                         __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2712                                                   TCP_NAGLE_OFF);
2713                         if (after(tp->snd_nxt, tp->high_seq))
2714                                 return; /* Step 2.b */
2715                         tp->frto = 0;
2716                 }
2717         }
2718 
2719         if (recovered) {
2720                 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2721                 tcp_try_undo_recovery(sk);
2722                 return;
2723         }
2724         if (tcp_is_reno(tp)) {
2725                 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2726                  * delivered. Lower inflight to clock out (re)tranmissions.
2727                  */
2728                 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2729                         tcp_add_reno_sack(sk);
2730                 else if (flag & FLAG_SND_UNA_ADVANCED)
2731                         tcp_reset_reno_sack(tp);
2732         }
2733         tcp_xmit_retransmit_queue(sk);
2734 }
2735 
2736 /* Undo during fast recovery after partial ACK. */
2737 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2738                                  const int prior_unsacked)
2739 {
2740         struct tcp_sock *tp = tcp_sk(sk);
2741 
2742         if (tp->undo_marker && tcp_packet_delayed(tp)) {
2743                 /* Plain luck! Hole if filled with delayed
2744                  * packet, rather than with a retransmit.
2745                  */
2746                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2747 
2748                 /* We are getting evidence that the reordering degree is higher
2749                  * than we realized. If there are no retransmits out then we
2750                  * can undo. Otherwise we clock out new packets but do not
2751                  * mark more packets lost or retransmit more.
2752                  */
2753                 if (tp->retrans_out) {
2754                         tcp_cwnd_reduction(sk, prior_unsacked, 0);
2755                         return true;
2756                 }
2757 
2758                 if (!tcp_any_retrans_done(sk))
2759                         tp->retrans_stamp = 0;
2760 
2761                 DBGUNDO(sk, "partial recovery");
2762                 tcp_undo_cwnd_reduction(sk, true);
2763                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2764                 tcp_try_keep_open(sk);
2765                 return true;
2766         }
2767         return false;
2768 }
2769 
2770 /* Process an event, which can update packets-in-flight not trivially.
2771  * Main goal of this function is to calculate new estimate for left_out,
2772  * taking into account both packets sitting in receiver's buffer and
2773  * packets lost by network.
2774  *
2775  * Besides that it does CWND reduction, when packet loss is detected
2776  * and changes state of machine.
2777  *
2778  * It does _not_ decide what to send, it is made in function
2779  * tcp_xmit_retransmit_queue().
2780  */
2781 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2782                                   const int prior_unsacked,
2783                                   bool is_dupack, int flag)
2784 {
2785         struct inet_connection_sock *icsk = inet_csk(sk);
2786         struct tcp_sock *tp = tcp_sk(sk);
2787         bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2788                                     (tcp_fackets_out(tp) > tp->reordering));
2789         int fast_rexmit = 0;
2790 
2791         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2792                 tp->sacked_out = 0;
2793         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2794                 tp->fackets_out = 0;
2795 
2796         /* Now state machine starts.
2797          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2798         if (flag & FLAG_ECE)
2799                 tp->prior_ssthresh = 0;
2800 
2801         /* B. In all the states check for reneging SACKs. */
2802         if (tcp_check_sack_reneging(sk, flag))
2803                 return;
2804 
2805         /* C. Check consistency of the current state. */
2806         tcp_verify_left_out(tp);
2807 
2808         /* D. Check state exit conditions. State can be terminated
2809          *    when high_seq is ACKed. */
2810         if (icsk->icsk_ca_state == TCP_CA_Open) {
2811                 WARN_ON(tp->retrans_out != 0);
2812                 tp->retrans_stamp = 0;
2813         } else if (!before(tp->snd_una, tp->high_seq)) {
2814                 switch (icsk->icsk_ca_state) {
2815                 case TCP_CA_CWR:
2816                         /* CWR is to be held something *above* high_seq
2817                          * is ACKed for CWR bit to reach receiver. */
2818                         if (tp->snd_una != tp->high_seq) {
2819                                 tcp_end_cwnd_reduction(sk);
2820                                 tcp_set_ca_state(sk, TCP_CA_Open);
2821                         }
2822                         break;
2823 
2824                 case TCP_CA_Recovery:
2825                         if (tcp_is_reno(tp))
2826                                 tcp_reset_reno_sack(tp);
2827                         if (tcp_try_undo_recovery(sk))
2828                                 return;
2829                         tcp_end_cwnd_reduction(sk);
2830                         break;
2831                 }
2832         }
2833 
2834         /* E. Process state. */
2835         switch (icsk->icsk_ca_state) {
2836         case TCP_CA_Recovery:
2837                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2838                         if (tcp_is_reno(tp) && is_dupack)
2839                                 tcp_add_reno_sack(sk);
2840                 } else {
2841                         if (tcp_try_undo_partial(sk, acked, prior_unsacked))
2842                                 return;
2843                         /* Partial ACK arrived. Force fast retransmit. */
2844                         do_lost = tcp_is_reno(tp) ||
2845                                   tcp_fackets_out(tp) > tp->reordering;
2846                 }
2847                 if (tcp_try_undo_dsack(sk)) {
2848                         tcp_try_keep_open(sk);
2849                         return;
2850                 }
2851                 break;
2852         case TCP_CA_Loss:
2853                 tcp_process_loss(sk, flag, is_dupack);
2854                 if (icsk->icsk_ca_state != TCP_CA_Open)
2855                         return;
2856                 /* Fall through to processing in Open state. */
2857         default:
2858                 if (tcp_is_reno(tp)) {
2859                         if (flag & FLAG_SND_UNA_ADVANCED)
2860                                 tcp_reset_reno_sack(tp);
2861                         if (is_dupack)
2862                                 tcp_add_reno_sack(sk);
2863                 }
2864 
2865                 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2866                         tcp_try_undo_dsack(sk);
2867 
2868                 if (!tcp_time_to_recover(sk, flag)) {
2869                         tcp_try_to_open(sk, flag, prior_unsacked);
2870                         return;
2871                 }
2872 
2873                 /* MTU probe failure: don't reduce cwnd */
2874                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2875                     icsk->icsk_mtup.probe_size &&
2876                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2877                         tcp_mtup_probe_failed(sk);
2878                         /* Restores the reduction we did in tcp_mtup_probe() */
2879                         tp->snd_cwnd++;
2880                         tcp_simple_retransmit(sk);
2881                         return;
2882                 }
2883 
2884                 /* Otherwise enter Recovery state */
2885                 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2886                 fast_rexmit = 1;
2887         }
2888 
2889         if (do_lost)
2890                 tcp_update_scoreboard(sk, fast_rexmit);
2891         tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
2892         tcp_xmit_retransmit_queue(sk);
2893 }
2894 
2895 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2896                                       long seq_rtt_us, long sack_rtt_us)
2897 {
2898         const struct tcp_sock *tp = tcp_sk(sk);
2899 
2900         /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2901          * broken middle-boxes or peers may corrupt TS-ECR fields. But
2902          * Karn's algorithm forbids taking RTT if some retransmitted data
2903          * is acked (RFC6298).
2904          */
2905         if (flag & FLAG_RETRANS_DATA_ACKED)
2906                 seq_rtt_us = -1L;
2907 
2908         if (seq_rtt_us < 0)
2909                 seq_rtt_us = sack_rtt_us;
2910 
2911         /* RTTM Rule: A TSecr value received in a segment is used to
2912          * update the averaged RTT measurement only if the segment
2913          * acknowledges some new data, i.e., only if it advances the
2914          * left edge of the send window.
2915          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2916          */
2917         if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2918             flag & FLAG_ACKED)
2919                 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2920 
2921         if (seq_rtt_us < 0)
2922                 return false;
2923 
2924         tcp_rtt_estimator(sk, seq_rtt_us);
2925         tcp_set_rto(sk);
2926 
2927         /* RFC6298: only reset backoff on valid RTT measurement. */
2928         inet_csk(sk)->icsk_backoff = 0;
2929         return true;
2930 }
2931 
2932 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2933 static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
2934 {
2935         struct tcp_sock *tp = tcp_sk(sk);
2936         long seq_rtt_us = -1L;
2937 
2938         if (synack_stamp && !tp->total_retrans)
2939                 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - synack_stamp);
2940 
2941         /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2942          * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2943          */
2944         if (!tp->srtt_us)
2945                 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt_us, -1L);
2946 }
2947 
2948 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2949 {
2950         const struct inet_connection_sock *icsk = inet_csk(sk);
2951 
2952         icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2953         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2954 }
2955 
2956 /* Restart timer after forward progress on connection.
2957  * RFC2988 recommends to restart timer to now+rto.
2958  */
2959 void tcp_rearm_rto(struct sock *sk)
2960 {
2961         const struct inet_connection_sock *icsk = inet_csk(sk);
2962         struct tcp_sock *tp = tcp_sk(sk);
2963 
2964         /* If the retrans timer is currently being used by Fast Open
2965          * for SYN-ACK retrans purpose, stay put.
2966          */
2967         if (tp->fastopen_rsk)
2968                 return;
2969 
2970         if (!tp->packets_out) {
2971                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2972         } else {
2973                 u32 rto = inet_csk(sk)->icsk_rto;
2974                 /* Offset the time elapsed after installing regular RTO */
2975                 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2976                     icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2977                         struct sk_buff *skb = tcp_write_queue_head(sk);
2978                         const u32 rto_time_stamp =
2979                                 tcp_skb_timestamp(skb) + rto;
2980                         s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2981                         /* delta may not be positive if the socket is locked
2982                          * when the retrans timer fires and is rescheduled.
2983                          */
2984                         if (delta > 0)
2985                                 rto = delta;
2986                 }
2987                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2988                                           TCP_RTO_MAX);
2989         }
2990 }
2991 
2992 /* This function is called when the delayed ER timer fires. TCP enters
2993  * fast recovery and performs fast-retransmit.
2994  */
2995 void tcp_resume_early_retransmit(struct sock *sk)
2996 {
2997         struct tcp_sock *tp = tcp_sk(sk);
2998 
2999         tcp_rearm_rto(sk);
3000 
3001         /* Stop if ER is disabled after the delayed ER timer is scheduled */
3002         if (!tp->do_early_retrans)
3003                 return;
3004 
3005         tcp_enter_recovery(sk, false);
3006         tcp_update_scoreboard(sk, 1);
3007         tcp_xmit_retransmit_queue(sk);
3008 }
3009 
3010 /* If we get here, the whole TSO packet has not been acked. */
3011 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3012 {
3013         struct tcp_sock *tp = tcp_sk(sk);
3014         u32 packets_acked;
3015 
3016         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3017 
3018         packets_acked = tcp_skb_pcount(skb);
3019         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3020                 return 0;
3021         packets_acked -= tcp_skb_pcount(skb);
3022 
3023         if (packets_acked) {
3024                 BUG_ON(tcp_skb_pcount(skb) == 0);
3025                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3026         }
3027 
3028         return packets_acked;
3029 }
3030 
3031 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3032                            u32 prior_snd_una)
3033 {
3034         const struct skb_shared_info *shinfo;
3035 
3036         /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3037         if (likely(!(sk->sk_tsflags & SOF_TIMESTAMPING_TX_ACK)))
3038                 return;
3039 
3040         shinfo = skb_shinfo(skb);
3041         if ((shinfo->tx_flags & SKBTX_ACK_TSTAMP) &&
3042             between(shinfo->tskey, prior_snd_una, tcp_sk(sk)->snd_una - 1))
3043                 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3044 }
3045 
3046 /* Remove acknowledged frames from the retransmission queue. If our packet
3047  * is before the ack sequence we can discard it as it's confirmed to have
3048  * arrived at the other end.
3049  */
3050 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3051                                u32 prior_snd_una,
3052                                struct tcp_sacktag_state *sack)
3053 {
3054         const struct inet_connection_sock *icsk = inet_csk(sk);
3055         struct skb_mstamp first_ackt, last_ackt, now;
3056         struct tcp_sock *tp = tcp_sk(sk);
3057         u32 prior_sacked = tp->sacked_out;
3058         u32 reord = tp->packets_out;
3059         bool fully_acked = true;
3060         long sack_rtt_us = -1L;
3061         long seq_rtt_us = -1L;
3062         long ca_rtt_us = -1L;
3063         struct sk_buff *skb;
3064         u32 pkts_acked = 0;
3065         bool rtt_update;
3066         int flag = 0;
3067 
3068         first_ackt.v64 = 0;
3069 
3070         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3071                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3072                 u8 sacked = scb->sacked;
3073                 u32 acked_pcount;
3074 
3075                 tcp_ack_tstamp(sk, skb, prior_snd_una);
3076 
3077                 /* Determine how many packets and what bytes were acked, tso and else */
3078                 if (after(scb->end_seq, tp->snd_una)) {
3079                         if (tcp_skb_pcount(skb) == 1 ||
3080                             !after(tp->snd_una, scb->seq))
3081                                 break;
3082 
3083                         acked_pcount = tcp_tso_acked(sk, skb);
3084                         if (!acked_pcount)
3085                                 break;
3086 
3087                         fully_acked = false;
3088                 } else {
3089                         /* Speedup tcp_unlink_write_queue() and next loop */
3090                         prefetchw(skb->next);
3091                         acked_pcount = tcp_skb_pcount(skb);
3092                 }
3093 
3094                 if (unlikely(sacked & TCPCB_RETRANS)) {
3095                         if (sacked & TCPCB_SACKED_RETRANS)
3096                                 tp->retrans_out -= acked_pcount;
3097                         flag |= FLAG_RETRANS_DATA_ACKED;
3098                 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3099                         last_ackt = skb->skb_mstamp;
3100                         WARN_ON_ONCE(last_ackt.v64 == 0);
3101                         if (!first_ackt.v64)
3102                                 first_ackt = last_ackt;
3103 
3104                         reord = min(pkts_acked, reord);
3105                         if (!after(scb->end_seq, tp->high_seq))
3106                                 flag |= FLAG_ORIG_SACK_ACKED;
3107                 }
3108 
3109                 if (sacked & TCPCB_SACKED_ACKED)
3110                         tp->sacked_out -= acked_pcount;
3111                 if (sacked & TCPCB_LOST)
3112                         tp->lost_out -= acked_pcount;
3113 
3114                 tp->packets_out -= acked_pcount;
3115                 pkts_acked += acked_pcount;
3116 
3117                 /* Initial outgoing SYN's get put onto the write_queue
3118                  * just like anything else we transmit.  It is not
3119                  * true data, and if we misinform our callers that
3120                  * this ACK acks real data, we will erroneously exit
3121                  * connection startup slow start one packet too
3122                  * quickly.  This is severely frowned upon behavior.
3123                  */
3124                 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3125                         flag |= FLAG_DATA_ACKED;
3126                 } else {
3127                         flag |= FLAG_SYN_ACKED;
3128                         tp->retrans_stamp = 0;
3129                 }
3130 
3131                 if (!fully_acked)
3132                         break;
3133 
3134                 tcp_unlink_write_queue(skb, sk);
3135                 sk_wmem_free_skb(sk, skb);
3136                 if (unlikely(skb == tp->retransmit_skb_hint))
3137                         tp->retransmit_skb_hint = NULL;
3138                 if (unlikely(skb == tp->lost_skb_hint))
3139                         tp->lost_skb_hint = NULL;
3140         }
3141 
3142         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3143                 tp->snd_up = tp->snd_una;
3144 
3145         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3146                 flag |= FLAG_SACK_RENEGING;
3147 
3148         skb_mstamp_get(&now);
3149         if (likely(first_ackt.v64)) {
3150                 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3151                 ca_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3152         }
3153         if (sack->first_sackt.v64) {
3154                 sack_rtt_us = skb_mstamp_us_delta(&now, &sack->first_sackt);
3155                 ca_rtt_us = skb_mstamp_us_delta(&now, &sack->last_sackt);
3156         }
3157 
3158         rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us);
3159 
3160         if (flag & FLAG_ACKED) {
3161                 tcp_rearm_rto(sk);
3162                 if (unlikely(icsk->icsk_mtup.probe_size &&
3163                              !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3164                         tcp_mtup_probe_success(sk);
3165                 }
3166 
3167                 if (tcp_is_reno(tp)) {
3168                         tcp_remove_reno_sacks(sk, pkts_acked);
3169                 } else {
3170                         int delta;
3171 
3172                         /* Non-retransmitted hole got filled? That's reordering */
3173                         if (reord < prior_fackets)
3174                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3175 
3176                         delta = tcp_is_fack(tp) ? pkts_acked :
3177                                                   prior_sacked - tp->sacked_out;
3178                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3179                 }
3180 
3181                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3182 
3183         } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3184                    sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3185                 /* Do not re-arm RTO if the sack RTT is measured from data sent
3186                  * after when the head was last (re)transmitted. Otherwise the
3187                  * timeout may continue to extend in loss recovery.
3188                  */
3189                 tcp_rearm_rto(sk);
3190         }
3191 
3192         if (icsk->icsk_ca_ops->pkts_acked)
3193                 icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked, ca_rtt_us);
3194 
3195 #if FASTRETRANS_DEBUG > 0
3196         WARN_ON((int)tp->sacked_out < 0);
3197         WARN_ON((int)tp->lost_out < 0);
3198         WARN_ON((int)tp->retrans_out < 0);
3199         if (!tp->packets_out && tcp_is_sack(tp)) {
3200                 icsk = inet_csk(sk);
3201                 if (tp->lost_out) {
3202                         pr_debug("Leak l=%u %d\n",
3203                                  tp->lost_out, icsk->icsk_ca_state);
3204                         tp->lost_out = 0;
3205                 }
3206                 if (tp->sacked_out) {
3207                         pr_debug("Leak s=%u %d\n",
3208                                  tp->sacked_out, icsk->icsk_ca_state);
3209                         tp->sacked_out = 0;
3210                 }
3211                 if (tp->retrans_out) {
3212                         pr_debug("Leak r=%u %d\n",
3213                                  tp->retrans_out, icsk->icsk_ca_state);
3214                         tp->retrans_out = 0;
3215                 }
3216         }
3217 #endif
3218         return flag;
3219 }
3220 
3221 static void tcp_ack_probe(struct sock *sk)
3222 {
3223         const struct tcp_sock *tp = tcp_sk(sk);
3224         struct inet_connection_sock *icsk = inet_csk(sk);
3225 
3226         /* Was it a usable window open? */
3227 
3228         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3229                 icsk->icsk_backoff = 0;
3230                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3231                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3232                  * This function is not for random using!
3233                  */
3234         } else {
3235                 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3236 
3237                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3238                                           when, TCP_RTO_MAX);
3239         }
3240 }
3241 
3242 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3243 {
3244         return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3245                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3246 }
3247 
3248 /* Decide wheather to run the increase function of congestion control. */
3249 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3250 {
3251         if (tcp_in_cwnd_reduction(sk))
3252                 return false;
3253 
3254         /* If reordering is high then always grow cwnd whenever data is
3255          * delivered regardless of its ordering. Otherwise stay conservative
3256          * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3257          * new SACK or ECE mark may first advance cwnd here and later reduce
3258          * cwnd in tcp_fastretrans_alert() based on more states.
3259          */
3260         if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3261                 return flag & FLAG_FORWARD_PROGRESS;
3262 
3263         return flag & FLAG_DATA_ACKED;
3264 }
3265 
3266 /* Check that window update is acceptable.
3267  * The function assumes that snd_una<=ack<=snd_next.
3268  */
3269 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3270                                         const u32 ack, const u32 ack_seq,
3271                                         const u32 nwin)
3272 {
3273         return  after(ack, tp->snd_una) ||
3274                 after(ack_seq, tp->snd_wl1) ||
3275                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3276 }
3277 
3278 /* If we update tp->snd_una, also update tp->bytes_acked */
3279 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3280 {
3281         u32 delta = ack - tp->snd_una;
3282 
3283         u64_stats_update_begin(&tp->syncp);
3284         tp->bytes_acked += delta;
3285         u64_stats_update_end(&tp->syncp);
3286         tp->snd_una = ack;
3287 }
3288 
3289 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3290 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3291 {
3292         u32 delta = seq - tp->rcv_nxt;
3293 
3294         u64_stats_update_begin(&tp->syncp);
3295         tp->bytes_received += delta;
3296         u64_stats_update_end(&tp->syncp);
3297         tp->rcv_nxt = seq;
3298 }
3299 
3300 /* Update our send window.
3301  *
3302  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3303  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3304  */
3305 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3306                                  u32 ack_seq)
3307 {
3308         struct tcp_sock *tp = tcp_sk(sk);
3309         int flag = 0;
3310         u32 nwin = ntohs(tcp_hdr(skb)->window);
3311 
3312         if (likely(!tcp_hdr(skb)->syn))
3313                 nwin <<= tp->rx_opt.snd_wscale;
3314 
3315         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3316                 flag |= FLAG_WIN_UPDATE;
3317                 tcp_update_wl(tp, ack_seq);
3318 
3319                 if (tp->snd_wnd != nwin) {
3320                         tp->snd_wnd = nwin;
3321 
3322                         /* Note, it is the only place, where
3323                          * fast path is recovered for sending TCP.
3324                          */
3325                         tp->pred_flags = 0;
3326                         tcp_fast_path_check(sk);
3327 
3328                         if (nwin > tp->max_window) {
3329                                 tp->max_window = nwin;
3330                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3331                         }
3332                 }
3333         }
3334 
3335         tcp_snd_una_update(tp, ack);
3336 
3337         return flag;
3338 }
3339 
3340 /* Return true if we're currently rate-limiting out-of-window ACKs and
3341  * thus shouldn't send a dupack right now. We rate-limit dupacks in
3342  * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3343  * attacks that send repeated SYNs or ACKs for the same connection. To
3344  * do this, we do not send a duplicate SYNACK or ACK if the remote
3345  * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3346  */
3347 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3348                           int mib_idx, u32 *last_oow_ack_time)
3349 {
3350         /* Data packets without SYNs are not likely part of an ACK loop. */
3351         if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3352             !tcp_hdr(skb)->syn)
3353                 goto not_rate_limited;
3354 
3355         if (*last_oow_ack_time) {
3356                 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3357 
3358                 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3359                         NET_INC_STATS_BH(net, mib_idx);
3360                         return true;    /* rate-limited: don't send yet! */
3361                 }
3362         }
3363 
3364         *last_oow_ack_time = tcp_time_stamp;
3365 
3366 not_rate_limited:
3367         return false;   /* not rate-limited: go ahead, send dupack now! */
3368 }
3369 
3370 /* RFC 5961 7 [ACK Throttling] */
3371 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3372 {
3373         /* unprotected vars, we dont care of overwrites */
3374         static u32 challenge_timestamp;
3375         static unsigned int challenge_count;
3376         struct tcp_sock *tp = tcp_sk(sk);
3377         u32 now;
3378 
3379         /* First check our per-socket dupack rate limit. */
3380         if (tcp_oow_rate_limited(sock_net(sk), skb,
3381                                  LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3382                                  &tp->last_oow_ack_time))
3383                 return;
3384 
3385         /* Then check the check host-wide RFC 5961 rate limit. */
3386         now = jiffies / HZ;
3387         if (now != challenge_timestamp) {
3388                 challenge_timestamp = now;
3389                 challenge_count = 0;
3390         }
3391         if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3392                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3393                 tcp_send_ack(sk);
3394         }
3395 }
3396 
3397 static void tcp_store_ts_recent(struct tcp_sock *tp)
3398 {
3399         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3400         tp->rx_opt.ts_recent_stamp = get_seconds();
3401 }
3402 
3403 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3404 {
3405         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3406                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3407                  * extra check below makes sure this can only happen
3408                  * for pure ACK frames.  -DaveM
3409                  *
3410                  * Not only, also it occurs for expired timestamps.
3411                  */
3412 
3413                 if (tcp_paws_check(&tp->rx_opt, 0))
3414                         tcp_store_ts_recent(tp);
3415         }
3416 }
3417 
3418 /* This routine deals with acks during a TLP episode.
3419  * We mark the end of a TLP episode on receiving TLP dupack or when
3420  * ack is after tlp_high_seq.
3421  * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3422  */
3423 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3424 {
3425         struct tcp_sock *tp = tcp_sk(sk);
3426 
3427         if (before(ack, tp->tlp_high_seq))
3428                 return;
3429 
3430         if (flag & FLAG_DSACKING_ACK) {
3431                 /* This DSACK means original and TLP probe arrived; no loss */
3432                 tp->tlp_high_seq = 0;
3433         } else if (after(ack, tp->tlp_high_seq)) {
3434                 /* ACK advances: there was a loss, so reduce cwnd. Reset
3435                  * tlp_high_seq in tcp_init_cwnd_reduction()
3436                  */
3437                 tcp_init_cwnd_reduction(sk);
3438                 tcp_set_ca_state(sk, TCP_CA_CWR);
3439                 tcp_end_cwnd_reduction(sk);
3440                 tcp_try_keep_open(sk);
3441                 NET_INC_STATS_BH(sock_net(sk),
3442                                  LINUX_MIB_TCPLOSSPROBERECOVERY);
3443         } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3444                              FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3445                 /* Pure dupack: original and TLP probe arrived; no loss */
3446                 tp->tlp_high_seq = 0;
3447         }
3448 }
3449 
3450 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3451 {
3452         const struct inet_connection_sock *icsk = inet_csk(sk);
3453 
3454         if (icsk->icsk_ca_ops->in_ack_event)
3455                 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3456 }
3457 
3458 /* This routine deals with incoming acks, but not outgoing ones. */
3459 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3460 {
3461         struct inet_connection_sock *icsk = inet_csk(sk);
3462         struct tcp_sock *tp = tcp_sk(sk);
3463         struct tcp_sacktag_state sack_state;
3464         u32 prior_snd_una = tp->snd_una;
3465         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3466         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3467         bool is_dupack = false;
3468         u32 prior_fackets;
3469         int prior_packets = tp->packets_out;
3470         const int prior_unsacked = tp->packets_out - tp->sacked_out;
3471         int acked = 0; /* Number of packets newly acked */
3472 
3473         sack_state.first_sackt.v64 = 0;
3474 
3475         /* We very likely will need to access write queue head. */
3476         prefetchw(sk->sk_write_queue.next);
3477 
3478         /* If the ack is older than previous acks
3479          * then we can probably ignore it.
3480          */
3481         if (before(ack, prior_snd_una)) {
3482                 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3483                 if (before(ack, prior_snd_una - tp->max_window)) {
3484                         tcp_send_challenge_ack(sk, skb);
3485                         return -1;
3486                 }
3487                 goto old_ack;
3488         }
3489 
3490         /* If the ack includes data we haven't sent yet, discard
3491          * this segment (RFC793 Section 3.9).
3492          */
3493         if (after(ack, tp->snd_nxt))
3494                 goto invalid_ack;
3495 
3496         if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3497             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3498                 tcp_rearm_rto(sk);
3499 
3500         if (after(ack, prior_snd_una)) {
3501                 flag |= FLAG_SND_UNA_ADVANCED;
3502                 icsk->icsk_retransmits = 0;
3503         }
3504 
3505         prior_fackets = tp->fackets_out;
3506 
3507         /* ts_recent update must be made after we are sure that the packet
3508          * is in window.
3509          */
3510         if (flag & FLAG_UPDATE_TS_RECENT)
3511                 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3512 
3513         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3514                 /* Window is constant, pure forward advance.
3515                  * No more checks are required.
3516                  * Note, we use the fact that SND.UNA>=SND.WL2.
3517                  */
3518                 tcp_update_wl(tp, ack_seq);
3519                 tcp_snd_una_update(tp, ack);
3520                 flag |= FLAG_WIN_UPDATE;
3521 
3522                 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3523 
3524                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3525         } else {
3526                 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3527 
3528                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3529                         flag |= FLAG_DATA;
3530                 else
3531                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3532 
3533                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3534 
3535                 if (TCP_SKB_CB(skb)->sacked)
3536                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3537                                                         &sack_state);
3538 
3539                 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3540                         flag |= FLAG_ECE;
3541                         ack_ev_flags |= CA_ACK_ECE;
3542                 }
3543 
3544                 if (flag & FLAG_WIN_UPDATE)
3545                         ack_ev_flags |= CA_ACK_WIN_UPDATE;
3546 
3547                 tcp_in_ack_event(sk, ack_ev_flags);
3548         }
3549 
3550         /* We passed data and got it acked, remove any soft error
3551          * log. Something worked...
3552          */
3553         sk->sk_err_soft = 0;
3554         icsk->icsk_probes_out = 0;
3555         tp->rcv_tstamp = tcp_time_stamp;
3556         if (!prior_packets)
3557                 goto no_queue;
3558 
3559         /* See if we can take anything off of the retransmit queue. */
3560         acked = tp->packets_out;
3561         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
3562                                     &sack_state);
3563         acked -= tp->packets_out;
3564 
3565         /* Advance cwnd if state allows */
3566         if (tcp_may_raise_cwnd(sk, flag))
3567                 tcp_cong_avoid(sk, ack, acked);
3568 
3569         if (tcp_ack_is_dubious(sk, flag)) {
3570                 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3571                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3572                                       is_dupack, flag);
3573         }
3574         if (tp->tlp_high_seq)
3575                 tcp_process_tlp_ack(sk, ack, flag);
3576 
3577         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3578                 struct dst_entry *dst = __sk_dst_get(sk);
3579                 if (dst)
3580                         dst_confirm(dst);
3581         }
3582 
3583         if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3584                 tcp_schedule_loss_probe(sk);
3585         tcp_update_pacing_rate(sk);
3586         return 1;
3587 
3588 no_queue:
3589         /* If data was DSACKed, see if we can undo a cwnd reduction. */
3590         if (flag & FLAG_DSACKING_ACK)
3591                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3592                                       is_dupack, flag);
3593         /* If this ack opens up a zero window, clear backoff.  It was
3594          * being used to time the probes, and is probably far higher than
3595          * it needs to be for normal retransmission.
3596          */
3597         if (tcp_send_head(sk))
3598                 tcp_ack_probe(sk);
3599 
3600         if (tp->tlp_high_seq)
3601                 tcp_process_tlp_ack(sk, ack, flag);
3602         return 1;
3603 
3604 invalid_ack:
3605         SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3606         return -1;
3607 
3608 old_ack:
3609         /* If data was SACKed, tag it and see if we should send more data.
3610          * If data was DSACKed, see if we can undo a cwnd reduction.
3611          */
3612         if (TCP_SKB_CB(skb)->sacked) {
3613                 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3614                                                 &sack_state);
3615                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3616                                       is_dupack, flag);
3617         }
3618 
3619         SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3620         return 0;
3621 }
3622 
3623 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3624                                       bool syn, struct tcp_fastopen_cookie *foc,
3625                                       bool exp_opt)
3626 {
3627         /* Valid only in SYN or SYN-ACK with an even length.  */
3628         if (!foc || !syn || len < 0 || (len & 1))
3629                 return;
3630 
3631         if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3632             len <= TCP_FASTOPEN_COOKIE_MAX)
3633                 memcpy(foc->val, cookie, len);
3634         else if (len != 0)
3635                 len = -1;
3636         foc->len = len;
3637         foc->exp = exp_opt;
3638 }
3639 
3640 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3641  * But, this can also be called on packets in the established flow when
3642  * the fast version below fails.
3643  */
3644 void tcp_parse_options(const struct sk_buff *skb,
3645                        struct tcp_options_received *opt_rx, int estab,
3646                        struct tcp_fastopen_cookie *foc)
3647 {
3648         const unsigned char *ptr;
3649         const struct tcphdr *th = tcp_hdr(skb);
3650         int length = (th->doff * 4) - sizeof(struct tcphdr);
3651 
3652         ptr = (const unsigned char *)(th + 1);
3653         opt_rx->saw_tstamp = 0;
3654 
3655         while (length > 0) {
3656                 int opcode = *ptr++;
3657                 int opsize;
3658 
3659                 switch (opcode) {
3660                 case TCPOPT_EOL:
3661                         return;
3662                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3663                         length--;
3664                         continue;
3665                 default:
3666                         opsize = *ptr++;
3667                         if (opsize < 2) /* "silly options" */
3668                                 return;
3669                         if (opsize > length)
3670                                 return; /* don't parse partial options */
3671                         switch (opcode) {
3672                         case TCPOPT_MSS:
3673                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3674                                         u16 in_mss = get_unaligned_be16(ptr);
3675                                         if (in_mss) {
3676                                                 if (opt_rx->user_mss &&
3677                                                     opt_rx->user_mss < in_mss)
3678                                                         in_mss = opt_rx->user_mss;
3679                                                 opt_rx->mss_clamp = in_mss;
3680                                         }
3681                                 }
3682                                 break;
3683                         case TCPOPT_WINDOW:
3684                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3685                                     !estab && sysctl_tcp_window_scaling) {
3686                                         __u8 snd_wscale = *(__u8 *)ptr;
3687                                         opt_rx->wscale_ok = 1;
3688                                         if (snd_wscale > 14) {
3689                                                 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3690                                                                      __func__,
3691                                                                      snd_wscale);
3692                                                 snd_wscale = 14;
3693                                         }
3694                                         opt_rx->snd_wscale = snd_wscale;
3695                                 }
3696                                 break;
3697                         case TCPOPT_TIMESTAMP:
3698                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3699                                     ((estab && opt_rx->tstamp_ok) ||
3700                                      (!estab && sysctl_tcp_timestamps))) {
3701                                         opt_rx->saw_tstamp = 1;
3702                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3703                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3704                                 }
3705                                 break;
3706                         case TCPOPT_SACK_PERM:
3707                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3708                                     !estab && sysctl_tcp_sack) {
3709                                         opt_rx->sack_ok = TCP_SACK_SEEN;
3710                                         tcp_sack_reset(opt_rx);
3711                                 }
3712                                 break;
3713 
3714                         case TCPOPT_SACK:
3715                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3716                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3717                                    opt_rx->sack_ok) {
3718                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3719                                 }
3720                                 break;
3721 #ifdef CONFIG_TCP_MD5SIG
3722                         case TCPOPT_MD5SIG:
3723                                 /*
3724                                  * The MD5 Hash has already been
3725                                  * checked (see tcp_v{4,6}_do_rcv()).
3726                                  */
3727                                 break;
3728 #endif
3729                         case TCPOPT_FASTOPEN:
3730                                 tcp_parse_fastopen_option(
3731                                         opsize - TCPOLEN_FASTOPEN_BASE,
3732                                         ptr, th->syn, foc, false);
3733                                 break;
3734 
3735                         case TCPOPT_EXP:
3736                                 /* Fast Open option shares code 254 using a
3737                                  * 16 bits magic number.
3738                                  */
3739                                 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3740                                     get_unaligned_be16(ptr) ==
3741                                     TCPOPT_FASTOPEN_MAGIC)
3742                                         tcp_parse_fastopen_option(opsize -
3743                                                 TCPOLEN_EXP_FASTOPEN_BASE,
3744                                                 ptr + 2, th->syn, foc, true);
3745                                 break;
3746 
3747                         }
3748                         ptr += opsize-2;
3749                         length -= opsize;
3750                 }
3751         }
3752 }
3753 EXPORT_SYMBOL(tcp_parse_options);
3754 
3755 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3756 {
3757         const __be32 *ptr = (const __be32 *)(th + 1);
3758 
3759         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3760                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3761                 tp->rx_opt.saw_tstamp = 1;
3762                 ++ptr;
3763                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3764                 ++ptr;
3765                 if (*ptr)
3766                         tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3767                 else
3768                         tp->rx_opt.rcv_tsecr = 0;
3769                 return true;
3770         }
3771         return false;
3772 }
3773 
3774 /* Fast parse options. This hopes to only see timestamps.
3775  * If it is wrong it falls back on tcp_parse_options().
3776  */
3777 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3778                                    const struct tcphdr *th, struct tcp_sock *tp)
3779 {
3780         /* In the spirit of fast parsing, compare doff directly to constant
3781          * values.  Because equality is used, short doff can be ignored here.
3782          */
3783         if (th->doff == (sizeof(*th) / 4)) {
3784                 tp->rx_opt.saw_tstamp = 0;
3785                 return false;
3786         } else if (tp->rx_opt.tstamp_ok &&
3787                    th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3788                 if (tcp_parse_aligned_timestamp(tp, th))
3789                         return true;
3790         }
3791 
3792         tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3793         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3794                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3795 
3796         return true;
3797 }
3798 
3799 #ifdef CONFIG_TCP_MD5SIG
3800 /*
3801  * Parse MD5 Signature option
3802  */
3803 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3804 {
3805         int length = (th->doff << 2) - sizeof(*th);
3806         const u8 *ptr = (const u8 *)(th + 1);
3807 
3808         /* If the TCP option is too short, we can short cut */
3809         if (length < TCPOLEN_MD5SIG)
3810                 return NULL;
3811 
3812         while (length > 0) {
3813                 int opcode = *ptr++;
3814                 int opsize;
3815 
3816                 switch (opcode) {
3817                 case TCPOPT_EOL:
3818                         return NULL;
3819                 case TCPOPT_NOP:
3820                         length--;
3821                         continue;
3822                 default:
3823                         opsize = *ptr++;
3824                         if (opsize < 2 || opsize > length)
3825                                 return NULL;
3826                         if (opcode == TCPOPT_MD5SIG)
3827                                 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3828                 }
3829                 ptr += opsize - 2;
3830                 length -= opsize;
3831         }
3832         return NULL;
3833 }
3834 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3835 #endif
3836 
3837 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3838  *
3839  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3840  * it can pass through stack. So, the following predicate verifies that
3841  * this segment is not used for anything but congestion avoidance or
3842  * fast retransmit. Moreover, we even are able to eliminate most of such
3843  * second order effects, if we apply some small "replay" window (~RTO)
3844  * to timestamp space.
3845  *
3846  * All these measures still do not guarantee that we reject wrapped ACKs
3847  * on networks with high bandwidth, when sequence space is recycled fastly,
3848  * but it guarantees that such events will be very rare and do not affect
3849  * connection seriously. This doesn't look nice, but alas, PAWS is really
3850  * buggy extension.
3851  *
3852  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3853  * states that events when retransmit arrives after original data are rare.
3854  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3855  * the biggest problem on large power networks even with minor reordering.
3856  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3857  * up to bandwidth of 18Gigabit/sec. 8) ]
3858  */
3859 
3860 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3861 {
3862         const struct tcp_sock *tp = tcp_sk(sk);
3863         const struct tcphdr *th = tcp_hdr(skb);
3864         u32 seq = TCP_SKB_CB(skb)->seq;
3865         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3866 
3867         return (/* 1. Pure ACK with correct sequence number. */
3868                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3869 
3870                 /* 2. ... and duplicate ACK. */
3871                 ack == tp->snd_una &&
3872 
3873                 /* 3. ... and does not update window. */
3874                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3875 
3876                 /* 4. ... and sits in replay window. */
3877                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3878 }
3879 
3880 static inline bool tcp_paws_discard(const struct sock *sk,
3881                                    const struct sk_buff *skb)
3882 {
3883         const struct tcp_sock *tp = tcp_sk(sk);
3884 
3885         return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3886                !tcp_disordered_ack(sk, skb);
3887 }
3888 
3889 /* Check segment sequence number for validity.
3890  *
3891  * Segment controls are considered valid, if the segment
3892  * fits to the window after truncation to the window. Acceptability
3893  * of data (and SYN, FIN, of course) is checked separately.
3894  * See tcp_data_queue(), for example.
3895  *
3896  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3897  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3898  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3899  * (borrowed from freebsd)
3900  */
3901 
3902 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3903 {
3904         return  !before(end_seq, tp->rcv_wup) &&
3905                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3906 }
3907 
3908 /* When we get a reset we do this. */
3909 void tcp_reset(struct sock *sk)
3910 {
3911         /* We want the right error as BSD sees it (and indeed as we do). */
3912         switch (sk->sk_state) {
3913         case TCP_SYN_SENT:
3914                 sk->sk_err = ECONNREFUSED;
3915                 break;
3916         case TCP_CLOSE_WAIT:
3917                 sk->sk_err = EPIPE;
3918                 break;
3919         case TCP_CLOSE:
3920                 return;
3921         default:
3922                 sk->sk_err = ECONNRESET;
3923         }
3924         /* This barrier is coupled with smp_rmb() in tcp_poll() */
3925         smp_wmb();
3926 
3927         if (!sock_flag(sk, SOCK_DEAD))
3928                 sk->sk_error_report(sk);
3929 
3930         tcp_done(sk);
3931 }
3932 
3933 /*
3934  *      Process the FIN bit. This now behaves as it is supposed to work
3935  *      and the FIN takes effect when it is validly part of sequence
3936  *      space. Not before when we get holes.
3937  *
3938  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3939  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3940  *      TIME-WAIT)
3941  *
3942  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3943  *      close and we go into CLOSING (and later onto TIME-WAIT)
3944  *
3945  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3946  */
3947 static void tcp_fin(struct sock *sk)
3948 {
3949         struct tcp_sock *tp = tcp_sk(sk);
3950         const struct dst_entry *dst;
3951 
3952         inet_csk_schedule_ack(sk);
3953 
3954         sk->sk_shutdown |= RCV_SHUTDOWN;
3955         sock_set_flag(sk, SOCK_DONE);
3956 
3957         switch (sk->sk_state) {
3958         case TCP_SYN_RECV:
3959         case TCP_ESTABLISHED:
3960                 /* Move to CLOSE_WAIT */
3961                 tcp_set_state(sk, TCP_CLOSE_WAIT);
3962                 dst = __sk_dst_get(sk);
3963                 if (!dst || !dst_metric(dst, RTAX_QUICKACK))
3964                         inet_csk(sk)->icsk_ack.pingpong = 1;
3965                 break;
3966 
3967         case TCP_CLOSE_WAIT:
3968         case TCP_CLOSING:
3969                 /* Received a retransmission of the FIN, do
3970                  * nothing.
3971                  */
3972                 break;
3973         case TCP_LAST_ACK:
3974                 /* RFC793: Remain in the LAST-ACK state. */
3975                 break;
3976 
3977         case TCP_FIN_WAIT1:
3978                 /* This case occurs when a simultaneous close
3979                  * happens, we must ack the received FIN and
3980                  * enter the CLOSING state.
3981                  */
3982                 tcp_send_ack(sk);
3983                 tcp_set_state(sk, TCP_CLOSING);
3984                 break;
3985         case TCP_FIN_WAIT2:
3986                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3987                 tcp_send_ack(sk);
3988                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3989                 break;
3990         default:
3991                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3992                  * cases we should never reach this piece of code.
3993                  */
3994                 pr_err("%s: Impossible, sk->sk_state=%d\n",
3995                        __func__, sk->sk_state);
3996                 break;
3997         }
3998 
3999         /* It _is_ possible, that we have something out-of-order _after_ FIN.
4000          * Probably, we should reset in this case. For now drop them.
4001          */
4002         __skb_queue_purge(&tp->out_of_order_queue);
4003         if (tcp_is_sack(tp))
4004                 tcp_sack_reset(&tp->rx_opt);
4005         sk_mem_reclaim(sk);
4006 
4007         if (!sock_flag(sk, SOCK_DEAD)) {
4008                 sk->sk_state_change(sk);
4009 
4010                 /* Do not send POLL_HUP for half duplex close. */
4011                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4012                     sk->sk_state == TCP_CLOSE)
4013                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4014                 else
4015                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4016         }
4017 }
4018 
4019 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4020                                   u32 end_seq)
4021 {
4022         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4023                 if (before(seq, sp->start_seq))
4024                         sp->start_seq = seq;
4025                 if (after(end_seq, sp->end_seq))
4026                         sp->end_seq = end_seq;
4027                 return true;
4028         }
4029         return false;
4030 }
4031 
4032 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4033 {
4034         struct tcp_sock *tp = tcp_sk(sk);
4035 
4036         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4037                 int mib_idx;
4038 
4039                 if (before(seq, tp->rcv_nxt))
4040                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4041                 else
4042                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4043 
4044                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4045 
4046                 tp->rx_opt.dsack = 1;
4047                 tp->duplicate_sack[0].start_seq = seq;
4048                 tp->duplicate_sack[0].end_seq = end_seq;
4049         }
4050 }
4051 
4052 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4053 {
4054         struct tcp_sock *tp = tcp_sk(sk);
4055 
4056         if (!tp->rx_opt.dsack)
4057                 tcp_dsack_set(sk, seq, end_seq);
4058         else
4059                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4060 }
4061 
4062 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4063 {
4064         struct tcp_sock *tp = tcp_sk(sk);
4065 
4066         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4067             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4068                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4069                 tcp_enter_quickack_mode(sk);
4070 
4071                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4072                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4073 
4074                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4075                                 end_seq = tp->rcv_nxt;
4076                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4077                 }
4078         }
4079 
4080         tcp_send_ack(sk);
4081 }
4082 
4083 /* These routines update the SACK block as out-of-order packets arrive or
4084  * in-order packets close up the sequence space.
4085  */
4086 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4087 {
4088         int this_sack;
4089         struct tcp_sack_block *sp = &tp->selective_acks[0];
4090         struct tcp_sack_block *swalk = sp + 1;
4091 
4092         /* See if the recent change to the first SACK eats into
4093          * or hits the sequence space of other SACK blocks, if so coalesce.
4094          */
4095         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4096                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4097                         int i;
4098 
4099                         /* Zap SWALK, by moving every further SACK up by one slot.
4100                          * Decrease num_sacks.
4101                          */
4102                         tp->rx_opt.num_sacks--;
4103                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4104                                 sp[i] = sp[i + 1];
4105                         continue;
4106                 }
4107                 this_sack++, swalk++;
4108         }
4109 }
4110 
4111 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4112 {
4113         struct tcp_sock *tp = tcp_sk(sk);
4114         struct tcp_sack_block *sp = &tp->selective_acks[0];
4115         int cur_sacks = tp->rx_opt.num_sacks;
4116         int this_sack;
4117 
4118         if (!cur_sacks)
4119                 goto new_sack;
4120 
4121         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4122                 if (tcp_sack_extend(sp, seq, end_seq)) {
4123                         /* Rotate this_sack to the first one. */
4124                         for (; this_sack > 0; this_sack--, sp--)
4125                                 swap(*sp, *(sp - 1));
4126                         if (cur_sacks > 1)
4127                                 tcp_sack_maybe_coalesce(tp);
4128                         return;
4129                 }
4130         }
4131 
4132         /* Could not find an adjacent existing SACK, build a new one,
4133          * put it at the front, and shift everyone else down.  We
4134          * always know there is at least one SACK present already here.
4135          *
4136          * If the sack array is full, forget about the last one.
4137          */
4138         if (this_sack >= TCP_NUM_SACKS) {
4139                 this_sack--;
4140                 tp->rx_opt.num_sacks--;
4141                 sp--;
4142         }
4143         for (; this_sack > 0; this_sack--, sp--)
4144                 *sp = *(sp - 1);
4145 
4146 new_sack:
4147         /* Build the new head SACK, and we're done. */
4148         sp->start_seq = seq;
4149         sp->end_seq = end_seq;
4150         tp->rx_opt.num_sacks++;
4151 }
4152 
4153 /* RCV.NXT advances, some SACKs should be eaten. */
4154 
4155 static void tcp_sack_remove(struct tcp_sock *tp)
4156 {
4157         struct tcp_sack_block *sp = &tp->selective_acks[0];
4158         int num_sacks = tp->rx_opt.num_sacks;
4159         int this_sack;
4160 
4161         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4162         if (skb_queue_empty(&tp->out_of_order_queue)) {
4163                 tp->rx_opt.num_sacks = 0;
4164                 return;
4165         }
4166 
4167         for (this_sack = 0; this_sack < num_sacks;) {
4168                 /* Check if the start of the sack is covered by RCV.NXT. */
4169                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4170                         int i;
4171 
4172                         /* RCV.NXT must cover all the block! */
4173                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4174 
4175                         /* Zap this SACK, by moving forward any other SACKS. */
4176                         for (i = this_sack+1; i < num_sacks; i++)
4177                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4178                         num_sacks--;
4179                         continue;
4180                 }
4181                 this_sack++;
4182                 sp++;
4183         }
4184         tp->rx_opt.num_sacks = num_sacks;
4185 }
4186 
4187 /**
4188  * tcp_try_coalesce - try to merge skb to prior one
4189  * @sk: socket
4190  * @to: prior buffer
4191  * @from: buffer to add in queue
4192  * @fragstolen: pointer to boolean
4193  *
4194  * Before queueing skb @from after @to, try to merge them
4195  * to reduce overall memory use and queue lengths, if cost is small.
4196  * Packets in ofo or receive queues can stay a long time.
4197  * Better try to coalesce them right now to avoid future collapses.
4198  * Returns true if caller should free @from instead of queueing it
4199  */
4200 static bool tcp_try_coalesce(struct sock *sk,
4201                              struct sk_buff *to,
4202                              struct sk_buff *from,
4203                              bool *fragstolen)
4204 {
4205         int delta;
4206 
4207         *fragstolen = false;
4208 
4209         /* Its possible this segment overlaps with prior segment in queue */
4210         if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4211                 return false;
4212 
4213         if (!skb_try_coalesce(to, from, fragstolen, &delta))
4214                 return false;
4215 
4216         atomic_add(delta, &sk->sk_rmem_alloc);
4217         sk_mem_charge(sk, delta);
4218         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4219         TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4220         TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4221         TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4222         return true;
4223 }
4224 
4225 /* This one checks to see if we can put data from the
4226  * out_of_order queue into the receive_queue.
4227  */
4228 static void tcp_ofo_queue(struct sock *sk)
4229 {
4230         struct tcp_sock *tp = tcp_sk(sk);
4231         __u32 dsack_high = tp->rcv_nxt;
4232         struct sk_buff *skb, *tail;
4233         bool fragstolen, eaten;
4234 
4235         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4236                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4237                         break;
4238 
4239                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4240                         __u32 dsack = dsack_high;
4241                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4242                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4243                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4244                 }
4245 
4246                 __skb_unlink(skb, &tp->out_of_order_queue);
4247                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4248                         SOCK_DEBUG(sk, "ofo packet was already received\n");
4249                         __kfree_skb(skb);
4250                         continue;
4251                 }
4252                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4253                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4254                            TCP_SKB_CB(skb)->end_seq);
4255 
4256                 tail = skb_peek_tail(&sk->sk_receive_queue);
4257                 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4258                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4259                 if (!eaten)
4260                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4261                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4262                         tcp_fin(sk);
4263                 if (eaten)
4264                         kfree_skb_partial(skb, fragstolen);
4265         }
4266 }
4267 
4268 static bool tcp_prune_ofo_queue(struct sock *sk);
4269 static int tcp_prune_queue(struct sock *sk);
4270 
4271 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4272                                  unsigned int size)
4273 {
4274         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4275             !sk_rmem_schedule(sk, skb, size)) {
4276 
4277                 if (tcp_prune_queue(sk) < 0)
4278                         return -1;
4279 
4280                 if (!sk_rmem_schedule(sk, skb, size)) {
4281                         if (!tcp_prune_ofo_queue(sk))
4282                                 return -1;
4283 
4284                         if (!sk_rmem_schedule(sk, skb, size))
4285                                 return -1;
4286                 }
4287         }
4288         return 0;
4289 }
4290 
4291 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4292 {
4293         struct tcp_sock *tp = tcp_sk(sk);
4294         struct sk_buff *skb1;
4295         u32 seq, end_seq;
4296 
4297         tcp_ecn_check_ce(tp, skb);
4298 
4299         if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4300                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4301                 __kfree_skb(skb);
4302                 return;
4303         }
4304 
4305         /* Disable header prediction. */
4306         tp->pred_flags = 0;
4307         inet_csk_schedule_ack(sk);
4308 
4309         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4310         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4311                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4312 
4313         skb1 = skb_peek_tail(&tp->out_of_order_queue);
4314         if (!skb1) {
4315                 /* Initial out of order segment, build 1 SACK. */
4316                 if (tcp_is_sack(tp)) {
4317                         tp->rx_opt.num_sacks = 1;
4318                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4319                         tp->selective_acks[0].end_seq =
4320                                                 TCP_SKB_CB(skb)->end_seq;
4321                 }
4322                 __skb_queue_head(&tp->out_of_order_queue, skb);
4323                 goto end;
4324         }
4325 
4326         seq = TCP_SKB_CB(skb)->seq;
4327         end_seq = TCP_SKB_CB(skb)->end_seq;
4328 
4329         if (seq == TCP_SKB_CB(skb1)->end_seq) {
4330                 bool fragstolen;
4331 
4332                 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4333                         __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4334                 } else {
4335                         tcp_grow_window(sk, skb);
4336                         kfree_skb_partial(skb, fragstolen);
4337                         skb = NULL;
4338                 }
4339 
4340                 if (!tp->rx_opt.num_sacks ||
4341                     tp->selective_acks[0].end_seq != seq)
4342                         goto add_sack;
4343 
4344                 /* Common case: data arrive in order after hole. */
4345                 tp->selective_acks[0].end_seq = end_seq;
4346                 goto end;
4347         }
4348 
4349         /* Find place to insert this segment. */
4350         while (1) {
4351                 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4352                         break;
4353                 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4354                         skb1 = NULL;
4355                         break;
4356                 }
4357                 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4358         }
4359 
4360         /* Do skb overlap to previous one? */
4361         if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4362                 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4363                         /* All the bits are present. Drop. */
4364                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4365                         __kfree_skb(skb);
4366                         skb = NULL;
4367                         tcp_dsack_set(sk, seq, end_seq);
4368                         goto add_sack;
4369                 }
4370                 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4371                         /* Partial overlap. */
4372                         tcp_dsack_set(sk, seq,
4373                                       TCP_SKB_CB(skb1)->end_seq);
4374                 } else {
4375                         if (skb_queue_is_first(&tp->out_of_order_queue,
4376                                                skb1))
4377                                 skb1 = NULL;
4378                         else
4379                                 skb1 = skb_queue_prev(
4380                                         &tp->out_of_order_queue,
4381                                         skb1);
4382                 }
4383         }
4384         if (!skb1)
4385                 __skb_queue_head(&tp->out_of_order_queue, skb);
4386         else
4387                 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4388 
4389         /* And clean segments covered by new one as whole. */
4390         while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4391                 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4392 
4393                 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4394                         break;
4395                 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4396                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4397                                          end_seq);
4398                         break;
4399                 }
4400                 __skb_unlink(skb1, &tp->out_of_order_queue);
4401                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4402                                  TCP_SKB_CB(skb1)->end_seq);
4403                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4404                 __kfree_skb(skb1);
4405         }
4406 
4407 add_sack:
4408         if (tcp_is_sack(tp))
4409                 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4410 end:
4411         if (skb) {
4412                 tcp_grow_window(sk, skb);
4413                 skb_set_owner_r(skb, sk);
4414         }
4415 }
4416 
4417 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4418                   bool *fragstolen)
4419 {
4420         int eaten;
4421         struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4422 
4423         __skb_pull(skb, hdrlen);
4424         eaten = (tail &&
4425                  tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4426         tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4427         if (!eaten) {
4428                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4429                 skb_set_owner_r(skb, sk);
4430         }
4431         return eaten;
4432 }
4433 
4434 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4435 {
4436         struct sk_buff *skb;
4437         bool fragstolen;
4438 
4439         if (size == 0)
4440                 return 0;
4441 
4442         skb = alloc_skb(size, sk->sk_allocation);
4443         if (!skb)
4444                 goto err;
4445 
4446         if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4447                 goto err_free;
4448 
4449         if (memcpy_from_msg(skb_put(skb, size), msg, size))
4450                 goto err_free;
4451 
4452         TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4453         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4454         TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4455 
4456         if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4457                 WARN_ON_ONCE(fragstolen); /* should not happen */
4458                 __kfree_skb(skb);
4459         }
4460         return size;
4461 
4462 err_free:
4463         kfree_skb(skb);
4464 err:
4465         return -ENOMEM;
4466 }
4467 
4468 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4469 {
4470         struct tcp_sock *tp = tcp_sk(sk);
4471         int eaten = -1;
4472         bool fragstolen = false;
4473 
4474         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4475                 goto drop;
4476 
4477         skb_dst_drop(skb);
4478         __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4479 
4480         tcp_ecn_accept_cwr(tp, skb);
4481 
4482         tp->rx_opt.dsack = 0;
4483 
4484         /*  Queue data for delivery to the user.
4485          *  Packets in sequence go to the receive queue.
4486          *  Out of sequence packets to the out_of_order_queue.
4487          */
4488         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4489                 if (tcp_receive_window(tp) == 0)
4490                         goto out_of_window;
4491 
4492                 /* Ok. In sequence. In window. */
4493                 if (tp->ucopy.task == current &&
4494                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4495                     sock_owned_by_user(sk) && !tp->urg_data) {
4496                         int chunk = min_t(unsigned int, skb->len,
4497                                           tp->ucopy.len);
4498 
4499                         __set_current_state(TASK_RUNNING);
4500 
4501                         local_bh_enable();
4502                         if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4503                                 tp->ucopy.len -= chunk;
4504                                 tp->copied_seq += chunk;
4505                                 eaten = (chunk == skb->len);
4506                                 tcp_rcv_space_adjust(sk);
4507                         }
4508                         local_bh_disable();
4509                 }
4510 
4511                 if (eaten <= 0) {
4512 queue_and_out:
4513                         if (eaten < 0) {
4514                                 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4515                                         sk_forced_mem_schedule(sk, skb->truesize);
4516                                 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4517                                         goto drop;
4518                         }
4519                         eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4520                 }
4521                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4522                 if (skb->len)
4523                         tcp_event_data_recv(sk, skb);
4524                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4525                         tcp_fin(sk);
4526 
4527                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4528                         tcp_ofo_queue(sk);
4529 
4530                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4531                          * gap in queue is filled.
4532                          */
4533                         if (skb_queue_empty(&tp->out_of_order_queue))
4534                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4535                 }
4536 
4537                 if (tp->rx_opt.num_sacks)
4538                         tcp_sack_remove(tp);
4539 
4540                 tcp_fast_path_check(sk);
4541 
4542                 if (eaten > 0)
4543                         kfree_skb_partial(skb, fragstolen);
4544                 if (!sock_flag(sk, SOCK_DEAD))
4545                         sk->sk_data_ready(sk);
4546                 return;
4547         }
4548 
4549         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4550                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4551                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4552                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4553 
4554 out_of_window:
4555                 tcp_enter_quickack_mode(sk);
4556                 inet_csk_schedule_ack(sk);
4557 drop:
4558                 __kfree_skb(skb);
4559                 return;
4560         }
4561 
4562         /* Out of window. F.e. zero window probe. */
4563         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4564                 goto out_of_window;
4565 
4566         tcp_enter_quickack_mode(sk);
4567 
4568         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4569                 /* Partial packet, seq < rcv_next < end_seq */
4570                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4571                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4572                            TCP_SKB_CB(skb)->end_seq);
4573 
4574                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4575 
4576                 /* If window is closed, drop tail of packet. But after
4577                  * remembering D-SACK for its head made in previous line.
4578                  */
4579                 if (!tcp_receive_window(tp))
4580                         goto out_of_window;
4581                 goto queue_and_out;
4582         }
4583 
4584         tcp_data_queue_ofo(sk, skb);
4585 }
4586 
4587 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4588                                         struct sk_buff_head *list)
4589 {
4590         struct sk_buff *next = NULL;
4591 
4592         if (!skb_queue_is_last(list, skb))
4593                 next = skb_queue_next(list, skb);
4594 
4595         __skb_unlink(skb, list);
4596         __kfree_skb(skb);
4597         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4598 
4599         return next;
4600 }
4601 
4602 /* Collapse contiguous sequence of skbs head..tail with
4603  * sequence numbers start..end.
4604  *
4605  * If tail is NULL, this means until the end of the list.
4606  *
4607  * Segments with FIN/SYN are not collapsed (only because this
4608  * simplifies code)
4609  */
4610 static void
4611 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4612              struct sk_buff *head, struct sk_buff *tail,
4613              u32 start, u32 end)
4614 {
4615         struct sk_buff *skb, *n;
4616         bool end_of_skbs;
4617 
4618         /* First, check that queue is collapsible and find
4619          * the point where collapsing can be useful. */
4620         skb = head;
4621 restart:
4622         end_of_skbs = true;
4623         skb_queue_walk_from_safe(list, skb, n) {
4624                 if (skb == tail)
4625                         break;
4626                 /* No new bits? It is possible on ofo queue. */
4627                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4628                         skb = tcp_collapse_one(sk, skb, list);
4629                         if (!skb)
4630                                 break;
4631                         goto restart;
4632                 }
4633 
4634                 /* The first skb to collapse is:
4635                  * - not SYN/FIN and
4636                  * - bloated or contains data before "start" or
4637                  *   overlaps to the next one.
4638                  */
4639                 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4640                     (tcp_win_from_space(skb->truesize) > skb->len ||
4641                      before(TCP_SKB_CB(skb)->seq, start))) {
4642                         end_of_skbs = false;
4643                         break;
4644                 }
4645 
4646                 if (!skb_queue_is_last(list, skb)) {
4647                         struct sk_buff *next = skb_queue_next(list, skb);
4648                         if (next != tail &&
4649                             TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4650                                 end_of_skbs = false;
4651                                 break;
4652                         }
4653                 }
4654 
4655                 /* Decided to skip this, advance start seq. */
4656                 start = TCP_SKB_CB(skb)->end_seq;
4657         }
4658         if (end_of_skbs ||
4659             (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4660                 return;
4661 
4662         while (before(start, end)) {
4663                 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4664                 struct sk_buff *nskb;
4665 
4666                 nskb = alloc_skb(copy, GFP_ATOMIC);
4667                 if (!nskb)
4668                         return;
4669 
4670                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4671                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4672                 __skb_queue_before(list, skb, nskb);
4673                 skb_set_owner_r(nskb, sk);
4674 
4675                 /* Copy data, releasing collapsed skbs. */
4676                 while (copy > 0) {
4677                         int offset = start - TCP_SKB_CB(skb)->seq;
4678                         int size = TCP_SKB_CB(skb)->end_seq - start;
4679 
4680                         BUG_ON(offset < 0);
4681                         if (size > 0) {
4682                                 size = min(copy, size);
4683                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4684                                         BUG();
4685                                 TCP_SKB_CB(nskb)->end_seq += size;
4686                                 copy -= size;
4687                                 start += size;
4688                         }
4689                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4690                                 skb = tcp_collapse_one(sk, skb, list);
4691                                 if (!skb ||
4692                                     skb == tail ||
4693                                     (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4694                                         return;
4695                         }
4696                 }
4697         }
4698 }
4699 
4700 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4701  * and tcp_collapse() them until all the queue is collapsed.
4702  */
4703 static void tcp_collapse_ofo_queue(struct sock *sk)
4704 {
4705         struct tcp_sock *tp = tcp_sk(sk);
4706         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4707         struct sk_buff *head;
4708         u32 start, end;
4709 
4710         if (!skb)
4711                 return;
4712 
4713         start = TCP_SKB_CB(skb)->seq;
4714         end = TCP_SKB_CB(skb)->end_seq;
4715         head = skb;
4716 
4717         for (;;) {
4718                 struct sk_buff *next = NULL;
4719 
4720                 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4721                         next = skb_queue_next(&tp->out_of_order_queue, skb);
4722                 skb = next;
4723 
4724                 /* Segment is terminated when we see gap or when
4725                  * we are at the end of all the queue. */
4726                 if (!skb ||
4727                     after(TCP_SKB_CB(skb)->seq, end) ||
4728                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4729                         tcp_collapse(sk, &tp->out_of_order_queue,
4730                                      head, skb, start, end);
4731                         head = skb;
4732                         if (!skb)
4733                                 break;
4734                         /* Start new segment */
4735                         start = TCP_SKB_CB(skb)->seq;
4736                         end = TCP_SKB_CB(skb)->end_seq;
4737                 } else {
4738                         if (before(TCP_SKB_CB(skb)->seq, start))
4739                                 start = TCP_SKB_CB(skb)->seq;
4740                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4741                                 end = TCP_SKB_CB(skb)->end_seq;
4742                 }
4743         }
4744 }
4745 
4746 /*
4747  * Purge the out-of-order queue.
4748  * Return true if queue was pruned.
4749  */
4750 static bool tcp_prune_ofo_queue(struct sock *sk)
4751 {
4752         struct tcp_sock *tp = tcp_sk(sk);
4753         bool res = false;
4754 
4755         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4756                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4757                 __skb_queue_purge(&tp->out_of_order_queue);
4758 
4759                 /* Reset SACK state.  A conforming SACK implementation will
4760                  * do the same at a timeout based retransmit.  When a connection
4761                  * is in a sad state like this, we care only about integrity
4762                  * of the connection not performance.
4763                  */
4764                 if (tp->rx_opt.sack_ok)
4765                         tcp_sack_reset(&tp->rx_opt);
4766                 sk_mem_reclaim(sk);
4767                 res = true;
4768         }
4769         return res;
4770 }
4771 
4772 /* Reduce allocated memory if we can, trying to get
4773  * the socket within its memory limits again.
4774  *
4775  * Return less than zero if we should start dropping frames
4776  * until the socket owning process reads some of the data
4777  * to stabilize the situation.
4778  */
4779 static int tcp_prune_queue(struct sock *sk)
4780 {
4781         struct tcp_sock *tp = tcp_sk(sk);
4782 
4783         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4784 
4785         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4786 
4787         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4788                 tcp_clamp_window(sk);
4789         else if (tcp_under_memory_pressure(sk))
4790                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4791 
4792         tcp_collapse_ofo_queue(sk);
4793         if (!skb_queue_empty(&sk->sk_receive_queue))
4794                 tcp_collapse(sk, &sk->sk_receive_queue,
4795                              skb_peek(&sk->sk_receive_queue),
4796                              NULL,
4797                              tp->copied_seq, tp->rcv_nxt);
4798         sk_mem_reclaim(sk);
4799 
4800         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4801                 return 0;
4802 
4803         /* Collapsing did not help, destructive actions follow.
4804          * This must not ever occur. */
4805 
4806         tcp_prune_ofo_queue(sk);
4807 
4808         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4809                 return 0;
4810 
4811         /* If we are really being abused, tell the caller to silently
4812          * drop receive data on the floor.  It will get retransmitted
4813          * and hopefully then we'll have sufficient space.
4814          */
4815         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4816 
4817         /* Massive buffer overcommit. */
4818         tp->pred_flags = 0;
4819         return -1;
4820 }
4821 
4822 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4823 {
4824         const struct tcp_sock *tp = tcp_sk(sk);
4825 
4826         /* If the user specified a specific send buffer setting, do
4827          * not modify it.
4828          */
4829         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4830                 return false;
4831 
4832         /* If we are under global TCP memory pressure, do not expand.  */
4833         if (tcp_under_memory_pressure(sk))
4834                 return false;
4835 
4836         /* If we are under soft global TCP memory pressure, do not expand.  */
4837         if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4838                 return false;
4839 
4840         /* If we filled the congestion window, do not expand.  */
4841         if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4842                 return false;
4843 
4844         return true;
4845 }
4846 
4847 /* When incoming ACK allowed to free some skb from write_queue,
4848  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4849  * on the exit from tcp input handler.
4850  *
4851  * PROBLEM: sndbuf expansion does not work well with largesend.
4852  */
4853 static void tcp_new_space(struct sock *sk)
4854 {
4855         struct tcp_sock *tp = tcp_sk(sk);
4856 
4857         if (tcp_should_expand_sndbuf(sk)) {
4858                 tcp_sndbuf_expand(sk);
4859                 tp->snd_cwnd_stamp = tcp_time_stamp;
4860         }
4861 
4862         sk->sk_write_space(sk);
4863 }
4864 
4865 static void tcp_check_space(struct sock *sk)
4866 {
4867         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4868                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4869                 /* pairs with tcp_poll() */
4870                 smp_mb__after_atomic();
4871                 if (sk->sk_socket &&
4872                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4873                         tcp_new_space(sk);
4874         }
4875 }
4876 
4877 static inline void tcp_data_snd_check(struct sock *sk)
4878 {
4879         tcp_push_pending_frames(sk);
4880         tcp_check_space(sk);
4881 }
4882 
4883 /*
4884  * Check if sending an ack is needed.
4885  */
4886 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4887 {
4888         struct tcp_sock *tp = tcp_sk(sk);
4889 
4890             /* More than one full frame received... */
4891         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4892              /* ... and right edge of window advances far enough.
4893               * (tcp_recvmsg() will send ACK otherwise). Or...
4894               */
4895              __tcp_select_window(sk) >= tp->rcv_wnd) ||
4896             /* We ACK each frame or... */
4897             tcp_in_quickack_mode(sk) ||
4898             /* We have out of order data. */
4899             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4900                 /* Then ack it now */
4901                 tcp_send_ack(sk);
4902         } else {
4903                 /* Else, send delayed ack. */
4904                 tcp_send_delayed_ack(sk);
4905         }
4906 }
4907 
4908 static inline void tcp_ack_snd_check(struct sock *sk)
4909 {
4910         if (!inet_csk_ack_scheduled(sk)) {
4911                 /* We sent a data segment already. */
4912                 return;
4913         }
4914         __tcp_ack_snd_check(sk, 1);
4915 }
4916 
4917 /*
4918  *      This routine is only called when we have urgent data
4919  *      signaled. Its the 'slow' part of tcp_urg. It could be
4920  *      moved inline now as tcp_urg is only called from one
4921  *      place. We handle URGent data wrong. We have to - as
4922  *      BSD still doesn't use the correction from RFC961.
4923  *      For 1003.1g we should support a new option TCP_STDURG to permit
4924  *      either form (or just set the sysctl tcp_stdurg).
4925  */
4926 
4927 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4928 {
4929         struct tcp_sock *tp = tcp_sk(sk);
4930         u32 ptr = ntohs(th->urg_ptr);
4931 
4932         if (ptr && !sysctl_tcp_stdurg)
4933                 ptr--;
4934         ptr += ntohl(th->seq);
4935 
4936         /* Ignore urgent data that we've already seen and read. */
4937         if (after(tp->copied_seq, ptr))
4938                 return;
4939 
4940         /* Do not replay urg ptr.
4941          *
4942          * NOTE: interesting situation not covered by specs.
4943          * Misbehaving sender may send urg ptr, pointing to segment,
4944          * which we already have in ofo queue. We are not able to fetch
4945          * such data and will stay in TCP_URG_NOTYET until will be eaten
4946          * by recvmsg(). Seems, we are not obliged to handle such wicked
4947          * situations. But it is worth to think about possibility of some
4948          * DoSes using some hypothetical application level deadlock.
4949          */
4950         if (before(ptr, tp->rcv_nxt))
4951                 return;
4952 
4953         /* Do we already have a newer (or duplicate) urgent pointer? */
4954         if (tp->urg_data && !after(ptr, tp->urg_seq))
4955                 return;
4956 
4957         /* Tell the world about our new urgent pointer. */
4958         sk_send_sigurg(sk);
4959 
4960         /* We may be adding urgent data when the last byte read was
4961          * urgent. To do this requires some care. We cannot just ignore
4962          * tp->copied_seq since we would read the last urgent byte again
4963          * as data, nor can we alter copied_seq until this data arrives
4964          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4965          *
4966          * NOTE. Double Dutch. Rendering to plain English: author of comment
4967          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4968          * and expect that both A and B disappear from stream. This is _wrong_.
4969          * Though this happens in BSD with high probability, this is occasional.
4970          * Any application relying on this is buggy. Note also, that fix "works"
4971          * only in this artificial test. Insert some normal data between A and B and we will
4972          * decline of BSD again. Verdict: it is better to remove to trap
4973          * buggy users.
4974          */
4975         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4976             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4977                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4978                 tp->copied_seq++;
4979                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4980                         __skb_unlink(skb, &sk->sk_receive_queue);
4981                         __kfree_skb(skb);
4982                 }
4983         }
4984 
4985         tp->urg_data = TCP_URG_NOTYET;
4986         tp->urg_seq = ptr;
4987 
4988         /* Disable header prediction. */
4989         tp->pred_flags = 0;
4990 }
4991 
4992 /* This is the 'fast' part of urgent handling. */
4993 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4994 {
4995         struct tcp_sock *tp = tcp_sk(sk);
4996 
4997         /* Check if we get a new urgent pointer - normally not. */
4998         if (th->urg)
4999                 tcp_check_urg(sk, th);
5000 
5001         /* Do we wait for any urgent data? - normally not... */
5002         if (tp->urg_data == TCP_URG_NOTYET) {
5003                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5004                           th->syn;
5005 
5006                 /* Is the urgent pointer pointing into this packet? */
5007                 if (ptr < skb->len) {
5008                         u8 tmp;
5009                         if (skb_copy_bits(skb, ptr, &tmp, 1))
5010                                 BUG();
5011                         tp->urg_data = TCP_URG_VALID | tmp;
5012                         if (!sock_flag(sk, SOCK_DEAD))
5013                                 sk->sk_data_ready(sk);
5014                 }
5015         }
5016 }
5017 
5018 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5019 {
5020         struct tcp_sock *tp = tcp_sk(sk);
5021         int chunk = skb->len - hlen;
5022         int err;
5023 
5024         local_bh_enable();
5025         if (skb_csum_unnecessary(skb))
5026                 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5027         else
5028                 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5029 
5030         if (!err) {
5031                 tp->ucopy.len -= chunk;
5032                 tp->copied_seq += chunk;
5033                 tcp_rcv_space_adjust(sk);
5034         }
5035 
5036         local_bh_disable();
5037         return err;
5038 }
5039 
5040 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5041                                             struct sk_buff *skb)
5042 {
5043         __sum16 result;
5044 
5045         if (sock_owned_by_user(sk)) {
5046                 local_bh_enable();
5047                 result = __tcp_checksum_complete(skb);
5048                 local_bh_disable();
5049         } else {
5050                 result = __tcp_checksum_complete(skb);
5051         }
5052         return result;
5053 }
5054 
5055 static inline bool tcp_checksum_complete_user(struct sock *sk,
5056                                              struct sk_buff *skb)
5057 {
5058         return !skb_csum_unnecessary(skb) &&
5059                __tcp_checksum_complete_user(sk, skb);
5060 }
5061 
5062 /* Does PAWS and seqno based validation of an incoming segment, flags will
5063  * play significant role here.
5064  */
5065 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5066                                   const struct tcphdr *th, int syn_inerr)
5067 {
5068         struct tcp_sock *tp = tcp_sk(sk);
5069 
5070         /* RFC1323: H1. Apply PAWS check first. */
5071         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5072             tcp_paws_discard(sk, skb)) {
5073                 if (!th->rst) {
5074                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5075                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5076                                                   LINUX_MIB_TCPACKSKIPPEDPAWS,
5077                                                   &tp->last_oow_ack_time))
5078                                 tcp_send_dupack(sk, skb);
5079                         goto discard;
5080                 }
5081                 /* Reset is accepted even if it did not pass PAWS. */
5082         }
5083 
5084         /* Step 1: check sequence number */
5085         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5086                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5087                  * (RST) segments are validated by checking their SEQ-fields."
5088                  * And page 69: "If an incoming segment is not acceptable,
5089                  * an acknowledgment should be sent in reply (unless the RST
5090                  * bit is set, if so drop the segment and return)".
5091                  */
5092                 if (!th->rst) {
5093                         if (th->syn)
5094                                 goto syn_challenge;
5095                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5096                                                   LINUX_MIB_TCPACKSKIPPEDSEQ,
5097                                                   &tp->last_oow_ack_time))
5098                                 tcp_send_dupack(sk, skb);
5099                 }
5100                 goto discard;
5101         }
5102 
5103         /* Step 2: check RST bit */
5104         if (th->rst) {
5105                 /* RFC 5961 3.2 :
5106                  * If sequence number exactly matches RCV.NXT, then
5107                  *     RESET the connection
5108                  * else
5109                  *     Send a challenge ACK
5110                  */
5111                 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5112                         tcp_reset(sk);
5113                 else
5114                         tcp_send_challenge_ack(sk, skb);
5115                 goto discard;
5116         }
5117 
5118         /* step 3: check security and precedence [ignored] */
5119 
5120         /* step 4: Check for a SYN
5121          * RFC 5961 4.2 : Send a challenge ack
5122          */
5123         if (th->syn) {
5124 syn_challenge:
5125                 if (syn_inerr)
5126                         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5127                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5128                 tcp_send_challenge_ack(sk, skb);
5129                 goto discard;
5130         }
5131 
5132         return true;
5133 
5134 discard:
5135         __kfree_skb(skb);
5136         return false;
5137 }
5138 
5139 /*
5140  *      TCP receive function for the ESTABLISHED state.
5141  *
5142  *      It is split into a fast path and a slow path. The fast path is
5143  *      disabled when:
5144  *      - A zero window was announced from us - zero window probing
5145  *        is only handled properly in the slow path.
5146  *      - Out of order segments arrived.
5147  *      - Urgent data is expected.
5148  *      - There is no buffer space left
5149  *      - Unexpected TCP flags/window values/header lengths are received
5150  *        (detected by checking the TCP header against pred_flags)
5151  *      - Data is sent in both directions. Fast path only supports pure senders
5152  *        or pure receivers (this means either the sequence number or the ack
5153  *        value must stay constant)
5154  *      - Unexpected TCP option.
5155  *
5156  *      When these conditions are not satisfied it drops into a standard
5157  *      receive procedure patterned after RFC793 to handle all cases.
5158  *      The first three cases are guaranteed by proper pred_flags setting,
5159  *      the rest is checked inline. Fast processing is turned on in
5160  *      tcp_data_queue when everything is OK.
5161  */
5162 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5163                          const struct tcphdr *th, unsigned int len)
5164 {
5165         struct tcp_sock *tp = tcp_sk(sk);
5166 
5167         if (unlikely(!sk->sk_rx_dst))
5168                 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5169         /*
5170          *      Header prediction.
5171          *      The code loosely follows the one in the famous
5172          *      "30 instruction TCP receive" Van Jacobson mail.
5173          *
5174          *      Van's trick is to deposit buffers into socket queue
5175          *      on a device interrupt, to call tcp_recv function
5176          *      on the receive process context and checksum and copy
5177          *      the buffer to user space. smart...
5178          *
5179          *      Our current scheme is not silly either but we take the
5180          *      extra cost of the net_bh soft interrupt processing...
5181          *      We do checksum and copy also but from device to kernel.
5182          */
5183 
5184         tp->rx_opt.saw_tstamp = 0;
5185 
5186         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5187          *      if header_prediction is to be made
5188          *      'S' will always be tp->tcp_header_len >> 2
5189          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5190          *  turn it off (when there are holes in the receive
5191          *       space for instance)
5192          *      PSH flag is ignored.
5193          */
5194 
5195         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5196             TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5197             !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5198                 int tcp_header_len = tp->tcp_header_len;
5199 
5200                 /* Timestamp header prediction: tcp_header_len
5201                  * is automatically equal to th->doff*4 due to pred_flags
5202                  * match.
5203                  */
5204 
5205                 /* Check timestamp */
5206                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5207                         /* No? Slow path! */
5208                         if (!tcp_parse_aligned_timestamp(tp, th))
5209                                 goto slow_path;
5210 
5211                         /* If PAWS failed, check it more carefully in slow path */
5212                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5213                                 goto slow_path;
5214 
5215                         /* DO NOT update ts_recent here, if checksum fails
5216                          * and timestamp was corrupted part, it will result
5217                          * in a hung connection since we will drop all
5218                          * future packets due to the PAWS test.
5219                          */
5220                 }
5221 
5222                 if (len <= tcp_header_len) {
5223                         /* Bulk data transfer: sender */
5224                         if (len == tcp_header_len) {
5225                                 /* Predicted packet is in window by definition.
5226                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5227                                  * Hence, check seq<=rcv_wup reduces to:
5228                                  */
5229                                 if (tcp_header_len ==
5230                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5231                                     tp->rcv_nxt == tp->rcv_wup)
5232                                         tcp_store_ts_recent(tp);
5233 
5234                                 /* We know that such packets are checksummed
5235                                  * on entry.
5236                                  */
5237                                 tcp_ack(sk, skb, 0);
5238                                 __kfree_skb(skb);
5239                                 tcp_data_snd_check(sk);
5240                                 return;
5241                         } else { /* Header too small */
5242                                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5243                                 goto discard;
5244                         }
5245                 } else {
5246                         int eaten = 0;
5247                         bool fragstolen = false;
5248 
5249                         if (tp->ucopy.task == current &&
5250                             tp->copied_seq == tp->rcv_nxt &&
5251                             len - tcp_header_len <= tp->ucopy.len &&
5252                             sock_owned_by_user(sk)) {
5253                                 __set_current_state(TASK_RUNNING);
5254 
5255                                 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5256                                         /* Predicted packet is in window by definition.
5257                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5258                                          * Hence, check seq<=rcv_wup reduces to:
5259                                          */
5260                                         if (tcp_header_len ==
5261                                             (sizeof(struct tcphdr) +
5262                                              TCPOLEN_TSTAMP_ALIGNED) &&
5263                                             tp->rcv_nxt == tp->rcv_wup)
5264                                                 tcp_store_ts_recent(tp);
5265 
5266                                         tcp_rcv_rtt_measure_ts(sk, skb);
5267 
5268                                         __skb_pull(skb, tcp_header_len);
5269                                         tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5270                                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5271                                         eaten = 1;
5272                                 }
5273                         }
5274                         if (!eaten) {
5275                                 if (tcp_checksum_complete_user(sk, skb))
5276                                         goto csum_error;
5277 
5278                                 if ((int)skb->truesize > sk->sk_forward_alloc)
5279                                         goto step5;
5280 
5281                                 /* Predicted packet is in window by definition.
5282                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5283                                  * Hence, check seq<=rcv_wup reduces to:
5284                                  */
5285                                 if (tcp_header_len ==
5286                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5287                                     tp->rcv_nxt == tp->rcv_wup)
5288                                         tcp_store_ts_recent(tp);
5289 
5290                                 tcp_rcv_rtt_measure_ts(sk, skb);
5291 
5292                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5293 
5294                                 /* Bulk data transfer: receiver */
5295                                 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5296                                                       &fragstolen);
5297                         }
5298 
5299                         tcp_event_data_recv(sk, skb);
5300 
5301                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5302                                 /* Well, only one small jumplet in fast path... */
5303                                 tcp_ack(sk, skb, FLAG_DATA);
5304                                 tcp_data_snd_check(sk);
5305                                 if (!inet_csk_ack_scheduled(sk))
5306                                         goto no_ack;
5307                         }
5308 
5309                         __tcp_ack_snd_check(sk, 0);
5310 no_ack:
5311                         if (eaten)
5312                                 kfree_skb_partial(skb, fragstolen);
5313                         sk->sk_data_ready(sk);
5314                         return;
5315                 }
5316         }
5317 
5318 slow_path:
5319         if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5320                 goto csum_error;
5321 
5322         if (!th->ack && !th->rst && !th->syn)
5323                 goto discard;
5324 
5325         /*
5326          *      Standard slow path.
5327          */
5328 
5329         if (!tcp_validate_incoming(sk, skb, th, 1))
5330                 return;
5331 
5332 step5:
5333         if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5334                 goto discard;
5335 
5336         tcp_rcv_rtt_measure_ts(sk, skb);
5337 
5338         /* Process urgent data. */
5339         tcp_urg(sk, skb, th);
5340 
5341         /* step 7: process the segment text */
5342         tcp_data_queue(sk, skb);
5343 
5344         tcp_data_snd_check(sk);
5345         tcp_ack_snd_check(sk);
5346         return;
5347 
5348 csum_error:
5349         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5350         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5351 
5352 discard:
5353         __kfree_skb(skb);
5354 }
5355 EXPORT_SYMBOL(tcp_rcv_established);
5356 
5357 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5358 {
5359         struct tcp_sock *tp = tcp_sk(sk);
5360         struct inet_connection_sock *icsk = inet_csk(sk);
5361 
5362         tcp_set_state(sk, TCP_ESTABLISHED);
5363 
5364         if (skb) {
5365                 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5366                 security_inet_conn_established(sk, skb);
5367         }
5368 
5369         /* Make sure socket is routed, for correct metrics.  */
5370         icsk->icsk_af_ops->rebuild_header(sk);
5371 
5372         tcp_init_metrics(sk);
5373 
5374         tcp_init_congestion_control(sk);
5375 
5376         /* Prevent spurious tcp_cwnd_restart() on first data
5377          * packet.
5378          */
5379         tp->lsndtime = tcp_time_stamp;
5380 
5381         tcp_init_buffer_space(sk);
5382 
5383         if (sock_flag(sk, SOCK_KEEPOPEN))
5384                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5385 
5386         if (!tp->rx_opt.snd_wscale)
5387                 __tcp_fast_path_on(tp, tp->snd_wnd);
5388         else
5389                 tp->pred_flags = 0;
5390 
5391         if (!sock_flag(sk, SOCK_DEAD)) {
5392                 sk->sk_state_change(sk);
5393                 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5394         }
5395 }
5396 
5397 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5398                                     struct tcp_fastopen_cookie *cookie)
5399 {
5400         struct tcp_sock *tp = tcp_sk(sk);
5401         struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5402         u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5403         bool syn_drop = false;
5404 
5405         if (mss == tp->rx_opt.user_mss) {
5406                 struct tcp_options_received opt;
5407 
5408                 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5409                 tcp_clear_options(&opt);
5410                 opt.user_mss = opt.mss_clamp = 0;
5411                 tcp_parse_options(synack, &opt, 0, NULL);
5412                 mss = opt.mss_clamp;
5413         }
5414 
5415         if (!tp->syn_fastopen) {
5416                 /* Ignore an unsolicited cookie */
5417                 cookie->len = -1;
5418         } else if (tp->total_retrans) {
5419                 /* SYN timed out and the SYN-ACK neither has a cookie nor
5420                  * acknowledges data. Presumably the remote received only
5421                  * the retransmitted (regular) SYNs: either the original
5422                  * SYN-data or the corresponding SYN-ACK was dropped.
5423                  */
5424                 syn_drop = (cookie->len < 0 && data);
5425         } else if (cookie->len < 0 && !tp->syn_data) {
5426                 /* We requested a cookie but didn't get it. If we did not use
5427                  * the (old) exp opt format then try so next time (try_exp=1).
5428                  * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5429                  */
5430                 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5431         }
5432 
5433         tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5434 
5435         if (data) { /* Retransmit unacked data in SYN */
5436                 tcp_for_write_queue_from(data, sk) {
5437                         if (data == tcp_send_head(sk) ||
5438                             __tcp_retransmit_skb(sk, data))
5439                                 break;
5440                 }
5441                 tcp_rearm_rto(sk);
5442                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5443                 return true;
5444         }
5445         tp->syn_data_acked = tp->syn_data;
5446         if (tp->syn_data_acked)
5447                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5448         return false;
5449 }
5450 
5451 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5452                                          const struct tcphdr *th, unsigned int len)
5453 {
5454         struct inet_connection_sock *icsk = inet_csk(sk);
5455         struct tcp_sock *tp = tcp_sk(sk);
5456         struct tcp_fastopen_cookie foc = { .len = -1 };
5457         int saved_clamp = tp->rx_opt.mss_clamp;
5458 
5459         tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5460         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5461                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5462 
5463         if (th->ack) {
5464                 /* rfc793:
5465                  * "If the state is SYN-SENT then
5466                  *    first check the ACK bit
5467                  *      If the ACK bit is set
5468                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5469                  *        a reset (unless the RST bit is set, if so drop
5470                  *        the segment and return)"
5471                  */
5472                 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5473                     after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5474                         goto reset_and_undo;
5475 
5476                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5477                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5478                              tcp_time_stamp)) {
5479                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5480                         goto reset_and_undo;
5481                 }
5482 
5483                 /* Now ACK is acceptable.
5484                  *
5485                  * "If the RST bit is set
5486                  *    If the ACK was acceptable then signal the user "error:
5487                  *    connection reset", drop the segment, enter CLOSED state,
5488                  *    delete TCB, and return."
5489                  */
5490 
5491                 if (th->rst) {
5492                         tcp_reset(sk);
5493                         goto discard;
5494                 }
5495 
5496                 /* rfc793:
5497                  *   "fifth, if neither of the SYN or RST bits is set then
5498                  *    drop the segment and return."
5499                  *
5500                  *    See note below!
5501                  *                                        --ANK(990513)
5502                  */
5503                 if (!th->syn)
5504                         goto discard_and_undo;
5505 
5506                 /* rfc793:
5507                  *   "If the SYN bit is on ...
5508                  *    are acceptable then ...
5509                  *    (our SYN has been ACKed), change the connection
5510                  *    state to ESTABLISHED..."
5511                  */
5512 
5513                 tcp_ecn_rcv_synack(tp, th);
5514 
5515                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5516                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5517 
5518                 /* Ok.. it's good. Set up sequence numbers and
5519                  * move to established.
5520                  */
5521                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5522                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5523 
5524                 /* RFC1323: The window in SYN & SYN/ACK segments is
5525                  * never scaled.
5526                  */
5527                 tp->snd_wnd = ntohs(th->window);
5528 
5529                 if (!tp->rx_opt.wscale_ok) {
5530                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5531                         tp->window_clamp = min(tp->window_clamp, 65535U);
5532                 }
5533 
5534                 if (tp->rx_opt.saw_tstamp) {
5535                         tp->rx_opt.tstamp_ok       = 1;
5536                         tp->tcp_header_len =
5537                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5538                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5539                         tcp_store_ts_recent(tp);
5540                 } else {
5541                         tp->tcp_header_len = sizeof(struct tcphdr);
5542                 }
5543 
5544                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5545                         tcp_enable_fack(tp);
5546 
5547                 tcp_mtup_init(sk);
5548                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5549                 tcp_initialize_rcv_mss(sk);
5550 
5551                 /* Remember, tcp_poll() does not lock socket!
5552                  * Change state from SYN-SENT only after copied_seq
5553                  * is initialized. */
5554                 tp->copied_seq = tp->rcv_nxt;
5555 
5556                 smp_mb();
5557 
5558                 tcp_finish_connect(sk, skb);
5559 
5560                 if ((tp->syn_fastopen || tp->syn_data) &&
5561                     tcp_rcv_fastopen_synack(sk, skb, &foc))
5562                         return -1;
5563 
5564                 if (sk->sk_write_pending ||
5565                     icsk->icsk_accept_queue.rskq_defer_accept ||
5566                     icsk->icsk_ack.pingpong) {
5567                         /* Save one ACK. Data will be ready after
5568                          * several ticks, if write_pending is set.
5569                          *
5570                          * It may be deleted, but with this feature tcpdumps
5571                          * look so _wonderfully_ clever, that I was not able
5572                          * to stand against the temptation 8)     --ANK
5573                          */
5574                         inet_csk_schedule_ack(sk);
5575                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5576                         tcp_enter_quickack_mode(sk);
5577                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5578                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5579 
5580 discard:
5581                         __kfree_skb(skb);
5582                         return 0;
5583                 } else {
5584                         tcp_send_ack(sk);
5585                 }
5586                 return -1;
5587         }
5588 
5589         /* No ACK in the segment */
5590 
5591         if (th->rst) {
5592                 /* rfc793:
5593                  * "If the RST bit is set
5594                  *
5595                  *      Otherwise (no ACK) drop the segment and return."
5596                  */
5597 
5598                 goto discard_and_undo;
5599         }
5600 
5601         /* PAWS check. */
5602         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5603             tcp_paws_reject(&tp->rx_opt, 0))
5604                 goto discard_and_undo;
5605 
5606         if (th->syn) {
5607                 /* We see SYN without ACK. It is attempt of
5608                  * simultaneous connect with crossed SYNs.
5609                  * Particularly, it can be connect to self.
5610                  */
5611                 tcp_set_state(sk, TCP_SYN_RECV);
5612 
5613                 if (tp->rx_opt.saw_tstamp) {
5614                         tp->rx_opt.tstamp_ok = 1;
5615                         tcp_store_ts_recent(tp);
5616                         tp->tcp_header_len =
5617                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5618                 } else {
5619                         tp->tcp_header_len = sizeof(struct tcphdr);
5620                 }
5621 
5622                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5623                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5624 
5625                 /* RFC1323: The window in SYN & SYN/ACK segments is
5626                  * never scaled.
5627                  */
5628                 tp->snd_wnd    = ntohs(th->window);
5629                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5630                 tp->max_window = tp->snd_wnd;
5631 
5632                 tcp_ecn_rcv_syn(tp, th);
5633 
5634                 tcp_mtup_init(sk);
5635                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5636                 tcp_initialize_rcv_mss(sk);
5637 
5638                 tcp_send_synack(sk);
5639 #if 0
5640                 /* Note, we could accept data and URG from this segment.
5641                  * There are no obstacles to make this (except that we must
5642                  * either change tcp_recvmsg() to prevent it from returning data
5643                  * before 3WHS completes per RFC793, or employ TCP Fast Open).
5644                  *
5645                  * However, if we ignore data in ACKless segments sometimes,
5646                  * we have no reasons to accept it sometimes.
5647                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5648                  * is not flawless. So, discard packet for sanity.
5649                  * Uncomment this return to process the data.
5650                  */
5651                 return -1;
5652 #else
5653                 goto discard;
5654 #endif
5655         }
5656         /* "fifth, if neither of the SYN or RST bits is set then
5657          * drop the segment and return."
5658          */
5659 
5660 discard_and_undo:
5661         tcp_clear_options(&tp->rx_opt);
5662         tp->rx_opt.mss_clamp = saved_clamp;
5663         goto discard;
5664 
5665 reset_and_undo:
5666         tcp_clear_options(&tp->rx_opt);
5667         tp->rx_opt.mss_clamp = saved_clamp;
5668         return 1;
5669 }
5670 
5671 /*
5672  *      This function implements the receiving procedure of RFC 793 for
5673  *      all states except ESTABLISHED and TIME_WAIT.
5674  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5675  *      address independent.
5676  */
5677 
5678 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5679                           const struct tcphdr *th, unsigned int len)
5680 {
5681         struct tcp_sock *tp = tcp_sk(sk);
5682         struct inet_connection_sock *icsk = inet_csk(sk);
5683         struct request_sock *req;
5684         int queued = 0;
5685         bool acceptable;
5686         u32 synack_stamp;
5687 
5688         tp->rx_opt.saw_tstamp = 0;
5689 
5690         switch (sk->sk_state) {
5691         case TCP_CLOSE:
5692                 goto discard;
5693 
5694         case TCP_LISTEN:
5695                 if (th->ack)
5696                         return 1;
5697 
5698                 if (th->rst)
5699                         goto discard;
5700 
5701                 if (th->syn) {
5702                         if (th->fin)
5703                                 goto discard;
5704                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5705                                 return 1;
5706 
5707                         /* Now we have several options: In theory there is
5708                          * nothing else in the frame. KA9Q has an option to
5709                          * send data with the syn, BSD accepts data with the
5710                          * syn up to the [to be] advertised window and
5711                          * Solaris 2.1 gives you a protocol error. For now
5712                          * we just ignore it, that fits the spec precisely
5713                          * and avoids incompatibilities. It would be nice in
5714                          * future to drop through and process the data.
5715                          *
5716                          * Now that TTCP is starting to be used we ought to
5717                          * queue this data.
5718                          * But, this leaves one open to an easy denial of
5719                          * service attack, and SYN cookies can't defend
5720                          * against this problem. So, we drop the data
5721                          * in the interest of security over speed unless
5722                          * it's still in use.
5723                          */
5724                         kfree_skb(skb);
5725                         return 0;
5726                 }
5727                 goto discard;
5728 
5729         case TCP_SYN_SENT:
5730                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5731                 if (queued >= 0)
5732                         return queued;
5733 
5734                 /* Do step6 onward by hand. */
5735                 tcp_urg(sk, skb, th);
5736                 __kfree_skb(skb);
5737                 tcp_data_snd_check(sk);
5738                 return 0;
5739         }
5740 
5741         req = tp->fastopen_rsk;
5742         if (req) {
5743                 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5744                     sk->sk_state != TCP_FIN_WAIT1);
5745 
5746                 if (!tcp_check_req(sk, skb, req, true))
5747                         goto discard;
5748         }
5749 
5750         if (!th->ack && !th->rst && !th->syn)
5751                 goto discard;
5752 
5753         if (!tcp_validate_incoming(sk, skb, th, 0))
5754                 return 0;
5755 
5756         /* step 5: check the ACK field */
5757         acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5758                                       FLAG_UPDATE_TS_RECENT) > 0;
5759 
5760         switch (sk->sk_state) {
5761         case TCP_SYN_RECV:
5762                 if (!acceptable)
5763                         return 1;
5764 
5765                 /* Once we leave TCP_SYN_RECV, we no longer need req
5766                  * so release it.
5767                  */
5768                 if (req) {
5769                         synack_stamp = tcp_rsk(req)->snt_synack;
5770                         tp->total_retrans = req->num_retrans;
5771                         reqsk_fastopen_remove(sk, req, false);
5772                 } else {
5773                         synack_stamp = tp->lsndtime;
5774                         /* Make sure socket is routed, for correct metrics. */
5775                         icsk->icsk_af_ops->rebuild_header(sk);
5776                         tcp_init_congestion_control(sk);
5777 
5778                         tcp_mtup_init(sk);
5779                         tp->copied_seq = tp->rcv_nxt;
5780                         tcp_init_buffer_space(sk);
5781                 }
5782                 smp_mb();
5783                 tcp_set_state(sk, TCP_ESTABLISHED);
5784                 sk->sk_state_change(sk);
5785 
5786                 /* Note, that this wakeup is only for marginal crossed SYN case.
5787                  * Passively open sockets are not waked up, because
5788                  * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5789                  */
5790                 if (sk->sk_socket)
5791                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5792 
5793                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5794                 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5795                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5796                 tcp_synack_rtt_meas(sk, synack_stamp);
5797 
5798                 if (tp->rx_opt.tstamp_ok)
5799                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5800 
5801                 if (req) {
5802                         /* Re-arm the timer because data may have been sent out.
5803                          * This is similar to the regular data transmission case
5804                          * when new data has just been ack'ed.
5805                          *
5806                          * (TFO) - we could try to be more aggressive and
5807                          * retransmitting any data sooner based on when they
5808                          * are sent out.
5809                          */
5810                         tcp_rearm_rto(sk);
5811                 } else
5812                         tcp_init_metrics(sk);
5813 
5814                 tcp_update_pacing_rate(sk);
5815 
5816                 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5817                 tp->lsndtime = tcp_time_stamp;
5818 
5819                 tcp_initialize_rcv_mss(sk);
5820                 tcp_fast_path_on(tp);
5821                 break;
5822 
5823         case TCP_FIN_WAIT1: {
5824                 struct dst_entry *dst;
5825                 int tmo;
5826 
5827                 /* If we enter the TCP_FIN_WAIT1 state and we are a
5828                  * Fast Open socket and this is the first acceptable
5829                  * ACK we have received, this would have acknowledged
5830                  * our SYNACK so stop the SYNACK timer.
5831                  */
5832                 if (req) {
5833                         /* Return RST if ack_seq is invalid.
5834                          * Note that RFC793 only says to generate a
5835                          * DUPACK for it but for TCP Fast Open it seems
5836                          * better to treat this case like TCP_SYN_RECV
5837                          * above.
5838                          */
5839                         if (!acceptable)
5840                                 return 1;
5841                         /* We no longer need the request sock. */
5842                         reqsk_fastopen_remove(sk, req, false);
5843                         tcp_rearm_rto(sk);
5844                 }
5845                 if (tp->snd_una != tp->write_seq)
5846                         break;
5847 
5848                 tcp_set_state(sk, TCP_FIN_WAIT2);
5849                 sk->sk_shutdown |= SEND_SHUTDOWN;
5850 
5851                 dst = __sk_dst_get(sk);
5852                 if (dst)
5853                         dst_confirm(dst);
5854 
5855                 if (!sock_flag(sk, SOCK_DEAD)) {
5856                         /* Wake up lingering close() */
5857                         sk->sk_state_change(sk);
5858                         break;
5859                 }
5860 
5861                 if (tp->linger2 < 0 ||
5862                     (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5863                      after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5864                         tcp_done(sk);
5865                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5866                         return 1;
5867                 }
5868 
5869                 tmo = tcp_fin_time(sk);
5870                 if (tmo > TCP_TIMEWAIT_LEN) {
5871                         inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5872                 } else if (th->fin || sock_owned_by_user(sk)) {
5873                         /* Bad case. We could lose such FIN otherwise.
5874                          * It is not a big problem, but it looks confusing
5875                          * and not so rare event. We still can lose it now,
5876                          * if it spins in bh_lock_sock(), but it is really
5877                          * marginal case.
5878                          */
5879                         inet_csk_reset_keepalive_timer(sk, tmo);
5880                 } else {
5881                         tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5882                         goto discard;
5883                 }
5884                 break;
5885         }
5886 
5887         case TCP_CLOSING:
5888                 if (tp->snd_una == tp->write_seq) {
5889                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5890                         goto discard;
5891                 }
5892                 break;
5893 
5894         case TCP_LAST_ACK:
5895                 if (tp->snd_una == tp->write_seq) {
5896                         tcp_update_metrics(sk);
5897                         tcp_done(sk);
5898                         goto discard;
5899                 }
5900                 break;
5901         }
5902 
5903         /* step 6: check the URG bit */
5904         tcp_urg(sk, skb, th);
5905 
5906         /* step 7: process the segment text */
5907         switch (sk->sk_state) {
5908         case TCP_CLOSE_WAIT:
5909         case TCP_CLOSING:
5910         case TCP_LAST_ACK:
5911                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5912                         break;
5913         case TCP_FIN_WAIT1:
5914         case TCP_FIN_WAIT2:
5915                 /* RFC 793 says to queue data in these states,
5916                  * RFC 1122 says we MUST send a reset.
5917                  * BSD 4.4 also does reset.
5918                  */
5919                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5920                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5921                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5922                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5923                                 tcp_reset(sk);
5924                                 return 1;
5925                         }
5926                 }
5927                 /* Fall through */
5928         case TCP_ESTABLISHED:
5929                 tcp_data_queue(sk, skb);
5930                 queued = 1;
5931                 break;
5932         }
5933 
5934         /* tcp_data could move socket to TIME-WAIT */
5935         if (sk->sk_state != TCP_CLOSE) {
5936                 tcp_data_snd_check(sk);
5937                 tcp_ack_snd_check(sk);
5938         }
5939 
5940         if (!queued) {
5941 discard:
5942                 __kfree_skb(skb);
5943         }
5944         return 0;
5945 }
5946 EXPORT_SYMBOL(tcp_rcv_state_process);
5947 
5948 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
5949 {
5950         struct inet_request_sock *ireq = inet_rsk(req);
5951 
5952         if (family == AF_INET)
5953                 net_dbg_ratelimited("drop open request from %pI4/%u\n",
5954                                     &ireq->ir_rmt_addr, port);
5955 #if IS_ENABLED(CONFIG_IPV6)
5956         else if (family == AF_INET6)
5957                 net_dbg_ratelimited("drop open request from %pI6/%u\n",
5958                                     &ireq->ir_v6_rmt_addr, port);
5959 #endif
5960 }
5961 
5962 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
5963  *
5964  * If we receive a SYN packet with these bits set, it means a
5965  * network is playing bad games with TOS bits. In order to
5966  * avoid possible false congestion notifications, we disable
5967  * TCP ECN negotiation.
5968  *
5969  * Exception: tcp_ca wants ECN. This is required for DCTCP
5970  * congestion control: Linux DCTCP asserts ECT on all packets,
5971  * including SYN, which is most optimal solution; however,
5972  * others, such as FreeBSD do not.
5973  */
5974 static void tcp_ecn_create_request(struct request_sock *req,
5975                                    const struct sk_buff *skb,
5976                                    const struct sock *listen_sk,
5977                                    const struct dst_entry *dst)
5978 {
5979         const struct tcphdr *th = tcp_hdr(skb);
5980         const struct net *net = sock_net(listen_sk);
5981         bool th_ecn = th->ece && th->cwr;
5982         bool ect, ecn_ok;
5983 
5984         if (!th_ecn)
5985                 return;
5986 
5987         ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
5988         ecn_ok = net->ipv4.sysctl_tcp_ecn || dst_feature(dst, RTAX_FEATURE_ECN);
5989 
5990         if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk))
5991                 inet_rsk(req)->ecn_ok = 1;
5992 }
5993 
5994 static void tcp_openreq_init(struct request_sock *req,
5995                              const struct tcp_options_received *rx_opt,
5996                              struct sk_buff *skb, const struct sock *sk)
5997 {
5998         struct inet_request_sock *ireq = inet_rsk(req);
5999 
6000         req->rcv_wnd = 0;               /* So that tcp_send_synack() knows! */
6001         req->cookie_ts = 0;
6002         tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6003         tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6004         tcp_rsk(req)->snt_synack = tcp_time_stamp;
6005         tcp_rsk(req)->last_oow_ack_time = 0;
6006         req->mss = rx_opt->mss_clamp;
6007         req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6008         ireq->tstamp_ok = rx_opt->tstamp_ok;
6009         ireq->sack_ok = rx_opt->sack_ok;
6010         ireq->snd_wscale = rx_opt->snd_wscale;
6011         ireq->wscale_ok = rx_opt->wscale_ok;
6012         ireq->acked = 0;
6013         ireq->ecn_ok = 0;
6014         ireq->ir_rmt_port = tcp_hdr(skb)->source;
6015         ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6016         ireq->ir_mark = inet_request_mark(sk, skb);
6017 }
6018 
6019 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6020                                       struct sock *sk_listener)
6021 {
6022         struct request_sock *req = reqsk_alloc(ops, sk_listener);
6023 
6024         if (req) {
6025                 struct inet_request_sock *ireq = inet_rsk(req);
6026 
6027                 kmemcheck_annotate_bitfield(ireq, flags);
6028                 ireq->opt = NULL;
6029                 atomic64_set(&ireq->ir_cookie, 0);
6030                 ireq->ireq_state = TCP_NEW_SYN_RECV;
6031                 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6032                 ireq->ireq_family = sk_listener->sk_family;
6033         }
6034 
6035         return req;
6036 }
6037 EXPORT_SYMBOL(inet_reqsk_alloc);
6038 
6039 /*
6040  * Return true if a syncookie should be sent
6041  */
6042 static bool tcp_syn_flood_action(struct sock *sk,
6043                                  const struct sk_buff *skb,
6044                                  const char *proto)
6045 {
6046         const char *msg = "Dropping request";
6047         bool want_cookie = false;
6048         struct listen_sock *lopt;
6049 
6050 #ifdef CONFIG_SYN_COOKIES
6051         if (sysctl_tcp_syncookies) {
6052                 msg = "Sending cookies";
6053                 want_cookie = true;
6054                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6055         } else
6056 #endif
6057                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6058 
6059         lopt = inet_csk(sk)->icsk_accept_queue.listen_opt;
6060         if (!lopt->synflood_warned && sysctl_tcp_syncookies != 2) {
6061                 lopt->synflood_warned = 1;
6062                 pr_info("%s: Possible SYN flooding on port %d. %s.  Check SNMP counters.\n",
6063                         proto, ntohs(tcp_hdr(skb)->dest), msg);
6064         }
6065         return want_cookie;
6066 }
6067 
6068 static void tcp_reqsk_record_syn(const struct sock *sk,
6069                                  struct request_sock *req,
6070                                  const struct sk_buff *skb)
6071 {
6072         if (tcp_sk(sk)->save_syn) {
6073                 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6074                 u32 *copy;
6075 
6076                 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6077                 if (copy) {
6078                         copy[0] = len;
6079                         memcpy(&copy[1], skb_network_header(skb), len);
6080                         req->saved_syn = copy;
6081                 }
6082         }
6083 }
6084 
6085 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6086                      const struct tcp_request_sock_ops *af_ops,
6087                      struct sock *sk, struct sk_buff *skb)
6088 {
6089         struct tcp_options_received tmp_opt;
6090         struct request_sock *req;
6091         struct tcp_sock *tp = tcp_sk(sk);
6092         struct dst_entry *dst = NULL;
6093         __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6094         bool want_cookie = false, fastopen;
6095         struct flowi fl;
6096         struct tcp_fastopen_cookie foc = { .len = -1 };
6097         int err;
6098 
6099 
6100         /* TW buckets are converted to open requests without
6101          * limitations, they conserve resources and peer is
6102          * evidently real one.
6103          */
6104         if ((sysctl_tcp_syncookies == 2 ||
6105              inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6106                 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6107                 if (!want_cookie)
6108                         goto drop;
6109         }
6110 
6111 
6112         /* Accept backlog is full. If we have already queued enough
6113          * of warm entries in syn queue, drop request. It is better than
6114          * clogging syn queue with openreqs with exponentially increasing
6115          * timeout.
6116          */
6117         if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
6118                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6119                 goto drop;
6120         }
6121 
6122         req = inet_reqsk_alloc(rsk_ops, sk);
6123         if (!req)
6124                 goto drop;
6125 
6126         tcp_rsk(req)->af_specific = af_ops;
6127 
6128         tcp_clear_options(&tmp_opt);
6129         tmp_opt.mss_clamp = af_ops->mss_clamp;
6130         tmp_opt.user_mss  = tp->rx_opt.user_mss;
6131         tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6132 
6133         if (want_cookie && !tmp_opt.saw_tstamp)
6134                 tcp_clear_options(&tmp_opt);
6135 
6136         tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6137         tcp_openreq_init(req, &tmp_opt, skb, sk);
6138 
6139         /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6140         inet_rsk(req)->ir_iif = sk->sk_bound_dev_if;
6141 
6142         af_ops->init_req(req, sk, skb);
6143 
6144         if (security_inet_conn_request(sk, skb, req))
6145                 goto drop_and_free;
6146 
6147         if (!want_cookie && !isn) {
6148                 /* VJ's idea. We save last timestamp seen
6149                  * from the destination in peer table, when entering
6150                  * state TIME-WAIT, and check against it before
6151                  * accepting new connection request.
6152                  *
6153                  * If "isn" is not zero, this request hit alive
6154                  * timewait bucket, so that all the necessary checks
6155                  * are made in the function processing timewait state.
6156                  */
6157                 if (tcp_death_row.sysctl_tw_recycle) {
6158                         bool strict;
6159 
6160                         dst = af_ops->route_req(sk, &fl, req, &strict);
6161 
6162                         if (dst && strict &&
6163                             !tcp_peer_is_proven(req, dst, true,
6164                                                 tmp_opt.saw_tstamp)) {
6165                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
6166                                 goto drop_and_release;
6167                         }
6168                 }
6169                 /* Kill the following clause, if you dislike this way. */
6170                 else if (!sysctl_tcp_syncookies &&
6171                          (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6172                           (sysctl_max_syn_backlog >> 2)) &&
6173                          !tcp_peer_is_proven(req, dst, false,
6174                                              tmp_opt.saw_tstamp)) {
6175                         /* Without syncookies last quarter of
6176                          * backlog is filled with destinations,
6177                          * proven to be alive.
6178                          * It means that we continue to communicate
6179                          * to destinations, already remembered
6180                          * to the moment of synflood.
6181                          */
6182                         pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6183                                     rsk_ops->family);
6184                         goto drop_and_release;
6185                 }
6186 
6187                 isn = af_ops->init_seq(skb);
6188         }
6189         if (!dst) {
6190                 dst = af_ops->route_req(sk, &fl, req, NULL);
6191                 if (!dst)
6192                         goto drop_and_free;
6193         }
6194 
6195         tcp_ecn_create_request(req, skb, sk, dst);
6196 
6197         if (want_cookie) {
6198                 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6199                 req->cookie_ts = tmp_opt.tstamp_ok;
6200                 if (!tmp_opt.tstamp_ok)
6201                         inet_rsk(req)->ecn_ok = 0;
6202         }
6203 
6204         tcp_rsk(req)->snt_isn = isn;
6205         tcp_openreq_init_rwin(req, sk, dst);
6206         fastopen = !want_cookie &&
6207                    tcp_try_fastopen(sk, skb, req, &foc, dst);
6208         err = af_ops->send_synack(sk, dst, &fl, req,
6209                                   skb_get_queue_mapping(skb), &foc);
6210         if (!fastopen) {
6211                 if (err || want_cookie)
6212                         goto drop_and_free;
6213 
6214                 tcp_rsk(req)->tfo_listener = false;
6215                 af_ops->queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6216         }
6217         tcp_reqsk_record_syn(sk, req, skb);
6218 
6219         return 0;
6220 
6221 drop_and_release:
6222         dst_release(dst);
6223 drop_and_free:
6224         reqsk_free(req);
6225 drop:
6226         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS);
6227         return 0;
6228 }
6229 EXPORT_SYMBOL(tcp_conn_request);
6230 

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