Version:  2.0.40 2.2.26 2.4.37 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 4.3 4.4 4.5 4.6 4.7

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_max_reordering __read_mostly = 300;
 84 int sysctl_tcp_dsack __read_mostly = 1;
 85 int sysctl_tcp_app_win __read_mostly = 31;
 86 int sysctl_tcp_adv_win_scale __read_mostly = 1;
 87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
 88 
 89 /* rfc5961 challenge ack rate limiting */
 90 int sysctl_tcp_challenge_ack_limit = 1000;
 91 
 92 int sysctl_tcp_stdurg __read_mostly;
 93 int sysctl_tcp_rfc1337 __read_mostly;
 94 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
 95 int sysctl_tcp_frto __read_mostly = 2;
 96 int sysctl_tcp_min_rtt_wlen __read_mostly = 300;
 97 
 98 int sysctl_tcp_thin_dupack __read_mostly;
 99 
100 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
101 int sysctl_tcp_early_retrans __read_mostly = 3;
102 int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
103 
104 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
105 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
106 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
107 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
108 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
109 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
110 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
111 #define FLAG_LOST_RETRANS       0x80 /* This ACK marks some retransmission lost */
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 #define REXMIT_NONE     0 /* no loss recovery to do */
128 #define REXMIT_LOST     1 /* retransmit packets marked lost */
129 #define REXMIT_NEW      2 /* FRTO-style transmit of unsent/new packets */
130 
131 /* Adapt the MSS value used to make delayed ack decision to the
132  * real world.
133  */
134 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
135 {
136         struct inet_connection_sock *icsk = inet_csk(sk);
137         const unsigned int lss = icsk->icsk_ack.last_seg_size;
138         unsigned int len;
139 
140         icsk->icsk_ack.last_seg_size = 0;
141 
142         /* skb->len may jitter because of SACKs, even if peer
143          * sends good full-sized frames.
144          */
145         len = skb_shinfo(skb)->gso_size ? : skb->len;
146         if (len >= icsk->icsk_ack.rcv_mss) {
147                 icsk->icsk_ack.rcv_mss = len;
148         } else {
149                 /* Otherwise, we make more careful check taking into account,
150                  * that SACKs block is variable.
151                  *
152                  * "len" is invariant segment length, including TCP header.
153                  */
154                 len += skb->data - skb_transport_header(skb);
155                 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
156                     /* If PSH is not set, packet should be
157                      * full sized, provided peer TCP is not badly broken.
158                      * This observation (if it is correct 8)) allows
159                      * to handle super-low mtu links fairly.
160                      */
161                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
162                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
163                         /* Subtract also invariant (if peer is RFC compliant),
164                          * tcp header plus fixed timestamp option length.
165                          * Resulting "len" is MSS free of SACK jitter.
166                          */
167                         len -= tcp_sk(sk)->tcp_header_len;
168                         icsk->icsk_ack.last_seg_size = len;
169                         if (len == lss) {
170                                 icsk->icsk_ack.rcv_mss = len;
171                                 return;
172                         }
173                 }
174                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
175                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
176                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
177         }
178 }
179 
180 static void tcp_incr_quickack(struct sock *sk)
181 {
182         struct inet_connection_sock *icsk = inet_csk(sk);
183         unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
184 
185         if (quickacks == 0)
186                 quickacks = 2;
187         if (quickacks > icsk->icsk_ack.quick)
188                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
189 }
190 
191 static void tcp_enter_quickack_mode(struct sock *sk)
192 {
193         struct inet_connection_sock *icsk = inet_csk(sk);
194         tcp_incr_quickack(sk);
195         icsk->icsk_ack.pingpong = 0;
196         icsk->icsk_ack.ato = TCP_ATO_MIN;
197 }
198 
199 /* Send ACKs quickly, if "quick" count is not exhausted
200  * and the session is not interactive.
201  */
202 
203 static bool tcp_in_quickack_mode(struct sock *sk)
204 {
205         const struct inet_connection_sock *icsk = inet_csk(sk);
206         const struct dst_entry *dst = __sk_dst_get(sk);
207 
208         return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
209                 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
210 }
211 
212 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
213 {
214         if (tp->ecn_flags & TCP_ECN_OK)
215                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
216 }
217 
218 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
219 {
220         if (tcp_hdr(skb)->cwr)
221                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
222 }
223 
224 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
225 {
226         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
227 }
228 
229 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
230 {
231         switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
232         case INET_ECN_NOT_ECT:
233                 /* Funny extension: if ECT is not set on a segment,
234                  * and we already seen ECT on a previous segment,
235                  * it is probably a retransmit.
236                  */
237                 if (tp->ecn_flags & TCP_ECN_SEEN)
238                         tcp_enter_quickack_mode((struct sock *)tp);
239                 break;
240         case INET_ECN_CE:
241                 if (tcp_ca_needs_ecn((struct sock *)tp))
242                         tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
243 
244                 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
245                         /* Better not delay acks, sender can have a very low cwnd */
246                         tcp_enter_quickack_mode((struct sock *)tp);
247                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
248                 }
249                 tp->ecn_flags |= TCP_ECN_SEEN;
250                 break;
251         default:
252                 if (tcp_ca_needs_ecn((struct sock *)tp))
253                         tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
254                 tp->ecn_flags |= TCP_ECN_SEEN;
255                 break;
256         }
257 }
258 
259 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
260 {
261         if (tp->ecn_flags & TCP_ECN_OK)
262                 __tcp_ecn_check_ce(tp, skb);
263 }
264 
265 static void tcp_ecn_rcv_synack(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 void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
272 {
273         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
274                 tp->ecn_flags &= ~TCP_ECN_OK;
275 }
276 
277 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
278 {
279         if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
280                 return true;
281         return false;
282 }
283 
284 /* Buffer size and advertised window tuning.
285  *
286  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
287  */
288 
289 static void tcp_sndbuf_expand(struct sock *sk)
290 {
291         const struct tcp_sock *tp = tcp_sk(sk);
292         int sndmem, per_mss;
293         u32 nr_segs;
294 
295         /* Worst case is non GSO/TSO : each frame consumes one skb
296          * and skb->head is kmalloced using power of two area of memory
297          */
298         per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
299                   MAX_TCP_HEADER +
300                   SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
301 
302         per_mss = roundup_pow_of_two(per_mss) +
303                   SKB_DATA_ALIGN(sizeof(struct sk_buff));
304 
305         nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
306         nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
307 
308         /* Fast Recovery (RFC 5681 3.2) :
309          * Cubic needs 1.7 factor, rounded to 2 to include
310          * extra cushion (application might react slowly to POLLOUT)
311          */
312         sndmem = 2 * nr_segs * per_mss;
313 
314         if (sk->sk_sndbuf < sndmem)
315                 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
316 }
317 
318 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
319  *
320  * All tcp_full_space() is split to two parts: "network" buffer, allocated
321  * forward and advertised in receiver window (tp->rcv_wnd) and
322  * "application buffer", required to isolate scheduling/application
323  * latencies from network.
324  * window_clamp is maximal advertised window. It can be less than
325  * tcp_full_space(), in this case tcp_full_space() - window_clamp
326  * is reserved for "application" buffer. The less window_clamp is
327  * the smoother our behaviour from viewpoint of network, but the lower
328  * throughput and the higher sensitivity of the connection to losses. 8)
329  *
330  * rcv_ssthresh is more strict window_clamp used at "slow start"
331  * phase to predict further behaviour of this connection.
332  * It is used for two goals:
333  * - to enforce header prediction at sender, even when application
334  *   requires some significant "application buffer". It is check #1.
335  * - to prevent pruning of receive queue because of misprediction
336  *   of receiver window. Check #2.
337  *
338  * The scheme does not work when sender sends good segments opening
339  * window and then starts to feed us spaghetti. But it should work
340  * in common situations. Otherwise, we have to rely on queue collapsing.
341  */
342 
343 /* Slow part of check#2. */
344 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
345 {
346         struct tcp_sock *tp = tcp_sk(sk);
347         /* Optimize this! */
348         int truesize = tcp_win_from_space(skb->truesize) >> 1;
349         int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
350 
351         while (tp->rcv_ssthresh <= window) {
352                 if (truesize <= skb->len)
353                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
354 
355                 truesize >>= 1;
356                 window >>= 1;
357         }
358         return 0;
359 }
360 
361 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
362 {
363         struct tcp_sock *tp = tcp_sk(sk);
364 
365         /* Check #1 */
366         if (tp->rcv_ssthresh < tp->window_clamp &&
367             (int)tp->rcv_ssthresh < tcp_space(sk) &&
368             !tcp_under_memory_pressure(sk)) {
369                 int incr;
370 
371                 /* Check #2. Increase window, if skb with such overhead
372                  * will fit to rcvbuf in future.
373                  */
374                 if (tcp_win_from_space(skb->truesize) <= skb->len)
375                         incr = 2 * tp->advmss;
376                 else
377                         incr = __tcp_grow_window(sk, skb);
378 
379                 if (incr) {
380                         incr = max_t(int, incr, 2 * skb->len);
381                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
382                                                tp->window_clamp);
383                         inet_csk(sk)->icsk_ack.quick |= 1;
384                 }
385         }
386 }
387 
388 /* 3. Tuning rcvbuf, when connection enters established state. */
389 static void tcp_fixup_rcvbuf(struct sock *sk)
390 {
391         u32 mss = tcp_sk(sk)->advmss;
392         int rcvmem;
393 
394         rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
395                  tcp_default_init_rwnd(mss);
396 
397         /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
398          * Allow enough cushion so that sender is not limited by our window
399          */
400         if (sysctl_tcp_moderate_rcvbuf)
401                 rcvmem <<= 2;
402 
403         if (sk->sk_rcvbuf < rcvmem)
404                 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
405 }
406 
407 /* 4. Try to fixup all. It is made immediately after connection enters
408  *    established state.
409  */
410 void tcp_init_buffer_space(struct sock *sk)
411 {
412         struct tcp_sock *tp = tcp_sk(sk);
413         int maxwin;
414 
415         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
416                 tcp_fixup_rcvbuf(sk);
417         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
418                 tcp_sndbuf_expand(sk);
419 
420         tp->rcvq_space.space = tp->rcv_wnd;
421         tp->rcvq_space.time = tcp_time_stamp;
422         tp->rcvq_space.seq = tp->copied_seq;
423 
424         maxwin = tcp_full_space(sk);
425 
426         if (tp->window_clamp >= maxwin) {
427                 tp->window_clamp = maxwin;
428 
429                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
430                         tp->window_clamp = max(maxwin -
431                                                (maxwin >> sysctl_tcp_app_win),
432                                                4 * tp->advmss);
433         }
434 
435         /* Force reservation of one segment. */
436         if (sysctl_tcp_app_win &&
437             tp->window_clamp > 2 * tp->advmss &&
438             tp->window_clamp + tp->advmss > maxwin)
439                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
440 
441         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
442         tp->snd_cwnd_stamp = tcp_time_stamp;
443 }
444 
445 /* 5. Recalculate window clamp after socket hit its memory bounds. */
446 static void tcp_clamp_window(struct sock *sk)
447 {
448         struct tcp_sock *tp = tcp_sk(sk);
449         struct inet_connection_sock *icsk = inet_csk(sk);
450 
451         icsk->icsk_ack.quick = 0;
452 
453         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
454             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
455             !tcp_under_memory_pressure(sk) &&
456             sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
457                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
458                                     sysctl_tcp_rmem[2]);
459         }
460         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
461                 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
462 }
463 
464 /* Initialize RCV_MSS value.
465  * RCV_MSS is an our guess about MSS used by the peer.
466  * We haven't any direct information about the MSS.
467  * It's better to underestimate the RCV_MSS rather than overestimate.
468  * Overestimations make us ACKing less frequently than needed.
469  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
470  */
471 void tcp_initialize_rcv_mss(struct sock *sk)
472 {
473         const struct tcp_sock *tp = tcp_sk(sk);
474         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
475 
476         hint = min(hint, tp->rcv_wnd / 2);
477         hint = min(hint, TCP_MSS_DEFAULT);
478         hint = max(hint, TCP_MIN_MSS);
479 
480         inet_csk(sk)->icsk_ack.rcv_mss = hint;
481 }
482 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
483 
484 /* Receiver "autotuning" code.
485  *
486  * The algorithm for RTT estimation w/o timestamps is based on
487  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
488  * <http://public.lanl.gov/radiant/pubs.html#DRS>
489  *
490  * More detail on this code can be found at
491  * <http://staff.psc.edu/jheffner/>,
492  * though this reference is out of date.  A new paper
493  * is pending.
494  */
495 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
496 {
497         u32 new_sample = tp->rcv_rtt_est.rtt;
498         long m = sample;
499 
500         if (m == 0)
501                 m = 1;
502 
503         if (new_sample != 0) {
504                 /* If we sample in larger samples in the non-timestamp
505                  * case, we could grossly overestimate the RTT especially
506                  * with chatty applications or bulk transfer apps which
507                  * are stalled on filesystem I/O.
508                  *
509                  * Also, since we are only going for a minimum in the
510                  * non-timestamp case, we do not smooth things out
511                  * else with timestamps disabled convergence takes too
512                  * long.
513                  */
514                 if (!win_dep) {
515                         m -= (new_sample >> 3);
516                         new_sample += m;
517                 } else {
518                         m <<= 3;
519                         if (m < new_sample)
520                                 new_sample = m;
521                 }
522         } else {
523                 /* No previous measure. */
524                 new_sample = m << 3;
525         }
526 
527         if (tp->rcv_rtt_est.rtt != new_sample)
528                 tp->rcv_rtt_est.rtt = new_sample;
529 }
530 
531 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
532 {
533         if (tp->rcv_rtt_est.time == 0)
534                 goto new_measure;
535         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
536                 return;
537         tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
538 
539 new_measure:
540         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
541         tp->rcv_rtt_est.time = tcp_time_stamp;
542 }
543 
544 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
545                                           const struct sk_buff *skb)
546 {
547         struct tcp_sock *tp = tcp_sk(sk);
548         if (tp->rx_opt.rcv_tsecr &&
549             (TCP_SKB_CB(skb)->end_seq -
550              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
551                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
552 }
553 
554 /*
555  * This function should be called every time data is copied to user space.
556  * It calculates the appropriate TCP receive buffer space.
557  */
558 void tcp_rcv_space_adjust(struct sock *sk)
559 {
560         struct tcp_sock *tp = tcp_sk(sk);
561         int time;
562         int copied;
563 
564         time = tcp_time_stamp - tp->rcvq_space.time;
565         if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
566                 return;
567 
568         /* Number of bytes copied to user in last RTT */
569         copied = tp->copied_seq - tp->rcvq_space.seq;
570         if (copied <= tp->rcvq_space.space)
571                 goto new_measure;
572 
573         /* A bit of theory :
574          * copied = bytes received in previous RTT, our base window
575          * To cope with packet losses, we need a 2x factor
576          * To cope with slow start, and sender growing its cwin by 100 %
577          * every RTT, we need a 4x factor, because the ACK we are sending
578          * now is for the next RTT, not the current one :
579          * <prev RTT . ><current RTT .. ><next RTT .... >
580          */
581 
582         if (sysctl_tcp_moderate_rcvbuf &&
583             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
584                 int rcvwin, rcvmem, rcvbuf;
585 
586                 /* minimal window to cope with packet losses, assuming
587                  * steady state. Add some cushion because of small variations.
588                  */
589                 rcvwin = (copied << 1) + 16 * tp->advmss;
590 
591                 /* If rate increased by 25%,
592                  *      assume slow start, rcvwin = 3 * copied
593                  * If rate increased by 50%,
594                  *      assume sender can use 2x growth, rcvwin = 4 * copied
595                  */
596                 if (copied >=
597                     tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
598                         if (copied >=
599                             tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
600                                 rcvwin <<= 1;
601                         else
602                                 rcvwin += (rcvwin >> 1);
603                 }
604 
605                 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
606                 while (tcp_win_from_space(rcvmem) < tp->advmss)
607                         rcvmem += 128;
608 
609                 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
610                 if (rcvbuf > sk->sk_rcvbuf) {
611                         sk->sk_rcvbuf = rcvbuf;
612 
613                         /* Make the window clamp follow along.  */
614                         tp->window_clamp = rcvwin;
615                 }
616         }
617         tp->rcvq_space.space = copied;
618 
619 new_measure:
620         tp->rcvq_space.seq = tp->copied_seq;
621         tp->rcvq_space.time = tcp_time_stamp;
622 }
623 
624 /* There is something which you must keep in mind when you analyze the
625  * behavior of the tp->ato delayed ack timeout interval.  When a
626  * connection starts up, we want to ack as quickly as possible.  The
627  * problem is that "good" TCP's do slow start at the beginning of data
628  * transmission.  The means that until we send the first few ACK's the
629  * sender will sit on his end and only queue most of his data, because
630  * he can only send snd_cwnd unacked packets at any given time.  For
631  * each ACK we send, he increments snd_cwnd and transmits more of his
632  * queue.  -DaveM
633  */
634 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
635 {
636         struct tcp_sock *tp = tcp_sk(sk);
637         struct inet_connection_sock *icsk = inet_csk(sk);
638         u32 now;
639 
640         inet_csk_schedule_ack(sk);
641 
642         tcp_measure_rcv_mss(sk, skb);
643 
644         tcp_rcv_rtt_measure(tp);
645 
646         now = tcp_time_stamp;
647 
648         if (!icsk->icsk_ack.ato) {
649                 /* The _first_ data packet received, initialize
650                  * delayed ACK engine.
651                  */
652                 tcp_incr_quickack(sk);
653                 icsk->icsk_ack.ato = TCP_ATO_MIN;
654         } else {
655                 int m = now - icsk->icsk_ack.lrcvtime;
656 
657                 if (m <= TCP_ATO_MIN / 2) {
658                         /* The fastest case is the first. */
659                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
660                 } else if (m < icsk->icsk_ack.ato) {
661                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
662                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
663                                 icsk->icsk_ack.ato = icsk->icsk_rto;
664                 } else if (m > icsk->icsk_rto) {
665                         /* Too long gap. Apparently sender failed to
666                          * restart window, so that we send ACKs quickly.
667                          */
668                         tcp_incr_quickack(sk);
669                         sk_mem_reclaim(sk);
670                 }
671         }
672         icsk->icsk_ack.lrcvtime = now;
673 
674         tcp_ecn_check_ce(tp, skb);
675 
676         if (skb->len >= 128)
677                 tcp_grow_window(sk, skb);
678 }
679 
680 /* Called to compute a smoothed rtt estimate. The data fed to this
681  * routine either comes from timestamps, or from segments that were
682  * known _not_ to have been retransmitted [see Karn/Partridge
683  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
684  * piece by Van Jacobson.
685  * NOTE: the next three routines used to be one big routine.
686  * To save cycles in the RFC 1323 implementation it was better to break
687  * it up into three procedures. -- erics
688  */
689 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
690 {
691         struct tcp_sock *tp = tcp_sk(sk);
692         long m = mrtt_us; /* RTT */
693         u32 srtt = tp->srtt_us;
694 
695         /*      The following amusing code comes from Jacobson's
696          *      article in SIGCOMM '88.  Note that rtt and mdev
697          *      are scaled versions of rtt and mean deviation.
698          *      This is designed to be as fast as possible
699          *      m stands for "measurement".
700          *
701          *      On a 1990 paper the rto value is changed to:
702          *      RTO = rtt + 4 * mdev
703          *
704          * Funny. This algorithm seems to be very broken.
705          * These formulae increase RTO, when it should be decreased, increase
706          * too slowly, when it should be increased quickly, decrease too quickly
707          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
708          * does not matter how to _calculate_ it. Seems, it was trap
709          * that VJ failed to avoid. 8)
710          */
711         if (srtt != 0) {
712                 m -= (srtt >> 3);       /* m is now error in rtt est */
713                 srtt += m;              /* rtt = 7/8 rtt + 1/8 new */
714                 if (m < 0) {
715                         m = -m;         /* m is now abs(error) */
716                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
717                         /* This is similar to one of Eifel findings.
718                          * Eifel blocks mdev updates when rtt decreases.
719                          * This solution is a bit different: we use finer gain
720                          * for mdev in this case (alpha*beta).
721                          * Like Eifel it also prevents growth of rto,
722                          * but also it limits too fast rto decreases,
723                          * happening in pure Eifel.
724                          */
725                         if (m > 0)
726                                 m >>= 3;
727                 } else {
728                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
729                 }
730                 tp->mdev_us += m;               /* mdev = 3/4 mdev + 1/4 new */
731                 if (tp->mdev_us > tp->mdev_max_us) {
732                         tp->mdev_max_us = tp->mdev_us;
733                         if (tp->mdev_max_us > tp->rttvar_us)
734                                 tp->rttvar_us = tp->mdev_max_us;
735                 }
736                 if (after(tp->snd_una, tp->rtt_seq)) {
737                         if (tp->mdev_max_us < tp->rttvar_us)
738                                 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
739                         tp->rtt_seq = tp->snd_nxt;
740                         tp->mdev_max_us = tcp_rto_min_us(sk);
741                 }
742         } else {
743                 /* no previous measure. */
744                 srtt = m << 3;          /* take the measured time to be rtt */
745                 tp->mdev_us = m << 1;   /* make sure rto = 3*rtt */
746                 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
747                 tp->mdev_max_us = tp->rttvar_us;
748                 tp->rtt_seq = tp->snd_nxt;
749         }
750         tp->srtt_us = max(1U, srtt);
751 }
752 
753 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
754  * Note: TCP stack does not yet implement pacing.
755  * FQ packet scheduler can be used to implement cheap but effective
756  * TCP pacing, to smooth the burst on large writes when packets
757  * in flight is significantly lower than cwnd (or rwin)
758  */
759 int sysctl_tcp_pacing_ss_ratio __read_mostly = 200;
760 int sysctl_tcp_pacing_ca_ratio __read_mostly = 120;
761 
762 static void tcp_update_pacing_rate(struct sock *sk)
763 {
764         const struct tcp_sock *tp = tcp_sk(sk);
765         u64 rate;
766 
767         /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
768         rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
769 
770         /* current rate is (cwnd * mss) / srtt
771          * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
772          * In Congestion Avoidance phase, set it to 120 % the current rate.
773          *
774          * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
775          *       If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
776          *       end of slow start and should slow down.
777          */
778         if (tp->snd_cwnd < tp->snd_ssthresh / 2)
779                 rate *= sysctl_tcp_pacing_ss_ratio;
780         else
781                 rate *= sysctl_tcp_pacing_ca_ratio;
782 
783         rate *= max(tp->snd_cwnd, tp->packets_out);
784 
785         if (likely(tp->srtt_us))
786                 do_div(rate, tp->srtt_us);
787 
788         /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
789          * without any lock. We want to make sure compiler wont store
790          * intermediate values in this location.
791          */
792         ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
793                                                 sk->sk_max_pacing_rate);
794 }
795 
796 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
797  * routine referred to above.
798  */
799 static void tcp_set_rto(struct sock *sk)
800 {
801         const struct tcp_sock *tp = tcp_sk(sk);
802         /* Old crap is replaced with new one. 8)
803          *
804          * More seriously:
805          * 1. If rtt variance happened to be less 50msec, it is hallucination.
806          *    It cannot be less due to utterly erratic ACK generation made
807          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
808          *    to do with delayed acks, because at cwnd>2 true delack timeout
809          *    is invisible. Actually, Linux-2.4 also generates erratic
810          *    ACKs in some circumstances.
811          */
812         inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
813 
814         /* 2. Fixups made earlier cannot be right.
815          *    If we do not estimate RTO correctly without them,
816          *    all the algo is pure shit and should be replaced
817          *    with correct one. It is exactly, which we pretend to do.
818          */
819 
820         /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
821          * guarantees that rto is higher.
822          */
823         tcp_bound_rto(sk);
824 }
825 
826 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
827 {
828         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
829 
830         if (!cwnd)
831                 cwnd = TCP_INIT_CWND;
832         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
833 }
834 
835 /*
836  * Packet counting of FACK is based on in-order assumptions, therefore TCP
837  * disables it when reordering is detected
838  */
839 void tcp_disable_fack(struct tcp_sock *tp)
840 {
841         /* RFC3517 uses different metric in lost marker => reset on change */
842         if (tcp_is_fack(tp))
843                 tp->lost_skb_hint = NULL;
844         tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
845 }
846 
847 /* Take a notice that peer is sending D-SACKs */
848 static void tcp_dsack_seen(struct tcp_sock *tp)
849 {
850         tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
851 }
852 
853 static void tcp_update_reordering(struct sock *sk, const int metric,
854                                   const int ts)
855 {
856         struct tcp_sock *tp = tcp_sk(sk);
857         if (metric > tp->reordering) {
858                 int mib_idx;
859 
860                 tp->reordering = min(sysctl_tcp_max_reordering, metric);
861 
862                 /* This exciting event is worth to be remembered. 8) */
863                 if (ts)
864                         mib_idx = LINUX_MIB_TCPTSREORDER;
865                 else if (tcp_is_reno(tp))
866                         mib_idx = LINUX_MIB_TCPRENOREORDER;
867                 else if (tcp_is_fack(tp))
868                         mib_idx = LINUX_MIB_TCPFACKREORDER;
869                 else
870                         mib_idx = LINUX_MIB_TCPSACKREORDER;
871 
872                 NET_INC_STATS(sock_net(sk), mib_idx);
873 #if FASTRETRANS_DEBUG > 1
874                 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
875                          tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
876                          tp->reordering,
877                          tp->fackets_out,
878                          tp->sacked_out,
879                          tp->undo_marker ? tp->undo_retrans : 0);
880 #endif
881                 tcp_disable_fack(tp);
882         }
883 
884         if (metric > 0)
885                 tcp_disable_early_retrans(tp);
886         tp->rack.reord = 1;
887 }
888 
889 /* This must be called before lost_out is incremented */
890 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
891 {
892         if (!tp->retransmit_skb_hint ||
893             before(TCP_SKB_CB(skb)->seq,
894                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
895                 tp->retransmit_skb_hint = skb;
896 
897         if (!tp->lost_out ||
898             after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
899                 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
900 }
901 
902 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
903 {
904         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
905                 tcp_verify_retransmit_hint(tp, skb);
906 
907                 tp->lost_out += tcp_skb_pcount(skb);
908                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
909         }
910 }
911 
912 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
913 {
914         tcp_verify_retransmit_hint(tp, skb);
915 
916         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
917                 tp->lost_out += tcp_skb_pcount(skb);
918                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
919         }
920 }
921 
922 /* This procedure tags the retransmission queue when SACKs arrive.
923  *
924  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
925  * Packets in queue with these bits set are counted in variables
926  * sacked_out, retrans_out and lost_out, correspondingly.
927  *
928  * Valid combinations are:
929  * Tag  InFlight        Description
930  * 0    1               - orig segment is in flight.
931  * S    0               - nothing flies, orig reached receiver.
932  * L    0               - nothing flies, orig lost by net.
933  * R    2               - both orig and retransmit are in flight.
934  * L|R  1               - orig is lost, retransmit is in flight.
935  * S|R  1               - orig reached receiver, retrans is still in flight.
936  * (L|S|R is logically valid, it could occur when L|R is sacked,
937  *  but it is equivalent to plain S and code short-curcuits it to S.
938  *  L|S is logically invalid, it would mean -1 packet in flight 8))
939  *
940  * These 6 states form finite state machine, controlled by the following events:
941  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
942  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
943  * 3. Loss detection event of two flavors:
944  *      A. Scoreboard estimator decided the packet is lost.
945  *         A'. Reno "three dupacks" marks head of queue lost.
946  *         A''. Its FACK modification, head until snd.fack is lost.
947  *      B. SACK arrives sacking SND.NXT at the moment, when the
948  *         segment was retransmitted.
949  * 4. D-SACK added new rule: D-SACK changes any tag to S.
950  *
951  * It is pleasant to note, that state diagram turns out to be commutative,
952  * so that we are allowed not to be bothered by order of our actions,
953  * when multiple events arrive simultaneously. (see the function below).
954  *
955  * Reordering detection.
956  * --------------------
957  * Reordering metric is maximal distance, which a packet can be displaced
958  * in packet stream. With SACKs we can estimate it:
959  *
960  * 1. SACK fills old hole and the corresponding segment was not
961  *    ever retransmitted -> reordering. Alas, we cannot use it
962  *    when segment was retransmitted.
963  * 2. The last flaw is solved with D-SACK. D-SACK arrives
964  *    for retransmitted and already SACKed segment -> reordering..
965  * Both of these heuristics are not used in Loss state, when we cannot
966  * account for retransmits accurately.
967  *
968  * SACK block validation.
969  * ----------------------
970  *
971  * SACK block range validation checks that the received SACK block fits to
972  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
973  * Note that SND.UNA is not included to the range though being valid because
974  * it means that the receiver is rather inconsistent with itself reporting
975  * SACK reneging when it should advance SND.UNA. Such SACK block this is
976  * perfectly valid, however, in light of RFC2018 which explicitly states
977  * that "SACK block MUST reflect the newest segment.  Even if the newest
978  * segment is going to be discarded ...", not that it looks very clever
979  * in case of head skb. Due to potentional receiver driven attacks, we
980  * choose to avoid immediate execution of a walk in write queue due to
981  * reneging and defer head skb's loss recovery to standard loss recovery
982  * procedure that will eventually trigger (nothing forbids us doing this).
983  *
984  * Implements also blockage to start_seq wrap-around. Problem lies in the
985  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
986  * there's no guarantee that it will be before snd_nxt (n). The problem
987  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
988  * wrap (s_w):
989  *
990  *         <- outs wnd ->                          <- wrapzone ->
991  *         u     e      n                         u_w   e_w  s n_w
992  *         |     |      |                          |     |   |  |
993  * |<------------+------+----- TCP seqno space --------------+---------->|
994  * ...-- <2^31 ->|                                           |<--------...
995  * ...---- >2^31 ------>|                                    |<--------...
996  *
997  * Current code wouldn't be vulnerable but it's better still to discard such
998  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
999  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1000  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1001  * equal to the ideal case (infinite seqno space without wrap caused issues).
1002  *
1003  * With D-SACK the lower bound is extended to cover sequence space below
1004  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1005  * again, D-SACK block must not to go across snd_una (for the same reason as
1006  * for the normal SACK blocks, explained above). But there all simplicity
1007  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1008  * fully below undo_marker they do not affect behavior in anyway and can
1009  * therefore be safely ignored. In rare cases (which are more or less
1010  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1011  * fragmentation and packet reordering past skb's retransmission. To consider
1012  * them correctly, the acceptable range must be extended even more though
1013  * the exact amount is rather hard to quantify. However, tp->max_window can
1014  * be used as an exaggerated estimate.
1015  */
1016 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1017                                    u32 start_seq, u32 end_seq)
1018 {
1019         /* Too far in future, or reversed (interpretation is ambiguous) */
1020         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1021                 return false;
1022 
1023         /* Nasty start_seq wrap-around check (see comments above) */
1024         if (!before(start_seq, tp->snd_nxt))
1025                 return false;
1026 
1027         /* In outstanding window? ...This is valid exit for D-SACKs too.
1028          * start_seq == snd_una is non-sensical (see comments above)
1029          */
1030         if (after(start_seq, tp->snd_una))
1031                 return true;
1032 
1033         if (!is_dsack || !tp->undo_marker)
1034                 return false;
1035 
1036         /* ...Then it's D-SACK, and must reside below snd_una completely */
1037         if (after(end_seq, tp->snd_una))
1038                 return false;
1039 
1040         if (!before(start_seq, tp->undo_marker))
1041                 return true;
1042 
1043         /* Too old */
1044         if (!after(end_seq, tp->undo_marker))
1045                 return false;
1046 
1047         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1048          *   start_seq < undo_marker and end_seq >= undo_marker.
1049          */
1050         return !before(start_seq, end_seq - tp->max_window);
1051 }
1052 
1053 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1054                             struct tcp_sack_block_wire *sp, int num_sacks,
1055                             u32 prior_snd_una)
1056 {
1057         struct tcp_sock *tp = tcp_sk(sk);
1058         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1059         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1060         bool dup_sack = false;
1061 
1062         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1063                 dup_sack = true;
1064                 tcp_dsack_seen(tp);
1065                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1066         } else if (num_sacks > 1) {
1067                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1068                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1069 
1070                 if (!after(end_seq_0, end_seq_1) &&
1071                     !before(start_seq_0, start_seq_1)) {
1072                         dup_sack = true;
1073                         tcp_dsack_seen(tp);
1074                         NET_INC_STATS(sock_net(sk),
1075                                         LINUX_MIB_TCPDSACKOFORECV);
1076                 }
1077         }
1078 
1079         /* D-SACK for already forgotten data... Do dumb counting. */
1080         if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1081             !after(end_seq_0, prior_snd_una) &&
1082             after(end_seq_0, tp->undo_marker))
1083                 tp->undo_retrans--;
1084 
1085         return dup_sack;
1086 }
1087 
1088 struct tcp_sacktag_state {
1089         int     reord;
1090         int     fack_count;
1091         /* Timestamps for earliest and latest never-retransmitted segment
1092          * that was SACKed. RTO needs the earliest RTT to stay conservative,
1093          * but congestion control should still get an accurate delay signal.
1094          */
1095         struct skb_mstamp first_sackt;
1096         struct skb_mstamp last_sackt;
1097         int     flag;
1098 };
1099 
1100 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1101  * the incoming SACK may not exactly match but we can find smaller MSS
1102  * aligned portion of it that matches. Therefore we might need to fragment
1103  * which may fail and creates some hassle (caller must handle error case
1104  * returns).
1105  *
1106  * FIXME: this could be merged to shift decision code
1107  */
1108 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1109                                   u32 start_seq, u32 end_seq)
1110 {
1111         int err;
1112         bool in_sack;
1113         unsigned int pkt_len;
1114         unsigned int mss;
1115 
1116         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1117                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1118 
1119         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1120             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1121                 mss = tcp_skb_mss(skb);
1122                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1123 
1124                 if (!in_sack) {
1125                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1126                         if (pkt_len < mss)
1127                                 pkt_len = mss;
1128                 } else {
1129                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1130                         if (pkt_len < mss)
1131                                 return -EINVAL;
1132                 }
1133 
1134                 /* Round if necessary so that SACKs cover only full MSSes
1135                  * and/or the remaining small portion (if present)
1136                  */
1137                 if (pkt_len > mss) {
1138                         unsigned int new_len = (pkt_len / mss) * mss;
1139                         if (!in_sack && new_len < pkt_len) {
1140                                 new_len += mss;
1141                                 if (new_len >= skb->len)
1142                                         return 0;
1143                         }
1144                         pkt_len = new_len;
1145                 }
1146                 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1147                 if (err < 0)
1148                         return err;
1149         }
1150 
1151         return in_sack;
1152 }
1153 
1154 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1155 static u8 tcp_sacktag_one(struct sock *sk,
1156                           struct tcp_sacktag_state *state, u8 sacked,
1157                           u32 start_seq, u32 end_seq,
1158                           int dup_sack, int pcount,
1159                           const struct skb_mstamp *xmit_time)
1160 {
1161         struct tcp_sock *tp = tcp_sk(sk);
1162         int fack_count = state->fack_count;
1163 
1164         /* Account D-SACK for retransmitted packet. */
1165         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1166                 if (tp->undo_marker && tp->undo_retrans > 0 &&
1167                     after(end_seq, tp->undo_marker))
1168                         tp->undo_retrans--;
1169                 if (sacked & TCPCB_SACKED_ACKED)
1170                         state->reord = min(fack_count, state->reord);
1171         }
1172 
1173         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1174         if (!after(end_seq, tp->snd_una))
1175                 return sacked;
1176 
1177         if (!(sacked & TCPCB_SACKED_ACKED)) {
1178                 tcp_rack_advance(tp, xmit_time, sacked);
1179 
1180                 if (sacked & TCPCB_SACKED_RETRANS) {
1181                         /* If the segment is not tagged as lost,
1182                          * we do not clear RETRANS, believing
1183                          * that retransmission is still in flight.
1184                          */
1185                         if (sacked & TCPCB_LOST) {
1186                                 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1187                                 tp->lost_out -= pcount;
1188                                 tp->retrans_out -= pcount;
1189                         }
1190                 } else {
1191                         if (!(sacked & TCPCB_RETRANS)) {
1192                                 /* New sack for not retransmitted frame,
1193                                  * which was in hole. It is reordering.
1194                                  */
1195                                 if (before(start_seq,
1196                                            tcp_highest_sack_seq(tp)))
1197                                         state->reord = min(fack_count,
1198                                                            state->reord);
1199                                 if (!after(end_seq, tp->high_seq))
1200                                         state->flag |= FLAG_ORIG_SACK_ACKED;
1201                                 if (state->first_sackt.v64 == 0)
1202                                         state->first_sackt = *xmit_time;
1203                                 state->last_sackt = *xmit_time;
1204                         }
1205 
1206                         if (sacked & TCPCB_LOST) {
1207                                 sacked &= ~TCPCB_LOST;
1208                                 tp->lost_out -= pcount;
1209                         }
1210                 }
1211 
1212                 sacked |= TCPCB_SACKED_ACKED;
1213                 state->flag |= FLAG_DATA_SACKED;
1214                 tp->sacked_out += pcount;
1215                 tp->delivered += pcount;  /* Out-of-order packets delivered */
1216 
1217                 fack_count += pcount;
1218 
1219                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1220                 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1221                     before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1222                         tp->lost_cnt_hint += pcount;
1223 
1224                 if (fack_count > tp->fackets_out)
1225                         tp->fackets_out = fack_count;
1226         }
1227 
1228         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1229          * frames and clear it. undo_retrans is decreased above, L|R frames
1230          * are accounted above as well.
1231          */
1232         if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1233                 sacked &= ~TCPCB_SACKED_RETRANS;
1234                 tp->retrans_out -= pcount;
1235         }
1236 
1237         return sacked;
1238 }
1239 
1240 /* Shift newly-SACKed bytes from this skb to the immediately previous
1241  * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1242  */
1243 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1244                             struct tcp_sacktag_state *state,
1245                             unsigned int pcount, int shifted, int mss,
1246                             bool dup_sack)
1247 {
1248         struct tcp_sock *tp = tcp_sk(sk);
1249         struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1250         u32 start_seq = TCP_SKB_CB(skb)->seq;   /* start of newly-SACKed */
1251         u32 end_seq = start_seq + shifted;      /* end of newly-SACKed */
1252 
1253         BUG_ON(!pcount);
1254 
1255         /* Adjust counters and hints for the newly sacked sequence
1256          * range but discard the return value since prev is already
1257          * marked. We must tag the range first because the seq
1258          * advancement below implicitly advances
1259          * tcp_highest_sack_seq() when skb is highest_sack.
1260          */
1261         tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1262                         start_seq, end_seq, dup_sack, pcount,
1263                         &skb->skb_mstamp);
1264 
1265         if (skb == tp->lost_skb_hint)
1266                 tp->lost_cnt_hint += pcount;
1267 
1268         TCP_SKB_CB(prev)->end_seq += shifted;
1269         TCP_SKB_CB(skb)->seq += shifted;
1270 
1271         tcp_skb_pcount_add(prev, pcount);
1272         BUG_ON(tcp_skb_pcount(skb) < pcount);
1273         tcp_skb_pcount_add(skb, -pcount);
1274 
1275         /* When we're adding to gso_segs == 1, gso_size will be zero,
1276          * in theory this shouldn't be necessary but as long as DSACK
1277          * code can come after this skb later on it's better to keep
1278          * setting gso_size to something.
1279          */
1280         if (!TCP_SKB_CB(prev)->tcp_gso_size)
1281                 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1282 
1283         /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1284         if (tcp_skb_pcount(skb) <= 1)
1285                 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1286 
1287         /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1288         TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1289 
1290         if (skb->len > 0) {
1291                 BUG_ON(!tcp_skb_pcount(skb));
1292                 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1293                 return false;
1294         }
1295 
1296         /* Whole SKB was eaten :-) */
1297 
1298         if (skb == tp->retransmit_skb_hint)
1299                 tp->retransmit_skb_hint = prev;
1300         if (skb == tp->lost_skb_hint) {
1301                 tp->lost_skb_hint = prev;
1302                 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1303         }
1304 
1305         TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1306         TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1307         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1308                 TCP_SKB_CB(prev)->end_seq++;
1309 
1310         if (skb == tcp_highest_sack(sk))
1311                 tcp_advance_highest_sack(sk, skb);
1312 
1313         tcp_skb_collapse_tstamp(prev, skb);
1314         tcp_unlink_write_queue(skb, sk);
1315         sk_wmem_free_skb(sk, skb);
1316 
1317         NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1318 
1319         return true;
1320 }
1321 
1322 /* I wish gso_size would have a bit more sane initialization than
1323  * something-or-zero which complicates things
1324  */
1325 static int tcp_skb_seglen(const struct sk_buff *skb)
1326 {
1327         return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1328 }
1329 
1330 /* Shifting pages past head area doesn't work */
1331 static int skb_can_shift(const struct sk_buff *skb)
1332 {
1333         return !skb_headlen(skb) && skb_is_nonlinear(skb);
1334 }
1335 
1336 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1337  * skb.
1338  */
1339 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1340                                           struct tcp_sacktag_state *state,
1341                                           u32 start_seq, u32 end_seq,
1342                                           bool dup_sack)
1343 {
1344         struct tcp_sock *tp = tcp_sk(sk);
1345         struct sk_buff *prev;
1346         int mss;
1347         int pcount = 0;
1348         int len;
1349         int in_sack;
1350 
1351         if (!sk_can_gso(sk))
1352                 goto fallback;
1353 
1354         /* Normally R but no L won't result in plain S */
1355         if (!dup_sack &&
1356             (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1357                 goto fallback;
1358         if (!skb_can_shift(skb))
1359                 goto fallback;
1360         /* This frame is about to be dropped (was ACKed). */
1361         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1362                 goto fallback;
1363 
1364         /* Can only happen with delayed DSACK + discard craziness */
1365         if (unlikely(skb == tcp_write_queue_head(sk)))
1366                 goto fallback;
1367         prev = tcp_write_queue_prev(sk, skb);
1368 
1369         if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1370                 goto fallback;
1371 
1372         if (!tcp_skb_can_collapse_to(prev))
1373                 goto fallback;
1374 
1375         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1376                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1377 
1378         if (in_sack) {
1379                 len = skb->len;
1380                 pcount = tcp_skb_pcount(skb);
1381                 mss = tcp_skb_seglen(skb);
1382 
1383                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1384                  * drop this restriction as unnecessary
1385                  */
1386                 if (mss != tcp_skb_seglen(prev))
1387                         goto fallback;
1388         } else {
1389                 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1390                         goto noop;
1391                 /* CHECKME: This is non-MSS split case only?, this will
1392                  * cause skipped skbs due to advancing loop btw, original
1393                  * has that feature too
1394                  */
1395                 if (tcp_skb_pcount(skb) <= 1)
1396                         goto noop;
1397 
1398                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1399                 if (!in_sack) {
1400                         /* TODO: head merge to next could be attempted here
1401                          * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1402                          * though it might not be worth of the additional hassle
1403                          *
1404                          * ...we can probably just fallback to what was done
1405                          * previously. We could try merging non-SACKed ones
1406                          * as well but it probably isn't going to buy off
1407                          * because later SACKs might again split them, and
1408                          * it would make skb timestamp tracking considerably
1409                          * harder problem.
1410                          */
1411                         goto fallback;
1412                 }
1413 
1414                 len = end_seq - TCP_SKB_CB(skb)->seq;
1415                 BUG_ON(len < 0);
1416                 BUG_ON(len > skb->len);
1417 
1418                 /* MSS boundaries should be honoured or else pcount will
1419                  * severely break even though it makes things bit trickier.
1420                  * Optimize common case to avoid most of the divides
1421                  */
1422                 mss = tcp_skb_mss(skb);
1423 
1424                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1425                  * drop this restriction as unnecessary
1426                  */
1427                 if (mss != tcp_skb_seglen(prev))
1428                         goto fallback;
1429 
1430                 if (len == mss) {
1431                         pcount = 1;
1432                 } else if (len < mss) {
1433                         goto noop;
1434                 } else {
1435                         pcount = len / mss;
1436                         len = pcount * mss;
1437                 }
1438         }
1439 
1440         /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1441         if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1442                 goto fallback;
1443 
1444         if (!skb_shift(prev, skb, len))
1445                 goto fallback;
1446         if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1447                 goto out;
1448 
1449         /* Hole filled allows collapsing with the next as well, this is very
1450          * useful when hole on every nth skb pattern happens
1451          */
1452         if (prev == tcp_write_queue_tail(sk))
1453                 goto out;
1454         skb = tcp_write_queue_next(sk, prev);
1455 
1456         if (!skb_can_shift(skb) ||
1457             (skb == tcp_send_head(sk)) ||
1458             ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1459             (mss != tcp_skb_seglen(skb)))
1460                 goto out;
1461 
1462         len = skb->len;
1463         if (skb_shift(prev, skb, len)) {
1464                 pcount += tcp_skb_pcount(skb);
1465                 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1466         }
1467 
1468 out:
1469         state->fack_count += pcount;
1470         return prev;
1471 
1472 noop:
1473         return skb;
1474 
1475 fallback:
1476         NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1477         return NULL;
1478 }
1479 
1480 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1481                                         struct tcp_sack_block *next_dup,
1482                                         struct tcp_sacktag_state *state,
1483                                         u32 start_seq, u32 end_seq,
1484                                         bool dup_sack_in)
1485 {
1486         struct tcp_sock *tp = tcp_sk(sk);
1487         struct sk_buff *tmp;
1488 
1489         tcp_for_write_queue_from(skb, sk) {
1490                 int in_sack = 0;
1491                 bool dup_sack = dup_sack_in;
1492 
1493                 if (skb == tcp_send_head(sk))
1494                         break;
1495 
1496                 /* queue is in-order => we can short-circuit the walk early */
1497                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1498                         break;
1499 
1500                 if (next_dup  &&
1501                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1502                         in_sack = tcp_match_skb_to_sack(sk, skb,
1503                                                         next_dup->start_seq,
1504                                                         next_dup->end_seq);
1505                         if (in_sack > 0)
1506                                 dup_sack = true;
1507                 }
1508 
1509                 /* skb reference here is a bit tricky to get right, since
1510                  * shifting can eat and free both this skb and the next,
1511                  * so not even _safe variant of the loop is enough.
1512                  */
1513                 if (in_sack <= 0) {
1514                         tmp = tcp_shift_skb_data(sk, skb, state,
1515                                                  start_seq, end_seq, dup_sack);
1516                         if (tmp) {
1517                                 if (tmp != skb) {
1518                                         skb = tmp;
1519                                         continue;
1520                                 }
1521 
1522                                 in_sack = 0;
1523                         } else {
1524                                 in_sack = tcp_match_skb_to_sack(sk, skb,
1525                                                                 start_seq,
1526                                                                 end_seq);
1527                         }
1528                 }
1529 
1530                 if (unlikely(in_sack < 0))
1531                         break;
1532 
1533                 if (in_sack) {
1534                         TCP_SKB_CB(skb)->sacked =
1535                                 tcp_sacktag_one(sk,
1536                                                 state,
1537                                                 TCP_SKB_CB(skb)->sacked,
1538                                                 TCP_SKB_CB(skb)->seq,
1539                                                 TCP_SKB_CB(skb)->end_seq,
1540                                                 dup_sack,
1541                                                 tcp_skb_pcount(skb),
1542                                                 &skb->skb_mstamp);
1543 
1544                         if (!before(TCP_SKB_CB(skb)->seq,
1545                                     tcp_highest_sack_seq(tp)))
1546                                 tcp_advance_highest_sack(sk, skb);
1547                 }
1548 
1549                 state->fack_count += tcp_skb_pcount(skb);
1550         }
1551         return skb;
1552 }
1553 
1554 /* Avoid all extra work that is being done by sacktag while walking in
1555  * a normal way
1556  */
1557 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1558                                         struct tcp_sacktag_state *state,
1559                                         u32 skip_to_seq)
1560 {
1561         tcp_for_write_queue_from(skb, sk) {
1562                 if (skb == tcp_send_head(sk))
1563                         break;
1564 
1565                 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1566                         break;
1567 
1568                 state->fack_count += tcp_skb_pcount(skb);
1569         }
1570         return skb;
1571 }
1572 
1573 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1574                                                 struct sock *sk,
1575                                                 struct tcp_sack_block *next_dup,
1576                                                 struct tcp_sacktag_state *state,
1577                                                 u32 skip_to_seq)
1578 {
1579         if (!next_dup)
1580                 return skb;
1581 
1582         if (before(next_dup->start_seq, skip_to_seq)) {
1583                 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1584                 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1585                                        next_dup->start_seq, next_dup->end_seq,
1586                                        1);
1587         }
1588 
1589         return skb;
1590 }
1591 
1592 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1593 {
1594         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1595 }
1596 
1597 static int
1598 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1599                         u32 prior_snd_una, struct tcp_sacktag_state *state)
1600 {
1601         struct tcp_sock *tp = tcp_sk(sk);
1602         const unsigned char *ptr = (skb_transport_header(ack_skb) +
1603                                     TCP_SKB_CB(ack_skb)->sacked);
1604         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1605         struct tcp_sack_block sp[TCP_NUM_SACKS];
1606         struct tcp_sack_block *cache;
1607         struct sk_buff *skb;
1608         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1609         int used_sacks;
1610         bool found_dup_sack = false;
1611         int i, j;
1612         int first_sack_index;
1613 
1614         state->flag = 0;
1615         state->reord = tp->packets_out;
1616 
1617         if (!tp->sacked_out) {
1618                 if (WARN_ON(tp->fackets_out))
1619                         tp->fackets_out = 0;
1620                 tcp_highest_sack_reset(sk);
1621         }
1622 
1623         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1624                                          num_sacks, prior_snd_una);
1625         if (found_dup_sack)
1626                 state->flag |= FLAG_DSACKING_ACK;
1627 
1628         /* Eliminate too old ACKs, but take into
1629          * account more or less fresh ones, they can
1630          * contain valid SACK info.
1631          */
1632         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1633                 return 0;
1634 
1635         if (!tp->packets_out)
1636                 goto out;
1637 
1638         used_sacks = 0;
1639         first_sack_index = 0;
1640         for (i = 0; i < num_sacks; i++) {
1641                 bool dup_sack = !i && found_dup_sack;
1642 
1643                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1644                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1645 
1646                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1647                                             sp[used_sacks].start_seq,
1648                                             sp[used_sacks].end_seq)) {
1649                         int mib_idx;
1650 
1651                         if (dup_sack) {
1652                                 if (!tp->undo_marker)
1653                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1654                                 else
1655                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1656                         } else {
1657                                 /* Don't count olds caused by ACK reordering */
1658                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1659                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1660                                         continue;
1661                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1662                         }
1663 
1664                         NET_INC_STATS(sock_net(sk), mib_idx);
1665                         if (i == 0)
1666                                 first_sack_index = -1;
1667                         continue;
1668                 }
1669 
1670                 /* Ignore very old stuff early */
1671                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1672                         continue;
1673 
1674                 used_sacks++;
1675         }
1676 
1677         /* order SACK blocks to allow in order walk of the retrans queue */
1678         for (i = used_sacks - 1; i > 0; i--) {
1679                 for (j = 0; j < i; j++) {
1680                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1681                                 swap(sp[j], sp[j + 1]);
1682 
1683                                 /* Track where the first SACK block goes to */
1684                                 if (j == first_sack_index)
1685                                         first_sack_index = j + 1;
1686                         }
1687                 }
1688         }
1689 
1690         skb = tcp_write_queue_head(sk);
1691         state->fack_count = 0;
1692         i = 0;
1693 
1694         if (!tp->sacked_out) {
1695                 /* It's already past, so skip checking against it */
1696                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1697         } else {
1698                 cache = tp->recv_sack_cache;
1699                 /* Skip empty blocks in at head of the cache */
1700                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1701                        !cache->end_seq)
1702                         cache++;
1703         }
1704 
1705         while (i < used_sacks) {
1706                 u32 start_seq = sp[i].start_seq;
1707                 u32 end_seq = sp[i].end_seq;
1708                 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1709                 struct tcp_sack_block *next_dup = NULL;
1710 
1711                 if (found_dup_sack && ((i + 1) == first_sack_index))
1712                         next_dup = &sp[i + 1];
1713 
1714                 /* Skip too early cached blocks */
1715                 while (tcp_sack_cache_ok(tp, cache) &&
1716                        !before(start_seq, cache->end_seq))
1717                         cache++;
1718 
1719                 /* Can skip some work by looking recv_sack_cache? */
1720                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1721                     after(end_seq, cache->start_seq)) {
1722 
1723                         /* Head todo? */
1724                         if (before(start_seq, cache->start_seq)) {
1725                                 skb = tcp_sacktag_skip(skb, sk, state,
1726                                                        start_seq);
1727                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1728                                                        state,
1729                                                        start_seq,
1730                                                        cache->start_seq,
1731                                                        dup_sack);
1732                         }
1733 
1734                         /* Rest of the block already fully processed? */
1735                         if (!after(end_seq, cache->end_seq))
1736                                 goto advance_sp;
1737 
1738                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1739                                                        state,
1740                                                        cache->end_seq);
1741 
1742                         /* ...tail remains todo... */
1743                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1744                                 /* ...but better entrypoint exists! */
1745                                 skb = tcp_highest_sack(sk);
1746                                 if (!skb)
1747                                         break;
1748                                 state->fack_count = tp->fackets_out;
1749                                 cache++;
1750                                 goto walk;
1751                         }
1752 
1753                         skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1754                         /* Check overlap against next cached too (past this one already) */
1755                         cache++;
1756                         continue;
1757                 }
1758 
1759                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1760                         skb = tcp_highest_sack(sk);
1761                         if (!skb)
1762                                 break;
1763                         state->fack_count = tp->fackets_out;
1764                 }
1765                 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1766 
1767 walk:
1768                 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1769                                        start_seq, end_seq, dup_sack);
1770 
1771 advance_sp:
1772                 i++;
1773         }
1774 
1775         /* Clear the head of the cache sack blocks so we can skip it next time */
1776         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1777                 tp->recv_sack_cache[i].start_seq = 0;
1778                 tp->recv_sack_cache[i].end_seq = 0;
1779         }
1780         for (j = 0; j < used_sacks; j++)
1781                 tp->recv_sack_cache[i++] = sp[j];
1782 
1783         if ((state->reord < tp->fackets_out) &&
1784             ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1785                 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
1786 
1787         tcp_verify_left_out(tp);
1788 out:
1789 
1790 #if FASTRETRANS_DEBUG > 0
1791         WARN_ON((int)tp->sacked_out < 0);
1792         WARN_ON((int)tp->lost_out < 0);
1793         WARN_ON((int)tp->retrans_out < 0);
1794         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1795 #endif
1796         return state->flag;
1797 }
1798 
1799 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1800  * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1801  */
1802 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1803 {
1804         u32 holes;
1805 
1806         holes = max(tp->lost_out, 1U);
1807         holes = min(holes, tp->packets_out);
1808 
1809         if ((tp->sacked_out + holes) > tp->packets_out) {
1810                 tp->sacked_out = tp->packets_out - holes;
1811                 return true;
1812         }
1813         return false;
1814 }
1815 
1816 /* If we receive more dupacks than we expected counting segments
1817  * in assumption of absent reordering, interpret this as reordering.
1818  * The only another reason could be bug in receiver TCP.
1819  */
1820 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1821 {
1822         struct tcp_sock *tp = tcp_sk(sk);
1823         if (tcp_limit_reno_sacked(tp))
1824                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1825 }
1826 
1827 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1828 
1829 static void tcp_add_reno_sack(struct sock *sk)
1830 {
1831         struct tcp_sock *tp = tcp_sk(sk);
1832         u32 prior_sacked = tp->sacked_out;
1833 
1834         tp->sacked_out++;
1835         tcp_check_reno_reordering(sk, 0);
1836         if (tp->sacked_out > prior_sacked)
1837                 tp->delivered++; /* Some out-of-order packet is delivered */
1838         tcp_verify_left_out(tp);
1839 }
1840 
1841 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1842 
1843 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1844 {
1845         struct tcp_sock *tp = tcp_sk(sk);
1846 
1847         if (acked > 0) {
1848                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1849                 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1850                 if (acked - 1 >= tp->sacked_out)
1851                         tp->sacked_out = 0;
1852                 else
1853                         tp->sacked_out -= acked - 1;
1854         }
1855         tcp_check_reno_reordering(sk, acked);
1856         tcp_verify_left_out(tp);
1857 }
1858 
1859 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1860 {
1861         tp->sacked_out = 0;
1862 }
1863 
1864 void tcp_clear_retrans(struct tcp_sock *tp)
1865 {
1866         tp->retrans_out = 0;
1867         tp->lost_out = 0;
1868         tp->undo_marker = 0;
1869         tp->undo_retrans = -1;
1870         tp->fackets_out = 0;
1871         tp->sacked_out = 0;
1872 }
1873 
1874 static inline void tcp_init_undo(struct tcp_sock *tp)
1875 {
1876         tp->undo_marker = tp->snd_una;
1877         /* Retransmission still in flight may cause DSACKs later. */
1878         tp->undo_retrans = tp->retrans_out ? : -1;
1879 }
1880 
1881 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1882  * and reset tags completely, otherwise preserve SACKs. If receiver
1883  * dropped its ofo queue, we will know this due to reneging detection.
1884  */
1885 void tcp_enter_loss(struct sock *sk)
1886 {
1887         const struct inet_connection_sock *icsk = inet_csk(sk);
1888         struct tcp_sock *tp = tcp_sk(sk);
1889         struct net *net = sock_net(sk);
1890         struct sk_buff *skb;
1891         bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1892         bool is_reneg;                  /* is receiver reneging on SACKs? */
1893 
1894         /* Reduce ssthresh if it has not yet been made inside this window. */
1895         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1896             !after(tp->high_seq, tp->snd_una) ||
1897             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1898                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1899                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1900                 tcp_ca_event(sk, CA_EVENT_LOSS);
1901                 tcp_init_undo(tp);
1902         }
1903         tp->snd_cwnd       = 1;
1904         tp->snd_cwnd_cnt   = 0;
1905         tp->snd_cwnd_stamp = tcp_time_stamp;
1906 
1907         tp->retrans_out = 0;
1908         tp->lost_out = 0;
1909 
1910         if (tcp_is_reno(tp))
1911                 tcp_reset_reno_sack(tp);
1912 
1913         skb = tcp_write_queue_head(sk);
1914         is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1915         if (is_reneg) {
1916                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1917                 tp->sacked_out = 0;
1918                 tp->fackets_out = 0;
1919         }
1920         tcp_clear_all_retrans_hints(tp);
1921 
1922         tcp_for_write_queue(skb, sk) {
1923                 if (skb == tcp_send_head(sk))
1924                         break;
1925 
1926                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1927                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
1928                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1929                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1930                         tp->lost_out += tcp_skb_pcount(skb);
1931                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1932                 }
1933         }
1934         tcp_verify_left_out(tp);
1935 
1936         /* Timeout in disordered state after receiving substantial DUPACKs
1937          * suggests that the degree of reordering is over-estimated.
1938          */
1939         if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1940             tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
1941                 tp->reordering = min_t(unsigned int, tp->reordering,
1942                                        net->ipv4.sysctl_tcp_reordering);
1943         tcp_set_ca_state(sk, TCP_CA_Loss);
1944         tp->high_seq = tp->snd_nxt;
1945         tcp_ecn_queue_cwr(tp);
1946 
1947         /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1948          * loss recovery is underway except recurring timeout(s) on
1949          * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1950          */
1951         tp->frto = sysctl_tcp_frto &&
1952                    (new_recovery || icsk->icsk_retransmits) &&
1953                    !inet_csk(sk)->icsk_mtup.probe_size;
1954 }
1955 
1956 /* If ACK arrived pointing to a remembered SACK, it means that our
1957  * remembered SACKs do not reflect real state of receiver i.e.
1958  * receiver _host_ is heavily congested (or buggy).
1959  *
1960  * To avoid big spurious retransmission bursts due to transient SACK
1961  * scoreboard oddities that look like reneging, we give the receiver a
1962  * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1963  * restore sanity to the SACK scoreboard. If the apparent reneging
1964  * persists until this RTO then we'll clear the SACK scoreboard.
1965  */
1966 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1967 {
1968         if (flag & FLAG_SACK_RENEGING) {
1969                 struct tcp_sock *tp = tcp_sk(sk);
1970                 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
1971                                           msecs_to_jiffies(10));
1972 
1973                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1974                                           delay, TCP_RTO_MAX);
1975                 return true;
1976         }
1977         return false;
1978 }
1979 
1980 static inline int tcp_fackets_out(const struct tcp_sock *tp)
1981 {
1982         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
1983 }
1984 
1985 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1986  * counter when SACK is enabled (without SACK, sacked_out is used for
1987  * that purpose).
1988  *
1989  * Instead, with FACK TCP uses fackets_out that includes both SACKed
1990  * segments up to the highest received SACK block so far and holes in
1991  * between them.
1992  *
1993  * With reordering, holes may still be in flight, so RFC3517 recovery
1994  * uses pure sacked_out (total number of SACKed segments) even though
1995  * it violates the RFC that uses duplicate ACKs, often these are equal
1996  * but when e.g. out-of-window ACKs or packet duplication occurs,
1997  * they differ. Since neither occurs due to loss, TCP should really
1998  * ignore them.
1999  */
2000 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2001 {
2002         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2003 }
2004 
2005 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2006 {
2007         struct tcp_sock *tp = tcp_sk(sk);
2008         unsigned long delay;
2009 
2010         /* Delay early retransmit and entering fast recovery for
2011          * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2012          * available, or RTO is scheduled to fire first.
2013          */
2014         if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2015             (flag & FLAG_ECE) || !tp->srtt_us)
2016                 return false;
2017 
2018         delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2019                     msecs_to_jiffies(2));
2020 
2021         if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2022                 return false;
2023 
2024         inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2025                                   TCP_RTO_MAX);
2026         return true;
2027 }
2028 
2029 /* Linux NewReno/SACK/FACK/ECN state machine.
2030  * --------------------------------------
2031  *
2032  * "Open"       Normal state, no dubious events, fast path.
2033  * "Disorder"   In all the respects it is "Open",
2034  *              but requires a bit more attention. It is entered when
2035  *              we see some SACKs or dupacks. It is split of "Open"
2036  *              mainly to move some processing from fast path to slow one.
2037  * "CWR"        CWND was reduced due to some Congestion Notification event.
2038  *              It can be ECN, ICMP source quench, local device congestion.
2039  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2040  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2041  *
2042  * tcp_fastretrans_alert() is entered:
2043  * - each incoming ACK, if state is not "Open"
2044  * - when arrived ACK is unusual, namely:
2045  *      * SACK
2046  *      * Duplicate ACK.
2047  *      * ECN ECE.
2048  *
2049  * Counting packets in flight is pretty simple.
2050  *
2051  *      in_flight = packets_out - left_out + retrans_out
2052  *
2053  *      packets_out is SND.NXT-SND.UNA counted in packets.
2054  *
2055  *      retrans_out is number of retransmitted segments.
2056  *
2057  *      left_out is number of segments left network, but not ACKed yet.
2058  *
2059  *              left_out = sacked_out + lost_out
2060  *
2061  *     sacked_out: Packets, which arrived to receiver out of order
2062  *                 and hence not ACKed. With SACKs this number is simply
2063  *                 amount of SACKed data. Even without SACKs
2064  *                 it is easy to give pretty reliable estimate of this number,
2065  *                 counting duplicate ACKs.
2066  *
2067  *       lost_out: Packets lost by network. TCP has no explicit
2068  *                 "loss notification" feedback from network (for now).
2069  *                 It means that this number can be only _guessed_.
2070  *                 Actually, it is the heuristics to predict lossage that
2071  *                 distinguishes different algorithms.
2072  *
2073  *      F.e. after RTO, when all the queue is considered as lost,
2074  *      lost_out = packets_out and in_flight = retrans_out.
2075  *
2076  *              Essentially, we have now two algorithms counting
2077  *              lost packets.
2078  *
2079  *              FACK: It is the simplest heuristics. As soon as we decided
2080  *              that something is lost, we decide that _all_ not SACKed
2081  *              packets until the most forward SACK are lost. I.e.
2082  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2083  *              It is absolutely correct estimate, if network does not reorder
2084  *              packets. And it loses any connection to reality when reordering
2085  *              takes place. We use FACK by default until reordering
2086  *              is suspected on the path to this destination.
2087  *
2088  *              NewReno: when Recovery is entered, we assume that one segment
2089  *              is lost (classic Reno). While we are in Recovery and
2090  *              a partial ACK arrives, we assume that one more packet
2091  *              is lost (NewReno). This heuristics are the same in NewReno
2092  *              and SACK.
2093  *
2094  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2095  *  deflation etc. CWND is real congestion window, never inflated, changes
2096  *  only according to classic VJ rules.
2097  *
2098  * Really tricky (and requiring careful tuning) part of algorithm
2099  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2100  * The first determines the moment _when_ we should reduce CWND and,
2101  * hence, slow down forward transmission. In fact, it determines the moment
2102  * when we decide that hole is caused by loss, rather than by a reorder.
2103  *
2104  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2105  * holes, caused by lost packets.
2106  *
2107  * And the most logically complicated part of algorithm is undo
2108  * heuristics. We detect false retransmits due to both too early
2109  * fast retransmit (reordering) and underestimated RTO, analyzing
2110  * timestamps and D-SACKs. When we detect that some segments were
2111  * retransmitted by mistake and CWND reduction was wrong, we undo
2112  * window reduction and abort recovery phase. This logic is hidden
2113  * inside several functions named tcp_try_undo_<something>.
2114  */
2115 
2116 /* This function decides, when we should leave Disordered state
2117  * and enter Recovery phase, reducing congestion window.
2118  *
2119  * Main question: may we further continue forward transmission
2120  * with the same cwnd?
2121  */
2122 static bool tcp_time_to_recover(struct sock *sk, int flag)
2123 {
2124         struct tcp_sock *tp = tcp_sk(sk);
2125         __u32 packets_out;
2126         int tcp_reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
2127 
2128         /* Trick#1: The loss is proven. */
2129         if (tp->lost_out)
2130                 return true;
2131 
2132         /* Not-A-Trick#2 : Classic rule... */
2133         if (tcp_dupack_heuristics(tp) > tp->reordering)
2134                 return true;
2135 
2136         /* Trick#4: It is still not OK... But will it be useful to delay
2137          * recovery more?
2138          */
2139         packets_out = tp->packets_out;
2140         if (packets_out <= tp->reordering &&
2141             tp->sacked_out >= max_t(__u32, packets_out/2, tcp_reordering) &&
2142             !tcp_may_send_now(sk)) {
2143                 /* We have nothing to send. This connection is limited
2144                  * either by receiver window or by application.
2145                  */
2146                 return true;
2147         }
2148 
2149         /* If a thin stream is detected, retransmit after first
2150          * received dupack. Employ only if SACK is supported in order
2151          * to avoid possible corner-case series of spurious retransmissions
2152          * Use only if there are no unsent data.
2153          */
2154         if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2155             tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2156             tcp_is_sack(tp) && !tcp_send_head(sk))
2157                 return true;
2158 
2159         /* Trick#6: TCP early retransmit, per RFC5827.  To avoid spurious
2160          * retransmissions due to small network reorderings, we implement
2161          * Mitigation A.3 in the RFC and delay the retransmission for a short
2162          * interval if appropriate.
2163          */
2164         if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2165             (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2166             !tcp_may_send_now(sk))
2167                 return !tcp_pause_early_retransmit(sk, flag);
2168 
2169         return false;
2170 }
2171 
2172 /* Detect loss in event "A" above by marking head of queue up as lost.
2173  * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2174  * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2175  * has at least tp->reordering SACKed seqments above it; "packets" refers to
2176  * the maximum SACKed segments to pass before reaching this limit.
2177  */
2178 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2179 {
2180         struct tcp_sock *tp = tcp_sk(sk);
2181         struct sk_buff *skb;
2182         int cnt, oldcnt, lost;
2183         unsigned int mss;
2184         /* Use SACK to deduce losses of new sequences sent during recovery */
2185         const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2186 
2187         WARN_ON(packets > tp->packets_out);
2188         if (tp->lost_skb_hint) {
2189                 skb = tp->lost_skb_hint;
2190                 cnt = tp->lost_cnt_hint;
2191                 /* Head already handled? */
2192                 if (mark_head && skb != tcp_write_queue_head(sk))
2193                         return;
2194         } else {
2195                 skb = tcp_write_queue_head(sk);
2196                 cnt = 0;
2197         }
2198 
2199         tcp_for_write_queue_from(skb, sk) {
2200                 if (skb == tcp_send_head(sk))
2201                         break;
2202                 /* TODO: do this better */
2203                 /* this is not the most efficient way to do this... */
2204                 tp->lost_skb_hint = skb;
2205                 tp->lost_cnt_hint = cnt;
2206 
2207                 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2208                         break;
2209 
2210                 oldcnt = cnt;
2211                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2212                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2213                         cnt += tcp_skb_pcount(skb);
2214 
2215                 if (cnt > packets) {
2216                         if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2217                             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2218                             (oldcnt >= packets))
2219                                 break;
2220 
2221                         mss = tcp_skb_mss(skb);
2222                         /* If needed, chop off the prefix to mark as lost. */
2223                         lost = (packets - oldcnt) * mss;
2224                         if (lost < skb->len &&
2225                             tcp_fragment(sk, skb, lost, mss, GFP_ATOMIC) < 0)
2226                                 break;
2227                         cnt = packets;
2228                 }
2229 
2230                 tcp_skb_mark_lost(tp, skb);
2231 
2232                 if (mark_head)
2233                         break;
2234         }
2235         tcp_verify_left_out(tp);
2236 }
2237 
2238 /* Account newly detected lost packet(s) */
2239 
2240 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2241 {
2242         struct tcp_sock *tp = tcp_sk(sk);
2243 
2244         if (tcp_is_reno(tp)) {
2245                 tcp_mark_head_lost(sk, 1, 1);
2246         } else if (tcp_is_fack(tp)) {
2247                 int lost = tp->fackets_out - tp->reordering;
2248                 if (lost <= 0)
2249                         lost = 1;
2250                 tcp_mark_head_lost(sk, lost, 0);
2251         } else {
2252                 int sacked_upto = tp->sacked_out - tp->reordering;
2253                 if (sacked_upto >= 0)
2254                         tcp_mark_head_lost(sk, sacked_upto, 0);
2255                 else if (fast_rexmit)
2256                         tcp_mark_head_lost(sk, 1, 1);
2257         }
2258 }
2259 
2260 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2261 {
2262         return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2263                before(tp->rx_opt.rcv_tsecr, when);
2264 }
2265 
2266 /* skb is spurious retransmitted if the returned timestamp echo
2267  * reply is prior to the skb transmission time
2268  */
2269 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2270                                      const struct sk_buff *skb)
2271 {
2272         return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2273                tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2274 }
2275 
2276 /* Nothing was retransmitted or returned timestamp is less
2277  * than timestamp of the first retransmission.
2278  */
2279 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2280 {
2281         return !tp->retrans_stamp ||
2282                tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2283 }
2284 
2285 /* Undo procedures. */
2286 
2287 /* We can clear retrans_stamp when there are no retransmissions in the
2288  * window. It would seem that it is trivially available for us in
2289  * tp->retrans_out, however, that kind of assumptions doesn't consider
2290  * what will happen if errors occur when sending retransmission for the
2291  * second time. ...It could the that such segment has only
2292  * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2293  * the head skb is enough except for some reneging corner cases that
2294  * are not worth the effort.
2295  *
2296  * Main reason for all this complexity is the fact that connection dying
2297  * time now depends on the validity of the retrans_stamp, in particular,
2298  * that successive retransmissions of a segment must not advance
2299  * retrans_stamp under any conditions.
2300  */
2301 static bool tcp_any_retrans_done(const struct sock *sk)
2302 {
2303         const struct tcp_sock *tp = tcp_sk(sk);
2304         struct sk_buff *skb;
2305 
2306         if (tp->retrans_out)
2307                 return true;
2308 
2309         skb = tcp_write_queue_head(sk);
2310         if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2311                 return true;
2312 
2313         return false;
2314 }
2315 
2316 #if FASTRETRANS_DEBUG > 1
2317 static void DBGUNDO(struct sock *sk, const char *msg)
2318 {
2319         struct tcp_sock *tp = tcp_sk(sk);
2320         struct inet_sock *inet = inet_sk(sk);
2321 
2322         if (sk->sk_family == AF_INET) {
2323                 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2324                          msg,
2325                          &inet->inet_daddr, ntohs(inet->inet_dport),
2326                          tp->snd_cwnd, tcp_left_out(tp),
2327                          tp->snd_ssthresh, tp->prior_ssthresh,
2328                          tp->packets_out);
2329         }
2330 #if IS_ENABLED(CONFIG_IPV6)
2331         else if (sk->sk_family == AF_INET6) {
2332                 struct ipv6_pinfo *np = inet6_sk(sk);
2333                 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2334                          msg,
2335                          &np->daddr, ntohs(inet->inet_dport),
2336                          tp->snd_cwnd, tcp_left_out(tp),
2337                          tp->snd_ssthresh, tp->prior_ssthresh,
2338                          tp->packets_out);
2339         }
2340 #endif
2341 }
2342 #else
2343 #define DBGUNDO(x...) do { } while (0)
2344 #endif
2345 
2346 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2347 {
2348         struct tcp_sock *tp = tcp_sk(sk);
2349 
2350         if (unmark_loss) {
2351                 struct sk_buff *skb;
2352 
2353                 tcp_for_write_queue(skb, sk) {
2354                         if (skb == tcp_send_head(sk))
2355                                 break;
2356                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2357                 }
2358                 tp->lost_out = 0;
2359                 tcp_clear_all_retrans_hints(tp);
2360         }
2361 
2362         if (tp->prior_ssthresh) {
2363                 const struct inet_connection_sock *icsk = inet_csk(sk);
2364 
2365                 if (icsk->icsk_ca_ops->undo_cwnd)
2366                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2367                 else
2368                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2369 
2370                 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2371                         tp->snd_ssthresh = tp->prior_ssthresh;
2372                         tcp_ecn_withdraw_cwr(tp);
2373                 }
2374         }
2375         tp->snd_cwnd_stamp = tcp_time_stamp;
2376         tp->undo_marker = 0;
2377 }
2378 
2379 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2380 {
2381         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2382 }
2383 
2384 /* People celebrate: "We love our President!" */
2385 static bool tcp_try_undo_recovery(struct sock *sk)
2386 {
2387         struct tcp_sock *tp = tcp_sk(sk);
2388 
2389         if (tcp_may_undo(tp)) {
2390                 int mib_idx;
2391 
2392                 /* Happy end! We did not retransmit anything
2393                  * or our original transmission succeeded.
2394                  */
2395                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2396                 tcp_undo_cwnd_reduction(sk, false);
2397                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2398                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2399                 else
2400                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2401 
2402                 NET_INC_STATS(sock_net(sk), mib_idx);
2403         }
2404         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2405                 /* Hold old state until something *above* high_seq
2406                  * is ACKed. For Reno it is MUST to prevent false
2407                  * fast retransmits (RFC2582). SACK TCP is safe. */
2408                 if (!tcp_any_retrans_done(sk))
2409                         tp->retrans_stamp = 0;
2410                 return true;
2411         }
2412         tcp_set_ca_state(sk, TCP_CA_Open);
2413         return false;
2414 }
2415 
2416 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2417 static bool tcp_try_undo_dsack(struct sock *sk)
2418 {
2419         struct tcp_sock *tp = tcp_sk(sk);
2420 
2421         if (tp->undo_marker && !tp->undo_retrans) {
2422                 DBGUNDO(sk, "D-SACK");
2423                 tcp_undo_cwnd_reduction(sk, false);
2424                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2425                 return true;
2426         }
2427         return false;
2428 }
2429 
2430 /* Undo during loss recovery after partial ACK or using F-RTO. */
2431 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2432 {
2433         struct tcp_sock *tp = tcp_sk(sk);
2434 
2435         if (frto_undo || tcp_may_undo(tp)) {
2436                 tcp_undo_cwnd_reduction(sk, true);
2437 
2438                 DBGUNDO(sk, "partial loss");
2439                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2440                 if (frto_undo)
2441                         NET_INC_STATS(sock_net(sk),
2442                                         LINUX_MIB_TCPSPURIOUSRTOS);
2443                 inet_csk(sk)->icsk_retransmits = 0;
2444                 if (frto_undo || tcp_is_sack(tp))
2445                         tcp_set_ca_state(sk, TCP_CA_Open);
2446                 return true;
2447         }
2448         return false;
2449 }
2450 
2451 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2452  * It computes the number of packets to send (sndcnt) based on packets newly
2453  * delivered:
2454  *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2455  *      cwnd reductions across a full RTT.
2456  *   2) Otherwise PRR uses packet conservation to send as much as delivered.
2457  *      But when the retransmits are acked without further losses, PRR
2458  *      slow starts cwnd up to ssthresh to speed up the recovery.
2459  */
2460 static void tcp_init_cwnd_reduction(struct sock *sk)
2461 {
2462         struct tcp_sock *tp = tcp_sk(sk);
2463 
2464         tp->high_seq = tp->snd_nxt;
2465         tp->tlp_high_seq = 0;
2466         tp->snd_cwnd_cnt = 0;
2467         tp->prior_cwnd = tp->snd_cwnd;
2468         tp->prr_delivered = 0;
2469         tp->prr_out = 0;
2470         tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2471         tcp_ecn_queue_cwr(tp);
2472 }
2473 
2474 static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked,
2475                                int flag)
2476 {
2477         struct tcp_sock *tp = tcp_sk(sk);
2478         int sndcnt = 0;
2479         int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2480 
2481         if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2482                 return;
2483 
2484         tp->prr_delivered += newly_acked_sacked;
2485         if (delta < 0) {
2486                 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2487                                tp->prior_cwnd - 1;
2488                 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2489         } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2490                    !(flag & FLAG_LOST_RETRANS)) {
2491                 sndcnt = min_t(int, delta,
2492                                max_t(int, tp->prr_delivered - tp->prr_out,
2493                                      newly_acked_sacked) + 1);
2494         } else {
2495                 sndcnt = min(delta, newly_acked_sacked);
2496         }
2497         /* Force a fast retransmit upon entering fast recovery */
2498         sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2499         tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2500 }
2501 
2502 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2503 {
2504         struct tcp_sock *tp = tcp_sk(sk);
2505 
2506         /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2507         if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2508             (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2509                 tp->snd_cwnd = tp->snd_ssthresh;
2510                 tp->snd_cwnd_stamp = tcp_time_stamp;
2511         }
2512         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2513 }
2514 
2515 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2516 void tcp_enter_cwr(struct sock *sk)
2517 {
2518         struct tcp_sock *tp = tcp_sk(sk);
2519 
2520         tp->prior_ssthresh = 0;
2521         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2522                 tp->undo_marker = 0;
2523                 tcp_init_cwnd_reduction(sk);
2524                 tcp_set_ca_state(sk, TCP_CA_CWR);
2525         }
2526 }
2527 EXPORT_SYMBOL(tcp_enter_cwr);
2528 
2529 static void tcp_try_keep_open(struct sock *sk)
2530 {
2531         struct tcp_sock *tp = tcp_sk(sk);
2532         int state = TCP_CA_Open;
2533 
2534         if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2535                 state = TCP_CA_Disorder;
2536 
2537         if (inet_csk(sk)->icsk_ca_state != state) {
2538                 tcp_set_ca_state(sk, state);
2539                 tp->high_seq = tp->snd_nxt;
2540         }
2541 }
2542 
2543 static void tcp_try_to_open(struct sock *sk, int flag)
2544 {
2545         struct tcp_sock *tp = tcp_sk(sk);
2546 
2547         tcp_verify_left_out(tp);
2548 
2549         if (!tcp_any_retrans_done(sk))
2550                 tp->retrans_stamp = 0;
2551 
2552         if (flag & FLAG_ECE)
2553                 tcp_enter_cwr(sk);
2554 
2555         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2556                 tcp_try_keep_open(sk);
2557         }
2558 }
2559 
2560 static void tcp_mtup_probe_failed(struct sock *sk)
2561 {
2562         struct inet_connection_sock *icsk = inet_csk(sk);
2563 
2564         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2565         icsk->icsk_mtup.probe_size = 0;
2566         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2567 }
2568 
2569 static void tcp_mtup_probe_success(struct sock *sk)
2570 {
2571         struct tcp_sock *tp = tcp_sk(sk);
2572         struct inet_connection_sock *icsk = inet_csk(sk);
2573 
2574         /* FIXME: breaks with very large cwnd */
2575         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2576         tp->snd_cwnd = tp->snd_cwnd *
2577                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2578                        icsk->icsk_mtup.probe_size;
2579         tp->snd_cwnd_cnt = 0;
2580         tp->snd_cwnd_stamp = tcp_time_stamp;
2581         tp->snd_ssthresh = tcp_current_ssthresh(sk);
2582 
2583         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2584         icsk->icsk_mtup.probe_size = 0;
2585         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2586         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2587 }
2588 
2589 /* Do a simple retransmit without using the backoff mechanisms in
2590  * tcp_timer. This is used for path mtu discovery.
2591  * The socket is already locked here.
2592  */
2593 void tcp_simple_retransmit(struct sock *sk)
2594 {
2595         const struct inet_connection_sock *icsk = inet_csk(sk);
2596         struct tcp_sock *tp = tcp_sk(sk);
2597         struct sk_buff *skb;
2598         unsigned int mss = tcp_current_mss(sk);
2599         u32 prior_lost = tp->lost_out;
2600 
2601         tcp_for_write_queue(skb, sk) {
2602                 if (skb == tcp_send_head(sk))
2603                         break;
2604                 if (tcp_skb_seglen(skb) > mss &&
2605                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2606                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2607                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2608                                 tp->retrans_out -= tcp_skb_pcount(skb);
2609                         }
2610                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2611                 }
2612         }
2613 
2614         tcp_clear_retrans_hints_partial(tp);
2615 
2616         if (prior_lost == tp->lost_out)
2617                 return;
2618 
2619         if (tcp_is_reno(tp))
2620                 tcp_limit_reno_sacked(tp);
2621 
2622         tcp_verify_left_out(tp);
2623 
2624         /* Don't muck with the congestion window here.
2625          * Reason is that we do not increase amount of _data_
2626          * in network, but units changed and effective
2627          * cwnd/ssthresh really reduced now.
2628          */
2629         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2630                 tp->high_seq = tp->snd_nxt;
2631                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2632                 tp->prior_ssthresh = 0;
2633                 tp->undo_marker = 0;
2634                 tcp_set_ca_state(sk, TCP_CA_Loss);
2635         }
2636         tcp_xmit_retransmit_queue(sk);
2637 }
2638 EXPORT_SYMBOL(tcp_simple_retransmit);
2639 
2640 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2641 {
2642         struct tcp_sock *tp = tcp_sk(sk);
2643         int mib_idx;
2644 
2645         if (tcp_is_reno(tp))
2646                 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2647         else
2648                 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2649 
2650         NET_INC_STATS(sock_net(sk), mib_idx);
2651 
2652         tp->prior_ssthresh = 0;
2653         tcp_init_undo(tp);
2654 
2655         if (!tcp_in_cwnd_reduction(sk)) {
2656                 if (!ece_ack)
2657                         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2658                 tcp_init_cwnd_reduction(sk);
2659         }
2660         tcp_set_ca_state(sk, TCP_CA_Recovery);
2661 }
2662 
2663 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2664  * recovered or spurious. Otherwise retransmits more on partial ACKs.
2665  */
2666 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2667                              int *rexmit)
2668 {
2669         struct tcp_sock *tp = tcp_sk(sk);
2670         bool recovered = !before(tp->snd_una, tp->high_seq);
2671 
2672         if ((flag & FLAG_SND_UNA_ADVANCED) &&
2673             tcp_try_undo_loss(sk, false))
2674                 return;
2675 
2676         if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2677                 /* Step 3.b. A timeout is spurious if not all data are
2678                  * lost, i.e., never-retransmitted data are (s)acked.
2679                  */
2680                 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2681                     tcp_try_undo_loss(sk, true))
2682                         return;
2683 
2684                 if (after(tp->snd_nxt, tp->high_seq)) {
2685                         if (flag & FLAG_DATA_SACKED || is_dupack)
2686                                 tp->frto = 0; /* Step 3.a. loss was real */
2687                 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2688                         tp->high_seq = tp->snd_nxt;
2689                         /* Step 2.b. Try send new data (but deferred until cwnd
2690                          * is updated in tcp_ack()). Otherwise fall back to
2691                          * the conventional recovery.
2692                          */
2693                         if (tcp_send_head(sk) &&
2694                             after(tcp_wnd_end(tp), tp->snd_nxt)) {
2695                                 *rexmit = REXMIT_NEW;
2696                                 return;
2697                         }
2698                         tp->frto = 0;
2699                 }
2700         }
2701 
2702         if (recovered) {
2703                 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2704                 tcp_try_undo_recovery(sk);
2705                 return;
2706         }
2707         if (tcp_is_reno(tp)) {
2708                 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2709                  * delivered. Lower inflight to clock out (re)tranmissions.
2710                  */
2711                 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2712                         tcp_add_reno_sack(sk);
2713                 else if (flag & FLAG_SND_UNA_ADVANCED)
2714                         tcp_reset_reno_sack(tp);
2715         }
2716         *rexmit = REXMIT_LOST;
2717 }
2718 
2719 /* Undo during fast recovery after partial ACK. */
2720 static bool tcp_try_undo_partial(struct sock *sk, const int acked)
2721 {
2722         struct tcp_sock *tp = tcp_sk(sk);
2723 
2724         if (tp->undo_marker && tcp_packet_delayed(tp)) {
2725                 /* Plain luck! Hole if filled with delayed
2726                  * packet, rather than with a retransmit.
2727                  */
2728                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2729 
2730                 /* We are getting evidence that the reordering degree is higher
2731                  * than we realized. If there are no retransmits out then we
2732                  * can undo. Otherwise we clock out new packets but do not
2733                  * mark more packets lost or retransmit more.
2734                  */
2735                 if (tp->retrans_out)
2736                         return true;
2737 
2738                 if (!tcp_any_retrans_done(sk))
2739                         tp->retrans_stamp = 0;
2740 
2741                 DBGUNDO(sk, "partial recovery");
2742                 tcp_undo_cwnd_reduction(sk, true);
2743                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2744                 tcp_try_keep_open(sk);
2745                 return true;
2746         }
2747         return false;
2748 }
2749 
2750 /* Process an event, which can update packets-in-flight not trivially.
2751  * Main goal of this function is to calculate new estimate for left_out,
2752  * taking into account both packets sitting in receiver's buffer and
2753  * packets lost by network.
2754  *
2755  * Besides that it updates the congestion state when packet loss or ECN
2756  * is detected. But it does not reduce the cwnd, it is done by the
2757  * congestion control later.
2758  *
2759  * It does _not_ decide what to send, it is made in function
2760  * tcp_xmit_retransmit_queue().
2761  */
2762 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2763                                   bool is_dupack, int *ack_flag, int *rexmit)
2764 {
2765         struct inet_connection_sock *icsk = inet_csk(sk);
2766         struct tcp_sock *tp = tcp_sk(sk);
2767         int fast_rexmit = 0, flag = *ack_flag;
2768         bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2769                                     (tcp_fackets_out(tp) > tp->reordering));
2770 
2771         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2772                 tp->sacked_out = 0;
2773         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2774                 tp->fackets_out = 0;
2775 
2776         /* Now state machine starts.
2777          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2778         if (flag & FLAG_ECE)
2779                 tp->prior_ssthresh = 0;
2780 
2781         /* B. In all the states check for reneging SACKs. */
2782         if (tcp_check_sack_reneging(sk, flag))
2783                 return;
2784 
2785         /* C. Check consistency of the current state. */
2786         tcp_verify_left_out(tp);
2787 
2788         /* D. Check state exit conditions. State can be terminated
2789          *    when high_seq is ACKed. */
2790         if (icsk->icsk_ca_state == TCP_CA_Open) {
2791                 WARN_ON(tp->retrans_out != 0);
2792                 tp->retrans_stamp = 0;
2793         } else if (!before(tp->snd_una, tp->high_seq)) {
2794                 switch (icsk->icsk_ca_state) {
2795                 case TCP_CA_CWR:
2796                         /* CWR is to be held something *above* high_seq
2797                          * is ACKed for CWR bit to reach receiver. */
2798                         if (tp->snd_una != tp->high_seq) {
2799                                 tcp_end_cwnd_reduction(sk);
2800                                 tcp_set_ca_state(sk, TCP_CA_Open);
2801                         }
2802                         break;
2803 
2804                 case TCP_CA_Recovery:
2805                         if (tcp_is_reno(tp))
2806                                 tcp_reset_reno_sack(tp);
2807                         if (tcp_try_undo_recovery(sk))
2808                                 return;
2809                         tcp_end_cwnd_reduction(sk);
2810                         break;
2811                 }
2812         }
2813 
2814         /* Use RACK to detect loss */
2815         if (sysctl_tcp_recovery & TCP_RACK_LOST_RETRANS &&
2816             tcp_rack_mark_lost(sk)) {
2817                 flag |= FLAG_LOST_RETRANS;
2818                 *ack_flag |= FLAG_LOST_RETRANS;
2819         }
2820 
2821         /* E. Process state. */
2822         switch (icsk->icsk_ca_state) {
2823         case TCP_CA_Recovery:
2824                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2825                         if (tcp_is_reno(tp) && is_dupack)
2826                                 tcp_add_reno_sack(sk);
2827                 } else {
2828                         if (tcp_try_undo_partial(sk, acked))
2829                                 return;
2830                         /* Partial ACK arrived. Force fast retransmit. */
2831                         do_lost = tcp_is_reno(tp) ||
2832                                   tcp_fackets_out(tp) > tp->reordering;
2833                 }
2834                 if (tcp_try_undo_dsack(sk)) {
2835                         tcp_try_keep_open(sk);
2836                         return;
2837                 }
2838                 break;
2839         case TCP_CA_Loss:
2840                 tcp_process_loss(sk, flag, is_dupack, rexmit);
2841                 if (icsk->icsk_ca_state != TCP_CA_Open &&
2842                     !(flag & FLAG_LOST_RETRANS))
2843                         return;
2844                 /* Change state if cwnd is undone or retransmits are lost */
2845         default:
2846                 if (tcp_is_reno(tp)) {
2847                         if (flag & FLAG_SND_UNA_ADVANCED)
2848                                 tcp_reset_reno_sack(tp);
2849                         if (is_dupack)
2850                                 tcp_add_reno_sack(sk);
2851                 }
2852 
2853                 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2854                         tcp_try_undo_dsack(sk);
2855 
2856                 if (!tcp_time_to_recover(sk, flag)) {
2857                         tcp_try_to_open(sk, flag);
2858                         return;
2859                 }
2860 
2861                 /* MTU probe failure: don't reduce cwnd */
2862                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2863                     icsk->icsk_mtup.probe_size &&
2864                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2865                         tcp_mtup_probe_failed(sk);
2866                         /* Restores the reduction we did in tcp_mtup_probe() */
2867                         tp->snd_cwnd++;
2868                         tcp_simple_retransmit(sk);
2869                         return;
2870                 }
2871 
2872                 /* Otherwise enter Recovery state */
2873                 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2874                 fast_rexmit = 1;
2875         }
2876 
2877         if (do_lost)
2878                 tcp_update_scoreboard(sk, fast_rexmit);
2879         *rexmit = REXMIT_LOST;
2880 }
2881 
2882 /* Kathleen Nichols' algorithm for tracking the minimum value of
2883  * a data stream over some fixed time interval. (E.g., the minimum
2884  * RTT over the past five minutes.) It uses constant space and constant
2885  * time per update yet almost always delivers the same minimum as an
2886  * implementation that has to keep all the data in the window.
2887  *
2888  * The algorithm keeps track of the best, 2nd best & 3rd best min
2889  * values, maintaining an invariant that the measurement time of the
2890  * n'th best >= n-1'th best. It also makes sure that the three values
2891  * are widely separated in the time window since that bounds the worse
2892  * case error when that data is monotonically increasing over the window.
2893  *
2894  * Upon getting a new min, we can forget everything earlier because it
2895  * has no value - the new min is <= everything else in the window by
2896  * definition and it's the most recent. So we restart fresh on every new min
2897  * and overwrites 2nd & 3rd choices. The same property holds for 2nd & 3rd
2898  * best.
2899  */
2900 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2901 {
2902         const u32 now = tcp_time_stamp, wlen = sysctl_tcp_min_rtt_wlen * HZ;
2903         struct rtt_meas *m = tcp_sk(sk)->rtt_min;
2904         struct rtt_meas rttm = {
2905                 .rtt = likely(rtt_us) ? rtt_us : jiffies_to_usecs(1),
2906                 .ts = now,
2907         };
2908         u32 elapsed;
2909 
2910         /* Check if the new measurement updates the 1st, 2nd, or 3rd choices */
2911         if (unlikely(rttm.rtt <= m[0].rtt))
2912                 m[0] = m[1] = m[2] = rttm;
2913         else if (rttm.rtt <= m[1].rtt)
2914                 m[1] = m[2] = rttm;
2915         else if (rttm.rtt <= m[2].rtt)
2916                 m[2] = rttm;
2917 
2918         elapsed = now - m[0].ts;
2919         if (unlikely(elapsed > wlen)) {
2920                 /* Passed entire window without a new min so make 2nd choice
2921                  * the new min & 3rd choice the new 2nd. So forth and so on.
2922                  */
2923                 m[0] = m[1];
2924                 m[1] = m[2];
2925                 m[2] = rttm;
2926                 if (now - m[0].ts > wlen) {
2927                         m[0] = m[1];
2928                         m[1] = rttm;
2929                         if (now - m[0].ts > wlen)
2930                                 m[0] = rttm;
2931                 }
2932         } else if (m[1].ts == m[0].ts && elapsed > wlen / 4) {
2933                 /* Passed a quarter of the window without a new min so
2934                  * take 2nd choice from the 2nd quarter of the window.
2935                  */
2936                 m[2] = m[1] = rttm;
2937         } else if (m[2].ts == m[1].ts && elapsed > wlen / 2) {
2938                 /* Passed half the window without a new min so take the 3rd
2939                  * choice from the last half of the window.
2940                  */
2941                 m[2] = rttm;
2942         }
2943 }
2944 
2945 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2946                                       long seq_rtt_us, long sack_rtt_us,
2947                                       long ca_rtt_us)
2948 {
2949         const struct tcp_sock *tp = tcp_sk(sk);
2950 
2951         /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2952          * broken middle-boxes or peers may corrupt TS-ECR fields. But
2953          * Karn's algorithm forbids taking RTT if some retransmitted data
2954          * is acked (RFC6298).
2955          */
2956         if (seq_rtt_us < 0)
2957                 seq_rtt_us = sack_rtt_us;
2958 
2959         /* RTTM Rule: A TSecr value received in a segment is used to
2960          * update the averaged RTT measurement only if the segment
2961          * acknowledges some new data, i.e., only if it advances the
2962          * left edge of the send window.
2963          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2964          */
2965         if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2966             flag & FLAG_ACKED)
2967                 seq_rtt_us = ca_rtt_us = jiffies_to_usecs(tcp_time_stamp -
2968                                                           tp->rx_opt.rcv_tsecr);
2969         if (seq_rtt_us < 0)
2970                 return false;
2971 
2972         /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2973          * always taken together with ACK, SACK, or TS-opts. Any negative
2974          * values will be skipped with the seq_rtt_us < 0 check above.
2975          */
2976         tcp_update_rtt_min(sk, ca_rtt_us);
2977         tcp_rtt_estimator(sk, seq_rtt_us);
2978         tcp_set_rto(sk);
2979 
2980         /* RFC6298: only reset backoff on valid RTT measurement. */
2981         inet_csk(sk)->icsk_backoff = 0;
2982         return true;
2983 }
2984 
2985 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2986 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2987 {
2988         long rtt_us = -1L;
2989 
2990         if (req && !req->num_retrans && tcp_rsk(req)->snt_synack.v64) {
2991                 struct skb_mstamp now;
2992 
2993                 skb_mstamp_get(&now);
2994                 rtt_us = skb_mstamp_us_delta(&now, &tcp_rsk(req)->snt_synack);
2995         }
2996 
2997         tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us);
2998 }
2999 
3000 
3001 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3002 {
3003         const struct inet_connection_sock *icsk = inet_csk(sk);
3004 
3005         icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3006         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3007 }
3008 
3009 /* Restart timer after forward progress on connection.
3010  * RFC2988 recommends to restart timer to now+rto.
3011  */
3012 void tcp_rearm_rto(struct sock *sk)
3013 {
3014         const struct inet_connection_sock *icsk = inet_csk(sk);
3015         struct tcp_sock *tp = tcp_sk(sk);
3016 
3017         /* If the retrans timer is currently being used by Fast Open
3018          * for SYN-ACK retrans purpose, stay put.
3019          */
3020         if (tp->fastopen_rsk)
3021                 return;
3022 
3023         if (!tp->packets_out) {
3024                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3025         } else {
3026                 u32 rto = inet_csk(sk)->icsk_rto;
3027                 /* Offset the time elapsed after installing regular RTO */
3028                 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3029                     icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3030                         struct sk_buff *skb = tcp_write_queue_head(sk);
3031                         const u32 rto_time_stamp =
3032                                 tcp_skb_timestamp(skb) + rto;
3033                         s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3034                         /* delta may not be positive if the socket is locked
3035                          * when the retrans timer fires and is rescheduled.
3036                          */
3037                         if (delta > 0)
3038                                 rto = delta;
3039                 }
3040                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3041                                           TCP_RTO_MAX);
3042         }
3043 }
3044 
3045 /* This function is called when the delayed ER timer fires. TCP enters
3046  * fast recovery and performs fast-retransmit.
3047  */
3048 void tcp_resume_early_retransmit(struct sock *sk)
3049 {
3050         struct tcp_sock *tp = tcp_sk(sk);
3051 
3052         tcp_rearm_rto(sk);
3053 
3054         /* Stop if ER is disabled after the delayed ER timer is scheduled */
3055         if (!tp->do_early_retrans)
3056                 return;
3057 
3058         tcp_enter_recovery(sk, false);
3059         tcp_update_scoreboard(sk, 1);
3060         tcp_xmit_retransmit_queue(sk);
3061 }
3062 
3063 /* If we get here, the whole TSO packet has not been acked. */
3064 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3065 {
3066         struct tcp_sock *tp = tcp_sk(sk);
3067         u32 packets_acked;
3068 
3069         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3070 
3071         packets_acked = tcp_skb_pcount(skb);
3072         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3073                 return 0;
3074         packets_acked -= tcp_skb_pcount(skb);
3075 
3076         if (packets_acked) {
3077                 BUG_ON(tcp_skb_pcount(skb) == 0);
3078                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3079         }
3080 
3081         return packets_acked;
3082 }
3083 
3084 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3085                            u32 prior_snd_una)
3086 {
3087         const struct skb_shared_info *shinfo;
3088 
3089         /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3090         if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3091                 return;
3092 
3093         shinfo = skb_shinfo(skb);
3094         if (!before(shinfo->tskey, prior_snd_una) &&
3095             before(shinfo->tskey, tcp_sk(sk)->snd_una))
3096                 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3097 }
3098 
3099 /* Remove acknowledged frames from the retransmission queue. If our packet
3100  * is before the ack sequence we can discard it as it's confirmed to have
3101  * arrived at the other end.
3102  */
3103 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3104                                u32 prior_snd_una, int *acked,
3105                                struct tcp_sacktag_state *sack)
3106 {
3107         const struct inet_connection_sock *icsk = inet_csk(sk);
3108         struct skb_mstamp first_ackt, last_ackt, now;
3109         struct tcp_sock *tp = tcp_sk(sk);
3110         u32 prior_sacked = tp->sacked_out;
3111         u32 reord = tp->packets_out;
3112         bool fully_acked = true;
3113         long sack_rtt_us = -1L;
3114         long seq_rtt_us = -1L;
3115         long ca_rtt_us = -1L;
3116         struct sk_buff *skb;
3117         u32 pkts_acked = 0;
3118         bool rtt_update;
3119         int flag = 0;
3120 
3121         first_ackt.v64 = 0;
3122 
3123         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3124                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3125                 u8 sacked = scb->sacked;
3126                 u32 acked_pcount;
3127 
3128                 tcp_ack_tstamp(sk, skb, prior_snd_una);
3129 
3130                 /* Determine how many packets and what bytes were acked, tso and else */
3131                 if (after(scb->end_seq, tp->snd_una)) {
3132                         if (tcp_skb_pcount(skb) == 1 ||
3133                             !after(tp->snd_una, scb->seq))
3134                                 break;
3135 
3136                         acked_pcount = tcp_tso_acked(sk, skb);
3137                         if (!acked_pcount)
3138                                 break;
3139 
3140                         fully_acked = false;
3141                 } else {
3142                         /* Speedup tcp_unlink_write_queue() and next loop */
3143                         prefetchw(skb->next);
3144                         acked_pcount = tcp_skb_pcount(skb);
3145                 }
3146 
3147                 if (unlikely(sacked & TCPCB_RETRANS)) {
3148                         if (sacked & TCPCB_SACKED_RETRANS)
3149                                 tp->retrans_out -= acked_pcount;
3150                         flag |= FLAG_RETRANS_DATA_ACKED;
3151                 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3152                         last_ackt = skb->skb_mstamp;
3153                         WARN_ON_ONCE(last_ackt.v64 == 0);
3154                         if (!first_ackt.v64)
3155                                 first_ackt = last_ackt;
3156 
3157                         reord = min(pkts_acked, reord);
3158                         if (!after(scb->end_seq, tp->high_seq))
3159                                 flag |= FLAG_ORIG_SACK_ACKED;
3160                 }
3161 
3162                 if (sacked & TCPCB_SACKED_ACKED) {
3163                         tp->sacked_out -= acked_pcount;
3164                 } else if (tcp_is_sack(tp)) {
3165                         tp->delivered += acked_pcount;
3166                         if (!tcp_skb_spurious_retrans(tp, skb))
3167                                 tcp_rack_advance(tp, &skb->skb_mstamp, sacked);
3168                 }
3169                 if (sacked & TCPCB_LOST)
3170                         tp->lost_out -= acked_pcount;
3171 
3172                 tp->packets_out -= acked_pcount;
3173                 pkts_acked += acked_pcount;
3174 
3175                 /* Initial outgoing SYN's get put onto the write_queue
3176                  * just like anything else we transmit.  It is not
3177                  * true data, and if we misinform our callers that
3178                  * this ACK acks real data, we will erroneously exit
3179                  * connection startup slow start one packet too
3180                  * quickly.  This is severely frowned upon behavior.
3181                  */
3182                 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3183                         flag |= FLAG_DATA_ACKED;
3184                 } else {
3185                         flag |= FLAG_SYN_ACKED;
3186                         tp->retrans_stamp = 0;
3187                 }
3188 
3189                 if (!fully_acked)
3190                         break;
3191 
3192                 tcp_unlink_write_queue(skb, sk);
3193                 sk_wmem_free_skb(sk, skb);
3194                 if (unlikely(skb == tp->retransmit_skb_hint))
3195                         tp->retransmit_skb_hint = NULL;
3196                 if (unlikely(skb == tp->lost_skb_hint))
3197                         tp->lost_skb_hint = NULL;
3198         }
3199 
3200         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3201                 tp->snd_up = tp->snd_una;
3202 
3203         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3204                 flag |= FLAG_SACK_RENEGING;
3205 
3206         skb_mstamp_get(&now);
3207         if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3208                 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3209                 ca_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3210         }
3211         if (sack->first_sackt.v64) {
3212                 sack_rtt_us = skb_mstamp_us_delta(&now, &sack->first_sackt);
3213                 ca_rtt_us = skb_mstamp_us_delta(&now, &sack->last_sackt);
3214         }
3215 
3216         rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3217                                         ca_rtt_us);
3218 
3219         if (flag & FLAG_ACKED) {
3220                 tcp_rearm_rto(sk);
3221                 if (unlikely(icsk->icsk_mtup.probe_size &&
3222                              !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3223                         tcp_mtup_probe_success(sk);
3224                 }
3225 
3226                 if (tcp_is_reno(tp)) {
3227                         tcp_remove_reno_sacks(sk, pkts_acked);
3228                 } else {
3229                         int delta;
3230 
3231                         /* Non-retransmitted hole got filled? That's reordering */
3232                         if (reord < prior_fackets)
3233                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3234 
3235                         delta = tcp_is_fack(tp) ? pkts_acked :
3236                                                   prior_sacked - tp->sacked_out;
3237                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3238                 }
3239 
3240                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3241 
3242         } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3243                    sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3244                 /* Do not re-arm RTO if the sack RTT is measured from data sent
3245                  * after when the head was last (re)transmitted. Otherwise the
3246                  * timeout may continue to extend in loss recovery.
3247                  */
3248                 tcp_rearm_rto(sk);
3249         }
3250 
3251         if (icsk->icsk_ca_ops->pkts_acked) {
3252                 struct ack_sample sample = { .pkts_acked = pkts_acked,
3253                                              .rtt_us = ca_rtt_us };
3254 
3255                 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3256         }
3257 
3258 #if FASTRETRANS_DEBUG > 0
3259         WARN_ON((int)tp->sacked_out < 0);
3260         WARN_ON((int)tp->lost_out < 0);
3261         WARN_ON((int)tp->retrans_out < 0);
3262         if (!tp->packets_out && tcp_is_sack(tp)) {
3263                 icsk = inet_csk(sk);
3264                 if (tp->lost_out) {
3265                         pr_debug("Leak l=%u %d\n",
3266                                  tp->lost_out, icsk->icsk_ca_state);
3267                         tp->lost_out = 0;
3268                 }
3269                 if (tp->sacked_out) {
3270                         pr_debug("Leak s=%u %d\n",
3271                                  tp->sacked_out, icsk->icsk_ca_state);
3272                         tp->sacked_out = 0;
3273                 }
3274                 if (tp->retrans_out) {
3275                         pr_debug("Leak r=%u %d\n",
3276                                  tp->retrans_out, icsk->icsk_ca_state);
3277                         tp->retrans_out = 0;
3278                 }
3279         }
3280 #endif
3281         *acked = pkts_acked;
3282         return flag;
3283 }
3284 
3285 static void tcp_ack_probe(struct sock *sk)
3286 {
3287         const struct tcp_sock *tp = tcp_sk(sk);
3288         struct inet_connection_sock *icsk = inet_csk(sk);
3289 
3290         /* Was it a usable window open? */
3291 
3292         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3293                 icsk->icsk_backoff = 0;
3294                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3295                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3296                  * This function is not for random using!
3297                  */
3298         } else {
3299                 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3300 
3301                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3302                                           when, TCP_RTO_MAX);
3303         }
3304 }
3305 
3306 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3307 {
3308         return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3309                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3310 }
3311 
3312 /* Decide wheather to run the increase function of congestion control. */
3313 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3314 {
3315         /* If reordering is high then always grow cwnd whenever data is
3316          * delivered regardless of its ordering. Otherwise stay conservative
3317          * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3318          * new SACK or ECE mark may first advance cwnd here and later reduce
3319          * cwnd in tcp_fastretrans_alert() based on more states.
3320          */
3321         if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3322                 return flag & FLAG_FORWARD_PROGRESS;
3323 
3324         return flag & FLAG_DATA_ACKED;
3325 }
3326 
3327 /* The "ultimate" congestion control function that aims to replace the rigid
3328  * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3329  * It's called toward the end of processing an ACK with precise rate
3330  * information. All transmission or retransmission are delayed afterwards.
3331  */
3332 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3333                              int flag)
3334 {
3335         if (tcp_in_cwnd_reduction(sk)) {
3336                 /* Reduce cwnd if state mandates */
3337                 tcp_cwnd_reduction(sk, acked_sacked, flag);
3338         } else if (tcp_may_raise_cwnd(sk, flag)) {
3339                 /* Advance cwnd if state allows */
3340                 tcp_cong_avoid(sk, ack, acked_sacked);
3341         }
3342         tcp_update_pacing_rate(sk);
3343 }
3344 
3345 /* Check that window update is acceptable.
3346  * The function assumes that snd_una<=ack<=snd_next.
3347  */
3348 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3349                                         const u32 ack, const u32 ack_seq,
3350                                         const u32 nwin)
3351 {
3352         return  after(ack, tp->snd_una) ||
3353                 after(ack_seq, tp->snd_wl1) ||
3354                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3355 }
3356 
3357 /* If we update tp->snd_una, also update tp->bytes_acked */
3358 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3359 {
3360         u32 delta = ack - tp->snd_una;
3361 
3362         sock_owned_by_me((struct sock *)tp);
3363         u64_stats_update_begin_raw(&tp->syncp);
3364         tp->bytes_acked += delta;
3365         u64_stats_update_end_raw(&tp->syncp);
3366         tp->snd_una = ack;
3367 }
3368 
3369 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3370 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3371 {
3372         u32 delta = seq - tp->rcv_nxt;
3373 
3374         sock_owned_by_me((struct sock *)tp);
3375         u64_stats_update_begin_raw(&tp->syncp);
3376         tp->bytes_received += delta;
3377         u64_stats_update_end_raw(&tp->syncp);
3378         tp->rcv_nxt = seq;
3379 }
3380 
3381 /* Update our send window.
3382  *
3383  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3384  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3385  */
3386 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3387                                  u32 ack_seq)
3388 {
3389         struct tcp_sock *tp = tcp_sk(sk);
3390         int flag = 0;
3391         u32 nwin = ntohs(tcp_hdr(skb)->window);
3392 
3393         if (likely(!tcp_hdr(skb)->syn))
3394                 nwin <<= tp->rx_opt.snd_wscale;
3395 
3396         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3397                 flag |= FLAG_WIN_UPDATE;
3398                 tcp_update_wl(tp, ack_seq);
3399 
3400                 if (tp->snd_wnd != nwin) {
3401                         tp->snd_wnd = nwin;
3402 
3403                         /* Note, it is the only place, where
3404                          * fast path is recovered for sending TCP.
3405                          */
3406                         tp->pred_flags = 0;
3407                         tcp_fast_path_check(sk);
3408 
3409                         if (tcp_send_head(sk))
3410                                 tcp_slow_start_after_idle_check(sk);
3411 
3412                         if (nwin > tp->max_window) {
3413                                 tp->max_window = nwin;
3414                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3415                         }
3416                 }
3417         }
3418 
3419         tcp_snd_una_update(tp, ack);
3420 
3421         return flag;
3422 }
3423 
3424 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3425                                    u32 *last_oow_ack_time)
3426 {
3427         if (*last_oow_ack_time) {
3428                 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3429 
3430                 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3431                         NET_INC_STATS(net, mib_idx);
3432                         return true;    /* rate-limited: don't send yet! */
3433                 }
3434         }
3435 
3436         *last_oow_ack_time = tcp_time_stamp;
3437 
3438         return false;   /* not rate-limited: go ahead, send dupack now! */
3439 }
3440 
3441 /* Return true if we're currently rate-limiting out-of-window ACKs and
3442  * thus shouldn't send a dupack right now. We rate-limit dupacks in
3443  * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3444  * attacks that send repeated SYNs or ACKs for the same connection. To
3445  * do this, we do not send a duplicate SYNACK or ACK if the remote
3446  * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3447  */
3448 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3449                           int mib_idx, u32 *last_oow_ack_time)
3450 {
3451         /* Data packets without SYNs are not likely part of an ACK loop. */
3452         if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3453             !tcp_hdr(skb)->syn)
3454                 return false;
3455 
3456         return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3457 }
3458 
3459 /* RFC 5961 7 [ACK Throttling] */
3460 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3461 {
3462         /* unprotected vars, we dont care of overwrites */
3463         static u32 challenge_timestamp;
3464         static unsigned int challenge_count;
3465         struct tcp_sock *tp = tcp_sk(sk);
3466         u32 count, now;
3467 
3468         /* First check our per-socket dupack rate limit. */
3469         if (__tcp_oow_rate_limited(sock_net(sk),
3470                                    LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3471                                    &tp->last_oow_ack_time))
3472                 return;
3473 
3474         /* Then check host-wide RFC 5961 rate limit. */
3475         now = jiffies / HZ;
3476         if (now != challenge_timestamp) {
3477                 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3478 
3479                 challenge_timestamp = now;
3480                 WRITE_ONCE(challenge_count, half +
3481                            prandom_u32_max(sysctl_tcp_challenge_ack_limit));
3482         }
3483         count = READ_ONCE(challenge_count);
3484         if (count > 0) {
3485                 WRITE_ONCE(challenge_count, count - 1);
3486                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3487                 tcp_send_ack(sk);
3488         }
3489 }
3490 
3491 static void tcp_store_ts_recent(struct tcp_sock *tp)
3492 {
3493         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3494         tp->rx_opt.ts_recent_stamp = get_seconds();
3495 }
3496 
3497 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3498 {
3499         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3500                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3501                  * extra check below makes sure this can only happen
3502                  * for pure ACK frames.  -DaveM
3503                  *
3504                  * Not only, also it occurs for expired timestamps.
3505                  */
3506 
3507                 if (tcp_paws_check(&tp->rx_opt, 0))
3508                         tcp_store_ts_recent(tp);
3509         }
3510 }
3511 
3512 /* This routine deals with acks during a TLP episode.
3513  * We mark the end of a TLP episode on receiving TLP dupack or when
3514  * ack is after tlp_high_seq.
3515  * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3516  */
3517 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3518 {
3519         struct tcp_sock *tp = tcp_sk(sk);
3520 
3521         if (before(ack, tp->tlp_high_seq))
3522                 return;
3523 
3524         if (flag & FLAG_DSACKING_ACK) {
3525                 /* This DSACK means original and TLP probe arrived; no loss */
3526                 tp->tlp_high_seq = 0;
3527         } else if (after(ack, tp->tlp_high_seq)) {
3528                 /* ACK advances: there was a loss, so reduce cwnd. Reset
3529                  * tlp_high_seq in tcp_init_cwnd_reduction()
3530                  */
3531                 tcp_init_cwnd_reduction(sk);
3532                 tcp_set_ca_state(sk, TCP_CA_CWR);
3533                 tcp_end_cwnd_reduction(sk);
3534                 tcp_try_keep_open(sk);
3535                 NET_INC_STATS(sock_net(sk),
3536                                 LINUX_MIB_TCPLOSSPROBERECOVERY);
3537         } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3538                              FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3539                 /* Pure dupack: original and TLP probe arrived; no loss */
3540                 tp->tlp_high_seq = 0;
3541         }
3542 }
3543 
3544 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3545 {
3546         const struct inet_connection_sock *icsk = inet_csk(sk);
3547 
3548         if (icsk->icsk_ca_ops->in_ack_event)
3549                 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3550 }
3551 
3552 /* Congestion control has updated the cwnd already. So if we're in
3553  * loss recovery then now we do any new sends (for FRTO) or
3554  * retransmits (for CA_Loss or CA_recovery) that make sense.
3555  */
3556 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3557 {
3558         struct tcp_sock *tp = tcp_sk(sk);
3559 
3560         if (rexmit == REXMIT_NONE)
3561                 return;
3562 
3563         if (unlikely(rexmit == 2)) {
3564                 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3565                                           TCP_NAGLE_OFF);
3566                 if (after(tp->snd_nxt, tp->high_seq))
3567                         return;
3568                 tp->frto = 0;
3569         }
3570         tcp_xmit_retransmit_queue(sk);
3571 }
3572 
3573 /* This routine deals with incoming acks, but not outgoing ones. */
3574 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3575 {
3576         struct inet_connection_sock *icsk = inet_csk(sk);
3577         struct tcp_sock *tp = tcp_sk(sk);
3578         struct tcp_sacktag_state sack_state;
3579         u32 prior_snd_una = tp->snd_una;
3580         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3581         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3582         bool is_dupack = false;
3583         u32 prior_fackets;
3584         int prior_packets = tp->packets_out;
3585         u32 prior_delivered = tp->delivered;
3586         int acked = 0; /* Number of packets newly acked */
3587         int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3588 
3589         sack_state.first_sackt.v64 = 0;
3590 
3591         /* We very likely will need to access write queue head. */
3592         prefetchw(sk->sk_write_queue.next);
3593 
3594         /* If the ack is older than previous acks
3595          * then we can probably ignore it.
3596          */
3597         if (before(ack, prior_snd_una)) {
3598                 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3599                 if (before(ack, prior_snd_una - tp->max_window)) {
3600                         tcp_send_challenge_ack(sk, skb);
3601                         return -1;
3602                 }
3603                 goto old_ack;
3604         }
3605 
3606         /* If the ack includes data we haven't sent yet, discard
3607          * this segment (RFC793 Section 3.9).
3608          */
3609         if (after(ack, tp->snd_nxt))
3610                 goto invalid_ack;
3611 
3612         if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3613             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3614                 tcp_rearm_rto(sk);
3615 
3616         if (after(ack, prior_snd_una)) {
3617                 flag |= FLAG_SND_UNA_ADVANCED;
3618                 icsk->icsk_retransmits = 0;
3619         }
3620 
3621         prior_fackets = tp->fackets_out;
3622 
3623         /* ts_recent update must be made after we are sure that the packet
3624          * is in window.
3625          */
3626         if (flag & FLAG_UPDATE_TS_RECENT)
3627                 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3628 
3629         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3630                 /* Window is constant, pure forward advance.
3631                  * No more checks are required.
3632                  * Note, we use the fact that SND.UNA>=SND.WL2.
3633                  */
3634                 tcp_update_wl(tp, ack_seq);
3635                 tcp_snd_una_update(tp, ack);
3636                 flag |= FLAG_WIN_UPDATE;
3637 
3638                 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3639 
3640                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3641         } else {
3642                 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3643 
3644                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3645                         flag |= FLAG_DATA;
3646                 else
3647                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3648 
3649                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3650 
3651                 if (TCP_SKB_CB(skb)->sacked)
3652                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3653                                                         &sack_state);
3654 
3655                 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3656                         flag |= FLAG_ECE;
3657                         ack_ev_flags |= CA_ACK_ECE;
3658                 }
3659 
3660                 if (flag & FLAG_WIN_UPDATE)
3661                         ack_ev_flags |= CA_ACK_WIN_UPDATE;
3662 
3663                 tcp_in_ack_event(sk, ack_ev_flags);
3664         }
3665 
3666         /* We passed data and got it acked, remove any soft error
3667          * log. Something worked...
3668          */
3669         sk->sk_err_soft = 0;
3670         icsk->icsk_probes_out = 0;
3671         tp->rcv_tstamp = tcp_time_stamp;
3672         if (!prior_packets)
3673                 goto no_queue;
3674 
3675         /* See if we can take anything off of the retransmit queue. */
3676         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, &acked,
3677                                     &sack_state);
3678 
3679         if (tcp_ack_is_dubious(sk, flag)) {
3680                 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3681                 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3682         }
3683         if (tp->tlp_high_seq)
3684                 tcp_process_tlp_ack(sk, ack, flag);
3685 
3686         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3687                 struct dst_entry *dst = __sk_dst_get(sk);
3688                 if (dst)
3689                         dst_confirm(dst);
3690         }
3691 
3692         if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3693                 tcp_schedule_loss_probe(sk);
3694         tcp_cong_control(sk, ack, tp->delivered - prior_delivered, flag);
3695         tcp_xmit_recovery(sk, rexmit);
3696         return 1;
3697 
3698 no_queue:
3699         /* If data was DSACKed, see if we can undo a cwnd reduction. */
3700         if (flag & FLAG_DSACKING_ACK)
3701                 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3702         /* If this ack opens up a zero window, clear backoff.  It was
3703          * being used to time the probes, and is probably far higher than
3704          * it needs to be for normal retransmission.
3705          */
3706         if (tcp_send_head(sk))
3707                 tcp_ack_probe(sk);
3708 
3709         if (tp->tlp_high_seq)
3710                 tcp_process_tlp_ack(sk, ack, flag);
3711         return 1;
3712 
3713 invalid_ack:
3714         SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3715         return -1;
3716 
3717 old_ack:
3718         /* If data was SACKed, tag it and see if we should send more data.
3719          * If data was DSACKed, see if we can undo a cwnd reduction.
3720          */
3721         if (TCP_SKB_CB(skb)->sacked) {
3722                 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3723                                                 &sack_state);
3724                 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3725                 tcp_xmit_recovery(sk, rexmit);
3726         }
3727 
3728         SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3729         return 0;
3730 }
3731 
3732 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3733                                       bool syn, struct tcp_fastopen_cookie *foc,
3734                                       bool exp_opt)
3735 {
3736         /* Valid only in SYN or SYN-ACK with an even length.  */
3737         if (!foc || !syn || len < 0 || (len & 1))
3738                 return;
3739 
3740         if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3741             len <= TCP_FASTOPEN_COOKIE_MAX)
3742                 memcpy(foc->val, cookie, len);
3743         else if (len != 0)
3744                 len = -1;
3745         foc->len = len;
3746         foc->exp = exp_opt;
3747 }
3748 
3749 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3750  * But, this can also be called on packets in the established flow when
3751  * the fast version below fails.
3752  */
3753 void tcp_parse_options(const struct sk_buff *skb,
3754                        struct tcp_options_received *opt_rx, int estab,
3755                        struct tcp_fastopen_cookie *foc)
3756 {
3757         const unsigned char *ptr;
3758         const struct tcphdr *th = tcp_hdr(skb);
3759         int length = (th->doff * 4) - sizeof(struct tcphdr);
3760 
3761         ptr = (const unsigned char *)(th + 1);
3762         opt_rx->saw_tstamp = 0;
3763 
3764         while (length > 0) {
3765                 int opcode = *ptr++;
3766                 int opsize;
3767 
3768                 switch (opcode) {
3769                 case TCPOPT_EOL:
3770                         return;
3771                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3772                         length--;
3773                         continue;
3774                 default:
3775                         opsize = *ptr++;
3776                         if (opsize < 2) /* "silly options" */
3777                                 return;
3778                         if (opsize > length)
3779                                 return; /* don't parse partial options */
3780                         switch (opcode) {
3781                         case TCPOPT_MSS:
3782                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3783                                         u16 in_mss = get_unaligned_be16(ptr);
3784                                         if (in_mss) {
3785                                                 if (opt_rx->user_mss &&
3786                                                     opt_rx->user_mss < in_mss)
3787                                                         in_mss = opt_rx->user_mss;
3788                                                 opt_rx->mss_clamp = in_mss;
3789                                         }
3790                                 }
3791                                 break;
3792                         case TCPOPT_WINDOW:
3793                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3794                                     !estab && sysctl_tcp_window_scaling) {
3795                                         __u8 snd_wscale = *(__u8 *)ptr;
3796                                         opt_rx->wscale_ok = 1;
3797                                         if (snd_wscale > 14) {
3798                                                 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3799                                                                      __func__,
3800                                                                      snd_wscale);
3801                                                 snd_wscale = 14;
3802                                         }
3803                                         opt_rx->snd_wscale = snd_wscale;
3804                                 }
3805                                 break;
3806                         case TCPOPT_TIMESTAMP:
3807                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3808                                     ((estab && opt_rx->tstamp_ok) ||
3809                                      (!estab && sysctl_tcp_timestamps))) {
3810                                         opt_rx->saw_tstamp = 1;
3811                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3812                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3813                                 }
3814                                 break;
3815                         case TCPOPT_SACK_PERM:
3816                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3817                                     !estab && sysctl_tcp_sack) {
3818                                         opt_rx->sack_ok = TCP_SACK_SEEN;
3819                                         tcp_sack_reset(opt_rx);
3820                                 }
3821                                 break;
3822 
3823                         case TCPOPT_SACK:
3824                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3825                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3826                                    opt_rx->sack_ok) {
3827                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3828                                 }
3829                                 break;
3830 #ifdef CONFIG_TCP_MD5SIG
3831                         case TCPOPT_MD5SIG:
3832                                 /*
3833                                  * The MD5 Hash has already been
3834                                  * checked (see tcp_v{4,6}_do_rcv()).
3835                                  */
3836                                 break;
3837 #endif
3838                         case TCPOPT_FASTOPEN:
3839                                 tcp_parse_fastopen_option(
3840                                         opsize - TCPOLEN_FASTOPEN_BASE,
3841                                         ptr, th->syn, foc, false);
3842                                 break;
3843 
3844                         case TCPOPT_EXP:
3845                                 /* Fast Open option shares code 254 using a
3846                                  * 16 bits magic number.
3847                                  */
3848                                 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3849                                     get_unaligned_be16(ptr) ==
3850                                     TCPOPT_FASTOPEN_MAGIC)
3851                                         tcp_parse_fastopen_option(opsize -
3852                                                 TCPOLEN_EXP_FASTOPEN_BASE,
3853                                                 ptr + 2, th->syn, foc, true);
3854                                 break;
3855 
3856                         }
3857                         ptr += opsize-2;
3858                         length -= opsize;
3859                 }
3860         }
3861 }
3862 EXPORT_SYMBOL(tcp_parse_options);
3863 
3864 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3865 {
3866         const __be32 *ptr = (const __be32 *)(th + 1);
3867 
3868         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3869                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3870                 tp->rx_opt.saw_tstamp = 1;
3871                 ++ptr;
3872                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3873                 ++ptr;
3874                 if (*ptr)
3875                         tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3876                 else
3877                         tp->rx_opt.rcv_tsecr = 0;
3878                 return true;
3879         }
3880         return false;
3881 }
3882 
3883 /* Fast parse options. This hopes to only see timestamps.
3884  * If it is wrong it falls back on tcp_parse_options().
3885  */
3886 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3887                                    const struct tcphdr *th, struct tcp_sock *tp)
3888 {
3889         /* In the spirit of fast parsing, compare doff directly to constant
3890          * values.  Because equality is used, short doff can be ignored here.
3891          */
3892         if (th->doff == (sizeof(*th) / 4)) {
3893                 tp->rx_opt.saw_tstamp = 0;
3894                 return false;
3895         } else if (tp->rx_opt.tstamp_ok &&
3896                    th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3897                 if (tcp_parse_aligned_timestamp(tp, th))
3898                         return true;
3899         }
3900 
3901         tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3902         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3903                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3904 
3905         return true;
3906 }
3907 
3908 #ifdef CONFIG_TCP_MD5SIG
3909 /*
3910  * Parse MD5 Signature option
3911  */
3912 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3913 {
3914         int length = (th->doff << 2) - sizeof(*th);
3915         const u8 *ptr = (const u8 *)(th + 1);
3916 
3917         /* If the TCP option is too short, we can short cut */
3918         if (length < TCPOLEN_MD5SIG)
3919                 return NULL;
3920 
3921         while (length > 0) {
3922                 int opcode = *ptr++;
3923                 int opsize;
3924 
3925                 switch (opcode) {
3926                 case TCPOPT_EOL:
3927                         return NULL;
3928                 case TCPOPT_NOP:
3929                         length--;
3930                         continue;
3931                 default:
3932                         opsize = *ptr++;
3933                         if (opsize < 2 || opsize > length)
3934                                 return NULL;
3935                         if (opcode == TCPOPT_MD5SIG)
3936                                 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3937                 }
3938                 ptr += opsize - 2;
3939                 length -= opsize;
3940         }
3941         return NULL;
3942 }
3943 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3944 #endif
3945 
3946 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3947  *
3948  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3949  * it can pass through stack. So, the following predicate verifies that
3950  * this segment is not used for anything but congestion avoidance or
3951  * fast retransmit. Moreover, we even are able to eliminate most of such
3952  * second order effects, if we apply some small "replay" window (~RTO)
3953  * to timestamp space.
3954  *
3955  * All these measures still do not guarantee that we reject wrapped ACKs
3956  * on networks with high bandwidth, when sequence space is recycled fastly,
3957  * but it guarantees that such events will be very rare and do not affect
3958  * connection seriously. This doesn't look nice, but alas, PAWS is really
3959  * buggy extension.
3960  *
3961  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3962  * states that events when retransmit arrives after original data are rare.
3963  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3964  * the biggest problem on large power networks even with minor reordering.
3965  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3966  * up to bandwidth of 18Gigabit/sec. 8) ]
3967  */
3968 
3969 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3970 {
3971         const struct tcp_sock *tp = tcp_sk(sk);
3972         const struct tcphdr *th = tcp_hdr(skb);
3973         u32 seq = TCP_SKB_CB(skb)->seq;
3974         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3975 
3976         return (/* 1. Pure ACK with correct sequence number. */
3977                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3978 
3979                 /* 2. ... and duplicate ACK. */
3980                 ack == tp->snd_una &&
3981 
3982                 /* 3. ... and does not update window. */
3983                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3984 
3985                 /* 4. ... and sits in replay window. */
3986                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3987 }
3988 
3989 static inline bool tcp_paws_discard(const struct sock *sk,
3990                                    const struct sk_buff *skb)
3991 {
3992         const struct tcp_sock *tp = tcp_sk(sk);
3993 
3994         return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3995                !tcp_disordered_ack(sk, skb);
3996 }
3997 
3998 /* Check segment sequence number for validity.
3999  *
4000  * Segment controls are considered valid, if the segment
4001  * fits to the window after truncation to the window. Acceptability
4002  * of data (and SYN, FIN, of course) is checked separately.
4003  * See tcp_data_queue(), for example.
4004  *
4005  * Also, controls (RST is main one) are accepted using RCV.WUP instead
4006  * of RCV.NXT. Peer still did not advance his SND.UNA when we
4007  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4008  * (borrowed from freebsd)
4009  */
4010 
4011 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4012 {
4013         return  !before(end_seq, tp->rcv_wup) &&
4014                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4015 }
4016 
4017 /* When we get a reset we do this. */
4018 void tcp_reset(struct sock *sk)
4019 {
4020         /* We want the right error as BSD sees it (and indeed as we do). */
4021         switch (sk->sk_state) {
4022         case TCP_SYN_SENT:
4023                 sk->sk_err = ECONNREFUSED;
4024                 break;
4025         case TCP_CLOSE_WAIT:
4026                 sk->sk_err = EPIPE;
4027                 break;
4028         case TCP_CLOSE:
4029                 return;
4030         default:
4031                 sk->sk_err = ECONNRESET;
4032         }
4033         /* This barrier is coupled with smp_rmb() in tcp_poll() */
4034         smp_wmb();
4035 
4036         if (!sock_flag(sk, SOCK_DEAD))
4037                 sk->sk_error_report(sk);
4038 
4039         tcp_done(sk);
4040 }
4041 
4042 /*
4043  *      Process the FIN bit. This now behaves as it is supposed to work
4044  *      and the FIN takes effect when it is validly part of sequence
4045  *      space. Not before when we get holes.
4046  *
4047  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4048  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
4049  *      TIME-WAIT)
4050  *
4051  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
4052  *      close and we go into CLOSING (and later onto TIME-WAIT)
4053  *
4054  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4055  */
4056 void tcp_fin(struct sock *sk)
4057 {
4058         struct tcp_sock *tp = tcp_sk(sk);
4059 
4060         inet_csk_schedule_ack(sk);
4061 
4062         sk->sk_shutdown |= RCV_SHUTDOWN;
4063         sock_set_flag(sk, SOCK_DONE);
4064 
4065         switch (sk->sk_state) {
4066         case TCP_SYN_RECV:
4067         case TCP_ESTABLISHED:
4068                 /* Move to CLOSE_WAIT */
4069                 tcp_set_state(sk, TCP_CLOSE_WAIT);
4070                 inet_csk(sk)->icsk_ack.pingpong = 1;
4071                 break;
4072 
4073         case TCP_CLOSE_WAIT:
4074         case TCP_CLOSING:
4075                 /* Received a retransmission of the FIN, do
4076                  * nothing.
4077                  */
4078                 break;
4079         case TCP_LAST_ACK:
4080                 /* RFC793: Remain in the LAST-ACK state. */
4081                 break;
4082 
4083         case TCP_FIN_WAIT1:
4084                 /* This case occurs when a simultaneous close
4085                  * happens, we must ack the received FIN and
4086                  * enter the CLOSING state.
4087                  */
4088                 tcp_send_ack(sk);
4089                 tcp_set_state(sk, TCP_CLOSING);
4090                 break;
4091         case TCP_FIN_WAIT2:
4092                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4093                 tcp_send_ack(sk);
4094                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4095                 break;
4096         default:
4097                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4098                  * cases we should never reach this piece of code.
4099                  */
4100                 pr_err("%s: Impossible, sk->sk_state=%d\n",
4101                        __func__, sk->sk_state);
4102                 break;
4103         }
4104 
4105         /* It _is_ possible, that we have something out-of-order _after_ FIN.
4106          * Probably, we should reset in this case. For now drop them.
4107          */
4108         __skb_queue_purge(&tp->out_of_order_queue);
4109         if (tcp_is_sack(tp))
4110                 tcp_sack_reset(&tp->rx_opt);
4111         sk_mem_reclaim(sk);
4112 
4113         if (!sock_flag(sk, SOCK_DEAD)) {
4114                 sk->sk_state_change(sk);
4115 
4116                 /* Do not send POLL_HUP for half duplex close. */
4117                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4118                     sk->sk_state == TCP_CLOSE)
4119                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4120                 else
4121                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4122         }
4123 }
4124 
4125 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4126                                   u32 end_seq)
4127 {
4128         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4129                 if (before(seq, sp->start_seq))
4130                         sp->start_seq = seq;
4131                 if (after(end_seq, sp->end_seq))
4132                         sp->end_seq = end_seq;
4133                 return true;
4134         }
4135         return false;
4136 }
4137 
4138 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4139 {
4140         struct tcp_sock *tp = tcp_sk(sk);
4141 
4142         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4143                 int mib_idx;
4144 
4145                 if (before(seq, tp->rcv_nxt))
4146                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4147                 else
4148                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4149 
4150                 NET_INC_STATS(sock_net(sk), mib_idx);
4151 
4152                 tp->rx_opt.dsack = 1;
4153                 tp->duplicate_sack[0].start_seq = seq;
4154                 tp->duplicate_sack[0].end_seq = end_seq;
4155         }
4156 }
4157 
4158 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4159 {
4160         struct tcp_sock *tp = tcp_sk(sk);
4161 
4162         if (!tp->rx_opt.dsack)
4163                 tcp_dsack_set(sk, seq, end_seq);
4164         else
4165                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4166 }
4167 
4168 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4169 {
4170         struct tcp_sock *tp = tcp_sk(sk);
4171 
4172         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4173             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4174                 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4175                 tcp_enter_quickack_mode(sk);
4176 
4177                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4178                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4179 
4180                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4181                                 end_seq = tp->rcv_nxt;
4182                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4183                 }
4184         }
4185 
4186         tcp_send_ack(sk);
4187 }
4188 
4189 /* These routines update the SACK block as out-of-order packets arrive or
4190  * in-order packets close up the sequence space.
4191  */
4192 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4193 {
4194         int this_sack;
4195         struct tcp_sack_block *sp = &tp->selective_acks[0];
4196         struct tcp_sack_block *swalk = sp + 1;
4197 
4198         /* See if the recent change to the first SACK eats into
4199          * or hits the sequence space of other SACK blocks, if so coalesce.
4200          */
4201         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4202                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4203                         int i;
4204 
4205                         /* Zap SWALK, by moving every further SACK up by one slot.
4206                          * Decrease num_sacks.
4207                          */
4208                         tp->rx_opt.num_sacks--;
4209                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4210                                 sp[i] = sp[i + 1];
4211                         continue;
4212                 }
4213                 this_sack++, swalk++;
4214         }
4215 }
4216 
4217 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4218 {
4219         struct tcp_sock *tp = tcp_sk(sk);
4220         struct tcp_sack_block *sp = &tp->selective_acks[0];
4221         int cur_sacks = tp->rx_opt.num_sacks;
4222         int this_sack;
4223 
4224         if (!cur_sacks)
4225                 goto new_sack;
4226 
4227         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4228                 if (tcp_sack_extend(sp, seq, end_seq)) {
4229                         /* Rotate this_sack to the first one. */
4230                         for (; this_sack > 0; this_sack--, sp--)
4231                                 swap(*sp, *(sp - 1));
4232                         if (cur_sacks > 1)
4233                                 tcp_sack_maybe_coalesce(tp);
4234                         return;
4235                 }
4236         }
4237 
4238         /* Could not find an adjacent existing SACK, build a new one,
4239          * put it at the front, and shift everyone else down.  We
4240          * always know there is at least one SACK present already here.
4241          *
4242          * If the sack array is full, forget about the last one.
4243          */
4244         if (this_sack >= TCP_NUM_SACKS) {
4245                 this_sack--;
4246                 tp->rx_opt.num_sacks--;
4247                 sp--;
4248         }
4249         for (; this_sack > 0; this_sack--, sp--)
4250                 *sp = *(sp - 1);
4251 
4252 new_sack:
4253         /* Build the new head SACK, and we're done. */
4254         sp->start_seq = seq;
4255         sp->end_seq = end_seq;
4256         tp->rx_opt.num_sacks++;
4257 }
4258 
4259 /* RCV.NXT advances, some SACKs should be eaten. */
4260 
4261 static void tcp_sack_remove(struct tcp_sock *tp)
4262 {
4263         struct tcp_sack_block *sp = &tp->selective_acks[0];
4264         int num_sacks = tp->rx_opt.num_sacks;
4265         int this_sack;
4266 
4267         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4268         if (skb_queue_empty(&tp->out_of_order_queue)) {
4269                 tp->rx_opt.num_sacks = 0;
4270                 return;
4271         }
4272 
4273         for (this_sack = 0; this_sack < num_sacks;) {
4274                 /* Check if the start of the sack is covered by RCV.NXT. */
4275                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4276                         int i;
4277 
4278                         /* RCV.NXT must cover all the block! */
4279                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4280 
4281                         /* Zap this SACK, by moving forward any other SACKS. */
4282                         for (i = this_sack+1; i < num_sacks; i++)
4283                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4284                         num_sacks--;
4285                         continue;
4286                 }
4287                 this_sack++;
4288                 sp++;
4289         }
4290         tp->rx_opt.num_sacks = num_sacks;
4291 }
4292 
4293 /**
4294  * tcp_try_coalesce - try to merge skb to prior one
4295  * @sk: socket
4296  * @to: prior buffer
4297  * @from: buffer to add in queue
4298  * @fragstolen: pointer to boolean
4299  *
4300  * Before queueing skb @from after @to, try to merge them
4301  * to reduce overall memory use and queue lengths, if cost is small.
4302  * Packets in ofo or receive queues can stay a long time.
4303  * Better try to coalesce them right now to avoid future collapses.
4304  * Returns true if caller should free @from instead of queueing it
4305  */
4306 static bool tcp_try_coalesce(struct sock *sk,
4307                              struct sk_buff *to,
4308                              struct sk_buff *from,
4309                              bool *fragstolen)
4310 {
4311         int delta;
4312 
4313         *fragstolen = false;
4314 
4315         /* Its possible this segment overlaps with prior segment in queue */
4316         if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4317                 return false;
4318 
4319         if (!skb_try_coalesce(to, from, fragstolen, &delta))
4320                 return false;
4321 
4322         atomic_add(delta, &sk->sk_rmem_alloc);
4323         sk_mem_charge(sk, delta);
4324         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4325         TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4326         TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4327         TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4328         return true;
4329 }
4330 
4331 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4332 {
4333         sk_drops_add(sk, skb);
4334         __kfree_skb(skb);
4335 }
4336 
4337 /* This one checks to see if we can put data from the
4338  * out_of_order queue into the receive_queue.
4339  */
4340 static void tcp_ofo_queue(struct sock *sk)
4341 {
4342         struct tcp_sock *tp = tcp_sk(sk);
4343         __u32 dsack_high = tp->rcv_nxt;
4344         struct sk_buff *skb, *tail;
4345         bool fragstolen, eaten;
4346 
4347         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4348                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4349                         break;
4350 
4351                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4352                         __u32 dsack = dsack_high;
4353                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4354                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4355                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4356                 }
4357 
4358                 __skb_unlink(skb, &tp->out_of_order_queue);
4359                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4360                         SOCK_DEBUG(sk, "ofo packet was already received\n");
4361                         tcp_drop(sk, skb);
4362                         continue;
4363                 }
4364                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4365                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4366                            TCP_SKB_CB(skb)->end_seq);
4367 
4368                 tail = skb_peek_tail(&sk->sk_receive_queue);
4369                 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4370                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4371                 if (!eaten)
4372                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4373                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4374                         tcp_fin(sk);
4375                 if (eaten)
4376                         kfree_skb_partial(skb, fragstolen);
4377         }
4378 }
4379 
4380 static bool tcp_prune_ofo_queue(struct sock *sk);
4381 static int tcp_prune_queue(struct sock *sk);
4382 
4383 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4384                                  unsigned int size)
4385 {
4386         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4387             !sk_rmem_schedule(sk, skb, size)) {
4388 
4389                 if (tcp_prune_queue(sk) < 0)
4390                         return -1;
4391 
4392                 if (!sk_rmem_schedule(sk, skb, size)) {
4393                         if (!tcp_prune_ofo_queue(sk))
4394                                 return -1;
4395 
4396                         if (!sk_rmem_schedule(sk, skb, size))
4397                                 return -1;
4398                 }
4399         }
4400         return 0;
4401 }
4402 
4403 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4404 {
4405         struct tcp_sock *tp = tcp_sk(sk);
4406         struct sk_buff *skb1;
4407         u32 seq, end_seq;
4408 
4409         tcp_ecn_check_ce(tp, skb);
4410 
4411         if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4412                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4413                 tcp_drop(sk, skb);
4414                 return;
4415         }
4416 
4417         /* Disable header prediction. */
4418         tp->pred_flags = 0;
4419         inet_csk_schedule_ack(sk);
4420 
4421         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4422         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4423                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4424 
4425         skb1 = skb_peek_tail(&tp->out_of_order_queue);
4426         if (!skb1) {
4427                 /* Initial out of order segment, build 1 SACK. */
4428                 if (tcp_is_sack(tp)) {
4429                         tp->rx_opt.num_sacks = 1;
4430                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4431                         tp->selective_acks[0].end_seq =
4432                                                 TCP_SKB_CB(skb)->end_seq;
4433                 }
4434                 __skb_queue_head(&tp->out_of_order_queue, skb);
4435                 goto end;
4436         }
4437 
4438         seq = TCP_SKB_CB(skb)->seq;
4439         end_seq = TCP_SKB_CB(skb)->end_seq;
4440 
4441         if (seq == TCP_SKB_CB(skb1)->end_seq) {
4442                 bool fragstolen;
4443 
4444                 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4445                         __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4446                 } else {
4447                         tcp_grow_window(sk, skb);
4448                         kfree_skb_partial(skb, fragstolen);
4449                         skb = NULL;
4450                 }
4451 
4452                 if (!tp->rx_opt.num_sacks ||
4453                     tp->selective_acks[0].end_seq != seq)
4454                         goto add_sack;
4455 
4456                 /* Common case: data arrive in order after hole. */
4457                 tp->selective_acks[0].end_seq = end_seq;
4458                 goto end;
4459         }
4460 
4461         /* Find place to insert this segment. */
4462         while (1) {
4463                 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4464                         break;
4465                 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4466                         skb1 = NULL;
4467                         break;
4468                 }
4469                 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4470         }
4471 
4472         /* Do skb overlap to previous one? */
4473         if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4474                 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4475                         /* All the bits are present. Drop. */
4476                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4477                         tcp_drop(sk, skb);
4478                         skb = NULL;
4479                         tcp_dsack_set(sk, seq, end_seq);
4480                         goto add_sack;
4481                 }
4482                 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4483                         /* Partial overlap. */
4484                         tcp_dsack_set(sk, seq,
4485                                       TCP_SKB_CB(skb1)->end_seq);
4486                 } else {
4487                         if (skb_queue_is_first(&tp->out_of_order_queue,
4488                                                skb1))
4489                                 skb1 = NULL;
4490                         else
4491                                 skb1 = skb_queue_prev(
4492                                         &tp->out_of_order_queue,
4493                                         skb1);
4494                 }
4495         }
4496         if (!skb1)
4497                 __skb_queue_head(&tp->out_of_order_queue, skb);
4498         else
4499                 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4500 
4501         /* And clean segments covered by new one as whole. */
4502         while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4503                 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4504 
4505                 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4506                         break;
4507                 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4508                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4509                                          end_seq);
4510                         break;
4511                 }
4512                 __skb_unlink(skb1, &tp->out_of_order_queue);
4513                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4514                                  TCP_SKB_CB(skb1)->end_seq);
4515                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4516                 tcp_drop(sk, skb1);
4517         }
4518 
4519 add_sack:
4520         if (tcp_is_sack(tp))
4521                 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4522 end:
4523         if (skb) {
4524                 tcp_grow_window(sk, skb);
4525                 skb_set_owner_r(skb, sk);
4526         }
4527 }
4528 
4529 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4530                   bool *fragstolen)
4531 {
4532         int eaten;
4533         struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4534 
4535         __skb_pull(skb, hdrlen);
4536         eaten = (tail &&
4537                  tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4538         tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4539         if (!eaten) {
4540                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4541                 skb_set_owner_r(skb, sk);
4542         }
4543         return eaten;
4544 }
4545 
4546 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4547 {
4548         struct sk_buff *skb;
4549         int err = -ENOMEM;
4550         int data_len = 0;
4551         bool fragstolen;
4552 
4553         if (size == 0)
4554                 return 0;
4555 
4556         if (size > PAGE_SIZE) {
4557                 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4558 
4559                 data_len = npages << PAGE_SHIFT;
4560                 size = data_len + (size & ~PAGE_MASK);
4561         }
4562         skb = alloc_skb_with_frags(size - data_len, data_len,
4563                                    PAGE_ALLOC_COSTLY_ORDER,
4564                                    &err, sk->sk_allocation);
4565         if (!skb)
4566                 goto err;
4567 
4568         skb_put(skb, size - data_len);
4569         skb->data_len = data_len;
4570         skb->len = size;
4571 
4572         if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4573                 goto err_free;
4574 
4575         err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4576         if (err)
4577                 goto err_free;
4578 
4579         TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4580         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4581         TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4582 
4583         if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4584                 WARN_ON_ONCE(fragstolen); /* should not happen */
4585                 __kfree_skb(skb);
4586         }
4587         return size;
4588 
4589 err_free:
4590         kfree_skb(skb);
4591 err:
4592         return err;
4593 
4594 }
4595 
4596 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4597 {
4598         struct tcp_sock *tp = tcp_sk(sk);
4599         bool fragstolen = false;
4600         int eaten = -1;
4601 
4602         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4603                 __kfree_skb(skb);
4604                 return;
4605         }
4606         skb_dst_drop(skb);
4607         __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4608 
4609         tcp_ecn_accept_cwr(tp, skb);
4610 
4611         tp->rx_opt.dsack = 0;
4612 
4613         /*  Queue data for delivery to the user.
4614          *  Packets in sequence go to the receive queue.
4615          *  Out of sequence packets to the out_of_order_queue.
4616          */
4617         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4618                 if (tcp_receive_window(tp) == 0)
4619                         goto out_of_window;
4620 
4621                 /* Ok. In sequence. In window. */
4622                 if (tp->ucopy.task == current &&
4623                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4624                     sock_owned_by_user(sk) && !tp->urg_data) {
4625                         int chunk = min_t(unsigned int, skb->len,
4626                                           tp->ucopy.len);
4627 
4628                         __set_current_state(TASK_RUNNING);
4629 
4630                         if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4631                                 tp->ucopy.len -= chunk;
4632                                 tp->copied_seq += chunk;
4633                                 eaten = (chunk == skb->len);
4634                                 tcp_rcv_space_adjust(sk);
4635                         }
4636                 }
4637 
4638                 if (eaten <= 0) {
4639 queue_and_out:
4640                         if (eaten < 0) {
4641                                 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4642                                         sk_forced_mem_schedule(sk, skb->truesize);
4643                                 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4644                                         goto drop;
4645                         }
4646                         eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4647                 }
4648                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4649                 if (skb->len)
4650                         tcp_event_data_recv(sk, skb);
4651                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4652                         tcp_fin(sk);
4653 
4654                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4655                         tcp_ofo_queue(sk);
4656 
4657                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4658                          * gap in queue is filled.
4659                          */
4660                         if (skb_queue_empty(&tp->out_of_order_queue))
4661                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4662                 }
4663 
4664                 if (tp->rx_opt.num_sacks)
4665                         tcp_sack_remove(tp);
4666 
4667                 tcp_fast_path_check(sk);
4668 
4669                 if (eaten > 0)
4670                         kfree_skb_partial(skb, fragstolen);
4671                 if (!sock_flag(sk, SOCK_DEAD))
4672                         sk->sk_data_ready(sk);
4673                 return;
4674         }
4675 
4676         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4677                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4678                 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4679                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4680 
4681 out_of_window:
4682                 tcp_enter_quickack_mode(sk);
4683                 inet_csk_schedule_ack(sk);
4684 drop:
4685                 tcp_drop(sk, skb);
4686                 return;
4687         }
4688 
4689         /* Out of window. F.e. zero window probe. */
4690         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4691                 goto out_of_window;
4692 
4693         tcp_enter_quickack_mode(sk);
4694 
4695         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4696                 /* Partial packet, seq < rcv_next < end_seq */
4697                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4698                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4699                            TCP_SKB_CB(skb)->end_seq);
4700 
4701                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4702 
4703                 /* If window is closed, drop tail of packet. But after
4704                  * remembering D-SACK for its head made in previous line.
4705                  */
4706                 if (!tcp_receive_window(tp))
4707                         goto out_of_window;
4708                 goto queue_and_out;
4709         }
4710 
4711         tcp_data_queue_ofo(sk, skb);
4712 }
4713 
4714 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4715                                         struct sk_buff_head *list)
4716 {
4717         struct sk_buff *next = NULL;
4718 
4719         if (!skb_queue_is_last(list, skb))
4720                 next = skb_queue_next(list, skb);
4721 
4722         __skb_unlink(skb, list);
4723         __kfree_skb(skb);
4724         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4725 
4726         return next;
4727 }
4728 
4729 /* Collapse contiguous sequence of skbs head..tail with
4730  * sequence numbers start..end.
4731  *
4732  * If tail is NULL, this means until the end of the list.
4733  *
4734  * Segments with FIN/SYN are not collapsed (only because this
4735  * simplifies code)
4736  */
4737 static void
4738 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4739              struct sk_buff *head, struct sk_buff *tail,
4740              u32 start, u32 end)
4741 {
4742         struct sk_buff *skb, *n;
4743         bool end_of_skbs;
4744 
4745         /* First, check that queue is collapsible and find
4746          * the point where collapsing can be useful. */
4747         skb = head;
4748 restart:
4749         end_of_skbs = true;
4750         skb_queue_walk_from_safe(list, skb, n) {
4751                 if (skb == tail)
4752                         break;
4753                 /* No new bits? It is possible on ofo queue. */
4754                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4755                         skb = tcp_collapse_one(sk, skb, list);
4756                         if (!skb)
4757                                 break;
4758                         goto restart;
4759                 }
4760 
4761                 /* The first skb to collapse is:
4762                  * - not SYN/FIN and
4763                  * - bloated or contains data before "start" or
4764                  *   overlaps to the next one.
4765                  */
4766                 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4767                     (tcp_win_from_space(skb->truesize) > skb->len ||
4768                      before(TCP_SKB_CB(skb)->seq, start))) {
4769                         end_of_skbs = false;
4770                         break;
4771                 }
4772 
4773                 if (!skb_queue_is_last(list, skb)) {
4774                         struct sk_buff *next = skb_queue_next(list, skb);
4775                         if (next != tail &&
4776                             TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4777                                 end_of_skbs = false;
4778                                 break;
4779                         }
4780                 }
4781 
4782                 /* Decided to skip this, advance start seq. */
4783                 start = TCP_SKB_CB(skb)->end_seq;
4784         }
4785         if (end_of_skbs ||
4786             (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4787                 return;
4788 
4789         while (before(start, end)) {
4790                 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4791                 struct sk_buff *nskb;
4792 
4793                 nskb = alloc_skb(copy, GFP_ATOMIC);
4794                 if (!nskb)
4795                         return;
4796 
4797                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4798                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4799                 __skb_queue_before(list, skb, nskb);
4800                 skb_set_owner_r(nskb, sk);
4801 
4802                 /* Copy data, releasing collapsed skbs. */
4803                 while (copy > 0) {
4804                         int offset = start - TCP_SKB_CB(skb)->seq;
4805                         int size = TCP_SKB_CB(skb)->end_seq - start;
4806 
4807                         BUG_ON(offset < 0);
4808                         if (size > 0) {
4809                                 size = min(copy, size);
4810                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4811                                         BUG();
4812                                 TCP_SKB_CB(nskb)->end_seq += size;
4813                                 copy -= size;
4814                                 start += size;
4815                         }
4816                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4817                                 skb = tcp_collapse_one(sk, skb, list);
4818                                 if (!skb ||
4819                                     skb == tail ||
4820                                     (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4821                                         return;
4822                         }
4823                 }
4824         }
4825 }
4826 
4827 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4828  * and tcp_collapse() them until all the queue is collapsed.
4829  */
4830 static void tcp_collapse_ofo_queue(struct sock *sk)
4831 {
4832         struct tcp_sock *tp = tcp_sk(sk);
4833         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4834         struct sk_buff *head;
4835         u32 start, end;
4836 
4837         if (!skb)
4838                 return;
4839 
4840         start = TCP_SKB_CB(skb)->seq;
4841         end = TCP_SKB_CB(skb)->end_seq;
4842         head = skb;
4843 
4844         for (;;) {
4845                 struct sk_buff *next = NULL;
4846 
4847                 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4848                         next = skb_queue_next(&tp->out_of_order_queue, skb);
4849                 skb = next;
4850 
4851                 /* Segment is terminated when we see gap or when
4852                  * we are at the end of all the queue. */
4853                 if (!skb ||
4854                     after(TCP_SKB_CB(skb)->seq, end) ||
4855                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4856                         tcp_collapse(sk, &tp->out_of_order_queue,
4857                                      head, skb, start, end);
4858                         head = skb;
4859                         if (!skb)
4860                                 break;
4861                         /* Start new segment */
4862                         start = TCP_SKB_CB(skb)->seq;
4863                         end = TCP_SKB_CB(skb)->end_seq;
4864                 } else {
4865                         if (before(TCP_SKB_CB(skb)->seq, start))
4866                                 start = TCP_SKB_CB(skb)->seq;
4867                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4868                                 end = TCP_SKB_CB(skb)->end_seq;
4869                 }
4870         }
4871 }
4872 
4873 /*
4874  * Purge the out-of-order queue.
4875  * Return true if queue was pruned.
4876  */
4877 static bool tcp_prune_ofo_queue(struct sock *sk)
4878 {
4879         struct tcp_sock *tp = tcp_sk(sk);
4880         bool res = false;
4881 
4882         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4883                 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
4884                 __skb_queue_purge(&tp->out_of_order_queue);
4885 
4886                 /* Reset SACK state.  A conforming SACK implementation will
4887                  * do the same at a timeout based retransmit.  When a connection
4888                  * is in a sad state like this, we care only about integrity
4889                  * of the connection not performance.
4890                  */
4891                 if (tp->rx_opt.sack_ok)
4892                         tcp_sack_reset(&tp->rx_opt);
4893                 sk_mem_reclaim(sk);
4894                 res = true;
4895         }
4896         return res;
4897 }
4898 
4899 /* Reduce allocated memory if we can, trying to get
4900  * the socket within its memory limits again.
4901  *
4902  * Return less than zero if we should start dropping frames
4903  * until the socket owning process reads some of the data
4904  * to stabilize the situation.
4905  */
4906 static int tcp_prune_queue(struct sock *sk)
4907 {
4908         struct tcp_sock *tp = tcp_sk(sk);
4909 
4910         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4911 
4912         NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
4913 
4914         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4915                 tcp_clamp_window(sk);
4916         else if (tcp_under_memory_pressure(sk))
4917                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4918 
4919         tcp_collapse_ofo_queue(sk);
4920         if (!skb_queue_empty(&sk->sk_receive_queue))
4921                 tcp_collapse(sk, &sk->sk_receive_queue,
4922                              skb_peek(&sk->sk_receive_queue),
4923                              NULL,
4924                              tp->copied_seq, tp->rcv_nxt);
4925         sk_mem_reclaim(sk);
4926 
4927         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4928                 return 0;
4929 
4930         /* Collapsing did not help, destructive actions follow.
4931          * This must not ever occur. */
4932 
4933         tcp_prune_ofo_queue(sk);
4934 
4935         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4936                 return 0;
4937 
4938         /* If we are really being abused, tell the caller to silently
4939          * drop receive data on the floor.  It will get retransmitted
4940          * and hopefully then we'll have sufficient space.
4941          */
4942         NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
4943 
4944         /* Massive buffer overcommit. */
4945         tp->pred_flags = 0;
4946         return -1;
4947 }
4948 
4949 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4950 {
4951         const struct tcp_sock *tp = tcp_sk(sk);
4952 
4953         /* If the user specified a specific send buffer setting, do
4954          * not modify it.
4955          */
4956         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4957                 return false;
4958 
4959         /* If we are under global TCP memory pressure, do not expand.  */
4960         if (tcp_under_memory_pressure(sk))
4961                 return false;
4962 
4963         /* If we are under soft global TCP memory pressure, do not expand.  */
4964         if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4965                 return false;
4966 
4967         /* If we filled the congestion window, do not expand.  */
4968         if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4969                 return false;
4970 
4971         return true;
4972 }
4973 
4974 /* When incoming ACK allowed to free some skb from write_queue,
4975  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4976  * on the exit from tcp input handler.
4977  *
4978  * PROBLEM: sndbuf expansion does not work well with largesend.
4979  */
4980 static void tcp_new_space(struct sock *sk)
4981 {
4982         struct tcp_sock *tp = tcp_sk(sk);
4983 
4984         if (tcp_should_expand_sndbuf(sk)) {
4985                 tcp_sndbuf_expand(sk);
4986                 tp->snd_cwnd_stamp = tcp_time_stamp;
4987         }
4988 
4989         sk->sk_write_space(sk);
4990 }
4991 
4992 static void tcp_check_space(struct sock *sk)
4993 {
4994         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4995                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4996                 /* pairs with tcp_poll() */
4997                 smp_mb__after_atomic();
4998                 if (sk->sk_socket &&
4999                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5000                         tcp_new_space(sk);
5001         }
5002 }
5003 
5004 static inline void tcp_data_snd_check(struct sock *sk)
5005 {
5006         tcp_push_pending_frames(sk);
5007         tcp_check_space(sk);
5008 }
5009 
5010 /*
5011  * Check if sending an ack is needed.
5012  */
5013 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5014 {
5015         struct tcp_sock *tp = tcp_sk(sk);
5016 
5017             /* More than one full frame received... */
5018         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5019              /* ... and right edge of window advances far enough.
5020               * (tcp_recvmsg() will send ACK otherwise). Or...
5021               */
5022              __tcp_select_window(sk) >= tp->rcv_wnd) ||
5023             /* We ACK each frame or... */
5024             tcp_in_quickack_mode(sk) ||
5025             /* We have out of order data. */
5026             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
5027                 /* Then ack it now */
5028                 tcp_send_ack(sk);
5029         } else {
5030                 /* Else, send delayed ack. */
5031                 tcp_send_delayed_ack(sk);
5032         }
5033 }
5034 
5035 static inline void tcp_ack_snd_check(struct sock *sk)
5036 {
5037         if (!inet_csk_ack_scheduled(sk)) {
5038                 /* We sent a data segment already. */
5039                 return;
5040         }
5041         __tcp_ack_snd_check(sk, 1);
5042 }
5043 
5044 /*
5045  *      This routine is only called when we have urgent data
5046  *      signaled. Its the 'slow' part of tcp_urg. It could be
5047  *      moved inline now as tcp_urg is only called from one
5048  *      place. We handle URGent data wrong. We have to - as
5049  *      BSD still doesn't use the correction from RFC961.
5050  *      For 1003.1g we should support a new option TCP_STDURG to permit
5051  *      either form (or just set the sysctl tcp_stdurg).
5052  */
5053 
5054 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5055 {
5056         struct tcp_sock *tp = tcp_sk(sk);
5057         u32 ptr = ntohs(th->urg_ptr);
5058 
5059         if (ptr && !sysctl_tcp_stdurg)
5060                 ptr--;
5061         ptr += ntohl(th->seq);
5062 
5063         /* Ignore urgent data that we've already seen and read. */
5064         if (after(tp->copied_seq, ptr))
5065                 return;
5066 
5067         /* Do not replay urg ptr.
5068          *
5069          * NOTE: interesting situation not covered by specs.
5070          * Misbehaving sender may send urg ptr, pointing to segment,
5071          * which we already have in ofo queue. We are not able to fetch
5072          * such data and will stay in TCP_URG_NOTYET until will be eaten
5073          * by recvmsg(). Seems, we are not obliged to handle such wicked
5074          * situations. But it is worth to think about possibility of some
5075          * DoSes using some hypothetical application level deadlock.
5076          */
5077         if (before(ptr, tp->rcv_nxt))
5078                 return;
5079 
5080         /* Do we already have a newer (or duplicate) urgent pointer? */
5081         if (tp->urg_data && !after(ptr, tp->urg_seq))
5082                 return;
5083 
5084         /* Tell the world about our new urgent pointer. */
5085         sk_send_sigurg(sk);
5086 
5087         /* We may be adding urgent data when the last byte read was
5088          * urgent. To do this requires some care. We cannot just ignore
5089          * tp->copied_seq since we would read the last urgent byte again
5090          * as data, nor can we alter copied_seq until this data arrives
5091          * or we break the semantics of SIOCATMARK (and thus sockatmark())
5092          *
5093          * NOTE. Double Dutch. Rendering to plain English: author of comment
5094          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
5095          * and expect that both A and B disappear from stream. This is _wrong_.
5096          * Though this happens in BSD with high probability, this is occasional.
5097          * Any application relying on this is buggy. Note also, that fix "works"
5098          * only in this artificial test. Insert some normal data between A and B and we will
5099          * decline of BSD again. Verdict: it is better to remove to trap
5100          * buggy users.
5101          */
5102         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5103             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5104                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5105                 tp->copied_seq++;
5106                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5107                         __skb_unlink(skb, &sk->sk_receive_queue);
5108                         __kfree_skb(skb);
5109                 }
5110         }
5111 
5112         tp->urg_data = TCP_URG_NOTYET;
5113         tp->urg_seq = ptr;
5114 
5115         /* Disable header prediction. */
5116         tp->pred_flags = 0;
5117 }
5118 
5119 /* This is the 'fast' part of urgent handling. */
5120 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5121 {
5122         struct tcp_sock *tp = tcp_sk(sk);
5123 
5124         /* Check if we get a new urgent pointer - normally not. */
5125         if (th->urg)
5126                 tcp_check_urg(sk, th);
5127 
5128         /* Do we wait for any urgent data? - normally not... */
5129         if (tp->urg_data == TCP_URG_NOTYET) {
5130                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5131                           th->syn;
5132 
5133                 /* Is the urgent pointer pointing into this packet? */
5134                 if (ptr < skb->len) {
5135                         u8 tmp;
5136                         if (skb_copy_bits(skb, ptr, &tmp, 1))
5137                                 BUG();
5138                         tp->urg_data = TCP_URG_VALID | tmp;
5139                         if (!sock_flag(sk, SOCK_DEAD))
5140                                 sk->sk_data_ready(sk);
5141                 }
5142         }
5143 }
5144 
5145 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5146 {
5147         struct tcp_sock *tp = tcp_sk(sk);
5148         int chunk = skb->len - hlen;
5149         int err;
5150 
5151         if (skb_csum_unnecessary(skb))
5152                 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5153         else
5154                 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5155 
5156         if (!err) {
5157                 tp->ucopy.len -= chunk;
5158                 tp->copied_seq += chunk;
5159                 tcp_rcv_space_adjust(sk);
5160         }
5161 
5162         return err;
5163 }
5164 
5165 /* Does PAWS and seqno based validation of an incoming segment, flags will
5166  * play significant role here.
5167  */
5168 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5169                                   const struct tcphdr *th, int syn_inerr)
5170 {
5171         struct tcp_sock *tp = tcp_sk(sk);
5172 
5173         /* RFC1323: H1. Apply PAWS check first. */
5174         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5175             tcp_paws_discard(sk, skb)) {
5176                 if (!th->rst) {
5177                         NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5178                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5179                                                   LINUX_MIB_TCPACKSKIPPEDPAWS,
5180                                                   &tp->last_oow_ack_time))
5181                                 tcp_send_dupack(sk, skb);
5182                         goto discard;
5183                 }
5184                 /* Reset is accepted even if it did not pass PAWS. */
5185         }
5186 
5187         /* Step 1: check sequence number */
5188         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5189                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5190                  * (RST) segments are validated by checking their SEQ-fields."
5191                  * And page 69: "If an incoming segment is not acceptable,
5192                  * an acknowledgment should be sent in reply (unless the RST
5193                  * bit is set, if so drop the segment and return)".
5194                  */
5195                 if (!th->rst) {
5196                         if (th->syn)
5197                                 goto syn_challenge;
5198                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5199                                                   LINUX_MIB_TCPACKSKIPPEDSEQ,
5200                                                   &tp->last_oow_ack_time))
5201                                 tcp_send_dupack(sk, skb);
5202                 }
5203                 goto discard;
5204         }
5205 
5206         /* Step 2: check RST bit */
5207         if (th->rst) {
5208                 /* RFC 5961 3.2 :
5209                  * If sequence number exactly matches RCV.NXT, then
5210                  *     RESET the connection
5211                  * else
5212                  *     Send a challenge ACK
5213                  */
5214                 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5215                         tcp_reset(sk);
5216                 else
5217                         tcp_send_challenge_ack(sk, skb);
5218                 goto discard;
5219         }
5220 
5221         /* step 3: check security and precedence [ignored] */
5222 
5223         /* step 4: Check for a SYN
5224          * RFC 5961 4.2 : Send a challenge ack
5225          */
5226         if (th->syn) {
5227 syn_challenge:
5228                 if (syn_inerr)
5229                         TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5230                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5231                 tcp_send_challenge_ack(sk, skb);
5232                 goto discard;
5233         }
5234 
5235         return true;
5236 
5237 discard:
5238         tcp_drop(sk, skb);
5239         return false;
5240 }
5241 
5242 /*
5243  *      TCP receive function for the ESTABLISHED state.
5244  *
5245  *      It is split into a fast path and a slow path. The fast path is
5246  *      disabled when:
5247  *      - A zero window was announced from us - zero window probing
5248  *        is only handled properly in the slow path.
5249  *      - Out of order segments arrived.
5250  *      - Urgent data is expected.
5251  *      - There is no buffer space left
5252  *      - Unexpected TCP flags/window values/header lengths are received
5253  *        (detected by checking the TCP header against pred_flags)
5254  *      - Data is sent in both directions. Fast path only supports pure senders
5255  *        or pure receivers (this means either the sequence number or the ack
5256  *        value must stay constant)
5257  *      - Unexpected TCP option.
5258  *
5259  *      When these conditions are not satisfied it drops into a standard
5260  *      receive procedure patterned after RFC793 to handle all cases.
5261  *      The first three cases are guaranteed by proper pred_flags setting,
5262  *      the rest is checked inline. Fast processing is turned on in
5263  *      tcp_data_queue when everything is OK.
5264  */
5265 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5266                          const struct tcphdr *th, unsigned int len)
5267 {
5268         struct tcp_sock *tp = tcp_sk(sk);
5269 
5270         if (unlikely(!sk->sk_rx_dst))
5271                 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5272         /*
5273          *      Header prediction.
5274          *      The code loosely follows the one in the famous
5275          *      "30 instruction TCP receive" Van Jacobson mail.
5276          *
5277          *      Van's trick is to deposit buffers into socket queue
5278          *      on a device interrupt, to call tcp_recv function
5279          *      on the receive process context and checksum and copy
5280          *      the buffer to user space. smart...
5281          *
5282          *      Our current scheme is not silly either but we take the
5283          *      extra cost of the net_bh soft interrupt processing...
5284          *      We do checksum and copy also but from device to kernel.
5285          */
5286 
5287         tp->rx_opt.saw_tstamp = 0;
5288 
5289         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5290          *      if header_prediction is to be made
5291          *      'S' will always be tp->tcp_header_len >> 2
5292          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5293          *  turn it off (when there are holes in the receive
5294          *       space for instance)
5295          *      PSH flag is ignored.
5296          */
5297 
5298         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5299             TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5300             !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5301                 int tcp_header_len = tp->tcp_header_len;
5302 
5303                 /* Timestamp header prediction: tcp_header_len
5304                  * is automatically equal to th->doff*4 due to pred_flags
5305                  * match.
5306                  */
5307 
5308                 /* Check timestamp */
5309                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5310                         /* No? Slow path! */
5311                         if (!tcp_parse_aligned_timestamp(tp, th))
5312                                 goto slow_path;
5313 
5314                         /* If PAWS failed, check it more carefully in slow path */
5315                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5316                                 goto slow_path;
5317 
5318                         /* DO NOT update ts_recent here, if checksum fails
5319                          * and timestamp was corrupted part, it will result
5320                          * in a hung connection since we will drop all
5321                          * future packets due to the PAWS test.
5322                          */
5323                 }
5324 
5325                 if (len <= tcp_header_len) {
5326                         /* Bulk data transfer: sender */
5327                         if (len == tcp_header_len) {
5328                                 /* Predicted packet is in window by definition.
5329                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5330                                  * Hence, check seq<=rcv_wup reduces to:
5331                                  */
5332                                 if (tcp_header_len ==
5333                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5334                                     tp->rcv_nxt == tp->rcv_wup)
5335                                         tcp_store_ts_recent(tp);
5336 
5337                                 /* We know that such packets are checksummed
5338                                  * on entry.
5339                                  */
5340                                 tcp_ack(sk, skb, 0);
5341                                 __kfree_skb(skb);
5342                                 tcp_data_snd_check(sk);
5343                                 return;
5344                         } else { /* Header too small */
5345                                 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5346                                 goto discard;
5347                         }
5348                 } else {
5349                         int eaten = 0;
5350                         bool fragstolen = false;
5351 
5352                         if (tp->ucopy.task == current &&
5353                             tp->copied_seq == tp->rcv_nxt &&
5354                             len - tcp_header_len <= tp->ucopy.len &&
5355                             sock_owned_by_user(sk)) {
5356                                 __set_current_state(TASK_RUNNING);
5357 
5358                                 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5359                                         /* Predicted packet is in window by definition.
5360                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5361                                          * Hence, check seq<=rcv_wup reduces to:
5362                                          */
5363                                         if (tcp_header_len ==
5364                                             (sizeof(struct tcphdr) +
5365                                              TCPOLEN_TSTAMP_ALIGNED) &&
5366                                             tp->rcv_nxt == tp->rcv_wup)
5367                                                 tcp_store_ts_recent(tp);
5368 
5369                                         tcp_rcv_rtt_measure_ts(sk, skb);
5370 
5371                                         __skb_pull(skb, tcp_header_len);
5372                                         tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5373                                         NET_INC_STATS(sock_net(sk),
5374                                                         LINUX_MIB_TCPHPHITSTOUSER);
5375                                         eaten = 1;
5376                                 }
5377                         }
5378                         if (!eaten) {
5379                                 if (tcp_checksum_complete(skb))
5380                                         goto csum_error;
5381 
5382                                 if ((int)skb->truesize > sk->sk_forward_alloc)
5383                                         goto step5;
5384 
5385                                 /* Predicted packet is in window by definition.
5386                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5387                                  * Hence, check seq<=rcv_wup reduces to:
5388                                  */
5389                                 if (tcp_header_len ==
5390                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5391                                     tp->rcv_nxt == tp->rcv_wup)
5392                                         tcp_store_ts_recent(tp);
5393 
5394                                 tcp_rcv_rtt_measure_ts(sk, skb);
5395 
5396                                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5397 
5398                                 /* Bulk data transfer: receiver */
5399                                 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5400                                                       &fragstolen);
5401                         }
5402 
5403                         tcp_event_data_recv(sk, skb);
5404 
5405                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5406                                 /* Well, only one small jumplet in fast path... */
5407                                 tcp_ack(sk, skb, FLAG_DATA);
5408                                 tcp_data_snd_check(sk);
5409                                 if (!inet_csk_ack_scheduled(sk))
5410                                         goto no_ack;
5411                         }
5412 
5413                         __tcp_ack_snd_check(sk, 0);
5414 no_ack:
5415                         if (eaten)
5416                                 kfree_skb_partial(skb, fragstolen);
5417                         sk->sk_data_ready(sk);
5418                         return;
5419                 }
5420         }
5421 
5422 slow_path:
5423         if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5424                 goto csum_error;
5425 
5426         if (!th->ack && !th->rst && !th->syn)
5427                 goto discard;
5428 
5429         /*
5430          *      Standard slow path.
5431          */
5432 
5433         if (!tcp_validate_incoming(sk, skb, th, 1))
5434                 return;
5435 
5436 step5:
5437         if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5438                 goto discard;
5439 
5440         tcp_rcv_rtt_measure_ts(sk, skb);
5441 
5442         /* Process urgent data. */
5443         tcp_urg(sk, skb, th);
5444 
5445         /* step 7: process the segment text */
5446         tcp_data_queue(sk, skb);
5447 
5448         tcp_data_snd_check(sk);
5449         tcp_ack_snd_check(sk);
5450         return;
5451 
5452 csum_error:
5453         TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5454         TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5455 
5456 discard:
5457         tcp_drop(sk, skb);
5458 }
5459 EXPORT_SYMBOL(tcp_rcv_established);
5460 
5461 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5462 {
5463         struct tcp_sock *tp = tcp_sk(sk);
5464         struct inet_connection_sock *icsk = inet_csk(sk);
5465 
5466         tcp_set_state(sk, TCP_ESTABLISHED);
5467 
5468         if (skb) {
5469                 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5470                 security_inet_conn_established(sk, skb);
5471         }
5472 
5473         /* Make sure socket is routed, for correct metrics.  */
5474         icsk->icsk_af_ops->rebuild_header(sk);
5475 
5476         tcp_init_metrics(sk);
5477 
5478         tcp_init_congestion_control(sk);
5479 
5480         /* Prevent spurious tcp_cwnd_restart() on first data
5481          * packet.
5482          */
5483         tp->lsndtime = tcp_time_stamp;
5484 
5485         tcp_init_buffer_space(sk);
5486 
5487         if (sock_flag(sk, SOCK_KEEPOPEN))
5488                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5489 
5490         if (!tp->rx_opt.snd_wscale)
5491                 __tcp_fast_path_on(tp, tp->snd_wnd);
5492         else
5493                 tp->pred_flags = 0;
5494 
5495         if (!sock_flag(sk, SOCK_DEAD)) {
5496                 sk->sk_state_change(sk);
5497                 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5498         }
5499 }
5500 
5501 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5502                                     struct tcp_fastopen_cookie *cookie)
5503 {
5504         struct tcp_sock *tp = tcp_sk(sk);
5505         struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5506         u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5507         bool syn_drop = false;
5508 
5509         if (mss == tp->rx_opt.user_mss) {
5510                 struct tcp_options_received opt;
5511 
5512                 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5513                 tcp_clear_options(&opt);
5514                 opt.user_mss = opt.mss_clamp = 0;
5515                 tcp_parse_options(synack, &opt, 0, NULL);
5516                 mss = opt.mss_clamp;
5517         }
5518 
5519         if (!tp->syn_fastopen) {
5520                 /* Ignore an unsolicited cookie */
5521                 cookie->len = -1;
5522         } else if (tp->total_retrans) {
5523                 /* SYN timed out and the SYN-ACK neither has a cookie nor
5524                  * acknowledges data. Presumably the remote received only
5525                  * the retransmitted (regular) SYNs: either the original
5526                  * SYN-data or the corresponding SYN-ACK was dropped.
5527                  */
5528                 syn_drop = (cookie->len < 0 && data);
5529         } else if (cookie->len < 0 && !tp->syn_data) {
5530                 /* We requested a cookie but didn't get it. If we did not use
5531                  * the (old) exp opt format then try so next time (try_exp=1).
5532                  * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5533                  */
5534                 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5535         }
5536 
5537         tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5538 
5539         if (data) { /* Retransmit unacked data in SYN */
5540                 tcp_for_write_queue_from(data, sk) {
5541                         if (data == tcp_send_head(sk) ||
5542                             __tcp_retransmit_skb(sk, data, 1))
5543                                 break;
5544                 }
5545                 tcp_rearm_rto(sk);
5546                 NET_INC_STATS(sock_net(sk),
5547                                 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5548                 return true;
5549         }
5550         tp->syn_data_acked = tp->syn_data;
5551         if (tp->syn_data_acked)
5552                 NET_INC_STATS(sock_net(sk),
5553                                 LINUX_MIB_TCPFASTOPENACTIVE);
5554 
5555         tcp_fastopen_add_skb(sk, synack);
5556 
5557         return false;
5558 }
5559 
5560 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5561                                          const struct tcphdr *th)
5562 {
5563         struct inet_connection_sock *icsk = inet_csk(sk);
5564         struct tcp_sock *tp = tcp_sk(sk);
5565         struct tcp_fastopen_cookie foc = { .len = -1 };
5566         int saved_clamp = tp->rx_opt.mss_clamp;
5567 
5568         tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5569         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5570                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5571 
5572         if (th->ack) {
5573                 /* rfc793:
5574                  * "If the state is SYN-SENT then
5575                  *    first check the ACK bit
5576                  *      If the ACK bit is set
5577                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5578                  *        a reset (unless the RST bit is set, if so drop
5579                  *        the segment and return)"
5580                  */
5581                 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5582                     after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5583                         goto reset_and_undo;
5584 
5585                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5586                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5587                              tcp_time_stamp)) {
5588                         NET_INC_STATS(sock_net(sk),
5589                                         LINUX_MIB_PAWSACTIVEREJECTED);
5590                         goto reset_and_undo;
5591                 }
5592 
5593                 /* Now ACK is acceptable.
5594                  *
5595                  * "If the RST bit is set
5596                  *    If the ACK was acceptable then signal the user "error:
5597                  *    connection reset", drop the segment, enter CLOSED state,
5598                  *    delete TCB, and return."
5599                  */
5600 
5601                 if (th->rst) {
5602                         tcp_reset(sk);
5603                         goto discard;
5604                 }
5605 
5606                 /* rfc793:
5607                  *   "fifth, if neither of the SYN or RST bits is set then
5608                  *    drop the segment and return."
5609                  *
5610                  *    See note below!
5611                  *                                        --ANK(990513)
5612                  */
5613                 if (!th->syn)
5614                         goto discard_and_undo;
5615 
5616                 /* rfc793:
5617                  *   "If the SYN bit is on ...
5618                  *    are acceptable then ...
5619                  *    (our SYN has been ACKed), change the connection
5620                  *    state to ESTABLISHED..."
5621                  */
5622 
5623                 tcp_ecn_rcv_synack(tp, th);
5624 
5625                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5626                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5627 
5628                 /* Ok.. it's good. Set up sequence numbers and
5629                  * move to established.
5630                  */
5631                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5632                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5633 
5634                 /* RFC1323: The window in SYN & SYN/ACK segments is
5635                  * never scaled.
5636                  */
5637                 tp->snd_wnd = ntohs(th->window);
5638 
5639                 if (!tp->rx_opt.wscale_ok) {
5640                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5641                         tp->window_clamp = min(tp->window_clamp, 65535U);
5642                 }
5643 
5644                 if (tp->rx_opt.saw_tstamp) {
5645                         tp->rx_opt.tstamp_ok       = 1;
5646                         tp->tcp_header_len =
5647                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5648                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5649                         tcp_store_ts_recent(tp);
5650                 } else {
5651                         tp->tcp_header_len = sizeof(struct tcphdr);
5652                 }
5653 
5654                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5655                         tcp_enable_fack(tp);
5656 
5657                 tcp_mtup_init(sk);
5658                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5659                 tcp_initialize_rcv_mss(sk);
5660 
5661                 /* Remember, tcp_poll() does not lock socket!
5662                  * Change state from SYN-SENT only after copied_seq
5663                  * is initialized. */
5664                 tp->copied_seq = tp->rcv_nxt;
5665 
5666                 smp_mb();
5667 
5668                 tcp_finish_connect(sk, skb);
5669 
5670                 if ((tp->syn_fastopen || tp->syn_data) &&
5671                     tcp_rcv_fastopen_synack(sk, skb, &foc))
5672                         return -1;
5673 
5674                 if (sk->sk_write_pending ||
5675                     icsk->icsk_accept_queue.rskq_defer_accept ||
5676                     icsk->icsk_ack.pingpong) {
5677                         /* Save one ACK. Data will be ready after
5678                          * several ticks, if write_pending is set.
5679                          *
5680                          * It may be deleted, but with this feature tcpdumps
5681                          * look so _wonderfully_ clever, that I was not able
5682                          * to stand against the temptation 8)     --ANK
5683                          */
5684                         inet_csk_schedule_ack(sk);
5685                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5686                         tcp_enter_quickack_mode(sk);
5687                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5688                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5689 
5690 discard:
5691                         tcp_drop(sk, skb);
5692                         return 0;
5693                 } else {
5694                         tcp_send_ack(sk);
5695                 }
5696                 return -1;
5697         }
5698 
5699         /* No ACK in the segment */
5700 
5701         if (th->rst) {
5702                 /* rfc793:
5703                  * "If the RST bit is set
5704                  *
5705                  *      Otherwise (no ACK) drop the segment and return."
5706                  */
5707 
5708                 goto discard_and_undo;
5709         }
5710 
5711         /* PAWS check. */
5712         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5713             tcp_paws_reject(&tp->rx_opt, 0))
5714                 goto discard_and_undo;
5715 
5716         if (th->syn) {
5717                 /* We see SYN without ACK. It is attempt of
5718                  * simultaneous connect with crossed SYNs.
5719                  * Particularly, it can be connect to self.
5720                  */
5721                 tcp_set_state(sk, TCP_SYN_RECV);
5722 
5723                 if (tp->rx_opt.saw_tstamp) {
5724                         tp->rx_opt.tstamp_ok = 1;
5725                         tcp_store_ts_recent(tp);
5726                         tp->tcp_header_len =
5727                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5728                 } else {
5729                         tp->tcp_header_len = sizeof(struct tcphdr);
5730                 }
5731 
5732                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5733                 tp->copied_seq = tp->rcv_nxt;
5734                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5735 
5736                 /* RFC1323: The window in SYN & SYN/ACK segments is
5737                  * never scaled.
5738                  */
5739                 tp->snd_wnd    = ntohs(th->window);
5740                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5741                 tp->max_window = tp->snd_wnd;
5742 
5743                 tcp_ecn_rcv_syn(tp, th);
5744 
5745                 tcp_mtup_init(sk);
5746                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5747                 tcp_initialize_rcv_mss(sk);
5748 
5749                 tcp_send_synack(sk);
5750 #if 0
5751                 /* Note, we could accept data and URG from this segment.
5752                  * There are no obstacles to make this (except that we must
5753                  * either change tcp_recvmsg() to prevent it from returning data
5754                  * before 3WHS completes per RFC793, or employ TCP Fast Open).
5755                  *
5756                  * However, if we ignore data in ACKless segments sometimes,
5757                  * we have no reasons to accept it sometimes.
5758                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5759                  * is not flawless. So, discard packet for sanity.
5760                  * Uncomment this return to process the data.
5761                  */
5762                 return -1;
5763 #else
5764                 goto discard;
5765 #endif
5766         }
5767         /* "fifth, if neither of the SYN or RST bits is set then
5768          * drop the segment and return."
5769          */
5770 
5771 discard_and_undo:
5772         tcp_clear_options(&tp->rx_opt);
5773         tp->rx_opt.mss_clamp = saved_clamp;
5774         goto discard;
5775 
5776 reset_and_undo:
5777         tcp_clear_options(&tp->rx_opt);
5778         tp->rx_opt.mss_clamp = saved_clamp;
5779         return 1;
5780 }
5781 
5782 /*
5783  *      This function implements the receiving procedure of RFC 793 for
5784  *      all states except ESTABLISHED and TIME_WAIT.
5785  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5786  *      address independent.
5787  */
5788 
5789 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5790 {
5791         struct tcp_sock *tp = tcp_sk(sk);
5792         struct inet_connection_sock *icsk = inet_csk(sk);
5793         const struct tcphdr *th = tcp_hdr(skb);
5794         struct request_sock *req;
5795         int queued = 0;
5796         bool acceptable;
5797 
5798         switch (sk->sk_state) {
5799         case TCP_CLOSE:
5800                 goto discard;
5801 
5802         case TCP_LISTEN:
5803                 if (th->ack)
5804                         return 1;
5805 
5806                 if (th->rst)
5807                         goto discard;
5808 
5809                 if (th->syn) {
5810                         if (th->fin)
5811                                 goto discard;
5812                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5813                                 return 1;
5814 
5815                         consume_skb(skb);
5816                         return 0;
5817                 }
5818                 goto discard;
5819 
5820         case TCP_SYN_SENT:
5821                 tp->rx_opt.saw_tstamp = 0;
5822                 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5823                 if (queued >= 0)
5824                         return queued;
5825 
5826                 /* Do step6 onward by hand. */
5827                 tcp_urg(sk, skb, th);
5828                 __kfree_skb(skb);
5829                 tcp_data_snd_check(sk);
5830                 return 0;
5831         }
5832 
5833         tp->rx_opt.saw_tstamp = 0;
5834         req = tp->fastopen_rsk;
5835         if (req) {
5836                 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5837                     sk->sk_state != TCP_FIN_WAIT1);
5838 
5839                 if (!tcp_check_req(sk, skb, req, true))
5840                         goto discard;
5841         }
5842 
5843         if (!th->ack && !th->rst && !th->syn)
5844                 goto discard;
5845 
5846         if (!tcp_validate_incoming(sk, skb, th, 0))
5847                 return 0;
5848 
5849         /* step 5: check the ACK field */
5850         acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5851                                       FLAG_UPDATE_TS_RECENT) > 0;
5852 
5853         switch (sk->sk_state) {
5854         case TCP_SYN_RECV:
5855                 if (!acceptable)
5856                         return 1;
5857 
5858                 if (!tp->srtt_us)
5859                         tcp_synack_rtt_meas(sk, req);
5860 
5861                 /* Once we leave TCP_SYN_RECV, we no longer need req
5862                  * so release it.
5863                  */
5864                 if (req) {
5865                         tp->total_retrans = req->num_retrans;
5866                         reqsk_fastopen_remove(sk, req, false);
5867                 } else {
5868                         /* Make sure socket is routed, for correct metrics. */
5869                         icsk->icsk_af_ops->rebuild_header(sk);
5870                         tcp_init_congestion_control(sk);
5871 
5872                         tcp_mtup_init(sk);
5873                         tp->copied_seq = tp->rcv_nxt;
5874                         tcp_init_buffer_space(sk);
5875                 }
5876                 smp_mb();
5877                 tcp_set_state(sk, TCP_ESTABLISHED);
5878                 sk->sk_state_change(sk);
5879 
5880                 /* Note, that this wakeup is only for marginal crossed SYN case.
5881                  * Passively open sockets are not waked up, because
5882                  * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5883                  */
5884                 if (sk->sk_socket)
5885                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5886 
5887                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5888                 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5889                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5890 
5891                 if (tp->rx_opt.tstamp_ok)
5892                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5893 
5894                 if (req) {
5895                         /* Re-arm the timer because data may have been sent out.
5896                          * This is similar to the regular data transmission case
5897                          * when new data has just been ack'ed.
5898                          *
5899                          * (TFO) - we could try to be more aggressive and
5900                          * retransmitting any data sooner based on when they
5901                          * are sent out.
5902                          */
5903                         tcp_rearm_rto(sk);
5904                 } else
5905                         tcp_init_metrics(sk);
5906 
5907                 tcp_update_pacing_rate(sk);
5908 
5909                 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5910                 tp->lsndtime = tcp_time_stamp;
5911 
5912                 tcp_initialize_rcv_mss(sk);
5913                 tcp_fast_path_on(tp);
5914                 break;
5915 
5916         case TCP_FIN_WAIT1: {
5917                 struct dst_entry *dst;
5918                 int tmo;
5919 
5920                 /* If we enter the TCP_FIN_WAIT1 state and we are a
5921                  * Fast Open socket and this is the first acceptable
5922                  * ACK we have received, this would have acknowledged
5923                  * our SYNACK so stop the SYNACK timer.
5924                  */
5925                 if (req) {
5926                         /* Return RST if ack_seq is invalid.
5927                          * Note that RFC793 only says to generate a
5928                          * DUPACK for it but for TCP Fast Open it seems
5929                          * better to treat this case like TCP_SYN_RECV
5930                          * above.
5931                          */
5932                         if (!acceptable)
5933                                 return 1;
5934                         /* We no longer need the request sock. */
5935                         reqsk_fastopen_remove(sk, req, false);
5936                         tcp_rearm_rto(sk);
5937                 }
5938                 if (tp->snd_una != tp->write_seq)
5939                         break;
5940 
5941                 tcp_set_state(sk, TCP_FIN_WAIT2);
5942                 sk->sk_shutdown |= SEND_SHUTDOWN;
5943 
5944                 dst = __sk_dst_get(sk);
5945                 if (dst)
5946                         dst_confirm(dst);
5947 
5948                 if (!sock_flag(sk, SOCK_DEAD)) {
5949                         /* Wake up lingering close() */
5950                         sk->sk_state_change(sk);
5951                         break;
5952                 }
5953 
5954                 if (tp->linger2 < 0 ||
5955                     (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5956                      after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5957                         tcp_done(sk);
5958                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5959                         return 1;
5960                 }
5961 
5962                 tmo = tcp_fin_time(sk);
5963                 if (tmo > TCP_TIMEWAIT_LEN) {
5964                         inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5965                 } else if (th->fin || sock_owned_by_user(sk)) {
5966                         /* Bad case. We could lose such FIN otherwise.
5967                          * It is not a big problem, but it looks confusing
5968                          * and not so rare event. We still can lose it now,
5969                          * if it spins in bh_lock_sock(), but it is really
5970                          * marginal case.
5971                          */
5972                         inet_csk_reset_keepalive_timer(sk, tmo);
5973                 } else {
5974                         tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5975                         goto discard;
5976                 }
5977                 break;
5978         }
5979 
5980         case TCP_CLOSING:
5981                 if (tp->snd_una == tp->write_seq) {
5982                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5983                         goto discard;
5984                 }
5985                 break;
5986 
5987         case TCP_LAST_ACK:
5988                 if (tp->snd_una == tp->write_seq) {
5989                         tcp_update_metrics(sk);
5990                         tcp_done(sk);
5991                         goto discard;
5992                 }
5993                 break;
5994         }
5995 
5996         /* step 6: check the URG bit */
5997         tcp_urg(sk, skb, th);
5998 
5999         /* step 7: process the segment text */
6000         switch (sk->sk_state) {
6001         case TCP_CLOSE_WAIT:
6002         case TCP_CLOSING:
6003         case TCP_LAST_ACK:
6004                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6005                         break;
6006         case TCP_FIN_WAIT1:
6007         case TCP_FIN_WAIT2:
6008                 /* RFC 793 says to queue data in these states,
6009                  * RFC 1122 says we MUST send a reset.
6010                  * BSD 4.4 also does reset.
6011                  */
6012                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6013                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6014                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6015                                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6016                                 tcp_reset(sk);
6017                                 return 1;
6018                         }
6019                 }
6020                 /* Fall through */
6021         case TCP_ESTABLISHED:
6022                 tcp_data_queue(sk, skb);
6023                 queued = 1;
6024                 break;
6025         }
6026 
6027         /* tcp_data could move socket to TIME-WAIT */
6028         if (sk->sk_state != TCP_CLOSE) {
6029                 tcp_data_snd_check(sk);
6030                 tcp_ack_snd_check(sk);
6031         }
6032 
6033         if (!queued) {
6034 discard:
6035                 tcp_drop(sk, skb);
6036         }
6037         return 0;
6038 }
6039 EXPORT_SYMBOL(tcp_rcv_state_process);
6040 
6041 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6042 {
6043         struct inet_request_sock *ireq = inet_rsk(req);
6044 
6045         if (family == AF_INET)
6046                 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6047                                     &ireq->ir_rmt_addr, port);
6048 #if IS_ENABLED(CONFIG_IPV6)
6049         else if (family == AF_INET6)
6050                 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6051                                     &ireq->ir_v6_rmt_addr, port);
6052 #endif
6053 }
6054 
6055 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6056  *
6057  * If we receive a SYN packet with these bits set, it means a
6058  * network is playing bad games with TOS bits. In order to
6059  * avoid possible false congestion notifications, we disable
6060  * TCP ECN negotiation.
6061  *
6062  * Exception: tcp_ca wants ECN. This is required for DCTCP
6063  * congestion control: Linux DCTCP asserts ECT on all packets,
6064  * including SYN, which is most optimal solution; however,
6065  * others, such as FreeBSD do not.
6066  */
6067 static void tcp_ecn_create_request(struct request_sock *req,
6068                                    const struct sk_buff *skb,
6069                                    const struct sock *listen_sk,
6070                                    const struct dst_entry *dst)
6071 {
6072         const struct tcphdr *th = tcp_hdr(skb);
6073         const struct net *net = sock_net(listen_sk);
6074         bool th_ecn = th->ece && th->cwr;
6075         bool ect, ecn_ok;
6076         u32 ecn_ok_dst;
6077 
6078         if (!th_ecn)
6079                 return;
6080 
6081         ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6082         ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6083         ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6084 
6085         if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6086             (ecn_ok_dst & DST_FEATURE_ECN_CA))
6087                 inet_rsk(req)->ecn_ok = 1;
6088 }
6089 
6090 static void tcp_openreq_init(struct request_sock *req,
6091                              const struct tcp_options_received *rx_opt,
6092                              struct sk_buff *skb, const struct sock *sk)
6093 {
6094         struct inet_request_sock *ireq = inet_rsk(req);
6095 
6096         req->rsk_rcv_wnd = 0;           /* So that tcp_send_synack() knows! */
6097         req->cookie_ts = 0;
6098         tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6099         tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6100         skb_mstamp_get(&tcp_rsk(req)->snt_synack);
6101         tcp_rsk(req)->last_oow_ack_time = 0;
6102         req->mss = rx_opt->mss_clamp;
6103         req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6104         ireq->tstamp_ok = rx_opt->tstamp_ok;
6105         ireq->sack_ok = rx_opt->sack_ok;
6106         ireq->snd_wscale = rx_opt->snd_wscale;
6107         ireq->wscale_ok = rx_opt->wscale_ok;
6108         ireq->acked = 0;
6109         ireq->ecn_ok = 0;
6110         ireq->ir_rmt_port = tcp_hdr(skb)->source;
6111         ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6112         ireq->ir_mark = inet_request_mark(sk, skb);
6113 }
6114 
6115 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6116                                       struct sock *sk_listener,
6117                                       bool attach_listener)
6118 {
6119         struct request_sock *req = reqsk_alloc(ops, sk_listener,
6120                                                attach_listener);
6121 
6122         if (req) {
6123                 struct inet_request_sock *ireq = inet_rsk(req);
6124 
6125                 kmemcheck_annotate_bitfield(ireq, flags);
6126                 ireq->opt = NULL;
6127                 atomic64_set(&ireq->ir_cookie, 0);
6128                 ireq->ireq_state = TCP_NEW_SYN_RECV;
6129                 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6130                 ireq->ireq_family = sk_listener->sk_family;
6131         }
6132 
6133         return req;
6134 }
6135 EXPORT_SYMBOL(inet_reqsk_alloc);
6136 
6137 /*
6138  * Return true if a syncookie should be sent
6139  */
6140 static bool tcp_syn_flood_action(const struct sock *sk,
6141                                  const struct sk_buff *skb,
6142                                  const char *proto)
6143 {
6144         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6145         const char *msg = "Dropping request";
6146         bool want_cookie = false;
6147         struct net *net = sock_net(sk);
6148 
6149 #ifdef CONFIG_SYN_COOKIES
6150         if (net->ipv4.sysctl_tcp_syncookies) {
6151                 msg = "Sending cookies";
6152                 want_cookie = true;
6153                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6154         } else
6155 #endif
6156                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6157 
6158         if (!queue->synflood_warned &&
6159             net->ipv4.sysctl_tcp_syncookies != 2 &&
6160             xchg(&queue->synflood_warned, 1) == 0)
6161                 pr_info("%s: Possible SYN flooding on port %d. %s.  Check SNMP counters.\n",
6162                         proto, ntohs(tcp_hdr(skb)->dest), msg);
6163 
6164         return want_cookie;
6165 }
6166 
6167 static void tcp_reqsk_record_syn(const struct sock *sk,
6168                                  struct request_sock *req,
6169                                  const struct sk_buff *skb)
6170 {
6171         if (tcp_sk(sk)->save_syn) {
6172                 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6173                 u32 *copy;
6174 
6175                 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6176                 if (copy) {
6177                         copy[0] = len;
6178                         memcpy(&copy[1], skb_network_header(skb), len);
6179                         req->saved_syn = copy;
6180                 }
6181         }
6182 }
6183 
6184 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6185                      const struct tcp_request_sock_ops *af_ops,
6186                      struct sock *sk, struct sk_buff *skb)
6187 {
6188         struct tcp_fastopen_cookie foc = { .len = -1 };
6189         __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6190         struct tcp_options_received tmp_opt;
6191         struct tcp_sock *tp = tcp_sk(sk);
6192         struct net *net = sock_net(sk);
6193         struct sock *fastopen_sk = NULL;
6194         struct dst_entry *dst = NULL;
6195         struct request_sock *req;
6196         bool want_cookie = false;
6197         struct flowi fl;
6198 
6199         /* TW buckets are converted to open requests without
6200          * limitations, they conserve resources and peer is
6201          * evidently real one.
6202          */
6203         if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6204              inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6205                 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6206                 if (!want_cookie)
6207                         goto drop;
6208         }
6209 
6210 
6211         /* Accept backlog is full. If we have already queued enough
6212          * of warm entries in syn queue, drop request. It is better than
6213          * clogging syn queue with openreqs with exponentially increasing
6214          * timeout.
6215          */
6216         if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
6217                 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6218                 goto drop;
6219         }
6220 
6221         req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6222         if (!req)
6223                 goto drop;
6224 
6225         tcp_rsk(req)->af_specific = af_ops;
6226 
6227         tcp_clear_options(&tmp_opt);
6228         tmp_opt.mss_clamp = af_ops->mss_clamp;
6229         tmp_opt.user_mss  = tp->rx_opt.user_mss;
6230         tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6231 
6232         if (want_cookie && !tmp_opt.saw_tstamp)
6233                 tcp_clear_options(&tmp_opt);
6234 
6235         tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6236         tcp_openreq_init(req, &tmp_opt, skb, sk);
6237 
6238         /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6239         inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6240 
6241         af_ops->init_req(req, sk, skb);
6242 
6243         if (security_inet_conn_request(sk, skb, req))
6244                 goto drop_and_free;
6245 
6246         if (!want_cookie && !isn) {
6247                 /* VJ's idea. We save last timestamp seen
6248                  * from the destination in peer table, when entering
6249                  * state TIME-WAIT, and check against it before
6250                  * accepting new connection request.
6251                  *
6252                  * If "isn" is not zero, this request hit alive
6253                  * timewait bucket, so that all the necessary checks
6254                  * are made in the function processing timewait state.
6255                  */
6256                 if (tcp_death_row.sysctl_tw_recycle) {
6257                         bool strict;
6258 
6259                         dst = af_ops->route_req(sk, &fl, req, &strict);
6260 
6261                         if (dst && strict &&
6262                             !tcp_peer_is_proven(req, dst, true,
6263                                                 tmp_opt.saw_tstamp)) {
6264                                 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
6265                                 goto drop_and_release;
6266                         }
6267                 }
6268                 /* Kill the following clause, if you dislike this way. */
6269                 else if (!net->ipv4.sysctl_tcp_syncookies &&
6270                          (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6271                           (sysctl_max_syn_backlog >> 2)) &&
6272                          !tcp_peer_is_proven(req, dst, false,
6273                                              tmp_opt.saw_tstamp)) {
6274                         /* Without syncookies last quarter of
6275                          * backlog is filled with destinations,
6276                          * proven to be alive.
6277                          * It means that we continue to communicate
6278                          * to destinations, already remembered
6279                          * to the moment of synflood.
6280                          */
6281                         pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6282                                     rsk_ops->family);
6283                         goto drop_and_release;
6284                 }
6285 
6286                 isn = af_ops->init_seq(skb);
6287         }
6288         if (!dst) {
6289                 dst = af_ops->route_req(sk, &fl, req, NULL);
6290                 if (!dst)
6291                         goto drop_and_free;
6292         }
6293 
6294         tcp_ecn_create_request(req, skb, sk, dst);
6295 
6296         if (want_cookie) {
6297                 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6298                 req->cookie_ts = tmp_opt.tstamp_ok;
6299                 if (!tmp_opt.tstamp_ok)
6300                         inet_rsk(req)->ecn_ok = 0;
6301         }
6302 
6303         tcp_rsk(req)->snt_isn = isn;
6304         tcp_rsk(req)->txhash = net_tx_rndhash();
6305         tcp_openreq_init_rwin(req, sk, dst);
6306         if (!want_cookie) {
6307                 tcp_reqsk_record_syn(sk, req, skb);
6308                 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6309         }
6310         if (fastopen_sk) {
6311                 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6312                                     &foc, TCP_SYNACK_FASTOPEN);
6313                 /* Add the child socket directly into the accept queue */
6314                 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6315                 sk->sk_data_ready(sk);
6316                 bh_unlock_sock(fastopen_sk);
6317                 sock_put(fastopen_sk);
6318         } else {
6319                 tcp_rsk(req)->tfo_listener = false;
6320                 if (!want_cookie)
6321                         inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6322                 af_ops->send_synack(sk, dst, &fl, req, &foc,
6323                                     !want_cookie ? TCP_SYNACK_NORMAL :
6324                                                    TCP_SYNACK_COOKIE);
6325                 if (want_cookie) {
6326                         reqsk_free(req);
6327                         return 0;
6328                 }
6329         }
6330         reqsk_put(req);
6331         return 0;
6332 
6333 drop_and_release:
6334         dst_release(dst);
6335 drop_and_free:
6336         reqsk_free(req);
6337 drop:
6338         tcp_listendrop(sk);
6339         return 0;
6340 }
6341 EXPORT_SYMBOL(tcp_conn_request);
6342 

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