Version:  2.0.40 2.2.26 2.4.37 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 4.8

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                 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2333                          msg,
2334                          &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2335                          tp->snd_cwnd, tcp_left_out(tp),
2336                          tp->snd_ssthresh, tp->prior_ssthresh,
2337                          tp->packets_out);
2338         }
2339 #endif
2340 }
2341 #else
2342 #define DBGUNDO(x...) do { } while (0)
2343 #endif
2344 
2345 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2346 {
2347         struct tcp_sock *tp = tcp_sk(sk);
2348 
2349         if (unmark_loss) {
2350                 struct sk_buff *skb;
2351 
2352                 tcp_for_write_queue(skb, sk) {
2353                         if (skb == tcp_send_head(sk))
2354                                 break;
2355                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2356                 }
2357                 tp->lost_out = 0;
2358                 tcp_clear_all_retrans_hints(tp);
2359         }
2360 
2361         if (tp->prior_ssthresh) {
2362                 const struct inet_connection_sock *icsk = inet_csk(sk);
2363 
2364                 if (icsk->icsk_ca_ops->undo_cwnd)
2365                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2366                 else
2367                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2368 
2369                 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2370                         tp->snd_ssthresh = tp->prior_ssthresh;
2371                         tcp_ecn_withdraw_cwr(tp);
2372                 }
2373         }
2374         tp->snd_cwnd_stamp = tcp_time_stamp;
2375         tp->undo_marker = 0;
2376 }
2377 
2378 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2379 {
2380         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2381 }
2382 
2383 /* People celebrate: "We love our President!" */
2384 static bool tcp_try_undo_recovery(struct sock *sk)
2385 {
2386         struct tcp_sock *tp = tcp_sk(sk);
2387 
2388         if (tcp_may_undo(tp)) {
2389                 int mib_idx;
2390 
2391                 /* Happy end! We did not retransmit anything
2392                  * or our original transmission succeeded.
2393                  */
2394                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2395                 tcp_undo_cwnd_reduction(sk, false);
2396                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2397                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2398                 else
2399                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2400 
2401                 NET_INC_STATS(sock_net(sk), mib_idx);
2402         }
2403         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2404                 /* Hold old state until something *above* high_seq
2405                  * is ACKed. For Reno it is MUST to prevent false
2406                  * fast retransmits (RFC2582). SACK TCP is safe. */
2407                 if (!tcp_any_retrans_done(sk))
2408                         tp->retrans_stamp = 0;
2409                 return true;
2410         }
2411         tcp_set_ca_state(sk, TCP_CA_Open);
2412         return false;
2413 }
2414 
2415 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2416 static bool tcp_try_undo_dsack(struct sock *sk)
2417 {
2418         struct tcp_sock *tp = tcp_sk(sk);
2419 
2420         if (tp->undo_marker && !tp->undo_retrans) {
2421                 DBGUNDO(sk, "D-SACK");
2422                 tcp_undo_cwnd_reduction(sk, false);
2423                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2424                 return true;
2425         }
2426         return false;
2427 }
2428 
2429 /* Undo during loss recovery after partial ACK or using F-RTO. */
2430 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2431 {
2432         struct tcp_sock *tp = tcp_sk(sk);
2433 
2434         if (frto_undo || tcp_may_undo(tp)) {
2435                 tcp_undo_cwnd_reduction(sk, true);
2436 
2437                 DBGUNDO(sk, "partial loss");
2438                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2439                 if (frto_undo)
2440                         NET_INC_STATS(sock_net(sk),
2441                                         LINUX_MIB_TCPSPURIOUSRTOS);
2442                 inet_csk(sk)->icsk_retransmits = 0;
2443                 if (frto_undo || tcp_is_sack(tp))
2444                         tcp_set_ca_state(sk, TCP_CA_Open);
2445                 return true;
2446         }
2447         return false;
2448 }
2449 
2450 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2451  * It computes the number of packets to send (sndcnt) based on packets newly
2452  * delivered:
2453  *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2454  *      cwnd reductions across a full RTT.
2455  *   2) Otherwise PRR uses packet conservation to send as much as delivered.
2456  *      But when the retransmits are acked without further losses, PRR
2457  *      slow starts cwnd up to ssthresh to speed up the recovery.
2458  */
2459 static void tcp_init_cwnd_reduction(struct sock *sk)
2460 {
2461         struct tcp_sock *tp = tcp_sk(sk);
2462 
2463         tp->high_seq = tp->snd_nxt;
2464         tp->tlp_high_seq = 0;
2465         tp->snd_cwnd_cnt = 0;
2466         tp->prior_cwnd = tp->snd_cwnd;
2467         tp->prr_delivered = 0;
2468         tp->prr_out = 0;
2469         tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2470         tcp_ecn_queue_cwr(tp);
2471 }
2472 
2473 static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked,
2474                                int flag)
2475 {
2476         struct tcp_sock *tp = tcp_sk(sk);
2477         int sndcnt = 0;
2478         int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2479 
2480         if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2481                 return;
2482 
2483         tp->prr_delivered += newly_acked_sacked;
2484         if (delta < 0) {
2485                 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2486                                tp->prior_cwnd - 1;
2487                 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2488         } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2489                    !(flag & FLAG_LOST_RETRANS)) {
2490                 sndcnt = min_t(int, delta,
2491                                max_t(int, tp->prr_delivered - tp->prr_out,
2492                                      newly_acked_sacked) + 1);
2493         } else {
2494                 sndcnt = min(delta, newly_acked_sacked);
2495         }
2496         /* Force a fast retransmit upon entering fast recovery */
2497         sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2498         tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2499 }
2500 
2501 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2502 {
2503         struct tcp_sock *tp = tcp_sk(sk);
2504 
2505         /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2506         if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2507             (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2508                 tp->snd_cwnd = tp->snd_ssthresh;
2509                 tp->snd_cwnd_stamp = tcp_time_stamp;
2510         }
2511         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2512 }
2513 
2514 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2515 void tcp_enter_cwr(struct sock *sk)
2516 {
2517         struct tcp_sock *tp = tcp_sk(sk);
2518 
2519         tp->prior_ssthresh = 0;
2520         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2521                 tp->undo_marker = 0;
2522                 tcp_init_cwnd_reduction(sk);
2523                 tcp_set_ca_state(sk, TCP_CA_CWR);
2524         }
2525 }
2526 EXPORT_SYMBOL(tcp_enter_cwr);
2527 
2528 static void tcp_try_keep_open(struct sock *sk)
2529 {
2530         struct tcp_sock *tp = tcp_sk(sk);
2531         int state = TCP_CA_Open;
2532 
2533         if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2534                 state = TCP_CA_Disorder;
2535 
2536         if (inet_csk(sk)->icsk_ca_state != state) {
2537                 tcp_set_ca_state(sk, state);
2538                 tp->high_seq = tp->snd_nxt;
2539         }
2540 }
2541 
2542 static void tcp_try_to_open(struct sock *sk, int flag)
2543 {
2544         struct tcp_sock *tp = tcp_sk(sk);
2545 
2546         tcp_verify_left_out(tp);
2547 
2548         if (!tcp_any_retrans_done(sk))
2549                 tp->retrans_stamp = 0;
2550 
2551         if (flag & FLAG_ECE)
2552                 tcp_enter_cwr(sk);
2553 
2554         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2555                 tcp_try_keep_open(sk);
2556         }
2557 }
2558 
2559 static void tcp_mtup_probe_failed(struct sock *sk)
2560 {
2561         struct inet_connection_sock *icsk = inet_csk(sk);
2562 
2563         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2564         icsk->icsk_mtup.probe_size = 0;
2565         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2566 }
2567 
2568 static void tcp_mtup_probe_success(struct sock *sk)
2569 {
2570         struct tcp_sock *tp = tcp_sk(sk);
2571         struct inet_connection_sock *icsk = inet_csk(sk);
2572 
2573         /* FIXME: breaks with very large cwnd */
2574         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2575         tp->snd_cwnd = tp->snd_cwnd *
2576                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2577                        icsk->icsk_mtup.probe_size;
2578         tp->snd_cwnd_cnt = 0;
2579         tp->snd_cwnd_stamp = tcp_time_stamp;
2580         tp->snd_ssthresh = tcp_current_ssthresh(sk);
2581 
2582         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2583         icsk->icsk_mtup.probe_size = 0;
2584         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2585         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2586 }
2587 
2588 /* Do a simple retransmit without using the backoff mechanisms in
2589  * tcp_timer. This is used for path mtu discovery.
2590  * The socket is already locked here.
2591  */
2592 void tcp_simple_retransmit(struct sock *sk)
2593 {
2594         const struct inet_connection_sock *icsk = inet_csk(sk);
2595         struct tcp_sock *tp = tcp_sk(sk);
2596         struct sk_buff *skb;
2597         unsigned int mss = tcp_current_mss(sk);
2598         u32 prior_lost = tp->lost_out;
2599 
2600         tcp_for_write_queue(skb, sk) {
2601                 if (skb == tcp_send_head(sk))
2602                         break;
2603                 if (tcp_skb_seglen(skb) > mss &&
2604                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2605                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2606                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2607                                 tp->retrans_out -= tcp_skb_pcount(skb);
2608                         }
2609                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2610                 }
2611         }
2612 
2613         tcp_clear_retrans_hints_partial(tp);
2614 
2615         if (prior_lost == tp->lost_out)
2616                 return;
2617 
2618         if (tcp_is_reno(tp))
2619                 tcp_limit_reno_sacked(tp);
2620 
2621         tcp_verify_left_out(tp);
2622 
2623         /* Don't muck with the congestion window here.
2624          * Reason is that we do not increase amount of _data_
2625          * in network, but units changed and effective
2626          * cwnd/ssthresh really reduced now.
2627          */
2628         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2629                 tp->high_seq = tp->snd_nxt;
2630                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2631                 tp->prior_ssthresh = 0;
2632                 tp->undo_marker = 0;
2633                 tcp_set_ca_state(sk, TCP_CA_Loss);
2634         }
2635         tcp_xmit_retransmit_queue(sk);
2636 }
2637 EXPORT_SYMBOL(tcp_simple_retransmit);
2638 
2639 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2640 {
2641         struct tcp_sock *tp = tcp_sk(sk);
2642         int mib_idx;
2643 
2644         if (tcp_is_reno(tp))
2645                 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2646         else
2647                 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2648 
2649         NET_INC_STATS(sock_net(sk), mib_idx);
2650 
2651         tp->prior_ssthresh = 0;
2652         tcp_init_undo(tp);
2653 
2654         if (!tcp_in_cwnd_reduction(sk)) {
2655                 if (!ece_ack)
2656                         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2657                 tcp_init_cwnd_reduction(sk);
2658         }
2659         tcp_set_ca_state(sk, TCP_CA_Recovery);
2660 }
2661 
2662 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2663  * recovered or spurious. Otherwise retransmits more on partial ACKs.
2664  */
2665 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2666                              int *rexmit)
2667 {
2668         struct tcp_sock *tp = tcp_sk(sk);
2669         bool recovered = !before(tp->snd_una, tp->high_seq);
2670 
2671         if ((flag & FLAG_SND_UNA_ADVANCED) &&
2672             tcp_try_undo_loss(sk, false))
2673                 return;
2674 
2675         if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2676                 /* Step 3.b. A timeout is spurious if not all data are
2677                  * lost, i.e., never-retransmitted data are (s)acked.
2678                  */
2679                 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2680                     tcp_try_undo_loss(sk, true))
2681                         return;
2682 
2683                 if (after(tp->snd_nxt, tp->high_seq)) {
2684                         if (flag & FLAG_DATA_SACKED || is_dupack)
2685                                 tp->frto = 0; /* Step 3.a. loss was real */
2686                 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2687                         tp->high_seq = tp->snd_nxt;
2688                         /* Step 2.b. Try send new data (but deferred until cwnd
2689                          * is updated in tcp_ack()). Otherwise fall back to
2690                          * the conventional recovery.
2691                          */
2692                         if (tcp_send_head(sk) &&
2693                             after(tcp_wnd_end(tp), tp->snd_nxt)) {
2694                                 *rexmit = REXMIT_NEW;
2695                                 return;
2696                         }
2697                         tp->frto = 0;
2698                 }
2699         }
2700 
2701         if (recovered) {
2702                 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2703                 tcp_try_undo_recovery(sk);
2704                 return;
2705         }
2706         if (tcp_is_reno(tp)) {
2707                 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2708                  * delivered. Lower inflight to clock out (re)tranmissions.
2709                  */
2710                 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2711                         tcp_add_reno_sack(sk);
2712                 else if (flag & FLAG_SND_UNA_ADVANCED)
2713                         tcp_reset_reno_sack(tp);
2714         }
2715         *rexmit = REXMIT_LOST;
2716 }
2717 
2718 /* Undo during fast recovery after partial ACK. */
2719 static bool tcp_try_undo_partial(struct sock *sk, const int acked)
2720 {
2721         struct tcp_sock *tp = tcp_sk(sk);
2722 
2723         if (tp->undo_marker && tcp_packet_delayed(tp)) {
2724                 /* Plain luck! Hole if filled with delayed
2725                  * packet, rather than with a retransmit.
2726                  */
2727                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2728 
2729                 /* We are getting evidence that the reordering degree is higher
2730                  * than we realized. If there are no retransmits out then we
2731                  * can undo. Otherwise we clock out new packets but do not
2732                  * mark more packets lost or retransmit more.
2733                  */
2734                 if (tp->retrans_out)
2735                         return true;
2736 
2737                 if (!tcp_any_retrans_done(sk))
2738                         tp->retrans_stamp = 0;
2739 
2740                 DBGUNDO(sk, "partial recovery");
2741                 tcp_undo_cwnd_reduction(sk, true);
2742                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2743                 tcp_try_keep_open(sk);
2744                 return true;
2745         }
2746         return false;
2747 }
2748 
2749 /* Process an event, which can update packets-in-flight not trivially.
2750  * Main goal of this function is to calculate new estimate for left_out,
2751  * taking into account both packets sitting in receiver's buffer and
2752  * packets lost by network.
2753  *
2754  * Besides that it updates the congestion state when packet loss or ECN
2755  * is detected. But it does not reduce the cwnd, it is done by the
2756  * congestion control later.
2757  *
2758  * It does _not_ decide what to send, it is made in function
2759  * tcp_xmit_retransmit_queue().
2760  */
2761 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2762                                   bool is_dupack, int *ack_flag, int *rexmit)
2763 {
2764         struct inet_connection_sock *icsk = inet_csk(sk);
2765         struct tcp_sock *tp = tcp_sk(sk);
2766         int fast_rexmit = 0, flag = *ack_flag;
2767         bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2768                                     (tcp_fackets_out(tp) > tp->reordering));
2769 
2770         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2771                 tp->sacked_out = 0;
2772         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2773                 tp->fackets_out = 0;
2774 
2775         /* Now state machine starts.
2776          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2777         if (flag & FLAG_ECE)
2778                 tp->prior_ssthresh = 0;
2779 
2780         /* B. In all the states check for reneging SACKs. */
2781         if (tcp_check_sack_reneging(sk, flag))
2782                 return;
2783 
2784         /* C. Check consistency of the current state. */
2785         tcp_verify_left_out(tp);
2786 
2787         /* D. Check state exit conditions. State can be terminated
2788          *    when high_seq is ACKed. */
2789         if (icsk->icsk_ca_state == TCP_CA_Open) {
2790                 WARN_ON(tp->retrans_out != 0);
2791                 tp->retrans_stamp = 0;
2792         } else if (!before(tp->snd_una, tp->high_seq)) {
2793                 switch (icsk->icsk_ca_state) {
2794                 case TCP_CA_CWR:
2795                         /* CWR is to be held something *above* high_seq
2796                          * is ACKed for CWR bit to reach receiver. */
2797                         if (tp->snd_una != tp->high_seq) {
2798                                 tcp_end_cwnd_reduction(sk);
2799                                 tcp_set_ca_state(sk, TCP_CA_Open);
2800                         }
2801                         break;
2802 
2803                 case TCP_CA_Recovery:
2804                         if (tcp_is_reno(tp))
2805                                 tcp_reset_reno_sack(tp);
2806                         if (tcp_try_undo_recovery(sk))
2807                                 return;
2808                         tcp_end_cwnd_reduction(sk);
2809                         break;
2810                 }
2811         }
2812 
2813         /* Use RACK to detect loss */
2814         if (sysctl_tcp_recovery & TCP_RACK_LOST_RETRANS &&
2815             tcp_rack_mark_lost(sk)) {
2816                 flag |= FLAG_LOST_RETRANS;
2817                 *ack_flag |= FLAG_LOST_RETRANS;
2818         }
2819 
2820         /* E. Process state. */
2821         switch (icsk->icsk_ca_state) {
2822         case TCP_CA_Recovery:
2823                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2824                         if (tcp_is_reno(tp) && is_dupack)
2825                                 tcp_add_reno_sack(sk);
2826                 } else {
2827                         if (tcp_try_undo_partial(sk, acked))
2828                                 return;
2829                         /* Partial ACK arrived. Force fast retransmit. */
2830                         do_lost = tcp_is_reno(tp) ||
2831                                   tcp_fackets_out(tp) > tp->reordering;
2832                 }
2833                 if (tcp_try_undo_dsack(sk)) {
2834                         tcp_try_keep_open(sk);
2835                         return;
2836                 }
2837                 break;
2838         case TCP_CA_Loss:
2839                 tcp_process_loss(sk, flag, is_dupack, rexmit);
2840                 if (icsk->icsk_ca_state != TCP_CA_Open &&
2841                     !(flag & FLAG_LOST_RETRANS))
2842                         return;
2843                 /* Change state if cwnd is undone or retransmits are lost */
2844         default:
2845                 if (tcp_is_reno(tp)) {
2846                         if (flag & FLAG_SND_UNA_ADVANCED)
2847                                 tcp_reset_reno_sack(tp);
2848                         if (is_dupack)
2849                                 tcp_add_reno_sack(sk);
2850                 }
2851 
2852                 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2853                         tcp_try_undo_dsack(sk);
2854 
2855                 if (!tcp_time_to_recover(sk, flag)) {
2856                         tcp_try_to_open(sk, flag);
2857                         return;
2858                 }
2859 
2860                 /* MTU probe failure: don't reduce cwnd */
2861                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2862                     icsk->icsk_mtup.probe_size &&
2863                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2864                         tcp_mtup_probe_failed(sk);
2865                         /* Restores the reduction we did in tcp_mtup_probe() */
2866                         tp->snd_cwnd++;
2867                         tcp_simple_retransmit(sk);
2868                         return;
2869                 }
2870 
2871                 /* Otherwise enter Recovery state */
2872                 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2873                 fast_rexmit = 1;
2874         }
2875 
2876         if (do_lost)
2877                 tcp_update_scoreboard(sk, fast_rexmit);
2878         *rexmit = REXMIT_LOST;
2879 }
2880 
2881 /* Kathleen Nichols' algorithm for tracking the minimum value of
2882  * a data stream over some fixed time interval. (E.g., the minimum
2883  * RTT over the past five minutes.) It uses constant space and constant
2884  * time per update yet almost always delivers the same minimum as an
2885  * implementation that has to keep all the data in the window.
2886  *
2887  * The algorithm keeps track of the best, 2nd best & 3rd best min
2888  * values, maintaining an invariant that the measurement time of the
2889  * n'th best >= n-1'th best. It also makes sure that the three values
2890  * are widely separated in the time window since that bounds the worse
2891  * case error when that data is monotonically increasing over the window.
2892  *
2893  * Upon getting a new min, we can forget everything earlier because it
2894  * has no value - the new min is <= everything else in the window by
2895  * definition and it's the most recent. So we restart fresh on every new min
2896  * and overwrites 2nd & 3rd choices. The same property holds for 2nd & 3rd
2897  * best.
2898  */
2899 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2900 {
2901         const u32 now = tcp_time_stamp, wlen = sysctl_tcp_min_rtt_wlen * HZ;
2902         struct rtt_meas *m = tcp_sk(sk)->rtt_min;
2903         struct rtt_meas rttm = {
2904                 .rtt = likely(rtt_us) ? rtt_us : jiffies_to_usecs(1),
2905                 .ts = now,
2906         };
2907         u32 elapsed;
2908 
2909         /* Check if the new measurement updates the 1st, 2nd, or 3rd choices */
2910         if (unlikely(rttm.rtt <= m[0].rtt))
2911                 m[0] = m[1] = m[2] = rttm;
2912         else if (rttm.rtt <= m[1].rtt)
2913                 m[1] = m[2] = rttm;
2914         else if (rttm.rtt <= m[2].rtt)
2915                 m[2] = rttm;
2916 
2917         elapsed = now - m[0].ts;
2918         if (unlikely(elapsed > wlen)) {
2919                 /* Passed entire window without a new min so make 2nd choice
2920                  * the new min & 3rd choice the new 2nd. So forth and so on.
2921                  */
2922                 m[0] = m[1];
2923                 m[1] = m[2];
2924                 m[2] = rttm;
2925                 if (now - m[0].ts > wlen) {
2926                         m[0] = m[1];
2927                         m[1] = rttm;
2928                         if (now - m[0].ts > wlen)
2929                                 m[0] = rttm;
2930                 }
2931         } else if (m[1].ts == m[0].ts && elapsed > wlen / 4) {
2932                 /* Passed a quarter of the window without a new min so
2933                  * take 2nd choice from the 2nd quarter of the window.
2934                  */
2935                 m[2] = m[1] = rttm;
2936         } else if (m[2].ts == m[1].ts && elapsed > wlen / 2) {
2937                 /* Passed half the window without a new min so take the 3rd
2938                  * choice from the last half of the window.
2939                  */
2940                 m[2] = rttm;
2941         }
2942 }
2943 
2944 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2945                                       long seq_rtt_us, long sack_rtt_us,
2946                                       long ca_rtt_us)
2947 {
2948         const struct tcp_sock *tp = tcp_sk(sk);
2949 
2950         /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2951          * broken middle-boxes or peers may corrupt TS-ECR fields. But
2952          * Karn's algorithm forbids taking RTT if some retransmitted data
2953          * is acked (RFC6298).
2954          */
2955         if (seq_rtt_us < 0)
2956                 seq_rtt_us = sack_rtt_us;
2957 
2958         /* RTTM Rule: A TSecr value received in a segment is used to
2959          * update the averaged RTT measurement only if the segment
2960          * acknowledges some new data, i.e., only if it advances the
2961          * left edge of the send window.
2962          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2963          */
2964         if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2965             flag & FLAG_ACKED)
2966                 seq_rtt_us = ca_rtt_us = jiffies_to_usecs(tcp_time_stamp -
2967                                                           tp->rx_opt.rcv_tsecr);
2968         if (seq_rtt_us < 0)
2969                 return false;
2970 
2971         /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2972          * always taken together with ACK, SACK, or TS-opts. Any negative
2973          * values will be skipped with the seq_rtt_us < 0 check above.
2974          */
2975         tcp_update_rtt_min(sk, ca_rtt_us);
2976         tcp_rtt_estimator(sk, seq_rtt_us);
2977         tcp_set_rto(sk);
2978 
2979         /* RFC6298: only reset backoff on valid RTT measurement. */
2980         inet_csk(sk)->icsk_backoff = 0;
2981         return true;
2982 }
2983 
2984 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2985 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2986 {
2987         long rtt_us = -1L;
2988 
2989         if (req && !req->num_retrans && tcp_rsk(req)->snt_synack.v64) {
2990                 struct skb_mstamp now;
2991 
2992                 skb_mstamp_get(&now);
2993                 rtt_us = skb_mstamp_us_delta(&now, &tcp_rsk(req)->snt_synack);
2994         }
2995 
2996         tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us);
2997 }
2998 
2999 
3000 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3001 {
3002         const struct inet_connection_sock *icsk = inet_csk(sk);
3003 
3004         icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3005         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3006 }
3007 
3008 /* Restart timer after forward progress on connection.
3009  * RFC2988 recommends to restart timer to now+rto.
3010  */
3011 void tcp_rearm_rto(struct sock *sk)
3012 {
3013         const struct inet_connection_sock *icsk = inet_csk(sk);
3014         struct tcp_sock *tp = tcp_sk(sk);
3015 
3016         /* If the retrans timer is currently being used by Fast Open
3017          * for SYN-ACK retrans purpose, stay put.
3018          */
3019         if (tp->fastopen_rsk)
3020                 return;
3021 
3022         if (!tp->packets_out) {
3023                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3024         } else {
3025                 u32 rto = inet_csk(sk)->icsk_rto;
3026                 /* Offset the time elapsed after installing regular RTO */
3027                 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3028                     icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3029                         struct sk_buff *skb = tcp_write_queue_head(sk);
3030                         const u32 rto_time_stamp =
3031                                 tcp_skb_timestamp(skb) + rto;
3032                         s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3033                         /* delta may not be positive if the socket is locked
3034                          * when the retrans timer fires and is rescheduled.
3035                          */
3036                         if (delta > 0)
3037                                 rto = delta;
3038                 }
3039                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3040                                           TCP_RTO_MAX);
3041         }
3042 }
3043 
3044 /* This function is called when the delayed ER timer fires. TCP enters
3045  * fast recovery and performs fast-retransmit.
3046  */
3047 void tcp_resume_early_retransmit(struct sock *sk)
3048 {
3049         struct tcp_sock *tp = tcp_sk(sk);
3050 
3051         tcp_rearm_rto(sk);
3052 
3053         /* Stop if ER is disabled after the delayed ER timer is scheduled */
3054         if (!tp->do_early_retrans)
3055                 return;
3056 
3057         tcp_enter_recovery(sk, false);
3058         tcp_update_scoreboard(sk, 1);
3059         tcp_xmit_retransmit_queue(sk);
3060 }
3061 
3062 /* If we get here, the whole TSO packet has not been acked. */
3063 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3064 {
3065         struct tcp_sock *tp = tcp_sk(sk);
3066         u32 packets_acked;
3067 
3068         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3069 
3070         packets_acked = tcp_skb_pcount(skb);
3071         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3072                 return 0;
3073         packets_acked -= tcp_skb_pcount(skb);
3074 
3075         if (packets_acked) {
3076                 BUG_ON(tcp_skb_pcount(skb) == 0);
3077                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3078         }
3079 
3080         return packets_acked;
3081 }
3082 
3083 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3084                            u32 prior_snd_una)
3085 {
3086         const struct skb_shared_info *shinfo;
3087 
3088         /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3089         if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3090                 return;
3091 
3092         shinfo = skb_shinfo(skb);
3093         if (!before(shinfo->tskey, prior_snd_una) &&
3094             before(shinfo->tskey, tcp_sk(sk)->snd_una))
3095                 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3096 }
3097 
3098 /* Remove acknowledged frames from the retransmission queue. If our packet
3099  * is before the ack sequence we can discard it as it's confirmed to have
3100  * arrived at the other end.
3101  */
3102 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3103                                u32 prior_snd_una, int *acked,
3104                                struct tcp_sacktag_state *sack)
3105 {
3106         const struct inet_connection_sock *icsk = inet_csk(sk);
3107         struct skb_mstamp first_ackt, last_ackt, now;
3108         struct tcp_sock *tp = tcp_sk(sk);
3109         u32 prior_sacked = tp->sacked_out;
3110         u32 reord = tp->packets_out;
3111         bool fully_acked = true;
3112         long sack_rtt_us = -1L;
3113         long seq_rtt_us = -1L;
3114         long ca_rtt_us = -1L;
3115         struct sk_buff *skb;
3116         u32 pkts_acked = 0;
3117         u32 last_in_flight = 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                         last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3158                         reord = min(pkts_acked, reord);
3159                         if (!after(scb->end_seq, tp->high_seq))
3160                                 flag |= FLAG_ORIG_SACK_ACKED;
3161                 }
3162 
3163                 if (sacked & TCPCB_SACKED_ACKED) {
3164                         tp->sacked_out -= acked_pcount;
3165                 } else if (tcp_is_sack(tp)) {
3166                         tp->delivered += acked_pcount;
3167                         if (!tcp_skb_spurious_retrans(tp, skb))
3168                                 tcp_rack_advance(tp, &skb->skb_mstamp, sacked);
3169                 }
3170                 if (sacked & TCPCB_LOST)
3171                         tp->lost_out -= acked_pcount;
3172 
3173                 tp->packets_out -= acked_pcount;
3174                 pkts_acked += acked_pcount;
3175 
3176                 /* Initial outgoing SYN's get put onto the write_queue
3177                  * just like anything else we transmit.  It is not
3178                  * true data, and if we misinform our callers that
3179                  * this ACK acks real data, we will erroneously exit
3180                  * connection startup slow start one packet too
3181                  * quickly.  This is severely frowned upon behavior.
3182                  */
3183                 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3184                         flag |= FLAG_DATA_ACKED;
3185                 } else {
3186                         flag |= FLAG_SYN_ACKED;
3187                         tp->retrans_stamp = 0;
3188                 }
3189 
3190                 if (!fully_acked)
3191                         break;
3192 
3193                 tcp_unlink_write_queue(skb, sk);
3194                 sk_wmem_free_skb(sk, skb);
3195                 if (unlikely(skb == tp->retransmit_skb_hint))
3196                         tp->retransmit_skb_hint = NULL;
3197                 if (unlikely(skb == tp->lost_skb_hint))
3198                         tp->lost_skb_hint = NULL;
3199         }
3200 
3201         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3202                 tp->snd_up = tp->snd_una;
3203 
3204         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3205                 flag |= FLAG_SACK_RENEGING;
3206 
3207         skb_mstamp_get(&now);
3208         if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3209                 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3210                 ca_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3211         }
3212         if (sack->first_sackt.v64) {
3213                 sack_rtt_us = skb_mstamp_us_delta(&now, &sack->first_sackt);
3214                 ca_rtt_us = skb_mstamp_us_delta(&now, &sack->last_sackt);
3215         }
3216 
3217         rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3218                                         ca_rtt_us);
3219 
3220         if (flag & FLAG_ACKED) {
3221                 tcp_rearm_rto(sk);
3222                 if (unlikely(icsk->icsk_mtup.probe_size &&
3223                              !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3224                         tcp_mtup_probe_success(sk);
3225                 }
3226 
3227                 if (tcp_is_reno(tp)) {
3228                         tcp_remove_reno_sacks(sk, pkts_acked);
3229                 } else {
3230                         int delta;
3231 
3232                         /* Non-retransmitted hole got filled? That's reordering */
3233                         if (reord < prior_fackets)
3234                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3235 
3236                         delta = tcp_is_fack(tp) ? pkts_acked :
3237                                                   prior_sacked - tp->sacked_out;
3238                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3239                 }
3240 
3241                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3242 
3243         } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3244                    sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3245                 /* Do not re-arm RTO if the sack RTT is measured from data sent
3246                  * after when the head was last (re)transmitted. Otherwise the
3247                  * timeout may continue to extend in loss recovery.
3248                  */
3249                 tcp_rearm_rto(sk);
3250         }
3251 
3252         if (icsk->icsk_ca_ops->pkts_acked) {
3253                 struct ack_sample sample = { .pkts_acked = pkts_acked,
3254                                              .rtt_us = ca_rtt_us,
3255                                              .in_flight = last_in_flight };
3256 
3257                 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3258         }
3259 
3260 #if FASTRETRANS_DEBUG > 0
3261         WARN_ON((int)tp->sacked_out < 0);
3262         WARN_ON((int)tp->lost_out < 0);
3263         WARN_ON((int)tp->retrans_out < 0);
3264         if (!tp->packets_out && tcp_is_sack(tp)) {
3265                 icsk = inet_csk(sk);
3266                 if (tp->lost_out) {
3267                         pr_debug("Leak l=%u %d\n",
3268                                  tp->lost_out, icsk->icsk_ca_state);
3269                         tp->lost_out = 0;
3270                 }
3271                 if (tp->sacked_out) {
3272                         pr_debug("Leak s=%u %d\n",
3273                                  tp->sacked_out, icsk->icsk_ca_state);
3274                         tp->sacked_out = 0;
3275                 }
3276                 if (tp->retrans_out) {
3277                         pr_debug("Leak r=%u %d\n",
3278                                  tp->retrans_out, icsk->icsk_ca_state);
3279                         tp->retrans_out = 0;
3280                 }
3281         }
3282 #endif
3283         *acked = pkts_acked;
3284         return flag;
3285 }
3286 
3287 static void tcp_ack_probe(struct sock *sk)
3288 {
3289         const struct tcp_sock *tp = tcp_sk(sk);
3290         struct inet_connection_sock *icsk = inet_csk(sk);
3291 
3292         /* Was it a usable window open? */
3293 
3294         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3295                 icsk->icsk_backoff = 0;
3296                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3297                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3298                  * This function is not for random using!
3299                  */
3300         } else {
3301                 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3302 
3303                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3304                                           when, TCP_RTO_MAX);
3305         }
3306 }
3307 
3308 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3309 {
3310         return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3311                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3312 }
3313 
3314 /* Decide wheather to run the increase function of congestion control. */
3315 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3316 {
3317         /* If reordering is high then always grow cwnd whenever data is
3318          * delivered regardless of its ordering. Otherwise stay conservative
3319          * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3320          * new SACK or ECE mark may first advance cwnd here and later reduce
3321          * cwnd in tcp_fastretrans_alert() based on more states.
3322          */
3323         if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3324                 return flag & FLAG_FORWARD_PROGRESS;
3325 
3326         return flag & FLAG_DATA_ACKED;
3327 }
3328 
3329 /* The "ultimate" congestion control function that aims to replace the rigid
3330  * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3331  * It's called toward the end of processing an ACK with precise rate
3332  * information. All transmission or retransmission are delayed afterwards.
3333  */
3334 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3335                              int flag)
3336 {
3337         if (tcp_in_cwnd_reduction(sk)) {
3338                 /* Reduce cwnd if state mandates */
3339                 tcp_cwnd_reduction(sk, acked_sacked, flag);
3340         } else if (tcp_may_raise_cwnd(sk, flag)) {
3341                 /* Advance cwnd if state allows */
3342                 tcp_cong_avoid(sk, ack, acked_sacked);
3343         }
3344         tcp_update_pacing_rate(sk);
3345 }
3346 
3347 /* Check that window update is acceptable.
3348  * The function assumes that snd_una<=ack<=snd_next.
3349  */
3350 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3351                                         const u32 ack, const u32 ack_seq,
3352                                         const u32 nwin)
3353 {
3354         return  after(ack, tp->snd_una) ||
3355                 after(ack_seq, tp->snd_wl1) ||
3356                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3357 }
3358 
3359 /* If we update tp->snd_una, also update tp->bytes_acked */
3360 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3361 {
3362         u32 delta = ack - tp->snd_una;
3363 
3364         sock_owned_by_me((struct sock *)tp);
3365         u64_stats_update_begin_raw(&tp->syncp);
3366         tp->bytes_acked += delta;
3367         u64_stats_update_end_raw(&tp->syncp);
3368         tp->snd_una = ack;
3369 }
3370 
3371 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3372 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3373 {
3374         u32 delta = seq - tp->rcv_nxt;
3375 
3376         sock_owned_by_me((struct sock *)tp);
3377         u64_stats_update_begin_raw(&tp->syncp);
3378         tp->bytes_received += delta;
3379         u64_stats_update_end_raw(&tp->syncp);
3380         tp->rcv_nxt = seq;
3381 }
3382 
3383 /* Update our send window.
3384  *
3385  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3386  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3387  */
3388 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3389                                  u32 ack_seq)
3390 {
3391         struct tcp_sock *tp = tcp_sk(sk);
3392         int flag = 0;
3393         u32 nwin = ntohs(tcp_hdr(skb)->window);
3394 
3395         if (likely(!tcp_hdr(skb)->syn))
3396                 nwin <<= tp->rx_opt.snd_wscale;
3397 
3398         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3399                 flag |= FLAG_WIN_UPDATE;
3400                 tcp_update_wl(tp, ack_seq);
3401 
3402                 if (tp->snd_wnd != nwin) {
3403                         tp->snd_wnd = nwin;
3404 
3405                         /* Note, it is the only place, where
3406                          * fast path is recovered for sending TCP.
3407                          */
3408                         tp->pred_flags = 0;
3409                         tcp_fast_path_check(sk);
3410 
3411                         if (tcp_send_head(sk))
3412                                 tcp_slow_start_after_idle_check(sk);
3413 
3414                         if (nwin > tp->max_window) {
3415                                 tp->max_window = nwin;
3416                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3417                         }
3418                 }
3419         }
3420 
3421         tcp_snd_una_update(tp, ack);
3422 
3423         return flag;
3424 }
3425 
3426 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3427                                    u32 *last_oow_ack_time)
3428 {
3429         if (*last_oow_ack_time) {
3430                 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3431 
3432                 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3433                         NET_INC_STATS(net, mib_idx);
3434                         return true;    /* rate-limited: don't send yet! */
3435                 }
3436         }
3437 
3438         *last_oow_ack_time = tcp_time_stamp;
3439 
3440         return false;   /* not rate-limited: go ahead, send dupack now! */
3441 }
3442 
3443 /* Return true if we're currently rate-limiting out-of-window ACKs and
3444  * thus shouldn't send a dupack right now. We rate-limit dupacks in
3445  * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3446  * attacks that send repeated SYNs or ACKs for the same connection. To
3447  * do this, we do not send a duplicate SYNACK or ACK if the remote
3448  * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3449  */
3450 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3451                           int mib_idx, u32 *last_oow_ack_time)
3452 {
3453         /* Data packets without SYNs are not likely part of an ACK loop. */
3454         if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3455             !tcp_hdr(skb)->syn)
3456                 return false;
3457 
3458         return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3459 }
3460 
3461 /* RFC 5961 7 [ACK Throttling] */
3462 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3463 {
3464         /* unprotected vars, we dont care of overwrites */
3465         static u32 challenge_timestamp;
3466         static unsigned int challenge_count;
3467         struct tcp_sock *tp = tcp_sk(sk);
3468         u32 count, now;
3469 
3470         /* First check our per-socket dupack rate limit. */
3471         if (__tcp_oow_rate_limited(sock_net(sk),
3472                                    LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3473                                    &tp->last_oow_ack_time))
3474                 return;
3475 
3476         /* Then check host-wide RFC 5961 rate limit. */
3477         now = jiffies / HZ;
3478         if (now != challenge_timestamp) {
3479                 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3480 
3481                 challenge_timestamp = now;
3482                 WRITE_ONCE(challenge_count, half +
3483                            prandom_u32_max(sysctl_tcp_challenge_ack_limit));
3484         }
3485         count = READ_ONCE(challenge_count);
3486         if (count > 0) {
3487                 WRITE_ONCE(challenge_count, count - 1);
3488                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3489                 tcp_send_ack(sk);
3490         }
3491 }
3492 
3493 static void tcp_store_ts_recent(struct tcp_sock *tp)
3494 {
3495         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3496         tp->rx_opt.ts_recent_stamp = get_seconds();
3497 }
3498 
3499 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3500 {
3501         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3502                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3503                  * extra check below makes sure this can only happen
3504                  * for pure ACK frames.  -DaveM
3505                  *
3506                  * Not only, also it occurs for expired timestamps.
3507                  */
3508 
3509                 if (tcp_paws_check(&tp->rx_opt, 0))
3510                         tcp_store_ts_recent(tp);
3511         }
3512 }
3513 
3514 /* This routine deals with acks during a TLP episode.
3515  * We mark the end of a TLP episode on receiving TLP dupack or when
3516  * ack is after tlp_high_seq.
3517  * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3518  */
3519 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3520 {
3521         struct tcp_sock *tp = tcp_sk(sk);
3522 
3523         if (before(ack, tp->tlp_high_seq))
3524                 return;
3525 
3526         if (flag & FLAG_DSACKING_ACK) {
3527                 /* This DSACK means original and TLP probe arrived; no loss */
3528                 tp->tlp_high_seq = 0;
3529         } else if (after(ack, tp->tlp_high_seq)) {
3530                 /* ACK advances: there was a loss, so reduce cwnd. Reset
3531                  * tlp_high_seq in tcp_init_cwnd_reduction()
3532                  */
3533                 tcp_init_cwnd_reduction(sk);
3534                 tcp_set_ca_state(sk, TCP_CA_CWR);
3535                 tcp_end_cwnd_reduction(sk);
3536                 tcp_try_keep_open(sk);
3537                 NET_INC_STATS(sock_net(sk),
3538                                 LINUX_MIB_TCPLOSSPROBERECOVERY);
3539         } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3540                              FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3541                 /* Pure dupack: original and TLP probe arrived; no loss */
3542                 tp->tlp_high_seq = 0;
3543         }
3544 }
3545 
3546 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3547 {
3548         const struct inet_connection_sock *icsk = inet_csk(sk);
3549 
3550         if (icsk->icsk_ca_ops->in_ack_event)
3551                 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3552 }
3553 
3554 /* Congestion control has updated the cwnd already. So if we're in
3555  * loss recovery then now we do any new sends (for FRTO) or
3556  * retransmits (for CA_Loss or CA_recovery) that make sense.
3557  */
3558 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3559 {
3560         struct tcp_sock *tp = tcp_sk(sk);
3561 
3562         if (rexmit == REXMIT_NONE)
3563                 return;
3564 
3565         if (unlikely(rexmit == 2)) {
3566                 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3567                                           TCP_NAGLE_OFF);
3568                 if (after(tp->snd_nxt, tp->high_seq))
3569                         return;
3570                 tp->frto = 0;
3571         }
3572         tcp_xmit_retransmit_queue(sk);
3573 }
3574 
3575 /* This routine deals with incoming acks, but not outgoing ones. */
3576 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3577 {
3578         struct inet_connection_sock *icsk = inet_csk(sk);
3579         struct tcp_sock *tp = tcp_sk(sk);
3580         struct tcp_sacktag_state sack_state;
3581         u32 prior_snd_una = tp->snd_una;
3582         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3583         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3584         bool is_dupack = false;
3585         u32 prior_fackets;
3586         int prior_packets = tp->packets_out;
3587         u32 prior_delivered = tp->delivered;
3588         int acked = 0; /* Number of packets newly acked */
3589         int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3590 
3591         sack_state.first_sackt.v64 = 0;
3592 
3593         /* We very likely will need to access write queue head. */
3594         prefetchw(sk->sk_write_queue.next);
3595 
3596         /* If the ack is older than previous acks
3597          * then we can probably ignore it.
3598          */
3599         if (before(ack, prior_snd_una)) {
3600                 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3601                 if (before(ack, prior_snd_una - tp->max_window)) {
3602                         tcp_send_challenge_ack(sk, skb);
3603                         return -1;
3604                 }
3605                 goto old_ack;
3606         }
3607 
3608         /* If the ack includes data we haven't sent yet, discard
3609          * this segment (RFC793 Section 3.9).
3610          */
3611         if (after(ack, tp->snd_nxt))
3612                 goto invalid_ack;
3613 
3614         if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3615             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3616                 tcp_rearm_rto(sk);
3617 
3618         if (after(ack, prior_snd_una)) {
3619                 flag |= FLAG_SND_UNA_ADVANCED;
3620                 icsk->icsk_retransmits = 0;
3621         }
3622 
3623         prior_fackets = tp->fackets_out;
3624 
3625         /* ts_recent update must be made after we are sure that the packet
3626          * is in window.
3627          */
3628         if (flag & FLAG_UPDATE_TS_RECENT)
3629                 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3630 
3631         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3632                 /* Window is constant, pure forward advance.
3633                  * No more checks are required.
3634                  * Note, we use the fact that SND.UNA>=SND.WL2.
3635                  */
3636                 tcp_update_wl(tp, ack_seq);
3637                 tcp_snd_una_update(tp, ack);
3638                 flag |= FLAG_WIN_UPDATE;
3639 
3640                 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3641 
3642                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3643         } else {
3644                 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3645 
3646                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3647                         flag |= FLAG_DATA;
3648                 else
3649                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3650 
3651                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3652 
3653                 if (TCP_SKB_CB(skb)->sacked)
3654                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3655                                                         &sack_state);
3656 
3657                 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3658                         flag |= FLAG_ECE;
3659                         ack_ev_flags |= CA_ACK_ECE;
3660                 }
3661 
3662                 if (flag & FLAG_WIN_UPDATE)
3663                         ack_ev_flags |= CA_ACK_WIN_UPDATE;
3664 
3665                 tcp_in_ack_event(sk, ack_ev_flags);
3666         }
3667 
3668         /* We passed data and got it acked, remove any soft error
3669          * log. Something worked...
3670          */
3671         sk->sk_err_soft = 0;
3672         icsk->icsk_probes_out = 0;
3673         tp->rcv_tstamp = tcp_time_stamp;
3674         if (!prior_packets)
3675                 goto no_queue;
3676 
3677         /* See if we can take anything off of the retransmit queue. */
3678         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, &acked,
3679                                     &sack_state);
3680 
3681         if (tcp_ack_is_dubious(sk, flag)) {
3682                 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3683                 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3684         }
3685         if (tp->tlp_high_seq)
3686                 tcp_process_tlp_ack(sk, ack, flag);
3687 
3688         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3689                 struct dst_entry *dst = __sk_dst_get(sk);
3690                 if (dst)
3691                         dst_confirm(dst);
3692         }
3693 
3694         if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3695                 tcp_schedule_loss_probe(sk);
3696         tcp_cong_control(sk, ack, tp->delivered - prior_delivered, flag);
3697         tcp_xmit_recovery(sk, rexmit);
3698         return 1;
3699 
3700 no_queue:
3701         /* If data was DSACKed, see if we can undo a cwnd reduction. */
3702         if (flag & FLAG_DSACKING_ACK)
3703                 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3704         /* If this ack opens up a zero window, clear backoff.  It was
3705          * being used to time the probes, and is probably far higher than
3706          * it needs to be for normal retransmission.
3707          */
3708         if (tcp_send_head(sk))
3709                 tcp_ack_probe(sk);
3710 
3711         if (tp->tlp_high_seq)
3712                 tcp_process_tlp_ack(sk, ack, flag);
3713         return 1;
3714 
3715 invalid_ack:
3716         SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3717         return -1;
3718 
3719 old_ack:
3720         /* If data was SACKed, tag it and see if we should send more data.
3721          * If data was DSACKed, see if we can undo a cwnd reduction.
3722          */
3723         if (TCP_SKB_CB(skb)->sacked) {
3724                 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3725                                                 &sack_state);
3726                 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3727                 tcp_xmit_recovery(sk, rexmit);
3728         }
3729 
3730         SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3731         return 0;
3732 }
3733 
3734 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3735                                       bool syn, struct tcp_fastopen_cookie *foc,
3736                                       bool exp_opt)
3737 {
3738         /* Valid only in SYN or SYN-ACK with an even length.  */
3739         if (!foc || !syn || len < 0 || (len & 1))
3740                 return;
3741 
3742         if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3743             len <= TCP_FASTOPEN_COOKIE_MAX)
3744                 memcpy(foc->val, cookie, len);
3745         else if (len != 0)
3746                 len = -1;
3747         foc->len = len;
3748         foc->exp = exp_opt;
3749 }
3750 
3751 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3752  * But, this can also be called on packets in the established flow when
3753  * the fast version below fails.
3754  */
3755 void tcp_parse_options(const struct sk_buff *skb,
3756                        struct tcp_options_received *opt_rx, int estab,
3757                        struct tcp_fastopen_cookie *foc)
3758 {
3759         const unsigned char *ptr;
3760         const struct tcphdr *th = tcp_hdr(skb);
3761         int length = (th->doff * 4) - sizeof(struct tcphdr);
3762 
3763         ptr = (const unsigned char *)(th + 1);
3764         opt_rx->saw_tstamp = 0;
3765 
3766         while (length > 0) {
3767                 int opcode = *ptr++;
3768                 int opsize;
3769 
3770                 switch (opcode) {
3771                 case TCPOPT_EOL:
3772                         return;
3773                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3774                         length--;
3775                         continue;
3776                 default:
3777                         opsize = *ptr++;
3778                         if (opsize < 2) /* "silly options" */
3779                                 return;
3780                         if (opsize > length)
3781                                 return; /* don't parse partial options */
3782                         switch (opcode) {
3783                         case TCPOPT_MSS:
3784                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3785                                         u16 in_mss = get_unaligned_be16(ptr);
3786                                         if (in_mss) {
3787                                                 if (opt_rx->user_mss &&
3788                                                     opt_rx->user_mss < in_mss)
3789                                                         in_mss = opt_rx->user_mss;
3790                                                 opt_rx->mss_clamp = in_mss;
3791                                         }
3792                                 }
3793                                 break;
3794                         case TCPOPT_WINDOW:
3795                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3796                                     !estab && sysctl_tcp_window_scaling) {
3797                                         __u8 snd_wscale = *(__u8 *)ptr;
3798                                         opt_rx->wscale_ok = 1;
3799                                         if (snd_wscale > 14) {
3800                                                 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3801                                                                      __func__,
3802                                                                      snd_wscale);
3803                                                 snd_wscale = 14;
3804                                         }
3805                                         opt_rx->snd_wscale = snd_wscale;
3806                                 }
3807                                 break;
3808                         case TCPOPT_TIMESTAMP:
3809                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3810                                     ((estab && opt_rx->tstamp_ok) ||
3811                                      (!estab && sysctl_tcp_timestamps))) {
3812                                         opt_rx->saw_tstamp = 1;
3813                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3814                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3815                                 }
3816                                 break;
3817                         case TCPOPT_SACK_PERM:
3818                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3819                                     !estab && sysctl_tcp_sack) {
3820                                         opt_rx->sack_ok = TCP_SACK_SEEN;
3821                                         tcp_sack_reset(opt_rx);
3822                                 }
3823                                 break;
3824 
3825                         case TCPOPT_SACK:
3826                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3827                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3828                                    opt_rx->sack_ok) {
3829                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3830                                 }
3831                                 break;
3832 #ifdef CONFIG_TCP_MD5SIG
3833                         case TCPOPT_MD5SIG:
3834                                 /*
3835                                  * The MD5 Hash has already been
3836                                  * checked (see tcp_v{4,6}_do_rcv()).
3837                                  */
3838                                 break;
3839 #endif
3840                         case TCPOPT_FASTOPEN:
3841                                 tcp_parse_fastopen_option(
3842                                         opsize - TCPOLEN_FASTOPEN_BASE,
3843                                         ptr, th->syn, foc, false);
3844                                 break;
3845 
3846                         case TCPOPT_EXP:
3847                                 /* Fast Open option shares code 254 using a
3848                                  * 16 bits magic number.
3849                                  */
3850                                 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3851                                     get_unaligned_be16(ptr) ==
3852                                     TCPOPT_FASTOPEN_MAGIC)
3853                                         tcp_parse_fastopen_option(opsize -
3854                                                 TCPOLEN_EXP_FASTOPEN_BASE,
3855                                                 ptr + 2, th->syn, foc, true);
3856                                 break;
3857 
3858                         }
3859                         ptr += opsize-2;
3860                         length -= opsize;
3861                 }
3862         }
3863 }
3864 EXPORT_SYMBOL(tcp_parse_options);
3865 
3866 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3867 {
3868         const __be32 *ptr = (const __be32 *)(th + 1);
3869 
3870         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3871                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3872                 tp->rx_opt.saw_tstamp = 1;
3873                 ++ptr;
3874                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3875                 ++ptr;
3876                 if (*ptr)
3877                         tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3878                 else
3879                         tp->rx_opt.rcv_tsecr = 0;
3880                 return true;
3881         }
3882         return false;
3883 }
3884 
3885 /* Fast parse options. This hopes to only see timestamps.
3886  * If it is wrong it falls back on tcp_parse_options().
3887  */
3888 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3889                                    const struct tcphdr *th, struct tcp_sock *tp)
3890 {
3891         /* In the spirit of fast parsing, compare doff directly to constant
3892          * values.  Because equality is used, short doff can be ignored here.
3893          */
3894         if (th->doff == (sizeof(*th) / 4)) {
3895                 tp->rx_opt.saw_tstamp = 0;
3896                 return false;
3897         } else if (tp->rx_opt.tstamp_ok &&
3898                    th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3899                 if (tcp_parse_aligned_timestamp(tp, th))
3900                         return true;
3901         }
3902 
3903         tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3904         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3905                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3906 
3907         return true;
3908 }
3909 
3910 #ifdef CONFIG_TCP_MD5SIG
3911 /*
3912  * Parse MD5 Signature option
3913  */
3914 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3915 {
3916         int length = (th->doff << 2) - sizeof(*th);
3917         const u8 *ptr = (const u8 *)(th + 1);
3918 
3919         /* If the TCP option is too short, we can short cut */
3920         if (length < TCPOLEN_MD5SIG)
3921                 return NULL;
3922 
3923         while (length > 0) {
3924                 int opcode = *ptr++;
3925                 int opsize;
3926 
3927                 switch (opcode) {
3928                 case TCPOPT_EOL:
3929                         return NULL;
3930                 case TCPOPT_NOP:
3931                         length--;
3932                         continue;
3933                 default:
3934                         opsize = *ptr++;
3935                         if (opsize < 2 || opsize > length)
3936                                 return NULL;
3937                         if (opcode == TCPOPT_MD5SIG)
3938                                 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3939                 }
3940                 ptr += opsize - 2;
3941                 length -= opsize;
3942         }
3943         return NULL;
3944 }
3945 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3946 #endif
3947 
3948 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3949  *
3950  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3951  * it can pass through stack. So, the following predicate verifies that
3952  * this segment is not used for anything but congestion avoidance or
3953  * fast retransmit. Moreover, we even are able to eliminate most of such
3954  * second order effects, if we apply some small "replay" window (~RTO)
3955  * to timestamp space.
3956  *
3957  * All these measures still do not guarantee that we reject wrapped ACKs
3958  * on networks with high bandwidth, when sequence space is recycled fastly,
3959  * but it guarantees that such events will be very rare and do not affect
3960  * connection seriously. This doesn't look nice, but alas, PAWS is really
3961  * buggy extension.
3962  *
3963  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3964  * states that events when retransmit arrives after original data are rare.
3965  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3966  * the biggest problem on large power networks even with minor reordering.
3967  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3968  * up to bandwidth of 18Gigabit/sec. 8) ]
3969  */
3970 
3971 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3972 {
3973         const struct tcp_sock *tp = tcp_sk(sk);
3974         const struct tcphdr *th = tcp_hdr(skb);
3975         u32 seq = TCP_SKB_CB(skb)->seq;
3976         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3977 
3978         return (/* 1. Pure ACK with correct sequence number. */
3979                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3980 
3981                 /* 2. ... and duplicate ACK. */
3982                 ack == tp->snd_una &&
3983 
3984                 /* 3. ... and does not update window. */
3985                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3986 
3987                 /* 4. ... and sits in replay window. */
3988                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3989 }
3990 
3991 static inline bool tcp_paws_discard(const struct sock *sk,
3992                                    const struct sk_buff *skb)
3993 {
3994         const struct tcp_sock *tp = tcp_sk(sk);
3995 
3996         return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3997                !tcp_disordered_ack(sk, skb);
3998 }
3999 
4000 /* Check segment sequence number for validity.
4001  *
4002  * Segment controls are considered valid, if the segment
4003  * fits to the window after truncation to the window. Acceptability
4004  * of data (and SYN, FIN, of course) is checked separately.
4005  * See tcp_data_queue(), for example.
4006  *
4007  * Also, controls (RST is main one) are accepted using RCV.WUP instead
4008  * of RCV.NXT. Peer still did not advance his SND.UNA when we
4009  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4010  * (borrowed from freebsd)
4011  */
4012 
4013 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4014 {
4015         return  !before(end_seq, tp->rcv_wup) &&
4016                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4017 }
4018 
4019 /* When we get a reset we do this. */
4020 void tcp_reset(struct sock *sk)
4021 {
4022         /* We want the right error as BSD sees it (and indeed as we do). */
4023         switch (sk->sk_state) {
4024         case TCP_SYN_SENT:
4025                 sk->sk_err = ECONNREFUSED;
4026                 break;
4027         case TCP_CLOSE_WAIT:
4028                 sk->sk_err = EPIPE;
4029                 break;
4030         case TCP_CLOSE:
4031                 return;
4032         default:
4033                 sk->sk_err = ECONNRESET;
4034         }
4035         /* This barrier is coupled with smp_rmb() in tcp_poll() */
4036         smp_wmb();
4037 
4038         if (!sock_flag(sk, SOCK_DEAD))
4039                 sk->sk_error_report(sk);
4040 
4041         tcp_done(sk);
4042 }
4043 
4044 /*
4045  *      Process the FIN bit. This now behaves as it is supposed to work
4046  *      and the FIN takes effect when it is validly part of sequence
4047  *      space. Not before when we get holes.
4048  *
4049  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4050  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
4051  *      TIME-WAIT)
4052  *
4053  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
4054  *      close and we go into CLOSING (and later onto TIME-WAIT)
4055  *
4056  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4057  */
4058 void tcp_fin(struct sock *sk)
4059 {
4060         struct tcp_sock *tp = tcp_sk(sk);
4061 
4062         inet_csk_schedule_ack(sk);
4063 
4064         sk->sk_shutdown |= RCV_SHUTDOWN;
4065         sock_set_flag(sk, SOCK_DONE);
4066 
4067         switch (sk->sk_state) {
4068         case TCP_SYN_RECV:
4069         case TCP_ESTABLISHED:
4070                 /* Move to CLOSE_WAIT */
4071                 tcp_set_state(sk, TCP_CLOSE_WAIT);
4072                 inet_csk(sk)->icsk_ack.pingpong = 1;
4073                 break;
4074 
4075         case TCP_CLOSE_WAIT:
4076         case TCP_CLOSING:
4077                 /* Received a retransmission of the FIN, do
4078                  * nothing.
4079                  */
4080                 break;
4081         case TCP_LAST_ACK:
4082                 /* RFC793: Remain in the LAST-ACK state. */
4083                 break;
4084 
4085         case TCP_FIN_WAIT1:
4086                 /* This case occurs when a simultaneous close
4087                  * happens, we must ack the received FIN and
4088                  * enter the CLOSING state.
4089                  */
4090                 tcp_send_ack(sk);
4091                 tcp_set_state(sk, TCP_CLOSING);
4092                 break;
4093         case TCP_FIN_WAIT2:
4094                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4095                 tcp_send_ack(sk);
4096                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4097                 break;
4098         default:
4099                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4100                  * cases we should never reach this piece of code.
4101                  */
4102                 pr_err("%s: Impossible, sk->sk_state=%d\n",
4103                        __func__, sk->sk_state);
4104                 break;
4105         }
4106 
4107         /* It _is_ possible, that we have something out-of-order _after_ FIN.
4108          * Probably, we should reset in this case. For now drop them.
4109          */
4110         __skb_queue_purge(&tp->out_of_order_queue);
4111         if (tcp_is_sack(tp))
4112                 tcp_sack_reset(&tp->rx_opt);
4113         sk_mem_reclaim(sk);
4114 
4115         if (!sock_flag(sk, SOCK_DEAD)) {
4116                 sk->sk_state_change(sk);
4117 
4118                 /* Do not send POLL_HUP for half duplex close. */
4119                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4120                     sk->sk_state == TCP_CLOSE)
4121                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4122                 else
4123                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4124         }
4125 }
4126 
4127 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4128                                   u32 end_seq)
4129 {
4130         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4131                 if (before(seq, sp->start_seq))
4132                         sp->start_seq = seq;
4133                 if (after(end_seq, sp->end_seq))
4134                         sp->end_seq = end_seq;
4135                 return true;
4136         }
4137         return false;
4138 }
4139 
4140 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4141 {
4142         struct tcp_sock *tp = tcp_sk(sk);
4143 
4144         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4145                 int mib_idx;
4146 
4147                 if (before(seq, tp->rcv_nxt))
4148                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4149                 else
4150                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4151 
4152                 NET_INC_STATS(sock_net(sk), mib_idx);
4153 
4154                 tp->rx_opt.dsack = 1;
4155                 tp->duplicate_sack[0].start_seq = seq;
4156                 tp->duplicate_sack[0].end_seq = end_seq;
4157         }
4158 }
4159 
4160 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4161 {
4162         struct tcp_sock *tp = tcp_sk(sk);
4163 
4164         if (!tp->rx_opt.dsack)
4165                 tcp_dsack_set(sk, seq, end_seq);
4166         else
4167                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4168 }
4169 
4170 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4171 {
4172         struct tcp_sock *tp = tcp_sk(sk);
4173 
4174         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4175             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4176                 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4177                 tcp_enter_quickack_mode(sk);
4178 
4179                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4180                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4181 
4182                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4183                                 end_seq = tp->rcv_nxt;
4184                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4185                 }
4186         }
4187 
4188         tcp_send_ack(sk);
4189 }
4190 
4191 /* These routines update the SACK block as out-of-order packets arrive or
4192  * in-order packets close up the sequence space.
4193  */
4194 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4195 {
4196         int this_sack;
4197         struct tcp_sack_block *sp = &tp->selective_acks[0];
4198         struct tcp_sack_block *swalk = sp + 1;
4199 
4200         /* See if the recent change to the first SACK eats into
4201          * or hits the sequence space of other SACK blocks, if so coalesce.
4202          */
4203         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4204                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4205                         int i;
4206 
4207                         /* Zap SWALK, by moving every further SACK up by one slot.
4208                          * Decrease num_sacks.
4209                          */
4210                         tp->rx_opt.num_sacks--;
4211                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4212                                 sp[i] = sp[i + 1];
4213                         continue;
4214                 }
4215                 this_sack++, swalk++;
4216         }
4217 }
4218 
4219 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4220 {
4221         struct tcp_sock *tp = tcp_sk(sk);
4222         struct tcp_sack_block *sp = &tp->selective_acks[0];
4223         int cur_sacks = tp->rx_opt.num_sacks;
4224         int this_sack;
4225 
4226         if (!cur_sacks)
4227                 goto new_sack;
4228 
4229         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4230                 if (tcp_sack_extend(sp, seq, end_seq)) {
4231                         /* Rotate this_sack to the first one. */
4232                         for (; this_sack > 0; this_sack--, sp--)
4233                                 swap(*sp, *(sp - 1));
4234                         if (cur_sacks > 1)
4235                                 tcp_sack_maybe_coalesce(tp);
4236                         return;
4237                 }
4238         }
4239 
4240         /* Could not find an adjacent existing SACK, build a new one,
4241          * put it at the front, and shift everyone else down.  We
4242          * always know there is at least one SACK present already here.
4243          *
4244          * If the sack array is full, forget about the last one.
4245          */
4246         if (this_sack >= TCP_NUM_SACKS) {
4247                 this_sack--;
4248                 tp->rx_opt.num_sacks--;
4249                 sp--;
4250         }
4251         for (; this_sack > 0; this_sack--, sp--)
4252                 *sp = *(sp - 1);
4253 
4254 new_sack:
4255         /* Build the new head SACK, and we're done. */
4256         sp->start_seq = seq;
4257         sp->end_seq = end_seq;
4258         tp->rx_opt.num_sacks++;
4259 }
4260 
4261 /* RCV.NXT advances, some SACKs should be eaten. */
4262 
4263 static void tcp_sack_remove(struct tcp_sock *tp)
4264 {
4265         struct tcp_sack_block *sp = &tp->selective_acks[0];
4266         int num_sacks = tp->rx_opt.num_sacks;
4267         int this_sack;
4268 
4269         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4270         if (skb_queue_empty(&tp->out_of_order_queue)) {
4271                 tp->rx_opt.num_sacks = 0;
4272                 return;
4273         }
4274 
4275         for (this_sack = 0; this_sack < num_sacks;) {
4276                 /* Check if the start of the sack is covered by RCV.NXT. */
4277                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4278                         int i;
4279 
4280                         /* RCV.NXT must cover all the block! */
4281                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4282 
4283                         /* Zap this SACK, by moving forward any other SACKS. */
4284                         for (i = this_sack+1; i < num_sacks; i++)
4285                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4286                         num_sacks--;
4287                         continue;
4288                 }
4289                 this_sack++;
4290                 sp++;
4291         }
4292         tp->rx_opt.num_sacks = num_sacks;
4293 }
4294 
4295 /**
4296  * tcp_try_coalesce - try to merge skb to prior one
4297  * @sk: socket
4298  * @to: prior buffer
4299  * @from: buffer to add in queue
4300  * @fragstolen: pointer to boolean
4301  *
4302  * Before queueing skb @from after @to, try to merge them
4303  * to reduce overall memory use and queue lengths, if cost is small.
4304  * Packets in ofo or receive queues can stay a long time.
4305  * Better try to coalesce them right now to avoid future collapses.
4306  * Returns true if caller should free @from instead of queueing it
4307  */
4308 static bool tcp_try_coalesce(struct sock *sk,
4309                              struct sk_buff *to,
4310                              struct sk_buff *from,
4311                              bool *fragstolen)
4312 {
4313         int delta;
4314 
4315         *fragstolen = false;
4316 
4317         /* Its possible this segment overlaps with prior segment in queue */
4318         if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4319                 return false;
4320 
4321         if (!skb_try_coalesce(to, from, fragstolen, &delta))
4322                 return false;
4323 
4324         atomic_add(delta, &sk->sk_rmem_alloc);
4325         sk_mem_charge(sk, delta);
4326         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4327         TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4328         TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4329         TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4330         return true;
4331 }
4332 
4333 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4334 {
4335         sk_drops_add(sk, skb);
4336         __kfree_skb(skb);
4337 }
4338 
4339 /* This one checks to see if we can put data from the
4340  * out_of_order queue into the receive_queue.
4341  */
4342 static void tcp_ofo_queue(struct sock *sk)
4343 {
4344         struct tcp_sock *tp = tcp_sk(sk);
4345         __u32 dsack_high = tp->rcv_nxt;
4346         struct sk_buff *skb, *tail;
4347         bool fragstolen, eaten;
4348 
4349         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4350                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4351                         break;
4352 
4353                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4354                         __u32 dsack = dsack_high;
4355                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4356                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4357                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4358                 }
4359 
4360                 __skb_unlink(skb, &tp->out_of_order_queue);
4361                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4362                         SOCK_DEBUG(sk, "ofo packet was already received\n");
4363                         tcp_drop(sk, skb);
4364                         continue;
4365                 }
4366                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4367                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4368                            TCP_SKB_CB(skb)->end_seq);
4369 
4370                 tail = skb_peek_tail(&sk->sk_receive_queue);
4371                 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4372                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4373                 if (!eaten)
4374                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4375                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4376                         tcp_fin(sk);
4377                 if (eaten)
4378                         kfree_skb_partial(skb, fragstolen);
4379         }
4380 }
4381 
4382 static bool tcp_prune_ofo_queue(struct sock *sk);
4383 static int tcp_prune_queue(struct sock *sk);
4384 
4385 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4386                                  unsigned int size)
4387 {
4388         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4389             !sk_rmem_schedule(sk, skb, size)) {
4390 
4391                 if (tcp_prune_queue(sk) < 0)
4392                         return -1;
4393 
4394                 if (!sk_rmem_schedule(sk, skb, size)) {
4395                         if (!tcp_prune_ofo_queue(sk))
4396                                 return -1;
4397 
4398                         if (!sk_rmem_schedule(sk, skb, size))
4399                                 return -1;
4400                 }
4401         }
4402         return 0;
4403 }
4404 
4405 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4406 {
4407         struct tcp_sock *tp = tcp_sk(sk);
4408         struct sk_buff *skb1;
4409         u32 seq, end_seq;
4410 
4411         tcp_ecn_check_ce(tp, skb);
4412 
4413         if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4414                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4415                 tcp_drop(sk, skb);
4416                 return;
4417         }
4418 
4419         /* Disable header prediction. */
4420         tp->pred_flags = 0;
4421         inet_csk_schedule_ack(sk);
4422 
4423         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4424         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4425                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4426 
4427         skb1 = skb_peek_tail(&tp->out_of_order_queue);
4428         if (!skb1) {
4429                 /* Initial out of order segment, build 1 SACK. */
4430                 if (tcp_is_sack(tp)) {
4431                         tp->rx_opt.num_sacks = 1;
4432                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4433                         tp->selective_acks[0].end_seq =
4434                                                 TCP_SKB_CB(skb)->end_seq;
4435                 }
4436                 __skb_queue_head(&tp->out_of_order_queue, skb);
4437                 goto end;
4438         }
4439 
4440         seq = TCP_SKB_CB(skb)->seq;
4441         end_seq = TCP_SKB_CB(skb)->end_seq;
4442 
4443         if (seq == TCP_SKB_CB(skb1)->end_seq) {
4444                 bool fragstolen;
4445 
4446                 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4447                         __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4448                 } else {
4449                         tcp_grow_window(sk, skb);
4450                         kfree_skb_partial(skb, fragstolen);
4451                         skb = NULL;
4452                 }
4453 
4454                 if (!tp->rx_opt.num_sacks ||
4455                     tp->selective_acks[0].end_seq != seq)
4456                         goto add_sack;
4457 
4458                 /* Common case: data arrive in order after hole. */
4459                 tp->selective_acks[0].end_seq = end_seq;
4460                 goto end;
4461         }
4462 
4463         /* Find place to insert this segment. */
4464         while (1) {
4465                 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4466                         break;
4467                 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4468                         skb1 = NULL;
4469                         break;
4470                 }
4471                 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4472         }
4473 
4474         /* Do skb overlap to previous one? */
4475         if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4476                 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4477                         /* All the bits are present. Drop. */
4478                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4479                         tcp_drop(sk, skb);
4480                         skb = NULL;
4481                         tcp_dsack_set(sk, seq, end_seq);
4482                         goto add_sack;
4483                 }
4484                 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4485                         /* Partial overlap. */
4486                         tcp_dsack_set(sk, seq,
4487                                       TCP_SKB_CB(skb1)->end_seq);
4488                 } else {
4489                         if (skb_queue_is_first(&tp->out_of_order_queue,
4490                                                skb1))
4491                                 skb1 = NULL;
4492                         else
4493                                 skb1 = skb_queue_prev(
4494                                         &tp->out_of_order_queue,
4495                                         skb1);
4496                 }
4497         }
4498         if (!skb1)
4499                 __skb_queue_head(&tp->out_of_order_queue, skb);
4500         else
4501                 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4502 
4503         /* And clean segments covered by new one as whole. */
4504         while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4505                 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4506 
4507                 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4508                         break;
4509                 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4510                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4511                                          end_seq);
4512                         break;
4513                 }
4514                 __skb_unlink(skb1, &tp->out_of_order_queue);
4515                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4516                                  TCP_SKB_CB(skb1)->end_seq);
4517                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4518                 tcp_drop(sk, skb1);
4519         }
4520 
4521 add_sack:
4522         if (tcp_is_sack(tp))
4523                 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4524 end:
4525         if (skb) {
4526                 tcp_grow_window(sk, skb);
4527                 skb_set_owner_r(skb, sk);
4528         }
4529 }
4530 
4531 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4532                   bool *fragstolen)
4533 {
4534         int eaten;
4535         struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4536 
4537         __skb_pull(skb, hdrlen);
4538         eaten = (tail &&
4539                  tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4540         tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4541         if (!eaten) {
4542                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4543                 skb_set_owner_r(skb, sk);
4544         }
4545         return eaten;
4546 }
4547 
4548 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4549 {
4550         struct sk_buff *skb;
4551         int err = -ENOMEM;
4552         int data_len = 0;
4553         bool fragstolen;
4554 
4555         if (size == 0)
4556                 return 0;
4557 
4558         if (size > PAGE_SIZE) {
4559                 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4560 
4561                 data_len = npages << PAGE_SHIFT;
4562                 size = data_len + (size & ~PAGE_MASK);
4563         }
4564         skb = alloc_skb_with_frags(size - data_len, data_len,
4565                                    PAGE_ALLOC_COSTLY_ORDER,
4566                                    &err, sk->sk_allocation);
4567         if (!skb)
4568                 goto err;
4569 
4570         skb_put(skb, size - data_len);
4571         skb->data_len = data_len;
4572         skb->len = size;
4573 
4574         if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4575                 goto err_free;
4576 
4577         err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4578         if (err)
4579                 goto err_free;
4580 
4581         TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4582         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4583         TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4584 
4585         if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4586                 WARN_ON_ONCE(fragstolen); /* should not happen */
4587                 __kfree_skb(skb);
4588         }
4589         return size;
4590 
4591 err_free:
4592         kfree_skb(skb);
4593 err:
4594         return err;
4595 
4596 }
4597 
4598 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4599 {
4600         struct tcp_sock *tp = tcp_sk(sk);
4601         bool fragstolen = false;
4602         int eaten = -1;
4603 
4604         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4605                 __kfree_skb(skb);
4606                 return;
4607         }
4608         skb_dst_drop(skb);
4609         __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4610 
4611         tcp_ecn_accept_cwr(tp, skb);
4612 
4613         tp->rx_opt.dsack = 0;
4614 
4615         /*  Queue data for delivery to the user.
4616          *  Packets in sequence go to the receive queue.
4617          *  Out of sequence packets to the out_of_order_queue.
4618          */
4619         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4620                 if (tcp_receive_window(tp) == 0)
4621                         goto out_of_window;
4622 
4623                 /* Ok. In sequence. In window. */
4624                 if (tp->ucopy.task == current &&
4625                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4626                     sock_owned_by_user(sk) && !tp->urg_data) {
4627                         int chunk = min_t(unsigned int, skb->len,
4628                                           tp->ucopy.len);
4629 
4630                         __set_current_state(TASK_RUNNING);
4631 
4632                         if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4633                                 tp->ucopy.len -= chunk;
4634                                 tp->copied_seq += chunk;
4635                                 eaten = (chunk == skb->len);
4636                                 tcp_rcv_space_adjust(sk);
4637                         }
4638                 }
4639 
4640                 if (eaten <= 0) {
4641 queue_and_out:
4642                         if (eaten < 0) {
4643                                 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4644                                         sk_forced_mem_schedule(sk, skb->truesize);
4645                                 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4646                                         goto drop;
4647                         }
4648                         eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4649                 }
4650                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4651                 if (skb->len)
4652                         tcp_event_data_recv(sk, skb);
4653                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4654                         tcp_fin(sk);
4655 
4656                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4657                         tcp_ofo_queue(sk);
4658 
4659                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4660                          * gap in queue is filled.
4661                          */
4662                         if (skb_queue_empty(&tp->out_of_order_queue))
4663                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4664                 }
4665 
4666                 if (tp->rx_opt.num_sacks)
4667                         tcp_sack_remove(tp);
4668 
4669                 tcp_fast_path_check(sk);
4670 
4671                 if (eaten > 0)
4672                         kfree_skb_partial(skb, fragstolen);
4673                 if (!sock_flag(sk, SOCK_DEAD))
4674                         sk->sk_data_ready(sk);
4675                 return;
4676         }
4677 
4678         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4679                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4680                 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4681                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4682 
4683 out_of_window:
4684                 tcp_enter_quickack_mode(sk);
4685                 inet_csk_schedule_ack(sk);
4686 drop:
4687                 tcp_drop(sk, skb);
4688                 return;
4689         }
4690 
4691         /* Out of window. F.e. zero window probe. */
4692         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4693                 goto out_of_window;
4694 
4695         tcp_enter_quickack_mode(sk);
4696 
4697         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4698                 /* Partial packet, seq < rcv_next < end_seq */
4699                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4700                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4701                            TCP_SKB_CB(skb)->end_seq);
4702 
4703                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4704 
4705                 /* If window is closed, drop tail of packet. But after
4706                  * remembering D-SACK for its head made in previous line.
4707                  */
4708                 if (!tcp_receive_window(tp))
4709                         goto out_of_window;
4710                 goto queue_and_out;
4711         }
4712 
4713         tcp_data_queue_ofo(sk, skb);
4714 }
4715 
4716 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4717                                         struct sk_buff_head *list)
4718 {
4719         struct sk_buff *next = NULL;
4720 
4721         if (!skb_queue_is_last(list, skb))
4722                 next = skb_queue_next(list, skb);
4723 
4724         __skb_unlink(skb, list);
4725         __kfree_skb(skb);
4726         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4727 
4728         return next;
4729 }
4730 
4731 /* Collapse contiguous sequence of skbs head..tail with
4732  * sequence numbers start..end.
4733  *
4734  * If tail is NULL, this means until the end of the list.
4735  *
4736  * Segments with FIN/SYN are not collapsed (only because this
4737  * simplifies code)
4738  */
4739 static void
4740 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4741              struct sk_buff *head, struct sk_buff *tail,
4742              u32 start, u32 end)
4743 {
4744         struct sk_buff *skb, *n;
4745         bool end_of_skbs;
4746 
4747         /* First, check that queue is collapsible and find
4748          * the point where collapsing can be useful. */
4749         skb = head;
4750 restart:
4751         end_of_skbs = true;
4752         skb_queue_walk_from_safe(list, skb, n) {
4753                 if (skb == tail)
4754                         break;
4755                 /* No new bits? It is possible on ofo queue. */
4756                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4757                         skb = tcp_collapse_one(sk, skb, list);
4758                         if (!skb)
4759                                 break;
4760                         goto restart;
4761                 }
4762 
4763                 /* The first skb to collapse is:
4764                  * - not SYN/FIN and
4765                  * - bloated or contains data before "start" or
4766                  *   overlaps to the next one.
4767                  */
4768                 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4769                     (tcp_win_from_space(skb->truesize) > skb->len ||
4770                      before(TCP_SKB_CB(skb)->seq, start))) {
4771                         end_of_skbs = false;
4772                         break;
4773                 }
4774 
4775                 if (!skb_queue_is_last(list, skb)) {
4776                         struct sk_buff *next = skb_queue_next(list, skb);
4777                         if (next != tail &&
4778                             TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4779                                 end_of_skbs = false;
4780                                 break;
4781                         }
4782                 }
4783 
4784                 /* Decided to skip this, advance start seq. */
4785                 start = TCP_SKB_CB(skb)->end_seq;
4786         }
4787         if (end_of_skbs ||
4788             (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4789                 return;
4790 
4791         while (before(start, end)) {
4792                 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4793                 struct sk_buff *nskb;
4794 
4795                 nskb = alloc_skb(copy, GFP_ATOMIC);
4796                 if (!nskb)
4797                         return;
4798 
4799                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4800                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4801                 __skb_queue_before(list, skb, nskb);
4802                 skb_set_owner_r(nskb, sk);
4803 
4804                 /* Copy data, releasing collapsed skbs. */
4805                 while (copy > 0) {
4806                         int offset = start - TCP_SKB_CB(skb)->seq;
4807                         int size = TCP_SKB_CB(skb)->end_seq - start;
4808 
4809                         BUG_ON(offset < 0);
4810                         if (size > 0) {
4811                                 size = min(copy, size);
4812                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4813                                         BUG();
4814                                 TCP_SKB_CB(nskb)->end_seq += size;
4815                                 copy -= size;
4816                                 start += size;
4817                         }
4818                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4819                                 skb = tcp_collapse_one(sk, skb, list);
4820                                 if (!skb ||
4821                                     skb == tail ||
4822                                     (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4823                                         return;
4824                         }
4825                 }
4826         }
4827 }
4828 
4829 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4830  * and tcp_collapse() them until all the queue is collapsed.
4831  */
4832 static void tcp_collapse_ofo_queue(struct sock *sk)
4833 {
4834         struct tcp_sock *tp = tcp_sk(sk);
4835         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4836         struct sk_buff *head;
4837         u32 start, end;
4838 
4839         if (!skb)
4840                 return;
4841 
4842         start = TCP_SKB_CB(skb)->seq;
4843         end = TCP_SKB_CB(skb)->end_seq;
4844         head = skb;
4845 
4846         for (;;) {
4847                 struct sk_buff *next = NULL;
4848 
4849                 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4850                         next = skb_queue_next(&tp->out_of_order_queue, skb);
4851                 skb = next;
4852 
4853                 /* Segment is terminated when we see gap or when
4854                  * we are at the end of all the queue. */
4855                 if (!skb ||
4856                     after(TCP_SKB_CB(skb)->seq, end) ||
4857                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4858                         tcp_collapse(sk, &tp->out_of_order_queue,
4859                                      head, skb, start, end);
4860                         head = skb;
4861                         if (!skb)
4862                                 break;
4863                         /* Start new segment */
4864                         start = TCP_SKB_CB(skb)->seq;
4865                         end = TCP_SKB_CB(skb)->end_seq;
4866                 } else {
4867                         if (before(TCP_SKB_CB(skb)->seq, start))
4868                                 start = TCP_SKB_CB(skb)->seq;
4869                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4870                                 end = TCP_SKB_CB(skb)->end_seq;
4871                 }
4872         }
4873 }
4874 
4875 /*
4876  * Purge the out-of-order queue.
4877  * Return true if queue was pruned.
4878  */
4879 static bool tcp_prune_ofo_queue(struct sock *sk)
4880 {
4881         struct tcp_sock *tp = tcp_sk(sk);
4882         bool res = false;
4883 
4884         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4885                 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
4886                 __skb_queue_purge(&tp->out_of_order_queue);
4887 
4888                 /* Reset SACK state.  A conforming SACK implementation will
4889                  * do the same at a timeout based retransmit.  When a connection
4890                  * is in a sad state like this, we care only about integrity
4891                  * of the connection not performance.
4892                  */
4893                 if (tp->rx_opt.sack_ok)
4894                         tcp_sack_reset(&tp->rx_opt);
4895                 sk_mem_reclaim(sk);
4896                 res = true;
4897         }
4898         return res;
4899 }
4900 
4901 /* Reduce allocated memory if we can, trying to get
4902  * the socket within its memory limits again.
4903  *
4904  * Return less than zero if we should start dropping frames
4905  * until the socket owning process reads some of the data
4906  * to stabilize the situation.
4907  */
4908 static int tcp_prune_queue(struct sock *sk)
4909 {
4910         struct tcp_sock *tp = tcp_sk(sk);
4911 
4912         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4913 
4914         NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
4915 
4916         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4917                 tcp_clamp_window(sk);
4918         else if (tcp_under_memory_pressure(sk))
4919                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4920 
4921         tcp_collapse_ofo_queue(sk);
4922         if (!skb_queue_empty(&sk->sk_receive_queue))
4923                 tcp_collapse(sk, &sk->sk_receive_queue,
4924                              skb_peek(&sk->sk_receive_queue),
4925                              NULL,
4926                              tp->copied_seq, tp->rcv_nxt);
4927         sk_mem_reclaim(sk);
4928 
4929         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4930                 return 0;
4931 
4932         /* Collapsing did not help, destructive actions follow.
4933          * This must not ever occur. */
4934 
4935         tcp_prune_ofo_queue(sk);
4936 
4937         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4938                 return 0;
4939 
4940         /* If we are really being abused, tell the caller to silently
4941          * drop receive data on the floor.  It will get retransmitted
4942          * and hopefully then we'll have sufficient space.
4943          */
4944         NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
4945 
4946         /* Massive buffer overcommit. */
4947         tp->pred_flags = 0;
4948         return -1;
4949 }
4950 
4951 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4952 {
4953         const struct tcp_sock *tp = tcp_sk(sk);
4954 
4955         /* If the user specified a specific send buffer setting, do
4956          * not modify it.
4957          */
4958         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4959                 return false;
4960 
4961         /* If we are under global TCP memory pressure, do not expand.  */
4962         if (tcp_under_memory_pressure(sk))
4963                 return false;
4964 
4965         /* If we are under soft global TCP memory pressure, do not expand.  */
4966         if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4967                 return false;
4968 
4969         /* If we filled the congestion window, do not expand.  */
4970         if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4971                 return false;
4972 
4973         return true;
4974 }
4975 
4976 /* When incoming ACK allowed to free some skb from write_queue,
4977  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4978  * on the exit from tcp input handler.
4979  *
4980  * PROBLEM: sndbuf expansion does not work well with largesend.
4981  */
4982 static void tcp_new_space(struct sock *sk)
4983 {
4984         struct tcp_sock *tp = tcp_sk(sk);
4985 
4986         if (tcp_should_expand_sndbuf(sk)) {
4987                 tcp_sndbuf_expand(sk);
4988                 tp->snd_cwnd_stamp = tcp_time_stamp;
4989         }
4990 
4991         sk->sk_write_space(sk);
4992 }
4993 
4994 static void tcp_check_space(struct sock *sk)
4995 {
4996         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4997                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4998                 /* pairs with tcp_poll() */
4999                 smp_mb__after_atomic();
5000                 if (sk->sk_socket &&
5001                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5002                         tcp_new_space(sk);
5003         }
5004 }
5005 
5006 static inline void tcp_data_snd_check(struct sock *sk)
5007 {
5008         tcp_push_pending_frames(sk);
5009         tcp_check_space(sk);
5010 }
5011 
5012 /*
5013  * Check if sending an ack is needed.
5014  */
5015 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5016 {
5017         struct tcp_sock *tp = tcp_sk(sk);
5018 
5019             /* More than one full frame received... */
5020         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5021              /* ... and right edge of window advances far enough.
5022               * (tcp_recvmsg() will send ACK otherwise). Or...
5023               */
5024              __tcp_select_window(sk) >= tp->rcv_wnd) ||
5025             /* We ACK each frame or... */
5026             tcp_in_quickack_mode(sk) ||
5027             /* We have out of order data. */
5028             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
5029                 /* Then ack it now */
5030                 tcp_send_ack(sk);
5031         } else {
5032                 /* Else, send delayed ack. */
5033                 tcp_send_delayed_ack(sk);
5034         }
5035 }
5036 
5037 static inline void tcp_ack_snd_check(struct sock *sk)
5038 {
5039         if (!inet_csk_ack_scheduled(sk)) {
5040                 /* We sent a data segment already. */
5041                 return;
5042         }
5043         __tcp_ack_snd_check(sk, 1);
5044 }
5045 
5046 /*
5047  *      This routine is only called when we have urgent data
5048  *      signaled. Its the 'slow' part of tcp_urg. It could be
5049  *      moved inline now as tcp_urg is only called from one
5050  *      place. We handle URGent data wrong. We have to - as
5051  *      BSD still doesn't use the correction from RFC961.
5052  *      For 1003.1g we should support a new option TCP_STDURG to permit
5053  *      either form (or just set the sysctl tcp_stdurg).
5054  */
5055 
5056 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5057 {
5058         struct tcp_sock *tp = tcp_sk(sk);
5059         u32 ptr = ntohs(th->urg_ptr);
5060 
5061         if (ptr && !sysctl_tcp_stdurg)
5062                 ptr--;
5063         ptr += ntohl(th->seq);
5064 
5065         /* Ignore urgent data that we've already seen and read. */
5066         if (after(tp->copied_seq, ptr))
5067                 return;
5068 
5069         /* Do not replay urg ptr.
5070          *
5071          * NOTE: interesting situation not covered by specs.
5072          * Misbehaving sender may send urg ptr, pointing to segment,
5073          * which we already have in ofo queue. We are not able to fetch
5074          * such data and will stay in TCP_URG_NOTYET until will be eaten
5075          * by recvmsg(). Seems, we are not obliged to handle such wicked
5076          * situations. But it is worth to think about possibility of some
5077          * DoSes using some hypothetical application level deadlock.
5078          */
5079         if (before(ptr, tp->rcv_nxt))
5080                 return;
5081 
5082         /* Do we already have a newer (or duplicate) urgent pointer? */
5083         if (tp->urg_data && !after(ptr, tp->urg_seq))
5084                 return;
5085 
5086         /* Tell the world about our new urgent pointer. */
5087         sk_send_sigurg(sk);
5088 
5089         /* We may be adding urgent data when the last byte read was
5090          * urgent. To do this requires some care. We cannot just ignore
5091          * tp->copied_seq since we would read the last urgent byte again
5092          * as data, nor can we alter copied_seq until this data arrives
5093          * or we break the semantics of SIOCATMARK (and thus sockatmark())
5094          *
5095          * NOTE. Double Dutch. Rendering to plain English: author of comment
5096          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
5097          * and expect that both A and B disappear from stream. This is _wrong_.
5098          * Though this happens in BSD with high probability, this is occasional.
5099          * Any application relying on this is buggy. Note also, that fix "works"
5100          * only in this artificial test. Insert some normal data between A and B and we will
5101          * decline of BSD again. Verdict: it is better to remove to trap
5102          * buggy users.
5103          */
5104         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5105             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5106                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5107                 tp->copied_seq++;
5108                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5109                         __skb_unlink(skb, &sk->sk_receive_queue);
5110                         __kfree_skb(skb);
5111                 }
5112         }
5113 
5114         tp->urg_data = TCP_URG_NOTYET;
5115         tp->urg_seq = ptr;
5116 
5117         /* Disable header prediction. */
5118         tp->pred_flags = 0;
5119 }
5120 
5121 /* This is the 'fast' part of urgent handling. */
5122 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5123 {
5124         struct tcp_sock *tp = tcp_sk(sk);
5125 
5126         /* Check if we get a new urgent pointer - normally not. */
5127         if (th->urg)
5128                 tcp_check_urg(sk, th);
5129 
5130         /* Do we wait for any urgent data? - normally not... */
5131         if (tp->urg_data == TCP_URG_NOTYET) {
5132                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5133                           th->syn;
5134 
5135                 /* Is the urgent pointer pointing into this packet? */
5136                 if (ptr < skb->len) {
5137                         u8 tmp;
5138                         if (skb_copy_bits(skb, ptr, &tmp, 1))
5139                                 BUG();
5140                         tp->urg_data = TCP_URG_VALID | tmp;
5141                         if (!sock_flag(sk, SOCK_DEAD))
5142                                 sk->sk_data_ready(sk);
5143                 }
5144         }
5145 }
5146 
5147 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5148 {
5149         struct tcp_sock *tp = tcp_sk(sk);
5150         int chunk = skb->len - hlen;
5151         int err;
5152 
5153         if (skb_csum_unnecessary(skb))
5154                 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5155         else
5156                 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5157 
5158         if (!err) {
5159                 tp->ucopy.len -= chunk;
5160                 tp->copied_seq += chunk;
5161                 tcp_rcv_space_adjust(sk);
5162         }
5163 
5164         return err;
5165 }
5166 
5167 /* Does PAWS and seqno based validation of an incoming segment, flags will
5168  * play significant role here.
5169  */
5170 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5171                                   const struct tcphdr *th, int syn_inerr)
5172 {
5173         struct tcp_sock *tp = tcp_sk(sk);
5174         bool rst_seq_match = false;
5175 
5176         /* RFC1323: H1. Apply PAWS check first. */
5177         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5178             tcp_paws_discard(sk, skb)) {
5179                 if (!th->rst) {
5180                         NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5181                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5182                                                   LINUX_MIB_TCPACKSKIPPEDPAWS,
5183                                                   &tp->last_oow_ack_time))
5184                                 tcp_send_dupack(sk, skb);
5185                         goto discard;
5186                 }
5187                 /* Reset is accepted even if it did not pass PAWS. */
5188         }
5189 
5190         /* Step 1: check sequence number */
5191         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5192                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5193                  * (RST) segments are validated by checking their SEQ-fields."
5194                  * And page 69: "If an incoming segment is not acceptable,
5195                  * an acknowledgment should be sent in reply (unless the RST
5196                  * bit is set, if so drop the segment and return)".
5197                  */
5198                 if (!th->rst) {
5199                         if (th->syn)
5200                                 goto syn_challenge;
5201                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5202                                                   LINUX_MIB_TCPACKSKIPPEDSEQ,
5203                                                   &tp->last_oow_ack_time))
5204                                 tcp_send_dupack(sk, skb);
5205                 }
5206                 goto discard;
5207         }
5208 
5209         /* Step 2: check RST bit */
5210         if (th->rst) {
5211                 /* RFC 5961 3.2 (extend to match against SACK too if available):
5212                  * If seq num matches RCV.NXT or the right-most SACK block,
5213                  * then
5214                  *     RESET the connection
5215                  * else
5216                  *     Send a challenge ACK
5217                  */
5218                 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5219                         rst_seq_match = true;
5220                 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5221                         struct tcp_sack_block *sp = &tp->selective_acks[0];
5222                         int max_sack = sp[0].end_seq;
5223                         int this_sack;
5224 
5225                         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5226                              ++this_sack) {
5227                                 max_sack = after(sp[this_sack].end_seq,
5228                                                  max_sack) ?
5229                                         sp[this_sack].end_seq : max_sack;
5230                         }
5231 
5232                         if (TCP_SKB_CB(skb)->seq == max_sack)
5233                                 rst_seq_match = true;
5234                 }
5235 
5236                 if (rst_seq_match)
5237                         tcp_reset(sk);
5238                 else
5239                         tcp_send_challenge_ack(sk, skb);
5240                 goto discard;
5241         }
5242 
5243         /* step 3: check security and precedence [ignored] */
5244 
5245         /* step 4: Check for a SYN
5246          * RFC 5961 4.2 : Send a challenge ack
5247          */
5248         if (th->syn) {
5249 syn_challenge:
5250                 if (syn_inerr)
5251                         TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5252                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5253                 tcp_send_challenge_ack(sk, skb);
5254                 goto discard;
5255         }
5256 
5257         return true;
5258 
5259 discard:
5260         tcp_drop(sk, skb);
5261         return false;
5262 }
5263 
5264 /*
5265  *      TCP receive function for the ESTABLISHED state.
5266  *
5267  *      It is split into a fast path and a slow path. The fast path is
5268  *      disabled when:
5269  *      - A zero window was announced from us - zero window probing
5270  *        is only handled properly in the slow path.
5271  *      - Out of order segments arrived.
5272  *      - Urgent data is expected.
5273  *      - There is no buffer space left
5274  *      - Unexpected TCP flags/window values/header lengths are received
5275  *        (detected by checking the TCP header against pred_flags)
5276  *      - Data is sent in both directions. Fast path only supports pure senders
5277  *        or pure receivers (this means either the sequence number or the ack
5278  *        value must stay constant)
5279  *      - Unexpected TCP option.
5280  *
5281  *      When these conditions are not satisfied it drops into a standard
5282  *      receive procedure patterned after RFC793 to handle all cases.
5283  *      The first three cases are guaranteed by proper pred_flags setting,
5284  *      the rest is checked inline. Fast processing is turned on in
5285  *      tcp_data_queue when everything is OK.
5286  */
5287 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5288                          const struct tcphdr *th, unsigned int len)
5289 {
5290         struct tcp_sock *tp = tcp_sk(sk);
5291 
5292         if (unlikely(!sk->sk_rx_dst))
5293                 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5294         /*
5295          *      Header prediction.
5296          *      The code loosely follows the one in the famous
5297          *      "30 instruction TCP receive" Van Jacobson mail.
5298          *
5299          *      Van's trick is to deposit buffers into socket queue
5300          *      on a device interrupt, to call tcp_recv function
5301          *      on the receive process context and checksum and copy
5302          *      the buffer to user space. smart...
5303          *
5304          *      Our current scheme is not silly either but we take the
5305          *      extra cost of the net_bh soft interrupt processing...
5306          *      We do checksum and copy also but from device to kernel.
5307          */
5308 
5309         tp->rx_opt.saw_tstamp = 0;
5310 
5311         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5312          *      if header_prediction is to be made
5313          *      'S' will always be tp->tcp_header_len >> 2
5314          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5315          *  turn it off (when there are holes in the receive
5316          *       space for instance)
5317          *      PSH flag is ignored.
5318          */
5319 
5320         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5321             TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5322             !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5323                 int tcp_header_len = tp->tcp_header_len;
5324 
5325                 /* Timestamp header prediction: tcp_header_len
5326                  * is automatically equal to th->doff*4 due to pred_flags
5327                  * match.
5328                  */
5329 
5330                 /* Check timestamp */
5331                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5332                         /* No? Slow path! */
5333                         if (!tcp_parse_aligned_timestamp(tp, th))
5334                                 goto slow_path;
5335 
5336                         /* If PAWS failed, check it more carefully in slow path */
5337                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5338                                 goto slow_path;
5339 
5340                         /* DO NOT update ts_recent here, if checksum fails
5341                          * and timestamp was corrupted part, it will result
5342                          * in a hung connection since we will drop all
5343                          * future packets due to the PAWS test.
5344                          */
5345                 }
5346 
5347                 if (len <= tcp_header_len) {
5348                         /* Bulk data transfer: sender */
5349                         if (len == tcp_header_len) {
5350                                 /* Predicted packet is in window by definition.
5351                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5352                                  * Hence, check seq<=rcv_wup reduces to:
5353                                  */
5354                                 if (tcp_header_len ==
5355                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5356                                     tp->rcv_nxt == tp->rcv_wup)
5357                                         tcp_store_ts_recent(tp);
5358 
5359                                 /* We know that such packets are checksummed
5360                                  * on entry.
5361                                  */
5362                                 tcp_ack(sk, skb, 0);
5363                                 __kfree_skb(skb);
5364                                 tcp_data_snd_check(sk);
5365                                 return;
5366                         } else { /* Header too small */
5367                                 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5368                                 goto discard;
5369                         }
5370                 } else {
5371                         int eaten = 0;
5372                         bool fragstolen = false;
5373 
5374                         if (tp->ucopy.task == current &&
5375                             tp->copied_seq == tp->rcv_nxt &&
5376                             len - tcp_header_len <= tp->ucopy.len &&
5377                             sock_owned_by_user(sk)) {
5378                                 __set_current_state(TASK_RUNNING);
5379 
5380                                 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5381                                         /* Predicted packet is in window by definition.
5382                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5383                                          * Hence, check seq<=rcv_wup reduces to:
5384                                          */
5385                                         if (tcp_header_len ==
5386                                             (sizeof(struct tcphdr) +
5387                                              TCPOLEN_TSTAMP_ALIGNED) &&
5388                                             tp->rcv_nxt == tp->rcv_wup)
5389                                                 tcp_store_ts_recent(tp);
5390 
5391                                         tcp_rcv_rtt_measure_ts(sk, skb);
5392 
5393                                         __skb_pull(skb, tcp_header_len);
5394                                         tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5395                                         NET_INC_STATS(sock_net(sk),
5396                                                         LINUX_MIB_TCPHPHITSTOUSER);
5397                                         eaten = 1;
5398                                 }
5399                         }
5400                         if (!eaten) {
5401                                 if (tcp_checksum_complete(skb))
5402                                         goto csum_error;
5403 
5404                                 if ((int)skb->truesize > sk->sk_forward_alloc)
5405                                         goto step5;
5406 
5407                                 /* Predicted packet is in window by definition.
5408                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5409                                  * Hence, check seq<=rcv_wup reduces to:
5410                                  */
5411                                 if (tcp_header_len ==
5412                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5413                                     tp->rcv_nxt == tp->rcv_wup)
5414                                         tcp_store_ts_recent(tp);
5415 
5416                                 tcp_rcv_rtt_measure_ts(sk, skb);
5417 
5418                                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5419 
5420                                 /* Bulk data transfer: receiver */
5421                                 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5422                                                       &fragstolen);
5423                         }
5424 
5425                         tcp_event_data_recv(sk, skb);
5426 
5427                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5428                                 /* Well, only one small jumplet in fast path... */
5429                                 tcp_ack(sk, skb, FLAG_DATA);
5430                                 tcp_data_snd_check(sk);
5431                                 if (!inet_csk_ack_scheduled(sk))
5432                                         goto no_ack;
5433                         }
5434 
5435                         __tcp_ack_snd_check(sk, 0);
5436 no_ack:
5437                         if (eaten)
5438                                 kfree_skb_partial(skb, fragstolen);
5439                         sk->sk_data_ready(sk);
5440                         return;
5441                 }
5442         }
5443 
5444 slow_path:
5445         if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5446                 goto csum_error;
5447 
5448         if (!th->ack && !th->rst && !th->syn)
5449                 goto discard;
5450 
5451         /*
5452          *      Standard slow path.
5453          */
5454 
5455         if (!tcp_validate_incoming(sk, skb, th, 1))
5456                 return;
5457 
5458 step5:
5459         if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5460                 goto discard;
5461 
5462         tcp_rcv_rtt_measure_ts(sk, skb);
5463 
5464         /* Process urgent data. */
5465         tcp_urg(sk, skb, th);
5466 
5467         /* step 7: process the segment text */
5468         tcp_data_queue(sk, skb);
5469 
5470         tcp_data_snd_check(sk);
5471         tcp_ack_snd_check(sk);
5472         return;
5473 
5474 csum_error:
5475         TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5476         TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5477 
5478 discard:
5479         tcp_drop(sk, skb);
5480 }
5481 EXPORT_SYMBOL(tcp_rcv_established);
5482 
5483 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5484 {
5485         struct tcp_sock *tp = tcp_sk(sk);
5486         struct inet_connection_sock *icsk = inet_csk(sk);
5487 
5488         tcp_set_state(sk, TCP_ESTABLISHED);
5489 
5490         if (skb) {
5491                 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5492                 security_inet_conn_established(sk, skb);
5493         }
5494 
5495         /* Make sure socket is routed, for correct metrics.  */
5496         icsk->icsk_af_ops->rebuild_header(sk);
5497 
5498         tcp_init_metrics(sk);
5499 
5500         tcp_init_congestion_control(sk);
5501 
5502         /* Prevent spurious tcp_cwnd_restart() on first data
5503          * packet.
5504          */
5505         tp->lsndtime = tcp_time_stamp;
5506 
5507         tcp_init_buffer_space(sk);
5508 
5509         if (sock_flag(sk, SOCK_KEEPOPEN))
5510                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5511 
5512         if (!tp->rx_opt.snd_wscale)
5513                 __tcp_fast_path_on(tp, tp->snd_wnd);
5514         else
5515                 tp->pred_flags = 0;
5516 
5517         if (!sock_flag(sk, SOCK_DEAD)) {
5518                 sk->sk_state_change(sk);
5519                 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5520         }
5521 }
5522 
5523 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5524                                     struct tcp_fastopen_cookie *cookie)
5525 {
5526         struct tcp_sock *tp = tcp_sk(sk);
5527         struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5528         u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5529         bool syn_drop = false;
5530 
5531         if (mss == tp->rx_opt.user_mss) {
5532                 struct tcp_options_received opt;
5533 
5534                 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5535                 tcp_clear_options(&opt);
5536                 opt.user_mss = opt.mss_clamp = 0;
5537                 tcp_parse_options(synack, &opt, 0, NULL);
5538                 mss = opt.mss_clamp;
5539         }
5540 
5541         if (!tp->syn_fastopen) {
5542                 /* Ignore an unsolicited cookie */
5543                 cookie->len = -1;
5544         } else if (tp->total_retrans) {
5545                 /* SYN timed out and the SYN-ACK neither has a cookie nor
5546                  * acknowledges data. Presumably the remote received only
5547                  * the retransmitted (regular) SYNs: either the original
5548                  * SYN-data or the corresponding SYN-ACK was dropped.
5549                  */
5550                 syn_drop = (cookie->len < 0 && data);
5551         } else if (cookie->len < 0 && !tp->syn_data) {
5552                 /* We requested a cookie but didn't get it. If we did not use
5553                  * the (old) exp opt format then try so next time (try_exp=1).
5554                  * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5555                  */
5556                 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5557         }
5558 
5559         tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5560 
5561         if (data) { /* Retransmit unacked data in SYN */
5562                 tcp_for_write_queue_from(data, sk) {
5563                         if (data == tcp_send_head(sk) ||
5564                             __tcp_retransmit_skb(sk, data, 1))
5565                                 break;
5566                 }
5567                 tcp_rearm_rto(sk);
5568                 NET_INC_STATS(sock_net(sk),
5569                                 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5570                 return true;
5571         }
5572         tp->syn_data_acked = tp->syn_data;
5573         if (tp->syn_data_acked)
5574                 NET_INC_STATS(sock_net(sk),
5575                                 LINUX_MIB_TCPFASTOPENACTIVE);
5576 
5577         tcp_fastopen_add_skb(sk, synack);
5578 
5579         return false;
5580 }
5581 
5582 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5583                                          const struct tcphdr *th)
5584 {
5585         struct inet_connection_sock *icsk = inet_csk(sk);
5586         struct tcp_sock *tp = tcp_sk(sk);
5587         struct tcp_fastopen_cookie foc = { .len = -1 };
5588         int saved_clamp = tp->rx_opt.mss_clamp;
5589 
5590         tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5591         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5592                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5593 
5594         if (th->ack) {
5595                 /* rfc793:
5596                  * "If the state is SYN-SENT then
5597                  *    first check the ACK bit
5598                  *      If the ACK bit is set
5599                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5600                  *        a reset (unless the RST bit is set, if so drop
5601                  *        the segment and return)"
5602                  */
5603                 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5604                     after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5605                         goto reset_and_undo;
5606 
5607                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5608                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5609                              tcp_time_stamp)) {
5610                         NET_INC_STATS(sock_net(sk),
5611                                         LINUX_MIB_PAWSACTIVEREJECTED);
5612                         goto reset_and_undo;
5613                 }
5614 
5615                 /* Now ACK is acceptable.
5616                  *
5617                  * "If the RST bit is set
5618                  *    If the ACK was acceptable then signal the user "error:
5619                  *    connection reset", drop the segment, enter CLOSED state,
5620                  *    delete TCB, and return."
5621                  */
5622 
5623                 if (th->rst) {
5624                         tcp_reset(sk);
5625                         goto discard;
5626                 }
5627 
5628                 /* rfc793:
5629                  *   "fifth, if neither of the SYN or RST bits is set then
5630                  *    drop the segment and return."
5631                  *
5632                  *    See note below!
5633                  *                                        --ANK(990513)
5634                  */
5635                 if (!th->syn)
5636                         goto discard_and_undo;
5637 
5638                 /* rfc793:
5639                  *   "If the SYN bit is on ...
5640                  *    are acceptable then ...
5641                  *    (our SYN has been ACKed), change the connection
5642                  *    state to ESTABLISHED..."
5643                  */
5644 
5645                 tcp_ecn_rcv_synack(tp, th);
5646 
5647                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5648                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5649 
5650                 /* Ok.. it's good. Set up sequence numbers and
5651                  * move to established.
5652                  */
5653                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5654                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5655 
5656                 /* RFC1323: The window in SYN & SYN/ACK segments is
5657                  * never scaled.
5658                  */
5659                 tp->snd_wnd = ntohs(th->window);
5660 
5661                 if (!tp->rx_opt.wscale_ok) {
5662                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5663                         tp->window_clamp = min(tp->window_clamp, 65535U);
5664                 }
5665 
5666                 if (tp->rx_opt.saw_tstamp) {
5667                         tp->rx_opt.tstamp_ok       = 1;
5668                         tp->tcp_header_len =
5669                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5670                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5671                         tcp_store_ts_recent(tp);
5672                 } else {
5673                         tp->tcp_header_len = sizeof(struct tcphdr);
5674                 }
5675 
5676                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5677                         tcp_enable_fack(tp);
5678 
5679                 tcp_mtup_init(sk);
5680                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5681                 tcp_initialize_rcv_mss(sk);
5682 
5683                 /* Remember, tcp_poll() does not lock socket!
5684                  * Change state from SYN-SENT only after copied_seq
5685                  * is initialized. */
5686                 tp->copied_seq = tp->rcv_nxt;
5687 
5688                 smp_mb();
5689 
5690                 tcp_finish_connect(sk, skb);
5691 
5692                 if ((tp->syn_fastopen || tp->syn_data) &&
5693                     tcp_rcv_fastopen_synack(sk, skb, &foc))
5694                         return -1;
5695 
5696                 if (sk->sk_write_pending ||
5697                     icsk->icsk_accept_queue.rskq_defer_accept ||
5698                     icsk->icsk_ack.pingpong) {
5699                         /* Save one ACK. Data will be ready after
5700                          * several ticks, if write_pending is set.
5701                          *
5702                          * It may be deleted, but with this feature tcpdumps
5703                          * look so _wonderfully_ clever, that I was not able
5704                          * to stand against the temptation 8)     --ANK
5705                          */
5706                         inet_csk_schedule_ack(sk);
5707                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5708                         tcp_enter_quickack_mode(sk);
5709                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5710                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5711 
5712 discard:
5713                         tcp_drop(sk, skb);
5714                         return 0;
5715                 } else {
5716                         tcp_send_ack(sk);
5717                 }
5718                 return -1;
5719         }
5720 
5721         /* No ACK in the segment */
5722 
5723         if (th->rst) {
5724                 /* rfc793:
5725                  * "If the RST bit is set
5726                  *
5727                  *      Otherwise (no ACK) drop the segment and return."
5728                  */
5729 
5730                 goto discard_and_undo;
5731         }
5732 
5733         /* PAWS check. */
5734         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5735             tcp_paws_reject(&tp->rx_opt, 0))
5736                 goto discard_and_undo;
5737 
5738         if (th->syn) {
5739                 /* We see SYN without ACK. It is attempt of
5740                  * simultaneous connect with crossed SYNs.
5741                  * Particularly, it can be connect to self.
5742                  */
5743                 tcp_set_state(sk, TCP_SYN_RECV);
5744 
5745                 if (tp->rx_opt.saw_tstamp) {
5746                         tp->rx_opt.tstamp_ok = 1;
5747                         tcp_store_ts_recent(tp);
5748                         tp->tcp_header_len =
5749                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5750                 } else {
5751                         tp->tcp_header_len = sizeof(struct tcphdr);
5752                 }
5753 
5754                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5755                 tp->copied_seq = tp->rcv_nxt;
5756                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5757 
5758                 /* RFC1323: The window in SYN & SYN/ACK segments is
5759                  * never scaled.
5760                  */
5761                 tp->snd_wnd    = ntohs(th->window);
5762                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5763                 tp->max_window = tp->snd_wnd;
5764 
5765                 tcp_ecn_rcv_syn(tp, th);
5766 
5767                 tcp_mtup_init(sk);
5768                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5769                 tcp_initialize_rcv_mss(sk);
5770 
5771                 tcp_send_synack(sk);
5772 #if 0
5773                 /* Note, we could accept data and URG from this segment.
5774                  * There are no obstacles to make this (except that we must
5775                  * either change tcp_recvmsg() to prevent it from returning data
5776                  * before 3WHS completes per RFC793, or employ TCP Fast Open).
5777                  *
5778                  * However, if we ignore data in ACKless segments sometimes,
5779                  * we have no reasons to accept it sometimes.
5780                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5781                  * is not flawless. So, discard packet for sanity.
5782                  * Uncomment this return to process the data.
5783                  */
5784                 return -1;
5785 #else
5786                 goto discard;
5787 #endif
5788         }
5789         /* "fifth, if neither of the SYN or RST bits is set then
5790          * drop the segment and return."
5791          */
5792 
5793 discard_and_undo:
5794         tcp_clear_options(&tp->rx_opt);
5795         tp->rx_opt.mss_clamp = saved_clamp;
5796         goto discard;
5797 
5798 reset_and_undo:
5799         tcp_clear_options(&tp->rx_opt);
5800         tp->rx_opt.mss_clamp = saved_clamp;
5801         return 1;
5802 }
5803 
5804 /*
5805  *      This function implements the receiving procedure of RFC 793 for
5806  *      all states except ESTABLISHED and TIME_WAIT.
5807  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5808  *      address independent.
5809  */
5810 
5811 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5812 {
5813         struct tcp_sock *tp = tcp_sk(sk);
5814         struct inet_connection_sock *icsk = inet_csk(sk);
5815         const struct tcphdr *th = tcp_hdr(skb);
5816         struct request_sock *req;
5817         int queued = 0;
5818         bool acceptable;
5819 
5820         switch (sk->sk_state) {
5821         case TCP_CLOSE:
5822                 goto discard;
5823 
5824         case TCP_LISTEN:
5825                 if (th->ack)
5826                         return 1;
5827 
5828                 if (th->rst)
5829                         goto discard;
5830 
5831                 if (th->syn) {
5832                         if (th->fin)
5833                                 goto discard;
5834                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5835                                 return 1;
5836 
5837                         consume_skb(skb);
5838                         return 0;
5839                 }
5840                 goto discard;
5841 
5842         case TCP_SYN_SENT:
5843                 tp->rx_opt.saw_tstamp = 0;
5844                 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5845                 if (queued >= 0)
5846                         return queued;
5847 
5848                 /* Do step6 onward by hand. */
5849                 tcp_urg(sk, skb, th);
5850                 __kfree_skb(skb);
5851                 tcp_data_snd_check(sk);
5852                 return 0;
5853         }
5854 
5855         tp->rx_opt.saw_tstamp = 0;
5856         req = tp->fastopen_rsk;
5857         if (req) {
5858                 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5859                     sk->sk_state != TCP_FIN_WAIT1);
5860 
5861                 if (!tcp_check_req(sk, skb, req, true))
5862                         goto discard;
5863         }
5864 
5865         if (!th->ack && !th->rst && !th->syn)
5866                 goto discard;
5867 
5868         if (!tcp_validate_incoming(sk, skb, th, 0))
5869                 return 0;
5870 
5871         /* step 5: check the ACK field */
5872         acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5873                                       FLAG_UPDATE_TS_RECENT) > 0;
5874 
5875         switch (sk->sk_state) {
5876         case TCP_SYN_RECV:
5877                 if (!acceptable)
5878                         return 1;
5879 
5880                 if (!tp->srtt_us)
5881                         tcp_synack_rtt_meas(sk, req);
5882 
5883                 /* Once we leave TCP_SYN_RECV, we no longer need req
5884                  * so release it.
5885                  */
5886                 if (req) {
5887                         inet_csk(sk)->icsk_retransmits = 0;
5888                         reqsk_fastopen_remove(sk, req, false);
5889                 } else {
5890                         /* Make sure socket is routed, for correct metrics. */
5891                         icsk->icsk_af_ops->rebuild_header(sk);
5892                         tcp_init_congestion_control(sk);
5893 
5894                         tcp_mtup_init(sk);
5895                         tp->copied_seq = tp->rcv_nxt;
5896                         tcp_init_buffer_space(sk);
5897                 }
5898                 smp_mb();
5899                 tcp_set_state(sk, TCP_ESTABLISHED);
5900                 sk->sk_state_change(sk);
5901 
5902                 /* Note, that this wakeup is only for marginal crossed SYN case.
5903                  * Passively open sockets are not waked up, because
5904                  * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5905                  */
5906                 if (sk->sk_socket)
5907                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5908 
5909                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5910                 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5911                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5912 
5913                 if (tp->rx_opt.tstamp_ok)
5914                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5915 
5916                 if (req) {
5917                         /* Re-arm the timer because data may have been sent out.
5918                          * This is similar to the regular data transmission case
5919                          * when new data has just been ack'ed.
5920                          *
5921                          * (TFO) - we could try to be more aggressive and
5922                          * retransmitting any data sooner based on when they
5923                          * are sent out.
5924                          */
5925                         tcp_rearm_rto(sk);
5926                 } else
5927                         tcp_init_metrics(sk);
5928 
5929                 tcp_update_pacing_rate(sk);
5930 
5931                 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5932                 tp->lsndtime = tcp_time_stamp;
5933 
5934                 tcp_initialize_rcv_mss(sk);
5935                 tcp_fast_path_on(tp);
5936                 break;
5937 
5938         case TCP_FIN_WAIT1: {
5939                 struct dst_entry *dst;
5940                 int tmo;
5941 
5942                 /* If we enter the TCP_FIN_WAIT1 state and we are a
5943                  * Fast Open socket and this is the first acceptable
5944                  * ACK we have received, this would have acknowledged
5945                  * our SYNACK so stop the SYNACK timer.
5946                  */
5947                 if (req) {
5948                         /* Return RST if ack_seq is invalid.
5949                          * Note that RFC793 only says to generate a
5950                          * DUPACK for it but for TCP Fast Open it seems
5951                          * better to treat this case like TCP_SYN_RECV
5952                          * above.
5953                          */
5954                         if (!acceptable)
5955                                 return 1;
5956                         /* We no longer need the request sock. */
5957                         reqsk_fastopen_remove(sk, req, false);
5958                         tcp_rearm_rto(sk);
5959                 }
5960                 if (tp->snd_una != tp->write_seq)
5961                         break;
5962 
5963                 tcp_set_state(sk, TCP_FIN_WAIT2);
5964                 sk->sk_shutdown |= SEND_SHUTDOWN;
5965 
5966                 dst = __sk_dst_get(sk);
5967                 if (dst)
5968                         dst_confirm(dst);
5969 
5970                 if (!sock_flag(sk, SOCK_DEAD)) {
5971                         /* Wake up lingering close() */
5972                         sk->sk_state_change(sk);
5973                         break;
5974                 }
5975 
5976                 if (tp->linger2 < 0 ||
5977                     (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5978                      after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5979                         tcp_done(sk);
5980                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5981                         return 1;
5982                 }
5983 
5984                 tmo = tcp_fin_time(sk);
5985                 if (tmo > TCP_TIMEWAIT_LEN) {
5986                         inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5987                 } else if (th->fin || sock_owned_by_user(sk)) {
5988                         /* Bad case. We could lose such FIN otherwise.
5989                          * It is not a big problem, but it looks confusing
5990                          * and not so rare event. We still can lose it now,
5991                          * if it spins in bh_lock_sock(), but it is really
5992                          * marginal case.
5993                          */
5994                         inet_csk_reset_keepalive_timer(sk, tmo);
5995                 } else {
5996                         tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5997                         goto discard;
5998                 }
5999                 break;
6000         }
6001 
6002         case TCP_CLOSING:
6003                 if (tp->snd_una == tp->write_seq) {
6004                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6005                         goto discard;
6006                 }
6007                 break;
6008 
6009         case TCP_LAST_ACK:
6010                 if (tp->snd_una == tp->write_seq) {
6011                         tcp_update_metrics(sk);
6012                         tcp_done(sk);
6013                         goto discard;
6014                 }
6015                 break;
6016         }
6017 
6018         /* step 6: check the URG bit */
6019         tcp_urg(sk, skb, th);
6020 
6021         /* step 7: process the segment text */
6022         switch (sk->sk_state) {
6023         case TCP_CLOSE_WAIT:
6024         case TCP_CLOSING:
6025         case TCP_LAST_ACK:
6026                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6027                         break;
6028         case TCP_FIN_WAIT1:
6029         case TCP_FIN_WAIT2:
6030                 /* RFC 793 says to queue data in these states,
6031                  * RFC 1122 says we MUST send a reset.
6032                  * BSD 4.4 also does reset.
6033                  */
6034                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6035                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6036                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6037                                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6038                                 tcp_reset(sk);
6039                                 return 1;
6040                         }
6041                 }
6042                 /* Fall through */
6043         case TCP_ESTABLISHED:
6044                 tcp_data_queue(sk, skb);
6045                 queued = 1;
6046                 break;
6047         }
6048 
6049         /* tcp_data could move socket to TIME-WAIT */
6050         if (sk->sk_state != TCP_CLOSE) {
6051                 tcp_data_snd_check(sk);
6052                 tcp_ack_snd_check(sk);
6053         }
6054 
6055         if (!queued) {
6056 discard:
6057                 tcp_drop(sk, skb);
6058         }
6059         return 0;
6060 }
6061 EXPORT_SYMBOL(tcp_rcv_state_process);
6062 
6063 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6064 {
6065         struct inet_request_sock *ireq = inet_rsk(req);
6066 
6067         if (family == AF_INET)
6068                 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6069                                     &ireq->ir_rmt_addr, port);
6070 #if IS_ENABLED(CONFIG_IPV6)
6071         else if (family == AF_INET6)
6072                 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6073                                     &ireq->ir_v6_rmt_addr, port);
6074 #endif
6075 }
6076 
6077 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6078  *
6079  * If we receive a SYN packet with these bits set, it means a
6080  * network is playing bad games with TOS bits. In order to
6081  * avoid possible false congestion notifications, we disable
6082  * TCP ECN negotiation.
6083  *
6084  * Exception: tcp_ca wants ECN. This is required for DCTCP
6085  * congestion control: Linux DCTCP asserts ECT on all packets,
6086  * including SYN, which is most optimal solution; however,
6087  * others, such as FreeBSD do not.
6088  */
6089 static void tcp_ecn_create_request(struct request_sock *req,
6090                                    const struct sk_buff *skb,
6091                                    const struct sock *listen_sk,
6092                                    const struct dst_entry *dst)
6093 {
6094         const struct tcphdr *th = tcp_hdr(skb);
6095         const struct net *net = sock_net(listen_sk);
6096         bool th_ecn = th->ece && th->cwr;
6097         bool ect, ecn_ok;
6098         u32 ecn_ok_dst;
6099 
6100         if (!th_ecn)
6101                 return;
6102 
6103         ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6104         ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6105         ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6106 
6107         if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6108             (ecn_ok_dst & DST_FEATURE_ECN_CA))
6109                 inet_rsk(req)->ecn_ok = 1;
6110 }
6111 
6112 static void tcp_openreq_init(struct request_sock *req,
6113                              const struct tcp_options_received *rx_opt,
6114                              struct sk_buff *skb, const struct sock *sk)
6115 {
6116         struct inet_request_sock *ireq = inet_rsk(req);
6117 
6118         req->rsk_rcv_wnd = 0;           /* So that tcp_send_synack() knows! */
6119         req->cookie_ts = 0;
6120         tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6121         tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6122         skb_mstamp_get(&tcp_rsk(req)->snt_synack);
6123         tcp_rsk(req)->last_oow_ack_time = 0;
6124         req->mss = rx_opt->mss_clamp;
6125         req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6126         ireq->tstamp_ok = rx_opt->tstamp_ok;
6127         ireq->sack_ok = rx_opt->sack_ok;
6128         ireq->snd_wscale = rx_opt->snd_wscale;
6129         ireq->wscale_ok = rx_opt->wscale_ok;
6130         ireq->acked = 0;
6131         ireq->ecn_ok = 0;
6132         ireq->ir_rmt_port = tcp_hdr(skb)->source;
6133         ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6134         ireq->ir_mark = inet_request_mark(sk, skb);
6135 }
6136 
6137 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6138                                       struct sock *sk_listener,
6139                                       bool attach_listener)
6140 {
6141         struct request_sock *req = reqsk_alloc(ops, sk_listener,
6142                                                attach_listener);
6143 
6144         if (req) {
6145                 struct inet_request_sock *ireq = inet_rsk(req);
6146 
6147                 kmemcheck_annotate_bitfield(ireq, flags);
6148                 ireq->opt = NULL;
6149 #if IS_ENABLED(CONFIG_IPV6)
6150                 ireq->pktopts = NULL;
6151 #endif
6152                 atomic64_set(&ireq->ir_cookie, 0);
6153                 ireq->ireq_state = TCP_NEW_SYN_RECV;
6154                 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6155                 ireq->ireq_family = sk_listener->sk_family;
6156         }
6157 
6158         return req;
6159 }
6160 EXPORT_SYMBOL(inet_reqsk_alloc);
6161 
6162 /*
6163  * Return true if a syncookie should be sent
6164  */
6165 static bool tcp_syn_flood_action(const struct sock *sk,
6166                                  const struct sk_buff *skb,
6167                                  const char *proto)
6168 {
6169         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6170         const char *msg = "Dropping request";
6171         bool want_cookie = false;
6172         struct net *net = sock_net(sk);
6173 
6174 #ifdef CONFIG_SYN_COOKIES
6175         if (net->ipv4.sysctl_tcp_syncookies) {
6176                 msg = "Sending cookies";
6177                 want_cookie = true;
6178                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6179         } else
6180 #endif
6181                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6182 
6183         if (!queue->synflood_warned &&
6184             net->ipv4.sysctl_tcp_syncookies != 2 &&
6185             xchg(&queue->synflood_warned, 1) == 0)
6186                 pr_info("%s: Possible SYN flooding on port %d. %s.  Check SNMP counters.\n",
6187                         proto, ntohs(tcp_hdr(skb)->dest), msg);
6188 
6189         return want_cookie;
6190 }
6191 
6192 static void tcp_reqsk_record_syn(const struct sock *sk,
6193                                  struct request_sock *req,
6194                                  const struct sk_buff *skb)
6195 {
6196         if (tcp_sk(sk)->save_syn) {
6197                 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6198                 u32 *copy;
6199 
6200                 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6201                 if (copy) {
6202                         copy[0] = len;
6203                         memcpy(&copy[1], skb_network_header(skb), len);
6204                         req->saved_syn = copy;
6205                 }
6206         }
6207 }
6208 
6209 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6210                      const struct tcp_request_sock_ops *af_ops,
6211                      struct sock *sk, struct sk_buff *skb)
6212 {
6213         struct tcp_fastopen_cookie foc = { .len = -1 };
6214         __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6215         struct tcp_options_received tmp_opt;
6216         struct tcp_sock *tp = tcp_sk(sk);
6217         struct net *net = sock_net(sk);
6218         struct sock *fastopen_sk = NULL;
6219         struct dst_entry *dst = NULL;
6220         struct request_sock *req;
6221         bool want_cookie = false;
6222         struct flowi fl;
6223 
6224         /* TW buckets are converted to open requests without
6225          * limitations, they conserve resources and peer is
6226          * evidently real one.
6227          */
6228         if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6229              inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6230                 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6231                 if (!want_cookie)
6232                         goto drop;
6233         }
6234 
6235 
6236         /* Accept backlog is full. If we have already queued enough
6237          * of warm entries in syn queue, drop request. It is better than
6238          * clogging syn queue with openreqs with exponentially increasing
6239          * timeout.
6240          */
6241         if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
6242                 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6243                 goto drop;
6244         }
6245 
6246         req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6247         if (!req)
6248                 goto drop;
6249 
6250         tcp_rsk(req)->af_specific = af_ops;
6251 
6252         tcp_clear_options(&tmp_opt);
6253         tmp_opt.mss_clamp = af_ops->mss_clamp;
6254         tmp_opt.user_mss  = tp->rx_opt.user_mss;
6255         tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6256 
6257         if (want_cookie && !tmp_opt.saw_tstamp)
6258                 tcp_clear_options(&tmp_opt);
6259 
6260         tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6261         tcp_openreq_init(req, &tmp_opt, skb, sk);
6262 
6263         /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6264         inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6265 
6266         af_ops->init_req(req, sk, skb);
6267 
6268         if (security_inet_conn_request(sk, skb, req))
6269                 goto drop_and_free;
6270 
6271         if (!want_cookie && !isn) {
6272                 /* VJ's idea. We save last timestamp seen
6273                  * from the destination in peer table, when entering
6274                  * state TIME-WAIT, and check against it before
6275                  * accepting new connection request.
6276                  *
6277                  * If "isn" is not zero, this request hit alive
6278                  * timewait bucket, so that all the necessary checks
6279                  * are made in the function processing timewait state.
6280                  */
6281                 if (tcp_death_row.sysctl_tw_recycle) {
6282                         bool strict;
6283 
6284                         dst = af_ops->route_req(sk, &fl, req, &strict);
6285 
6286                         if (dst && strict &&
6287                             !tcp_peer_is_proven(req, dst, true,
6288                                                 tmp_opt.saw_tstamp)) {
6289                                 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
6290                                 goto drop_and_release;
6291                         }
6292                 }
6293                 /* Kill the following clause, if you dislike this way. */
6294                 else if (!net->ipv4.sysctl_tcp_syncookies &&
6295                          (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6296                           (sysctl_max_syn_backlog >> 2)) &&
6297                          !tcp_peer_is_proven(req, dst, false,
6298                                              tmp_opt.saw_tstamp)) {
6299                         /* Without syncookies last quarter of
6300                          * backlog is filled with destinations,
6301                          * proven to be alive.
6302                          * It means that we continue to communicate
6303                          * to destinations, already remembered
6304                          * to the moment of synflood.
6305                          */
6306                         pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6307                                     rsk_ops->family);
6308                         goto drop_and_release;
6309                 }
6310 
6311                 isn = af_ops->init_seq(skb);
6312         }
6313         if (!dst) {
6314                 dst = af_ops->route_req(sk, &fl, req, NULL);
6315                 if (!dst)
6316                         goto drop_and_free;
6317         }
6318 
6319         tcp_ecn_create_request(req, skb, sk, dst);
6320 
6321         if (want_cookie) {
6322                 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6323                 req->cookie_ts = tmp_opt.tstamp_ok;
6324                 if (!tmp_opt.tstamp_ok)
6325                         inet_rsk(req)->ecn_ok = 0;
6326         }
6327 
6328         tcp_rsk(req)->snt_isn = isn;
6329         tcp_rsk(req)->txhash = net_tx_rndhash();
6330         tcp_openreq_init_rwin(req, sk, dst);
6331         if (!want_cookie) {
6332                 tcp_reqsk_record_syn(sk, req, skb);
6333                 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6334         }
6335         if (fastopen_sk) {
6336                 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6337                                     &foc, TCP_SYNACK_FASTOPEN);
6338                 /* Add the child socket directly into the accept queue */
6339                 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6340                 sk->sk_data_ready(sk);
6341                 bh_unlock_sock(fastopen_sk);
6342                 sock_put(fastopen_sk);
6343         } else {
6344                 tcp_rsk(req)->tfo_listener = false;
6345                 if (!want_cookie)
6346                         inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6347                 af_ops->send_synack(sk, dst, &fl, req, &foc,
6348                                     !want_cookie ? TCP_SYNACK_NORMAL :
6349                                                    TCP_SYNACK_COOKIE);
6350                 if (want_cookie) {
6351                         reqsk_free(req);
6352                         return 0;
6353                 }
6354         }
6355         reqsk_put(req);
6356         return 0;
6357 
6358 drop_and_release:
6359         dst_release(dst);
6360 drop_and_free:
6361         reqsk_free(req);
6362 drop:
6363         tcp_listendrop(sk);
6364         return 0;
6365 }
6366 EXPORT_SYMBOL(tcp_conn_request);
6367 

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