Version:  2.0.40 2.2.26 2.4.37 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 4.9

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 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb)
132 {
133         static bool __once __read_mostly;
134 
135         if (!__once) {
136                 struct net_device *dev;
137 
138                 __once = true;
139 
140                 rcu_read_lock();
141                 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
142                 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
143                         dev ? dev->name : "Unknown driver");
144                 rcu_read_unlock();
145         }
146 }
147 
148 /* Adapt the MSS value used to make delayed ack decision to the
149  * real world.
150  */
151 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
152 {
153         struct inet_connection_sock *icsk = inet_csk(sk);
154         const unsigned int lss = icsk->icsk_ack.last_seg_size;
155         unsigned int len;
156 
157         icsk->icsk_ack.last_seg_size = 0;
158 
159         /* skb->len may jitter because of SACKs, even if peer
160          * sends good full-sized frames.
161          */
162         len = skb_shinfo(skb)->gso_size ? : skb->len;
163         if (len >= icsk->icsk_ack.rcv_mss) {
164                 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
165                                                tcp_sk(sk)->advmss);
166                 if (unlikely(icsk->icsk_ack.rcv_mss != len))
167                         tcp_gro_dev_warn(sk, skb);
168         } else {
169                 /* Otherwise, we make more careful check taking into account,
170                  * that SACKs block is variable.
171                  *
172                  * "len" is invariant segment length, including TCP header.
173                  */
174                 len += skb->data - skb_transport_header(skb);
175                 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
176                     /* If PSH is not set, packet should be
177                      * full sized, provided peer TCP is not badly broken.
178                      * This observation (if it is correct 8)) allows
179                      * to handle super-low mtu links fairly.
180                      */
181                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
182                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
183                         /* Subtract also invariant (if peer is RFC compliant),
184                          * tcp header plus fixed timestamp option length.
185                          * Resulting "len" is MSS free of SACK jitter.
186                          */
187                         len -= tcp_sk(sk)->tcp_header_len;
188                         icsk->icsk_ack.last_seg_size = len;
189                         if (len == lss) {
190                                 icsk->icsk_ack.rcv_mss = len;
191                                 return;
192                         }
193                 }
194                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
195                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
196                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
197         }
198 }
199 
200 static void tcp_incr_quickack(struct sock *sk)
201 {
202         struct inet_connection_sock *icsk = inet_csk(sk);
203         unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
204 
205         if (quickacks == 0)
206                 quickacks = 2;
207         if (quickacks > icsk->icsk_ack.quick)
208                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
209 }
210 
211 static void tcp_enter_quickack_mode(struct sock *sk)
212 {
213         struct inet_connection_sock *icsk = inet_csk(sk);
214         tcp_incr_quickack(sk);
215         icsk->icsk_ack.pingpong = 0;
216         icsk->icsk_ack.ato = TCP_ATO_MIN;
217 }
218 
219 /* Send ACKs quickly, if "quick" count is not exhausted
220  * and the session is not interactive.
221  */
222 
223 static bool tcp_in_quickack_mode(struct sock *sk)
224 {
225         const struct inet_connection_sock *icsk = inet_csk(sk);
226         const struct dst_entry *dst = __sk_dst_get(sk);
227 
228         return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
229                 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
230 }
231 
232 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
233 {
234         if (tp->ecn_flags & TCP_ECN_OK)
235                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
236 }
237 
238 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
239 {
240         if (tcp_hdr(skb)->cwr)
241                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
242 }
243 
244 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
245 {
246         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
247 }
248 
249 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
250 {
251         switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
252         case INET_ECN_NOT_ECT:
253                 /* Funny extension: if ECT is not set on a segment,
254                  * and we already seen ECT on a previous segment,
255                  * it is probably a retransmit.
256                  */
257                 if (tp->ecn_flags & TCP_ECN_SEEN)
258                         tcp_enter_quickack_mode((struct sock *)tp);
259                 break;
260         case INET_ECN_CE:
261                 if (tcp_ca_needs_ecn((struct sock *)tp))
262                         tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
263 
264                 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
265                         /* Better not delay acks, sender can have a very low cwnd */
266                         tcp_enter_quickack_mode((struct sock *)tp);
267                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
268                 }
269                 tp->ecn_flags |= TCP_ECN_SEEN;
270                 break;
271         default:
272                 if (tcp_ca_needs_ecn((struct sock *)tp))
273                         tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
274                 tp->ecn_flags |= TCP_ECN_SEEN;
275                 break;
276         }
277 }
278 
279 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
280 {
281         if (tp->ecn_flags & TCP_ECN_OK)
282                 __tcp_ecn_check_ce(tp, skb);
283 }
284 
285 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
286 {
287         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
288                 tp->ecn_flags &= ~TCP_ECN_OK;
289 }
290 
291 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
292 {
293         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
294                 tp->ecn_flags &= ~TCP_ECN_OK;
295 }
296 
297 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
298 {
299         if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
300                 return true;
301         return false;
302 }
303 
304 /* Buffer size and advertised window tuning.
305  *
306  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
307  */
308 
309 static void tcp_sndbuf_expand(struct sock *sk)
310 {
311         const struct tcp_sock *tp = tcp_sk(sk);
312         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
313         int sndmem, per_mss;
314         u32 nr_segs;
315 
316         /* Worst case is non GSO/TSO : each frame consumes one skb
317          * and skb->head is kmalloced using power of two area of memory
318          */
319         per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
320                   MAX_TCP_HEADER +
321                   SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
322 
323         per_mss = roundup_pow_of_two(per_mss) +
324                   SKB_DATA_ALIGN(sizeof(struct sk_buff));
325 
326         nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
327         nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
328 
329         /* Fast Recovery (RFC 5681 3.2) :
330          * Cubic needs 1.7 factor, rounded to 2 to include
331          * extra cushion (application might react slowly to POLLOUT)
332          */
333         sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
334         sndmem *= nr_segs * per_mss;
335 
336         if (sk->sk_sndbuf < sndmem)
337                 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
338 }
339 
340 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
341  *
342  * All tcp_full_space() is split to two parts: "network" buffer, allocated
343  * forward and advertised in receiver window (tp->rcv_wnd) and
344  * "application buffer", required to isolate scheduling/application
345  * latencies from network.
346  * window_clamp is maximal advertised window. It can be less than
347  * tcp_full_space(), in this case tcp_full_space() - window_clamp
348  * is reserved for "application" buffer. The less window_clamp is
349  * the smoother our behaviour from viewpoint of network, but the lower
350  * throughput and the higher sensitivity of the connection to losses. 8)
351  *
352  * rcv_ssthresh is more strict window_clamp used at "slow start"
353  * phase to predict further behaviour of this connection.
354  * It is used for two goals:
355  * - to enforce header prediction at sender, even when application
356  *   requires some significant "application buffer". It is check #1.
357  * - to prevent pruning of receive queue because of misprediction
358  *   of receiver window. Check #2.
359  *
360  * The scheme does not work when sender sends good segments opening
361  * window and then starts to feed us spaghetti. But it should work
362  * in common situations. Otherwise, we have to rely on queue collapsing.
363  */
364 
365 /* Slow part of check#2. */
366 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
367 {
368         struct tcp_sock *tp = tcp_sk(sk);
369         /* Optimize this! */
370         int truesize = tcp_win_from_space(skb->truesize) >> 1;
371         int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
372 
373         while (tp->rcv_ssthresh <= window) {
374                 if (truesize <= skb->len)
375                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
376 
377                 truesize >>= 1;
378                 window >>= 1;
379         }
380         return 0;
381 }
382 
383 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
384 {
385         struct tcp_sock *tp = tcp_sk(sk);
386 
387         /* Check #1 */
388         if (tp->rcv_ssthresh < tp->window_clamp &&
389             (int)tp->rcv_ssthresh < tcp_space(sk) &&
390             !tcp_under_memory_pressure(sk)) {
391                 int incr;
392 
393                 /* Check #2. Increase window, if skb with such overhead
394                  * will fit to rcvbuf in future.
395                  */
396                 if (tcp_win_from_space(skb->truesize) <= skb->len)
397                         incr = 2 * tp->advmss;
398                 else
399                         incr = __tcp_grow_window(sk, skb);
400 
401                 if (incr) {
402                         incr = max_t(int, incr, 2 * skb->len);
403                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
404                                                tp->window_clamp);
405                         inet_csk(sk)->icsk_ack.quick |= 1;
406                 }
407         }
408 }
409 
410 /* 3. Tuning rcvbuf, when connection enters established state. */
411 static void tcp_fixup_rcvbuf(struct sock *sk)
412 {
413         u32 mss = tcp_sk(sk)->advmss;
414         int rcvmem;
415 
416         rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
417                  tcp_default_init_rwnd(mss);
418 
419         /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
420          * Allow enough cushion so that sender is not limited by our window
421          */
422         if (sysctl_tcp_moderate_rcvbuf)
423                 rcvmem <<= 2;
424 
425         if (sk->sk_rcvbuf < rcvmem)
426                 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
427 }
428 
429 /* 4. Try to fixup all. It is made immediately after connection enters
430  *    established state.
431  */
432 void tcp_init_buffer_space(struct sock *sk)
433 {
434         struct tcp_sock *tp = tcp_sk(sk);
435         int maxwin;
436 
437         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
438                 tcp_fixup_rcvbuf(sk);
439         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
440                 tcp_sndbuf_expand(sk);
441 
442         tp->rcvq_space.space = tp->rcv_wnd;
443         tp->rcvq_space.time = tcp_time_stamp;
444         tp->rcvq_space.seq = tp->copied_seq;
445 
446         maxwin = tcp_full_space(sk);
447 
448         if (tp->window_clamp >= maxwin) {
449                 tp->window_clamp = maxwin;
450 
451                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
452                         tp->window_clamp = max(maxwin -
453                                                (maxwin >> sysctl_tcp_app_win),
454                                                4 * tp->advmss);
455         }
456 
457         /* Force reservation of one segment. */
458         if (sysctl_tcp_app_win &&
459             tp->window_clamp > 2 * tp->advmss &&
460             tp->window_clamp + tp->advmss > maxwin)
461                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
462 
463         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
464         tp->snd_cwnd_stamp = tcp_time_stamp;
465 }
466 
467 /* 5. Recalculate window clamp after socket hit its memory bounds. */
468 static void tcp_clamp_window(struct sock *sk)
469 {
470         struct tcp_sock *tp = tcp_sk(sk);
471         struct inet_connection_sock *icsk = inet_csk(sk);
472 
473         icsk->icsk_ack.quick = 0;
474 
475         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
476             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
477             !tcp_under_memory_pressure(sk) &&
478             sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
479                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
480                                     sysctl_tcp_rmem[2]);
481         }
482         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
483                 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
484 }
485 
486 /* Initialize RCV_MSS value.
487  * RCV_MSS is an our guess about MSS used by the peer.
488  * We haven't any direct information about the MSS.
489  * It's better to underestimate the RCV_MSS rather than overestimate.
490  * Overestimations make us ACKing less frequently than needed.
491  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
492  */
493 void tcp_initialize_rcv_mss(struct sock *sk)
494 {
495         const struct tcp_sock *tp = tcp_sk(sk);
496         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
497 
498         hint = min(hint, tp->rcv_wnd / 2);
499         hint = min(hint, TCP_MSS_DEFAULT);
500         hint = max(hint, TCP_MIN_MSS);
501 
502         inet_csk(sk)->icsk_ack.rcv_mss = hint;
503 }
504 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
505 
506 /* Receiver "autotuning" code.
507  *
508  * The algorithm for RTT estimation w/o timestamps is based on
509  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
510  * <http://public.lanl.gov/radiant/pubs.html#DRS>
511  *
512  * More detail on this code can be found at
513  * <http://staff.psc.edu/jheffner/>,
514  * though this reference is out of date.  A new paper
515  * is pending.
516  */
517 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
518 {
519         u32 new_sample = tp->rcv_rtt_est.rtt;
520         long m = sample;
521 
522         if (m == 0)
523                 m = 1;
524 
525         if (new_sample != 0) {
526                 /* If we sample in larger samples in the non-timestamp
527                  * case, we could grossly overestimate the RTT especially
528                  * with chatty applications or bulk transfer apps which
529                  * are stalled on filesystem I/O.
530                  *
531                  * Also, since we are only going for a minimum in the
532                  * non-timestamp case, we do not smooth things out
533                  * else with timestamps disabled convergence takes too
534                  * long.
535                  */
536                 if (!win_dep) {
537                         m -= (new_sample >> 3);
538                         new_sample += m;
539                 } else {
540                         m <<= 3;
541                         if (m < new_sample)
542                                 new_sample = m;
543                 }
544         } else {
545                 /* No previous measure. */
546                 new_sample = m << 3;
547         }
548 
549         if (tp->rcv_rtt_est.rtt != new_sample)
550                 tp->rcv_rtt_est.rtt = new_sample;
551 }
552 
553 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
554 {
555         if (tp->rcv_rtt_est.time == 0)
556                 goto new_measure;
557         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
558                 return;
559         tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
560 
561 new_measure:
562         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
563         tp->rcv_rtt_est.time = tcp_time_stamp;
564 }
565 
566 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
567                                           const struct sk_buff *skb)
568 {
569         struct tcp_sock *tp = tcp_sk(sk);
570         if (tp->rx_opt.rcv_tsecr &&
571             (TCP_SKB_CB(skb)->end_seq -
572              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
573                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
574 }
575 
576 /*
577  * This function should be called every time data is copied to user space.
578  * It calculates the appropriate TCP receive buffer space.
579  */
580 void tcp_rcv_space_adjust(struct sock *sk)
581 {
582         struct tcp_sock *tp = tcp_sk(sk);
583         int time;
584         int copied;
585 
586         time = tcp_time_stamp - tp->rcvq_space.time;
587         if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
588                 return;
589 
590         /* Number of bytes copied to user in last RTT */
591         copied = tp->copied_seq - tp->rcvq_space.seq;
592         if (copied <= tp->rcvq_space.space)
593                 goto new_measure;
594 
595         /* A bit of theory :
596          * copied = bytes received in previous RTT, our base window
597          * To cope with packet losses, we need a 2x factor
598          * To cope with slow start, and sender growing its cwin by 100 %
599          * every RTT, we need a 4x factor, because the ACK we are sending
600          * now is for the next RTT, not the current one :
601          * <prev RTT . ><current RTT .. ><next RTT .... >
602          */
603 
604         if (sysctl_tcp_moderate_rcvbuf &&
605             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
606                 int rcvwin, rcvmem, rcvbuf;
607 
608                 /* minimal window to cope with packet losses, assuming
609                  * steady state. Add some cushion because of small variations.
610                  */
611                 rcvwin = (copied << 1) + 16 * tp->advmss;
612 
613                 /* If rate increased by 25%,
614                  *      assume slow start, rcvwin = 3 * copied
615                  * If rate increased by 50%,
616                  *      assume sender can use 2x growth, rcvwin = 4 * copied
617                  */
618                 if (copied >=
619                     tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
620                         if (copied >=
621                             tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
622                                 rcvwin <<= 1;
623                         else
624                                 rcvwin += (rcvwin >> 1);
625                 }
626 
627                 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
628                 while (tcp_win_from_space(rcvmem) < tp->advmss)
629                         rcvmem += 128;
630 
631                 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
632                 if (rcvbuf > sk->sk_rcvbuf) {
633                         sk->sk_rcvbuf = rcvbuf;
634 
635                         /* Make the window clamp follow along.  */
636                         tp->window_clamp = rcvwin;
637                 }
638         }
639         tp->rcvq_space.space = copied;
640 
641 new_measure:
642         tp->rcvq_space.seq = tp->copied_seq;
643         tp->rcvq_space.time = tcp_time_stamp;
644 }
645 
646 /* There is something which you must keep in mind when you analyze the
647  * behavior of the tp->ato delayed ack timeout interval.  When a
648  * connection starts up, we want to ack as quickly as possible.  The
649  * problem is that "good" TCP's do slow start at the beginning of data
650  * transmission.  The means that until we send the first few ACK's the
651  * sender will sit on his end and only queue most of his data, because
652  * he can only send snd_cwnd unacked packets at any given time.  For
653  * each ACK we send, he increments snd_cwnd and transmits more of his
654  * queue.  -DaveM
655  */
656 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
657 {
658         struct tcp_sock *tp = tcp_sk(sk);
659         struct inet_connection_sock *icsk = inet_csk(sk);
660         u32 now;
661 
662         inet_csk_schedule_ack(sk);
663 
664         tcp_measure_rcv_mss(sk, skb);
665 
666         tcp_rcv_rtt_measure(tp);
667 
668         now = tcp_time_stamp;
669 
670         if (!icsk->icsk_ack.ato) {
671                 /* The _first_ data packet received, initialize
672                  * delayed ACK engine.
673                  */
674                 tcp_incr_quickack(sk);
675                 icsk->icsk_ack.ato = TCP_ATO_MIN;
676         } else {
677                 int m = now - icsk->icsk_ack.lrcvtime;
678 
679                 if (m <= TCP_ATO_MIN / 2) {
680                         /* The fastest case is the first. */
681                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
682                 } else if (m < icsk->icsk_ack.ato) {
683                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
684                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
685                                 icsk->icsk_ack.ato = icsk->icsk_rto;
686                 } else if (m > icsk->icsk_rto) {
687                         /* Too long gap. Apparently sender failed to
688                          * restart window, so that we send ACKs quickly.
689                          */
690                         tcp_incr_quickack(sk);
691                         sk_mem_reclaim(sk);
692                 }
693         }
694         icsk->icsk_ack.lrcvtime = now;
695 
696         tcp_ecn_check_ce(tp, skb);
697 
698         if (skb->len >= 128)
699                 tcp_grow_window(sk, skb);
700 }
701 
702 /* Called to compute a smoothed rtt estimate. The data fed to this
703  * routine either comes from timestamps, or from segments that were
704  * known _not_ to have been retransmitted [see Karn/Partridge
705  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
706  * piece by Van Jacobson.
707  * NOTE: the next three routines used to be one big routine.
708  * To save cycles in the RFC 1323 implementation it was better to break
709  * it up into three procedures. -- erics
710  */
711 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
712 {
713         struct tcp_sock *tp = tcp_sk(sk);
714         long m = mrtt_us; /* RTT */
715         u32 srtt = tp->srtt_us;
716 
717         /*      The following amusing code comes from Jacobson's
718          *      article in SIGCOMM '88.  Note that rtt and mdev
719          *      are scaled versions of rtt and mean deviation.
720          *      This is designed to be as fast as possible
721          *      m stands for "measurement".
722          *
723          *      On a 1990 paper the rto value is changed to:
724          *      RTO = rtt + 4 * mdev
725          *
726          * Funny. This algorithm seems to be very broken.
727          * These formulae increase RTO, when it should be decreased, increase
728          * too slowly, when it should be increased quickly, decrease too quickly
729          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
730          * does not matter how to _calculate_ it. Seems, it was trap
731          * that VJ failed to avoid. 8)
732          */
733         if (srtt != 0) {
734                 m -= (srtt >> 3);       /* m is now error in rtt est */
735                 srtt += m;              /* rtt = 7/8 rtt + 1/8 new */
736                 if (m < 0) {
737                         m = -m;         /* m is now abs(error) */
738                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
739                         /* This is similar to one of Eifel findings.
740                          * Eifel blocks mdev updates when rtt decreases.
741                          * This solution is a bit different: we use finer gain
742                          * for mdev in this case (alpha*beta).
743                          * Like Eifel it also prevents growth of rto,
744                          * but also it limits too fast rto decreases,
745                          * happening in pure Eifel.
746                          */
747                         if (m > 0)
748                                 m >>= 3;
749                 } else {
750                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
751                 }
752                 tp->mdev_us += m;               /* mdev = 3/4 mdev + 1/4 new */
753                 if (tp->mdev_us > tp->mdev_max_us) {
754                         tp->mdev_max_us = tp->mdev_us;
755                         if (tp->mdev_max_us > tp->rttvar_us)
756                                 tp->rttvar_us = tp->mdev_max_us;
757                 }
758                 if (after(tp->snd_una, tp->rtt_seq)) {
759                         if (tp->mdev_max_us < tp->rttvar_us)
760                                 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
761                         tp->rtt_seq = tp->snd_nxt;
762                         tp->mdev_max_us = tcp_rto_min_us(sk);
763                 }
764         } else {
765                 /* no previous measure. */
766                 srtt = m << 3;          /* take the measured time to be rtt */
767                 tp->mdev_us = m << 1;   /* make sure rto = 3*rtt */
768                 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
769                 tp->mdev_max_us = tp->rttvar_us;
770                 tp->rtt_seq = tp->snd_nxt;
771         }
772         tp->srtt_us = max(1U, srtt);
773 }
774 
775 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
776  * Note: TCP stack does not yet implement pacing.
777  * FQ packet scheduler can be used to implement cheap but effective
778  * TCP pacing, to smooth the burst on large writes when packets
779  * in flight is significantly lower than cwnd (or rwin)
780  */
781 int sysctl_tcp_pacing_ss_ratio __read_mostly = 200;
782 int sysctl_tcp_pacing_ca_ratio __read_mostly = 120;
783 
784 static void tcp_update_pacing_rate(struct sock *sk)
785 {
786         const struct tcp_sock *tp = tcp_sk(sk);
787         u64 rate;
788 
789         /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
790         rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
791 
792         /* current rate is (cwnd * mss) / srtt
793          * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
794          * In Congestion Avoidance phase, set it to 120 % the current rate.
795          *
796          * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
797          *       If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
798          *       end of slow start and should slow down.
799          */
800         if (tp->snd_cwnd < tp->snd_ssthresh / 2)
801                 rate *= sysctl_tcp_pacing_ss_ratio;
802         else
803                 rate *= sysctl_tcp_pacing_ca_ratio;
804 
805         rate *= max(tp->snd_cwnd, tp->packets_out);
806 
807         if (likely(tp->srtt_us))
808                 do_div(rate, tp->srtt_us);
809 
810         /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
811          * without any lock. We want to make sure compiler wont store
812          * intermediate values in this location.
813          */
814         ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
815                                                 sk->sk_max_pacing_rate);
816 }
817 
818 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
819  * routine referred to above.
820  */
821 static void tcp_set_rto(struct sock *sk)
822 {
823         const struct tcp_sock *tp = tcp_sk(sk);
824         /* Old crap is replaced with new one. 8)
825          *
826          * More seriously:
827          * 1. If rtt variance happened to be less 50msec, it is hallucination.
828          *    It cannot be less due to utterly erratic ACK generation made
829          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
830          *    to do with delayed acks, because at cwnd>2 true delack timeout
831          *    is invisible. Actually, Linux-2.4 also generates erratic
832          *    ACKs in some circumstances.
833          */
834         inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
835 
836         /* 2. Fixups made earlier cannot be right.
837          *    If we do not estimate RTO correctly without them,
838          *    all the algo is pure shit and should be replaced
839          *    with correct one. It is exactly, which we pretend to do.
840          */
841 
842         /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
843          * guarantees that rto is higher.
844          */
845         tcp_bound_rto(sk);
846 }
847 
848 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
849 {
850         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
851 
852         if (!cwnd)
853                 cwnd = TCP_INIT_CWND;
854         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
855 }
856 
857 /*
858  * Packet counting of FACK is based on in-order assumptions, therefore TCP
859  * disables it when reordering is detected
860  */
861 void tcp_disable_fack(struct tcp_sock *tp)
862 {
863         /* RFC3517 uses different metric in lost marker => reset on change */
864         if (tcp_is_fack(tp))
865                 tp->lost_skb_hint = NULL;
866         tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
867 }
868 
869 /* Take a notice that peer is sending D-SACKs */
870 static void tcp_dsack_seen(struct tcp_sock *tp)
871 {
872         tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
873 }
874 
875 static void tcp_update_reordering(struct sock *sk, const int metric,
876                                   const int ts)
877 {
878         struct tcp_sock *tp = tcp_sk(sk);
879         if (metric > tp->reordering) {
880                 int mib_idx;
881 
882                 tp->reordering = min(sysctl_tcp_max_reordering, metric);
883 
884                 /* This exciting event is worth to be remembered. 8) */
885                 if (ts)
886                         mib_idx = LINUX_MIB_TCPTSREORDER;
887                 else if (tcp_is_reno(tp))
888                         mib_idx = LINUX_MIB_TCPRENOREORDER;
889                 else if (tcp_is_fack(tp))
890                         mib_idx = LINUX_MIB_TCPFACKREORDER;
891                 else
892                         mib_idx = LINUX_MIB_TCPSACKREORDER;
893 
894                 NET_INC_STATS(sock_net(sk), mib_idx);
895 #if FASTRETRANS_DEBUG > 1
896                 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
897                          tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
898                          tp->reordering,
899                          tp->fackets_out,
900                          tp->sacked_out,
901                          tp->undo_marker ? tp->undo_retrans : 0);
902 #endif
903                 tcp_disable_fack(tp);
904         }
905 
906         if (metric > 0)
907                 tcp_disable_early_retrans(tp);
908         tp->rack.reord = 1;
909 }
910 
911 /* This must be called before lost_out is incremented */
912 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
913 {
914         if (!tp->retransmit_skb_hint ||
915             before(TCP_SKB_CB(skb)->seq,
916                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
917                 tp->retransmit_skb_hint = skb;
918 
919         if (!tp->lost_out ||
920             after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
921                 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
922 }
923 
924 /* Sum the number of packets on the wire we have marked as lost.
925  * There are two cases we care about here:
926  * a) Packet hasn't been marked lost (nor retransmitted),
927  *    and this is the first loss.
928  * b) Packet has been marked both lost and retransmitted,
929  *    and this means we think it was lost again.
930  */
931 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
932 {
933         __u8 sacked = TCP_SKB_CB(skb)->sacked;
934 
935         if (!(sacked & TCPCB_LOST) ||
936             ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
937                 tp->lost += tcp_skb_pcount(skb);
938 }
939 
940 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
941 {
942         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
943                 tcp_verify_retransmit_hint(tp, skb);
944 
945                 tp->lost_out += tcp_skb_pcount(skb);
946                 tcp_sum_lost(tp, skb);
947                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
948         }
949 }
950 
951 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
952 {
953         tcp_verify_retransmit_hint(tp, skb);
954 
955         tcp_sum_lost(tp, skb);
956         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
957                 tp->lost_out += tcp_skb_pcount(skb);
958                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
959         }
960 }
961 
962 /* This procedure tags the retransmission queue when SACKs arrive.
963  *
964  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
965  * Packets in queue with these bits set are counted in variables
966  * sacked_out, retrans_out and lost_out, correspondingly.
967  *
968  * Valid combinations are:
969  * Tag  InFlight        Description
970  * 0    1               - orig segment is in flight.
971  * S    0               - nothing flies, orig reached receiver.
972  * L    0               - nothing flies, orig lost by net.
973  * R    2               - both orig and retransmit are in flight.
974  * L|R  1               - orig is lost, retransmit is in flight.
975  * S|R  1               - orig reached receiver, retrans is still in flight.
976  * (L|S|R is logically valid, it could occur when L|R is sacked,
977  *  but it is equivalent to plain S and code short-curcuits it to S.
978  *  L|S is logically invalid, it would mean -1 packet in flight 8))
979  *
980  * These 6 states form finite state machine, controlled by the following events:
981  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
982  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
983  * 3. Loss detection event of two flavors:
984  *      A. Scoreboard estimator decided the packet is lost.
985  *         A'. Reno "three dupacks" marks head of queue lost.
986  *         A''. Its FACK modification, head until snd.fack is lost.
987  *      B. SACK arrives sacking SND.NXT at the moment, when the
988  *         segment was retransmitted.
989  * 4. D-SACK added new rule: D-SACK changes any tag to S.
990  *
991  * It is pleasant to note, that state diagram turns out to be commutative,
992  * so that we are allowed not to be bothered by order of our actions,
993  * when multiple events arrive simultaneously. (see the function below).
994  *
995  * Reordering detection.
996  * --------------------
997  * Reordering metric is maximal distance, which a packet can be displaced
998  * in packet stream. With SACKs we can estimate it:
999  *
1000  * 1. SACK fills old hole and the corresponding segment was not
1001  *    ever retransmitted -> reordering. Alas, we cannot use it
1002  *    when segment was retransmitted.
1003  * 2. The last flaw is solved with D-SACK. D-SACK arrives
1004  *    for retransmitted and already SACKed segment -> reordering..
1005  * Both of these heuristics are not used in Loss state, when we cannot
1006  * account for retransmits accurately.
1007  *
1008  * SACK block validation.
1009  * ----------------------
1010  *
1011  * SACK block range validation checks that the received SACK block fits to
1012  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1013  * Note that SND.UNA is not included to the range though being valid because
1014  * it means that the receiver is rather inconsistent with itself reporting
1015  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1016  * perfectly valid, however, in light of RFC2018 which explicitly states
1017  * that "SACK block MUST reflect the newest segment.  Even if the newest
1018  * segment is going to be discarded ...", not that it looks very clever
1019  * in case of head skb. Due to potentional receiver driven attacks, we
1020  * choose to avoid immediate execution of a walk in write queue due to
1021  * reneging and defer head skb's loss recovery to standard loss recovery
1022  * procedure that will eventually trigger (nothing forbids us doing this).
1023  *
1024  * Implements also blockage to start_seq wrap-around. Problem lies in the
1025  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1026  * there's no guarantee that it will be before snd_nxt (n). The problem
1027  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1028  * wrap (s_w):
1029  *
1030  *         <- outs wnd ->                          <- wrapzone ->
1031  *         u     e      n                         u_w   e_w  s n_w
1032  *         |     |      |                          |     |   |  |
1033  * |<------------+------+----- TCP seqno space --------------+---------->|
1034  * ...-- <2^31 ->|                                           |<--------...
1035  * ...---- >2^31 ------>|                                    |<--------...
1036  *
1037  * Current code wouldn't be vulnerable but it's better still to discard such
1038  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1039  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1040  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1041  * equal to the ideal case (infinite seqno space without wrap caused issues).
1042  *
1043  * With D-SACK the lower bound is extended to cover sequence space below
1044  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1045  * again, D-SACK block must not to go across snd_una (for the same reason as
1046  * for the normal SACK blocks, explained above). But there all simplicity
1047  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1048  * fully below undo_marker they do not affect behavior in anyway and can
1049  * therefore be safely ignored. In rare cases (which are more or less
1050  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1051  * fragmentation and packet reordering past skb's retransmission. To consider
1052  * them correctly, the acceptable range must be extended even more though
1053  * the exact amount is rather hard to quantify. However, tp->max_window can
1054  * be used as an exaggerated estimate.
1055  */
1056 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1057                                    u32 start_seq, u32 end_seq)
1058 {
1059         /* Too far in future, or reversed (interpretation is ambiguous) */
1060         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1061                 return false;
1062 
1063         /* Nasty start_seq wrap-around check (see comments above) */
1064         if (!before(start_seq, tp->snd_nxt))
1065                 return false;
1066 
1067         /* In outstanding window? ...This is valid exit for D-SACKs too.
1068          * start_seq == snd_una is non-sensical (see comments above)
1069          */
1070         if (after(start_seq, tp->snd_una))
1071                 return true;
1072 
1073         if (!is_dsack || !tp->undo_marker)
1074                 return false;
1075 
1076         /* ...Then it's D-SACK, and must reside below snd_una completely */
1077         if (after(end_seq, tp->snd_una))
1078                 return false;
1079 
1080         if (!before(start_seq, tp->undo_marker))
1081                 return true;
1082 
1083         /* Too old */
1084         if (!after(end_seq, tp->undo_marker))
1085                 return false;
1086 
1087         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1088          *   start_seq < undo_marker and end_seq >= undo_marker.
1089          */
1090         return !before(start_seq, end_seq - tp->max_window);
1091 }
1092 
1093 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1094                             struct tcp_sack_block_wire *sp, int num_sacks,
1095                             u32 prior_snd_una)
1096 {
1097         struct tcp_sock *tp = tcp_sk(sk);
1098         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1099         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1100         bool dup_sack = false;
1101 
1102         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1103                 dup_sack = true;
1104                 tcp_dsack_seen(tp);
1105                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1106         } else if (num_sacks > 1) {
1107                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1108                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1109 
1110                 if (!after(end_seq_0, end_seq_1) &&
1111                     !before(start_seq_0, start_seq_1)) {
1112                         dup_sack = true;
1113                         tcp_dsack_seen(tp);
1114                         NET_INC_STATS(sock_net(sk),
1115                                         LINUX_MIB_TCPDSACKOFORECV);
1116                 }
1117         }
1118 
1119         /* D-SACK for already forgotten data... Do dumb counting. */
1120         if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1121             !after(end_seq_0, prior_snd_una) &&
1122             after(end_seq_0, tp->undo_marker))
1123                 tp->undo_retrans--;
1124 
1125         return dup_sack;
1126 }
1127 
1128 struct tcp_sacktag_state {
1129         int     reord;
1130         int     fack_count;
1131         /* Timestamps for earliest and latest never-retransmitted segment
1132          * that was SACKed. RTO needs the earliest RTT to stay conservative,
1133          * but congestion control should still get an accurate delay signal.
1134          */
1135         struct skb_mstamp first_sackt;
1136         struct skb_mstamp last_sackt;
1137         struct rate_sample *rate;
1138         int     flag;
1139 };
1140 
1141 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1142  * the incoming SACK may not exactly match but we can find smaller MSS
1143  * aligned portion of it that matches. Therefore we might need to fragment
1144  * which may fail and creates some hassle (caller must handle error case
1145  * returns).
1146  *
1147  * FIXME: this could be merged to shift decision code
1148  */
1149 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1150                                   u32 start_seq, u32 end_seq)
1151 {
1152         int err;
1153         bool in_sack;
1154         unsigned int pkt_len;
1155         unsigned int mss;
1156 
1157         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1158                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1159 
1160         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1161             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1162                 mss = tcp_skb_mss(skb);
1163                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1164 
1165                 if (!in_sack) {
1166                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1167                         if (pkt_len < mss)
1168                                 pkt_len = mss;
1169                 } else {
1170                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1171                         if (pkt_len < mss)
1172                                 return -EINVAL;
1173                 }
1174 
1175                 /* Round if necessary so that SACKs cover only full MSSes
1176                  * and/or the remaining small portion (if present)
1177                  */
1178                 if (pkt_len > mss) {
1179                         unsigned int new_len = (pkt_len / mss) * mss;
1180                         if (!in_sack && new_len < pkt_len) {
1181                                 new_len += mss;
1182                                 if (new_len >= skb->len)
1183                                         return 0;
1184                         }
1185                         pkt_len = new_len;
1186                 }
1187                 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1188                 if (err < 0)
1189                         return err;
1190         }
1191 
1192         return in_sack;
1193 }
1194 
1195 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1196 static u8 tcp_sacktag_one(struct sock *sk,
1197                           struct tcp_sacktag_state *state, u8 sacked,
1198                           u32 start_seq, u32 end_seq,
1199                           int dup_sack, int pcount,
1200                           const struct skb_mstamp *xmit_time)
1201 {
1202         struct tcp_sock *tp = tcp_sk(sk);
1203         int fack_count = state->fack_count;
1204 
1205         /* Account D-SACK for retransmitted packet. */
1206         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1207                 if (tp->undo_marker && tp->undo_retrans > 0 &&
1208                     after(end_seq, tp->undo_marker))
1209                         tp->undo_retrans--;
1210                 if (sacked & TCPCB_SACKED_ACKED)
1211                         state->reord = min(fack_count, state->reord);
1212         }
1213 
1214         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1215         if (!after(end_seq, tp->snd_una))
1216                 return sacked;
1217 
1218         if (!(sacked & TCPCB_SACKED_ACKED)) {
1219                 tcp_rack_advance(tp, xmit_time, sacked);
1220 
1221                 if (sacked & TCPCB_SACKED_RETRANS) {
1222                         /* If the segment is not tagged as lost,
1223                          * we do not clear RETRANS, believing
1224                          * that retransmission is still in flight.
1225                          */
1226                         if (sacked & TCPCB_LOST) {
1227                                 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1228                                 tp->lost_out -= pcount;
1229                                 tp->retrans_out -= pcount;
1230                         }
1231                 } else {
1232                         if (!(sacked & TCPCB_RETRANS)) {
1233                                 /* New sack for not retransmitted frame,
1234                                  * which was in hole. It is reordering.
1235                                  */
1236                                 if (before(start_seq,
1237                                            tcp_highest_sack_seq(tp)))
1238                                         state->reord = min(fack_count,
1239                                                            state->reord);
1240                                 if (!after(end_seq, tp->high_seq))
1241                                         state->flag |= FLAG_ORIG_SACK_ACKED;
1242                                 if (state->first_sackt.v64 == 0)
1243                                         state->first_sackt = *xmit_time;
1244                                 state->last_sackt = *xmit_time;
1245                         }
1246 
1247                         if (sacked & TCPCB_LOST) {
1248                                 sacked &= ~TCPCB_LOST;
1249                                 tp->lost_out -= pcount;
1250                         }
1251                 }
1252 
1253                 sacked |= TCPCB_SACKED_ACKED;
1254                 state->flag |= FLAG_DATA_SACKED;
1255                 tp->sacked_out += pcount;
1256                 tp->delivered += pcount;  /* Out-of-order packets delivered */
1257 
1258                 fack_count += pcount;
1259 
1260                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1261                 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1262                     before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1263                         tp->lost_cnt_hint += pcount;
1264 
1265                 if (fack_count > tp->fackets_out)
1266                         tp->fackets_out = fack_count;
1267         }
1268 
1269         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1270          * frames and clear it. undo_retrans is decreased above, L|R frames
1271          * are accounted above as well.
1272          */
1273         if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1274                 sacked &= ~TCPCB_SACKED_RETRANS;
1275                 tp->retrans_out -= pcount;
1276         }
1277 
1278         return sacked;
1279 }
1280 
1281 /* Shift newly-SACKed bytes from this skb to the immediately previous
1282  * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1283  */
1284 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1285                             struct tcp_sacktag_state *state,
1286                             unsigned int pcount, int shifted, int mss,
1287                             bool dup_sack)
1288 {
1289         struct tcp_sock *tp = tcp_sk(sk);
1290         struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1291         u32 start_seq = TCP_SKB_CB(skb)->seq;   /* start of newly-SACKed */
1292         u32 end_seq = start_seq + shifted;      /* end of newly-SACKed */
1293 
1294         BUG_ON(!pcount);
1295 
1296         /* Adjust counters and hints for the newly sacked sequence
1297          * range but discard the return value since prev is already
1298          * marked. We must tag the range first because the seq
1299          * advancement below implicitly advances
1300          * tcp_highest_sack_seq() when skb is highest_sack.
1301          */
1302         tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1303                         start_seq, end_seq, dup_sack, pcount,
1304                         &skb->skb_mstamp);
1305         tcp_rate_skb_delivered(sk, skb, state->rate);
1306 
1307         if (skb == tp->lost_skb_hint)
1308                 tp->lost_cnt_hint += pcount;
1309 
1310         TCP_SKB_CB(prev)->end_seq += shifted;
1311         TCP_SKB_CB(skb)->seq += shifted;
1312 
1313         tcp_skb_pcount_add(prev, pcount);
1314         BUG_ON(tcp_skb_pcount(skb) < pcount);
1315         tcp_skb_pcount_add(skb, -pcount);
1316 
1317         /* When we're adding to gso_segs == 1, gso_size will be zero,
1318          * in theory this shouldn't be necessary but as long as DSACK
1319          * code can come after this skb later on it's better to keep
1320          * setting gso_size to something.
1321          */
1322         if (!TCP_SKB_CB(prev)->tcp_gso_size)
1323                 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1324 
1325         /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1326         if (tcp_skb_pcount(skb) <= 1)
1327                 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1328 
1329         /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1330         TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1331 
1332         if (skb->len > 0) {
1333                 BUG_ON(!tcp_skb_pcount(skb));
1334                 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1335                 return false;
1336         }
1337 
1338         /* Whole SKB was eaten :-) */
1339 
1340         if (skb == tp->retransmit_skb_hint)
1341                 tp->retransmit_skb_hint = prev;
1342         if (skb == tp->lost_skb_hint) {
1343                 tp->lost_skb_hint = prev;
1344                 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1345         }
1346 
1347         TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1348         TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1349         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1350                 TCP_SKB_CB(prev)->end_seq++;
1351 
1352         if (skb == tcp_highest_sack(sk))
1353                 tcp_advance_highest_sack(sk, skb);
1354 
1355         tcp_skb_collapse_tstamp(prev, skb);
1356         if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp.v64))
1357                 TCP_SKB_CB(prev)->tx.delivered_mstamp.v64 = 0;
1358 
1359         tcp_unlink_write_queue(skb, sk);
1360         sk_wmem_free_skb(sk, skb);
1361 
1362         NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1363 
1364         return true;
1365 }
1366 
1367 /* I wish gso_size would have a bit more sane initialization than
1368  * something-or-zero which complicates things
1369  */
1370 static int tcp_skb_seglen(const struct sk_buff *skb)
1371 {
1372         return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1373 }
1374 
1375 /* Shifting pages past head area doesn't work */
1376 static int skb_can_shift(const struct sk_buff *skb)
1377 {
1378         return !skb_headlen(skb) && skb_is_nonlinear(skb);
1379 }
1380 
1381 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1382  * skb.
1383  */
1384 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1385                                           struct tcp_sacktag_state *state,
1386                                           u32 start_seq, u32 end_seq,
1387                                           bool dup_sack)
1388 {
1389         struct tcp_sock *tp = tcp_sk(sk);
1390         struct sk_buff *prev;
1391         int mss;
1392         int pcount = 0;
1393         int len;
1394         int in_sack;
1395 
1396         if (!sk_can_gso(sk))
1397                 goto fallback;
1398 
1399         /* Normally R but no L won't result in plain S */
1400         if (!dup_sack &&
1401             (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1402                 goto fallback;
1403         if (!skb_can_shift(skb))
1404                 goto fallback;
1405         /* This frame is about to be dropped (was ACKed). */
1406         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1407                 goto fallback;
1408 
1409         /* Can only happen with delayed DSACK + discard craziness */
1410         if (unlikely(skb == tcp_write_queue_head(sk)))
1411                 goto fallback;
1412         prev = tcp_write_queue_prev(sk, skb);
1413 
1414         if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1415                 goto fallback;
1416 
1417         if (!tcp_skb_can_collapse_to(prev))
1418                 goto fallback;
1419 
1420         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1421                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1422 
1423         if (in_sack) {
1424                 len = skb->len;
1425                 pcount = tcp_skb_pcount(skb);
1426                 mss = tcp_skb_seglen(skb);
1427 
1428                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1429                  * drop this restriction as unnecessary
1430                  */
1431                 if (mss != tcp_skb_seglen(prev))
1432                         goto fallback;
1433         } else {
1434                 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1435                         goto noop;
1436                 /* CHECKME: This is non-MSS split case only?, this will
1437                  * cause skipped skbs due to advancing loop btw, original
1438                  * has that feature too
1439                  */
1440                 if (tcp_skb_pcount(skb) <= 1)
1441                         goto noop;
1442 
1443                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1444                 if (!in_sack) {
1445                         /* TODO: head merge to next could be attempted here
1446                          * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1447                          * though it might not be worth of the additional hassle
1448                          *
1449                          * ...we can probably just fallback to what was done
1450                          * previously. We could try merging non-SACKed ones
1451                          * as well but it probably isn't going to buy off
1452                          * because later SACKs might again split them, and
1453                          * it would make skb timestamp tracking considerably
1454                          * harder problem.
1455                          */
1456                         goto fallback;
1457                 }
1458 
1459                 len = end_seq - TCP_SKB_CB(skb)->seq;
1460                 BUG_ON(len < 0);
1461                 BUG_ON(len > skb->len);
1462 
1463                 /* MSS boundaries should be honoured or else pcount will
1464                  * severely break even though it makes things bit trickier.
1465                  * Optimize common case to avoid most of the divides
1466                  */
1467                 mss = tcp_skb_mss(skb);
1468 
1469                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1470                  * drop this restriction as unnecessary
1471                  */
1472                 if (mss != tcp_skb_seglen(prev))
1473                         goto fallback;
1474 
1475                 if (len == mss) {
1476                         pcount = 1;
1477                 } else if (len < mss) {
1478                         goto noop;
1479                 } else {
1480                         pcount = len / mss;
1481                         len = pcount * mss;
1482                 }
1483         }
1484 
1485         /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1486         if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1487                 goto fallback;
1488 
1489         if (!skb_shift(prev, skb, len))
1490                 goto fallback;
1491         if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1492                 goto out;
1493 
1494         /* Hole filled allows collapsing with the next as well, this is very
1495          * useful when hole on every nth skb pattern happens
1496          */
1497         if (prev == tcp_write_queue_tail(sk))
1498                 goto out;
1499         skb = tcp_write_queue_next(sk, prev);
1500 
1501         if (!skb_can_shift(skb) ||
1502             (skb == tcp_send_head(sk)) ||
1503             ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1504             (mss != tcp_skb_seglen(skb)))
1505                 goto out;
1506 
1507         len = skb->len;
1508         if (skb_shift(prev, skb, len)) {
1509                 pcount += tcp_skb_pcount(skb);
1510                 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1511         }
1512 
1513 out:
1514         state->fack_count += pcount;
1515         return prev;
1516 
1517 noop:
1518         return skb;
1519 
1520 fallback:
1521         NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1522         return NULL;
1523 }
1524 
1525 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1526                                         struct tcp_sack_block *next_dup,
1527                                         struct tcp_sacktag_state *state,
1528                                         u32 start_seq, u32 end_seq,
1529                                         bool dup_sack_in)
1530 {
1531         struct tcp_sock *tp = tcp_sk(sk);
1532         struct sk_buff *tmp;
1533 
1534         tcp_for_write_queue_from(skb, sk) {
1535                 int in_sack = 0;
1536                 bool dup_sack = dup_sack_in;
1537 
1538                 if (skb == tcp_send_head(sk))
1539                         break;
1540 
1541                 /* queue is in-order => we can short-circuit the walk early */
1542                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1543                         break;
1544 
1545                 if (next_dup  &&
1546                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1547                         in_sack = tcp_match_skb_to_sack(sk, skb,
1548                                                         next_dup->start_seq,
1549                                                         next_dup->end_seq);
1550                         if (in_sack > 0)
1551                                 dup_sack = true;
1552                 }
1553 
1554                 /* skb reference here is a bit tricky to get right, since
1555                  * shifting can eat and free both this skb and the next,
1556                  * so not even _safe variant of the loop is enough.
1557                  */
1558                 if (in_sack <= 0) {
1559                         tmp = tcp_shift_skb_data(sk, skb, state,
1560                                                  start_seq, end_seq, dup_sack);
1561                         if (tmp) {
1562                                 if (tmp != skb) {
1563                                         skb = tmp;
1564                                         continue;
1565                                 }
1566 
1567                                 in_sack = 0;
1568                         } else {
1569                                 in_sack = tcp_match_skb_to_sack(sk, skb,
1570                                                                 start_seq,
1571                                                                 end_seq);
1572                         }
1573                 }
1574 
1575                 if (unlikely(in_sack < 0))
1576                         break;
1577 
1578                 if (in_sack) {
1579                         TCP_SKB_CB(skb)->sacked =
1580                                 tcp_sacktag_one(sk,
1581                                                 state,
1582                                                 TCP_SKB_CB(skb)->sacked,
1583                                                 TCP_SKB_CB(skb)->seq,
1584                                                 TCP_SKB_CB(skb)->end_seq,
1585                                                 dup_sack,
1586                                                 tcp_skb_pcount(skb),
1587                                                 &skb->skb_mstamp);
1588                         tcp_rate_skb_delivered(sk, skb, state->rate);
1589 
1590                         if (!before(TCP_SKB_CB(skb)->seq,
1591                                     tcp_highest_sack_seq(tp)))
1592                                 tcp_advance_highest_sack(sk, skb);
1593                 }
1594 
1595                 state->fack_count += tcp_skb_pcount(skb);
1596         }
1597         return skb;
1598 }
1599 
1600 /* Avoid all extra work that is being done by sacktag while walking in
1601  * a normal way
1602  */
1603 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1604                                         struct tcp_sacktag_state *state,
1605                                         u32 skip_to_seq)
1606 {
1607         tcp_for_write_queue_from(skb, sk) {
1608                 if (skb == tcp_send_head(sk))
1609                         break;
1610 
1611                 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1612                         break;
1613 
1614                 state->fack_count += tcp_skb_pcount(skb);
1615         }
1616         return skb;
1617 }
1618 
1619 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1620                                                 struct sock *sk,
1621                                                 struct tcp_sack_block *next_dup,
1622                                                 struct tcp_sacktag_state *state,
1623                                                 u32 skip_to_seq)
1624 {
1625         if (!next_dup)
1626                 return skb;
1627 
1628         if (before(next_dup->start_seq, skip_to_seq)) {
1629                 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1630                 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1631                                        next_dup->start_seq, next_dup->end_seq,
1632                                        1);
1633         }
1634 
1635         return skb;
1636 }
1637 
1638 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1639 {
1640         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1641 }
1642 
1643 static int
1644 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1645                         u32 prior_snd_una, struct tcp_sacktag_state *state)
1646 {
1647         struct tcp_sock *tp = tcp_sk(sk);
1648         const unsigned char *ptr = (skb_transport_header(ack_skb) +
1649                                     TCP_SKB_CB(ack_skb)->sacked);
1650         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1651         struct tcp_sack_block sp[TCP_NUM_SACKS];
1652         struct tcp_sack_block *cache;
1653         struct sk_buff *skb;
1654         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1655         int used_sacks;
1656         bool found_dup_sack = false;
1657         int i, j;
1658         int first_sack_index;
1659 
1660         state->flag = 0;
1661         state->reord = tp->packets_out;
1662 
1663         if (!tp->sacked_out) {
1664                 if (WARN_ON(tp->fackets_out))
1665                         tp->fackets_out = 0;
1666                 tcp_highest_sack_reset(sk);
1667         }
1668 
1669         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1670                                          num_sacks, prior_snd_una);
1671         if (found_dup_sack) {
1672                 state->flag |= FLAG_DSACKING_ACK;
1673                 tp->delivered++; /* A spurious retransmission is delivered */
1674         }
1675 
1676         /* Eliminate too old ACKs, but take into
1677          * account more or less fresh ones, they can
1678          * contain valid SACK info.
1679          */
1680         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1681                 return 0;
1682 
1683         if (!tp->packets_out)
1684                 goto out;
1685 
1686         used_sacks = 0;
1687         first_sack_index = 0;
1688         for (i = 0; i < num_sacks; i++) {
1689                 bool dup_sack = !i && found_dup_sack;
1690 
1691                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1692                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1693 
1694                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1695                                             sp[used_sacks].start_seq,
1696                                             sp[used_sacks].end_seq)) {
1697                         int mib_idx;
1698 
1699                         if (dup_sack) {
1700                                 if (!tp->undo_marker)
1701                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1702                                 else
1703                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1704                         } else {
1705                                 /* Don't count olds caused by ACK reordering */
1706                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1707                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1708                                         continue;
1709                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1710                         }
1711 
1712                         NET_INC_STATS(sock_net(sk), mib_idx);
1713                         if (i == 0)
1714                                 first_sack_index = -1;
1715                         continue;
1716                 }
1717 
1718                 /* Ignore very old stuff early */
1719                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1720                         continue;
1721 
1722                 used_sacks++;
1723         }
1724 
1725         /* order SACK blocks to allow in order walk of the retrans queue */
1726         for (i = used_sacks - 1; i > 0; i--) {
1727                 for (j = 0; j < i; j++) {
1728                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1729                                 swap(sp[j], sp[j + 1]);
1730 
1731                                 /* Track where the first SACK block goes to */
1732                                 if (j == first_sack_index)
1733                                         first_sack_index = j + 1;
1734                         }
1735                 }
1736         }
1737 
1738         skb = tcp_write_queue_head(sk);
1739         state->fack_count = 0;
1740         i = 0;
1741 
1742         if (!tp->sacked_out) {
1743                 /* It's already past, so skip checking against it */
1744                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1745         } else {
1746                 cache = tp->recv_sack_cache;
1747                 /* Skip empty blocks in at head of the cache */
1748                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1749                        !cache->end_seq)
1750                         cache++;
1751         }
1752 
1753         while (i < used_sacks) {
1754                 u32 start_seq = sp[i].start_seq;
1755                 u32 end_seq = sp[i].end_seq;
1756                 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1757                 struct tcp_sack_block *next_dup = NULL;
1758 
1759                 if (found_dup_sack && ((i + 1) == first_sack_index))
1760                         next_dup = &sp[i + 1];
1761 
1762                 /* Skip too early cached blocks */
1763                 while (tcp_sack_cache_ok(tp, cache) &&
1764                        !before(start_seq, cache->end_seq))
1765                         cache++;
1766 
1767                 /* Can skip some work by looking recv_sack_cache? */
1768                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1769                     after(end_seq, cache->start_seq)) {
1770 
1771                         /* Head todo? */
1772                         if (before(start_seq, cache->start_seq)) {
1773                                 skb = tcp_sacktag_skip(skb, sk, state,
1774                                                        start_seq);
1775                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1776                                                        state,
1777                                                        start_seq,
1778                                                        cache->start_seq,
1779                                                        dup_sack);
1780                         }
1781 
1782                         /* Rest of the block already fully processed? */
1783                         if (!after(end_seq, cache->end_seq))
1784                                 goto advance_sp;
1785 
1786                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1787                                                        state,
1788                                                        cache->end_seq);
1789 
1790                         /* ...tail remains todo... */
1791                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1792                                 /* ...but better entrypoint exists! */
1793                                 skb = tcp_highest_sack(sk);
1794                                 if (!skb)
1795                                         break;
1796                                 state->fack_count = tp->fackets_out;
1797                                 cache++;
1798                                 goto walk;
1799                         }
1800 
1801                         skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1802                         /* Check overlap against next cached too (past this one already) */
1803                         cache++;
1804                         continue;
1805                 }
1806 
1807                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1808                         skb = tcp_highest_sack(sk);
1809                         if (!skb)
1810                                 break;
1811                         state->fack_count = tp->fackets_out;
1812                 }
1813                 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1814 
1815 walk:
1816                 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1817                                        start_seq, end_seq, dup_sack);
1818 
1819 advance_sp:
1820                 i++;
1821         }
1822 
1823         /* Clear the head of the cache sack blocks so we can skip it next time */
1824         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1825                 tp->recv_sack_cache[i].start_seq = 0;
1826                 tp->recv_sack_cache[i].end_seq = 0;
1827         }
1828         for (j = 0; j < used_sacks; j++)
1829                 tp->recv_sack_cache[i++] = sp[j];
1830 
1831         if ((state->reord < tp->fackets_out) &&
1832             ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1833                 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
1834 
1835         tcp_verify_left_out(tp);
1836 out:
1837 
1838 #if FASTRETRANS_DEBUG > 0
1839         WARN_ON((int)tp->sacked_out < 0);
1840         WARN_ON((int)tp->lost_out < 0);
1841         WARN_ON((int)tp->retrans_out < 0);
1842         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1843 #endif
1844         return state->flag;
1845 }
1846 
1847 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1848  * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1849  */
1850 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1851 {
1852         u32 holes;
1853 
1854         holes = max(tp->lost_out, 1U);
1855         holes = min(holes, tp->packets_out);
1856 
1857         if ((tp->sacked_out + holes) > tp->packets_out) {
1858                 tp->sacked_out = tp->packets_out - holes;
1859                 return true;
1860         }
1861         return false;
1862 }
1863 
1864 /* If we receive more dupacks than we expected counting segments
1865  * in assumption of absent reordering, interpret this as reordering.
1866  * The only another reason could be bug in receiver TCP.
1867  */
1868 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1869 {
1870         struct tcp_sock *tp = tcp_sk(sk);
1871         if (tcp_limit_reno_sacked(tp))
1872                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1873 }
1874 
1875 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1876 
1877 static void tcp_add_reno_sack(struct sock *sk)
1878 {
1879         struct tcp_sock *tp = tcp_sk(sk);
1880         u32 prior_sacked = tp->sacked_out;
1881 
1882         tp->sacked_out++;
1883         tcp_check_reno_reordering(sk, 0);
1884         if (tp->sacked_out > prior_sacked)
1885                 tp->delivered++; /* Some out-of-order packet is delivered */
1886         tcp_verify_left_out(tp);
1887 }
1888 
1889 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1890 
1891 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1892 {
1893         struct tcp_sock *tp = tcp_sk(sk);
1894 
1895         if (acked > 0) {
1896                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1897                 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1898                 if (acked - 1 >= tp->sacked_out)
1899                         tp->sacked_out = 0;
1900                 else
1901                         tp->sacked_out -= acked - 1;
1902         }
1903         tcp_check_reno_reordering(sk, acked);
1904         tcp_verify_left_out(tp);
1905 }
1906 
1907 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1908 {
1909         tp->sacked_out = 0;
1910 }
1911 
1912 void tcp_clear_retrans(struct tcp_sock *tp)
1913 {
1914         tp->retrans_out = 0;
1915         tp->lost_out = 0;
1916         tp->undo_marker = 0;
1917         tp->undo_retrans = -1;
1918         tp->fackets_out = 0;
1919         tp->sacked_out = 0;
1920 }
1921 
1922 static inline void tcp_init_undo(struct tcp_sock *tp)
1923 {
1924         tp->undo_marker = tp->snd_una;
1925         /* Retransmission still in flight may cause DSACKs later. */
1926         tp->undo_retrans = tp->retrans_out ? : -1;
1927 }
1928 
1929 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1930  * and reset tags completely, otherwise preserve SACKs. If receiver
1931  * dropped its ofo queue, we will know this due to reneging detection.
1932  */
1933 void tcp_enter_loss(struct sock *sk)
1934 {
1935         const struct inet_connection_sock *icsk = inet_csk(sk);
1936         struct tcp_sock *tp = tcp_sk(sk);
1937         struct net *net = sock_net(sk);
1938         struct sk_buff *skb;
1939         bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1940         bool is_reneg;                  /* is receiver reneging on SACKs? */
1941         bool mark_lost;
1942 
1943         /* Reduce ssthresh if it has not yet been made inside this window. */
1944         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1945             !after(tp->high_seq, tp->snd_una) ||
1946             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1947                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1948                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1949                 tcp_ca_event(sk, CA_EVENT_LOSS);
1950                 tcp_init_undo(tp);
1951         }
1952         tp->snd_cwnd       = 1;
1953         tp->snd_cwnd_cnt   = 0;
1954         tp->snd_cwnd_stamp = tcp_time_stamp;
1955 
1956         tp->retrans_out = 0;
1957         tp->lost_out = 0;
1958 
1959         if (tcp_is_reno(tp))
1960                 tcp_reset_reno_sack(tp);
1961 
1962         skb = tcp_write_queue_head(sk);
1963         is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1964         if (is_reneg) {
1965                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1966                 tp->sacked_out = 0;
1967                 tp->fackets_out = 0;
1968         }
1969         tcp_clear_all_retrans_hints(tp);
1970 
1971         tcp_for_write_queue(skb, sk) {
1972                 if (skb == tcp_send_head(sk))
1973                         break;
1974 
1975                 mark_lost = (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
1976                              is_reneg);
1977                 if (mark_lost)
1978                         tcp_sum_lost(tp, skb);
1979                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1980                 if (mark_lost) {
1981                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1982                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1983                         tp->lost_out += tcp_skb_pcount(skb);
1984                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1985                 }
1986         }
1987         tcp_verify_left_out(tp);
1988 
1989         /* Timeout in disordered state after receiving substantial DUPACKs
1990          * suggests that the degree of reordering is over-estimated.
1991          */
1992         if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1993             tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
1994                 tp->reordering = min_t(unsigned int, tp->reordering,
1995                                        net->ipv4.sysctl_tcp_reordering);
1996         tcp_set_ca_state(sk, TCP_CA_Loss);
1997         tp->high_seq = tp->snd_nxt;
1998         tcp_ecn_queue_cwr(tp);
1999 
2000         /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2001          * loss recovery is underway except recurring timeout(s) on
2002          * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2003          */
2004         tp->frto = sysctl_tcp_frto &&
2005                    (new_recovery || icsk->icsk_retransmits) &&
2006                    !inet_csk(sk)->icsk_mtup.probe_size;
2007 }
2008 
2009 /* If ACK arrived pointing to a remembered SACK, it means that our
2010  * remembered SACKs do not reflect real state of receiver i.e.
2011  * receiver _host_ is heavily congested (or buggy).
2012  *
2013  * To avoid big spurious retransmission bursts due to transient SACK
2014  * scoreboard oddities that look like reneging, we give the receiver a
2015  * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2016  * restore sanity to the SACK scoreboard. If the apparent reneging
2017  * persists until this RTO then we'll clear the SACK scoreboard.
2018  */
2019 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2020 {
2021         if (flag & FLAG_SACK_RENEGING) {
2022                 struct tcp_sock *tp = tcp_sk(sk);
2023                 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2024                                           msecs_to_jiffies(10));
2025 
2026                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2027                                           delay, TCP_RTO_MAX);
2028                 return true;
2029         }
2030         return false;
2031 }
2032 
2033 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2034 {
2035         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2036 }
2037 
2038 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2039  * counter when SACK is enabled (without SACK, sacked_out is used for
2040  * that purpose).
2041  *
2042  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2043  * segments up to the highest received SACK block so far and holes in
2044  * between them.
2045  *
2046  * With reordering, holes may still be in flight, so RFC3517 recovery
2047  * uses pure sacked_out (total number of SACKed segments) even though
2048  * it violates the RFC that uses duplicate ACKs, often these are equal
2049  * but when e.g. out-of-window ACKs or packet duplication occurs,
2050  * they differ. Since neither occurs due to loss, TCP should really
2051  * ignore them.
2052  */
2053 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2054 {
2055         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2056 }
2057 
2058 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2059 {
2060         struct tcp_sock *tp = tcp_sk(sk);
2061         unsigned long delay;
2062 
2063         /* Delay early retransmit and entering fast recovery for
2064          * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2065          * available, or RTO is scheduled to fire first.
2066          */
2067         if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2068             (flag & FLAG_ECE) || !tp->srtt_us)
2069                 return false;
2070 
2071         delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2072                     msecs_to_jiffies(2));
2073 
2074         if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2075                 return false;
2076 
2077         inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2078                                   TCP_RTO_MAX);
2079         return true;
2080 }
2081 
2082 /* Linux NewReno/SACK/FACK/ECN state machine.
2083  * --------------------------------------
2084  *
2085  * "Open"       Normal state, no dubious events, fast path.
2086  * "Disorder"   In all the respects it is "Open",
2087  *              but requires a bit more attention. It is entered when
2088  *              we see some SACKs or dupacks. It is split of "Open"
2089  *              mainly to move some processing from fast path to slow one.
2090  * "CWR"        CWND was reduced due to some Congestion Notification event.
2091  *              It can be ECN, ICMP source quench, local device congestion.
2092  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2093  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2094  *
2095  * tcp_fastretrans_alert() is entered:
2096  * - each incoming ACK, if state is not "Open"
2097  * - when arrived ACK is unusual, namely:
2098  *      * SACK
2099  *      * Duplicate ACK.
2100  *      * ECN ECE.
2101  *
2102  * Counting packets in flight is pretty simple.
2103  *
2104  *      in_flight = packets_out - left_out + retrans_out
2105  *
2106  *      packets_out is SND.NXT-SND.UNA counted in packets.
2107  *
2108  *      retrans_out is number of retransmitted segments.
2109  *
2110  *      left_out is number of segments left network, but not ACKed yet.
2111  *
2112  *              left_out = sacked_out + lost_out
2113  *
2114  *     sacked_out: Packets, which arrived to receiver out of order
2115  *                 and hence not ACKed. With SACKs this number is simply
2116  *                 amount of SACKed data. Even without SACKs
2117  *                 it is easy to give pretty reliable estimate of this number,
2118  *                 counting duplicate ACKs.
2119  *
2120  *       lost_out: Packets lost by network. TCP has no explicit
2121  *                 "loss notification" feedback from network (for now).
2122  *                 It means that this number can be only _guessed_.
2123  *                 Actually, it is the heuristics to predict lossage that
2124  *                 distinguishes different algorithms.
2125  *
2126  *      F.e. after RTO, when all the queue is considered as lost,
2127  *      lost_out = packets_out and in_flight = retrans_out.
2128  *
2129  *              Essentially, we have now two algorithms counting
2130  *              lost packets.
2131  *
2132  *              FACK: It is the simplest heuristics. As soon as we decided
2133  *              that something is lost, we decide that _all_ not SACKed
2134  *              packets until the most forward SACK are lost. I.e.
2135  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2136  *              It is absolutely correct estimate, if network does not reorder
2137  *              packets. And it loses any connection to reality when reordering
2138  *              takes place. We use FACK by default until reordering
2139  *              is suspected on the path to this destination.
2140  *
2141  *              NewReno: when Recovery is entered, we assume that one segment
2142  *              is lost (classic Reno). While we are in Recovery and
2143  *              a partial ACK arrives, we assume that one more packet
2144  *              is lost (NewReno). This heuristics are the same in NewReno
2145  *              and SACK.
2146  *
2147  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2148  *  deflation etc. CWND is real congestion window, never inflated, changes
2149  *  only according to classic VJ rules.
2150  *
2151  * Really tricky (and requiring careful tuning) part of algorithm
2152  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2153  * The first determines the moment _when_ we should reduce CWND and,
2154  * hence, slow down forward transmission. In fact, it determines the moment
2155  * when we decide that hole is caused by loss, rather than by a reorder.
2156  *
2157  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2158  * holes, caused by lost packets.
2159  *
2160  * And the most logically complicated part of algorithm is undo
2161  * heuristics. We detect false retransmits due to both too early
2162  * fast retransmit (reordering) and underestimated RTO, analyzing
2163  * timestamps and D-SACKs. When we detect that some segments were
2164  * retransmitted by mistake and CWND reduction was wrong, we undo
2165  * window reduction and abort recovery phase. This logic is hidden
2166  * inside several functions named tcp_try_undo_<something>.
2167  */
2168 
2169 /* This function decides, when we should leave Disordered state
2170  * and enter Recovery phase, reducing congestion window.
2171  *
2172  * Main question: may we further continue forward transmission
2173  * with the same cwnd?
2174  */
2175 static bool tcp_time_to_recover(struct sock *sk, int flag)
2176 {
2177         struct tcp_sock *tp = tcp_sk(sk);
2178         __u32 packets_out;
2179         int tcp_reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
2180 
2181         /* Trick#1: The loss is proven. */
2182         if (tp->lost_out)
2183                 return true;
2184 
2185         /* Not-A-Trick#2 : Classic rule... */
2186         if (tcp_dupack_heuristics(tp) > tp->reordering)
2187                 return true;
2188 
2189         /* Trick#4: It is still not OK... But will it be useful to delay
2190          * recovery more?
2191          */
2192         packets_out = tp->packets_out;
2193         if (packets_out <= tp->reordering &&
2194             tp->sacked_out >= max_t(__u32, packets_out/2, tcp_reordering) &&
2195             !tcp_may_send_now(sk)) {
2196                 /* We have nothing to send. This connection is limited
2197                  * either by receiver window or by application.
2198                  */
2199                 return true;
2200         }
2201 
2202         /* If a thin stream is detected, retransmit after first
2203          * received dupack. Employ only if SACK is supported in order
2204          * to avoid possible corner-case series of spurious retransmissions
2205          * Use only if there are no unsent data.
2206          */
2207         if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2208             tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2209             tcp_is_sack(tp) && !tcp_send_head(sk))
2210                 return true;
2211 
2212         /* Trick#6: TCP early retransmit, per RFC5827.  To avoid spurious
2213          * retransmissions due to small network reorderings, we implement
2214          * Mitigation A.3 in the RFC and delay the retransmission for a short
2215          * interval if appropriate.
2216          */
2217         if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2218             (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2219             !tcp_may_send_now(sk))
2220                 return !tcp_pause_early_retransmit(sk, flag);
2221 
2222         return false;
2223 }
2224 
2225 /* Detect loss in event "A" above by marking head of queue up as lost.
2226  * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2227  * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2228  * has at least tp->reordering SACKed seqments above it; "packets" refers to
2229  * the maximum SACKed segments to pass before reaching this limit.
2230  */
2231 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2232 {
2233         struct tcp_sock *tp = tcp_sk(sk);
2234         struct sk_buff *skb;
2235         int cnt, oldcnt, lost;
2236         unsigned int mss;
2237         /* Use SACK to deduce losses of new sequences sent during recovery */
2238         const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2239 
2240         WARN_ON(packets > tp->packets_out);
2241         if (tp->lost_skb_hint) {
2242                 skb = tp->lost_skb_hint;
2243                 cnt = tp->lost_cnt_hint;
2244                 /* Head already handled? */
2245                 if (mark_head && skb != tcp_write_queue_head(sk))
2246                         return;
2247         } else {
2248                 skb = tcp_write_queue_head(sk);
2249                 cnt = 0;
2250         }
2251 
2252         tcp_for_write_queue_from(skb, sk) {
2253                 if (skb == tcp_send_head(sk))
2254                         break;
2255                 /* TODO: do this better */
2256                 /* this is not the most efficient way to do this... */
2257                 tp->lost_skb_hint = skb;
2258                 tp->lost_cnt_hint = cnt;
2259 
2260                 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2261                         break;
2262 
2263                 oldcnt = cnt;
2264                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2265                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2266                         cnt += tcp_skb_pcount(skb);
2267 
2268                 if (cnt > packets) {
2269                         if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2270                             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2271                             (oldcnt >= packets))
2272                                 break;
2273 
2274                         mss = tcp_skb_mss(skb);
2275                         /* If needed, chop off the prefix to mark as lost. */
2276                         lost = (packets - oldcnt) * mss;
2277                         if (lost < skb->len &&
2278                             tcp_fragment(sk, skb, lost, mss, GFP_ATOMIC) < 0)
2279                                 break;
2280                         cnt = packets;
2281                 }
2282 
2283                 tcp_skb_mark_lost(tp, skb);
2284 
2285                 if (mark_head)
2286                         break;
2287         }
2288         tcp_verify_left_out(tp);
2289 }
2290 
2291 /* Account newly detected lost packet(s) */
2292 
2293 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2294 {
2295         struct tcp_sock *tp = tcp_sk(sk);
2296 
2297         if (tcp_is_reno(tp)) {
2298                 tcp_mark_head_lost(sk, 1, 1);
2299         } else if (tcp_is_fack(tp)) {
2300                 int lost = tp->fackets_out - tp->reordering;
2301                 if (lost <= 0)
2302                         lost = 1;
2303                 tcp_mark_head_lost(sk, lost, 0);
2304         } else {
2305                 int sacked_upto = tp->sacked_out - tp->reordering;
2306                 if (sacked_upto >= 0)
2307                         tcp_mark_head_lost(sk, sacked_upto, 0);
2308                 else if (fast_rexmit)
2309                         tcp_mark_head_lost(sk, 1, 1);
2310         }
2311 }
2312 
2313 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2314 {
2315         return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2316                before(tp->rx_opt.rcv_tsecr, when);
2317 }
2318 
2319 /* skb is spurious retransmitted if the returned timestamp echo
2320  * reply is prior to the skb transmission time
2321  */
2322 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2323                                      const struct sk_buff *skb)
2324 {
2325         return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2326                tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2327 }
2328 
2329 /* Nothing was retransmitted or returned timestamp is less
2330  * than timestamp of the first retransmission.
2331  */
2332 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2333 {
2334         return !tp->retrans_stamp ||
2335                tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2336 }
2337 
2338 /* Undo procedures. */
2339 
2340 /* We can clear retrans_stamp when there are no retransmissions in the
2341  * window. It would seem that it is trivially available for us in
2342  * tp->retrans_out, however, that kind of assumptions doesn't consider
2343  * what will happen if errors occur when sending retransmission for the
2344  * second time. ...It could the that such segment has only
2345  * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2346  * the head skb is enough except for some reneging corner cases that
2347  * are not worth the effort.
2348  *
2349  * Main reason for all this complexity is the fact that connection dying
2350  * time now depends on the validity of the retrans_stamp, in particular,
2351  * that successive retransmissions of a segment must not advance
2352  * retrans_stamp under any conditions.
2353  */
2354 static bool tcp_any_retrans_done(const struct sock *sk)
2355 {
2356         const struct tcp_sock *tp = tcp_sk(sk);
2357         struct sk_buff *skb;
2358 
2359         if (tp->retrans_out)
2360                 return true;
2361 
2362         skb = tcp_write_queue_head(sk);
2363         if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2364                 return true;
2365 
2366         return false;
2367 }
2368 
2369 #if FASTRETRANS_DEBUG > 1
2370 static void DBGUNDO(struct sock *sk, const char *msg)
2371 {
2372         struct tcp_sock *tp = tcp_sk(sk);
2373         struct inet_sock *inet = inet_sk(sk);
2374 
2375         if (sk->sk_family == AF_INET) {
2376                 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2377                          msg,
2378                          &inet->inet_daddr, ntohs(inet->inet_dport),
2379                          tp->snd_cwnd, tcp_left_out(tp),
2380                          tp->snd_ssthresh, tp->prior_ssthresh,
2381                          tp->packets_out);
2382         }
2383 #if IS_ENABLED(CONFIG_IPV6)
2384         else if (sk->sk_family == AF_INET6) {
2385                 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2386                          msg,
2387                          &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2388                          tp->snd_cwnd, tcp_left_out(tp),
2389                          tp->snd_ssthresh, tp->prior_ssthresh,
2390                          tp->packets_out);
2391         }
2392 #endif
2393 }
2394 #else
2395 #define DBGUNDO(x...) do { } while (0)
2396 #endif
2397 
2398 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2399 {
2400         struct tcp_sock *tp = tcp_sk(sk);
2401 
2402         if (unmark_loss) {
2403                 struct sk_buff *skb;
2404 
2405                 tcp_for_write_queue(skb, sk) {
2406                         if (skb == tcp_send_head(sk))
2407                                 break;
2408                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2409                 }
2410                 tp->lost_out = 0;
2411                 tcp_clear_all_retrans_hints(tp);
2412         }
2413 
2414         if (tp->prior_ssthresh) {
2415                 const struct inet_connection_sock *icsk = inet_csk(sk);
2416 
2417                 if (icsk->icsk_ca_ops->undo_cwnd)
2418                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2419                 else
2420                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2421 
2422                 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2423                         tp->snd_ssthresh = tp->prior_ssthresh;
2424                         tcp_ecn_withdraw_cwr(tp);
2425                 }
2426         }
2427         tp->snd_cwnd_stamp = tcp_time_stamp;
2428         tp->undo_marker = 0;
2429 }
2430 
2431 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2432 {
2433         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2434 }
2435 
2436 /* People celebrate: "We love our President!" */
2437 static bool tcp_try_undo_recovery(struct sock *sk)
2438 {
2439         struct tcp_sock *tp = tcp_sk(sk);
2440 
2441         if (tcp_may_undo(tp)) {
2442                 int mib_idx;
2443 
2444                 /* Happy end! We did not retransmit anything
2445                  * or our original transmission succeeded.
2446                  */
2447                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2448                 tcp_undo_cwnd_reduction(sk, false);
2449                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2450                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2451                 else
2452                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2453 
2454                 NET_INC_STATS(sock_net(sk), mib_idx);
2455         }
2456         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2457                 /* Hold old state until something *above* high_seq
2458                  * is ACKed. For Reno it is MUST to prevent false
2459                  * fast retransmits (RFC2582). SACK TCP is safe. */
2460                 if (!tcp_any_retrans_done(sk))
2461                         tp->retrans_stamp = 0;
2462                 return true;
2463         }
2464         tcp_set_ca_state(sk, TCP_CA_Open);
2465         return false;
2466 }
2467 
2468 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2469 static bool tcp_try_undo_dsack(struct sock *sk)
2470 {
2471         struct tcp_sock *tp = tcp_sk(sk);
2472 
2473         if (tp->undo_marker && !tp->undo_retrans) {
2474                 DBGUNDO(sk, "D-SACK");
2475                 tcp_undo_cwnd_reduction(sk, false);
2476                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2477                 return true;
2478         }
2479         return false;
2480 }
2481 
2482 /* Undo during loss recovery after partial ACK or using F-RTO. */
2483 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2484 {
2485         struct tcp_sock *tp = tcp_sk(sk);
2486 
2487         if (frto_undo || tcp_may_undo(tp)) {
2488                 tcp_undo_cwnd_reduction(sk, true);
2489 
2490                 DBGUNDO(sk, "partial loss");
2491                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2492                 if (frto_undo)
2493                         NET_INC_STATS(sock_net(sk),
2494                                         LINUX_MIB_TCPSPURIOUSRTOS);
2495                 inet_csk(sk)->icsk_retransmits = 0;
2496                 if (frto_undo || tcp_is_sack(tp))
2497                         tcp_set_ca_state(sk, TCP_CA_Open);
2498                 return true;
2499         }
2500         return false;
2501 }
2502 
2503 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2504  * It computes the number of packets to send (sndcnt) based on packets newly
2505  * delivered:
2506  *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2507  *      cwnd reductions across a full RTT.
2508  *   2) Otherwise PRR uses packet conservation to send as much as delivered.
2509  *      But when the retransmits are acked without further losses, PRR
2510  *      slow starts cwnd up to ssthresh to speed up the recovery.
2511  */
2512 static void tcp_init_cwnd_reduction(struct sock *sk)
2513 {
2514         struct tcp_sock *tp = tcp_sk(sk);
2515 
2516         tp->high_seq = tp->snd_nxt;
2517         tp->tlp_high_seq = 0;
2518         tp->snd_cwnd_cnt = 0;
2519         tp->prior_cwnd = tp->snd_cwnd;
2520         tp->prr_delivered = 0;
2521         tp->prr_out = 0;
2522         tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2523         tcp_ecn_queue_cwr(tp);
2524 }
2525 
2526 static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked,
2527                                int flag)
2528 {
2529         struct tcp_sock *tp = tcp_sk(sk);
2530         int sndcnt = 0;
2531         int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2532 
2533         if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2534                 return;
2535 
2536         tp->prr_delivered += newly_acked_sacked;
2537         if (delta < 0) {
2538                 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2539                                tp->prior_cwnd - 1;
2540                 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2541         } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2542                    !(flag & FLAG_LOST_RETRANS)) {
2543                 sndcnt = min_t(int, delta,
2544                                max_t(int, tp->prr_delivered - tp->prr_out,
2545                                      newly_acked_sacked) + 1);
2546         } else {
2547                 sndcnt = min(delta, newly_acked_sacked);
2548         }
2549         /* Force a fast retransmit upon entering fast recovery */
2550         sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2551         tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2552 }
2553 
2554 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2555 {
2556         struct tcp_sock *tp = tcp_sk(sk);
2557 
2558         if (inet_csk(sk)->icsk_ca_ops->cong_control)
2559                 return;
2560 
2561         /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2562         if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2563             (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2564                 tp->snd_cwnd = tp->snd_ssthresh;
2565                 tp->snd_cwnd_stamp = tcp_time_stamp;
2566         }
2567         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2568 }
2569 
2570 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2571 void tcp_enter_cwr(struct sock *sk)
2572 {
2573         struct tcp_sock *tp = tcp_sk(sk);
2574 
2575         tp->prior_ssthresh = 0;
2576         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2577                 tp->undo_marker = 0;
2578                 tcp_init_cwnd_reduction(sk);
2579                 tcp_set_ca_state(sk, TCP_CA_CWR);
2580         }
2581 }
2582 EXPORT_SYMBOL(tcp_enter_cwr);
2583 
2584 static void tcp_try_keep_open(struct sock *sk)
2585 {
2586         struct tcp_sock *tp = tcp_sk(sk);
2587         int state = TCP_CA_Open;
2588 
2589         if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2590                 state = TCP_CA_Disorder;
2591 
2592         if (inet_csk(sk)->icsk_ca_state != state) {
2593                 tcp_set_ca_state(sk, state);
2594                 tp->high_seq = tp->snd_nxt;
2595         }
2596 }
2597 
2598 static void tcp_try_to_open(struct sock *sk, int flag)
2599 {
2600         struct tcp_sock *tp = tcp_sk(sk);
2601 
2602         tcp_verify_left_out(tp);
2603 
2604         if (!tcp_any_retrans_done(sk))
2605                 tp->retrans_stamp = 0;
2606 
2607         if (flag & FLAG_ECE)
2608                 tcp_enter_cwr(sk);
2609 
2610         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2611                 tcp_try_keep_open(sk);
2612         }
2613 }
2614 
2615 static void tcp_mtup_probe_failed(struct sock *sk)
2616 {
2617         struct inet_connection_sock *icsk = inet_csk(sk);
2618 
2619         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2620         icsk->icsk_mtup.probe_size = 0;
2621         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2622 }
2623 
2624 static void tcp_mtup_probe_success(struct sock *sk)
2625 {
2626         struct tcp_sock *tp = tcp_sk(sk);
2627         struct inet_connection_sock *icsk = inet_csk(sk);
2628 
2629         /* FIXME: breaks with very large cwnd */
2630         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2631         tp->snd_cwnd = tp->snd_cwnd *
2632                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2633                        icsk->icsk_mtup.probe_size;
2634         tp->snd_cwnd_cnt = 0;
2635         tp->snd_cwnd_stamp = tcp_time_stamp;
2636         tp->snd_ssthresh = tcp_current_ssthresh(sk);
2637 
2638         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2639         icsk->icsk_mtup.probe_size = 0;
2640         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2641         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2642 }
2643 
2644 /* Do a simple retransmit without using the backoff mechanisms in
2645  * tcp_timer. This is used for path mtu discovery.
2646  * The socket is already locked here.
2647  */
2648 void tcp_simple_retransmit(struct sock *sk)
2649 {
2650         const struct inet_connection_sock *icsk = inet_csk(sk);
2651         struct tcp_sock *tp = tcp_sk(sk);
2652         struct sk_buff *skb;
2653         unsigned int mss = tcp_current_mss(sk);
2654         u32 prior_lost = tp->lost_out;
2655 
2656         tcp_for_write_queue(skb, sk) {
2657                 if (skb == tcp_send_head(sk))
2658                         break;
2659                 if (tcp_skb_seglen(skb) > mss &&
2660                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2661                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2662                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2663                                 tp->retrans_out -= tcp_skb_pcount(skb);
2664                         }
2665                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2666                 }
2667         }
2668 
2669         tcp_clear_retrans_hints_partial(tp);
2670 
2671         if (prior_lost == tp->lost_out)
2672                 return;
2673 
2674         if (tcp_is_reno(tp))
2675                 tcp_limit_reno_sacked(tp);
2676 
2677         tcp_verify_left_out(tp);
2678 
2679         /* Don't muck with the congestion window here.
2680          * Reason is that we do not increase amount of _data_
2681          * in network, but units changed and effective
2682          * cwnd/ssthresh really reduced now.
2683          */
2684         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2685                 tp->high_seq = tp->snd_nxt;
2686                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2687                 tp->prior_ssthresh = 0;
2688                 tp->undo_marker = 0;
2689                 tcp_set_ca_state(sk, TCP_CA_Loss);
2690         }
2691         tcp_xmit_retransmit_queue(sk);
2692 }
2693 EXPORT_SYMBOL(tcp_simple_retransmit);
2694 
2695 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2696 {
2697         struct tcp_sock *tp = tcp_sk(sk);
2698         int mib_idx;
2699 
2700         if (tcp_is_reno(tp))
2701                 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2702         else
2703                 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2704 
2705         NET_INC_STATS(sock_net(sk), mib_idx);
2706 
2707         tp->prior_ssthresh = 0;
2708         tcp_init_undo(tp);
2709 
2710         if (!tcp_in_cwnd_reduction(sk)) {
2711                 if (!ece_ack)
2712                         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2713                 tcp_init_cwnd_reduction(sk);
2714         }
2715         tcp_set_ca_state(sk, TCP_CA_Recovery);
2716 }
2717 
2718 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2719  * recovered or spurious. Otherwise retransmits more on partial ACKs.
2720  */
2721 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2722                              int *rexmit)
2723 {
2724         struct tcp_sock *tp = tcp_sk(sk);
2725         bool recovered = !before(tp->snd_una, tp->high_seq);
2726 
2727         if ((flag & FLAG_SND_UNA_ADVANCED) &&
2728             tcp_try_undo_loss(sk, false))
2729                 return;
2730 
2731         if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2732                 /* Step 3.b. A timeout is spurious if not all data are
2733                  * lost, i.e., never-retransmitted data are (s)acked.
2734                  */
2735                 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2736                     tcp_try_undo_loss(sk, true))
2737                         return;
2738 
2739                 if (after(tp->snd_nxt, tp->high_seq)) {
2740                         if (flag & FLAG_DATA_SACKED || is_dupack)
2741                                 tp->frto = 0; /* Step 3.a. loss was real */
2742                 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2743                         tp->high_seq = tp->snd_nxt;
2744                         /* Step 2.b. Try send new data (but deferred until cwnd
2745                          * is updated in tcp_ack()). Otherwise fall back to
2746                          * the conventional recovery.
2747                          */
2748                         if (tcp_send_head(sk) &&
2749                             after(tcp_wnd_end(tp), tp->snd_nxt)) {
2750                                 *rexmit = REXMIT_NEW;
2751                                 return;
2752                         }
2753                         tp->frto = 0;
2754                 }
2755         }
2756 
2757         if (recovered) {
2758                 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2759                 tcp_try_undo_recovery(sk);
2760                 return;
2761         }
2762         if (tcp_is_reno(tp)) {
2763                 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2764                  * delivered. Lower inflight to clock out (re)tranmissions.
2765                  */
2766                 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2767                         tcp_add_reno_sack(sk);
2768                 else if (flag & FLAG_SND_UNA_ADVANCED)
2769                         tcp_reset_reno_sack(tp);
2770         }
2771         *rexmit = REXMIT_LOST;
2772 }
2773 
2774 /* Undo during fast recovery after partial ACK. */
2775 static bool tcp_try_undo_partial(struct sock *sk, const int acked)
2776 {
2777         struct tcp_sock *tp = tcp_sk(sk);
2778 
2779         if (tp->undo_marker && tcp_packet_delayed(tp)) {
2780                 /* Plain luck! Hole if filled with delayed
2781                  * packet, rather than with a retransmit.
2782                  */
2783                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2784 
2785                 /* We are getting evidence that the reordering degree is higher
2786                  * than we realized. If there are no retransmits out then we
2787                  * can undo. Otherwise we clock out new packets but do not
2788                  * mark more packets lost or retransmit more.
2789                  */
2790                 if (tp->retrans_out)
2791                         return true;
2792 
2793                 if (!tcp_any_retrans_done(sk))
2794                         tp->retrans_stamp = 0;
2795 
2796                 DBGUNDO(sk, "partial recovery");
2797                 tcp_undo_cwnd_reduction(sk, true);
2798                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2799                 tcp_try_keep_open(sk);
2800                 return true;
2801         }
2802         return false;
2803 }
2804 
2805 /* Process an event, which can update packets-in-flight not trivially.
2806  * Main goal of this function is to calculate new estimate for left_out,
2807  * taking into account both packets sitting in receiver's buffer and
2808  * packets lost by network.
2809  *
2810  * Besides that it updates the congestion state when packet loss or ECN
2811  * is detected. But it does not reduce the cwnd, it is done by the
2812  * congestion control later.
2813  *
2814  * It does _not_ decide what to send, it is made in function
2815  * tcp_xmit_retransmit_queue().
2816  */
2817 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2818                                   bool is_dupack, int *ack_flag, int *rexmit)
2819 {
2820         struct inet_connection_sock *icsk = inet_csk(sk);
2821         struct tcp_sock *tp = tcp_sk(sk);
2822         int fast_rexmit = 0, flag = *ack_flag;
2823         bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2824                                     (tcp_fackets_out(tp) > tp->reordering));
2825 
2826         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2827                 tp->sacked_out = 0;
2828         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2829                 tp->fackets_out = 0;
2830 
2831         /* Now state machine starts.
2832          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2833         if (flag & FLAG_ECE)
2834                 tp->prior_ssthresh = 0;
2835 
2836         /* B. In all the states check for reneging SACKs. */
2837         if (tcp_check_sack_reneging(sk, flag))
2838                 return;
2839 
2840         /* C. Check consistency of the current state. */
2841         tcp_verify_left_out(tp);
2842 
2843         /* D. Check state exit conditions. State can be terminated
2844          *    when high_seq is ACKed. */
2845         if (icsk->icsk_ca_state == TCP_CA_Open) {
2846                 WARN_ON(tp->retrans_out != 0);
2847                 tp->retrans_stamp = 0;
2848         } else if (!before(tp->snd_una, tp->high_seq)) {
2849                 switch (icsk->icsk_ca_state) {
2850                 case TCP_CA_CWR:
2851                         /* CWR is to be held something *above* high_seq
2852                          * is ACKed for CWR bit to reach receiver. */
2853                         if (tp->snd_una != tp->high_seq) {
2854                                 tcp_end_cwnd_reduction(sk);
2855                                 tcp_set_ca_state(sk, TCP_CA_Open);
2856                         }
2857                         break;
2858 
2859                 case TCP_CA_Recovery:
2860                         if (tcp_is_reno(tp))
2861                                 tcp_reset_reno_sack(tp);
2862                         if (tcp_try_undo_recovery(sk))
2863                                 return;
2864                         tcp_end_cwnd_reduction(sk);
2865                         break;
2866                 }
2867         }
2868 
2869         /* Use RACK to detect loss */
2870         if (sysctl_tcp_recovery & TCP_RACK_LOST_RETRANS &&
2871             tcp_rack_mark_lost(sk)) {
2872                 flag |= FLAG_LOST_RETRANS;
2873                 *ack_flag |= FLAG_LOST_RETRANS;
2874         }
2875 
2876         /* E. Process state. */
2877         switch (icsk->icsk_ca_state) {
2878         case TCP_CA_Recovery:
2879                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2880                         if (tcp_is_reno(tp) && is_dupack)
2881                                 tcp_add_reno_sack(sk);
2882                 } else {
2883                         if (tcp_try_undo_partial(sk, acked))
2884                                 return;
2885                         /* Partial ACK arrived. Force fast retransmit. */
2886                         do_lost = tcp_is_reno(tp) ||
2887                                   tcp_fackets_out(tp) > tp->reordering;
2888                 }
2889                 if (tcp_try_undo_dsack(sk)) {
2890                         tcp_try_keep_open(sk);
2891                         return;
2892                 }
2893                 break;
2894         case TCP_CA_Loss:
2895                 tcp_process_loss(sk, flag, is_dupack, rexmit);
2896                 if (icsk->icsk_ca_state != TCP_CA_Open &&
2897                     !(flag & FLAG_LOST_RETRANS))
2898                         return;
2899                 /* Change state if cwnd is undone or retransmits are lost */
2900         default:
2901                 if (tcp_is_reno(tp)) {
2902                         if (flag & FLAG_SND_UNA_ADVANCED)
2903                                 tcp_reset_reno_sack(tp);
2904                         if (is_dupack)
2905                                 tcp_add_reno_sack(sk);
2906                 }
2907 
2908                 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2909                         tcp_try_undo_dsack(sk);
2910 
2911                 if (!tcp_time_to_recover(sk, flag)) {
2912                         tcp_try_to_open(sk, flag);
2913                         return;
2914                 }
2915 
2916                 /* MTU probe failure: don't reduce cwnd */
2917                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2918                     icsk->icsk_mtup.probe_size &&
2919                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2920                         tcp_mtup_probe_failed(sk);
2921                         /* Restores the reduction we did in tcp_mtup_probe() */
2922                         tp->snd_cwnd++;
2923                         tcp_simple_retransmit(sk);
2924                         return;
2925                 }
2926 
2927                 /* Otherwise enter Recovery state */
2928                 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2929                 fast_rexmit = 1;
2930         }
2931 
2932         if (do_lost)
2933                 tcp_update_scoreboard(sk, fast_rexmit);
2934         *rexmit = REXMIT_LOST;
2935 }
2936 
2937 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2938 {
2939         struct tcp_sock *tp = tcp_sk(sk);
2940         u32 wlen = sysctl_tcp_min_rtt_wlen * HZ;
2941 
2942         minmax_running_min(&tp->rtt_min, wlen, tcp_time_stamp,
2943                            rtt_us ? : jiffies_to_usecs(1));
2944 }
2945 
2946 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2947                                       long seq_rtt_us, long sack_rtt_us,
2948                                       long ca_rtt_us)
2949 {
2950         const struct tcp_sock *tp = tcp_sk(sk);
2951 
2952         /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2953          * broken middle-boxes or peers may corrupt TS-ECR fields. But
2954          * Karn's algorithm forbids taking RTT if some retransmitted data
2955          * is acked (RFC6298).
2956          */
2957         if (seq_rtt_us < 0)
2958                 seq_rtt_us = sack_rtt_us;
2959 
2960         /* RTTM Rule: A TSecr value received in a segment is used to
2961          * update the averaged RTT measurement only if the segment
2962          * acknowledges some new data, i.e., only if it advances the
2963          * left edge of the send window.
2964          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2965          */
2966         if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2967             flag & FLAG_ACKED)
2968                 seq_rtt_us = ca_rtt_us = jiffies_to_usecs(tcp_time_stamp -
2969                                                           tp->rx_opt.rcv_tsecr);
2970         if (seq_rtt_us < 0)
2971                 return false;
2972 
2973         /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2974          * always taken together with ACK, SACK, or TS-opts. Any negative
2975          * values will be skipped with the seq_rtt_us < 0 check above.
2976          */
2977         tcp_update_rtt_min(sk, ca_rtt_us);
2978         tcp_rtt_estimator(sk, seq_rtt_us);
2979         tcp_set_rto(sk);
2980 
2981         /* RFC6298: only reset backoff on valid RTT measurement. */
2982         inet_csk(sk)->icsk_backoff = 0;
2983         return true;
2984 }
2985 
2986 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2987 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2988 {
2989         long rtt_us = -1L;
2990 
2991         if (req && !req->num_retrans && tcp_rsk(req)->snt_synack.v64) {
2992                 struct skb_mstamp now;
2993 
2994                 skb_mstamp_get(&now);
2995                 rtt_us = skb_mstamp_us_delta(&now, &tcp_rsk(req)->snt_synack);
2996         }
2997 
2998         tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us);
2999 }
3000 
3001 
3002 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3003 {
3004         const struct inet_connection_sock *icsk = inet_csk(sk);
3005 
3006         icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3007         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3008 }
3009 
3010 /* Restart timer after forward progress on connection.
3011  * RFC2988 recommends to restart timer to now+rto.
3012  */
3013 void tcp_rearm_rto(struct sock *sk)
3014 {
3015         const struct inet_connection_sock *icsk = inet_csk(sk);
3016         struct tcp_sock *tp = tcp_sk(sk);
3017 
3018         /* If the retrans timer is currently being used by Fast Open
3019          * for SYN-ACK retrans purpose, stay put.
3020          */
3021         if (tp->fastopen_rsk)
3022                 return;
3023 
3024         if (!tp->packets_out) {
3025                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3026         } else {
3027                 u32 rto = inet_csk(sk)->icsk_rto;
3028                 /* Offset the time elapsed after installing regular RTO */
3029                 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3030                     icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3031                         struct sk_buff *skb = tcp_write_queue_head(sk);
3032                         const u32 rto_time_stamp =
3033                                 tcp_skb_timestamp(skb) + rto;
3034                         s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3035                         /* delta may not be positive if the socket is locked
3036                          * when the retrans timer fires and is rescheduled.
3037                          */
3038                         if (delta > 0)
3039                                 rto = delta;
3040                 }
3041                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3042                                           TCP_RTO_MAX);
3043         }
3044 }
3045 
3046 /* This function is called when the delayed ER timer fires. TCP enters
3047  * fast recovery and performs fast-retransmit.
3048  */
3049 void tcp_resume_early_retransmit(struct sock *sk)
3050 {
3051         struct tcp_sock *tp = tcp_sk(sk);
3052 
3053         tcp_rearm_rto(sk);
3054 
3055         /* Stop if ER is disabled after the delayed ER timer is scheduled */
3056         if (!tp->do_early_retrans)
3057                 return;
3058 
3059         tcp_enter_recovery(sk, false);
3060         tcp_update_scoreboard(sk, 1);
3061         tcp_xmit_retransmit_queue(sk);
3062 }
3063 
3064 /* If we get here, the whole TSO packet has not been acked. */
3065 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3066 {
3067         struct tcp_sock *tp = tcp_sk(sk);
3068         u32 packets_acked;
3069 
3070         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3071 
3072         packets_acked = tcp_skb_pcount(skb);
3073         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3074                 return 0;
3075         packets_acked -= tcp_skb_pcount(skb);
3076 
3077         if (packets_acked) {
3078                 BUG_ON(tcp_skb_pcount(skb) == 0);
3079                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3080         }
3081 
3082         return packets_acked;
3083 }
3084 
3085 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3086                            u32 prior_snd_una)
3087 {
3088         const struct skb_shared_info *shinfo;
3089 
3090         /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3091         if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3092                 return;
3093 
3094         shinfo = skb_shinfo(skb);
3095         if (!before(shinfo->tskey, prior_snd_una) &&
3096             before(shinfo->tskey, tcp_sk(sk)->snd_una))
3097                 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3098 }
3099 
3100 /* Remove acknowledged frames from the retransmission queue. If our packet
3101  * is before the ack sequence we can discard it as it's confirmed to have
3102  * arrived at the other end.
3103  */
3104 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3105                                u32 prior_snd_una, int *acked,
3106                                struct tcp_sacktag_state *sack,
3107                                struct skb_mstamp *now)
3108 {
3109         const struct inet_connection_sock *icsk = inet_csk(sk);
3110         struct skb_mstamp first_ackt, last_ackt;
3111         struct tcp_sock *tp = tcp_sk(sk);
3112         u32 prior_sacked = tp->sacked_out;
3113         u32 reord = tp->packets_out;
3114         bool fully_acked = true;
3115         long sack_rtt_us = -1L;
3116         long seq_rtt_us = -1L;
3117         long ca_rtt_us = -1L;
3118         struct sk_buff *skb;
3119         u32 pkts_acked = 0;
3120         u32 last_in_flight = 0;
3121         bool rtt_update;
3122         int flag = 0;
3123 
3124         first_ackt.v64 = 0;
3125 
3126         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3127                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3128                 u8 sacked = scb->sacked;
3129                 u32 acked_pcount;
3130 
3131                 tcp_ack_tstamp(sk, skb, prior_snd_una);
3132 
3133                 /* Determine how many packets and what bytes were acked, tso and else */
3134                 if (after(scb->end_seq, tp->snd_una)) {
3135                         if (tcp_skb_pcount(skb) == 1 ||
3136                             !after(tp->snd_una, scb->seq))
3137                                 break;
3138 
3139                         acked_pcount = tcp_tso_acked(sk, skb);
3140                         if (!acked_pcount)
3141                                 break;
3142                         fully_acked = false;
3143                 } else {
3144                         /* Speedup tcp_unlink_write_queue() and next loop */
3145                         prefetchw(skb->next);
3146                         acked_pcount = tcp_skb_pcount(skb);
3147                 }
3148 
3149                 if (unlikely(sacked & TCPCB_RETRANS)) {
3150                         if (sacked & TCPCB_SACKED_RETRANS)
3151                                 tp->retrans_out -= acked_pcount;
3152                         flag |= FLAG_RETRANS_DATA_ACKED;
3153                 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3154                         last_ackt = skb->skb_mstamp;
3155                         WARN_ON_ONCE(last_ackt.v64 == 0);
3156                         if (!first_ackt.v64)
3157                                 first_ackt = last_ackt;
3158 
3159                         last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3160                         reord = min(pkts_acked, reord);
3161                         if (!after(scb->end_seq, tp->high_seq))
3162                                 flag |= FLAG_ORIG_SACK_ACKED;
3163                 }
3164 
3165                 if (sacked & TCPCB_SACKED_ACKED) {
3166                         tp->sacked_out -= acked_pcount;
3167                 } else if (tcp_is_sack(tp)) {
3168                         tp->delivered += acked_pcount;
3169                         if (!tcp_skb_spurious_retrans(tp, skb))
3170                                 tcp_rack_advance(tp, &skb->skb_mstamp, sacked);
3171                 }
3172                 if (sacked & TCPCB_LOST)
3173                         tp->lost_out -= acked_pcount;
3174 
3175                 tp->packets_out -= acked_pcount;
3176                 pkts_acked += acked_pcount;
3177                 tcp_rate_skb_delivered(sk, skb, sack->rate);
3178 
3179                 /* Initial outgoing SYN's get put onto the write_queue
3180                  * just like anything else we transmit.  It is not
3181                  * true data, and if we misinform our callers that
3182                  * this ACK acks real data, we will erroneously exit
3183                  * connection startup slow start one packet too
3184                  * quickly.  This is severely frowned upon behavior.
3185                  */
3186                 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3187                         flag |= FLAG_DATA_ACKED;
3188                 } else {
3189                         flag |= FLAG_SYN_ACKED;
3190                         tp->retrans_stamp = 0;
3191                 }
3192 
3193                 if (!fully_acked)
3194                         break;
3195 
3196                 tcp_unlink_write_queue(skb, sk);
3197                 sk_wmem_free_skb(sk, skb);
3198                 if (unlikely(skb == tp->retransmit_skb_hint))
3199                         tp->retransmit_skb_hint = NULL;
3200                 if (unlikely(skb == tp->lost_skb_hint))
3201                         tp->lost_skb_hint = NULL;
3202         }
3203 
3204         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3205                 tp->snd_up = tp->snd_una;
3206 
3207         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3208                 flag |= FLAG_SACK_RENEGING;
3209 
3210         if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3211                 seq_rtt_us = skb_mstamp_us_delta(now, &first_ackt);
3212                 ca_rtt_us = skb_mstamp_us_delta(now, &last_ackt);
3213         }
3214         if (sack->first_sackt.v64) {
3215                 sack_rtt_us = skb_mstamp_us_delta(now, &sack->first_sackt);
3216                 ca_rtt_us = skb_mstamp_us_delta(now, &sack->last_sackt);
3217         }
3218         sack->rate->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet, or -1 */
3219         rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3220                                         ca_rtt_us);
3221 
3222         if (flag & FLAG_ACKED) {
3223                 tcp_rearm_rto(sk);
3224                 if (unlikely(icsk->icsk_mtup.probe_size &&
3225                              !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3226                         tcp_mtup_probe_success(sk);
3227                 }
3228 
3229                 if (tcp_is_reno(tp)) {
3230                         tcp_remove_reno_sacks(sk, pkts_acked);
3231                 } else {
3232                         int delta;
3233 
3234                         /* Non-retransmitted hole got filled? That's reordering */
3235                         if (reord < prior_fackets)
3236                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3237 
3238                         delta = tcp_is_fack(tp) ? pkts_acked :
3239                                                   prior_sacked - tp->sacked_out;
3240                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3241                 }
3242 
3243                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3244 
3245         } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3246                    sack_rtt_us > skb_mstamp_us_delta(now, &skb->skb_mstamp)) {
3247                 /* Do not re-arm RTO if the sack RTT is measured from data sent
3248                  * after when the head was last (re)transmitted. Otherwise the
3249                  * timeout may continue to extend in loss recovery.
3250                  */
3251                 tcp_rearm_rto(sk);
3252         }
3253 
3254         if (icsk->icsk_ca_ops->pkts_acked) {
3255                 struct ack_sample sample = { .pkts_acked = pkts_acked,
3256                                              .rtt_us = ca_rtt_us,
3257                                              .in_flight = last_in_flight };
3258 
3259                 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3260         }
3261 
3262 #if FASTRETRANS_DEBUG > 0
3263         WARN_ON((int)tp->sacked_out < 0);
3264         WARN_ON((int)tp->lost_out < 0);
3265         WARN_ON((int)tp->retrans_out < 0);
3266         if (!tp->packets_out && tcp_is_sack(tp)) {
3267                 icsk = inet_csk(sk);
3268                 if (tp->lost_out) {
3269                         pr_debug("Leak l=%u %d\n",
3270                                  tp->lost_out, icsk->icsk_ca_state);
3271                         tp->lost_out = 0;
3272                 }
3273                 if (tp->sacked_out) {
3274                         pr_debug("Leak s=%u %d\n",
3275                                  tp->sacked_out, icsk->icsk_ca_state);
3276                         tp->sacked_out = 0;
3277                 }
3278                 if (tp->retrans_out) {
3279                         pr_debug("Leak r=%u %d\n",
3280                                  tp->retrans_out, icsk->icsk_ca_state);
3281                         tp->retrans_out = 0;
3282                 }
3283         }
3284 #endif
3285         *acked = pkts_acked;
3286         return flag;
3287 }
3288 
3289 static void tcp_ack_probe(struct sock *sk)
3290 {
3291         const struct tcp_sock *tp = tcp_sk(sk);
3292         struct inet_connection_sock *icsk = inet_csk(sk);
3293 
3294         /* Was it a usable window open? */
3295 
3296         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3297                 icsk->icsk_backoff = 0;
3298                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3299                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3300                  * This function is not for random using!
3301                  */
3302         } else {
3303                 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3304 
3305                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3306                                           when, TCP_RTO_MAX);
3307         }
3308 }
3309 
3310 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3311 {
3312         return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3313                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3314 }
3315 
3316 /* Decide wheather to run the increase function of congestion control. */
3317 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3318 {
3319         /* If reordering is high then always grow cwnd whenever data is
3320          * delivered regardless of its ordering. Otherwise stay conservative
3321          * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3322          * new SACK or ECE mark may first advance cwnd here and later reduce
3323          * cwnd in tcp_fastretrans_alert() based on more states.
3324          */
3325         if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3326                 return flag & FLAG_FORWARD_PROGRESS;
3327 
3328         return flag & FLAG_DATA_ACKED;
3329 }
3330 
3331 /* The "ultimate" congestion control function that aims to replace the rigid
3332  * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3333  * It's called toward the end of processing an ACK with precise rate
3334  * information. All transmission or retransmission are delayed afterwards.
3335  */
3336 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3337                              int flag, const struct rate_sample *rs)
3338 {
3339         const struct inet_connection_sock *icsk = inet_csk(sk);
3340 
3341         if (icsk->icsk_ca_ops->cong_control) {
3342                 icsk->icsk_ca_ops->cong_control(sk, rs);
3343                 return;
3344         }
3345 
3346         if (tcp_in_cwnd_reduction(sk)) {
3347                 /* Reduce cwnd if state mandates */
3348                 tcp_cwnd_reduction(sk, acked_sacked, flag);
3349         } else if (tcp_may_raise_cwnd(sk, flag)) {
3350                 /* Advance cwnd if state allows */
3351                 tcp_cong_avoid(sk, ack, acked_sacked);
3352         }
3353         tcp_update_pacing_rate(sk);
3354 }
3355 
3356 /* Check that window update is acceptable.
3357  * The function assumes that snd_una<=ack<=snd_next.
3358  */
3359 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3360                                         const u32 ack, const u32 ack_seq,
3361                                         const u32 nwin)
3362 {
3363         return  after(ack, tp->snd_una) ||
3364                 after(ack_seq, tp->snd_wl1) ||
3365                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3366 }
3367 
3368 /* If we update tp->snd_una, also update tp->bytes_acked */
3369 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3370 {
3371         u32 delta = ack - tp->snd_una;
3372 
3373         sock_owned_by_me((struct sock *)tp);
3374         u64_stats_update_begin_raw(&tp->syncp);
3375         tp->bytes_acked += delta;
3376         u64_stats_update_end_raw(&tp->syncp);
3377         tp->snd_una = ack;
3378 }
3379 
3380 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3381 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3382 {
3383         u32 delta = seq - tp->rcv_nxt;
3384 
3385         sock_owned_by_me((struct sock *)tp);
3386         u64_stats_update_begin_raw(&tp->syncp);
3387         tp->bytes_received += delta;
3388         u64_stats_update_end_raw(&tp->syncp);
3389         tp->rcv_nxt = seq;
3390 }
3391 
3392 /* Update our send window.
3393  *
3394  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3395  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3396  */
3397 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3398                                  u32 ack_seq)
3399 {
3400         struct tcp_sock *tp = tcp_sk(sk);
3401         int flag = 0;
3402         u32 nwin = ntohs(tcp_hdr(skb)->window);
3403 
3404         if (likely(!tcp_hdr(skb)->syn))
3405                 nwin <<= tp->rx_opt.snd_wscale;
3406 
3407         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3408                 flag |= FLAG_WIN_UPDATE;
3409                 tcp_update_wl(tp, ack_seq);
3410 
3411                 if (tp->snd_wnd != nwin) {
3412                         tp->snd_wnd = nwin;
3413 
3414                         /* Note, it is the only place, where
3415                          * fast path is recovered for sending TCP.
3416                          */
3417                         tp->pred_flags = 0;
3418                         tcp_fast_path_check(sk);
3419 
3420                         if (tcp_send_head(sk))
3421                                 tcp_slow_start_after_idle_check(sk);
3422 
3423                         if (nwin > tp->max_window) {
3424                                 tp->max_window = nwin;
3425                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3426                         }
3427                 }
3428         }
3429 
3430         tcp_snd_una_update(tp, ack);
3431 
3432         return flag;
3433 }
3434 
3435 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3436                                    u32 *last_oow_ack_time)
3437 {
3438         if (*last_oow_ack_time) {
3439                 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3440 
3441                 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3442                         NET_INC_STATS(net, mib_idx);
3443                         return true;    /* rate-limited: don't send yet! */
3444                 }
3445         }
3446 
3447         *last_oow_ack_time = tcp_time_stamp;
3448 
3449         return false;   /* not rate-limited: go ahead, send dupack now! */
3450 }
3451 
3452 /* Return true if we're currently rate-limiting out-of-window ACKs and
3453  * thus shouldn't send a dupack right now. We rate-limit dupacks in
3454  * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3455  * attacks that send repeated SYNs or ACKs for the same connection. To
3456  * do this, we do not send a duplicate SYNACK or ACK if the remote
3457  * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3458  */
3459 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3460                           int mib_idx, u32 *last_oow_ack_time)
3461 {
3462         /* Data packets without SYNs are not likely part of an ACK loop. */
3463         if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3464             !tcp_hdr(skb)->syn)
3465                 return false;
3466 
3467         return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3468 }
3469 
3470 /* RFC 5961 7 [ACK Throttling] */
3471 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3472 {
3473         /* unprotected vars, we dont care of overwrites */
3474         static u32 challenge_timestamp;
3475         static unsigned int challenge_count;
3476         struct tcp_sock *tp = tcp_sk(sk);
3477         u32 count, now;
3478 
3479         /* First check our per-socket dupack rate limit. */
3480         if (__tcp_oow_rate_limited(sock_net(sk),
3481                                    LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3482                                    &tp->last_oow_ack_time))
3483                 return;
3484 
3485         /* Then check host-wide RFC 5961 rate limit. */
3486         now = jiffies / HZ;
3487         if (now != challenge_timestamp) {
3488                 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3489 
3490                 challenge_timestamp = now;
3491                 WRITE_ONCE(challenge_count, half +
3492                            prandom_u32_max(sysctl_tcp_challenge_ack_limit));
3493         }
3494         count = READ_ONCE(challenge_count);
3495         if (count > 0) {
3496                 WRITE_ONCE(challenge_count, count - 1);
3497                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3498                 tcp_send_ack(sk);
3499         }
3500 }
3501 
3502 static void tcp_store_ts_recent(struct tcp_sock *tp)
3503 {
3504         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3505         tp->rx_opt.ts_recent_stamp = get_seconds();
3506 }
3507 
3508 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3509 {
3510         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3511                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3512                  * extra check below makes sure this can only happen
3513                  * for pure ACK frames.  -DaveM
3514                  *
3515                  * Not only, also it occurs for expired timestamps.
3516                  */
3517 
3518                 if (tcp_paws_check(&tp->rx_opt, 0))
3519                         tcp_store_ts_recent(tp);
3520         }
3521 }
3522 
3523 /* This routine deals with acks during a TLP episode.
3524  * We mark the end of a TLP episode on receiving TLP dupack or when
3525  * ack is after tlp_high_seq.
3526  * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3527  */
3528 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3529 {
3530         struct tcp_sock *tp = tcp_sk(sk);
3531 
3532         if (before(ack, tp->tlp_high_seq))
3533                 return;
3534 
3535         if (flag & FLAG_DSACKING_ACK) {
3536                 /* This DSACK means original and TLP probe arrived; no loss */
3537                 tp->tlp_high_seq = 0;
3538         } else if (after(ack, tp->tlp_high_seq)) {
3539                 /* ACK advances: there was a loss, so reduce cwnd. Reset
3540                  * tlp_high_seq in tcp_init_cwnd_reduction()
3541                  */
3542                 tcp_init_cwnd_reduction(sk);
3543                 tcp_set_ca_state(sk, TCP_CA_CWR);
3544                 tcp_end_cwnd_reduction(sk);
3545                 tcp_try_keep_open(sk);
3546                 NET_INC_STATS(sock_net(sk),
3547                                 LINUX_MIB_TCPLOSSPROBERECOVERY);
3548         } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3549                              FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3550                 /* Pure dupack: original and TLP probe arrived; no loss */
3551                 tp->tlp_high_seq = 0;
3552         }
3553 }
3554 
3555 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3556 {
3557         const struct inet_connection_sock *icsk = inet_csk(sk);
3558 
3559         if (icsk->icsk_ca_ops->in_ack_event)
3560                 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3561 }
3562 
3563 /* Congestion control has updated the cwnd already. So if we're in
3564  * loss recovery then now we do any new sends (for FRTO) or
3565  * retransmits (for CA_Loss or CA_recovery) that make sense.
3566  */
3567 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3568 {
3569         struct tcp_sock *tp = tcp_sk(sk);
3570 
3571         if (rexmit == REXMIT_NONE)
3572                 return;
3573 
3574         if (unlikely(rexmit == 2)) {
3575                 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3576                                           TCP_NAGLE_OFF);
3577                 if (after(tp->snd_nxt, tp->high_seq))
3578                         return;
3579                 tp->frto = 0;
3580         }
3581         tcp_xmit_retransmit_queue(sk);
3582 }
3583 
3584 /* This routine deals with incoming acks, but not outgoing ones. */
3585 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3586 {
3587         struct inet_connection_sock *icsk = inet_csk(sk);
3588         struct tcp_sock *tp = tcp_sk(sk);
3589         struct tcp_sacktag_state sack_state;
3590         struct rate_sample rs = { .prior_delivered = 0 };
3591         u32 prior_snd_una = tp->snd_una;
3592         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3593         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3594         bool is_dupack = false;
3595         u32 prior_fackets;
3596         int prior_packets = tp->packets_out;
3597         u32 delivered = tp->delivered;
3598         u32 lost = tp->lost;
3599         int acked = 0; /* Number of packets newly acked */
3600         int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3601         struct skb_mstamp now;
3602 
3603         sack_state.first_sackt.v64 = 0;
3604         sack_state.rate = &rs;
3605 
3606         /* We very likely will need to access write queue head. */
3607         prefetchw(sk->sk_write_queue.next);
3608 
3609         /* If the ack is older than previous acks
3610          * then we can probably ignore it.
3611          */
3612         if (before(ack, prior_snd_una)) {
3613                 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3614                 if (before(ack, prior_snd_una - tp->max_window)) {
3615                         tcp_send_challenge_ack(sk, skb);
3616                         return -1;
3617                 }
3618                 goto old_ack;
3619         }
3620 
3621         /* If the ack includes data we haven't sent yet, discard
3622          * this segment (RFC793 Section 3.9).
3623          */
3624         if (after(ack, tp->snd_nxt))
3625                 goto invalid_ack;
3626 
3627         skb_mstamp_get(&now);
3628 
3629         if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3630             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3631                 tcp_rearm_rto(sk);
3632 
3633         if (after(ack, prior_snd_una)) {
3634                 flag |= FLAG_SND_UNA_ADVANCED;
3635                 icsk->icsk_retransmits = 0;
3636         }
3637 
3638         prior_fackets = tp->fackets_out;
3639         rs.prior_in_flight = tcp_packets_in_flight(tp);
3640 
3641         /* ts_recent update must be made after we are sure that the packet
3642          * is in window.
3643          */
3644         if (flag & FLAG_UPDATE_TS_RECENT)
3645                 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3646 
3647         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3648                 /* Window is constant, pure forward advance.
3649                  * No more checks are required.
3650                  * Note, we use the fact that SND.UNA>=SND.WL2.
3651                  */
3652                 tcp_update_wl(tp, ack_seq);
3653                 tcp_snd_una_update(tp, ack);
3654                 flag |= FLAG_WIN_UPDATE;
3655 
3656                 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3657 
3658                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3659         } else {
3660                 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3661 
3662                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3663                         flag |= FLAG_DATA;
3664                 else
3665                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3666 
3667                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3668 
3669                 if (TCP_SKB_CB(skb)->sacked)
3670                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3671                                                         &sack_state);
3672 
3673                 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3674                         flag |= FLAG_ECE;
3675                         ack_ev_flags |= CA_ACK_ECE;
3676                 }
3677 
3678                 if (flag & FLAG_WIN_UPDATE)
3679                         ack_ev_flags |= CA_ACK_WIN_UPDATE;
3680 
3681                 tcp_in_ack_event(sk, ack_ev_flags);
3682         }
3683 
3684         /* We passed data and got it acked, remove any soft error
3685          * log. Something worked...
3686          */
3687         sk->sk_err_soft = 0;
3688         icsk->icsk_probes_out = 0;
3689         tp->rcv_tstamp = tcp_time_stamp;
3690         if (!prior_packets)
3691                 goto no_queue;
3692 
3693         /* See if we can take anything off of the retransmit queue. */
3694         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, &acked,
3695                                     &sack_state, &now);
3696 
3697         if (tcp_ack_is_dubious(sk, flag)) {
3698                 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3699                 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3700         }
3701         if (tp->tlp_high_seq)
3702                 tcp_process_tlp_ack(sk, ack, flag);
3703 
3704         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3705                 struct dst_entry *dst = __sk_dst_get(sk);
3706                 if (dst)
3707                         dst_confirm(dst);
3708         }
3709 
3710         if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3711                 tcp_schedule_loss_probe(sk);
3712         delivered = tp->delivered - delivered;  /* freshly ACKed or SACKed */
3713         lost = tp->lost - lost;                 /* freshly marked lost */
3714         tcp_rate_gen(sk, delivered, lost, &now, &rs);
3715         tcp_cong_control(sk, ack, delivered, flag, &rs);
3716         tcp_xmit_recovery(sk, rexmit);
3717         return 1;
3718 
3719 no_queue:
3720         /* If data was DSACKed, see if we can undo a cwnd reduction. */
3721         if (flag & FLAG_DSACKING_ACK)
3722                 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3723         /* If this ack opens up a zero window, clear backoff.  It was
3724          * being used to time the probes, and is probably far higher than
3725          * it needs to be for normal retransmission.
3726          */
3727         if (tcp_send_head(sk))
3728                 tcp_ack_probe(sk);
3729 
3730         if (tp->tlp_high_seq)
3731                 tcp_process_tlp_ack(sk, ack, flag);
3732         return 1;
3733 
3734 invalid_ack:
3735         SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3736         return -1;
3737 
3738 old_ack:
3739         /* If data was SACKed, tag it and see if we should send more data.
3740          * If data was DSACKed, see if we can undo a cwnd reduction.
3741          */
3742         if (TCP_SKB_CB(skb)->sacked) {
3743                 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3744                                                 &sack_state);
3745                 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3746                 tcp_xmit_recovery(sk, rexmit);
3747         }
3748 
3749         SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3750         return 0;
3751 }
3752 
3753 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3754                                       bool syn, struct tcp_fastopen_cookie *foc,
3755                                       bool exp_opt)
3756 {
3757         /* Valid only in SYN or SYN-ACK with an even length.  */
3758         if (!foc || !syn || len < 0 || (len & 1))
3759                 return;
3760 
3761         if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3762             len <= TCP_FASTOPEN_COOKIE_MAX)
3763                 memcpy(foc->val, cookie, len);
3764         else if (len != 0)
3765                 len = -1;
3766         foc->len = len;
3767         foc->exp = exp_opt;
3768 }
3769 
3770 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3771  * But, this can also be called on packets in the established flow when
3772  * the fast version below fails.
3773  */
3774 void tcp_parse_options(const struct sk_buff *skb,
3775                        struct tcp_options_received *opt_rx, int estab,
3776                        struct tcp_fastopen_cookie *foc)
3777 {
3778         const unsigned char *ptr;
3779         const struct tcphdr *th = tcp_hdr(skb);
3780         int length = (th->doff * 4) - sizeof(struct tcphdr);
3781 
3782         ptr = (const unsigned char *)(th + 1);
3783         opt_rx->saw_tstamp = 0;
3784 
3785         while (length > 0) {
3786                 int opcode = *ptr++;
3787                 int opsize;
3788 
3789                 switch (opcode) {
3790                 case TCPOPT_EOL:
3791                         return;
3792                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3793                         length--;
3794                         continue;
3795                 default:
3796                         opsize = *ptr++;
3797                         if (opsize < 2) /* "silly options" */
3798                                 return;
3799                         if (opsize > length)
3800                                 return; /* don't parse partial options */
3801                         switch (opcode) {
3802                         case TCPOPT_MSS:
3803                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3804                                         u16 in_mss = get_unaligned_be16(ptr);
3805                                         if (in_mss) {
3806                                                 if (opt_rx->user_mss &&
3807                                                     opt_rx->user_mss < in_mss)
3808                                                         in_mss = opt_rx->user_mss;
3809                                                 opt_rx->mss_clamp = in_mss;
3810                                         }
3811                                 }
3812                                 break;
3813                         case TCPOPT_WINDOW:
3814                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3815                                     !estab && sysctl_tcp_window_scaling) {
3816                                         __u8 snd_wscale = *(__u8 *)ptr;
3817                                         opt_rx->wscale_ok = 1;
3818                                         if (snd_wscale > 14) {
3819                                                 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3820                                                                      __func__,
3821                                                                      snd_wscale);
3822                                                 snd_wscale = 14;
3823                                         }
3824                                         opt_rx->snd_wscale = snd_wscale;
3825                                 }
3826                                 break;
3827                         case TCPOPT_TIMESTAMP:
3828                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3829                                     ((estab && opt_rx->tstamp_ok) ||
3830                                      (!estab && sysctl_tcp_timestamps))) {
3831                                         opt_rx->saw_tstamp = 1;
3832                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3833                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3834                                 }
3835                                 break;
3836                         case TCPOPT_SACK_PERM:
3837                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3838                                     !estab && sysctl_tcp_sack) {
3839                                         opt_rx->sack_ok = TCP_SACK_SEEN;
3840                                         tcp_sack_reset(opt_rx);
3841                                 }
3842                                 break;
3843 
3844                         case TCPOPT_SACK:
3845                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3846                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3847                                    opt_rx->sack_ok) {
3848                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3849                                 }
3850                                 break;
3851 #ifdef CONFIG_TCP_MD5SIG
3852                         case TCPOPT_MD5SIG:
3853                                 /*
3854                                  * The MD5 Hash has already been
3855                                  * checked (see tcp_v{4,6}_do_rcv()).
3856                                  */
3857                                 break;
3858 #endif
3859                         case TCPOPT_FASTOPEN:
3860                                 tcp_parse_fastopen_option(
3861                                         opsize - TCPOLEN_FASTOPEN_BASE,
3862                                         ptr, th->syn, foc, false);
3863                                 break;
3864 
3865                         case TCPOPT_EXP:
3866                                 /* Fast Open option shares code 254 using a
3867                                  * 16 bits magic number.
3868                                  */
3869                                 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3870                                     get_unaligned_be16(ptr) ==
3871                                     TCPOPT_FASTOPEN_MAGIC)
3872                                         tcp_parse_fastopen_option(opsize -
3873                                                 TCPOLEN_EXP_FASTOPEN_BASE,
3874                                                 ptr + 2, th->syn, foc, true);
3875                                 break;
3876 
3877                         }
3878                         ptr += opsize-2;
3879                         length -= opsize;
3880                 }
3881         }
3882 }
3883 EXPORT_SYMBOL(tcp_parse_options);
3884 
3885 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3886 {
3887         const __be32 *ptr = (const __be32 *)(th + 1);
3888 
3889         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3890                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3891                 tp->rx_opt.saw_tstamp = 1;
3892                 ++ptr;
3893                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3894                 ++ptr;
3895                 if (*ptr)
3896                         tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3897                 else
3898                         tp->rx_opt.rcv_tsecr = 0;
3899                 return true;
3900         }
3901         return false;
3902 }
3903 
3904 /* Fast parse options. This hopes to only see timestamps.
3905  * If it is wrong it falls back on tcp_parse_options().
3906  */
3907 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3908                                    const struct tcphdr *th, struct tcp_sock *tp)
3909 {
3910         /* In the spirit of fast parsing, compare doff directly to constant
3911          * values.  Because equality is used, short doff can be ignored here.
3912          */
3913         if (th->doff == (sizeof(*th) / 4)) {
3914                 tp->rx_opt.saw_tstamp = 0;
3915                 return false;
3916         } else if (tp->rx_opt.tstamp_ok &&
3917                    th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3918                 if (tcp_parse_aligned_timestamp(tp, th))
3919                         return true;
3920         }
3921 
3922         tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3923         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3924                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3925 
3926         return true;
3927 }
3928 
3929 #ifdef CONFIG_TCP_MD5SIG
3930 /*
3931  * Parse MD5 Signature option
3932  */
3933 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3934 {
3935         int length = (th->doff << 2) - sizeof(*th);
3936         const u8 *ptr = (const u8 *)(th + 1);
3937 
3938         /* If the TCP option is too short, we can short cut */
3939         if (length < TCPOLEN_MD5SIG)
3940                 return NULL;
3941 
3942         while (length > 0) {
3943                 int opcode = *ptr++;
3944                 int opsize;
3945 
3946                 switch (opcode) {
3947                 case TCPOPT_EOL:
3948                         return NULL;
3949                 case TCPOPT_NOP:
3950                         length--;
3951                         continue;
3952                 default:
3953                         opsize = *ptr++;
3954                         if (opsize < 2 || opsize > length)
3955                                 return NULL;
3956                         if (opcode == TCPOPT_MD5SIG)
3957                                 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3958                 }
3959                 ptr += opsize - 2;
3960                 length -= opsize;
3961         }
3962         return NULL;
3963 }
3964 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3965 #endif
3966 
3967 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3968  *
3969  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3970  * it can pass through stack. So, the following predicate verifies that
3971  * this segment is not used for anything but congestion avoidance or
3972  * fast retransmit. Moreover, we even are able to eliminate most of such
3973  * second order effects, if we apply some small "replay" window (~RTO)
3974  * to timestamp space.
3975  *
3976  * All these measures still do not guarantee that we reject wrapped ACKs
3977  * on networks with high bandwidth, when sequence space is recycled fastly,
3978  * but it guarantees that such events will be very rare and do not affect
3979  * connection seriously. This doesn't look nice, but alas, PAWS is really
3980  * buggy extension.
3981  *
3982  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3983  * states that events when retransmit arrives after original data are rare.
3984  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3985  * the biggest problem on large power networks even with minor reordering.
3986  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3987  * up to bandwidth of 18Gigabit/sec. 8) ]
3988  */
3989 
3990 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3991 {
3992         const struct tcp_sock *tp = tcp_sk(sk);
3993         const struct tcphdr *th = tcp_hdr(skb);
3994         u32 seq = TCP_SKB_CB(skb)->seq;
3995         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3996 
3997         return (/* 1. Pure ACK with correct sequence number. */
3998                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3999 
4000                 /* 2. ... and duplicate ACK. */
4001                 ack == tp->snd_una &&
4002 
4003                 /* 3. ... and does not update window. */
4004                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4005 
4006                 /* 4. ... and sits in replay window. */
4007                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4008 }
4009 
4010 static inline bool tcp_paws_discard(const struct sock *sk,
4011                                    const struct sk_buff *skb)
4012 {
4013         const struct tcp_sock *tp = tcp_sk(sk);
4014 
4015         return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4016                !tcp_disordered_ack(sk, skb);
4017 }
4018 
4019 /* Check segment sequence number for validity.
4020  *
4021  * Segment controls are considered valid, if the segment
4022  * fits to the window after truncation to the window. Acceptability
4023  * of data (and SYN, FIN, of course) is checked separately.
4024  * See tcp_data_queue(), for example.
4025  *
4026  * Also, controls (RST is main one) are accepted using RCV.WUP instead
4027  * of RCV.NXT. Peer still did not advance his SND.UNA when we
4028  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4029  * (borrowed from freebsd)
4030  */
4031 
4032 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4033 {
4034         return  !before(end_seq, tp->rcv_wup) &&
4035                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4036 }
4037 
4038 /* When we get a reset we do this. */
4039 void tcp_reset(struct sock *sk)
4040 {
4041         /* We want the right error as BSD sees it (and indeed as we do). */
4042         switch (sk->sk_state) {
4043         case TCP_SYN_SENT:
4044                 sk->sk_err = ECONNREFUSED;
4045                 break;
4046         case TCP_CLOSE_WAIT:
4047                 sk->sk_err = EPIPE;
4048                 break;
4049         case TCP_CLOSE:
4050                 return;
4051         default:
4052                 sk->sk_err = ECONNRESET;
4053         }
4054         /* This barrier is coupled with smp_rmb() in tcp_poll() */
4055         smp_wmb();
4056 
4057         if (!sock_flag(sk, SOCK_DEAD))
4058                 sk->sk_error_report(sk);
4059 
4060         tcp_done(sk);
4061 }
4062 
4063 /*
4064  *      Process the FIN bit. This now behaves as it is supposed to work
4065  *      and the FIN takes effect when it is validly part of sequence
4066  *      space. Not before when we get holes.
4067  *
4068  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4069  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
4070  *      TIME-WAIT)
4071  *
4072  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
4073  *      close and we go into CLOSING (and later onto TIME-WAIT)
4074  *
4075  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4076  */
4077 void tcp_fin(struct sock *sk)
4078 {
4079         struct tcp_sock *tp = tcp_sk(sk);
4080 
4081         inet_csk_schedule_ack(sk);
4082 
4083         sk->sk_shutdown |= RCV_SHUTDOWN;
4084         sock_set_flag(sk, SOCK_DONE);
4085 
4086         switch (sk->sk_state) {
4087         case TCP_SYN_RECV:
4088         case TCP_ESTABLISHED:
4089                 /* Move to CLOSE_WAIT */
4090                 tcp_set_state(sk, TCP_CLOSE_WAIT);
4091                 inet_csk(sk)->icsk_ack.pingpong = 1;
4092                 break;
4093 
4094         case TCP_CLOSE_WAIT:
4095         case TCP_CLOSING:
4096                 /* Received a retransmission of the FIN, do
4097                  * nothing.
4098                  */
4099                 break;
4100         case TCP_LAST_ACK:
4101                 /* RFC793: Remain in the LAST-ACK state. */
4102                 break;
4103 
4104         case TCP_FIN_WAIT1:
4105                 /* This case occurs when a simultaneous close
4106                  * happens, we must ack the received FIN and
4107                  * enter the CLOSING state.
4108                  */
4109                 tcp_send_ack(sk);
4110                 tcp_set_state(sk, TCP_CLOSING);
4111                 break;
4112         case TCP_FIN_WAIT2:
4113                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4114                 tcp_send_ack(sk);
4115                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4116                 break;
4117         default:
4118                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4119                  * cases we should never reach this piece of code.
4120                  */
4121                 pr_err("%s: Impossible, sk->sk_state=%d\n",
4122                        __func__, sk->sk_state);
4123                 break;
4124         }
4125 
4126         /* It _is_ possible, that we have something out-of-order _after_ FIN.
4127          * Probably, we should reset in this case. For now drop them.
4128          */
4129         skb_rbtree_purge(&tp->out_of_order_queue);
4130         if (tcp_is_sack(tp))
4131                 tcp_sack_reset(&tp->rx_opt);
4132         sk_mem_reclaim(sk);
4133 
4134         if (!sock_flag(sk, SOCK_DEAD)) {
4135                 sk->sk_state_change(sk);
4136 
4137                 /* Do not send POLL_HUP for half duplex close. */
4138                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4139                     sk->sk_state == TCP_CLOSE)
4140                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4141                 else
4142                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4143         }
4144 }
4145 
4146 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4147                                   u32 end_seq)
4148 {
4149         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4150                 if (before(seq, sp->start_seq))
4151                         sp->start_seq = seq;
4152                 if (after(end_seq, sp->end_seq))
4153                         sp->end_seq = end_seq;
4154                 return true;
4155         }
4156         return false;
4157 }
4158 
4159 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4160 {
4161         struct tcp_sock *tp = tcp_sk(sk);
4162 
4163         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4164                 int mib_idx;
4165 
4166                 if (before(seq, tp->rcv_nxt))
4167                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4168                 else
4169                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4170 
4171                 NET_INC_STATS(sock_net(sk), mib_idx);
4172 
4173                 tp->rx_opt.dsack = 1;
4174                 tp->duplicate_sack[0].start_seq = seq;
4175                 tp->duplicate_sack[0].end_seq = end_seq;
4176         }
4177 }
4178 
4179 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4180 {
4181         struct tcp_sock *tp = tcp_sk(sk);
4182 
4183         if (!tp->rx_opt.dsack)
4184                 tcp_dsack_set(sk, seq, end_seq);
4185         else
4186                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4187 }
4188 
4189 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4190 {
4191         struct tcp_sock *tp = tcp_sk(sk);
4192 
4193         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4194             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4195                 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4196                 tcp_enter_quickack_mode(sk);
4197 
4198                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4199                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4200 
4201                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4202                                 end_seq = tp->rcv_nxt;
4203                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4204                 }
4205         }
4206 
4207         tcp_send_ack(sk);
4208 }
4209 
4210 /* These routines update the SACK block as out-of-order packets arrive or
4211  * in-order packets close up the sequence space.
4212  */
4213 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4214 {
4215         int this_sack;
4216         struct tcp_sack_block *sp = &tp->selective_acks[0];
4217         struct tcp_sack_block *swalk = sp + 1;
4218 
4219         /* See if the recent change to the first SACK eats into
4220          * or hits the sequence space of other SACK blocks, if so coalesce.
4221          */
4222         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4223                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4224                         int i;
4225 
4226                         /* Zap SWALK, by moving every further SACK up by one slot.
4227                          * Decrease num_sacks.
4228                          */
4229                         tp->rx_opt.num_sacks--;
4230                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4231                                 sp[i] = sp[i + 1];
4232                         continue;
4233                 }
4234                 this_sack++, swalk++;
4235         }
4236 }
4237 
4238 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4239 {
4240         struct tcp_sock *tp = tcp_sk(sk);
4241         struct tcp_sack_block *sp = &tp->selective_acks[0];
4242         int cur_sacks = tp->rx_opt.num_sacks;
4243         int this_sack;
4244 
4245         if (!cur_sacks)
4246                 goto new_sack;
4247 
4248         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4249                 if (tcp_sack_extend(sp, seq, end_seq)) {
4250                         /* Rotate this_sack to the first one. */
4251                         for (; this_sack > 0; this_sack--, sp--)
4252                                 swap(*sp, *(sp - 1));
4253                         if (cur_sacks > 1)
4254                                 tcp_sack_maybe_coalesce(tp);
4255                         return;
4256                 }
4257         }
4258 
4259         /* Could not find an adjacent existing SACK, build a new one,
4260          * put it at the front, and shift everyone else down.  We
4261          * always know there is at least one SACK present already here.
4262          *
4263          * If the sack array is full, forget about the last one.
4264          */
4265         if (this_sack >= TCP_NUM_SACKS) {
4266                 this_sack--;
4267                 tp->rx_opt.num_sacks--;
4268                 sp--;
4269         }
4270         for (; this_sack > 0; this_sack--, sp--)
4271                 *sp = *(sp - 1);
4272 
4273 new_sack:
4274         /* Build the new head SACK, and we're done. */
4275         sp->start_seq = seq;
4276         sp->end_seq = end_seq;
4277         tp->rx_opt.num_sacks++;
4278 }
4279 
4280 /* RCV.NXT advances, some SACKs should be eaten. */
4281 
4282 static void tcp_sack_remove(struct tcp_sock *tp)
4283 {
4284         struct tcp_sack_block *sp = &tp->selective_acks[0];
4285         int num_sacks = tp->rx_opt.num_sacks;
4286         int this_sack;
4287 
4288         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4289         if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4290                 tp->rx_opt.num_sacks = 0;
4291                 return;
4292         }
4293 
4294         for (this_sack = 0; this_sack < num_sacks;) {
4295                 /* Check if the start of the sack is covered by RCV.NXT. */
4296                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4297                         int i;
4298 
4299                         /* RCV.NXT must cover all the block! */
4300                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4301 
4302                         /* Zap this SACK, by moving forward any other SACKS. */
4303                         for (i = this_sack+1; i < num_sacks; i++)
4304                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4305                         num_sacks--;
4306                         continue;
4307                 }
4308                 this_sack++;
4309                 sp++;
4310         }
4311         tp->rx_opt.num_sacks = num_sacks;
4312 }
4313 
4314 /**
4315  * tcp_try_coalesce - try to merge skb to prior one
4316  * @sk: socket
4317  * @to: prior buffer
4318  * @from: buffer to add in queue
4319  * @fragstolen: pointer to boolean
4320  *
4321  * Before queueing skb @from after @to, try to merge them
4322  * to reduce overall memory use and queue lengths, if cost is small.
4323  * Packets in ofo or receive queues can stay a long time.
4324  * Better try to coalesce them right now to avoid future collapses.
4325  * Returns true if caller should free @from instead of queueing it
4326  */
4327 static bool tcp_try_coalesce(struct sock *sk,
4328                              struct sk_buff *to,
4329                              struct sk_buff *from,
4330                              bool *fragstolen)
4331 {
4332         int delta;
4333 
4334         *fragstolen = false;
4335 
4336         /* Its possible this segment overlaps with prior segment in queue */
4337         if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4338                 return false;
4339 
4340         if (!skb_try_coalesce(to, from, fragstolen, &delta))
4341                 return false;
4342 
4343         atomic_add(delta, &sk->sk_rmem_alloc);
4344         sk_mem_charge(sk, delta);
4345         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4346         TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4347         TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4348         TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4349         return true;
4350 }
4351 
4352 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4353 {
4354         sk_drops_add(sk, skb);
4355         __kfree_skb(skb);
4356 }
4357 
4358 /* This one checks to see if we can put data from the
4359  * out_of_order queue into the receive_queue.
4360  */
4361 static void tcp_ofo_queue(struct sock *sk)
4362 {
4363         struct tcp_sock *tp = tcp_sk(sk);
4364         __u32 dsack_high = tp->rcv_nxt;
4365         bool fin, fragstolen, eaten;
4366         struct sk_buff *skb, *tail;
4367         struct rb_node *p;
4368 
4369         p = rb_first(&tp->out_of_order_queue);
4370         while (p) {
4371                 skb = rb_entry(p, struct sk_buff, rbnode);
4372                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4373                         break;
4374 
4375                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4376                         __u32 dsack = dsack_high;
4377                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4378                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4379                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4380                 }
4381                 p = rb_next(p);
4382                 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4383 
4384                 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4385                         SOCK_DEBUG(sk, "ofo packet was already received\n");
4386                         tcp_drop(sk, skb);
4387                         continue;
4388                 }
4389                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4390                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4391                            TCP_SKB_CB(skb)->end_seq);
4392 
4393                 tail = skb_peek_tail(&sk->sk_receive_queue);
4394                 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4395                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4396                 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4397                 if (!eaten)
4398                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4399                 else
4400                         kfree_skb_partial(skb, fragstolen);
4401 
4402                 if (unlikely(fin)) {
4403                         tcp_fin(sk);
4404                         /* tcp_fin() purges tp->out_of_order_queue,
4405                          * so we must end this loop right now.
4406                          */
4407                         break;
4408                 }
4409         }
4410 }
4411 
4412 static bool tcp_prune_ofo_queue(struct sock *sk);
4413 static int tcp_prune_queue(struct sock *sk);
4414 
4415 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4416                                  unsigned int size)
4417 {
4418         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4419             !sk_rmem_schedule(sk, skb, size)) {
4420 
4421                 if (tcp_prune_queue(sk) < 0)
4422                         return -1;
4423 
4424                 while (!sk_rmem_schedule(sk, skb, size)) {
4425                         if (!tcp_prune_ofo_queue(sk))
4426                                 return -1;
4427                 }
4428         }
4429         return 0;
4430 }
4431 
4432 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4433 {
4434         struct tcp_sock *tp = tcp_sk(sk);
4435         struct rb_node **p, *q, *parent;
4436         struct sk_buff *skb1;
4437         u32 seq, end_seq;
4438         bool fragstolen;
4439 
4440         tcp_ecn_check_ce(tp, skb);
4441 
4442         if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4443                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4444                 tcp_drop(sk, skb);
4445                 return;
4446         }
4447 
4448         /* Disable header prediction. */
4449         tp->pred_flags = 0;
4450         inet_csk_schedule_ack(sk);
4451 
4452         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4453         seq = TCP_SKB_CB(skb)->seq;
4454         end_seq = TCP_SKB_CB(skb)->end_seq;
4455         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4456                    tp->rcv_nxt, seq, end_seq);
4457 
4458         p = &tp->out_of_order_queue.rb_node;
4459         if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4460                 /* Initial out of order segment, build 1 SACK. */
4461                 if (tcp_is_sack(tp)) {
4462                         tp->rx_opt.num_sacks = 1;
4463                         tp->selective_acks[0].start_seq = seq;
4464                         tp->selective_acks[0].end_seq = end_seq;
4465                 }
4466                 rb_link_node(&skb->rbnode, NULL, p);
4467                 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4468                 tp->ooo_last_skb = skb;
4469                 goto end;
4470         }
4471 
4472         /* In the typical case, we are adding an skb to the end of the list.
4473          * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4474          */
4475         if (tcp_try_coalesce(sk, tp->ooo_last_skb, skb, &fragstolen)) {
4476 coalesce_done:
4477                 tcp_grow_window(sk, skb);
4478                 kfree_skb_partial(skb, fragstolen);
4479                 skb = NULL;
4480                 goto add_sack;
4481         }
4482         /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4483         if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4484                 parent = &tp->ooo_last_skb->rbnode;
4485                 p = &parent->rb_right;
4486                 goto insert;
4487         }
4488 
4489         /* Find place to insert this segment. Handle overlaps on the way. */
4490         parent = NULL;
4491         while (*p) {
4492                 parent = *p;
4493                 skb1 = rb_entry(parent, struct sk_buff, rbnode);
4494                 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4495                         p = &parent->rb_left;
4496                         continue;
4497                 }
4498                 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4499                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4500                                 /* All the bits are present. Drop. */
4501                                 NET_INC_STATS(sock_net(sk),
4502                                               LINUX_MIB_TCPOFOMERGE);
4503                                 __kfree_skb(skb);
4504                                 skb = NULL;
4505                                 tcp_dsack_set(sk, seq, end_seq);
4506                                 goto add_sack;
4507                         }
4508                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4509                                 /* Partial overlap. */
4510                                 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4511                         } else {
4512                                 /* skb's seq == skb1's seq and skb covers skb1.
4513                                  * Replace skb1 with skb.
4514                                  */
4515                                 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4516                                                 &tp->out_of_order_queue);
4517                                 tcp_dsack_extend(sk,
4518                                                  TCP_SKB_CB(skb1)->seq,
4519                                                  TCP_SKB_CB(skb1)->end_seq);
4520                                 NET_INC_STATS(sock_net(sk),
4521                                               LINUX_MIB_TCPOFOMERGE);
4522                                 __kfree_skb(skb1);
4523                                 goto merge_right;
4524                         }
4525                 } else if (tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4526                         goto coalesce_done;
4527                 }
4528                 p = &parent->rb_right;
4529         }
4530 insert:
4531         /* Insert segment into RB tree. */
4532         rb_link_node(&skb->rbnode, parent, p);
4533         rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4534 
4535 merge_right:
4536         /* Remove other segments covered by skb. */
4537         while ((q = rb_next(&skb->rbnode)) != NULL) {
4538                 skb1 = rb_entry(q, struct sk_buff, rbnode);
4539 
4540                 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4541                         break;
4542                 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4543                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4544                                          end_seq);
4545                         break;
4546                 }
4547                 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4548                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4549                                  TCP_SKB_CB(skb1)->end_seq);
4550                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4551                 tcp_drop(sk, skb1);
4552         }
4553         /* If there is no skb after us, we are the last_skb ! */
4554         if (!q)
4555                 tp->ooo_last_skb = skb;
4556 
4557 add_sack:
4558         if (tcp_is_sack(tp))
4559                 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4560 end:
4561         if (skb) {
4562                 tcp_grow_window(sk, skb);
4563                 skb_set_owner_r(skb, sk);
4564         }
4565 }
4566 
4567 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4568                   bool *fragstolen)
4569 {
4570         int eaten;
4571         struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4572 
4573         __skb_pull(skb, hdrlen);
4574         eaten = (tail &&
4575                  tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4576         tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4577         if (!eaten) {
4578                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4579                 skb_set_owner_r(skb, sk);
4580         }
4581         return eaten;
4582 }
4583 
4584 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4585 {
4586         struct sk_buff *skb;
4587         int err = -ENOMEM;
4588         int data_len = 0;
4589         bool fragstolen;
4590 
4591         if (size == 0)
4592                 return 0;
4593 
4594         if (size > PAGE_SIZE) {
4595                 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4596 
4597                 data_len = npages << PAGE_SHIFT;
4598                 size = data_len + (size & ~PAGE_MASK);
4599         }
4600         skb = alloc_skb_with_frags(size - data_len, data_len,
4601                                    PAGE_ALLOC_COSTLY_ORDER,
4602                                    &err, sk->sk_allocation);
4603         if (!skb)
4604                 goto err;
4605 
4606         skb_put(skb, size - data_len);
4607         skb->data_len = data_len;
4608         skb->len = size;
4609 
4610         if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4611                 goto err_free;
4612 
4613         err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4614         if (err)
4615                 goto err_free;
4616 
4617         TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4618         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4619         TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4620 
4621         if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4622                 WARN_ON_ONCE(fragstolen); /* should not happen */
4623                 __kfree_skb(skb);
4624         }
4625         return size;
4626 
4627 err_free:
4628         kfree_skb(skb);
4629 err:
4630         return err;
4631 
4632 }
4633 
4634 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4635 {
4636         struct tcp_sock *tp = tcp_sk(sk);
4637         bool fragstolen = false;
4638         int eaten = -1;
4639 
4640         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4641                 __kfree_skb(skb);
4642                 return;
4643         }
4644         skb_dst_drop(skb);
4645         __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4646 
4647         tcp_ecn_accept_cwr(tp, skb);
4648 
4649         tp->rx_opt.dsack = 0;
4650 
4651         /*  Queue data for delivery to the user.
4652          *  Packets in sequence go to the receive queue.
4653          *  Out of sequence packets to the out_of_order_queue.
4654          */
4655         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4656                 if (tcp_receive_window(tp) == 0)
4657                         goto out_of_window;
4658 
4659                 /* Ok. In sequence. In window. */
4660                 if (tp->ucopy.task == current &&
4661                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4662                     sock_owned_by_user(sk) && !tp->urg_data) {
4663                         int chunk = min_t(unsigned int, skb->len,
4664                                           tp->ucopy.len);
4665 
4666                         __set_current_state(TASK_RUNNING);
4667 
4668                         if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4669                                 tp->ucopy.len -= chunk;
4670                                 tp->copied_seq += chunk;
4671                                 eaten = (chunk == skb->len);
4672                                 tcp_rcv_space_adjust(sk);
4673                         }
4674                 }
4675 
4676                 if (eaten <= 0) {
4677 queue_and_out:
4678                         if (eaten < 0) {
4679                                 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4680                                         sk_forced_mem_schedule(sk, skb->truesize);
4681                                 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4682                                         goto drop;
4683                         }
4684                         eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4685                 }
4686                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4687                 if (skb->len)
4688                         tcp_event_data_recv(sk, skb);
4689                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4690                         tcp_fin(sk);
4691 
4692                 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4693                         tcp_ofo_queue(sk);
4694 
4695                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4696                          * gap in queue is filled.
4697                          */
4698                         if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4699                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4700                 }
4701 
4702                 if (tp->rx_opt.num_sacks)
4703                         tcp_sack_remove(tp);
4704 
4705                 tcp_fast_path_check(sk);
4706 
4707                 if (eaten > 0)
4708                         kfree_skb_partial(skb, fragstolen);
4709                 if (!sock_flag(sk, SOCK_DEAD))
4710                         sk->sk_data_ready(sk);
4711                 return;
4712         }
4713 
4714         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4715                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4716                 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4717                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4718 
4719 out_of_window:
4720                 tcp_enter_quickack_mode(sk);
4721                 inet_csk_schedule_ack(sk);
4722 drop:
4723                 tcp_drop(sk, skb);
4724                 return;
4725         }
4726 
4727         /* Out of window. F.e. zero window probe. */
4728         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4729                 goto out_of_window;
4730 
4731         tcp_enter_quickack_mode(sk);
4732 
4733         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4734                 /* Partial packet, seq < rcv_next < end_seq */
4735                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4736                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4737                            TCP_SKB_CB(skb)->end_seq);
4738 
4739                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4740 
4741                 /* If window is closed, drop tail of packet. But after
4742                  * remembering D-SACK for its head made in previous line.
4743                  */
4744                 if (!tcp_receive_window(tp))
4745                         goto out_of_window;
4746                 goto queue_and_out;
4747         }
4748 
4749         tcp_data_queue_ofo(sk, skb);
4750 }
4751 
4752 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4753 {
4754         if (list)
4755                 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4756 
4757         return rb_entry_safe(rb_next(&skb->rbnode), struct sk_buff, rbnode);
4758 }
4759 
4760 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4761                                         struct sk_buff_head *list,
4762                                         struct rb_root *root)
4763 {
4764         struct sk_buff *next = tcp_skb_next(skb, list);
4765 
4766         if (list)
4767                 __skb_unlink(skb, list);
4768         else
4769                 rb_erase(&skb->rbnode, root);
4770 
4771         __kfree_skb(skb);
4772         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4773 
4774         return next;
4775 }
4776 
4777 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4778 static void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4779 {
4780         struct rb_node **p = &root->rb_node;
4781         struct rb_node *parent = NULL;
4782         struct sk_buff *skb1;
4783 
4784         while (*p) {
4785                 parent = *p;
4786                 skb1 = rb_entry(parent, struct sk_buff, rbnode);
4787                 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4788                         p = &parent->rb_left;
4789                 else
4790                         p = &parent->rb_right;
4791         }
4792         rb_link_node(&skb->rbnode, parent, p);
4793         rb_insert_color(&skb->rbnode, root);
4794 }
4795 
4796 /* Collapse contiguous sequence of skbs head..tail with
4797  * sequence numbers start..end.
4798  *
4799  * If tail is NULL, this means until the end of the queue.
4800  *
4801  * Segments with FIN/SYN are not collapsed (only because this
4802  * simplifies code)
4803  */
4804 static void
4805 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4806              struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4807 {
4808         struct sk_buff *skb = head, *n;
4809         struct sk_buff_head tmp;
4810         bool end_of_skbs;
4811 
4812         /* First, check that queue is collapsible and find
4813          * the point where collapsing can be useful.
4814          */
4815 restart:
4816         for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4817                 n = tcp_skb_next(skb, list);
4818 
4819                 /* No new bits? It is possible on ofo queue. */
4820                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4821                         skb = tcp_collapse_one(sk, skb, list, root);
4822                         if (!skb)
4823                                 break;
4824                         goto restart;
4825                 }
4826 
4827                 /* The first skb to collapse is:
4828                  * - not SYN/FIN and
4829                  * - bloated or contains data before "start" or
4830                  *   overlaps to the next one.
4831                  */
4832                 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4833                     (tcp_win_from_space(skb->truesize) > skb->len ||
4834                      before(TCP_SKB_CB(skb)->seq, start))) {
4835                         end_of_skbs = false;
4836                         break;
4837                 }
4838 
4839                 if (n && n != tail &&
4840                     TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4841                         end_of_skbs = false;
4842                         break;
4843                 }
4844 
4845                 /* Decided to skip this, advance start seq. */
4846                 start = TCP_SKB_CB(skb)->end_seq;
4847         }
4848         if (end_of_skbs ||
4849             (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4850                 return;
4851 
4852         __skb_queue_head_init(&tmp);
4853 
4854         while (before(start, end)) {
4855                 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4856                 struct sk_buff *nskb;
4857 
4858                 nskb = alloc_skb(copy, GFP_ATOMIC);
4859                 if (!nskb)
4860                         break;
4861 
4862                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4863                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4864                 if (list)
4865                         __skb_queue_before(list, skb, nskb);
4866                 else
4867                         __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4868                 skb_set_owner_r(nskb, sk);
4869 
4870                 /* Copy data, releasing collapsed skbs. */
4871                 while (copy > 0) {
4872                         int offset = start - TCP_SKB_CB(skb)->seq;
4873                         int size = TCP_SKB_CB(skb)->end_seq - start;
4874 
4875                         BUG_ON(offset < 0);
4876                         if (size > 0) {
4877                                 size = min(copy, size);
4878                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4879                                         BUG();
4880                                 TCP_SKB_CB(nskb)->end_seq += size;
4881                                 copy -= size;
4882                                 start += size;
4883                         }
4884                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4885                                 skb = tcp_collapse_one(sk, skb, list, root);
4886                                 if (!skb ||
4887                                     skb == tail ||
4888                                     (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4889                                         goto end;
4890                         }
4891                 }
4892         }
4893 end:
4894         skb_queue_walk_safe(&tmp, skb, n)
4895                 tcp_rbtree_insert(root, skb);
4896 }
4897 
4898 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4899  * and tcp_collapse() them until all the queue is collapsed.
4900  */
4901 static void tcp_collapse_ofo_queue(struct sock *sk)
4902 {
4903         struct tcp_sock *tp = tcp_sk(sk);
4904         struct sk_buff *skb, *head;
4905         struct rb_node *p;
4906         u32 start, end;
4907 
4908         p = rb_first(&tp->out_of_order_queue);
4909         skb = rb_entry_safe(p, struct sk_buff, rbnode);
4910 new_range:
4911         if (!skb) {
4912                 p = rb_last(&tp->out_of_order_queue);
4913                 /* Note: This is possible p is NULL here. We do not
4914                  * use rb_entry_safe(), as ooo_last_skb is valid only
4915                  * if rbtree is not empty.
4916                  */
4917                 tp->ooo_last_skb = rb_entry(p, struct sk_buff, rbnode);
4918                 return;
4919         }
4920         start = TCP_SKB_CB(skb)->seq;
4921         end = TCP_SKB_CB(skb)->end_seq;
4922 
4923         for (head = skb;;) {
4924                 skb = tcp_skb_next(skb, NULL);
4925 
4926                 /* Range is terminated when we see a gap or when
4927                  * we are at the queue end.
4928                  */
4929                 if (!skb ||
4930                     after(TCP_SKB_CB(skb)->seq, end) ||
4931                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4932                         tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4933                                      head, skb, start, end);
4934                         goto new_range;
4935                 }
4936 
4937                 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4938                         start = TCP_SKB_CB(skb)->seq;
4939                 if (after(TCP_SKB_CB(skb)->end_seq, end))
4940                         end = TCP_SKB_CB(skb)->end_seq;
4941         }
4942 }
4943 
4944 /*
4945  * Clean the out-of-order queue to make room.
4946  * We drop high sequences packets to :
4947  * 1) Let a chance for holes to be filled.
4948  * 2) not add too big latencies if thousands of packets sit there.
4949  *    (But if application shrinks SO_RCVBUF, we could still end up
4950  *     freeing whole queue here)
4951  *
4952  * Return true if queue has shrunk.
4953  */
4954 static bool tcp_prune_ofo_queue(struct sock *sk)
4955 {
4956         struct tcp_sock *tp = tcp_sk(sk);
4957         struct rb_node *node, *prev;
4958 
4959         if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4960                 return false;
4961 
4962         NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
4963         node = &tp->ooo_last_skb->rbnode;
4964         do {
4965                 prev = rb_prev(node);
4966                 rb_erase(node, &tp->out_of_order_queue);
4967                 tcp_drop(sk, rb_entry(node, struct sk_buff, rbnode));
4968                 sk_mem_reclaim(sk);
4969                 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
4970                     !tcp_under_memory_pressure(sk))
4971                         break;
4972                 node = prev;
4973         } while (node);
4974         tp->ooo_last_skb = rb_entry(prev, struct sk_buff, rbnode);
4975 
4976         /* Reset SACK state.  A conforming SACK implementation will
4977          * do the same at a timeout based retransmit.  When a connection
4978          * is in a sad state like this, we care only about integrity
4979          * of the connection not performance.
4980          */
4981         if (tp->rx_opt.sack_ok)
4982                 tcp_sack_reset(&tp->rx_opt);
4983         return true;
4984 }
4985 
4986 /* Reduce allocated memory if we can, trying to get
4987  * the socket within its memory limits again.
4988  *
4989  * Return less than zero if we should start dropping frames
4990  * until the socket owning process reads some of the data
4991  * to stabilize the situation.
4992  */
4993 static int tcp_prune_queue(struct sock *sk)
4994 {
4995         struct tcp_sock *tp = tcp_sk(sk);
4996 
4997         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4998 
4999         NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5000 
5001         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5002                 tcp_clamp_window(sk);
5003         else if (tcp_under_memory_pressure(sk))
5004                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5005 
5006         tcp_collapse_ofo_queue(sk);
5007         if (!skb_queue_empty(&sk->sk_receive_queue))
5008                 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5009                              skb_peek(&sk->sk_receive_queue),
5010                              NULL,
5011                              tp->copied_seq, tp->rcv_nxt);
5012         sk_mem_reclaim(sk);
5013 
5014         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5015                 return 0;
5016 
5017         /* Collapsing did not help, destructive actions follow.
5018          * This must not ever occur. */
5019 
5020         tcp_prune_ofo_queue(sk);
5021 
5022         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5023                 return 0;
5024 
5025         /* If we are really being abused, tell the caller to silently
5026          * drop receive data on the floor.  It will get retransmitted
5027          * and hopefully then we'll have sufficient space.
5028          */
5029         NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5030 
5031         /* Massive buffer overcommit. */
5032         tp->pred_flags = 0;
5033         return -1;
5034 }
5035 
5036 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5037 {
5038         const struct tcp_sock *tp = tcp_sk(sk);
5039 
5040         /* If the user specified a specific send buffer setting, do
5041          * not modify it.
5042          */
5043         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5044                 return false;
5045 
5046         /* If we are under global TCP memory pressure, do not expand.  */
5047         if (tcp_under_memory_pressure(sk))
5048                 return false;
5049 
5050         /* If we are under soft global TCP memory pressure, do not expand.  */
5051         if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5052                 return false;
5053 
5054         /* If we filled the congestion window, do not expand.  */
5055         if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5056                 return false;
5057 
5058         return true;
5059 }
5060 
5061 /* When incoming ACK allowed to free some skb from write_queue,
5062  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5063  * on the exit from tcp input handler.
5064  *
5065  * PROBLEM: sndbuf expansion does not work well with largesend.
5066  */
5067 static void tcp_new_space(struct sock *sk)
5068 {
5069         struct tcp_sock *tp = tcp_sk(sk);
5070 
5071         if (tcp_should_expand_sndbuf(sk)) {
5072                 tcp_sndbuf_expand(sk);
5073                 tp->snd_cwnd_stamp = tcp_time_stamp;
5074         }
5075 
5076         sk->sk_write_space(sk);
5077 }
5078 
5079 static void tcp_check_space(struct sock *sk)
5080 {
5081         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5082                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5083                 /* pairs with tcp_poll() */
5084                 smp_mb__after_atomic();
5085                 if (sk->sk_socket &&
5086                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5087                         tcp_new_space(sk);
5088         }
5089 }
5090 
5091 static inline void tcp_data_snd_check(struct sock *sk)
5092 {
5093         tcp_push_pending_frames(sk);
5094         tcp_check_space(sk);
5095 }
5096 
5097 /*
5098  * Check if sending an ack is needed.
5099  */
5100 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5101 {
5102         struct tcp_sock *tp = tcp_sk(sk);
5103 
5104             /* More than one full frame received... */
5105         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5106              /* ... and right edge of window advances far enough.
5107               * (tcp_recvmsg() will send ACK otherwise). Or...
5108               */
5109              __tcp_select_window(sk) >= tp->rcv_wnd) ||
5110             /* We ACK each frame or... */
5111             tcp_in_quickack_mode(sk) ||
5112             /* We have out of order data. */
5113             (ofo_possible && !RB_EMPTY_ROOT(&tp->out_of_order_queue))) {
5114                 /* Then ack it now */
5115                 tcp_send_ack(sk);
5116         } else {
5117                 /* Else, send delayed ack. */
5118                 tcp_send_delayed_ack(sk);
5119         }
5120 }
5121 
5122 static inline void tcp_ack_snd_check(struct sock *sk)
5123 {
5124         if (!inet_csk_ack_scheduled(sk)) {
5125                 /* We sent a data segment already. */
5126                 return;
5127         }
5128         __tcp_ack_snd_check(sk, 1);
5129 }
5130 
5131 /*
5132  *      This routine is only called when we have urgent data
5133  *      signaled. Its the 'slow' part of tcp_urg. It could be
5134  *      moved inline now as tcp_urg is only called from one
5135  *      place. We handle URGent data wrong. We have to - as
5136  *      BSD still doesn't use the correction from RFC961.
5137  *      For 1003.1g we should support a new option TCP_STDURG to permit
5138  *      either form (or just set the sysctl tcp_stdurg).
5139  */
5140 
5141 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5142 {
5143         struct tcp_sock *tp = tcp_sk(sk);
5144         u32 ptr = ntohs(th->urg_ptr);
5145 
5146         if (ptr && !sysctl_tcp_stdurg)
5147                 ptr--;
5148         ptr += ntohl(th->seq);
5149 
5150         /* Ignore urgent data that we've already seen and read. */
5151         if (after(tp->copied_seq, ptr))
5152                 return;
5153 
5154         /* Do not replay urg ptr.
5155          *
5156          * NOTE: interesting situation not covered by specs.
5157          * Misbehaving sender may send urg ptr, pointing to segment,
5158          * which we already have in ofo queue. We are not able to fetch
5159          * such data and will stay in TCP_URG_NOTYET until will be eaten
5160          * by recvmsg(). Seems, we are not obliged to handle such wicked
5161          * situations. But it is worth to think about possibility of some
5162          * DoSes using some hypothetical application level deadlock.
5163          */
5164         if (before(ptr, tp->rcv_nxt))
5165                 return;
5166 
5167         /* Do we already have a newer (or duplicate) urgent pointer? */
5168         if (tp->urg_data && !after(ptr, tp->urg_seq))
5169                 return;
5170 
5171         /* Tell the world about our new urgent pointer. */
5172         sk_send_sigurg(sk);
5173 
5174         /* We may be adding urgent data when the last byte read was
5175          * urgent. To do this requires some care. We cannot just ignore
5176          * tp->copied_seq since we would read the last urgent byte again
5177          * as data, nor can we alter copied_seq until this data arrives
5178          * or we break the semantics of SIOCATMARK (and thus sockatmark())
5179          *
5180          * NOTE. Double Dutch. Rendering to plain English: author of comment
5181          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
5182          * and expect that both A and B disappear from stream. This is _wrong_.
5183          * Though this happens in BSD with high probability, this is occasional.
5184          * Any application relying on this is buggy. Note also, that fix "works"
5185          * only in this artificial test. Insert some normal data between A and B and we will
5186          * decline of BSD again. Verdict: it is better to remove to trap
5187          * buggy users.
5188          */
5189         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5190             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5191                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5192                 tp->copied_seq++;
5193                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5194                         __skb_unlink(skb, &sk->sk_receive_queue);
5195                         __kfree_skb(skb);
5196                 }
5197         }
5198 
5199         tp->urg_data = TCP_URG_NOTYET;
5200         tp->urg_seq = ptr;
5201 
5202         /* Disable header prediction. */
5203         tp->pred_flags = 0;
5204 }
5205 
5206 /* This is the 'fast' part of urgent handling. */
5207 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5208 {
5209         struct tcp_sock *tp = tcp_sk(sk);
5210 
5211         /* Check if we get a new urgent pointer - normally not. */
5212         if (th->urg)
5213                 tcp_check_urg(sk, th);
5214 
5215         /* Do we wait for any urgent data? - normally not... */
5216         if (tp->urg_data == TCP_URG_NOTYET) {
5217                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5218                           th->syn;
5219 
5220                 /* Is the urgent pointer pointing into this packet? */
5221                 if (ptr < skb->len) {
5222                         u8 tmp;
5223                         if (skb_copy_bits(skb, ptr, &tmp, 1))
5224                                 BUG();
5225                         tp->urg_data = TCP_URG_VALID | tmp;
5226                         if (!sock_flag(sk, SOCK_DEAD))
5227                                 sk->sk_data_ready(sk);
5228                 }
5229         }
5230 }
5231 
5232 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5233 {
5234         struct tcp_sock *tp = tcp_sk(sk);
5235         int chunk = skb->len - hlen;
5236         int err;
5237 
5238         if (skb_csum_unnecessary(skb))
5239                 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5240         else
5241                 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5242 
5243         if (!err) {
5244                 tp->ucopy.len -= chunk;
5245                 tp->copied_seq += chunk;
5246                 tcp_rcv_space_adjust(sk);
5247         }
5248 
5249         return err;
5250 }
5251 
5252 /* Does PAWS and seqno based validation of an incoming segment, flags will
5253  * play significant role here.
5254  */
5255 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5256                                   const struct tcphdr *th, int syn_inerr)
5257 {
5258         struct tcp_sock *tp = tcp_sk(sk);
5259         bool rst_seq_match = false;
5260 
5261         /* RFC1323: H1. Apply PAWS check first. */
5262         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5263             tcp_paws_discard(sk, skb)) {
5264                 if (!th->rst) {
5265                         NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5266                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5267                                                   LINUX_MIB_TCPACKSKIPPEDPAWS,
5268                                                   &tp->last_oow_ack_time))
5269                                 tcp_send_dupack(sk, skb);
5270                         goto discard;
5271                 }
5272                 /* Reset is accepted even if it did not pass PAWS. */
5273         }
5274 
5275         /* Step 1: check sequence number */
5276         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5277                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5278                  * (RST) segments are validated by checking their SEQ-fields."
5279                  * And page 69: "If an incoming segment is not acceptable,
5280                  * an acknowledgment should be sent in reply (unless the RST
5281                  * bit is set, if so drop the segment and return)".
5282                  */
5283                 if (!th->rst) {
5284                         if (th->syn)
5285                                 goto syn_challenge;
5286                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5287                                                   LINUX_MIB_TCPACKSKIPPEDSEQ,
5288                                                   &tp->last_oow_ack_time))
5289                                 tcp_send_dupack(sk, skb);
5290                 }
5291                 goto discard;
5292         }
5293 
5294         /* Step 2: check RST bit */
5295         if (th->rst) {
5296                 /* RFC 5961 3.2 (extend to match against SACK too if available):
5297                  * If seq num matches RCV.NXT or the right-most SACK block,
5298                  * then
5299                  *     RESET the connection
5300                  * else
5301                  *     Send a challenge ACK
5302                  */
5303                 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5304                         rst_seq_match = true;
5305                 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5306                         struct tcp_sack_block *sp = &tp->selective_acks[0];
5307                         int max_sack = sp[0].end_seq;
5308                         int this_sack;
5309 
5310                         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5311                              ++this_sack) {
5312                                 max_sack = after(sp[this_sack].end_seq,
5313                                                  max_sack) ?
5314                                         sp[this_sack].end_seq : max_sack;
5315                         }
5316 
5317                         if (TCP_SKB_CB(skb)->seq == max_sack)
5318                                 rst_seq_match = true;
5319                 }
5320 
5321                 if (rst_seq_match)
5322                         tcp_reset(sk);
5323                 else
5324                         tcp_send_challenge_ack(sk, skb);
5325                 goto discard;
5326         }
5327 
5328         /* step 3: check security and precedence [ignored] */
5329 
5330         /* step 4: Check for a SYN
5331          * RFC 5961 4.2 : Send a challenge ack
5332          */
5333         if (th->syn) {
5334 syn_challenge:
5335                 if (syn_inerr)
5336                         TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5337                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5338                 tcp_send_challenge_ack(sk, skb);
5339                 goto discard;
5340         }
5341 
5342         return true;
5343 
5344 discard:
5345         tcp_drop(sk, skb);
5346         return false;
5347 }
5348 
5349 /*
5350  *      TCP receive function for the ESTABLISHED state.
5351  *
5352  *      It is split into a fast path and a slow path. The fast path is
5353  *      disabled when:
5354  *      - A zero window was announced from us - zero window probing
5355  *        is only handled properly in the slow path.
5356  *      - Out of order segments arrived.
5357  *      - Urgent data is expected.
5358  *      - There is no buffer space left
5359  *      - Unexpected TCP flags/window values/header lengths are received
5360  *        (detected by checking the TCP header against pred_flags)
5361  *      - Data is sent in both directions. Fast path only supports pure senders
5362  *        or pure receivers (this means either the sequence number or the ack
5363  *        value must stay constant)
5364  *      - Unexpected TCP option.
5365  *
5366  *      When these conditions are not satisfied it drops into a standard
5367  *      receive procedure patterned after RFC793 to handle all cases.
5368  *      The first three cases are guaranteed by proper pred_flags setting,
5369  *      the rest is checked inline. Fast processing is turned on in
5370  *      tcp_data_queue when everything is OK.
5371  */
5372 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5373                          const struct tcphdr *th, unsigned int len)
5374 {
5375         struct tcp_sock *tp = tcp_sk(sk);
5376 
5377         if (unlikely(!sk->sk_rx_dst))
5378                 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5379         /*
5380          *      Header prediction.
5381          *      The code loosely follows the one in the famous
5382          *      "30 instruction TCP receive" Van Jacobson mail.
5383          *
5384          *      Van's trick is to deposit buffers into socket queue
5385          *      on a device interrupt, to call tcp_recv function
5386          *      on the receive process context and checksum and copy
5387          *      the buffer to user space. smart...
5388          *
5389          *      Our current scheme is not silly either but we take the
5390          *      extra cost of the net_bh soft interrupt processing...
5391          *      We do checksum and copy also but from device to kernel.
5392          */
5393 
5394         tp->rx_opt.saw_tstamp = 0;
5395 
5396         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5397          *      if header_prediction is to be made
5398          *      'S' will always be tp->tcp_header_len >> 2
5399          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5400          *  turn it off (when there are holes in the receive
5401          *       space for instance)
5402          *      PSH flag is ignored.
5403          */
5404 
5405         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5406             TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5407             !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5408                 int tcp_header_len = tp->tcp_header_len;
5409 
5410                 /* Timestamp header prediction: tcp_header_len
5411                  * is automatically equal to th->doff*4 due to pred_flags
5412                  * match.
5413                  */
5414 
5415                 /* Check timestamp */
5416                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5417                         /* No? Slow path! */
5418                         if (!tcp_parse_aligned_timestamp(tp, th))
5419                                 goto slow_path;
5420 
5421                         /* If PAWS failed, check it more carefully in slow path */
5422                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5423                                 goto slow_path;
5424 
5425                         /* DO NOT update ts_recent here, if checksum fails
5426                          * and timestamp was corrupted part, it will result
5427                          * in a hung connection since we will drop all
5428                          * future packets due to the PAWS test.
5429                          */
5430                 }
5431 
5432                 if (len <= tcp_header_len) {
5433                         /* Bulk data transfer: sender */
5434                         if (len == tcp_header_len) {
5435                                 /* Predicted packet is in window by definition.
5436                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5437                                  * Hence, check seq<=rcv_wup reduces to:
5438                                  */
5439                                 if (tcp_header_len ==
5440                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5441                                     tp->rcv_nxt == tp->rcv_wup)
5442                                         tcp_store_ts_recent(tp);
5443 
5444                                 /* We know that such packets are checksummed
5445                                  * on entry.
5446                                  */
5447                                 tcp_ack(sk, skb, 0);
5448                                 __kfree_skb(skb);
5449                                 tcp_data_snd_check(sk);
5450                                 return;
5451                         } else { /* Header too small */
5452                                 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5453                                 goto discard;
5454                         }
5455                 } else {
5456                         int eaten = 0;
5457                         bool fragstolen = false;
5458 
5459                         if (tp->ucopy.task == current &&
5460                             tp->copied_seq == tp->rcv_nxt &&
5461                             len - tcp_header_len <= tp->ucopy.len &&
5462                             sock_owned_by_user(sk)) {
5463                                 __set_current_state(TASK_RUNNING);
5464 
5465                                 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5466                                         /* Predicted packet is in window by definition.
5467                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5468                                          * Hence, check seq<=rcv_wup reduces to:
5469                                          */
5470                                         if (tcp_header_len ==
5471                                             (sizeof(struct tcphdr) +
5472                                              TCPOLEN_TSTAMP_ALIGNED) &&
5473                                             tp->rcv_nxt == tp->rcv_wup)
5474                                                 tcp_store_ts_recent(tp);
5475 
5476                                         tcp_rcv_rtt_measure_ts(sk, skb);
5477 
5478                                         __skb_pull(skb, tcp_header_len);
5479                                         tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5480                                         NET_INC_STATS(sock_net(sk),
5481                                                         LINUX_MIB_TCPHPHITSTOUSER);
5482                                         eaten = 1;
5483                                 }
5484                         }
5485                         if (!eaten) {
5486                                 if (tcp_checksum_complete(skb))
5487                                         goto csum_error;
5488 
5489                                 if ((int)skb->truesize > sk->sk_forward_alloc)
5490                                         goto step5;
5491 
5492                                 /* Predicted packet is in window by definition.
5493                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5494                                  * Hence, check seq<=rcv_wup reduces to:
5495                                  */
5496                                 if (tcp_header_len ==
5497                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5498                                     tp->rcv_nxt == tp->rcv_wup)
5499                                         tcp_store_ts_recent(tp);
5500 
5501                                 tcp_rcv_rtt_measure_ts(sk, skb);
5502 
5503                                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5504 
5505                                 /* Bulk data transfer: receiver */
5506                                 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5507                                                       &fragstolen);
5508                         }
5509 
5510                         tcp_event_data_recv(sk, skb);
5511 
5512                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5513                                 /* Well, only one small jumplet in fast path... */
5514                                 tcp_ack(sk, skb, FLAG_DATA);
5515                                 tcp_data_snd_check(sk);
5516                                 if (!inet_csk_ack_scheduled(sk))
5517                                         goto no_ack;
5518                         }
5519 
5520                         __tcp_ack_snd_check(sk, 0);
5521 no_ack:
5522                         if (eaten)
5523                                 kfree_skb_partial(skb, fragstolen);
5524                         sk->sk_data_ready(sk);
5525                         return;
5526                 }
5527         }
5528 
5529 slow_path:
5530         if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5531                 goto csum_error;
5532 
5533         if (!th->ack && !th->rst && !th->syn)
5534                 goto discard;
5535 
5536         /*
5537          *      Standard slow path.
5538          */
5539 
5540         if (!tcp_validate_incoming(sk, skb, th, 1))
5541                 return;
5542 
5543 step5:
5544         if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5545                 goto discard;
5546 
5547         tcp_rcv_rtt_measure_ts(sk, skb);
5548 
5549         /* Process urgent data. */
5550         tcp_urg(sk, skb, th);
5551 
5552         /* step 7: process the segment text */
5553         tcp_data_queue(sk, skb);
5554 
5555         tcp_data_snd_check(sk);
5556         tcp_ack_snd_check(sk);
5557         return;
5558 
5559 csum_error:
5560         TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5561         TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5562 
5563 discard:
5564         tcp_drop(sk, skb);
5565 }
5566 EXPORT_SYMBOL(tcp_rcv_established);
5567 
5568 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5569 {
5570         struct tcp_sock *tp = tcp_sk(sk);
5571         struct inet_connection_sock *icsk = inet_csk(sk);
5572 
5573         tcp_set_state(sk, TCP_ESTABLISHED);
5574 
5575         if (skb) {
5576                 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5577                 security_inet_conn_established(sk, skb);
5578         }
5579 
5580         /* Make sure socket is routed, for correct metrics.  */
5581         icsk->icsk_af_ops->rebuild_header(sk);
5582 
5583         tcp_init_metrics(sk);
5584 
5585         tcp_init_congestion_control(sk);
5586 
5587         /* Prevent spurious tcp_cwnd_restart() on first data
5588          * packet.
5589          */
5590         tp->lsndtime = tcp_time_stamp;
5591 
5592         tcp_init_buffer_space(sk);
5593 
5594         if (sock_flag(sk, SOCK_KEEPOPEN))
5595                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5596 
5597         if (!tp->rx_opt.snd_wscale)
5598                 __tcp_fast_path_on(tp, tp->snd_wnd);
5599         else
5600                 tp->pred_flags = 0;
5601 
5602         if (!sock_flag(sk, SOCK_DEAD)) {
5603                 sk->sk_state_change(sk);
5604                 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5605         }
5606 }
5607 
5608 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5609                                     struct tcp_fastopen_cookie *cookie)
5610 {
5611         struct tcp_sock *tp = tcp_sk(sk);
5612         struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5613         u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5614         bool syn_drop = false;
5615 
5616         if (mss == tp->rx_opt.user_mss) {
5617                 struct tcp_options_received opt;
5618 
5619                 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5620                 tcp_clear_options(&opt);
5621                 opt.user_mss = opt.mss_clamp = 0;
5622                 tcp_parse_options(synack, &opt, 0, NULL);
5623                 mss = opt.mss_clamp;
5624         }
5625 
5626         if (!tp->syn_fastopen) {
5627                 /* Ignore an unsolicited cookie */
5628                 cookie->len = -1;
5629         } else if (tp->total_retrans) {
5630                 /* SYN timed out and the SYN-ACK neither has a cookie nor
5631                  * acknowledges data. Presumably the remote received only
5632                  * the retransmitted (regular) SYNs: either the original
5633                  * SYN-data or the corresponding SYN-ACK was dropped.
5634                  */
5635                 syn_drop = (cookie->len < 0 && data);
5636         } else if (cookie->len < 0 && !tp->syn_data) {
5637                 /* We requested a cookie but didn't get it. If we did not use
5638                  * the (old) exp opt format then try so next time (try_exp=1).
5639                  * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5640                  */
5641                 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5642         }
5643 
5644         tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5645 
5646         if (data) { /* Retransmit unacked data in SYN */
5647                 tcp_for_write_queue_from(data, sk) {
5648                         if (data == tcp_send_head(sk) ||
5649                             __tcp_retransmit_skb(sk, data, 1))
5650                                 break;
5651                 }
5652                 tcp_rearm_rto(sk);
5653                 NET_INC_STATS(sock_net(sk),
5654                                 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5655                 return true;
5656         }
5657         tp->syn_data_acked = tp->syn_data;
5658         if (tp->syn_data_acked)
5659                 NET_INC_STATS(sock_net(sk),
5660                                 LINUX_MIB_TCPFASTOPENACTIVE);
5661 
5662         tcp_fastopen_add_skb(sk, synack);
5663 
5664         return false;
5665 }
5666 
5667 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5668                                          const struct tcphdr *th)
5669 {
5670         struct inet_connection_sock *icsk = inet_csk(sk);
5671         struct tcp_sock *tp = tcp_sk(sk);
5672         struct tcp_fastopen_cookie foc = { .len = -1 };
5673         int saved_clamp = tp->rx_opt.mss_clamp;
5674 
5675         tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5676         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5677                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5678 
5679         if (th->ack) {
5680                 /* rfc793:
5681                  * "If the state is SYN-SENT then
5682                  *    first check the ACK bit
5683                  *      If the ACK bit is set
5684                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5685                  *        a reset (unless the RST bit is set, if so drop
5686                  *        the segment and return)"
5687                  */
5688                 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5689                     after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5690                         goto reset_and_undo;
5691 
5692                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5693                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5694                              tcp_time_stamp)) {
5695                         NET_INC_STATS(sock_net(sk),
5696                                         LINUX_MIB_PAWSACTIVEREJECTED);
5697                         goto reset_and_undo;
5698                 }
5699 
5700                 /* Now ACK is acceptable.
5701                  *
5702                  * "If the RST bit is set
5703                  *    If the ACK was acceptable then signal the user "error:
5704                  *    connection reset", drop the segment, enter CLOSED state,
5705                  *    delete TCB, and return."
5706                  */
5707 
5708                 if (th->rst) {
5709                         tcp_reset(sk);
5710                         goto discard;
5711                 }
5712 
5713                 /* rfc793:
5714                  *   "fifth, if neither of the SYN or RST bits is set then
5715                  *    drop the segment and return."
5716                  *
5717                  *    See note below!
5718                  *                                        --ANK(990513)
5719                  */
5720                 if (!th->syn)
5721                         goto discard_and_undo;
5722 
5723                 /* rfc793:
5724                  *   "If the SYN bit is on ...
5725                  *    are acceptable then ...
5726                  *    (our SYN has been ACKed), change the connection
5727                  *    state to ESTABLISHED..."
5728                  */
5729 
5730                 tcp_ecn_rcv_synack(tp, th);
5731 
5732                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5733                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5734 
5735                 /* Ok.. it's good. Set up sequence numbers and
5736                  * move to established.
5737                  */
5738                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5739                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5740 
5741                 /* RFC1323: The window in SYN & SYN/ACK segments is
5742                  * never scaled.
5743                  */
5744                 tp->snd_wnd = ntohs(th->window);
5745 
5746                 if (!tp->rx_opt.wscale_ok) {
5747                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5748                         tp->window_clamp = min(tp->window_clamp, 65535U);
5749                 }
5750 
5751                 if (tp->rx_opt.saw_tstamp) {
5752                         tp->rx_opt.tstamp_ok       = 1;
5753                         tp->tcp_header_len =
5754                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5755                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5756                         tcp_store_ts_recent(tp);
5757                 } else {
5758                         tp->tcp_header_len = sizeof(struct tcphdr);
5759                 }
5760 
5761                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5762                         tcp_enable_fack(tp);
5763 
5764                 tcp_mtup_init(sk);
5765                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5766                 tcp_initialize_rcv_mss(sk);
5767 
5768                 /* Remember, tcp_poll() does not lock socket!
5769                  * Change state from SYN-SENT only after copied_seq
5770                  * is initialized. */
5771                 tp->copied_seq = tp->rcv_nxt;
5772 
5773                 smp_mb();
5774 
5775                 tcp_finish_connect(sk, skb);
5776 
5777                 if ((tp->syn_fastopen || tp->syn_data) &&
5778                     tcp_rcv_fastopen_synack(sk, skb, &foc))
5779                         return -1;
5780 
5781                 if (sk->sk_write_pending ||
5782                     icsk->icsk_accept_queue.rskq_defer_accept ||
5783                     icsk->icsk_ack.pingpong) {
5784                         /* Save one ACK. Data will be ready after
5785                          * several ticks, if write_pending is set.
5786                          *
5787                          * It may be deleted, but with this feature tcpdumps
5788                          * look so _wonderfully_ clever, that I was not able
5789                          * to stand against the temptation 8)     --ANK
5790                          */
5791                         inet_csk_schedule_ack(sk);
5792                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5793                         tcp_enter_quickack_mode(sk);
5794                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5795                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5796 
5797 discard:
5798                         tcp_drop(sk, skb);
5799                         return 0;
5800                 } else {
5801                         tcp_send_ack(sk);
5802                 }
5803                 return -1;
5804         }
5805 
5806         /* No ACK in the segment */
5807 
5808         if (th->rst) {
5809                 /* rfc793:
5810                  * "If the RST bit is set
5811                  *
5812                  *      Otherwise (no ACK) drop the segment and return."
5813                  */
5814 
5815                 goto discard_and_undo;
5816         }
5817 
5818         /* PAWS check. */
5819         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5820             tcp_paws_reject(&tp->rx_opt, 0))
5821                 goto discard_and_undo;
5822 
5823         if (th->syn) {
5824                 /* We see SYN without ACK. It is attempt of
5825                  * simultaneous connect with crossed SYNs.
5826                  * Particularly, it can be connect to self.
5827                  */
5828                 tcp_set_state(sk, TCP_SYN_RECV);
5829 
5830                 if (tp->rx_opt.saw_tstamp) {
5831                         tp->rx_opt.tstamp_ok = 1;
5832                         tcp_store_ts_recent(tp);
5833                         tp->tcp_header_len =
5834                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5835                 } else {
5836                         tp->tcp_header_len = sizeof(struct tcphdr);
5837                 }
5838 
5839                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5840                 tp->copied_seq = tp->rcv_nxt;
5841                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5842 
5843                 /* RFC1323: The window in SYN & SYN/ACK segments is
5844                  * never scaled.
5845                  */
5846                 tp->snd_wnd    = ntohs(th->window);
5847                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5848                 tp->max_window = tp->snd_wnd;
5849 
5850                 tcp_ecn_rcv_syn(tp, th);
5851 
5852                 tcp_mtup_init(sk);
5853                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5854                 tcp_initialize_rcv_mss(sk);
5855 
5856                 tcp_send_synack(sk);
5857 #if 0
5858                 /* Note, we could accept data and URG from this segment.
5859                  * There are no obstacles to make this (except that we must
5860                  * either change tcp_recvmsg() to prevent it from returning data
5861                  * before 3WHS completes per RFC793, or employ TCP Fast Open).
5862                  *
5863                  * However, if we ignore data in ACKless segments sometimes,
5864                  * we have no reasons to accept it sometimes.
5865                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5866                  * is not flawless. So, discard packet for sanity.
5867                  * Uncomment this return to process the data.
5868                  */
5869                 return -1;
5870 #else
5871                 goto discard;
5872 #endif
5873         }
5874         /* "fifth, if neither of the SYN or RST bits is set then
5875          * drop the segment and return."
5876          */
5877 
5878 discard_and_undo:
5879         tcp_clear_options(&tp->rx_opt);
5880         tp->rx_opt.mss_clamp = saved_clamp;
5881         goto discard;
5882 
5883 reset_and_undo:
5884         tcp_clear_options(&tp->rx_opt);
5885         tp->rx_opt.mss_clamp = saved_clamp;
5886         return 1;
5887 }
5888 
5889 /*
5890  *      This function implements the receiving procedure of RFC 793 for
5891  *      all states except ESTABLISHED and TIME_WAIT.
5892  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5893  *      address independent.
5894  */
5895 
5896 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5897 {
5898         struct tcp_sock *tp = tcp_sk(sk);
5899         struct inet_connection_sock *icsk = inet_csk(sk);
5900         const struct tcphdr *th = tcp_hdr(skb);
5901         struct request_sock *req;
5902         int queued = 0;
5903         bool acceptable;
5904 
5905         switch (sk->sk_state) {
5906         case TCP_CLOSE:
5907                 goto discard;
5908 
5909         case TCP_LISTEN:
5910                 if (th->ack)
5911                         return 1;
5912 
5913                 if (th->rst)
5914                         goto discard;
5915 
5916                 if (th->syn) {
5917                         if (th->fin)
5918                                 goto discard;
5919                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5920                                 return 1;
5921 
5922                         consume_skb(skb);
5923                         return 0;
5924                 }
5925                 goto discard;
5926 
5927         case TCP_SYN_SENT:
5928                 tp->rx_opt.saw_tstamp = 0;
5929                 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5930                 if (queued >= 0)
5931                         return queued;
5932 
5933                 /* Do step6 onward by hand. */
5934                 tcp_urg(sk, skb, th);
5935                 __kfree_skb(skb);
5936                 tcp_data_snd_check(sk);
5937                 return 0;
5938         }
5939 
5940         tp->rx_opt.saw_tstamp = 0;
5941         req = tp->fastopen_rsk;
5942         if (req) {
5943                 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5944                     sk->sk_state != TCP_FIN_WAIT1);
5945 
5946                 if (!tcp_check_req(sk, skb, req, true))
5947                         goto discard;
5948         }
5949 
5950         if (!th->ack && !th->rst && !th->syn)
5951                 goto discard;
5952 
5953         if (!tcp_validate_incoming(sk, skb, th, 0))
5954                 return 0;
5955 
5956         /* step 5: check the ACK field */
5957         acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5958                                       FLAG_UPDATE_TS_RECENT) > 0;
5959 
5960         switch (sk->sk_state) {
5961         case TCP_SYN_RECV:
5962                 if (!acceptable)
5963                         return 1;
5964 
5965                 if (!tp->srtt_us)
5966                         tcp_synack_rtt_meas(sk, req);
5967 
5968                 /* Once we leave TCP_SYN_RECV, we no longer need req
5969                  * so release it.
5970                  */
5971                 if (req) {
5972                         inet_csk(sk)->icsk_retransmits = 0;
5973                         reqsk_fastopen_remove(sk, req, false);
5974                 } else {
5975                         /* Make sure socket is routed, for correct metrics. */
5976                         icsk->icsk_af_ops->rebuild_header(sk);
5977                         tcp_init_congestion_control(sk);
5978 
5979                         tcp_mtup_init(sk);
5980                         tp->copied_seq = tp->rcv_nxt;
5981                         tcp_init_buffer_space(sk);
5982                 }
5983                 smp_mb();
5984                 tcp_set_state(sk, TCP_ESTABLISHED);
5985                 sk->sk_state_change(sk);
5986 
5987                 /* Note, that this wakeup is only for marginal crossed SYN case.
5988                  * Passively open sockets are not waked up, because
5989                  * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5990                  */
5991                 if (sk->sk_socket)
5992                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5993 
5994                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5995                 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5996                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5997 
5998                 if (tp->rx_opt.tstamp_ok)
5999                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6000 
6001                 if (req) {
6002                         /* Re-arm the timer because data may have been sent out.
6003                          * This is similar to the regular data transmission case
6004                          * when new data has just been ack'ed.
6005                          *
6006                          * (TFO) - we could try to be more aggressive and
6007                          * retransmitting any data sooner based on when they
6008                          * are sent out.
6009                          */
6010                         tcp_rearm_rto(sk);
6011                 } else
6012                         tcp_init_metrics(sk);
6013 
6014                 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6015                         tcp_update_pacing_rate(sk);
6016 
6017                 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6018                 tp->lsndtime = tcp_time_stamp;
6019 
6020                 tcp_initialize_rcv_mss(sk);
6021                 tcp_fast_path_on(tp);
6022                 break;
6023 
6024         case TCP_FIN_WAIT1: {
6025                 struct dst_entry *dst;
6026                 int tmo;
6027 
6028                 /* If we enter the TCP_FIN_WAIT1 state and we are a
6029                  * Fast Open socket and this is the first acceptable
6030                  * ACK we have received, this would have acknowledged
6031                  * our SYNACK so stop the SYNACK timer.
6032                  */
6033                 if (req) {
6034                         /* Return RST if ack_seq is invalid.
6035                          * Note that RFC793 only says to generate a
6036                          * DUPACK for it but for TCP Fast Open it seems
6037                          * better to treat this case like TCP_SYN_RECV
6038                          * above.
6039                          */
6040                         if (!acceptable)
6041                                 return 1;
6042                         /* We no longer need the request sock. */
6043                         reqsk_fastopen_remove(sk, req, false);
6044                         tcp_rearm_rto(sk);
6045                 }
6046                 if (tp->snd_una != tp->write_seq)
6047                         break;
6048 
6049                 tcp_set_state(sk, TCP_FIN_WAIT2);
6050                 sk->sk_shutdown |= SEND_SHUTDOWN;
6051 
6052                 dst = __sk_dst_get(sk);
6053                 if (dst)
6054                         dst_confirm(dst);
6055 
6056                 if (!sock_flag(sk, SOCK_DEAD)) {
6057                         /* Wake up lingering close() */
6058                         sk->sk_state_change(sk);
6059                         break;
6060                 }
6061 
6062                 if (tp->linger2 < 0 ||
6063                     (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6064                      after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
6065                         tcp_done(sk);
6066                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6067                         return 1;
6068                 }
6069 
6070                 tmo = tcp_fin_time(sk);
6071                 if (tmo > TCP_TIMEWAIT_LEN) {
6072                         inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6073                 } else if (th->fin || sock_owned_by_user(sk)) {
6074                         /* Bad case. We could lose such FIN otherwise.
6075                          * It is not a big problem, but it looks confusing
6076                          * and not so rare event. We still can lose it now,
6077                          * if it spins in bh_lock_sock(), but it is really
6078                          * marginal case.
6079                          */
6080                         inet_csk_reset_keepalive_timer(sk, tmo);
6081                 } else {
6082                         tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6083                         goto discard;
6084                 }
6085                 break;
6086         }
6087 
6088         case TCP_CLOSING:
6089                 if (tp->snd_una == tp->write_seq) {
6090                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6091                         goto discard;
6092                 }
6093                 break;
6094 
6095         case TCP_LAST_ACK:
6096                 if (tp->snd_una == tp->write_seq) {
6097                         tcp_update_metrics(sk);
6098                         tcp_done(sk);
6099                         goto discard;
6100                 }
6101                 break;
6102         }
6103 
6104         /* step 6: check the URG bit */
6105         tcp_urg(sk, skb, th);
6106 
6107         /* step 7: process the segment text */
6108         switch (sk->sk_state) {
6109         case TCP_CLOSE_WAIT:
6110         case TCP_CLOSING:
6111         case TCP_LAST_ACK:
6112                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6113                         break;
6114         case TCP_FIN_WAIT1:
6115         case TCP_FIN_WAIT2:
6116                 /* RFC 793 says to queue data in these states,
6117                  * RFC 1122 says we MUST send a reset.
6118                  * BSD 4.4 also does reset.
6119                  */
6120                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6121                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6122                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6123                                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6124                                 tcp_reset(sk);
6125                                 return 1;
6126                         }
6127                 }
6128                 /* Fall through */
6129         case TCP_ESTABLISHED:
6130                 tcp_data_queue(sk, skb);
6131                 queued = 1;
6132                 break;
6133         }
6134 
6135         /* tcp_data could move socket to TIME-WAIT */
6136         if (sk->sk_state != TCP_CLOSE) {
6137                 tcp_data_snd_check(sk);
6138                 tcp_ack_snd_check(sk);
6139         }
6140 
6141         if (!queued) {
6142 discard:
6143                 tcp_drop(sk, skb);
6144         }
6145         return 0;
6146 }
6147 EXPORT_SYMBOL(tcp_rcv_state_process);
6148 
6149 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6150 {
6151         struct inet_request_sock *ireq = inet_rsk(req);
6152 
6153         if (family == AF_INET)
6154                 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6155                                     &ireq->ir_rmt_addr, port);
6156 #if IS_ENABLED(CONFIG_IPV6)
6157         else if (family == AF_INET6)
6158                 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6159                                     &ireq->ir_v6_rmt_addr, port);
6160 #endif
6161 }
6162 
6163 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6164  *
6165  * If we receive a SYN packet with these bits set, it means a
6166  * network is playing bad games with TOS bits. In order to
6167  * avoid possible false congestion notifications, we disable
6168  * TCP ECN negotiation.
6169  *
6170  * Exception: tcp_ca wants ECN. This is required for DCTCP
6171  * congestion control: Linux DCTCP asserts ECT on all packets,
6172  * including SYN, which is most optimal solution; however,
6173  * others, such as FreeBSD do not.
6174  */
6175 static void tcp_ecn_create_request(struct request_sock *req,
6176                                    const struct sk_buff *skb,
6177                                    const struct sock *listen_sk,
6178                                    const struct dst_entry *dst)
6179 {
6180         const struct tcphdr *th = tcp_hdr(skb);
6181         const struct net *net = sock_net(listen_sk);
6182         bool th_ecn = th->ece && th->cwr;
6183         bool ect, ecn_ok;
6184         u32 ecn_ok_dst;
6185 
6186         if (!th_ecn)
6187                 return;
6188 
6189         ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6190         ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6191         ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6192 
6193         if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6194             (ecn_ok_dst & DST_FEATURE_ECN_CA))
6195                 inet_rsk(req)->ecn_ok = 1;
6196 }
6197 
6198 static void tcp_openreq_init(struct request_sock *req,
6199                              const struct tcp_options_received *rx_opt,
6200                              struct sk_buff *skb, const struct sock *sk)
6201 {
6202         struct inet_request_sock *ireq = inet_rsk(req);
6203 
6204         req->rsk_rcv_wnd = 0;           /* So that tcp_send_synack() knows! */
6205         req->cookie_ts = 0;
6206         tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6207         tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6208         skb_mstamp_get(&tcp_rsk(req)->snt_synack);
6209         tcp_rsk(req)->last_oow_ack_time = 0;
6210         req->mss = rx_opt->mss_clamp;
6211         req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6212         ireq->tstamp_ok = rx_opt->tstamp_ok;
6213         ireq->sack_ok = rx_opt->sack_ok;
6214         ireq->snd_wscale = rx_opt->snd_wscale;
6215         ireq->wscale_ok = rx_opt->wscale_ok;
6216         ireq->acked = 0;
6217         ireq->ecn_ok = 0;
6218         ireq->ir_rmt_port = tcp_hdr(skb)->source;
6219         ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6220         ireq->ir_mark = inet_request_mark(sk, skb);
6221 }
6222 
6223 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6224                                       struct sock *sk_listener,
6225                                       bool attach_listener)
6226 {
6227         struct request_sock *req = reqsk_alloc(ops, sk_listener,
6228                                                attach_listener);
6229 
6230         if (req) {
6231                 struct inet_request_sock *ireq = inet_rsk(req);
6232 
6233                 kmemcheck_annotate_bitfield(ireq, flags);
6234                 ireq->opt = NULL;
6235 #if IS_ENABLED(CONFIG_IPV6)
6236                 ireq->pktopts = NULL;
6237 #endif
6238                 atomic64_set(&ireq->ir_cookie, 0);
6239                 ireq->ireq_state = TCP_NEW_SYN_RECV;
6240                 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6241                 ireq->ireq_family = sk_listener->sk_family;
6242         }
6243 
6244         return req;
6245 }
6246 EXPORT_SYMBOL(inet_reqsk_alloc);
6247 
6248 /*
6249  * Return true if a syncookie should be sent
6250  */
6251 static bool tcp_syn_flood_action(const struct sock *sk,
6252                                  const struct sk_buff *skb,
6253                                  const char *proto)
6254 {
6255         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6256         const char *msg = "Dropping request";
6257         bool want_cookie = false;
6258         struct net *net = sock_net(sk);
6259 
6260 #ifdef CONFIG_SYN_COOKIES
6261         if (net->ipv4.sysctl_tcp_syncookies) {
6262                 msg = "Sending cookies";
6263                 want_cookie = true;
6264                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6265         } else
6266 #endif
6267                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6268 
6269         if (!queue->synflood_warned &&
6270             net->ipv4.sysctl_tcp_syncookies != 2 &&
6271             xchg(&queue->synflood_warned, 1) == 0)
6272                 pr_info("%s: Possible SYN flooding on port %d. %s.  Check SNMP counters.\n",
6273                         proto, ntohs(tcp_hdr(skb)->dest), msg);
6274 
6275         return want_cookie;
6276 }
6277 
6278 static void tcp_reqsk_record_syn(const struct sock *sk,
6279                                  struct request_sock *req,
6280                                  const struct sk_buff *skb)
6281 {
6282         if (tcp_sk(sk)->save_syn) {
6283                 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6284                 u32 *copy;
6285 
6286                 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6287                 if (copy) {
6288                         copy[0] = len;
6289                         memcpy(&copy[1], skb_network_header(skb), len);
6290                         req->saved_syn = copy;
6291                 }
6292         }
6293 }
6294 
6295 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6296                      const struct tcp_request_sock_ops *af_ops,
6297                      struct sock *sk, struct sk_buff *skb)
6298 {
6299         struct tcp_fastopen_cookie foc = { .len = -1 };
6300         __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6301         struct tcp_options_received tmp_opt;
6302         struct tcp_sock *tp = tcp_sk(sk);
6303         struct net *net = sock_net(sk);
6304         struct sock *fastopen_sk = NULL;
6305         struct dst_entry *dst = NULL;
6306         struct request_sock *req;
6307         bool want_cookie = false;
6308         struct flowi fl;
6309 
6310         /* TW buckets are converted to open requests without
6311          * limitations, they conserve resources and peer is
6312          * evidently real one.
6313          */
6314         if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6315              inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6316                 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6317                 if (!want_cookie)
6318                         goto drop;
6319         }
6320 
6321 
6322         /* Accept backlog is full. If we have already queued enough
6323          * of warm entries in syn queue, drop request. It is better than
6324          * clogging syn queue with openreqs with exponentially increasing
6325          * timeout.
6326          */
6327         if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
6328                 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6329                 goto drop;
6330         }
6331 
6332         req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6333         if (!req)
6334                 goto drop;
6335 
6336         tcp_rsk(req)->af_specific = af_ops;
6337 
6338         tcp_clear_options(&tmp_opt);
6339         tmp_opt.mss_clamp = af_ops->mss_clamp;
6340         tmp_opt.user_mss  = tp->rx_opt.user_mss;
6341         tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6342 
6343         if (want_cookie && !tmp_opt.saw_tstamp)
6344                 tcp_clear_options(&tmp_opt);
6345 
6346         tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6347         tcp_openreq_init(req, &tmp_opt, skb, sk);
6348         inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6349 
6350         /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6351         inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6352 
6353         af_ops->init_req(req, sk, skb);
6354 
6355         if (security_inet_conn_request(sk, skb, req))
6356                 goto drop_and_free;
6357 
6358         if (!want_cookie && !isn) {
6359                 /* VJ's idea. We save last timestamp seen
6360                  * from the destination in peer table, when entering
6361                  * state TIME-WAIT, and check against it before
6362                  * accepting new connection request.
6363                  *
6364                  * If "isn" is not zero, this request hit alive
6365                  * timewait bucket, so that all the necessary checks
6366                  * are made in the function processing timewait state.
6367                  */
6368                 if (tcp_death_row.sysctl_tw_recycle) {
6369                         bool strict;
6370 
6371                         dst = af_ops->route_req(sk, &fl, req, &strict);
6372 
6373                         if (dst && strict &&
6374                             !tcp_peer_is_proven(req, dst, true,
6375                                                 tmp_opt.saw_tstamp)) {
6376                                 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
6377                                 goto drop_and_release;
6378                         }
6379                 }
6380                 /* Kill the following clause, if you dislike this way. */
6381                 else if (!net->ipv4.sysctl_tcp_syncookies &&
6382                          (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6383                           (sysctl_max_syn_backlog >> 2)) &&
6384                          !tcp_peer_is_proven(req, dst, false,
6385                                              tmp_opt.saw_tstamp)) {
6386                         /* Without syncookies last quarter of
6387                          * backlog is filled with destinations,
6388                          * proven to be alive.
6389                          * It means that we continue to communicate
6390                          * to destinations, already remembered
6391                          * to the moment of synflood.
6392                          */
6393                         pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6394                                     rsk_ops->family);
6395                         goto drop_and_release;
6396                 }
6397 
6398                 isn = af_ops->init_seq(skb);
6399         }
6400         if (!dst) {
6401                 dst = af_ops->route_req(sk, &fl, req, NULL);
6402                 if (!dst)
6403                         goto drop_and_free;
6404         }
6405 
6406         tcp_ecn_create_request(req, skb, sk, dst);
6407 
6408         if (want_cookie) {
6409                 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6410                 req->cookie_ts = tmp_opt.tstamp_ok;
6411                 if (!tmp_opt.tstamp_ok)
6412                         inet_rsk(req)->ecn_ok = 0;
6413         }
6414 
6415         tcp_rsk(req)->snt_isn = isn;
6416         tcp_rsk(req)->txhash = net_tx_rndhash();
6417         tcp_openreq_init_rwin(req, sk, dst);
6418         if (!want_cookie) {
6419                 tcp_reqsk_record_syn(sk, req, skb);
6420                 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6421         }
6422         if (fastopen_sk) {
6423                 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6424                                     &foc, TCP_SYNACK_FASTOPEN);
6425                 /* Add the child socket directly into the accept queue */
6426                 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6427                 sk->sk_data_ready(sk);
6428                 bh_unlock_sock(fastopen_sk);
6429                 sock_put(fastopen_sk);
6430         } else {
6431                 tcp_rsk(req)->tfo_listener = false;
6432                 if (!want_cookie)
6433                         inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6434                 af_ops->send_synack(sk, dst, &fl, req, &foc,
6435                                     !want_cookie ? TCP_SYNACK_NORMAL :
6436                                                    TCP_SYNACK_COOKIE);
6437                 if (want_cookie) {
6438                         reqsk_free(req);
6439                         return 0;
6440                 }
6441         }
6442         reqsk_put(req);
6443         return 0;
6444 
6445 drop_and_release:
6446         dst_release(dst);
6447 drop_and_free:
6448         reqsk_free(req);
6449 drop:
6450         tcp_listendrop(sk);
6451         return 0;
6452 }
6453 EXPORT_SYMBOL(tcp_conn_request);
6454 

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