Version:  2.0.40 2.2.26 2.4.37 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 4.10

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 EXPORT_SYMBOL(sysctl_tcp_timestamps);
 89 
 90 /* rfc5961 challenge ack rate limiting */
 91 int sysctl_tcp_challenge_ack_limit = 1000;
 92 
 93 int sysctl_tcp_stdurg __read_mostly;
 94 int sysctl_tcp_rfc1337 __read_mostly;
 95 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
 96 int sysctl_tcp_frto __read_mostly = 2;
 97 int sysctl_tcp_min_rtt_wlen __read_mostly = 300;
 98 
 99 int sysctl_tcp_thin_dupack __read_mostly;
100 
101 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
102 int sysctl_tcp_early_retrans __read_mostly = 3;
103 int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
104 
105 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
106 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
107 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
108 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
109 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
110 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
111 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
112 #define FLAG_LOST_RETRANS       0x80 /* This ACK marks some retransmission lost */
113 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
114 #define FLAG_ORIG_SACK_ACKED    0x200 /* Never retransmitted data are (s)acked  */
115 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
116 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
117 #define FLAG_SACK_RENEGING      0x2000 /* snd_una advanced to a sacked seq */
118 #define FLAG_UPDATE_TS_RECENT   0x4000 /* tcp_replace_ts_recent() */
119 
120 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
121 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
122 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
123 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
124 
125 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
126 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
127 
128 #define REXMIT_NONE     0 /* no loss recovery to do */
129 #define REXMIT_LOST     1 /* retransmit packets marked lost */
130 #define REXMIT_NEW      2 /* FRTO-style transmit of unsent/new packets */
131 
132 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb)
133 {
134         static bool __once __read_mostly;
135 
136         if (!__once) {
137                 struct net_device *dev;
138 
139                 __once = true;
140 
141                 rcu_read_lock();
142                 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
143                 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
144                         dev ? dev->name : "Unknown driver");
145                 rcu_read_unlock();
146         }
147 }
148 
149 /* Adapt the MSS value used to make delayed ack decision to the
150  * real world.
151  */
152 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
153 {
154         struct inet_connection_sock *icsk = inet_csk(sk);
155         const unsigned int lss = icsk->icsk_ack.last_seg_size;
156         unsigned int len;
157 
158         icsk->icsk_ack.last_seg_size = 0;
159 
160         /* skb->len may jitter because of SACKs, even if peer
161          * sends good full-sized frames.
162          */
163         len = skb_shinfo(skb)->gso_size ? : skb->len;
164         if (len >= icsk->icsk_ack.rcv_mss) {
165                 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
166                                                tcp_sk(sk)->advmss);
167                 if (unlikely(icsk->icsk_ack.rcv_mss != len))
168                         tcp_gro_dev_warn(sk, skb);
169         } else {
170                 /* Otherwise, we make more careful check taking into account,
171                  * that SACKs block is variable.
172                  *
173                  * "len" is invariant segment length, including TCP header.
174                  */
175                 len += skb->data - skb_transport_header(skb);
176                 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
177                     /* If PSH is not set, packet should be
178                      * full sized, provided peer TCP is not badly broken.
179                      * This observation (if it is correct 8)) allows
180                      * to handle super-low mtu links fairly.
181                      */
182                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
183                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
184                         /* Subtract also invariant (if peer is RFC compliant),
185                          * tcp header plus fixed timestamp option length.
186                          * Resulting "len" is MSS free of SACK jitter.
187                          */
188                         len -= tcp_sk(sk)->tcp_header_len;
189                         icsk->icsk_ack.last_seg_size = len;
190                         if (len == lss) {
191                                 icsk->icsk_ack.rcv_mss = len;
192                                 return;
193                         }
194                 }
195                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
196                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
197                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
198         }
199 }
200 
201 static void tcp_incr_quickack(struct sock *sk)
202 {
203         struct inet_connection_sock *icsk = inet_csk(sk);
204         unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
205 
206         if (quickacks == 0)
207                 quickacks = 2;
208         if (quickacks > icsk->icsk_ack.quick)
209                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
210 }
211 
212 static void tcp_enter_quickack_mode(struct sock *sk)
213 {
214         struct inet_connection_sock *icsk = inet_csk(sk);
215         tcp_incr_quickack(sk);
216         icsk->icsk_ack.pingpong = 0;
217         icsk->icsk_ack.ato = TCP_ATO_MIN;
218 }
219 
220 /* Send ACKs quickly, if "quick" count is not exhausted
221  * and the session is not interactive.
222  */
223 
224 static bool tcp_in_quickack_mode(struct sock *sk)
225 {
226         const struct inet_connection_sock *icsk = inet_csk(sk);
227         const struct dst_entry *dst = __sk_dst_get(sk);
228 
229         return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
230                 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
231 }
232 
233 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
234 {
235         if (tp->ecn_flags & TCP_ECN_OK)
236                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
237 }
238 
239 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
240 {
241         if (tcp_hdr(skb)->cwr)
242                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
243 }
244 
245 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
246 {
247         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
248 }
249 
250 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
251 {
252         switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
253         case INET_ECN_NOT_ECT:
254                 /* Funny extension: if ECT is not set on a segment,
255                  * and we already seen ECT on a previous segment,
256                  * it is probably a retransmit.
257                  */
258                 if (tp->ecn_flags & TCP_ECN_SEEN)
259                         tcp_enter_quickack_mode((struct sock *)tp);
260                 break;
261         case INET_ECN_CE:
262                 if (tcp_ca_needs_ecn((struct sock *)tp))
263                         tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
264 
265                 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
266                         /* Better not delay acks, sender can have a very low cwnd */
267                         tcp_enter_quickack_mode((struct sock *)tp);
268                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
269                 }
270                 tp->ecn_flags |= TCP_ECN_SEEN;
271                 break;
272         default:
273                 if (tcp_ca_needs_ecn((struct sock *)tp))
274                         tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
275                 tp->ecn_flags |= TCP_ECN_SEEN;
276                 break;
277         }
278 }
279 
280 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
281 {
282         if (tp->ecn_flags & TCP_ECN_OK)
283                 __tcp_ecn_check_ce(tp, skb);
284 }
285 
286 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
287 {
288         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
289                 tp->ecn_flags &= ~TCP_ECN_OK;
290 }
291 
292 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
293 {
294         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
295                 tp->ecn_flags &= ~TCP_ECN_OK;
296 }
297 
298 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
299 {
300         if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
301                 return true;
302         return false;
303 }
304 
305 /* Buffer size and advertised window tuning.
306  *
307  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
308  */
309 
310 static void tcp_sndbuf_expand(struct sock *sk)
311 {
312         const struct tcp_sock *tp = tcp_sk(sk);
313         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
314         int sndmem, per_mss;
315         u32 nr_segs;
316 
317         /* Worst case is non GSO/TSO : each frame consumes one skb
318          * and skb->head is kmalloced using power of two area of memory
319          */
320         per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
321                   MAX_TCP_HEADER +
322                   SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
323 
324         per_mss = roundup_pow_of_two(per_mss) +
325                   SKB_DATA_ALIGN(sizeof(struct sk_buff));
326 
327         nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
328         nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
329 
330         /* Fast Recovery (RFC 5681 3.2) :
331          * Cubic needs 1.7 factor, rounded to 2 to include
332          * extra cushion (application might react slowly to POLLOUT)
333          */
334         sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
335         sndmem *= nr_segs * per_mss;
336 
337         if (sk->sk_sndbuf < sndmem)
338                 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
339 }
340 
341 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
342  *
343  * All tcp_full_space() is split to two parts: "network" buffer, allocated
344  * forward and advertised in receiver window (tp->rcv_wnd) and
345  * "application buffer", required to isolate scheduling/application
346  * latencies from network.
347  * window_clamp is maximal advertised window. It can be less than
348  * tcp_full_space(), in this case tcp_full_space() - window_clamp
349  * is reserved for "application" buffer. The less window_clamp is
350  * the smoother our behaviour from viewpoint of network, but the lower
351  * throughput and the higher sensitivity of the connection to losses. 8)
352  *
353  * rcv_ssthresh is more strict window_clamp used at "slow start"
354  * phase to predict further behaviour of this connection.
355  * It is used for two goals:
356  * - to enforce header prediction at sender, even when application
357  *   requires some significant "application buffer". It is check #1.
358  * - to prevent pruning of receive queue because of misprediction
359  *   of receiver window. Check #2.
360  *
361  * The scheme does not work when sender sends good segments opening
362  * window and then starts to feed us spaghetti. But it should work
363  * in common situations. Otherwise, we have to rely on queue collapsing.
364  */
365 
366 /* Slow part of check#2. */
367 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
368 {
369         struct tcp_sock *tp = tcp_sk(sk);
370         /* Optimize this! */
371         int truesize = tcp_win_from_space(skb->truesize) >> 1;
372         int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
373 
374         while (tp->rcv_ssthresh <= window) {
375                 if (truesize <= skb->len)
376                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
377 
378                 truesize >>= 1;
379                 window >>= 1;
380         }
381         return 0;
382 }
383 
384 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
385 {
386         struct tcp_sock *tp = tcp_sk(sk);
387 
388         /* Check #1 */
389         if (tp->rcv_ssthresh < tp->window_clamp &&
390             (int)tp->rcv_ssthresh < tcp_space(sk) &&
391             !tcp_under_memory_pressure(sk)) {
392                 int incr;
393 
394                 /* Check #2. Increase window, if skb with such overhead
395                  * will fit to rcvbuf in future.
396                  */
397                 if (tcp_win_from_space(skb->truesize) <= skb->len)
398                         incr = 2 * tp->advmss;
399                 else
400                         incr = __tcp_grow_window(sk, skb);
401 
402                 if (incr) {
403                         incr = max_t(int, incr, 2 * skb->len);
404                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
405                                                tp->window_clamp);
406                         inet_csk(sk)->icsk_ack.quick |= 1;
407                 }
408         }
409 }
410 
411 /* 3. Tuning rcvbuf, when connection enters established state. */
412 static void tcp_fixup_rcvbuf(struct sock *sk)
413 {
414         u32 mss = tcp_sk(sk)->advmss;
415         int rcvmem;
416 
417         rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
418                  tcp_default_init_rwnd(mss);
419 
420         /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
421          * Allow enough cushion so that sender is not limited by our window
422          */
423         if (sysctl_tcp_moderate_rcvbuf)
424                 rcvmem <<= 2;
425 
426         if (sk->sk_rcvbuf < rcvmem)
427                 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
428 }
429 
430 /* 4. Try to fixup all. It is made immediately after connection enters
431  *    established state.
432  */
433 void tcp_init_buffer_space(struct sock *sk)
434 {
435         struct tcp_sock *tp = tcp_sk(sk);
436         int maxwin;
437 
438         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
439                 tcp_fixup_rcvbuf(sk);
440         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
441                 tcp_sndbuf_expand(sk);
442 
443         tp->rcvq_space.space = tp->rcv_wnd;
444         tp->rcvq_space.time = tcp_time_stamp;
445         tp->rcvq_space.seq = tp->copied_seq;
446 
447         maxwin = tcp_full_space(sk);
448 
449         if (tp->window_clamp >= maxwin) {
450                 tp->window_clamp = maxwin;
451 
452                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
453                         tp->window_clamp = max(maxwin -
454                                                (maxwin >> sysctl_tcp_app_win),
455                                                4 * tp->advmss);
456         }
457 
458         /* Force reservation of one segment. */
459         if (sysctl_tcp_app_win &&
460             tp->window_clamp > 2 * tp->advmss &&
461             tp->window_clamp + tp->advmss > maxwin)
462                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
463 
464         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
465         tp->snd_cwnd_stamp = tcp_time_stamp;
466 }
467 
468 /* 5. Recalculate window clamp after socket hit its memory bounds. */
469 static void tcp_clamp_window(struct sock *sk)
470 {
471         struct tcp_sock *tp = tcp_sk(sk);
472         struct inet_connection_sock *icsk = inet_csk(sk);
473 
474         icsk->icsk_ack.quick = 0;
475 
476         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
477             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
478             !tcp_under_memory_pressure(sk) &&
479             sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
480                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
481                                     sysctl_tcp_rmem[2]);
482         }
483         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
484                 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
485 }
486 
487 /* Initialize RCV_MSS value.
488  * RCV_MSS is an our guess about MSS used by the peer.
489  * We haven't any direct information about the MSS.
490  * It's better to underestimate the RCV_MSS rather than overestimate.
491  * Overestimations make us ACKing less frequently than needed.
492  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
493  */
494 void tcp_initialize_rcv_mss(struct sock *sk)
495 {
496         const struct tcp_sock *tp = tcp_sk(sk);
497         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
498 
499         hint = min(hint, tp->rcv_wnd / 2);
500         hint = min(hint, TCP_MSS_DEFAULT);
501         hint = max(hint, TCP_MIN_MSS);
502 
503         inet_csk(sk)->icsk_ack.rcv_mss = hint;
504 }
505 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
506 
507 /* Receiver "autotuning" code.
508  *
509  * The algorithm for RTT estimation w/o timestamps is based on
510  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
511  * <http://public.lanl.gov/radiant/pubs.html#DRS>
512  *
513  * More detail on this code can be found at
514  * <http://staff.psc.edu/jheffner/>,
515  * though this reference is out of date.  A new paper
516  * is pending.
517  */
518 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
519 {
520         u32 new_sample = tp->rcv_rtt_est.rtt;
521         long m = sample;
522 
523         if (m == 0)
524                 m = 1;
525 
526         if (new_sample != 0) {
527                 /* If we sample in larger samples in the non-timestamp
528                  * case, we could grossly overestimate the RTT especially
529                  * with chatty applications or bulk transfer apps which
530                  * are stalled on filesystem I/O.
531                  *
532                  * Also, since we are only going for a minimum in the
533                  * non-timestamp case, we do not smooth things out
534                  * else with timestamps disabled convergence takes too
535                  * long.
536                  */
537                 if (!win_dep) {
538                         m -= (new_sample >> 3);
539                         new_sample += m;
540                 } else {
541                         m <<= 3;
542                         if (m < new_sample)
543                                 new_sample = m;
544                 }
545         } else {
546                 /* No previous measure. */
547                 new_sample = m << 3;
548         }
549 
550         if (tp->rcv_rtt_est.rtt != new_sample)
551                 tp->rcv_rtt_est.rtt = new_sample;
552 }
553 
554 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
555 {
556         if (tp->rcv_rtt_est.time == 0)
557                 goto new_measure;
558         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
559                 return;
560         tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
561 
562 new_measure:
563         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
564         tp->rcv_rtt_est.time = tcp_time_stamp;
565 }
566 
567 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
568                                           const struct sk_buff *skb)
569 {
570         struct tcp_sock *tp = tcp_sk(sk);
571         if (tp->rx_opt.rcv_tsecr &&
572             (TCP_SKB_CB(skb)->end_seq -
573              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
574                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
575 }
576 
577 /*
578  * This function should be called every time data is copied to user space.
579  * It calculates the appropriate TCP receive buffer space.
580  */
581 void tcp_rcv_space_adjust(struct sock *sk)
582 {
583         struct tcp_sock *tp = tcp_sk(sk);
584         int time;
585         int copied;
586 
587         time = tcp_time_stamp - tp->rcvq_space.time;
588         if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
589                 return;
590 
591         /* Number of bytes copied to user in last RTT */
592         copied = tp->copied_seq - tp->rcvq_space.seq;
593         if (copied <= tp->rcvq_space.space)
594                 goto new_measure;
595 
596         /* A bit of theory :
597          * copied = bytes received in previous RTT, our base window
598          * To cope with packet losses, we need a 2x factor
599          * To cope with slow start, and sender growing its cwin by 100 %
600          * every RTT, we need a 4x factor, because the ACK we are sending
601          * now is for the next RTT, not the current one :
602          * <prev RTT . ><current RTT .. ><next RTT .... >
603          */
604 
605         if (sysctl_tcp_moderate_rcvbuf &&
606             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
607                 int rcvwin, rcvmem, rcvbuf;
608 
609                 /* minimal window to cope with packet losses, assuming
610                  * steady state. Add some cushion because of small variations.
611                  */
612                 rcvwin = (copied << 1) + 16 * tp->advmss;
613 
614                 /* If rate increased by 25%,
615                  *      assume slow start, rcvwin = 3 * copied
616                  * If rate increased by 50%,
617                  *      assume sender can use 2x growth, rcvwin = 4 * copied
618                  */
619                 if (copied >=
620                     tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
621                         if (copied >=
622                             tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
623                                 rcvwin <<= 1;
624                         else
625                                 rcvwin += (rcvwin >> 1);
626                 }
627 
628                 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
629                 while (tcp_win_from_space(rcvmem) < tp->advmss)
630                         rcvmem += 128;
631 
632                 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
633                 if (rcvbuf > sk->sk_rcvbuf) {
634                         sk->sk_rcvbuf = rcvbuf;
635 
636                         /* Make the window clamp follow along.  */
637                         tp->window_clamp = rcvwin;
638                 }
639         }
640         tp->rcvq_space.space = copied;
641 
642 new_measure:
643         tp->rcvq_space.seq = tp->copied_seq;
644         tp->rcvq_space.time = tcp_time_stamp;
645 }
646 
647 /* There is something which you must keep in mind when you analyze the
648  * behavior of the tp->ato delayed ack timeout interval.  When a
649  * connection starts up, we want to ack as quickly as possible.  The
650  * problem is that "good" TCP's do slow start at the beginning of data
651  * transmission.  The means that until we send the first few ACK's the
652  * sender will sit on his end and only queue most of his data, because
653  * he can only send snd_cwnd unacked packets at any given time.  For
654  * each ACK we send, he increments snd_cwnd and transmits more of his
655  * queue.  -DaveM
656  */
657 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
658 {
659         struct tcp_sock *tp = tcp_sk(sk);
660         struct inet_connection_sock *icsk = inet_csk(sk);
661         u32 now;
662 
663         inet_csk_schedule_ack(sk);
664 
665         tcp_measure_rcv_mss(sk, skb);
666 
667         tcp_rcv_rtt_measure(tp);
668 
669         now = tcp_time_stamp;
670 
671         if (!icsk->icsk_ack.ato) {
672                 /* The _first_ data packet received, initialize
673                  * delayed ACK engine.
674                  */
675                 tcp_incr_quickack(sk);
676                 icsk->icsk_ack.ato = TCP_ATO_MIN;
677         } else {
678                 int m = now - icsk->icsk_ack.lrcvtime;
679 
680                 if (m <= TCP_ATO_MIN / 2) {
681                         /* The fastest case is the first. */
682                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
683                 } else if (m < icsk->icsk_ack.ato) {
684                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
685                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
686                                 icsk->icsk_ack.ato = icsk->icsk_rto;
687                 } else if (m > icsk->icsk_rto) {
688                         /* Too long gap. Apparently sender failed to
689                          * restart window, so that we send ACKs quickly.
690                          */
691                         tcp_incr_quickack(sk);
692                         sk_mem_reclaim(sk);
693                 }
694         }
695         icsk->icsk_ack.lrcvtime = now;
696 
697         tcp_ecn_check_ce(tp, skb);
698 
699         if (skb->len >= 128)
700                 tcp_grow_window(sk, skb);
701 }
702 
703 /* Called to compute a smoothed rtt estimate. The data fed to this
704  * routine either comes from timestamps, or from segments that were
705  * known _not_ to have been retransmitted [see Karn/Partridge
706  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
707  * piece by Van Jacobson.
708  * NOTE: the next three routines used to be one big routine.
709  * To save cycles in the RFC 1323 implementation it was better to break
710  * it up into three procedures. -- erics
711  */
712 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
713 {
714         struct tcp_sock *tp = tcp_sk(sk);
715         long m = mrtt_us; /* RTT */
716         u32 srtt = tp->srtt_us;
717 
718         /*      The following amusing code comes from Jacobson's
719          *      article in SIGCOMM '88.  Note that rtt and mdev
720          *      are scaled versions of rtt and mean deviation.
721          *      This is designed to be as fast as possible
722          *      m stands for "measurement".
723          *
724          *      On a 1990 paper the rto value is changed to:
725          *      RTO = rtt + 4 * mdev
726          *
727          * Funny. This algorithm seems to be very broken.
728          * These formulae increase RTO, when it should be decreased, increase
729          * too slowly, when it should be increased quickly, decrease too quickly
730          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
731          * does not matter how to _calculate_ it. Seems, it was trap
732          * that VJ failed to avoid. 8)
733          */
734         if (srtt != 0) {
735                 m -= (srtt >> 3);       /* m is now error in rtt est */
736                 srtt += m;              /* rtt = 7/8 rtt + 1/8 new */
737                 if (m < 0) {
738                         m = -m;         /* m is now abs(error) */
739                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
740                         /* This is similar to one of Eifel findings.
741                          * Eifel blocks mdev updates when rtt decreases.
742                          * This solution is a bit different: we use finer gain
743                          * for mdev in this case (alpha*beta).
744                          * Like Eifel it also prevents growth of rto,
745                          * but also it limits too fast rto decreases,
746                          * happening in pure Eifel.
747                          */
748                         if (m > 0)
749                                 m >>= 3;
750                 } else {
751                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
752                 }
753                 tp->mdev_us += m;               /* mdev = 3/4 mdev + 1/4 new */
754                 if (tp->mdev_us > tp->mdev_max_us) {
755                         tp->mdev_max_us = tp->mdev_us;
756                         if (tp->mdev_max_us > tp->rttvar_us)
757                                 tp->rttvar_us = tp->mdev_max_us;
758                 }
759                 if (after(tp->snd_una, tp->rtt_seq)) {
760                         if (tp->mdev_max_us < tp->rttvar_us)
761                                 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
762                         tp->rtt_seq = tp->snd_nxt;
763                         tp->mdev_max_us = tcp_rto_min_us(sk);
764                 }
765         } else {
766                 /* no previous measure. */
767                 srtt = m << 3;          /* take the measured time to be rtt */
768                 tp->mdev_us = m << 1;   /* make sure rto = 3*rtt */
769                 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
770                 tp->mdev_max_us = tp->rttvar_us;
771                 tp->rtt_seq = tp->snd_nxt;
772         }
773         tp->srtt_us = max(1U, srtt);
774 }
775 
776 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
777  * Note: TCP stack does not yet implement pacing.
778  * FQ packet scheduler can be used to implement cheap but effective
779  * TCP pacing, to smooth the burst on large writes when packets
780  * in flight is significantly lower than cwnd (or rwin)
781  */
782 int sysctl_tcp_pacing_ss_ratio __read_mostly = 200;
783 int sysctl_tcp_pacing_ca_ratio __read_mostly = 120;
784 
785 static void tcp_update_pacing_rate(struct sock *sk)
786 {
787         const struct tcp_sock *tp = tcp_sk(sk);
788         u64 rate;
789 
790         /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
791         rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
792 
793         /* current rate is (cwnd * mss) / srtt
794          * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
795          * In Congestion Avoidance phase, set it to 120 % the current rate.
796          *
797          * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
798          *       If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
799          *       end of slow start and should slow down.
800          */
801         if (tp->snd_cwnd < tp->snd_ssthresh / 2)
802                 rate *= sysctl_tcp_pacing_ss_ratio;
803         else
804                 rate *= sysctl_tcp_pacing_ca_ratio;
805 
806         rate *= max(tp->snd_cwnd, tp->packets_out);
807 
808         if (likely(tp->srtt_us))
809                 do_div(rate, tp->srtt_us);
810 
811         /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
812          * without any lock. We want to make sure compiler wont store
813          * intermediate values in this location.
814          */
815         ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
816                                                 sk->sk_max_pacing_rate);
817 }
818 
819 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
820  * routine referred to above.
821  */
822 static void tcp_set_rto(struct sock *sk)
823 {
824         const struct tcp_sock *tp = tcp_sk(sk);
825         /* Old crap is replaced with new one. 8)
826          *
827          * More seriously:
828          * 1. If rtt variance happened to be less 50msec, it is hallucination.
829          *    It cannot be less due to utterly erratic ACK generation made
830          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
831          *    to do with delayed acks, because at cwnd>2 true delack timeout
832          *    is invisible. Actually, Linux-2.4 also generates erratic
833          *    ACKs in some circumstances.
834          */
835         inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
836 
837         /* 2. Fixups made earlier cannot be right.
838          *    If we do not estimate RTO correctly without them,
839          *    all the algo is pure shit and should be replaced
840          *    with correct one. It is exactly, which we pretend to do.
841          */
842 
843         /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
844          * guarantees that rto is higher.
845          */
846         tcp_bound_rto(sk);
847 }
848 
849 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
850 {
851         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
852 
853         if (!cwnd)
854                 cwnd = TCP_INIT_CWND;
855         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
856 }
857 
858 /*
859  * Packet counting of FACK is based on in-order assumptions, therefore TCP
860  * disables it when reordering is detected
861  */
862 void tcp_disable_fack(struct tcp_sock *tp)
863 {
864         /* RFC3517 uses different metric in lost marker => reset on change */
865         if (tcp_is_fack(tp))
866                 tp->lost_skb_hint = NULL;
867         tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
868 }
869 
870 /* Take a notice that peer is sending D-SACKs */
871 static void tcp_dsack_seen(struct tcp_sock *tp)
872 {
873         tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
874 }
875 
876 static void tcp_update_reordering(struct sock *sk, const int metric,
877                                   const int ts)
878 {
879         struct tcp_sock *tp = tcp_sk(sk);
880         if (metric > tp->reordering) {
881                 int mib_idx;
882 
883                 tp->reordering = min(sysctl_tcp_max_reordering, metric);
884 
885                 /* This exciting event is worth to be remembered. 8) */
886                 if (ts)
887                         mib_idx = LINUX_MIB_TCPTSREORDER;
888                 else if (tcp_is_reno(tp))
889                         mib_idx = LINUX_MIB_TCPRENOREORDER;
890                 else if (tcp_is_fack(tp))
891                         mib_idx = LINUX_MIB_TCPFACKREORDER;
892                 else
893                         mib_idx = LINUX_MIB_TCPSACKREORDER;
894 
895                 NET_INC_STATS(sock_net(sk), mib_idx);
896 #if FASTRETRANS_DEBUG > 1
897                 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
898                          tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
899                          tp->reordering,
900                          tp->fackets_out,
901                          tp->sacked_out,
902                          tp->undo_marker ? tp->undo_retrans : 0);
903 #endif
904                 tcp_disable_fack(tp);
905         }
906 
907         if (metric > 0)
908                 tcp_disable_early_retrans(tp);
909         tp->rack.reord = 1;
910 }
911 
912 /* This must be called before lost_out is incremented */
913 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
914 {
915         if (!tp->retransmit_skb_hint ||
916             before(TCP_SKB_CB(skb)->seq,
917                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
918                 tp->retransmit_skb_hint = skb;
919 
920         if (!tp->lost_out ||
921             after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
922                 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
923 }
924 
925 /* Sum the number of packets on the wire we have marked as lost.
926  * There are two cases we care about here:
927  * a) Packet hasn't been marked lost (nor retransmitted),
928  *    and this is the first loss.
929  * b) Packet has been marked both lost and retransmitted,
930  *    and this means we think it was lost again.
931  */
932 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
933 {
934         __u8 sacked = TCP_SKB_CB(skb)->sacked;
935 
936         if (!(sacked & TCPCB_LOST) ||
937             ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
938                 tp->lost += tcp_skb_pcount(skb);
939 }
940 
941 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
942 {
943         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
944                 tcp_verify_retransmit_hint(tp, skb);
945 
946                 tp->lost_out += tcp_skb_pcount(skb);
947                 tcp_sum_lost(tp, skb);
948                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
949         }
950 }
951 
952 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
953 {
954         tcp_verify_retransmit_hint(tp, skb);
955 
956         tcp_sum_lost(tp, skb);
957         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
958                 tp->lost_out += tcp_skb_pcount(skb);
959                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
960         }
961 }
962 
963 /* This procedure tags the retransmission queue when SACKs arrive.
964  *
965  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
966  * Packets in queue with these bits set are counted in variables
967  * sacked_out, retrans_out and lost_out, correspondingly.
968  *
969  * Valid combinations are:
970  * Tag  InFlight        Description
971  * 0    1               - orig segment is in flight.
972  * S    0               - nothing flies, orig reached receiver.
973  * L    0               - nothing flies, orig lost by net.
974  * R    2               - both orig and retransmit are in flight.
975  * L|R  1               - orig is lost, retransmit is in flight.
976  * S|R  1               - orig reached receiver, retrans is still in flight.
977  * (L|S|R is logically valid, it could occur when L|R is sacked,
978  *  but it is equivalent to plain S and code short-curcuits it to S.
979  *  L|S is logically invalid, it would mean -1 packet in flight 8))
980  *
981  * These 6 states form finite state machine, controlled by the following events:
982  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
983  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
984  * 3. Loss detection event of two flavors:
985  *      A. Scoreboard estimator decided the packet is lost.
986  *         A'. Reno "three dupacks" marks head of queue lost.
987  *         A''. Its FACK modification, head until snd.fack is lost.
988  *      B. SACK arrives sacking SND.NXT at the moment, when the
989  *         segment was retransmitted.
990  * 4. D-SACK added new rule: D-SACK changes any tag to S.
991  *
992  * It is pleasant to note, that state diagram turns out to be commutative,
993  * so that we are allowed not to be bothered by order of our actions,
994  * when multiple events arrive simultaneously. (see the function below).
995  *
996  * Reordering detection.
997  * --------------------
998  * Reordering metric is maximal distance, which a packet can be displaced
999  * in packet stream. With SACKs we can estimate it:
1000  *
1001  * 1. SACK fills old hole and the corresponding segment was not
1002  *    ever retransmitted -> reordering. Alas, we cannot use it
1003  *    when segment was retransmitted.
1004  * 2. The last flaw is solved with D-SACK. D-SACK arrives
1005  *    for retransmitted and already SACKed segment -> reordering..
1006  * Both of these heuristics are not used in Loss state, when we cannot
1007  * account for retransmits accurately.
1008  *
1009  * SACK block validation.
1010  * ----------------------
1011  *
1012  * SACK block range validation checks that the received SACK block fits to
1013  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1014  * Note that SND.UNA is not included to the range though being valid because
1015  * it means that the receiver is rather inconsistent with itself reporting
1016  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1017  * perfectly valid, however, in light of RFC2018 which explicitly states
1018  * that "SACK block MUST reflect the newest segment.  Even if the newest
1019  * segment is going to be discarded ...", not that it looks very clever
1020  * in case of head skb. Due to potentional receiver driven attacks, we
1021  * choose to avoid immediate execution of a walk in write queue due to
1022  * reneging and defer head skb's loss recovery to standard loss recovery
1023  * procedure that will eventually trigger (nothing forbids us doing this).
1024  *
1025  * Implements also blockage to start_seq wrap-around. Problem lies in the
1026  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1027  * there's no guarantee that it will be before snd_nxt (n). The problem
1028  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1029  * wrap (s_w):
1030  *
1031  *         <- outs wnd ->                          <- wrapzone ->
1032  *         u     e      n                         u_w   e_w  s n_w
1033  *         |     |      |                          |     |   |  |
1034  * |<------------+------+----- TCP seqno space --------------+---------->|
1035  * ...-- <2^31 ->|                                           |<--------...
1036  * ...---- >2^31 ------>|                                    |<--------...
1037  *
1038  * Current code wouldn't be vulnerable but it's better still to discard such
1039  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1040  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1041  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1042  * equal to the ideal case (infinite seqno space without wrap caused issues).
1043  *
1044  * With D-SACK the lower bound is extended to cover sequence space below
1045  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1046  * again, D-SACK block must not to go across snd_una (for the same reason as
1047  * for the normal SACK blocks, explained above). But there all simplicity
1048  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1049  * fully below undo_marker they do not affect behavior in anyway and can
1050  * therefore be safely ignored. In rare cases (which are more or less
1051  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1052  * fragmentation and packet reordering past skb's retransmission. To consider
1053  * them correctly, the acceptable range must be extended even more though
1054  * the exact amount is rather hard to quantify. However, tp->max_window can
1055  * be used as an exaggerated estimate.
1056  */
1057 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1058                                    u32 start_seq, u32 end_seq)
1059 {
1060         /* Too far in future, or reversed (interpretation is ambiguous) */
1061         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1062                 return false;
1063 
1064         /* Nasty start_seq wrap-around check (see comments above) */
1065         if (!before(start_seq, tp->snd_nxt))
1066                 return false;
1067 
1068         /* In outstanding window? ...This is valid exit for D-SACKs too.
1069          * start_seq == snd_una is non-sensical (see comments above)
1070          */
1071         if (after(start_seq, tp->snd_una))
1072                 return true;
1073 
1074         if (!is_dsack || !tp->undo_marker)
1075                 return false;
1076 
1077         /* ...Then it's D-SACK, and must reside below snd_una completely */
1078         if (after(end_seq, tp->snd_una))
1079                 return false;
1080 
1081         if (!before(start_seq, tp->undo_marker))
1082                 return true;
1083 
1084         /* Too old */
1085         if (!after(end_seq, tp->undo_marker))
1086                 return false;
1087 
1088         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1089          *   start_seq < undo_marker and end_seq >= undo_marker.
1090          */
1091         return !before(start_seq, end_seq - tp->max_window);
1092 }
1093 
1094 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1095                             struct tcp_sack_block_wire *sp, int num_sacks,
1096                             u32 prior_snd_una)
1097 {
1098         struct tcp_sock *tp = tcp_sk(sk);
1099         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1100         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1101         bool dup_sack = false;
1102 
1103         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1104                 dup_sack = true;
1105                 tcp_dsack_seen(tp);
1106                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1107         } else if (num_sacks > 1) {
1108                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1109                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1110 
1111                 if (!after(end_seq_0, end_seq_1) &&
1112                     !before(start_seq_0, start_seq_1)) {
1113                         dup_sack = true;
1114                         tcp_dsack_seen(tp);
1115                         NET_INC_STATS(sock_net(sk),
1116                                         LINUX_MIB_TCPDSACKOFORECV);
1117                 }
1118         }
1119 
1120         /* D-SACK for already forgotten data... Do dumb counting. */
1121         if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1122             !after(end_seq_0, prior_snd_una) &&
1123             after(end_seq_0, tp->undo_marker))
1124                 tp->undo_retrans--;
1125 
1126         return dup_sack;
1127 }
1128 
1129 struct tcp_sacktag_state {
1130         int     reord;
1131         int     fack_count;
1132         /* Timestamps for earliest and latest never-retransmitted segment
1133          * that was SACKed. RTO needs the earliest RTT to stay conservative,
1134          * but congestion control should still get an accurate delay signal.
1135          */
1136         struct skb_mstamp first_sackt;
1137         struct skb_mstamp last_sackt;
1138         struct rate_sample *rate;
1139         int     flag;
1140 };
1141 
1142 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1143  * the incoming SACK may not exactly match but we can find smaller MSS
1144  * aligned portion of it that matches. Therefore we might need to fragment
1145  * which may fail and creates some hassle (caller must handle error case
1146  * returns).
1147  *
1148  * FIXME: this could be merged to shift decision code
1149  */
1150 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1151                                   u32 start_seq, u32 end_seq)
1152 {
1153         int err;
1154         bool in_sack;
1155         unsigned int pkt_len;
1156         unsigned int mss;
1157 
1158         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1159                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1160 
1161         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1162             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1163                 mss = tcp_skb_mss(skb);
1164                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1165 
1166                 if (!in_sack) {
1167                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1168                         if (pkt_len < mss)
1169                                 pkt_len = mss;
1170                 } else {
1171                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1172                         if (pkt_len < mss)
1173                                 return -EINVAL;
1174                 }
1175 
1176                 /* Round if necessary so that SACKs cover only full MSSes
1177                  * and/or the remaining small portion (if present)
1178                  */
1179                 if (pkt_len > mss) {
1180                         unsigned int new_len = (pkt_len / mss) * mss;
1181                         if (!in_sack && new_len < pkt_len) {
1182                                 new_len += mss;
1183                                 if (new_len >= skb->len)
1184                                         return 0;
1185                         }
1186                         pkt_len = new_len;
1187                 }
1188                 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1189                 if (err < 0)
1190                         return err;
1191         }
1192 
1193         return in_sack;
1194 }
1195 
1196 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1197 static u8 tcp_sacktag_one(struct sock *sk,
1198                           struct tcp_sacktag_state *state, u8 sacked,
1199                           u32 start_seq, u32 end_seq,
1200                           int dup_sack, int pcount,
1201                           const struct skb_mstamp *xmit_time)
1202 {
1203         struct tcp_sock *tp = tcp_sk(sk);
1204         int fack_count = state->fack_count;
1205 
1206         /* Account D-SACK for retransmitted packet. */
1207         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1208                 if (tp->undo_marker && tp->undo_retrans > 0 &&
1209                     after(end_seq, tp->undo_marker))
1210                         tp->undo_retrans--;
1211                 if (sacked & TCPCB_SACKED_ACKED)
1212                         state->reord = min(fack_count, state->reord);
1213         }
1214 
1215         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1216         if (!after(end_seq, tp->snd_una))
1217                 return sacked;
1218 
1219         if (!(sacked & TCPCB_SACKED_ACKED)) {
1220                 tcp_rack_advance(tp, xmit_time, sacked);
1221 
1222                 if (sacked & TCPCB_SACKED_RETRANS) {
1223                         /* If the segment is not tagged as lost,
1224                          * we do not clear RETRANS, believing
1225                          * that retransmission is still in flight.
1226                          */
1227                         if (sacked & TCPCB_LOST) {
1228                                 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1229                                 tp->lost_out -= pcount;
1230                                 tp->retrans_out -= pcount;
1231                         }
1232                 } else {
1233                         if (!(sacked & TCPCB_RETRANS)) {
1234                                 /* New sack for not retransmitted frame,
1235                                  * which was in hole. It is reordering.
1236                                  */
1237                                 if (before(start_seq,
1238                                            tcp_highest_sack_seq(tp)))
1239                                         state->reord = min(fack_count,
1240                                                            state->reord);
1241                                 if (!after(end_seq, tp->high_seq))
1242                                         state->flag |= FLAG_ORIG_SACK_ACKED;
1243                                 if (state->first_sackt.v64 == 0)
1244                                         state->first_sackt = *xmit_time;
1245                                 state->last_sackt = *xmit_time;
1246                         }
1247 
1248                         if (sacked & TCPCB_LOST) {
1249                                 sacked &= ~TCPCB_LOST;
1250                                 tp->lost_out -= pcount;
1251                         }
1252                 }
1253 
1254                 sacked |= TCPCB_SACKED_ACKED;
1255                 state->flag |= FLAG_DATA_SACKED;
1256                 tp->sacked_out += pcount;
1257                 tp->delivered += pcount;  /* Out-of-order packets delivered */
1258 
1259                 fack_count += pcount;
1260 
1261                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1262                 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1263                     before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1264                         tp->lost_cnt_hint += pcount;
1265 
1266                 if (fack_count > tp->fackets_out)
1267                         tp->fackets_out = fack_count;
1268         }
1269 
1270         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1271          * frames and clear it. undo_retrans is decreased above, L|R frames
1272          * are accounted above as well.
1273          */
1274         if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1275                 sacked &= ~TCPCB_SACKED_RETRANS;
1276                 tp->retrans_out -= pcount;
1277         }
1278 
1279         return sacked;
1280 }
1281 
1282 /* Shift newly-SACKed bytes from this skb to the immediately previous
1283  * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1284  */
1285 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1286                             struct tcp_sacktag_state *state,
1287                             unsigned int pcount, int shifted, int mss,
1288                             bool dup_sack)
1289 {
1290         struct tcp_sock *tp = tcp_sk(sk);
1291         struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1292         u32 start_seq = TCP_SKB_CB(skb)->seq;   /* start of newly-SACKed */
1293         u32 end_seq = start_seq + shifted;      /* end of newly-SACKed */
1294 
1295         BUG_ON(!pcount);
1296 
1297         /* Adjust counters and hints for the newly sacked sequence
1298          * range but discard the return value since prev is already
1299          * marked. We must tag the range first because the seq
1300          * advancement below implicitly advances
1301          * tcp_highest_sack_seq() when skb is highest_sack.
1302          */
1303         tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1304                         start_seq, end_seq, dup_sack, pcount,
1305                         &skb->skb_mstamp);
1306         tcp_rate_skb_delivered(sk, skb, state->rate);
1307 
1308         if (skb == tp->lost_skb_hint)
1309                 tp->lost_cnt_hint += pcount;
1310 
1311         TCP_SKB_CB(prev)->end_seq += shifted;
1312         TCP_SKB_CB(skb)->seq += shifted;
1313 
1314         tcp_skb_pcount_add(prev, pcount);
1315         BUG_ON(tcp_skb_pcount(skb) < pcount);
1316         tcp_skb_pcount_add(skb, -pcount);
1317 
1318         /* When we're adding to gso_segs == 1, gso_size will be zero,
1319          * in theory this shouldn't be necessary but as long as DSACK
1320          * code can come after this skb later on it's better to keep
1321          * setting gso_size to something.
1322          */
1323         if (!TCP_SKB_CB(prev)->tcp_gso_size)
1324                 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1325 
1326         /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1327         if (tcp_skb_pcount(skb) <= 1)
1328                 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1329 
1330         /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1331         TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1332 
1333         if (skb->len > 0) {
1334                 BUG_ON(!tcp_skb_pcount(skb));
1335                 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1336                 return false;
1337         }
1338 
1339         /* Whole SKB was eaten :-) */
1340 
1341         if (skb == tp->retransmit_skb_hint)
1342                 tp->retransmit_skb_hint = prev;
1343         if (skb == tp->lost_skb_hint) {
1344                 tp->lost_skb_hint = prev;
1345                 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1346         }
1347 
1348         TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1349         TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1350         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1351                 TCP_SKB_CB(prev)->end_seq++;
1352 
1353         if (skb == tcp_highest_sack(sk))
1354                 tcp_advance_highest_sack(sk, skb);
1355 
1356         tcp_skb_collapse_tstamp(prev, skb);
1357         if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp.v64))
1358                 TCP_SKB_CB(prev)->tx.delivered_mstamp.v64 = 0;
1359 
1360         tcp_unlink_write_queue(skb, sk);
1361         sk_wmem_free_skb(sk, skb);
1362 
1363         NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1364 
1365         return true;
1366 }
1367 
1368 /* I wish gso_size would have a bit more sane initialization than
1369  * something-or-zero which complicates things
1370  */
1371 static int tcp_skb_seglen(const struct sk_buff *skb)
1372 {
1373         return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1374 }
1375 
1376 /* Shifting pages past head area doesn't work */
1377 static int skb_can_shift(const struct sk_buff *skb)
1378 {
1379         return !skb_headlen(skb) && skb_is_nonlinear(skb);
1380 }
1381 
1382 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1383  * skb.
1384  */
1385 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1386                                           struct tcp_sacktag_state *state,
1387                                           u32 start_seq, u32 end_seq,
1388                                           bool dup_sack)
1389 {
1390         struct tcp_sock *tp = tcp_sk(sk);
1391         struct sk_buff *prev;
1392         int mss;
1393         int pcount = 0;
1394         int len;
1395         int in_sack;
1396 
1397         if (!sk_can_gso(sk))
1398                 goto fallback;
1399 
1400         /* Normally R but no L won't result in plain S */
1401         if (!dup_sack &&
1402             (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1403                 goto fallback;
1404         if (!skb_can_shift(skb))
1405                 goto fallback;
1406         /* This frame is about to be dropped (was ACKed). */
1407         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1408                 goto fallback;
1409 
1410         /* Can only happen with delayed DSACK + discard craziness */
1411         if (unlikely(skb == tcp_write_queue_head(sk)))
1412                 goto fallback;
1413         prev = tcp_write_queue_prev(sk, skb);
1414 
1415         if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1416                 goto fallback;
1417 
1418         if (!tcp_skb_can_collapse_to(prev))
1419                 goto fallback;
1420 
1421         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1422                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1423 
1424         if (in_sack) {
1425                 len = skb->len;
1426                 pcount = tcp_skb_pcount(skb);
1427                 mss = tcp_skb_seglen(skb);
1428 
1429                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1430                  * drop this restriction as unnecessary
1431                  */
1432                 if (mss != tcp_skb_seglen(prev))
1433                         goto fallback;
1434         } else {
1435                 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1436                         goto noop;
1437                 /* CHECKME: This is non-MSS split case only?, this will
1438                  * cause skipped skbs due to advancing loop btw, original
1439                  * has that feature too
1440                  */
1441                 if (tcp_skb_pcount(skb) <= 1)
1442                         goto noop;
1443 
1444                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1445                 if (!in_sack) {
1446                         /* TODO: head merge to next could be attempted here
1447                          * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1448                          * though it might not be worth of the additional hassle
1449                          *
1450                          * ...we can probably just fallback to what was done
1451                          * previously. We could try merging non-SACKed ones
1452                          * as well but it probably isn't going to buy off
1453                          * because later SACKs might again split them, and
1454                          * it would make skb timestamp tracking considerably
1455                          * harder problem.
1456                          */
1457                         goto fallback;
1458                 }
1459 
1460                 len = end_seq - TCP_SKB_CB(skb)->seq;
1461                 BUG_ON(len < 0);
1462                 BUG_ON(len > skb->len);
1463 
1464                 /* MSS boundaries should be honoured or else pcount will
1465                  * severely break even though it makes things bit trickier.
1466                  * Optimize common case to avoid most of the divides
1467                  */
1468                 mss = tcp_skb_mss(skb);
1469 
1470                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1471                  * drop this restriction as unnecessary
1472                  */
1473                 if (mss != tcp_skb_seglen(prev))
1474                         goto fallback;
1475 
1476                 if (len == mss) {
1477                         pcount = 1;
1478                 } else if (len < mss) {
1479                         goto noop;
1480                 } else {
1481                         pcount = len / mss;
1482                         len = pcount * mss;
1483                 }
1484         }
1485 
1486         /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1487         if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1488                 goto fallback;
1489 
1490         if (!skb_shift(prev, skb, len))
1491                 goto fallback;
1492         if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1493                 goto out;
1494 
1495         /* Hole filled allows collapsing with the next as well, this is very
1496          * useful when hole on every nth skb pattern happens
1497          */
1498         if (prev == tcp_write_queue_tail(sk))
1499                 goto out;
1500         skb = tcp_write_queue_next(sk, prev);
1501 
1502         if (!skb_can_shift(skb) ||
1503             (skb == tcp_send_head(sk)) ||
1504             ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1505             (mss != tcp_skb_seglen(skb)))
1506                 goto out;
1507 
1508         len = skb->len;
1509         if (skb_shift(prev, skb, len)) {
1510                 pcount += tcp_skb_pcount(skb);
1511                 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1512         }
1513 
1514 out:
1515         state->fack_count += pcount;
1516         return prev;
1517 
1518 noop:
1519         return skb;
1520 
1521 fallback:
1522         NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1523         return NULL;
1524 }
1525 
1526 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1527                                         struct tcp_sack_block *next_dup,
1528                                         struct tcp_sacktag_state *state,
1529                                         u32 start_seq, u32 end_seq,
1530                                         bool dup_sack_in)
1531 {
1532         struct tcp_sock *tp = tcp_sk(sk);
1533         struct sk_buff *tmp;
1534 
1535         tcp_for_write_queue_from(skb, sk) {
1536                 int in_sack = 0;
1537                 bool dup_sack = dup_sack_in;
1538 
1539                 if (skb == tcp_send_head(sk))
1540                         break;
1541 
1542                 /* queue is in-order => we can short-circuit the walk early */
1543                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1544                         break;
1545 
1546                 if (next_dup  &&
1547                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1548                         in_sack = tcp_match_skb_to_sack(sk, skb,
1549                                                         next_dup->start_seq,
1550                                                         next_dup->end_seq);
1551                         if (in_sack > 0)
1552                                 dup_sack = true;
1553                 }
1554 
1555                 /* skb reference here is a bit tricky to get right, since
1556                  * shifting can eat and free both this skb and the next,
1557                  * so not even _safe variant of the loop is enough.
1558                  */
1559                 if (in_sack <= 0) {
1560                         tmp = tcp_shift_skb_data(sk, skb, state,
1561                                                  start_seq, end_seq, dup_sack);
1562                         if (tmp) {
1563                                 if (tmp != skb) {
1564                                         skb = tmp;
1565                                         continue;
1566                                 }
1567 
1568                                 in_sack = 0;
1569                         } else {
1570                                 in_sack = tcp_match_skb_to_sack(sk, skb,
1571                                                                 start_seq,
1572                                                                 end_seq);
1573                         }
1574                 }
1575 
1576                 if (unlikely(in_sack < 0))
1577                         break;
1578 
1579                 if (in_sack) {
1580                         TCP_SKB_CB(skb)->sacked =
1581                                 tcp_sacktag_one(sk,
1582                                                 state,
1583                                                 TCP_SKB_CB(skb)->sacked,
1584                                                 TCP_SKB_CB(skb)->seq,
1585                                                 TCP_SKB_CB(skb)->end_seq,
1586                                                 dup_sack,
1587                                                 tcp_skb_pcount(skb),
1588                                                 &skb->skb_mstamp);
1589                         tcp_rate_skb_delivered(sk, skb, state->rate);
1590 
1591                         if (!before(TCP_SKB_CB(skb)->seq,
1592                                     tcp_highest_sack_seq(tp)))
1593                                 tcp_advance_highest_sack(sk, skb);
1594                 }
1595 
1596                 state->fack_count += tcp_skb_pcount(skb);
1597         }
1598         return skb;
1599 }
1600 
1601 /* Avoid all extra work that is being done by sacktag while walking in
1602  * a normal way
1603  */
1604 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1605                                         struct tcp_sacktag_state *state,
1606                                         u32 skip_to_seq)
1607 {
1608         tcp_for_write_queue_from(skb, sk) {
1609                 if (skb == tcp_send_head(sk))
1610                         break;
1611 
1612                 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1613                         break;
1614 
1615                 state->fack_count += tcp_skb_pcount(skb);
1616         }
1617         return skb;
1618 }
1619 
1620 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1621                                                 struct sock *sk,
1622                                                 struct tcp_sack_block *next_dup,
1623                                                 struct tcp_sacktag_state *state,
1624                                                 u32 skip_to_seq)
1625 {
1626         if (!next_dup)
1627                 return skb;
1628 
1629         if (before(next_dup->start_seq, skip_to_seq)) {
1630                 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1631                 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1632                                        next_dup->start_seq, next_dup->end_seq,
1633                                        1);
1634         }
1635 
1636         return skb;
1637 }
1638 
1639 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1640 {
1641         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1642 }
1643 
1644 static int
1645 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1646                         u32 prior_snd_una, struct tcp_sacktag_state *state)
1647 {
1648         struct tcp_sock *tp = tcp_sk(sk);
1649         const unsigned char *ptr = (skb_transport_header(ack_skb) +
1650                                     TCP_SKB_CB(ack_skb)->sacked);
1651         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1652         struct tcp_sack_block sp[TCP_NUM_SACKS];
1653         struct tcp_sack_block *cache;
1654         struct sk_buff *skb;
1655         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1656         int used_sacks;
1657         bool found_dup_sack = false;
1658         int i, j;
1659         int first_sack_index;
1660 
1661         state->flag = 0;
1662         state->reord = tp->packets_out;
1663 
1664         if (!tp->sacked_out) {
1665                 if (WARN_ON(tp->fackets_out))
1666                         tp->fackets_out = 0;
1667                 tcp_highest_sack_reset(sk);
1668         }
1669 
1670         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1671                                          num_sacks, prior_snd_una);
1672         if (found_dup_sack) {
1673                 state->flag |= FLAG_DSACKING_ACK;
1674                 tp->delivered++; /* A spurious retransmission is delivered */
1675         }
1676 
1677         /* Eliminate too old ACKs, but take into
1678          * account more or less fresh ones, they can
1679          * contain valid SACK info.
1680          */
1681         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1682                 return 0;
1683 
1684         if (!tp->packets_out)
1685                 goto out;
1686 
1687         used_sacks = 0;
1688         first_sack_index = 0;
1689         for (i = 0; i < num_sacks; i++) {
1690                 bool dup_sack = !i && found_dup_sack;
1691 
1692                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1693                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1694 
1695                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1696                                             sp[used_sacks].start_seq,
1697                                             sp[used_sacks].end_seq)) {
1698                         int mib_idx;
1699 
1700                         if (dup_sack) {
1701                                 if (!tp->undo_marker)
1702                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1703                                 else
1704                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1705                         } else {
1706                                 /* Don't count olds caused by ACK reordering */
1707                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1708                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1709                                         continue;
1710                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1711                         }
1712 
1713                         NET_INC_STATS(sock_net(sk), mib_idx);
1714                         if (i == 0)
1715                                 first_sack_index = -1;
1716                         continue;
1717                 }
1718 
1719                 /* Ignore very old stuff early */
1720                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1721                         continue;
1722 
1723                 used_sacks++;
1724         }
1725 
1726         /* order SACK blocks to allow in order walk of the retrans queue */
1727         for (i = used_sacks - 1; i > 0; i--) {
1728                 for (j = 0; j < i; j++) {
1729                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1730                                 swap(sp[j], sp[j + 1]);
1731 
1732                                 /* Track where the first SACK block goes to */
1733                                 if (j == first_sack_index)
1734                                         first_sack_index = j + 1;
1735                         }
1736                 }
1737         }
1738 
1739         skb = tcp_write_queue_head(sk);
1740         state->fack_count = 0;
1741         i = 0;
1742 
1743         if (!tp->sacked_out) {
1744                 /* It's already past, so skip checking against it */
1745                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1746         } else {
1747                 cache = tp->recv_sack_cache;
1748                 /* Skip empty blocks in at head of the cache */
1749                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1750                        !cache->end_seq)
1751                         cache++;
1752         }
1753 
1754         while (i < used_sacks) {
1755                 u32 start_seq = sp[i].start_seq;
1756                 u32 end_seq = sp[i].end_seq;
1757                 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1758                 struct tcp_sack_block *next_dup = NULL;
1759 
1760                 if (found_dup_sack && ((i + 1) == first_sack_index))
1761                         next_dup = &sp[i + 1];
1762 
1763                 /* Skip too early cached blocks */
1764                 while (tcp_sack_cache_ok(tp, cache) &&
1765                        !before(start_seq, cache->end_seq))
1766                         cache++;
1767 
1768                 /* Can skip some work by looking recv_sack_cache? */
1769                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1770                     after(end_seq, cache->start_seq)) {
1771 
1772                         /* Head todo? */
1773                         if (before(start_seq, cache->start_seq)) {
1774                                 skb = tcp_sacktag_skip(skb, sk, state,
1775                                                        start_seq);
1776                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1777                                                        state,
1778                                                        start_seq,
1779                                                        cache->start_seq,
1780                                                        dup_sack);
1781                         }
1782 
1783                         /* Rest of the block already fully processed? */
1784                         if (!after(end_seq, cache->end_seq))
1785                                 goto advance_sp;
1786 
1787                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1788                                                        state,
1789                                                        cache->end_seq);
1790 
1791                         /* ...tail remains todo... */
1792                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1793                                 /* ...but better entrypoint exists! */
1794                                 skb = tcp_highest_sack(sk);
1795                                 if (!skb)
1796                                         break;
1797                                 state->fack_count = tp->fackets_out;
1798                                 cache++;
1799                                 goto walk;
1800                         }
1801 
1802                         skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1803                         /* Check overlap against next cached too (past this one already) */
1804                         cache++;
1805                         continue;
1806                 }
1807 
1808                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1809                         skb = tcp_highest_sack(sk);
1810                         if (!skb)
1811                                 break;
1812                         state->fack_count = tp->fackets_out;
1813                 }
1814                 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1815 
1816 walk:
1817                 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1818                                        start_seq, end_seq, dup_sack);
1819 
1820 advance_sp:
1821                 i++;
1822         }
1823 
1824         /* Clear the head of the cache sack blocks so we can skip it next time */
1825         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1826                 tp->recv_sack_cache[i].start_seq = 0;
1827                 tp->recv_sack_cache[i].end_seq = 0;
1828         }
1829         for (j = 0; j < used_sacks; j++)
1830                 tp->recv_sack_cache[i++] = sp[j];
1831 
1832         if ((state->reord < tp->fackets_out) &&
1833             ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1834                 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
1835 
1836         tcp_verify_left_out(tp);
1837 out:
1838 
1839 #if FASTRETRANS_DEBUG > 0
1840         WARN_ON((int)tp->sacked_out < 0);
1841         WARN_ON((int)tp->lost_out < 0);
1842         WARN_ON((int)tp->retrans_out < 0);
1843         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1844 #endif
1845         return state->flag;
1846 }
1847 
1848 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1849  * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1850  */
1851 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1852 {
1853         u32 holes;
1854 
1855         holes = max(tp->lost_out, 1U);
1856         holes = min(holes, tp->packets_out);
1857 
1858         if ((tp->sacked_out + holes) > tp->packets_out) {
1859                 tp->sacked_out = tp->packets_out - holes;
1860                 return true;
1861         }
1862         return false;
1863 }
1864 
1865 /* If we receive more dupacks than we expected counting segments
1866  * in assumption of absent reordering, interpret this as reordering.
1867  * The only another reason could be bug in receiver TCP.
1868  */
1869 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1870 {
1871         struct tcp_sock *tp = tcp_sk(sk);
1872         if (tcp_limit_reno_sacked(tp))
1873                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1874 }
1875 
1876 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1877 
1878 static void tcp_add_reno_sack(struct sock *sk)
1879 {
1880         struct tcp_sock *tp = tcp_sk(sk);
1881         u32 prior_sacked = tp->sacked_out;
1882 
1883         tp->sacked_out++;
1884         tcp_check_reno_reordering(sk, 0);
1885         if (tp->sacked_out > prior_sacked)
1886                 tp->delivered++; /* Some out-of-order packet is delivered */
1887         tcp_verify_left_out(tp);
1888 }
1889 
1890 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1891 
1892 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1893 {
1894         struct tcp_sock *tp = tcp_sk(sk);
1895 
1896         if (acked > 0) {
1897                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1898                 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1899                 if (acked - 1 >= tp->sacked_out)
1900                         tp->sacked_out = 0;
1901                 else
1902                         tp->sacked_out -= acked - 1;
1903         }
1904         tcp_check_reno_reordering(sk, acked);
1905         tcp_verify_left_out(tp);
1906 }
1907 
1908 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1909 {
1910         tp->sacked_out = 0;
1911 }
1912 
1913 void tcp_clear_retrans(struct tcp_sock *tp)
1914 {
1915         tp->retrans_out = 0;
1916         tp->lost_out = 0;
1917         tp->undo_marker = 0;
1918         tp->undo_retrans = -1;
1919         tp->fackets_out = 0;
1920         tp->sacked_out = 0;
1921 }
1922 
1923 static inline void tcp_init_undo(struct tcp_sock *tp)
1924 {
1925         tp->undo_marker = tp->snd_una;
1926         /* Retransmission still in flight may cause DSACKs later. */
1927         tp->undo_retrans = tp->retrans_out ? : -1;
1928 }
1929 
1930 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1931  * and reset tags completely, otherwise preserve SACKs. If receiver
1932  * dropped its ofo queue, we will know this due to reneging detection.
1933  */
1934 void tcp_enter_loss(struct sock *sk)
1935 {
1936         const struct inet_connection_sock *icsk = inet_csk(sk);
1937         struct tcp_sock *tp = tcp_sk(sk);
1938         struct net *net = sock_net(sk);
1939         struct sk_buff *skb;
1940         bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1941         bool is_reneg;                  /* is receiver reneging on SACKs? */
1942         bool mark_lost;
1943 
1944         /* Reduce ssthresh if it has not yet been made inside this window. */
1945         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1946             !after(tp->high_seq, tp->snd_una) ||
1947             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1948                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1949                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1950                 tcp_ca_event(sk, CA_EVENT_LOSS);
1951                 tcp_init_undo(tp);
1952         }
1953         tp->snd_cwnd       = 1;
1954         tp->snd_cwnd_cnt   = 0;
1955         tp->snd_cwnd_stamp = tcp_time_stamp;
1956 
1957         tp->retrans_out = 0;
1958         tp->lost_out = 0;
1959 
1960         if (tcp_is_reno(tp))
1961                 tcp_reset_reno_sack(tp);
1962 
1963         skb = tcp_write_queue_head(sk);
1964         is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1965         if (is_reneg) {
1966                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1967                 tp->sacked_out = 0;
1968                 tp->fackets_out = 0;
1969         }
1970         tcp_clear_all_retrans_hints(tp);
1971 
1972         tcp_for_write_queue(skb, sk) {
1973                 if (skb == tcp_send_head(sk))
1974                         break;
1975 
1976                 mark_lost = (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
1977                              is_reneg);
1978                 if (mark_lost)
1979                         tcp_sum_lost(tp, skb);
1980                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1981                 if (mark_lost) {
1982                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1983                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1984                         tp->lost_out += tcp_skb_pcount(skb);
1985                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1986                 }
1987         }
1988         tcp_verify_left_out(tp);
1989 
1990         /* Timeout in disordered state after receiving substantial DUPACKs
1991          * suggests that the degree of reordering is over-estimated.
1992          */
1993         if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1994             tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
1995                 tp->reordering = min_t(unsigned int, tp->reordering,
1996                                        net->ipv4.sysctl_tcp_reordering);
1997         tcp_set_ca_state(sk, TCP_CA_Loss);
1998         tp->high_seq = tp->snd_nxt;
1999         tcp_ecn_queue_cwr(tp);
2000 
2001         /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2002          * loss recovery is underway except recurring timeout(s) on
2003          * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2004          */
2005         tp->frto = sysctl_tcp_frto &&
2006                    (new_recovery || icsk->icsk_retransmits) &&
2007                    !inet_csk(sk)->icsk_mtup.probe_size;
2008 }
2009 
2010 /* If ACK arrived pointing to a remembered SACK, it means that our
2011  * remembered SACKs do not reflect real state of receiver i.e.
2012  * receiver _host_ is heavily congested (or buggy).
2013  *
2014  * To avoid big spurious retransmission bursts due to transient SACK
2015  * scoreboard oddities that look like reneging, we give the receiver a
2016  * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2017  * restore sanity to the SACK scoreboard. If the apparent reneging
2018  * persists until this RTO then we'll clear the SACK scoreboard.
2019  */
2020 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2021 {
2022         if (flag & FLAG_SACK_RENEGING) {
2023                 struct tcp_sock *tp = tcp_sk(sk);
2024                 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2025                                           msecs_to_jiffies(10));
2026 
2027                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2028                                           delay, TCP_RTO_MAX);
2029                 return true;
2030         }
2031         return false;
2032 }
2033 
2034 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2035 {
2036         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2037 }
2038 
2039 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2040  * counter when SACK is enabled (without SACK, sacked_out is used for
2041  * that purpose).
2042  *
2043  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2044  * segments up to the highest received SACK block so far and holes in
2045  * between them.
2046  *
2047  * With reordering, holes may still be in flight, so RFC3517 recovery
2048  * uses pure sacked_out (total number of SACKed segments) even though
2049  * it violates the RFC that uses duplicate ACKs, often these are equal
2050  * but when e.g. out-of-window ACKs or packet duplication occurs,
2051  * they differ. Since neither occurs due to loss, TCP should really
2052  * ignore them.
2053  */
2054 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2055 {
2056         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2057 }
2058 
2059 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2060 {
2061         struct tcp_sock *tp = tcp_sk(sk);
2062         unsigned long delay;
2063 
2064         /* Delay early retransmit and entering fast recovery for
2065          * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2066          * available, or RTO is scheduled to fire first.
2067          */
2068         if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2069             (flag & FLAG_ECE) || !tp->srtt_us)
2070                 return false;
2071 
2072         delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2073                     msecs_to_jiffies(2));
2074 
2075         if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2076                 return false;
2077 
2078         inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2079                                   TCP_RTO_MAX);
2080         return true;
2081 }
2082 
2083 /* Linux NewReno/SACK/FACK/ECN state machine.
2084  * --------------------------------------
2085  *
2086  * "Open"       Normal state, no dubious events, fast path.
2087  * "Disorder"   In all the respects it is "Open",
2088  *              but requires a bit more attention. It is entered when
2089  *              we see some SACKs or dupacks. It is split of "Open"
2090  *              mainly to move some processing from fast path to slow one.
2091  * "CWR"        CWND was reduced due to some Congestion Notification event.
2092  *              It can be ECN, ICMP source quench, local device congestion.
2093  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2094  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2095  *
2096  * tcp_fastretrans_alert() is entered:
2097  * - each incoming ACK, if state is not "Open"
2098  * - when arrived ACK is unusual, namely:
2099  *      * SACK
2100  *      * Duplicate ACK.
2101  *      * ECN ECE.
2102  *
2103  * Counting packets in flight is pretty simple.
2104  *
2105  *      in_flight = packets_out - left_out + retrans_out
2106  *
2107  *      packets_out is SND.NXT-SND.UNA counted in packets.
2108  *
2109  *      retrans_out is number of retransmitted segments.
2110  *
2111  *      left_out is number of segments left network, but not ACKed yet.
2112  *
2113  *              left_out = sacked_out + lost_out
2114  *
2115  *     sacked_out: Packets, which arrived to receiver out of order
2116  *                 and hence not ACKed. With SACKs this number is simply
2117  *                 amount of SACKed data. Even without SACKs
2118  *                 it is easy to give pretty reliable estimate of this number,
2119  *                 counting duplicate ACKs.
2120  *
2121  *       lost_out: Packets lost by network. TCP has no explicit
2122  *                 "loss notification" feedback from network (for now).
2123  *                 It means that this number can be only _guessed_.
2124  *                 Actually, it is the heuristics to predict lossage that
2125  *                 distinguishes different algorithms.
2126  *
2127  *      F.e. after RTO, when all the queue is considered as lost,
2128  *      lost_out = packets_out and in_flight = retrans_out.
2129  *
2130  *              Essentially, we have now two algorithms counting
2131  *              lost packets.
2132  *
2133  *              FACK: It is the simplest heuristics. As soon as we decided
2134  *              that something is lost, we decide that _all_ not SACKed
2135  *              packets until the most forward SACK are lost. I.e.
2136  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2137  *              It is absolutely correct estimate, if network does not reorder
2138  *              packets. And it loses any connection to reality when reordering
2139  *              takes place. We use FACK by default until reordering
2140  *              is suspected on the path to this destination.
2141  *
2142  *              NewReno: when Recovery is entered, we assume that one segment
2143  *              is lost (classic Reno). While we are in Recovery and
2144  *              a partial ACK arrives, we assume that one more packet
2145  *              is lost (NewReno). This heuristics are the same in NewReno
2146  *              and SACK.
2147  *
2148  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2149  *  deflation etc. CWND is real congestion window, never inflated, changes
2150  *  only according to classic VJ rules.
2151  *
2152  * Really tricky (and requiring careful tuning) part of algorithm
2153  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2154  * The first determines the moment _when_ we should reduce CWND and,
2155  * hence, slow down forward transmission. In fact, it determines the moment
2156  * when we decide that hole is caused by loss, rather than by a reorder.
2157  *
2158  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2159  * holes, caused by lost packets.
2160  *
2161  * And the most logically complicated part of algorithm is undo
2162  * heuristics. We detect false retransmits due to both too early
2163  * fast retransmit (reordering) and underestimated RTO, analyzing
2164  * timestamps and D-SACKs. When we detect that some segments were
2165  * retransmitted by mistake and CWND reduction was wrong, we undo
2166  * window reduction and abort recovery phase. This logic is hidden
2167  * inside several functions named tcp_try_undo_<something>.
2168  */
2169 
2170 /* This function decides, when we should leave Disordered state
2171  * and enter Recovery phase, reducing congestion window.
2172  *
2173  * Main question: may we further continue forward transmission
2174  * with the same cwnd?
2175  */
2176 static bool tcp_time_to_recover(struct sock *sk, int flag)
2177 {
2178         struct tcp_sock *tp = tcp_sk(sk);
2179         __u32 packets_out;
2180         int tcp_reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
2181 
2182         /* Trick#1: The loss is proven. */
2183         if (tp->lost_out)
2184                 return true;
2185 
2186         /* Not-A-Trick#2 : Classic rule... */
2187         if (tcp_dupack_heuristics(tp) > tp->reordering)
2188                 return true;
2189 
2190         /* Trick#4: It is still not OK... But will it be useful to delay
2191          * recovery more?
2192          */
2193         packets_out = tp->packets_out;
2194         if (packets_out <= tp->reordering &&
2195             tp->sacked_out >= max_t(__u32, packets_out/2, tcp_reordering) &&
2196             !tcp_may_send_now(sk)) {
2197                 /* We have nothing to send. This connection is limited
2198                  * either by receiver window or by application.
2199                  */
2200                 return true;
2201         }
2202 
2203         /* If a thin stream is detected, retransmit after first
2204          * received dupack. Employ only if SACK is supported in order
2205          * to avoid possible corner-case series of spurious retransmissions
2206          * Use only if there are no unsent data.
2207          */
2208         if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2209             tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2210             tcp_is_sack(tp) && !tcp_send_head(sk))
2211                 return true;
2212 
2213         /* Trick#6: TCP early retransmit, per RFC5827.  To avoid spurious
2214          * retransmissions due to small network reorderings, we implement
2215          * Mitigation A.3 in the RFC and delay the retransmission for a short
2216          * interval if appropriate.
2217          */
2218         if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2219             (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2220             !tcp_may_send_now(sk))
2221                 return !tcp_pause_early_retransmit(sk, flag);
2222 
2223         return false;
2224 }
2225 
2226 /* Detect loss in event "A" above by marking head of queue up as lost.
2227  * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2228  * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2229  * has at least tp->reordering SACKed seqments above it; "packets" refers to
2230  * the maximum SACKed segments to pass before reaching this limit.
2231  */
2232 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2233 {
2234         struct tcp_sock *tp = tcp_sk(sk);
2235         struct sk_buff *skb;
2236         int cnt, oldcnt, lost;
2237         unsigned int mss;
2238         /* Use SACK to deduce losses of new sequences sent during recovery */
2239         const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2240 
2241         WARN_ON(packets > tp->packets_out);
2242         if (tp->lost_skb_hint) {
2243                 skb = tp->lost_skb_hint;
2244                 cnt = tp->lost_cnt_hint;
2245                 /* Head already handled? */
2246                 if (mark_head && skb != tcp_write_queue_head(sk))
2247                         return;
2248         } else {
2249                 skb = tcp_write_queue_head(sk);
2250                 cnt = 0;
2251         }
2252 
2253         tcp_for_write_queue_from(skb, sk) {
2254                 if (skb == tcp_send_head(sk))
2255                         break;
2256                 /* TODO: do this better */
2257                 /* this is not the most efficient way to do this... */
2258                 tp->lost_skb_hint = skb;
2259                 tp->lost_cnt_hint = cnt;
2260 
2261                 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2262                         break;
2263 
2264                 oldcnt = cnt;
2265                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2266                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2267                         cnt += tcp_skb_pcount(skb);
2268 
2269                 if (cnt > packets) {
2270                         if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2271                             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2272                             (oldcnt >= packets))
2273                                 break;
2274 
2275                         mss = tcp_skb_mss(skb);
2276                         /* If needed, chop off the prefix to mark as lost. */
2277                         lost = (packets - oldcnt) * mss;
2278                         if (lost < skb->len &&
2279                             tcp_fragment(sk, skb, lost, mss, GFP_ATOMIC) < 0)
2280                                 break;
2281                         cnt = packets;
2282                 }
2283 
2284                 tcp_skb_mark_lost(tp, skb);
2285 
2286                 if (mark_head)
2287                         break;
2288         }
2289         tcp_verify_left_out(tp);
2290 }
2291 
2292 /* Account newly detected lost packet(s) */
2293 
2294 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2295 {
2296         struct tcp_sock *tp = tcp_sk(sk);
2297 
2298         if (tcp_is_reno(tp)) {
2299                 tcp_mark_head_lost(sk, 1, 1);
2300         } else if (tcp_is_fack(tp)) {
2301                 int lost = tp->fackets_out - tp->reordering;
2302                 if (lost <= 0)
2303                         lost = 1;
2304                 tcp_mark_head_lost(sk, lost, 0);
2305         } else {
2306                 int sacked_upto = tp->sacked_out - tp->reordering;
2307                 if (sacked_upto >= 0)
2308                         tcp_mark_head_lost(sk, sacked_upto, 0);
2309                 else if (fast_rexmit)
2310                         tcp_mark_head_lost(sk, 1, 1);
2311         }
2312 }
2313 
2314 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2315 {
2316         return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2317                before(tp->rx_opt.rcv_tsecr, when);
2318 }
2319 
2320 /* skb is spurious retransmitted if the returned timestamp echo
2321  * reply is prior to the skb transmission time
2322  */
2323 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2324                                      const struct sk_buff *skb)
2325 {
2326         return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2327                tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2328 }
2329 
2330 /* Nothing was retransmitted or returned timestamp is less
2331  * than timestamp of the first retransmission.
2332  */
2333 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2334 {
2335         return !tp->retrans_stamp ||
2336                tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2337 }
2338 
2339 /* Undo procedures. */
2340 
2341 /* We can clear retrans_stamp when there are no retransmissions in the
2342  * window. It would seem that it is trivially available for us in
2343  * tp->retrans_out, however, that kind of assumptions doesn't consider
2344  * what will happen if errors occur when sending retransmission for the
2345  * second time. ...It could the that such segment has only
2346  * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2347  * the head skb is enough except for some reneging corner cases that
2348  * are not worth the effort.
2349  *
2350  * Main reason for all this complexity is the fact that connection dying
2351  * time now depends on the validity of the retrans_stamp, in particular,
2352  * that successive retransmissions of a segment must not advance
2353  * retrans_stamp under any conditions.
2354  */
2355 static bool tcp_any_retrans_done(const struct sock *sk)
2356 {
2357         const struct tcp_sock *tp = tcp_sk(sk);
2358         struct sk_buff *skb;
2359 
2360         if (tp->retrans_out)
2361                 return true;
2362 
2363         skb = tcp_write_queue_head(sk);
2364         if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2365                 return true;
2366 
2367         return false;
2368 }
2369 
2370 #if FASTRETRANS_DEBUG > 1
2371 static void DBGUNDO(struct sock *sk, const char *msg)
2372 {
2373         struct tcp_sock *tp = tcp_sk(sk);
2374         struct inet_sock *inet = inet_sk(sk);
2375 
2376         if (sk->sk_family == AF_INET) {
2377                 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2378                          msg,
2379                          &inet->inet_daddr, ntohs(inet->inet_dport),
2380                          tp->snd_cwnd, tcp_left_out(tp),
2381                          tp->snd_ssthresh, tp->prior_ssthresh,
2382                          tp->packets_out);
2383         }
2384 #if IS_ENABLED(CONFIG_IPV6)
2385         else if (sk->sk_family == AF_INET6) {
2386                 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2387                          msg,
2388                          &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2389                          tp->snd_cwnd, tcp_left_out(tp),
2390                          tp->snd_ssthresh, tp->prior_ssthresh,
2391                          tp->packets_out);
2392         }
2393 #endif
2394 }
2395 #else
2396 #define DBGUNDO(x...) do { } while (0)
2397 #endif
2398 
2399 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2400 {
2401         struct tcp_sock *tp = tcp_sk(sk);
2402 
2403         if (unmark_loss) {
2404                 struct sk_buff *skb;
2405 
2406                 tcp_for_write_queue(skb, sk) {
2407                         if (skb == tcp_send_head(sk))
2408                                 break;
2409                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2410                 }
2411                 tp->lost_out = 0;
2412                 tcp_clear_all_retrans_hints(tp);
2413         }
2414 
2415         if (tp->prior_ssthresh) {
2416                 const struct inet_connection_sock *icsk = inet_csk(sk);
2417 
2418                 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2419 
2420                 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2421                         tp->snd_ssthresh = tp->prior_ssthresh;
2422                         tcp_ecn_withdraw_cwr(tp);
2423                 }
2424         }
2425         tp->snd_cwnd_stamp = tcp_time_stamp;
2426         tp->undo_marker = 0;
2427 }
2428 
2429 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2430 {
2431         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2432 }
2433 
2434 /* People celebrate: "We love our President!" */
2435 static bool tcp_try_undo_recovery(struct sock *sk)
2436 {
2437         struct tcp_sock *tp = tcp_sk(sk);
2438 
2439         if (tcp_may_undo(tp)) {
2440                 int mib_idx;
2441 
2442                 /* Happy end! We did not retransmit anything
2443                  * or our original transmission succeeded.
2444                  */
2445                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2446                 tcp_undo_cwnd_reduction(sk, false);
2447                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2448                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2449                 else
2450                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2451 
2452                 NET_INC_STATS(sock_net(sk), mib_idx);
2453         }
2454         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2455                 /* Hold old state until something *above* high_seq
2456                  * is ACKed. For Reno it is MUST to prevent false
2457                  * fast retransmits (RFC2582). SACK TCP is safe. */
2458                 if (!tcp_any_retrans_done(sk))
2459                         tp->retrans_stamp = 0;
2460                 return true;
2461         }
2462         tcp_set_ca_state(sk, TCP_CA_Open);
2463         return false;
2464 }
2465 
2466 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2467 static bool tcp_try_undo_dsack(struct sock *sk)
2468 {
2469         struct tcp_sock *tp = tcp_sk(sk);
2470 
2471         if (tp->undo_marker && !tp->undo_retrans) {
2472                 DBGUNDO(sk, "D-SACK");
2473                 tcp_undo_cwnd_reduction(sk, false);
2474                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2475                 return true;
2476         }
2477         return false;
2478 }
2479 
2480 /* Undo during loss recovery after partial ACK or using F-RTO. */
2481 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2482 {
2483         struct tcp_sock *tp = tcp_sk(sk);
2484 
2485         if (frto_undo || tcp_may_undo(tp)) {
2486                 tcp_undo_cwnd_reduction(sk, true);
2487 
2488                 DBGUNDO(sk, "partial loss");
2489                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2490                 if (frto_undo)
2491                         NET_INC_STATS(sock_net(sk),
2492                                         LINUX_MIB_TCPSPURIOUSRTOS);
2493                 inet_csk(sk)->icsk_retransmits = 0;
2494                 if (frto_undo || tcp_is_sack(tp))
2495                         tcp_set_ca_state(sk, TCP_CA_Open);
2496                 return true;
2497         }
2498         return false;
2499 }
2500 
2501 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2502  * It computes the number of packets to send (sndcnt) based on packets newly
2503  * delivered:
2504  *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2505  *      cwnd reductions across a full RTT.
2506  *   2) Otherwise PRR uses packet conservation to send as much as delivered.
2507  *      But when the retransmits are acked without further losses, PRR
2508  *      slow starts cwnd up to ssthresh to speed up the recovery.
2509  */
2510 static void tcp_init_cwnd_reduction(struct sock *sk)
2511 {
2512         struct tcp_sock *tp = tcp_sk(sk);
2513 
2514         tp->high_seq = tp->snd_nxt;
2515         tp->tlp_high_seq = 0;
2516         tp->snd_cwnd_cnt = 0;
2517         tp->prior_cwnd = tp->snd_cwnd;
2518         tp->prr_delivered = 0;
2519         tp->prr_out = 0;
2520         tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2521         tcp_ecn_queue_cwr(tp);
2522 }
2523 
2524 static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked,
2525                                int flag)
2526 {
2527         struct tcp_sock *tp = tcp_sk(sk);
2528         int sndcnt = 0;
2529         int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2530 
2531         if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2532                 return;
2533 
2534         tp->prr_delivered += newly_acked_sacked;
2535         if (delta < 0) {
2536                 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2537                                tp->prior_cwnd - 1;
2538                 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2539         } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2540                    !(flag & FLAG_LOST_RETRANS)) {
2541                 sndcnt = min_t(int, delta,
2542                                max_t(int, tp->prr_delivered - tp->prr_out,
2543                                      newly_acked_sacked) + 1);
2544         } else {
2545                 sndcnt = min(delta, newly_acked_sacked);
2546         }
2547         /* Force a fast retransmit upon entering fast recovery */
2548         sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2549         tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2550 }
2551 
2552 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2553 {
2554         struct tcp_sock *tp = tcp_sk(sk);
2555 
2556         if (inet_csk(sk)->icsk_ca_ops->cong_control)
2557                 return;
2558 
2559         /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2560         if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2561             (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2562                 tp->snd_cwnd = tp->snd_ssthresh;
2563                 tp->snd_cwnd_stamp = tcp_time_stamp;
2564         }
2565         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2566 }
2567 
2568 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2569 void tcp_enter_cwr(struct sock *sk)
2570 {
2571         struct tcp_sock *tp = tcp_sk(sk);
2572 
2573         tp->prior_ssthresh = 0;
2574         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2575                 tp->undo_marker = 0;
2576                 tcp_init_cwnd_reduction(sk);
2577                 tcp_set_ca_state(sk, TCP_CA_CWR);
2578         }
2579 }
2580 EXPORT_SYMBOL(tcp_enter_cwr);
2581 
2582 static void tcp_try_keep_open(struct sock *sk)
2583 {
2584         struct tcp_sock *tp = tcp_sk(sk);
2585         int state = TCP_CA_Open;
2586 
2587         if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2588                 state = TCP_CA_Disorder;
2589 
2590         if (inet_csk(sk)->icsk_ca_state != state) {
2591                 tcp_set_ca_state(sk, state);
2592                 tp->high_seq = tp->snd_nxt;
2593         }
2594 }
2595 
2596 static void tcp_try_to_open(struct sock *sk, int flag)
2597 {
2598         struct tcp_sock *tp = tcp_sk(sk);
2599 
2600         tcp_verify_left_out(tp);
2601 
2602         if (!tcp_any_retrans_done(sk))
2603                 tp->retrans_stamp = 0;
2604 
2605         if (flag & FLAG_ECE)
2606                 tcp_enter_cwr(sk);
2607 
2608         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2609                 tcp_try_keep_open(sk);
2610         }
2611 }
2612 
2613 static void tcp_mtup_probe_failed(struct sock *sk)
2614 {
2615         struct inet_connection_sock *icsk = inet_csk(sk);
2616 
2617         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2618         icsk->icsk_mtup.probe_size = 0;
2619         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2620 }
2621 
2622 static void tcp_mtup_probe_success(struct sock *sk)
2623 {
2624         struct tcp_sock *tp = tcp_sk(sk);
2625         struct inet_connection_sock *icsk = inet_csk(sk);
2626 
2627         /* FIXME: breaks with very large cwnd */
2628         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2629         tp->snd_cwnd = tp->snd_cwnd *
2630                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2631                        icsk->icsk_mtup.probe_size;
2632         tp->snd_cwnd_cnt = 0;
2633         tp->snd_cwnd_stamp = tcp_time_stamp;
2634         tp->snd_ssthresh = tcp_current_ssthresh(sk);
2635 
2636         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2637         icsk->icsk_mtup.probe_size = 0;
2638         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2639         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2640 }
2641 
2642 /* Do a simple retransmit without using the backoff mechanisms in
2643  * tcp_timer. This is used for path mtu discovery.
2644  * The socket is already locked here.
2645  */
2646 void tcp_simple_retransmit(struct sock *sk)
2647 {
2648         const struct inet_connection_sock *icsk = inet_csk(sk);
2649         struct tcp_sock *tp = tcp_sk(sk);
2650         struct sk_buff *skb;
2651         unsigned int mss = tcp_current_mss(sk);
2652         u32 prior_lost = tp->lost_out;
2653 
2654         tcp_for_write_queue(skb, sk) {
2655                 if (skb == tcp_send_head(sk))
2656                         break;
2657                 if (tcp_skb_seglen(skb) > mss &&
2658                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2659                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2660                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2661                                 tp->retrans_out -= tcp_skb_pcount(skb);
2662                         }
2663                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2664                 }
2665         }
2666 
2667         tcp_clear_retrans_hints_partial(tp);
2668 
2669         if (prior_lost == tp->lost_out)
2670                 return;
2671 
2672         if (tcp_is_reno(tp))
2673                 tcp_limit_reno_sacked(tp);
2674 
2675         tcp_verify_left_out(tp);
2676 
2677         /* Don't muck with the congestion window here.
2678          * Reason is that we do not increase amount of _data_
2679          * in network, but units changed and effective
2680          * cwnd/ssthresh really reduced now.
2681          */
2682         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2683                 tp->high_seq = tp->snd_nxt;
2684                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2685                 tp->prior_ssthresh = 0;
2686                 tp->undo_marker = 0;
2687                 tcp_set_ca_state(sk, TCP_CA_Loss);
2688         }
2689         tcp_xmit_retransmit_queue(sk);
2690 }
2691 EXPORT_SYMBOL(tcp_simple_retransmit);
2692 
2693 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2694 {
2695         struct tcp_sock *tp = tcp_sk(sk);
2696         int mib_idx;
2697 
2698         if (tcp_is_reno(tp))
2699                 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2700         else
2701                 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2702 
2703         NET_INC_STATS(sock_net(sk), mib_idx);
2704 
2705         tp->prior_ssthresh = 0;
2706         tcp_init_undo(tp);
2707 
2708         if (!tcp_in_cwnd_reduction(sk)) {
2709                 if (!ece_ack)
2710                         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2711                 tcp_init_cwnd_reduction(sk);
2712         }
2713         tcp_set_ca_state(sk, TCP_CA_Recovery);
2714 }
2715 
2716 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2717  * recovered or spurious. Otherwise retransmits more on partial ACKs.
2718  */
2719 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2720                              int *rexmit)
2721 {
2722         struct tcp_sock *tp = tcp_sk(sk);
2723         bool recovered = !before(tp->snd_una, tp->high_seq);
2724 
2725         if ((flag & FLAG_SND_UNA_ADVANCED) &&
2726             tcp_try_undo_loss(sk, false))
2727                 return;
2728 
2729         if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2730                 /* Step 3.b. A timeout is spurious if not all data are
2731                  * lost, i.e., never-retransmitted data are (s)acked.
2732                  */
2733                 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2734                     tcp_try_undo_loss(sk, true))
2735                         return;
2736 
2737                 if (after(tp->snd_nxt, tp->high_seq)) {
2738                         if (flag & FLAG_DATA_SACKED || is_dupack)
2739                                 tp->frto = 0; /* Step 3.a. loss was real */
2740                 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2741                         tp->high_seq = tp->snd_nxt;
2742                         /* Step 2.b. Try send new data (but deferred until cwnd
2743                          * is updated in tcp_ack()). Otherwise fall back to
2744                          * the conventional recovery.
2745                          */
2746                         if (tcp_send_head(sk) &&
2747                             after(tcp_wnd_end(tp), tp->snd_nxt)) {
2748                                 *rexmit = REXMIT_NEW;
2749                                 return;
2750                         }
2751                         tp->frto = 0;
2752                 }
2753         }
2754 
2755         if (recovered) {
2756                 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2757                 tcp_try_undo_recovery(sk);
2758                 return;
2759         }
2760         if (tcp_is_reno(tp)) {
2761                 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2762                  * delivered. Lower inflight to clock out (re)tranmissions.
2763                  */
2764                 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2765                         tcp_add_reno_sack(sk);
2766                 else if (flag & FLAG_SND_UNA_ADVANCED)
2767                         tcp_reset_reno_sack(tp);
2768         }
2769         *rexmit = REXMIT_LOST;
2770 }
2771 
2772 /* Undo during fast recovery after partial ACK. */
2773 static bool tcp_try_undo_partial(struct sock *sk, const int acked)
2774 {
2775         struct tcp_sock *tp = tcp_sk(sk);
2776 
2777         if (tp->undo_marker && tcp_packet_delayed(tp)) {
2778                 /* Plain luck! Hole if filled with delayed
2779                  * packet, rather than with a retransmit.
2780                  */
2781                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2782 
2783                 /* We are getting evidence that the reordering degree is higher
2784                  * than we realized. If there are no retransmits out then we
2785                  * can undo. Otherwise we clock out new packets but do not
2786                  * mark more packets lost or retransmit more.
2787                  */
2788                 if (tp->retrans_out)
2789                         return true;
2790 
2791                 if (!tcp_any_retrans_done(sk))
2792                         tp->retrans_stamp = 0;
2793 
2794                 DBGUNDO(sk, "partial recovery");
2795                 tcp_undo_cwnd_reduction(sk, true);
2796                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2797                 tcp_try_keep_open(sk);
2798                 return true;
2799         }
2800         return false;
2801 }
2802 
2803 /* Process an event, which can update packets-in-flight not trivially.
2804  * Main goal of this function is to calculate new estimate for left_out,
2805  * taking into account both packets sitting in receiver's buffer and
2806  * packets lost by network.
2807  *
2808  * Besides that it updates the congestion state when packet loss or ECN
2809  * is detected. But it does not reduce the cwnd, it is done by the
2810  * congestion control later.
2811  *
2812  * It does _not_ decide what to send, it is made in function
2813  * tcp_xmit_retransmit_queue().
2814  */
2815 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2816                                   bool is_dupack, int *ack_flag, int *rexmit)
2817 {
2818         struct inet_connection_sock *icsk = inet_csk(sk);
2819         struct tcp_sock *tp = tcp_sk(sk);
2820         int fast_rexmit = 0, flag = *ack_flag;
2821         bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2822                                     (tcp_fackets_out(tp) > tp->reordering));
2823 
2824         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2825                 tp->sacked_out = 0;
2826         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2827                 tp->fackets_out = 0;
2828 
2829         /* Now state machine starts.
2830          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2831         if (flag & FLAG_ECE)
2832                 tp->prior_ssthresh = 0;
2833 
2834         /* B. In all the states check for reneging SACKs. */
2835         if (tcp_check_sack_reneging(sk, flag))
2836                 return;
2837 
2838         /* C. Check consistency of the current state. */
2839         tcp_verify_left_out(tp);
2840 
2841         /* D. Check state exit conditions. State can be terminated
2842          *    when high_seq is ACKed. */
2843         if (icsk->icsk_ca_state == TCP_CA_Open) {
2844                 WARN_ON(tp->retrans_out != 0);
2845                 tp->retrans_stamp = 0;
2846         } else if (!before(tp->snd_una, tp->high_seq)) {
2847                 switch (icsk->icsk_ca_state) {
2848                 case TCP_CA_CWR:
2849                         /* CWR is to be held something *above* high_seq
2850                          * is ACKed for CWR bit to reach receiver. */
2851                         if (tp->snd_una != tp->high_seq) {
2852                                 tcp_end_cwnd_reduction(sk);
2853                                 tcp_set_ca_state(sk, TCP_CA_Open);
2854                         }
2855                         break;
2856 
2857                 case TCP_CA_Recovery:
2858                         if (tcp_is_reno(tp))
2859                                 tcp_reset_reno_sack(tp);
2860                         if (tcp_try_undo_recovery(sk))
2861                                 return;
2862                         tcp_end_cwnd_reduction(sk);
2863                         break;
2864                 }
2865         }
2866 
2867         /* Use RACK to detect loss */
2868         if (sysctl_tcp_recovery & TCP_RACK_LOST_RETRANS &&
2869             tcp_rack_mark_lost(sk)) {
2870                 flag |= FLAG_LOST_RETRANS;
2871                 *ack_flag |= FLAG_LOST_RETRANS;
2872         }
2873 
2874         /* E. Process state. */
2875         switch (icsk->icsk_ca_state) {
2876         case TCP_CA_Recovery:
2877                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2878                         if (tcp_is_reno(tp) && is_dupack)
2879                                 tcp_add_reno_sack(sk);
2880                 } else {
2881                         if (tcp_try_undo_partial(sk, acked))
2882                                 return;
2883                         /* Partial ACK arrived. Force fast retransmit. */
2884                         do_lost = tcp_is_reno(tp) ||
2885                                   tcp_fackets_out(tp) > tp->reordering;
2886                 }
2887                 if (tcp_try_undo_dsack(sk)) {
2888                         tcp_try_keep_open(sk);
2889                         return;
2890                 }
2891                 break;
2892         case TCP_CA_Loss:
2893                 tcp_process_loss(sk, flag, is_dupack, rexmit);
2894                 if (icsk->icsk_ca_state != TCP_CA_Open &&
2895                     !(flag & FLAG_LOST_RETRANS))
2896                         return;
2897                 /* Change state if cwnd is undone or retransmits are lost */
2898         default:
2899                 if (tcp_is_reno(tp)) {
2900                         if (flag & FLAG_SND_UNA_ADVANCED)
2901                                 tcp_reset_reno_sack(tp);
2902                         if (is_dupack)
2903                                 tcp_add_reno_sack(sk);
2904                 }
2905 
2906                 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2907                         tcp_try_undo_dsack(sk);
2908 
2909                 if (!tcp_time_to_recover(sk, flag)) {
2910                         tcp_try_to_open(sk, flag);
2911                         return;
2912                 }
2913 
2914                 /* MTU probe failure: don't reduce cwnd */
2915                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2916                     icsk->icsk_mtup.probe_size &&
2917                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2918                         tcp_mtup_probe_failed(sk);
2919                         /* Restores the reduction we did in tcp_mtup_probe() */
2920                         tp->snd_cwnd++;
2921                         tcp_simple_retransmit(sk);
2922                         return;
2923                 }
2924 
2925                 /* Otherwise enter Recovery state */
2926                 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2927                 fast_rexmit = 1;
2928         }
2929 
2930         if (do_lost)
2931                 tcp_update_scoreboard(sk, fast_rexmit);
2932         *rexmit = REXMIT_LOST;
2933 }
2934 
2935 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2936 {
2937         struct tcp_sock *tp = tcp_sk(sk);
2938         u32 wlen = sysctl_tcp_min_rtt_wlen * HZ;
2939 
2940         minmax_running_min(&tp->rtt_min, wlen, tcp_time_stamp,
2941                            rtt_us ? : jiffies_to_usecs(1));
2942 }
2943 
2944 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2945                                       long seq_rtt_us, long sack_rtt_us,
2946                                       long ca_rtt_us)
2947 {
2948         const struct tcp_sock *tp = tcp_sk(sk);
2949 
2950         /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2951          * broken middle-boxes or peers may corrupt TS-ECR fields. But
2952          * Karn's algorithm forbids taking RTT if some retransmitted data
2953          * is acked (RFC6298).
2954          */
2955         if (seq_rtt_us < 0)
2956                 seq_rtt_us = sack_rtt_us;
2957 
2958         /* RTTM Rule: A TSecr value received in a segment is used to
2959          * update the averaged RTT measurement only if the segment
2960          * acknowledges some new data, i.e., only if it advances the
2961          * left edge of the send window.
2962          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2963          */
2964         if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2965             flag & FLAG_ACKED)
2966                 seq_rtt_us = ca_rtt_us = jiffies_to_usecs(tcp_time_stamp -
2967                                                           tp->rx_opt.rcv_tsecr);
2968         if (seq_rtt_us < 0)
2969                 return false;
2970 
2971         /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2972          * always taken together with ACK, SACK, or TS-opts. Any negative
2973          * values will be skipped with the seq_rtt_us < 0 check above.
2974          */
2975         tcp_update_rtt_min(sk, ca_rtt_us);
2976         tcp_rtt_estimator(sk, seq_rtt_us);
2977         tcp_set_rto(sk);
2978 
2979         /* RFC6298: only reset backoff on valid RTT measurement. */
2980         inet_csk(sk)->icsk_backoff = 0;
2981         return true;
2982 }
2983 
2984 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2985 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2986 {
2987         long rtt_us = -1L;
2988 
2989         if (req && !req->num_retrans && tcp_rsk(req)->snt_synack.v64) {
2990                 struct skb_mstamp now;
2991 
2992                 skb_mstamp_get(&now);
2993                 rtt_us = skb_mstamp_us_delta(&now, &tcp_rsk(req)->snt_synack);
2994         }
2995 
2996         tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us);
2997 }
2998 
2999 
3000 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3001 {
3002         const struct inet_connection_sock *icsk = inet_csk(sk);
3003 
3004         icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3005         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3006 }
3007 
3008 /* Restart timer after forward progress on connection.
3009  * RFC2988 recommends to restart timer to now+rto.
3010  */
3011 void tcp_rearm_rto(struct sock *sk)
3012 {
3013         const struct inet_connection_sock *icsk = inet_csk(sk);
3014         struct tcp_sock *tp = tcp_sk(sk);
3015 
3016         /* If the retrans timer is currently being used by Fast Open
3017          * for SYN-ACK retrans purpose, stay put.
3018          */
3019         if (tp->fastopen_rsk)
3020                 return;
3021 
3022         if (!tp->packets_out) {
3023                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3024         } else {
3025                 u32 rto = inet_csk(sk)->icsk_rto;
3026                 /* Offset the time elapsed after installing regular RTO */
3027                 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3028                     icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3029                         struct sk_buff *skb = tcp_write_queue_head(sk);
3030                         const u32 rto_time_stamp =
3031                                 tcp_skb_timestamp(skb) + rto;
3032                         s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3033                         /* delta may not be positive if the socket is locked
3034                          * when the retrans timer fires and is rescheduled.
3035                          */
3036                         if (delta > 0)
3037                                 rto = delta;
3038                 }
3039                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3040                                           TCP_RTO_MAX);
3041         }
3042 }
3043 
3044 /* This function is called when the delayed ER timer fires. TCP enters
3045  * fast recovery and performs fast-retransmit.
3046  */
3047 void tcp_resume_early_retransmit(struct sock *sk)
3048 {
3049         struct tcp_sock *tp = tcp_sk(sk);
3050 
3051         tcp_rearm_rto(sk);
3052 
3053         /* Stop if ER is disabled after the delayed ER timer is scheduled */
3054         if (!tp->do_early_retrans)
3055                 return;
3056 
3057         tcp_enter_recovery(sk, false);
3058         tcp_update_scoreboard(sk, 1);
3059         tcp_xmit_retransmit_queue(sk);
3060 }
3061 
3062 /* If we get here, the whole TSO packet has not been acked. */
3063 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3064 {
3065         struct tcp_sock *tp = tcp_sk(sk);
3066         u32 packets_acked;
3067 
3068         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3069 
3070         packets_acked = tcp_skb_pcount(skb);
3071         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3072                 return 0;
3073         packets_acked -= tcp_skb_pcount(skb);
3074 
3075         if (packets_acked) {
3076                 BUG_ON(tcp_skb_pcount(skb) == 0);
3077                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3078         }
3079 
3080         return packets_acked;
3081 }
3082 
3083 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3084                            u32 prior_snd_una)
3085 {
3086         const struct skb_shared_info *shinfo;
3087 
3088         /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3089         if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3090                 return;
3091 
3092         shinfo = skb_shinfo(skb);
3093         if (!before(shinfo->tskey, prior_snd_una) &&
3094             before(shinfo->tskey, tcp_sk(sk)->snd_una))
3095                 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3096 }
3097 
3098 /* Remove acknowledged frames from the retransmission queue. If our packet
3099  * is before the ack sequence we can discard it as it's confirmed to have
3100  * arrived at the other end.
3101  */
3102 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3103                                u32 prior_snd_una, int *acked,
3104                                struct tcp_sacktag_state *sack,
3105                                struct skb_mstamp *now)
3106 {
3107         const struct inet_connection_sock *icsk = inet_csk(sk);
3108         struct skb_mstamp first_ackt, last_ackt;
3109         struct tcp_sock *tp = tcp_sk(sk);
3110         u32 prior_sacked = tp->sacked_out;
3111         u32 reord = tp->packets_out;
3112         bool fully_acked = true;
3113         long sack_rtt_us = -1L;
3114         long seq_rtt_us = -1L;
3115         long ca_rtt_us = -1L;
3116         struct sk_buff *skb;
3117         u32 pkts_acked = 0;
3118         u32 last_in_flight = 0;
3119         bool rtt_update;
3120         int flag = 0;
3121 
3122         first_ackt.v64 = 0;
3123 
3124         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3125                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3126                 u8 sacked = scb->sacked;
3127                 u32 acked_pcount;
3128 
3129                 tcp_ack_tstamp(sk, skb, prior_snd_una);
3130 
3131                 /* Determine how many packets and what bytes were acked, tso and else */
3132                 if (after(scb->end_seq, tp->snd_una)) {
3133                         if (tcp_skb_pcount(skb) == 1 ||
3134                             !after(tp->snd_una, scb->seq))
3135                                 break;
3136 
3137                         acked_pcount = tcp_tso_acked(sk, skb);
3138                         if (!acked_pcount)
3139                                 break;
3140                         fully_acked = false;
3141                 } else {
3142                         /* Speedup tcp_unlink_write_queue() and next loop */
3143                         prefetchw(skb->next);
3144                         acked_pcount = tcp_skb_pcount(skb);
3145                 }
3146 
3147                 if (unlikely(sacked & TCPCB_RETRANS)) {
3148                         if (sacked & TCPCB_SACKED_RETRANS)
3149                                 tp->retrans_out -= acked_pcount;
3150                         flag |= FLAG_RETRANS_DATA_ACKED;
3151                 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3152                         last_ackt = skb->skb_mstamp;
3153                         WARN_ON_ONCE(last_ackt.v64 == 0);
3154                         if (!first_ackt.v64)
3155                                 first_ackt = last_ackt;
3156 
3157                         last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3158                         reord = min(pkts_acked, reord);
3159                         if (!after(scb->end_seq, tp->high_seq))
3160                                 flag |= FLAG_ORIG_SACK_ACKED;
3161                 }
3162 
3163                 if (sacked & TCPCB_SACKED_ACKED) {
3164                         tp->sacked_out -= acked_pcount;
3165                 } else if (tcp_is_sack(tp)) {
3166                         tp->delivered += acked_pcount;
3167                         if (!tcp_skb_spurious_retrans(tp, skb))
3168                                 tcp_rack_advance(tp, &skb->skb_mstamp, sacked);
3169                 }
3170                 if (sacked & TCPCB_LOST)
3171                         tp->lost_out -= acked_pcount;
3172 
3173                 tp->packets_out -= acked_pcount;
3174                 pkts_acked += acked_pcount;
3175                 tcp_rate_skb_delivered(sk, skb, sack->rate);
3176 
3177                 /* Initial outgoing SYN's get put onto the write_queue
3178                  * just like anything else we transmit.  It is not
3179                  * true data, and if we misinform our callers that
3180                  * this ACK acks real data, we will erroneously exit
3181                  * connection startup slow start one packet too
3182                  * quickly.  This is severely frowned upon behavior.
3183                  */
3184                 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3185                         flag |= FLAG_DATA_ACKED;
3186                 } else {
3187                         flag |= FLAG_SYN_ACKED;
3188                         tp->retrans_stamp = 0;
3189                 }
3190 
3191                 if (!fully_acked)
3192                         break;
3193 
3194                 tcp_unlink_write_queue(skb, sk);
3195                 sk_wmem_free_skb(sk, skb);
3196                 if (unlikely(skb == tp->retransmit_skb_hint))
3197                         tp->retransmit_skb_hint = NULL;
3198                 if (unlikely(skb == tp->lost_skb_hint))
3199                         tp->lost_skb_hint = NULL;
3200         }
3201 
3202         if (!skb)
3203                 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3204 
3205         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3206                 tp->snd_up = tp->snd_una;
3207 
3208         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3209                 flag |= FLAG_SACK_RENEGING;
3210 
3211         if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3212                 seq_rtt_us = skb_mstamp_us_delta(now, &first_ackt);
3213                 ca_rtt_us = skb_mstamp_us_delta(now, &last_ackt);
3214         }
3215         if (sack->first_sackt.v64) {
3216                 sack_rtt_us = skb_mstamp_us_delta(now, &sack->first_sackt);
3217                 ca_rtt_us = skb_mstamp_us_delta(now, &sack->last_sackt);
3218         }
3219         sack->rate->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet, or -1 */
3220         rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3221                                         ca_rtt_us);
3222 
3223         if (flag & FLAG_ACKED) {
3224                 tcp_rearm_rto(sk);
3225                 if (unlikely(icsk->icsk_mtup.probe_size &&
3226                              !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3227                         tcp_mtup_probe_success(sk);
3228                 }
3229 
3230                 if (tcp_is_reno(tp)) {
3231                         tcp_remove_reno_sacks(sk, pkts_acked);
3232                 } else {
3233                         int delta;
3234 
3235                         /* Non-retransmitted hole got filled? That's reordering */
3236                         if (reord < prior_fackets)
3237                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3238 
3239                         delta = tcp_is_fack(tp) ? pkts_acked :
3240                                                   prior_sacked - tp->sacked_out;
3241                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3242                 }
3243 
3244                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3245 
3246         } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3247                    sack_rtt_us > skb_mstamp_us_delta(now, &skb->skb_mstamp)) {
3248                 /* Do not re-arm RTO if the sack RTT is measured from data sent
3249                  * after when the head was last (re)transmitted. Otherwise the
3250                  * timeout may continue to extend in loss recovery.
3251                  */
3252                 tcp_rearm_rto(sk);
3253         }
3254 
3255         if (icsk->icsk_ca_ops->pkts_acked) {
3256                 struct ack_sample sample = { .pkts_acked = pkts_acked,
3257                                              .rtt_us = ca_rtt_us,
3258                                              .in_flight = last_in_flight };
3259 
3260                 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3261         }
3262 
3263 #if FASTRETRANS_DEBUG > 0
3264         WARN_ON((int)tp->sacked_out < 0);
3265         WARN_ON((int)tp->lost_out < 0);
3266         WARN_ON((int)tp->retrans_out < 0);
3267         if (!tp->packets_out && tcp_is_sack(tp)) {
3268                 icsk = inet_csk(sk);
3269                 if (tp->lost_out) {
3270                         pr_debug("Leak l=%u %d\n",
3271                                  tp->lost_out, icsk->icsk_ca_state);
3272                         tp->lost_out = 0;
3273                 }
3274                 if (tp->sacked_out) {
3275                         pr_debug("Leak s=%u %d\n",
3276                                  tp->sacked_out, icsk->icsk_ca_state);
3277                         tp->sacked_out = 0;
3278                 }
3279                 if (tp->retrans_out) {
3280                         pr_debug("Leak r=%u %d\n",
3281                                  tp->retrans_out, icsk->icsk_ca_state);
3282                         tp->retrans_out = 0;
3283                 }
3284         }
3285 #endif
3286         *acked = pkts_acked;
3287         return flag;
3288 }
3289 
3290 static void tcp_ack_probe(struct sock *sk)
3291 {
3292         const struct tcp_sock *tp = tcp_sk(sk);
3293         struct inet_connection_sock *icsk = inet_csk(sk);
3294 
3295         /* Was it a usable window open? */
3296 
3297         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3298                 icsk->icsk_backoff = 0;
3299                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3300                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3301                  * This function is not for random using!
3302                  */
3303         } else {
3304                 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3305 
3306                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3307                                           when, TCP_RTO_MAX);
3308         }
3309 }
3310 
3311 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3312 {
3313         return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3314                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3315 }
3316 
3317 /* Decide wheather to run the increase function of congestion control. */
3318 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3319 {
3320         /* If reordering is high then always grow cwnd whenever data is
3321          * delivered regardless of its ordering. Otherwise stay conservative
3322          * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3323          * new SACK or ECE mark may first advance cwnd here and later reduce
3324          * cwnd in tcp_fastretrans_alert() based on more states.
3325          */
3326         if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3327                 return flag & FLAG_FORWARD_PROGRESS;
3328 
3329         return flag & FLAG_DATA_ACKED;
3330 }
3331 
3332 /* The "ultimate" congestion control function that aims to replace the rigid
3333  * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3334  * It's called toward the end of processing an ACK with precise rate
3335  * information. All transmission or retransmission are delayed afterwards.
3336  */
3337 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3338                              int flag, const struct rate_sample *rs)
3339 {
3340         const struct inet_connection_sock *icsk = inet_csk(sk);
3341 
3342         if (icsk->icsk_ca_ops->cong_control) {
3343                 icsk->icsk_ca_ops->cong_control(sk, rs);
3344                 return;
3345         }
3346 
3347         if (tcp_in_cwnd_reduction(sk)) {
3348                 /* Reduce cwnd if state mandates */
3349                 tcp_cwnd_reduction(sk, acked_sacked, flag);
3350         } else if (tcp_may_raise_cwnd(sk, flag)) {
3351                 /* Advance cwnd if state allows */
3352                 tcp_cong_avoid(sk, ack, acked_sacked);
3353         }
3354         tcp_update_pacing_rate(sk);
3355 }
3356 
3357 /* Check that window update is acceptable.
3358  * The function assumes that snd_una<=ack<=snd_next.
3359  */
3360 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3361                                         const u32 ack, const u32 ack_seq,
3362                                         const u32 nwin)
3363 {
3364         return  after(ack, tp->snd_una) ||
3365                 after(ack_seq, tp->snd_wl1) ||
3366                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3367 }
3368 
3369 /* If we update tp->snd_una, also update tp->bytes_acked */
3370 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3371 {
3372         u32 delta = ack - tp->snd_una;
3373 
3374         sock_owned_by_me((struct sock *)tp);
3375         tp->bytes_acked += delta;
3376         tp->snd_una = ack;
3377 }
3378 
3379 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3380 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3381 {
3382         u32 delta = seq - tp->rcv_nxt;
3383 
3384         sock_owned_by_me((struct sock *)tp);
3385         tp->bytes_received += delta;
3386         tp->rcv_nxt = seq;
3387 }
3388 
3389 /* Update our send window.
3390  *
3391  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3392  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3393  */
3394 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3395                                  u32 ack_seq)
3396 {
3397         struct tcp_sock *tp = tcp_sk(sk);
3398         int flag = 0;
3399         u32 nwin = ntohs(tcp_hdr(skb)->window);
3400 
3401         if (likely(!tcp_hdr(skb)->syn))
3402                 nwin <<= tp->rx_opt.snd_wscale;
3403 
3404         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3405                 flag |= FLAG_WIN_UPDATE;
3406                 tcp_update_wl(tp, ack_seq);
3407 
3408                 if (tp->snd_wnd != nwin) {
3409                         tp->snd_wnd = nwin;
3410 
3411                         /* Note, it is the only place, where
3412                          * fast path is recovered for sending TCP.
3413                          */
3414                         tp->pred_flags = 0;
3415                         tcp_fast_path_check(sk);
3416 
3417                         if (tcp_send_head(sk))
3418                                 tcp_slow_start_after_idle_check(sk);
3419 
3420                         if (nwin > tp->max_window) {
3421                                 tp->max_window = nwin;
3422                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3423                         }
3424                 }
3425         }
3426 
3427         tcp_snd_una_update(tp, ack);
3428 
3429         return flag;
3430 }
3431 
3432 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3433                                    u32 *last_oow_ack_time)
3434 {
3435         if (*last_oow_ack_time) {
3436                 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3437 
3438                 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3439                         NET_INC_STATS(net, mib_idx);
3440                         return true;    /* rate-limited: don't send yet! */
3441                 }
3442         }
3443 
3444         *last_oow_ack_time = tcp_time_stamp;
3445 
3446         return false;   /* not rate-limited: go ahead, send dupack now! */
3447 }
3448 
3449 /* Return true if we're currently rate-limiting out-of-window ACKs and
3450  * thus shouldn't send a dupack right now. We rate-limit dupacks in
3451  * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3452  * attacks that send repeated SYNs or ACKs for the same connection. To
3453  * do this, we do not send a duplicate SYNACK or ACK if the remote
3454  * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3455  */
3456 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3457                           int mib_idx, u32 *last_oow_ack_time)
3458 {
3459         /* Data packets without SYNs are not likely part of an ACK loop. */
3460         if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3461             !tcp_hdr(skb)->syn)
3462                 return false;
3463 
3464         return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3465 }
3466 
3467 /* RFC 5961 7 [ACK Throttling] */
3468 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3469 {
3470         /* unprotected vars, we dont care of overwrites */
3471         static u32 challenge_timestamp;
3472         static unsigned int challenge_count;
3473         struct tcp_sock *tp = tcp_sk(sk);
3474         u32 count, now;
3475 
3476         /* First check our per-socket dupack rate limit. */
3477         if (__tcp_oow_rate_limited(sock_net(sk),
3478                                    LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3479                                    &tp->last_oow_ack_time))
3480                 return;
3481 
3482         /* Then check host-wide RFC 5961 rate limit. */
3483         now = jiffies / HZ;
3484         if (now != challenge_timestamp) {
3485                 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3486 
3487                 challenge_timestamp = now;
3488                 WRITE_ONCE(challenge_count, half +
3489                            prandom_u32_max(sysctl_tcp_challenge_ack_limit));
3490         }
3491         count = READ_ONCE(challenge_count);
3492         if (count > 0) {
3493                 WRITE_ONCE(challenge_count, count - 1);
3494                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3495                 tcp_send_ack(sk);
3496         }
3497 }
3498 
3499 static void tcp_store_ts_recent(struct tcp_sock *tp)
3500 {
3501         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3502         tp->rx_opt.ts_recent_stamp = get_seconds();
3503 }
3504 
3505 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3506 {
3507         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3508                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3509                  * extra check below makes sure this can only happen
3510                  * for pure ACK frames.  -DaveM
3511                  *
3512                  * Not only, also it occurs for expired timestamps.
3513                  */
3514 
3515                 if (tcp_paws_check(&tp->rx_opt, 0))
3516                         tcp_store_ts_recent(tp);
3517         }
3518 }
3519 
3520 /* This routine deals with acks during a TLP episode.
3521  * We mark the end of a TLP episode on receiving TLP dupack or when
3522  * ack is after tlp_high_seq.
3523  * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3524  */
3525 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3526 {
3527         struct tcp_sock *tp = tcp_sk(sk);
3528 
3529         if (before(ack, tp->tlp_high_seq))
3530                 return;
3531 
3532         if (flag & FLAG_DSACKING_ACK) {
3533                 /* This DSACK means original and TLP probe arrived; no loss */
3534                 tp->tlp_high_seq = 0;
3535         } else if (after(ack, tp->tlp_high_seq)) {
3536                 /* ACK advances: there was a loss, so reduce cwnd. Reset
3537                  * tlp_high_seq in tcp_init_cwnd_reduction()
3538                  */
3539                 tcp_init_cwnd_reduction(sk);
3540                 tcp_set_ca_state(sk, TCP_CA_CWR);
3541                 tcp_end_cwnd_reduction(sk);
3542                 tcp_try_keep_open(sk);
3543                 NET_INC_STATS(sock_net(sk),
3544                                 LINUX_MIB_TCPLOSSPROBERECOVERY);
3545         } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3546                              FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3547                 /* Pure dupack: original and TLP probe arrived; no loss */
3548                 tp->tlp_high_seq = 0;
3549         }
3550 }
3551 
3552 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3553 {
3554         const struct inet_connection_sock *icsk = inet_csk(sk);
3555 
3556         if (icsk->icsk_ca_ops->in_ack_event)
3557                 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3558 }
3559 
3560 /* Congestion control has updated the cwnd already. So if we're in
3561  * loss recovery then now we do any new sends (for FRTO) or
3562  * retransmits (for CA_Loss or CA_recovery) that make sense.
3563  */
3564 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3565 {
3566         struct tcp_sock *tp = tcp_sk(sk);
3567 
3568         if (rexmit == REXMIT_NONE)
3569                 return;
3570 
3571         if (unlikely(rexmit == 2)) {
3572                 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3573                                           TCP_NAGLE_OFF);
3574                 if (after(tp->snd_nxt, tp->high_seq))
3575                         return;
3576                 tp->frto = 0;
3577         }
3578         tcp_xmit_retransmit_queue(sk);
3579 }
3580 
3581 /* This routine deals with incoming acks, but not outgoing ones. */
3582 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3583 {
3584         struct inet_connection_sock *icsk = inet_csk(sk);
3585         struct tcp_sock *tp = tcp_sk(sk);
3586         struct tcp_sacktag_state sack_state;
3587         struct rate_sample rs = { .prior_delivered = 0 };
3588         u32 prior_snd_una = tp->snd_una;
3589         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3590         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3591         bool is_dupack = false;
3592         u32 prior_fackets;
3593         int prior_packets = tp->packets_out;
3594         u32 delivered = tp->delivered;
3595         u32 lost = tp->lost;
3596         int acked = 0; /* Number of packets newly acked */
3597         int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3598         struct skb_mstamp now;
3599 
3600         sack_state.first_sackt.v64 = 0;
3601         sack_state.rate = &rs;
3602 
3603         /* We very likely will need to access write queue head. */
3604         prefetchw(sk->sk_write_queue.next);
3605 
3606         /* If the ack is older than previous acks
3607          * then we can probably ignore it.
3608          */
3609         if (before(ack, prior_snd_una)) {
3610                 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3611                 if (before(ack, prior_snd_una - tp->max_window)) {
3612                         tcp_send_challenge_ack(sk, skb);
3613                         return -1;
3614                 }
3615                 goto old_ack;
3616         }
3617 
3618         /* If the ack includes data we haven't sent yet, discard
3619          * this segment (RFC793 Section 3.9).
3620          */
3621         if (after(ack, tp->snd_nxt))
3622                 goto invalid_ack;
3623 
3624         skb_mstamp_get(&now);
3625 
3626         if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3627             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3628                 tcp_rearm_rto(sk);
3629 
3630         if (after(ack, prior_snd_una)) {
3631                 flag |= FLAG_SND_UNA_ADVANCED;
3632                 icsk->icsk_retransmits = 0;
3633         }
3634 
3635         prior_fackets = tp->fackets_out;
3636         rs.prior_in_flight = tcp_packets_in_flight(tp);
3637 
3638         /* ts_recent update must be made after we are sure that the packet
3639          * is in window.
3640          */
3641         if (flag & FLAG_UPDATE_TS_RECENT)
3642                 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3643 
3644         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3645                 /* Window is constant, pure forward advance.
3646                  * No more checks are required.
3647                  * Note, we use the fact that SND.UNA>=SND.WL2.
3648                  */
3649                 tcp_update_wl(tp, ack_seq);
3650                 tcp_snd_una_update(tp, ack);
3651                 flag |= FLAG_WIN_UPDATE;
3652 
3653                 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3654 
3655                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3656         } else {
3657                 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3658 
3659                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3660                         flag |= FLAG_DATA;
3661                 else
3662                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3663 
3664                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3665 
3666                 if (TCP_SKB_CB(skb)->sacked)
3667                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3668                                                         &sack_state);
3669 
3670                 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3671                         flag |= FLAG_ECE;
3672                         ack_ev_flags |= CA_ACK_ECE;
3673                 }
3674 
3675                 if (flag & FLAG_WIN_UPDATE)
3676                         ack_ev_flags |= CA_ACK_WIN_UPDATE;
3677 
3678                 tcp_in_ack_event(sk, ack_ev_flags);
3679         }
3680 
3681         /* We passed data and got it acked, remove any soft error
3682          * log. Something worked...
3683          */
3684         sk->sk_err_soft = 0;
3685         icsk->icsk_probes_out = 0;
3686         tp->rcv_tstamp = tcp_time_stamp;
3687         if (!prior_packets)
3688                 goto no_queue;
3689 
3690         /* See if we can take anything off of the retransmit queue. */
3691         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, &acked,
3692                                     &sack_state, &now);
3693 
3694         if (tcp_ack_is_dubious(sk, flag)) {
3695                 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3696                 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3697         }
3698         if (tp->tlp_high_seq)
3699                 tcp_process_tlp_ack(sk, ack, flag);
3700 
3701         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3702                 struct dst_entry *dst = __sk_dst_get(sk);
3703                 if (dst)
3704                         dst_confirm(dst);
3705         }
3706 
3707         if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3708                 tcp_schedule_loss_probe(sk);
3709         delivered = tp->delivered - delivered;  /* freshly ACKed or SACKed */
3710         lost = tp->lost - lost;                 /* freshly marked lost */
3711         tcp_rate_gen(sk, delivered, lost, &now, &rs);
3712         tcp_cong_control(sk, ack, delivered, flag, &rs);
3713         tcp_xmit_recovery(sk, rexmit);
3714         return 1;
3715 
3716 no_queue:
3717         /* If data was DSACKed, see if we can undo a cwnd reduction. */
3718         if (flag & FLAG_DSACKING_ACK)
3719                 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3720         /* If this ack opens up a zero window, clear backoff.  It was
3721          * being used to time the probes, and is probably far higher than
3722          * it needs to be for normal retransmission.
3723          */
3724         if (tcp_send_head(sk))
3725                 tcp_ack_probe(sk);
3726 
3727         if (tp->tlp_high_seq)
3728                 tcp_process_tlp_ack(sk, ack, flag);
3729         return 1;
3730 
3731 invalid_ack:
3732         SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3733         return -1;
3734 
3735 old_ack:
3736         /* If data was SACKed, tag it and see if we should send more data.
3737          * If data was DSACKed, see if we can undo a cwnd reduction.
3738          */
3739         if (TCP_SKB_CB(skb)->sacked) {
3740                 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3741                                                 &sack_state);
3742                 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3743                 tcp_xmit_recovery(sk, rexmit);
3744         }
3745 
3746         SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3747         return 0;
3748 }
3749 
3750 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3751                                       bool syn, struct tcp_fastopen_cookie *foc,
3752                                       bool exp_opt)
3753 {
3754         /* Valid only in SYN or SYN-ACK with an even length.  */
3755         if (!foc || !syn || len < 0 || (len & 1))
3756                 return;
3757 
3758         if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3759             len <= TCP_FASTOPEN_COOKIE_MAX)
3760                 memcpy(foc->val, cookie, len);
3761         else if (len != 0)
3762                 len = -1;
3763         foc->len = len;
3764         foc->exp = exp_opt;
3765 }
3766 
3767 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3768  * But, this can also be called on packets in the established flow when
3769  * the fast version below fails.
3770  */
3771 void tcp_parse_options(const struct sk_buff *skb,
3772                        struct tcp_options_received *opt_rx, int estab,
3773                        struct tcp_fastopen_cookie *foc)
3774 {
3775         const unsigned char *ptr;
3776         const struct tcphdr *th = tcp_hdr(skb);
3777         int length = (th->doff * 4) - sizeof(struct tcphdr);
3778 
3779         ptr = (const unsigned char *)(th + 1);
3780         opt_rx->saw_tstamp = 0;
3781 
3782         while (length > 0) {
3783                 int opcode = *ptr++;
3784                 int opsize;
3785 
3786                 switch (opcode) {
3787                 case TCPOPT_EOL:
3788                         return;
3789                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3790                         length--;
3791                         continue;
3792                 default:
3793                         opsize = *ptr++;
3794                         if (opsize < 2) /* "silly options" */
3795                                 return;
3796                         if (opsize > length)
3797                                 return; /* don't parse partial options */
3798                         switch (opcode) {
3799                         case TCPOPT_MSS:
3800                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3801                                         u16 in_mss = get_unaligned_be16(ptr);
3802                                         if (in_mss) {
3803                                                 if (opt_rx->user_mss &&
3804                                                     opt_rx->user_mss < in_mss)
3805                                                         in_mss = opt_rx->user_mss;
3806                                                 opt_rx->mss_clamp = in_mss;
3807                                         }
3808                                 }
3809                                 break;
3810                         case TCPOPT_WINDOW:
3811                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3812                                     !estab && sysctl_tcp_window_scaling) {
3813                                         __u8 snd_wscale = *(__u8 *)ptr;
3814                                         opt_rx->wscale_ok = 1;
3815                                         if (snd_wscale > 14) {
3816                                                 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3817                                                                      __func__,
3818                                                                      snd_wscale);
3819                                                 snd_wscale = 14;
3820                                         }
3821                                         opt_rx->snd_wscale = snd_wscale;
3822                                 }
3823                                 break;
3824                         case TCPOPT_TIMESTAMP:
3825                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3826                                     ((estab && opt_rx->tstamp_ok) ||
3827                                      (!estab && sysctl_tcp_timestamps))) {
3828                                         opt_rx->saw_tstamp = 1;
3829                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3830                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3831                                 }
3832                                 break;
3833                         case TCPOPT_SACK_PERM:
3834                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3835                                     !estab && sysctl_tcp_sack) {
3836                                         opt_rx->sack_ok = TCP_SACK_SEEN;
3837                                         tcp_sack_reset(opt_rx);
3838                                 }
3839                                 break;
3840 
3841                         case TCPOPT_SACK:
3842                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3843                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3844                                    opt_rx->sack_ok) {
3845                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3846                                 }
3847                                 break;
3848 #ifdef CONFIG_TCP_MD5SIG
3849                         case TCPOPT_MD5SIG:
3850                                 /*
3851                                  * The MD5 Hash has already been
3852                                  * checked (see tcp_v{4,6}_do_rcv()).
3853                                  */
3854                                 break;
3855 #endif
3856                         case TCPOPT_FASTOPEN:
3857                                 tcp_parse_fastopen_option(
3858                                         opsize - TCPOLEN_FASTOPEN_BASE,
3859                                         ptr, th->syn, foc, false);
3860                                 break;
3861 
3862                         case TCPOPT_EXP:
3863                                 /* Fast Open option shares code 254 using a
3864                                  * 16 bits magic number.
3865                                  */
3866                                 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3867                                     get_unaligned_be16(ptr) ==
3868                                     TCPOPT_FASTOPEN_MAGIC)
3869                                         tcp_parse_fastopen_option(opsize -
3870                                                 TCPOLEN_EXP_FASTOPEN_BASE,
3871                                                 ptr + 2, th->syn, foc, true);
3872                                 break;
3873 
3874                         }
3875                         ptr += opsize-2;
3876                         length -= opsize;
3877                 }
3878         }
3879 }
3880 EXPORT_SYMBOL(tcp_parse_options);
3881 
3882 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3883 {
3884         const __be32 *ptr = (const __be32 *)(th + 1);
3885 
3886         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3887                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3888                 tp->rx_opt.saw_tstamp = 1;
3889                 ++ptr;
3890                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3891                 ++ptr;
3892                 if (*ptr)
3893                         tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3894                 else
3895                         tp->rx_opt.rcv_tsecr = 0;
3896                 return true;
3897         }
3898         return false;
3899 }
3900 
3901 /* Fast parse options. This hopes to only see timestamps.
3902  * If it is wrong it falls back on tcp_parse_options().
3903  */
3904 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3905                                    const struct tcphdr *th, struct tcp_sock *tp)
3906 {
3907         /* In the spirit of fast parsing, compare doff directly to constant
3908          * values.  Because equality is used, short doff can be ignored here.
3909          */
3910         if (th->doff == (sizeof(*th) / 4)) {
3911                 tp->rx_opt.saw_tstamp = 0;
3912                 return false;
3913         } else if (tp->rx_opt.tstamp_ok &&
3914                    th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3915                 if (tcp_parse_aligned_timestamp(tp, th))
3916                         return true;
3917         }
3918 
3919         tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3920         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3921                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3922 
3923         return true;
3924 }
3925 
3926 #ifdef CONFIG_TCP_MD5SIG
3927 /*
3928  * Parse MD5 Signature option
3929  */
3930 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3931 {
3932         int length = (th->doff << 2) - sizeof(*th);
3933         const u8 *ptr = (const u8 *)(th + 1);
3934 
3935         /* If the TCP option is too short, we can short cut */
3936         if (length < TCPOLEN_MD5SIG)
3937                 return NULL;
3938 
3939         while (length > 0) {
3940                 int opcode = *ptr++;
3941                 int opsize;
3942 
3943                 switch (opcode) {
3944                 case TCPOPT_EOL:
3945                         return NULL;
3946                 case TCPOPT_NOP:
3947                         length--;
3948                         continue;
3949                 default:
3950                         opsize = *ptr++;
3951                         if (opsize < 2 || opsize > length)
3952                                 return NULL;
3953                         if (opcode == TCPOPT_MD5SIG)
3954                                 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3955                 }
3956                 ptr += opsize - 2;
3957                 length -= opsize;
3958         }
3959         return NULL;
3960 }
3961 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3962 #endif
3963 
3964 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3965  *
3966  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3967  * it can pass through stack. So, the following predicate verifies that
3968  * this segment is not used for anything but congestion avoidance or
3969  * fast retransmit. Moreover, we even are able to eliminate most of such
3970  * second order effects, if we apply some small "replay" window (~RTO)
3971  * to timestamp space.
3972  *
3973  * All these measures still do not guarantee that we reject wrapped ACKs
3974  * on networks with high bandwidth, when sequence space is recycled fastly,
3975  * but it guarantees that such events will be very rare and do not affect
3976  * connection seriously. This doesn't look nice, but alas, PAWS is really
3977  * buggy extension.
3978  *
3979  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3980  * states that events when retransmit arrives after original data are rare.
3981  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3982  * the biggest problem on large power networks even with minor reordering.
3983  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3984  * up to bandwidth of 18Gigabit/sec. 8) ]
3985  */
3986 
3987 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3988 {
3989         const struct tcp_sock *tp = tcp_sk(sk);
3990         const struct tcphdr *th = tcp_hdr(skb);
3991         u32 seq = TCP_SKB_CB(skb)->seq;
3992         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3993 
3994         return (/* 1. Pure ACK with correct sequence number. */
3995                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3996 
3997                 /* 2. ... and duplicate ACK. */
3998                 ack == tp->snd_una &&
3999 
4000                 /* 3. ... and does not update window. */
4001                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4002 
4003                 /* 4. ... and sits in replay window. */
4004                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4005 }
4006 
4007 static inline bool tcp_paws_discard(const struct sock *sk,
4008                                    const struct sk_buff *skb)
4009 {
4010         const struct tcp_sock *tp = tcp_sk(sk);
4011 
4012         return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4013                !tcp_disordered_ack(sk, skb);
4014 }
4015 
4016 /* Check segment sequence number for validity.
4017  *
4018  * Segment controls are considered valid, if the segment
4019  * fits to the window after truncation to the window. Acceptability
4020  * of data (and SYN, FIN, of course) is checked separately.
4021  * See tcp_data_queue(), for example.
4022  *
4023  * Also, controls (RST is main one) are accepted using RCV.WUP instead
4024  * of RCV.NXT. Peer still did not advance his SND.UNA when we
4025  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4026  * (borrowed from freebsd)
4027  */
4028 
4029 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4030 {
4031         return  !before(end_seq, tp->rcv_wup) &&
4032                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4033 }
4034 
4035 /* When we get a reset we do this. */
4036 void tcp_reset(struct sock *sk)
4037 {
4038         /* We want the right error as BSD sees it (and indeed as we do). */
4039         switch (sk->sk_state) {
4040         case TCP_SYN_SENT:
4041                 sk->sk_err = ECONNREFUSED;
4042                 break;
4043         case TCP_CLOSE_WAIT:
4044                 sk->sk_err = EPIPE;
4045                 break;
4046         case TCP_CLOSE:
4047                 return;
4048         default:
4049                 sk->sk_err = ECONNRESET;
4050         }
4051         /* This barrier is coupled with smp_rmb() in tcp_poll() */
4052         smp_wmb();
4053 
4054         if (!sock_flag(sk, SOCK_DEAD))
4055                 sk->sk_error_report(sk);
4056 
4057         tcp_done(sk);
4058 }
4059 
4060 /*
4061  *      Process the FIN bit. This now behaves as it is supposed to work
4062  *      and the FIN takes effect when it is validly part of sequence
4063  *      space. Not before when we get holes.
4064  *
4065  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4066  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
4067  *      TIME-WAIT)
4068  *
4069  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
4070  *      close and we go into CLOSING (and later onto TIME-WAIT)
4071  *
4072  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4073  */
4074 void tcp_fin(struct sock *sk)
4075 {
4076         struct tcp_sock *tp = tcp_sk(sk);
4077 
4078         inet_csk_schedule_ack(sk);
4079 
4080         sk->sk_shutdown |= RCV_SHUTDOWN;
4081         sock_set_flag(sk, SOCK_DONE);
4082 
4083         switch (sk->sk_state) {
4084         case TCP_SYN_RECV:
4085         case TCP_ESTABLISHED:
4086                 /* Move to CLOSE_WAIT */
4087                 tcp_set_state(sk, TCP_CLOSE_WAIT);
4088                 inet_csk(sk)->icsk_ack.pingpong = 1;
4089                 break;
4090 
4091         case TCP_CLOSE_WAIT:
4092         case TCP_CLOSING:
4093                 /* Received a retransmission of the FIN, do
4094                  * nothing.
4095                  */
4096                 break;
4097         case TCP_LAST_ACK:
4098                 /* RFC793: Remain in the LAST-ACK state. */
4099                 break;
4100 
4101         case TCP_FIN_WAIT1:
4102                 /* This case occurs when a simultaneous close
4103                  * happens, we must ack the received FIN and
4104                  * enter the CLOSING state.
4105                  */
4106                 tcp_send_ack(sk);
4107                 tcp_set_state(sk, TCP_CLOSING);
4108                 break;
4109         case TCP_FIN_WAIT2:
4110                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4111                 tcp_send_ack(sk);
4112                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4113                 break;
4114         default:
4115                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4116                  * cases we should never reach this piece of code.
4117                  */
4118                 pr_err("%s: Impossible, sk->sk_state=%d\n",
4119                        __func__, sk->sk_state);
4120                 break;
4121         }
4122 
4123         /* It _is_ possible, that we have something out-of-order _after_ FIN.
4124          * Probably, we should reset in this case. For now drop them.
4125          */
4126         skb_rbtree_purge(&tp->out_of_order_queue);
4127         if (tcp_is_sack(tp))
4128                 tcp_sack_reset(&tp->rx_opt);
4129         sk_mem_reclaim(sk);
4130 
4131         if (!sock_flag(sk, SOCK_DEAD)) {
4132                 sk->sk_state_change(sk);
4133 
4134                 /* Do not send POLL_HUP for half duplex close. */
4135                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4136                     sk->sk_state == TCP_CLOSE)
4137                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4138                 else
4139                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4140         }
4141 }
4142 
4143 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4144                                   u32 end_seq)
4145 {
4146         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4147                 if (before(seq, sp->start_seq))
4148                         sp->start_seq = seq;
4149                 if (after(end_seq, sp->end_seq))
4150                         sp->end_seq = end_seq;
4151                 return true;
4152         }
4153         return false;
4154 }
4155 
4156 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4157 {
4158         struct tcp_sock *tp = tcp_sk(sk);
4159 
4160         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4161                 int mib_idx;
4162 
4163                 if (before(seq, tp->rcv_nxt))
4164                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4165                 else
4166                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4167 
4168                 NET_INC_STATS(sock_net(sk), mib_idx);
4169 
4170                 tp->rx_opt.dsack = 1;
4171                 tp->duplicate_sack[0].start_seq = seq;
4172                 tp->duplicate_sack[0].end_seq = end_seq;
4173         }
4174 }
4175 
4176 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4177 {
4178         struct tcp_sock *tp = tcp_sk(sk);
4179 
4180         if (!tp->rx_opt.dsack)
4181                 tcp_dsack_set(sk, seq, end_seq);
4182         else
4183                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4184 }
4185 
4186 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4187 {
4188         struct tcp_sock *tp = tcp_sk(sk);
4189 
4190         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4191             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4192                 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4193                 tcp_enter_quickack_mode(sk);
4194 
4195                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4196                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4197 
4198                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4199                                 end_seq = tp->rcv_nxt;
4200                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4201                 }
4202         }
4203 
4204         tcp_send_ack(sk);
4205 }
4206 
4207 /* These routines update the SACK block as out-of-order packets arrive or
4208  * in-order packets close up the sequence space.
4209  */
4210 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4211 {
4212         int this_sack;
4213         struct tcp_sack_block *sp = &tp->selective_acks[0];
4214         struct tcp_sack_block *swalk = sp + 1;
4215 
4216         /* See if the recent change to the first SACK eats into
4217          * or hits the sequence space of other SACK blocks, if so coalesce.
4218          */
4219         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4220                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4221                         int i;
4222 
4223                         /* Zap SWALK, by moving every further SACK up by one slot.
4224                          * Decrease num_sacks.
4225                          */
4226                         tp->rx_opt.num_sacks--;
4227                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4228                                 sp[i] = sp[i + 1];
4229                         continue;
4230                 }
4231                 this_sack++, swalk++;
4232         }
4233 }
4234 
4235 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4236 {
4237         struct tcp_sock *tp = tcp_sk(sk);
4238         struct tcp_sack_block *sp = &tp->selective_acks[0];
4239         int cur_sacks = tp->rx_opt.num_sacks;
4240         int this_sack;
4241 
4242         if (!cur_sacks)
4243                 goto new_sack;
4244 
4245         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4246                 if (tcp_sack_extend(sp, seq, end_seq)) {
4247                         /* Rotate this_sack to the first one. */
4248                         for (; this_sack > 0; this_sack--, sp--)
4249                                 swap(*sp, *(sp - 1));
4250                         if (cur_sacks > 1)
4251                                 tcp_sack_maybe_coalesce(tp);
4252                         return;
4253                 }
4254         }
4255 
4256         /* Could not find an adjacent existing SACK, build a new one,
4257          * put it at the front, and shift everyone else down.  We
4258          * always know there is at least one SACK present already here.
4259          *
4260          * If the sack array is full, forget about the last one.
4261          */
4262         if (this_sack >= TCP_NUM_SACKS) {
4263                 this_sack--;
4264                 tp->rx_opt.num_sacks--;
4265                 sp--;
4266         }
4267         for (; this_sack > 0; this_sack--, sp--)
4268                 *sp = *(sp - 1);
4269 
4270 new_sack:
4271         /* Build the new head SACK, and we're done. */
4272         sp->start_seq = seq;
4273         sp->end_seq = end_seq;
4274         tp->rx_opt.num_sacks++;
4275 }
4276 
4277 /* RCV.NXT advances, some SACKs should be eaten. */
4278 
4279 static void tcp_sack_remove(struct tcp_sock *tp)
4280 {
4281         struct tcp_sack_block *sp = &tp->selective_acks[0];
4282         int num_sacks = tp->rx_opt.num_sacks;
4283         int this_sack;
4284 
4285         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4286         if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4287                 tp->rx_opt.num_sacks = 0;
4288                 return;
4289         }
4290 
4291         for (this_sack = 0; this_sack < num_sacks;) {
4292                 /* Check if the start of the sack is covered by RCV.NXT. */
4293                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4294                         int i;
4295 
4296                         /* RCV.NXT must cover all the block! */
4297                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4298 
4299                         /* Zap this SACK, by moving forward any other SACKS. */
4300                         for (i = this_sack+1; i < num_sacks; i++)
4301                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4302                         num_sacks--;
4303                         continue;
4304                 }
4305                 this_sack++;
4306                 sp++;
4307         }
4308         tp->rx_opt.num_sacks = num_sacks;
4309 }
4310 
4311 /**
4312  * tcp_try_coalesce - try to merge skb to prior one
4313  * @sk: socket
4314  * @to: prior buffer
4315  * @from: buffer to add in queue
4316  * @fragstolen: pointer to boolean
4317  *
4318  * Before queueing skb @from after @to, try to merge them
4319  * to reduce overall memory use and queue lengths, if cost is small.
4320  * Packets in ofo or receive queues can stay a long time.
4321  * Better try to coalesce them right now to avoid future collapses.
4322  * Returns true if caller should free @from instead of queueing it
4323  */
4324 static bool tcp_try_coalesce(struct sock *sk,
4325                              struct sk_buff *to,
4326                              struct sk_buff *from,
4327                              bool *fragstolen)
4328 {
4329         int delta;
4330 
4331         *fragstolen = false;
4332 
4333         /* Its possible this segment overlaps with prior segment in queue */
4334         if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4335                 return false;
4336 
4337         if (!skb_try_coalesce(to, from, fragstolen, &delta))
4338                 return false;
4339 
4340         atomic_add(delta, &sk->sk_rmem_alloc);
4341         sk_mem_charge(sk, delta);
4342         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4343         TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4344         TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4345         TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4346         return true;
4347 }
4348 
4349 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4350 {
4351         sk_drops_add(sk, skb);
4352         __kfree_skb(skb);
4353 }
4354 
4355 /* This one checks to see if we can put data from the
4356  * out_of_order queue into the receive_queue.
4357  */
4358 static void tcp_ofo_queue(struct sock *sk)
4359 {
4360         struct tcp_sock *tp = tcp_sk(sk);
4361         __u32 dsack_high = tp->rcv_nxt;
4362         bool fin, fragstolen, eaten;
4363         struct sk_buff *skb, *tail;
4364         struct rb_node *p;
4365 
4366         p = rb_first(&tp->out_of_order_queue);
4367         while (p) {
4368                 skb = rb_entry(p, struct sk_buff, rbnode);
4369                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4370                         break;
4371 
4372                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4373                         __u32 dsack = dsack_high;
4374                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4375                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4376                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4377                 }
4378                 p = rb_next(p);
4379                 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4380 
4381                 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4382                         SOCK_DEBUG(sk, "ofo packet was already received\n");
4383                         tcp_drop(sk, skb);
4384                         continue;
4385                 }
4386                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4387                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4388                            TCP_SKB_CB(skb)->end_seq);
4389 
4390                 tail = skb_peek_tail(&sk->sk_receive_queue);
4391                 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4392                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4393                 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4394                 if (!eaten)
4395                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4396                 else
4397                         kfree_skb_partial(skb, fragstolen);
4398 
4399                 if (unlikely(fin)) {
4400                         tcp_fin(sk);
4401                         /* tcp_fin() purges tp->out_of_order_queue,
4402                          * so we must end this loop right now.
4403                          */
4404                         break;
4405                 }
4406         }
4407 }
4408 
4409 static bool tcp_prune_ofo_queue(struct sock *sk);
4410 static int tcp_prune_queue(struct sock *sk);
4411 
4412 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4413                                  unsigned int size)
4414 {
4415         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4416             !sk_rmem_schedule(sk, skb, size)) {
4417 
4418                 if (tcp_prune_queue(sk) < 0)
4419                         return -1;
4420 
4421                 while (!sk_rmem_schedule(sk, skb, size)) {
4422                         if (!tcp_prune_ofo_queue(sk))
4423                                 return -1;
4424                 }
4425         }
4426         return 0;
4427 }
4428 
4429 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4430 {
4431         struct tcp_sock *tp = tcp_sk(sk);
4432         struct rb_node **p, *q, *parent;
4433         struct sk_buff *skb1;
4434         u32 seq, end_seq;
4435         bool fragstolen;
4436 
4437         tcp_ecn_check_ce(tp, skb);
4438 
4439         if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4440                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4441                 tcp_drop(sk, skb);
4442                 return;
4443         }
4444 
4445         /* Disable header prediction. */
4446         tp->pred_flags = 0;
4447         inet_csk_schedule_ack(sk);
4448 
4449         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4450         seq = TCP_SKB_CB(skb)->seq;
4451         end_seq = TCP_SKB_CB(skb)->end_seq;
4452         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4453                    tp->rcv_nxt, seq, end_seq);
4454 
4455         p = &tp->out_of_order_queue.rb_node;
4456         if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4457                 /* Initial out of order segment, build 1 SACK. */
4458                 if (tcp_is_sack(tp)) {
4459                         tp->rx_opt.num_sacks = 1;
4460                         tp->selective_acks[0].start_seq = seq;
4461                         tp->selective_acks[0].end_seq = end_seq;
4462                 }
4463                 rb_link_node(&skb->rbnode, NULL, p);
4464                 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4465                 tp->ooo_last_skb = skb;
4466                 goto end;
4467         }
4468 
4469         /* In the typical case, we are adding an skb to the end of the list.
4470          * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4471          */
4472         if (tcp_try_coalesce(sk, tp->ooo_last_skb, skb, &fragstolen)) {
4473 coalesce_done:
4474                 tcp_grow_window(sk, skb);
4475                 kfree_skb_partial(skb, fragstolen);
4476                 skb = NULL;
4477                 goto add_sack;
4478         }
4479         /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4480         if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4481                 parent = &tp->ooo_last_skb->rbnode;
4482                 p = &parent->rb_right;
4483                 goto insert;
4484         }
4485 
4486         /* Find place to insert this segment. Handle overlaps on the way. */
4487         parent = NULL;
4488         while (*p) {
4489                 parent = *p;
4490                 skb1 = rb_entry(parent, struct sk_buff, rbnode);
4491                 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4492                         p = &parent->rb_left;
4493                         continue;
4494                 }
4495                 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4496                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4497                                 /* All the bits are present. Drop. */
4498                                 NET_INC_STATS(sock_net(sk),
4499                                               LINUX_MIB_TCPOFOMERGE);
4500                                 __kfree_skb(skb);
4501                                 skb = NULL;
4502                                 tcp_dsack_set(sk, seq, end_seq);
4503                                 goto add_sack;
4504                         }
4505                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4506                                 /* Partial overlap. */
4507                                 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4508                         } else {
4509                                 /* skb's seq == skb1's seq and skb covers skb1.
4510                                  * Replace skb1 with skb.
4511                                  */
4512                                 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4513                                                 &tp->out_of_order_queue);
4514                                 tcp_dsack_extend(sk,
4515                                                  TCP_SKB_CB(skb1)->seq,
4516                                                  TCP_SKB_CB(skb1)->end_seq);
4517                                 NET_INC_STATS(sock_net(sk),
4518                                               LINUX_MIB_TCPOFOMERGE);
4519                                 __kfree_skb(skb1);
4520                                 goto merge_right;
4521                         }
4522                 } else if (tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4523                         goto coalesce_done;
4524                 }
4525                 p = &parent->rb_right;
4526         }
4527 insert:
4528         /* Insert segment into RB tree. */
4529         rb_link_node(&skb->rbnode, parent, p);
4530         rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4531 
4532 merge_right:
4533         /* Remove other segments covered by skb. */
4534         while ((q = rb_next(&skb->rbnode)) != NULL) {
4535                 skb1 = rb_entry(q, struct sk_buff, rbnode);
4536 
4537                 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4538                         break;
4539                 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4540                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4541                                          end_seq);
4542                         break;
4543                 }
4544                 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4545                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4546                                  TCP_SKB_CB(skb1)->end_seq);
4547                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4548                 tcp_drop(sk, skb1);
4549         }
4550         /* If there is no skb after us, we are the last_skb ! */
4551         if (!q)
4552                 tp->ooo_last_skb = skb;
4553 
4554 add_sack:
4555         if (tcp_is_sack(tp))
4556                 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4557 end:
4558         if (skb) {
4559                 tcp_grow_window(sk, skb);
4560                 skb_set_owner_r(skb, sk);
4561         }
4562 }
4563 
4564 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4565                   bool *fragstolen)
4566 {
4567         int eaten;
4568         struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4569 
4570         __skb_pull(skb, hdrlen);
4571         eaten = (tail &&
4572                  tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4573         tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4574         if (!eaten) {
4575                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4576                 skb_set_owner_r(skb, sk);
4577         }
4578         return eaten;
4579 }
4580 
4581 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4582 {
4583         struct sk_buff *skb;
4584         int err = -ENOMEM;
4585         int data_len = 0;
4586         bool fragstolen;
4587 
4588         if (size == 0)
4589                 return 0;
4590 
4591         if (size > PAGE_SIZE) {
4592                 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4593 
4594                 data_len = npages << PAGE_SHIFT;
4595                 size = data_len + (size & ~PAGE_MASK);
4596         }
4597         skb = alloc_skb_with_frags(size - data_len, data_len,
4598                                    PAGE_ALLOC_COSTLY_ORDER,
4599                                    &err, sk->sk_allocation);
4600         if (!skb)
4601                 goto err;
4602 
4603         skb_put(skb, size - data_len);
4604         skb->data_len = data_len;
4605         skb->len = size;
4606 
4607         if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4608                 goto err_free;
4609 
4610         err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4611         if (err)
4612                 goto err_free;
4613 
4614         TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4615         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4616         TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4617 
4618         if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4619                 WARN_ON_ONCE(fragstolen); /* should not happen */
4620                 __kfree_skb(skb);
4621         }
4622         return size;
4623 
4624 err_free:
4625         kfree_skb(skb);
4626 err:
4627         return err;
4628 
4629 }
4630 
4631 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4632 {
4633         struct tcp_sock *tp = tcp_sk(sk);
4634         bool fragstolen = false;
4635         int eaten = -1;
4636 
4637         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4638                 __kfree_skb(skb);
4639                 return;
4640         }
4641         skb_dst_drop(skb);
4642         __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4643 
4644         tcp_ecn_accept_cwr(tp, skb);
4645 
4646         tp->rx_opt.dsack = 0;
4647 
4648         /*  Queue data for delivery to the user.
4649          *  Packets in sequence go to the receive queue.
4650          *  Out of sequence packets to the out_of_order_queue.
4651          */
4652         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4653                 if (tcp_receive_window(tp) == 0)
4654                         goto out_of_window;
4655 
4656                 /* Ok. In sequence. In window. */
4657                 if (tp->ucopy.task == current &&
4658                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4659                     sock_owned_by_user(sk) && !tp->urg_data) {
4660                         int chunk = min_t(unsigned int, skb->len,
4661                                           tp->ucopy.len);
4662 
4663                         __set_current_state(TASK_RUNNING);
4664 
4665                         if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4666                                 tp->ucopy.len -= chunk;
4667                                 tp->copied_seq += chunk;
4668                                 eaten = (chunk == skb->len);
4669                                 tcp_rcv_space_adjust(sk);
4670                         }
4671                 }
4672 
4673                 if (eaten <= 0) {
4674 queue_and_out:
4675                         if (eaten < 0) {
4676                                 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4677                                         sk_forced_mem_schedule(sk, skb->truesize);
4678                                 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4679                                         goto drop;
4680                         }
4681                         eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4682                 }
4683                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4684                 if (skb->len)
4685                         tcp_event_data_recv(sk, skb);
4686                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4687                         tcp_fin(sk);
4688 
4689                 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4690                         tcp_ofo_queue(sk);
4691 
4692                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4693                          * gap in queue is filled.
4694                          */
4695                         if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4696                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4697                 }
4698 
4699                 if (tp->rx_opt.num_sacks)
4700                         tcp_sack_remove(tp);
4701 
4702                 tcp_fast_path_check(sk);
4703 
4704                 if (eaten > 0)
4705                         kfree_skb_partial(skb, fragstolen);
4706                 if (!sock_flag(sk, SOCK_DEAD))
4707                         sk->sk_data_ready(sk);
4708                 return;
4709         }
4710 
4711         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4712                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4713                 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4714                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4715 
4716 out_of_window:
4717                 tcp_enter_quickack_mode(sk);
4718                 inet_csk_schedule_ack(sk);
4719 drop:
4720                 tcp_drop(sk, skb);
4721                 return;
4722         }
4723 
4724         /* Out of window. F.e. zero window probe. */
4725         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4726                 goto out_of_window;
4727 
4728         tcp_enter_quickack_mode(sk);
4729 
4730         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4731                 /* Partial packet, seq < rcv_next < end_seq */
4732                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4733                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4734                            TCP_SKB_CB(skb)->end_seq);
4735 
4736                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4737 
4738                 /* If window is closed, drop tail of packet. But after
4739                  * remembering D-SACK for its head made in previous line.
4740                  */
4741                 if (!tcp_receive_window(tp))
4742                         goto out_of_window;
4743                 goto queue_and_out;
4744         }
4745 
4746         tcp_data_queue_ofo(sk, skb);
4747 }
4748 
4749 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4750 {
4751         if (list)
4752                 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4753 
4754         return rb_entry_safe(rb_next(&skb->rbnode), struct sk_buff, rbnode);
4755 }
4756 
4757 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4758                                         struct sk_buff_head *list,
4759                                         struct rb_root *root)
4760 {
4761         struct sk_buff *next = tcp_skb_next(skb, list);
4762 
4763         if (list)
4764                 __skb_unlink(skb, list);
4765         else
4766                 rb_erase(&skb->rbnode, root);
4767 
4768         __kfree_skb(skb);
4769         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4770 
4771         return next;
4772 }
4773 
4774 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4775 static void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4776 {
4777         struct rb_node **p = &root->rb_node;
4778         struct rb_node *parent = NULL;
4779         struct sk_buff *skb1;
4780 
4781         while (*p) {
4782                 parent = *p;
4783                 skb1 = rb_entry(parent, struct sk_buff, rbnode);
4784                 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4785                         p = &parent->rb_left;
4786                 else
4787                         p = &parent->rb_right;
4788         }
4789         rb_link_node(&skb->rbnode, parent, p);
4790         rb_insert_color(&skb->rbnode, root);
4791 }
4792 
4793 /* Collapse contiguous sequence of skbs head..tail with
4794  * sequence numbers start..end.
4795  *
4796  * If tail is NULL, this means until the end of the queue.
4797  *
4798  * Segments with FIN/SYN are not collapsed (only because this
4799  * simplifies code)
4800  */
4801 static void
4802 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4803              struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4804 {
4805         struct sk_buff *skb = head, *n;
4806         struct sk_buff_head tmp;
4807         bool end_of_skbs;
4808 
4809         /* First, check that queue is collapsible and find
4810          * the point where collapsing can be useful.
4811          */
4812 restart:
4813         for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4814                 n = tcp_skb_next(skb, list);
4815 
4816                 /* No new bits? It is possible on ofo queue. */
4817                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4818                         skb = tcp_collapse_one(sk, skb, list, root);
4819                         if (!skb)
4820                                 break;
4821                         goto restart;
4822                 }
4823 
4824                 /* The first skb to collapse is:
4825                  * - not SYN/FIN and
4826                  * - bloated or contains data before "start" or
4827                  *   overlaps to the next one.
4828                  */
4829                 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4830                     (tcp_win_from_space(skb->truesize) > skb->len ||
4831                      before(TCP_SKB_CB(skb)->seq, start))) {
4832                         end_of_skbs = false;
4833                         break;
4834                 }
4835 
4836                 if (n && n != tail &&
4837                     TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4838                         end_of_skbs = false;
4839                         break;
4840                 }
4841 
4842                 /* Decided to skip this, advance start seq. */
4843                 start = TCP_SKB_CB(skb)->end_seq;
4844         }
4845         if (end_of_skbs ||
4846             (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4847                 return;
4848 
4849         __skb_queue_head_init(&tmp);
4850 
4851         while (before(start, end)) {
4852                 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4853                 struct sk_buff *nskb;
4854 
4855                 nskb = alloc_skb(copy, GFP_ATOMIC);
4856                 if (!nskb)
4857                         break;
4858 
4859                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4860                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4861                 if (list)
4862                         __skb_queue_before(list, skb, nskb);
4863                 else
4864                         __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4865                 skb_set_owner_r(nskb, sk);
4866 
4867                 /* Copy data, releasing collapsed skbs. */
4868                 while (copy > 0) {
4869                         int offset = start - TCP_SKB_CB(skb)->seq;
4870                         int size = TCP_SKB_CB(skb)->end_seq - start;
4871 
4872                         BUG_ON(offset < 0);
4873                         if (size > 0) {
4874                                 size = min(copy, size);
4875                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4876                                         BUG();
4877                                 TCP_SKB_CB(nskb)->end_seq += size;
4878                                 copy -= size;
4879                                 start += size;
4880                         }
4881                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4882                                 skb = tcp_collapse_one(sk, skb, list, root);
4883                                 if (!skb ||
4884                                     skb == tail ||
4885                                     (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4886                                         goto end;
4887                         }
4888                 }
4889         }
4890 end:
4891         skb_queue_walk_safe(&tmp, skb, n)
4892                 tcp_rbtree_insert(root, skb);
4893 }
4894 
4895 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4896  * and tcp_collapse() them until all the queue is collapsed.
4897  */
4898 static void tcp_collapse_ofo_queue(struct sock *sk)
4899 {
4900         struct tcp_sock *tp = tcp_sk(sk);
4901         struct sk_buff *skb, *head;
4902         struct rb_node *p;
4903         u32 start, end;
4904 
4905         p = rb_first(&tp->out_of_order_queue);
4906         skb = rb_entry_safe(p, struct sk_buff, rbnode);
4907 new_range:
4908         if (!skb) {
4909                 p = rb_last(&tp->out_of_order_queue);
4910                 /* Note: This is possible p is NULL here. We do not
4911                  * use rb_entry_safe(), as ooo_last_skb is valid only
4912                  * if rbtree is not empty.
4913                  */
4914                 tp->ooo_last_skb = rb_entry(p, struct sk_buff, rbnode);
4915                 return;
4916         }
4917         start = TCP_SKB_CB(skb)->seq;
4918         end = TCP_SKB_CB(skb)->end_seq;
4919 
4920         for (head = skb;;) {
4921                 skb = tcp_skb_next(skb, NULL);
4922 
4923                 /* Range is terminated when we see a gap or when
4924                  * we are at the queue end.
4925                  */
4926                 if (!skb ||
4927                     after(TCP_SKB_CB(skb)->seq, end) ||
4928                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4929                         tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4930                                      head, skb, start, end);
4931                         goto new_range;
4932                 }
4933 
4934                 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4935                         start = TCP_SKB_CB(skb)->seq;
4936                 if (after(TCP_SKB_CB(skb)->end_seq, end))
4937                         end = TCP_SKB_CB(skb)->end_seq;
4938         }
4939 }
4940 
4941 /*
4942  * Clean the out-of-order queue to make room.
4943  * We drop high sequences packets to :
4944  * 1) Let a chance for holes to be filled.
4945  * 2) not add too big latencies if thousands of packets sit there.
4946  *    (But if application shrinks SO_RCVBUF, we could still end up
4947  *     freeing whole queue here)
4948  *
4949  * Return true if queue has shrunk.
4950  */
4951 static bool tcp_prune_ofo_queue(struct sock *sk)
4952 {
4953         struct tcp_sock *tp = tcp_sk(sk);
4954         struct rb_node *node, *prev;
4955 
4956         if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4957                 return false;
4958 
4959         NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
4960         node = &tp->ooo_last_skb->rbnode;
4961         do {
4962                 prev = rb_prev(node);
4963                 rb_erase(node, &tp->out_of_order_queue);
4964                 tcp_drop(sk, rb_entry(node, struct sk_buff, rbnode));
4965                 sk_mem_reclaim(sk);
4966                 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
4967                     !tcp_under_memory_pressure(sk))
4968                         break;
4969                 node = prev;
4970         } while (node);
4971         tp->ooo_last_skb = rb_entry(prev, struct sk_buff, rbnode);
4972 
4973         /* Reset SACK state.  A conforming SACK implementation will
4974          * do the same at a timeout based retransmit.  When a connection
4975          * is in a sad state like this, we care only about integrity
4976          * of the connection not performance.
4977          */
4978         if (tp->rx_opt.sack_ok)
4979                 tcp_sack_reset(&tp->rx_opt);
4980         return true;
4981 }
4982 
4983 /* Reduce allocated memory if we can, trying to get
4984  * the socket within its memory limits again.
4985  *
4986  * Return less than zero if we should start dropping frames
4987  * until the socket owning process reads some of the data
4988  * to stabilize the situation.
4989  */
4990 static int tcp_prune_queue(struct sock *sk)
4991 {
4992         struct tcp_sock *tp = tcp_sk(sk);
4993 
4994         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4995 
4996         NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
4997 
4998         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4999                 tcp_clamp_window(sk);
5000         else if (tcp_under_memory_pressure(sk))
5001                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5002 
5003         tcp_collapse_ofo_queue(sk);
5004         if (!skb_queue_empty(&sk->sk_receive_queue))
5005                 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5006                              skb_peek(&sk->sk_receive_queue),
5007                              NULL,
5008                              tp->copied_seq, tp->rcv_nxt);
5009         sk_mem_reclaim(sk);
5010 
5011         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5012                 return 0;
5013 
5014         /* Collapsing did not help, destructive actions follow.
5015          * This must not ever occur. */
5016 
5017         tcp_prune_ofo_queue(sk);
5018 
5019         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5020                 return 0;
5021 
5022         /* If we are really being abused, tell the caller to silently
5023          * drop receive data on the floor.  It will get retransmitted
5024          * and hopefully then we'll have sufficient space.
5025          */
5026         NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5027 
5028         /* Massive buffer overcommit. */
5029         tp->pred_flags = 0;
5030         return -1;
5031 }
5032 
5033 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5034 {
5035         const struct tcp_sock *tp = tcp_sk(sk);
5036 
5037         /* If the user specified a specific send buffer setting, do
5038          * not modify it.
5039          */
5040         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5041                 return false;
5042 
5043         /* If we are under global TCP memory pressure, do not expand.  */
5044         if (tcp_under_memory_pressure(sk))
5045                 return false;
5046 
5047         /* If we are under soft global TCP memory pressure, do not expand.  */
5048         if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5049                 return false;
5050 
5051         /* If we filled the congestion window, do not expand.  */
5052         if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5053                 return false;
5054 
5055         return true;
5056 }
5057 
5058 /* When incoming ACK allowed to free some skb from write_queue,
5059  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5060  * on the exit from tcp input handler.
5061  *
5062  * PROBLEM: sndbuf expansion does not work well with largesend.
5063  */
5064 static void tcp_new_space(struct sock *sk)
5065 {
5066         struct tcp_sock *tp = tcp_sk(sk);
5067 
5068         if (tcp_should_expand_sndbuf(sk)) {
5069                 tcp_sndbuf_expand(sk);
5070                 tp->snd_cwnd_stamp = tcp_time_stamp;
5071         }
5072 
5073         sk->sk_write_space(sk);
5074 }
5075 
5076 static void tcp_check_space(struct sock *sk)
5077 {
5078         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5079                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5080                 /* pairs with tcp_poll() */
5081                 smp_mb();
5082                 if (sk->sk_socket &&
5083                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5084                         tcp_new_space(sk);
5085                         if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5086                                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5087                 }
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         if (sk_acceptq_is_full(sk)) {
6322                 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6323                 goto drop;
6324         }
6325 
6326         req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6327         if (!req)
6328                 goto drop;
6329 
6330         tcp_rsk(req)->af_specific = af_ops;
6331         tcp_rsk(req)->ts_off = 0;
6332 
6333         tcp_clear_options(&tmp_opt);
6334         tmp_opt.mss_clamp = af_ops->mss_clamp;
6335         tmp_opt.user_mss  = tp->rx_opt.user_mss;
6336         tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6337 
6338         if (want_cookie && !tmp_opt.saw_tstamp)
6339                 tcp_clear_options(&tmp_opt);
6340 
6341         tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6342         tcp_openreq_init(req, &tmp_opt, skb, sk);
6343         inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6344 
6345         /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6346         inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6347 
6348         af_ops->init_req(req, sk, skb);
6349 
6350         if (security_inet_conn_request(sk, skb, req))
6351                 goto drop_and_free;
6352 
6353         if (isn && tmp_opt.tstamp_ok)
6354                 af_ops->init_seq(skb, &tcp_rsk(req)->ts_off);
6355 
6356         if (!want_cookie && !isn) {
6357                 /* VJ's idea. We save last timestamp seen
6358                  * from the destination in peer table, when entering
6359                  * state TIME-WAIT, and check against it before
6360                  * accepting new connection request.
6361                  *
6362                  * If "isn" is not zero, this request hit alive
6363                  * timewait bucket, so that all the necessary checks
6364                  * are made in the function processing timewait state.
6365                  */
6366                 if (tcp_death_row.sysctl_tw_recycle) {
6367                         bool strict;
6368 
6369                         dst = af_ops->route_req(sk, &fl, req, &strict);
6370 
6371                         if (dst && strict &&
6372                             !tcp_peer_is_proven(req, dst, true,
6373                                                 tmp_opt.saw_tstamp)) {
6374                                 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
6375                                 goto drop_and_release;
6376                         }
6377                 }
6378                 /* Kill the following clause, if you dislike this way. */
6379                 else if (!net->ipv4.sysctl_tcp_syncookies &&
6380                          (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6381                           (sysctl_max_syn_backlog >> 2)) &&
6382                          !tcp_peer_is_proven(req, dst, false,
6383                                              tmp_opt.saw_tstamp)) {
6384                         /* Without syncookies last quarter of
6385                          * backlog is filled with destinations,
6386                          * proven to be alive.
6387                          * It means that we continue to communicate
6388                          * to destinations, already remembered
6389                          * to the moment of synflood.
6390                          */
6391                         pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6392                                     rsk_ops->family);
6393                         goto drop_and_release;
6394                 }
6395 
6396                 isn = af_ops->init_seq(skb, &tcp_rsk(req)->ts_off);
6397         }
6398         if (!dst) {
6399                 dst = af_ops->route_req(sk, &fl, req, NULL);
6400                 if (!dst)
6401                         goto drop_and_free;
6402         }
6403 
6404         tcp_ecn_create_request(req, skb, sk, dst);
6405 
6406         if (want_cookie) {
6407                 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6408                 tcp_rsk(req)->ts_off = 0;
6409                 req->cookie_ts = tmp_opt.tstamp_ok;
6410                 if (!tmp_opt.tstamp_ok)
6411                         inet_rsk(req)->ecn_ok = 0;
6412         }
6413 
6414         tcp_rsk(req)->snt_isn = isn;
6415         tcp_rsk(req)->txhash = net_tx_rndhash();
6416         tcp_openreq_init_rwin(req, sk, dst);
6417         if (!want_cookie) {
6418                 tcp_reqsk_record_syn(sk, req, skb);
6419                 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6420         }
6421         if (fastopen_sk) {
6422                 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6423                                     &foc, TCP_SYNACK_FASTOPEN);
6424                 /* Add the child socket directly into the accept queue */
6425                 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6426                 sk->sk_data_ready(sk);
6427                 bh_unlock_sock(fastopen_sk);
6428                 sock_put(fastopen_sk);
6429         } else {
6430                 tcp_rsk(req)->tfo_listener = false;
6431                 if (!want_cookie)
6432                         inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6433                 af_ops->send_synack(sk, dst, &fl, req, &foc,
6434                                     !want_cookie ? TCP_SYNACK_NORMAL :
6435                                                    TCP_SYNACK_COOKIE);
6436                 if (want_cookie) {
6437                         reqsk_free(req);
6438                         return 0;
6439                 }
6440         }
6441         reqsk_put(req);
6442         return 0;
6443 
6444 drop_and_release:
6445         dst_release(dst);
6446 drop_and_free:
6447         reqsk_free(req);
6448 drop:
6449         tcp_listendrop(sk);
6450         return 0;
6451 }
6452 EXPORT_SYMBOL(tcp_conn_request);
6453 

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