~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/net/ipv4/tcp_input.c

Version: ~ [ linux-5.8-rc3 ] ~ [ linux-5.7.5 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.48 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.129 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.185 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.228 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.228 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  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;
 83 int sysctl_tcp_max_reordering __read_mostly = 300;
 84 int sysctl_tcp_dsack __read_mostly = 1;
 85 int sysctl_tcp_app_win __read_mostly = 31;
 86 int sysctl_tcp_adv_win_scale __read_mostly = 1;
 87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
 88 
 89 /* rfc5961 challenge ack rate limiting */
 90 int sysctl_tcp_challenge_ack_limit = 1000;
 91 
 92 int sysctl_tcp_stdurg __read_mostly;
 93 int sysctl_tcp_rfc1337 __read_mostly;
 94 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
 95 int sysctl_tcp_frto __read_mostly = 2;
 96 int sysctl_tcp_min_rtt_wlen __read_mostly = 300;
 97 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
 98 int sysctl_tcp_early_retrans __read_mostly = 3;
 99 int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
100 
101 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
102 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
103 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
104 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
105 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
106 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
107 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
108 #define FLAG_LOST_RETRANS       0x80 /* This ACK marks some retransmission lost */
109 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_ORIG_SACK_ACKED    0x200 /* Never retransmitted data are (s)acked  */
111 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
113 #define FLAG_SACK_RENEGING      0x2000 /* snd_una advanced to a sacked seq */
114 #define FLAG_UPDATE_TS_RECENT   0x4000 /* tcp_replace_ts_recent() */
115 
116 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
120 
121 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
123 
124 #define REXMIT_NONE     0 /* no loss recovery to do */
125 #define REXMIT_LOST     1 /* retransmit packets marked lost */
126 #define REXMIT_NEW      2 /* FRTO-style transmit of unsent/new packets */
127 
128 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
129                              unsigned int len)
130 {
131         static bool __once __read_mostly;
132 
133         if (!__once) {
134                 struct net_device *dev;
135 
136                 __once = true;
137 
138                 rcu_read_lock();
139                 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
140                 if (!dev || len >= dev->mtu)
141                         pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
142                                 dev ? dev->name : "Unknown driver");
143                 rcu_read_unlock();
144         }
145 }
146 
147 /* Adapt the MSS value used to make delayed ack decision to the
148  * real world.
149  */
150 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
151 {
152         struct inet_connection_sock *icsk = inet_csk(sk);
153         const unsigned int lss = icsk->icsk_ack.last_seg_size;
154         unsigned int len;
155 
156         icsk->icsk_ack.last_seg_size = 0;
157 
158         /* skb->len may jitter because of SACKs, even if peer
159          * sends good full-sized frames.
160          */
161         len = skb_shinfo(skb)->gso_size ? : skb->len;
162         if (len >= icsk->icsk_ack.rcv_mss) {
163                 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
164                                                tcp_sk(sk)->advmss);
165                 /* Account for possibly-removed options */
166                 if (unlikely(len > icsk->icsk_ack.rcv_mss +
167                                    MAX_TCP_OPTION_SPACE))
168                         tcp_gro_dev_warn(sk, skb, len);
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         skb_mstamp_get(&tp->tcp_mstamp);
445         tp->rcvq_space.time = tp->tcp_mstamp;
446         tp->rcvq_space.seq = tp->copied_seq;
447 
448         maxwin = tcp_full_space(sk);
449 
450         if (tp->window_clamp >= maxwin) {
451                 tp->window_clamp = maxwin;
452 
453                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
454                         tp->window_clamp = max(maxwin -
455                                                (maxwin >> sysctl_tcp_app_win),
456                                                4 * tp->advmss);
457         }
458 
459         /* Force reservation of one segment. */
460         if (sysctl_tcp_app_win &&
461             tp->window_clamp > 2 * tp->advmss &&
462             tp->window_clamp + tp->advmss > maxwin)
463                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
464 
465         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
466         tp->snd_cwnd_stamp = tcp_time_stamp;
467 }
468 
469 /* 5. Recalculate window clamp after socket hit its memory bounds. */
470 static void tcp_clamp_window(struct sock *sk)
471 {
472         struct tcp_sock *tp = tcp_sk(sk);
473         struct inet_connection_sock *icsk = inet_csk(sk);
474 
475         icsk->icsk_ack.quick = 0;
476 
477         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
478             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
479             !tcp_under_memory_pressure(sk) &&
480             sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
481                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
482                                     sysctl_tcp_rmem[2]);
483         }
484         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
485                 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
486 }
487 
488 /* Initialize RCV_MSS value.
489  * RCV_MSS is an our guess about MSS used by the peer.
490  * We haven't any direct information about the MSS.
491  * It's better to underestimate the RCV_MSS rather than overestimate.
492  * Overestimations make us ACKing less frequently than needed.
493  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
494  */
495 void tcp_initialize_rcv_mss(struct sock *sk)
496 {
497         const struct tcp_sock *tp = tcp_sk(sk);
498         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
499 
500         hint = min(hint, tp->rcv_wnd / 2);
501         hint = min(hint, TCP_MSS_DEFAULT);
502         hint = max(hint, TCP_MIN_MSS);
503 
504         inet_csk(sk)->icsk_ack.rcv_mss = hint;
505 }
506 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
507 
508 /* Receiver "autotuning" code.
509  *
510  * The algorithm for RTT estimation w/o timestamps is based on
511  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
512  * <http://public.lanl.gov/radiant/pubs.html#DRS>
513  *
514  * More detail on this code can be found at
515  * <http://staff.psc.edu/jheffner/>,
516  * though this reference is out of date.  A new paper
517  * is pending.
518  */
519 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
520 {
521         u32 new_sample = tp->rcv_rtt_est.rtt_us;
522         long m = sample;
523 
524         if (m == 0)
525                 m = 1;
526 
527         if (new_sample != 0) {
528                 /* If we sample in larger samples in the non-timestamp
529                  * case, we could grossly overestimate the RTT especially
530                  * with chatty applications or bulk transfer apps which
531                  * are stalled on filesystem I/O.
532                  *
533                  * Also, since we are only going for a minimum in the
534                  * non-timestamp case, we do not smooth things out
535                  * else with timestamps disabled convergence takes too
536                  * long.
537                  */
538                 if (!win_dep) {
539                         m -= (new_sample >> 3);
540                         new_sample += m;
541                 } else {
542                         m <<= 3;
543                         if (m < new_sample)
544                                 new_sample = m;
545                 }
546         } else {
547                 /* No previous measure. */
548                 new_sample = m << 3;
549         }
550 
551         tp->rcv_rtt_est.rtt_us = new_sample;
552 }
553 
554 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
555 {
556         u32 delta_us;
557 
558         if (tp->rcv_rtt_est.time.v64 == 0)
559                 goto new_measure;
560         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
561                 return;
562         delta_us = skb_mstamp_us_delta(&tp->tcp_mstamp, &tp->rcv_rtt_est.time);
563         tcp_rcv_rtt_update(tp, delta_us, 1);
564 
565 new_measure:
566         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
567         tp->rcv_rtt_est.time = tp->tcp_mstamp;
568 }
569 
570 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
571                                           const struct sk_buff *skb)
572 {
573         struct tcp_sock *tp = tcp_sk(sk);
574         if (tp->rx_opt.rcv_tsecr &&
575             (TCP_SKB_CB(skb)->end_seq -
576              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
577                 tcp_rcv_rtt_update(tp,
578                                    jiffies_to_usecs(tcp_time_stamp -
579                                                     tp->rx_opt.rcv_tsecr),
580                                    0);
581 }
582 
583 /*
584  * This function should be called every time data is copied to user space.
585  * It calculates the appropriate TCP receive buffer space.
586  */
587 void tcp_rcv_space_adjust(struct sock *sk)
588 {
589         struct tcp_sock *tp = tcp_sk(sk);
590         int time;
591         int copied;
592 
593         time = skb_mstamp_us_delta(&tp->tcp_mstamp, &tp->rcvq_space.time);
594         if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
595                 return;
596 
597         /* Number of bytes copied to user in last RTT */
598         copied = tp->copied_seq - tp->rcvq_space.seq;
599         if (copied <= tp->rcvq_space.space)
600                 goto new_measure;
601 
602         /* A bit of theory :
603          * copied = bytes received in previous RTT, our base window
604          * To cope with packet losses, we need a 2x factor
605          * To cope with slow start, and sender growing its cwin by 100 %
606          * every RTT, we need a 4x factor, because the ACK we are sending
607          * now is for the next RTT, not the current one :
608          * <prev RTT . ><current RTT .. ><next RTT .... >
609          */
610 
611         if (sysctl_tcp_moderate_rcvbuf &&
612             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
613                 int rcvwin, rcvmem, rcvbuf;
614 
615                 /* minimal window to cope with packet losses, assuming
616                  * steady state. Add some cushion because of small variations.
617                  */
618                 rcvwin = (copied << 1) + 16 * tp->advmss;
619 
620                 /* If rate increased by 25%,
621                  *      assume slow start, rcvwin = 3 * copied
622                  * If rate increased by 50%,
623                  *      assume sender can use 2x growth, rcvwin = 4 * copied
624                  */
625                 if (copied >=
626                     tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
627                         if (copied >=
628                             tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
629                                 rcvwin <<= 1;
630                         else
631                                 rcvwin += (rcvwin >> 1);
632                 }
633 
634                 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
635                 while (tcp_win_from_space(rcvmem) < tp->advmss)
636                         rcvmem += 128;
637 
638                 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
639                 if (rcvbuf > sk->sk_rcvbuf) {
640                         sk->sk_rcvbuf = rcvbuf;
641 
642                         /* Make the window clamp follow along.  */
643                         tp->window_clamp = rcvwin;
644                 }
645         }
646         tp->rcvq_space.space = copied;
647 
648 new_measure:
649         tp->rcvq_space.seq = tp->copied_seq;
650         tp->rcvq_space.time = tp->tcp_mstamp;
651 }
652 
653 /* There is something which you must keep in mind when you analyze the
654  * behavior of the tp->ato delayed ack timeout interval.  When a
655  * connection starts up, we want to ack as quickly as possible.  The
656  * problem is that "good" TCP's do slow start at the beginning of data
657  * transmission.  The means that until we send the first few ACK's the
658  * sender will sit on his end and only queue most of his data, because
659  * he can only send snd_cwnd unacked packets at any given time.  For
660  * each ACK we send, he increments snd_cwnd and transmits more of his
661  * queue.  -DaveM
662  */
663 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
664 {
665         struct tcp_sock *tp = tcp_sk(sk);
666         struct inet_connection_sock *icsk = inet_csk(sk);
667         u32 now;
668 
669         inet_csk_schedule_ack(sk);
670 
671         tcp_measure_rcv_mss(sk, skb);
672 
673         tcp_rcv_rtt_measure(tp);
674 
675         now = tcp_time_stamp;
676 
677         if (!icsk->icsk_ack.ato) {
678                 /* The _first_ data packet received, initialize
679                  * delayed ACK engine.
680                  */
681                 tcp_incr_quickack(sk);
682                 icsk->icsk_ack.ato = TCP_ATO_MIN;
683         } else {
684                 int m = now - icsk->icsk_ack.lrcvtime;
685 
686                 if (m <= TCP_ATO_MIN / 2) {
687                         /* The fastest case is the first. */
688                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
689                 } else if (m < icsk->icsk_ack.ato) {
690                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
691                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
692                                 icsk->icsk_ack.ato = icsk->icsk_rto;
693                 } else if (m > icsk->icsk_rto) {
694                         /* Too long gap. Apparently sender failed to
695                          * restart window, so that we send ACKs quickly.
696                          */
697                         tcp_incr_quickack(sk);
698                         sk_mem_reclaim(sk);
699                 }
700         }
701         icsk->icsk_ack.lrcvtime = now;
702 
703         tcp_ecn_check_ce(tp, skb);
704 
705         if (skb->len >= 128)
706                 tcp_grow_window(sk, skb);
707 }
708 
709 /* Called to compute a smoothed rtt estimate. The data fed to this
710  * routine either comes from timestamps, or from segments that were
711  * known _not_ to have been retransmitted [see Karn/Partridge
712  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
713  * piece by Van Jacobson.
714  * NOTE: the next three routines used to be one big routine.
715  * To save cycles in the RFC 1323 implementation it was better to break
716  * it up into three procedures. -- erics
717  */
718 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
719 {
720         struct tcp_sock *tp = tcp_sk(sk);
721         long m = mrtt_us; /* RTT */
722         u32 srtt = tp->srtt_us;
723 
724         /*      The following amusing code comes from Jacobson's
725          *      article in SIGCOMM '88.  Note that rtt and mdev
726          *      are scaled versions of rtt and mean deviation.
727          *      This is designed to be as fast as possible
728          *      m stands for "measurement".
729          *
730          *      On a 1990 paper the rto value is changed to:
731          *      RTO = rtt + 4 * mdev
732          *
733          * Funny. This algorithm seems to be very broken.
734          * These formulae increase RTO, when it should be decreased, increase
735          * too slowly, when it should be increased quickly, decrease too quickly
736          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
737          * does not matter how to _calculate_ it. Seems, it was trap
738          * that VJ failed to avoid. 8)
739          */
740         if (srtt != 0) {
741                 m -= (srtt >> 3);       /* m is now error in rtt est */
742                 srtt += m;              /* rtt = 7/8 rtt + 1/8 new */
743                 if (m < 0) {
744                         m = -m;         /* m is now abs(error) */
745                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
746                         /* This is similar to one of Eifel findings.
747                          * Eifel blocks mdev updates when rtt decreases.
748                          * This solution is a bit different: we use finer gain
749                          * for mdev in this case (alpha*beta).
750                          * Like Eifel it also prevents growth of rto,
751                          * but also it limits too fast rto decreases,
752                          * happening in pure Eifel.
753                          */
754                         if (m > 0)
755                                 m >>= 3;
756                 } else {
757                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
758                 }
759                 tp->mdev_us += m;               /* mdev = 3/4 mdev + 1/4 new */
760                 if (tp->mdev_us > tp->mdev_max_us) {
761                         tp->mdev_max_us = tp->mdev_us;
762                         if (tp->mdev_max_us > tp->rttvar_us)
763                                 tp->rttvar_us = tp->mdev_max_us;
764                 }
765                 if (after(tp->snd_una, tp->rtt_seq)) {
766                         if (tp->mdev_max_us < tp->rttvar_us)
767                                 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
768                         tp->rtt_seq = tp->snd_nxt;
769                         tp->mdev_max_us = tcp_rto_min_us(sk);
770                 }
771         } else {
772                 /* no previous measure. */
773                 srtt = m << 3;          /* take the measured time to be rtt */
774                 tp->mdev_us = m << 1;   /* make sure rto = 3*rtt */
775                 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
776                 tp->mdev_max_us = tp->rttvar_us;
777                 tp->rtt_seq = tp->snd_nxt;
778         }
779         tp->srtt_us = max(1U, srtt);
780 }
781 
782 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
783  * Note: TCP stack does not yet implement pacing.
784  * FQ packet scheduler can be used to implement cheap but effective
785  * TCP pacing, to smooth the burst on large writes when packets
786  * in flight is significantly lower than cwnd (or rwin)
787  */
788 int sysctl_tcp_pacing_ss_ratio __read_mostly = 200;
789 int sysctl_tcp_pacing_ca_ratio __read_mostly = 120;
790 
791 static void tcp_update_pacing_rate(struct sock *sk)
792 {
793         const struct tcp_sock *tp = tcp_sk(sk);
794         u64 rate;
795 
796         /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
797         rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
798 
799         /* current rate is (cwnd * mss) / srtt
800          * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
801          * In Congestion Avoidance phase, set it to 120 % the current rate.
802          *
803          * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
804          *       If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
805          *       end of slow start and should slow down.
806          */
807         if (tp->snd_cwnd < tp->snd_ssthresh / 2)
808                 rate *= sysctl_tcp_pacing_ss_ratio;
809         else
810                 rate *= sysctl_tcp_pacing_ca_ratio;
811 
812         rate *= max(tp->snd_cwnd, tp->packets_out);
813 
814         if (likely(tp->srtt_us))
815                 do_div(rate, tp->srtt_us);
816 
817         /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
818          * without any lock. We want to make sure compiler wont store
819          * intermediate values in this location.
820          */
821         ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
822                                                 sk->sk_max_pacing_rate);
823 }
824 
825 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
826  * routine referred to above.
827  */
828 static void tcp_set_rto(struct sock *sk)
829 {
830         const struct tcp_sock *tp = tcp_sk(sk);
831         /* Old crap is replaced with new one. 8)
832          *
833          * More seriously:
834          * 1. If rtt variance happened to be less 50msec, it is hallucination.
835          *    It cannot be less due to utterly erratic ACK generation made
836          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
837          *    to do with delayed acks, because at cwnd>2 true delack timeout
838          *    is invisible. Actually, Linux-2.4 also generates erratic
839          *    ACKs in some circumstances.
840          */
841         inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
842 
843         /* 2. Fixups made earlier cannot be right.
844          *    If we do not estimate RTO correctly without them,
845          *    all the algo is pure shit and should be replaced
846          *    with correct one. It is exactly, which we pretend to do.
847          */
848 
849         /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
850          * guarantees that rto is higher.
851          */
852         tcp_bound_rto(sk);
853 }
854 
855 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
856 {
857         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
858 
859         if (!cwnd)
860                 cwnd = TCP_INIT_CWND;
861         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
862 }
863 
864 /*
865  * Packet counting of FACK is based on in-order assumptions, therefore TCP
866  * disables it when reordering is detected
867  */
868 void tcp_disable_fack(struct tcp_sock *tp)
869 {
870         /* RFC3517 uses different metric in lost marker => reset on change */
871         if (tcp_is_fack(tp))
872                 tp->lost_skb_hint = NULL;
873         tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
874 }
875 
876 /* Take a notice that peer is sending D-SACKs */
877 static void tcp_dsack_seen(struct tcp_sock *tp)
878 {
879         tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
880 }
881 
882 static void tcp_update_reordering(struct sock *sk, const int metric,
883                                   const int ts)
884 {
885         struct tcp_sock *tp = tcp_sk(sk);
886         int mib_idx;
887 
888         if (metric > tp->reordering) {
889                 tp->reordering = min(sysctl_tcp_max_reordering, metric);
890 
891 #if FASTRETRANS_DEBUG > 1
892                 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
893                          tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
894                          tp->reordering,
895                          tp->fackets_out,
896                          tp->sacked_out,
897                          tp->undo_marker ? tp->undo_retrans : 0);
898 #endif
899                 tcp_disable_fack(tp);
900         }
901 
902         tp->rack.reord = 1;
903 
904         /* This exciting event is worth to be remembered. 8) */
905         if (ts)
906                 mib_idx = LINUX_MIB_TCPTSREORDER;
907         else if (tcp_is_reno(tp))
908                 mib_idx = LINUX_MIB_TCPRENOREORDER;
909         else if (tcp_is_fack(tp))
910                 mib_idx = LINUX_MIB_TCPFACKREORDER;
911         else
912                 mib_idx = LINUX_MIB_TCPSACKREORDER;
913 
914         NET_INC_STATS(sock_net(sk), mib_idx);
915 }
916 
917 /* This must be called before lost_out is incremented */
918 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
919 {
920         if (!tp->retransmit_skb_hint ||
921             before(TCP_SKB_CB(skb)->seq,
922                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
923                 tp->retransmit_skb_hint = skb;
924 }
925 
926 /* Sum the number of packets on the wire we have marked as lost.
927  * There are two cases we care about here:
928  * a) Packet hasn't been marked lost (nor retransmitted),
929  *    and this is the first loss.
930  * b) Packet has been marked both lost and retransmitted,
931  *    and this means we think it was lost again.
932  */
933 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
934 {
935         __u8 sacked = TCP_SKB_CB(skb)->sacked;
936 
937         if (!(sacked & TCPCB_LOST) ||
938             ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
939                 tp->lost += tcp_skb_pcount(skb);
940 }
941 
942 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
943 {
944         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
945                 tcp_verify_retransmit_hint(tp, skb);
946 
947                 tp->lost_out += tcp_skb_pcount(skb);
948                 tcp_sum_lost(tp, skb);
949                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
950         }
951 }
952 
953 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
954 {
955         tcp_verify_retransmit_hint(tp, skb);
956 
957         tcp_sum_lost(tp, skb);
958         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
959                 tp->lost_out += tcp_skb_pcount(skb);
960                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
961         }
962 }
963 
964 /* This procedure tags the retransmission queue when SACKs arrive.
965  *
966  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
967  * Packets in queue with these bits set are counted in variables
968  * sacked_out, retrans_out and lost_out, correspondingly.
969  *
970  * Valid combinations are:
971  * Tag  InFlight        Description
972  * 0    1               - orig segment is in flight.
973  * S    0               - nothing flies, orig reached receiver.
974  * L    0               - nothing flies, orig lost by net.
975  * R    2               - both orig and retransmit are in flight.
976  * L|R  1               - orig is lost, retransmit is in flight.
977  * S|R  1               - orig reached receiver, retrans is still in flight.
978  * (L|S|R is logically valid, it could occur when L|R is sacked,
979  *  but it is equivalent to plain S and code short-curcuits it to S.
980  *  L|S is logically invalid, it would mean -1 packet in flight 8))
981  *
982  * These 6 states form finite state machine, controlled by the following events:
983  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
984  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
985  * 3. Loss detection event of two flavors:
986  *      A. Scoreboard estimator decided the packet is lost.
987  *         A'. Reno "three dupacks" marks head of queue lost.
988  *         A''. Its FACK modification, head until snd.fack is lost.
989  *      B. SACK arrives sacking SND.NXT at the moment, when the
990  *         segment was retransmitted.
991  * 4. D-SACK added new rule: D-SACK changes any tag to S.
992  *
993  * It is pleasant to note, that state diagram turns out to be commutative,
994  * so that we are allowed not to be bothered by order of our actions,
995  * when multiple events arrive simultaneously. (see the function below).
996  *
997  * Reordering detection.
998  * --------------------
999  * Reordering metric is maximal distance, which a packet can be displaced
1000  * in packet stream. With SACKs we can estimate it:
1001  *
1002  * 1. SACK fills old hole and the corresponding segment was not
1003  *    ever retransmitted -> reordering. Alas, we cannot use it
1004  *    when segment was retransmitted.
1005  * 2. The last flaw is solved with D-SACK. D-SACK arrives
1006  *    for retransmitted and already SACKed segment -> reordering..
1007  * Both of these heuristics are not used in Loss state, when we cannot
1008  * account for retransmits accurately.
1009  *
1010  * SACK block validation.
1011  * ----------------------
1012  *
1013  * SACK block range validation checks that the received SACK block fits to
1014  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1015  * Note that SND.UNA is not included to the range though being valid because
1016  * it means that the receiver is rather inconsistent with itself reporting
1017  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1018  * perfectly valid, however, in light of RFC2018 which explicitly states
1019  * that "SACK block MUST reflect the newest segment.  Even if the newest
1020  * segment is going to be discarded ...", not that it looks very clever
1021  * in case of head skb. Due to potentional receiver driven attacks, we
1022  * choose to avoid immediate execution of a walk in write queue due to
1023  * reneging and defer head skb's loss recovery to standard loss recovery
1024  * procedure that will eventually trigger (nothing forbids us doing this).
1025  *
1026  * Implements also blockage to start_seq wrap-around. Problem lies in the
1027  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1028  * there's no guarantee that it will be before snd_nxt (n). The problem
1029  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1030  * wrap (s_w):
1031  *
1032  *         <- outs wnd ->                          <- wrapzone ->
1033  *         u     e      n                         u_w   e_w  s n_w
1034  *         |     |      |                          |     |   |  |
1035  * |<------------+------+----- TCP seqno space --------------+---------->|
1036  * ...-- <2^31 ->|                                           |<--------...
1037  * ...---- >2^31 ------>|                                    |<--------...
1038  *
1039  * Current code wouldn't be vulnerable but it's better still to discard such
1040  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1041  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1042  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1043  * equal to the ideal case (infinite seqno space without wrap caused issues).
1044  *
1045  * With D-SACK the lower bound is extended to cover sequence space below
1046  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1047  * again, D-SACK block must not to go across snd_una (for the same reason as
1048  * for the normal SACK blocks, explained above). But there all simplicity
1049  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1050  * fully below undo_marker they do not affect behavior in anyway and can
1051  * therefore be safely ignored. In rare cases (which are more or less
1052  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1053  * fragmentation and packet reordering past skb's retransmission. To consider
1054  * them correctly, the acceptable range must be extended even more though
1055  * the exact amount is rather hard to quantify. However, tp->max_window can
1056  * be used as an exaggerated estimate.
1057  */
1058 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1059                                    u32 start_seq, u32 end_seq)
1060 {
1061         /* Too far in future, or reversed (interpretation is ambiguous) */
1062         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1063                 return false;
1064 
1065         /* Nasty start_seq wrap-around check (see comments above) */
1066         if (!before(start_seq, tp->snd_nxt))
1067                 return false;
1068 
1069         /* In outstanding window? ...This is valid exit for D-SACKs too.
1070          * start_seq == snd_una is non-sensical (see comments above)
1071          */
1072         if (after(start_seq, tp->snd_una))
1073                 return true;
1074 
1075         if (!is_dsack || !tp->undo_marker)
1076                 return false;
1077 
1078         /* ...Then it's D-SACK, and must reside below snd_una completely */
1079         if (after(end_seq, tp->snd_una))
1080                 return false;
1081 
1082         if (!before(start_seq, tp->undo_marker))
1083                 return true;
1084 
1085         /* Too old */
1086         if (!after(end_seq, tp->undo_marker))
1087                 return false;
1088 
1089         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1090          *   start_seq < undo_marker and end_seq >= undo_marker.
1091          */
1092         return !before(start_seq, end_seq - tp->max_window);
1093 }
1094 
1095 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1096                             struct tcp_sack_block_wire *sp, int num_sacks,
1097                             u32 prior_snd_una)
1098 {
1099         struct tcp_sock *tp = tcp_sk(sk);
1100         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1101         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1102         bool dup_sack = false;
1103 
1104         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1105                 dup_sack = true;
1106                 tcp_dsack_seen(tp);
1107                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1108         } else if (num_sacks > 1) {
1109                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1110                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1111 
1112                 if (!after(end_seq_0, end_seq_1) &&
1113                     !before(start_seq_0, start_seq_1)) {
1114                         dup_sack = true;
1115                         tcp_dsack_seen(tp);
1116                         NET_INC_STATS(sock_net(sk),
1117                                         LINUX_MIB_TCPDSACKOFORECV);
1118                 }
1119         }
1120 
1121         /* D-SACK for already forgotten data... Do dumb counting. */
1122         if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1123             !after(end_seq_0, prior_snd_una) &&
1124             after(end_seq_0, tp->undo_marker))
1125                 tp->undo_retrans--;
1126 
1127         return dup_sack;
1128 }
1129 
1130 struct tcp_sacktag_state {
1131         int     reord;
1132         int     fack_count;
1133         /* Timestamps for earliest and latest never-retransmitted segment
1134          * that was SACKed. RTO needs the earliest RTT to stay conservative,
1135          * but congestion control should still get an accurate delay signal.
1136          */
1137         struct skb_mstamp first_sackt;
1138         struct skb_mstamp last_sackt;
1139         struct rate_sample *rate;
1140         int     flag;
1141 };
1142 
1143 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1144  * the incoming SACK may not exactly match but we can find smaller MSS
1145  * aligned portion of it that matches. Therefore we might need to fragment
1146  * which may fail and creates some hassle (caller must handle error case
1147  * returns).
1148  *
1149  * FIXME: this could be merged to shift decision code
1150  */
1151 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1152                                   u32 start_seq, u32 end_seq)
1153 {
1154         int err;
1155         bool in_sack;
1156         unsigned int pkt_len;
1157         unsigned int mss;
1158 
1159         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1160                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1161 
1162         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1163             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1164                 mss = tcp_skb_mss(skb);
1165                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1166 
1167                 if (!in_sack) {
1168                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1169                         if (pkt_len < mss)
1170                                 pkt_len = mss;
1171                 } else {
1172                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1173                         if (pkt_len < mss)
1174                                 return -EINVAL;
1175                 }
1176 
1177                 /* Round if necessary so that SACKs cover only full MSSes
1178                  * and/or the remaining small portion (if present)
1179                  */
1180                 if (pkt_len > mss) {
1181                         unsigned int new_len = (pkt_len / mss) * mss;
1182                         if (!in_sack && new_len < pkt_len)
1183                                 new_len += mss;
1184                         pkt_len = new_len;
1185                 }
1186 
1187                 if (pkt_len >= skb->len && !in_sack)
1188                         return 0;
1189 
1190                 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1191                 if (err < 0)
1192                         return err;
1193         }
1194 
1195         return in_sack;
1196 }
1197 
1198 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1199 static u8 tcp_sacktag_one(struct sock *sk,
1200                           struct tcp_sacktag_state *state, u8 sacked,
1201                           u32 start_seq, u32 end_seq,
1202                           int dup_sack, int pcount,
1203                           const struct skb_mstamp *xmit_time)
1204 {
1205         struct tcp_sock *tp = tcp_sk(sk);
1206         int fack_count = state->fack_count;
1207 
1208         /* Account D-SACK for retransmitted packet. */
1209         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1210                 if (tp->undo_marker && tp->undo_retrans > 0 &&
1211                     after(end_seq, tp->undo_marker))
1212                         tp->undo_retrans--;
1213                 if (sacked & TCPCB_SACKED_ACKED)
1214                         state->reord = min(fack_count, state->reord);
1215         }
1216 
1217         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1218         if (!after(end_seq, tp->snd_una))
1219                 return sacked;
1220 
1221         if (!(sacked & TCPCB_SACKED_ACKED)) {
1222                 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1223 
1224                 if (sacked & TCPCB_SACKED_RETRANS) {
1225                         /* If the segment is not tagged as lost,
1226                          * we do not clear RETRANS, believing
1227                          * that retransmission is still in flight.
1228                          */
1229                         if (sacked & TCPCB_LOST) {
1230                                 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1231                                 tp->lost_out -= pcount;
1232                                 tp->retrans_out -= pcount;
1233                         }
1234                 } else {
1235                         if (!(sacked & TCPCB_RETRANS)) {
1236                                 /* New sack for not retransmitted frame,
1237                                  * which was in hole. It is reordering.
1238                                  */
1239                                 if (before(start_seq,
1240                                            tcp_highest_sack_seq(tp)))
1241                                         state->reord = min(fack_count,
1242                                                            state->reord);
1243                                 if (!after(end_seq, tp->high_seq))
1244                                         state->flag |= FLAG_ORIG_SACK_ACKED;
1245                                 if (state->first_sackt.v64 == 0)
1246                                         state->first_sackt = *xmit_time;
1247                                 state->last_sackt = *xmit_time;
1248                         }
1249 
1250                         if (sacked & TCPCB_LOST) {
1251                                 sacked &= ~TCPCB_LOST;
1252                                 tp->lost_out -= pcount;
1253                         }
1254                 }
1255 
1256                 sacked |= TCPCB_SACKED_ACKED;
1257                 state->flag |= FLAG_DATA_SACKED;
1258                 tp->sacked_out += pcount;
1259                 tp->delivered += pcount;  /* Out-of-order packets delivered */
1260 
1261                 fack_count += pcount;
1262 
1263                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1264                 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1265                     before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1266                         tp->lost_cnt_hint += pcount;
1267 
1268                 if (fack_count > tp->fackets_out)
1269                         tp->fackets_out = fack_count;
1270         }
1271 
1272         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1273          * frames and clear it. undo_retrans is decreased above, L|R frames
1274          * are accounted above as well.
1275          */
1276         if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1277                 sacked &= ~TCPCB_SACKED_RETRANS;
1278                 tp->retrans_out -= pcount;
1279         }
1280 
1281         return sacked;
1282 }
1283 
1284 /* Shift newly-SACKed bytes from this skb to the immediately previous
1285  * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1286  */
1287 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1288                             struct tcp_sacktag_state *state,
1289                             unsigned int pcount, int shifted, int mss,
1290                             bool dup_sack)
1291 {
1292         struct tcp_sock *tp = tcp_sk(sk);
1293         struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1294         u32 start_seq = TCP_SKB_CB(skb)->seq;   /* start of newly-SACKed */
1295         u32 end_seq = start_seq + shifted;      /* end of newly-SACKed */
1296 
1297         BUG_ON(!pcount);
1298 
1299         /* Adjust counters and hints for the newly sacked sequence
1300          * range but discard the return value since prev is already
1301          * marked. We must tag the range first because the seq
1302          * advancement below implicitly advances
1303          * tcp_highest_sack_seq() when skb is highest_sack.
1304          */
1305         tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1306                         start_seq, end_seq, dup_sack, pcount,
1307                         &skb->skb_mstamp);
1308         tcp_rate_skb_delivered(sk, skb, state->rate);
1309 
1310         if (skb == tp->lost_skb_hint)
1311                 tp->lost_cnt_hint += pcount;
1312 
1313         TCP_SKB_CB(prev)->end_seq += shifted;
1314         TCP_SKB_CB(skb)->seq += shifted;
1315 
1316         tcp_skb_pcount_add(prev, pcount);
1317         BUG_ON(tcp_skb_pcount(skb) < pcount);
1318         tcp_skb_pcount_add(skb, -pcount);
1319 
1320         /* When we're adding to gso_segs == 1, gso_size will be zero,
1321          * in theory this shouldn't be necessary but as long as DSACK
1322          * code can come after this skb later on it's better to keep
1323          * setting gso_size to something.
1324          */
1325         if (!TCP_SKB_CB(prev)->tcp_gso_size)
1326                 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1327 
1328         /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1329         if (tcp_skb_pcount(skb) <= 1)
1330                 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1331 
1332         /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1333         TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1334 
1335         if (skb->len > 0) {
1336                 BUG_ON(!tcp_skb_pcount(skb));
1337                 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1338                 return false;
1339         }
1340 
1341         /* Whole SKB was eaten :-) */
1342 
1343         if (skb == tp->retransmit_skb_hint)
1344                 tp->retransmit_skb_hint = prev;
1345         if (skb == tp->lost_skb_hint) {
1346                 tp->lost_skb_hint = prev;
1347                 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1348         }
1349 
1350         TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1351         TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1352         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1353                 TCP_SKB_CB(prev)->end_seq++;
1354 
1355         if (skb == tcp_highest_sack(sk))
1356                 tcp_advance_highest_sack(sk, skb);
1357 
1358         tcp_skb_collapse_tstamp(prev, skb);
1359         if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp.v64))
1360                 TCP_SKB_CB(prev)->tx.delivered_mstamp.v64 = 0;
1361 
1362         tcp_unlink_write_queue(skb, sk);
1363         sk_wmem_free_skb(sk, skb);
1364 
1365         NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1366 
1367         return true;
1368 }
1369 
1370 /* I wish gso_size would have a bit more sane initialization than
1371  * something-or-zero which complicates things
1372  */
1373 static int tcp_skb_seglen(const struct sk_buff *skb)
1374 {
1375         return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1376 }
1377 
1378 /* Shifting pages past head area doesn't work */
1379 static int skb_can_shift(const struct sk_buff *skb)
1380 {
1381         return !skb_headlen(skb) && skb_is_nonlinear(skb);
1382 }
1383 
1384 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1385  * skb.
1386  */
1387 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1388                                           struct tcp_sacktag_state *state,
1389                                           u32 start_seq, u32 end_seq,
1390                                           bool dup_sack)
1391 {
1392         struct tcp_sock *tp = tcp_sk(sk);
1393         struct sk_buff *prev;
1394         int mss;
1395         int pcount = 0;
1396         int len;
1397         int in_sack;
1398 
1399         if (!sk_can_gso(sk))
1400                 goto fallback;
1401 
1402         /* Normally R but no L won't result in plain S */
1403         if (!dup_sack &&
1404             (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1405                 goto fallback;
1406         if (!skb_can_shift(skb))
1407                 goto fallback;
1408         /* This frame is about to be dropped (was ACKed). */
1409         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1410                 goto fallback;
1411 
1412         /* Can only happen with delayed DSACK + discard craziness */
1413         if (unlikely(skb == tcp_write_queue_head(sk)))
1414                 goto fallback;
1415         prev = tcp_write_queue_prev(sk, skb);
1416 
1417         if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1418                 goto fallback;
1419 
1420         if (!tcp_skb_can_collapse_to(prev))
1421                 goto fallback;
1422 
1423         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1424                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1425 
1426         if (in_sack) {
1427                 len = skb->len;
1428                 pcount = tcp_skb_pcount(skb);
1429                 mss = tcp_skb_seglen(skb);
1430 
1431                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1432                  * drop this restriction as unnecessary
1433                  */
1434                 if (mss != tcp_skb_seglen(prev))
1435                         goto fallback;
1436         } else {
1437                 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1438                         goto noop;
1439                 /* CHECKME: This is non-MSS split case only?, this will
1440                  * cause skipped skbs due to advancing loop btw, original
1441                  * has that feature too
1442                  */
1443                 if (tcp_skb_pcount(skb) <= 1)
1444                         goto noop;
1445 
1446                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1447                 if (!in_sack) {
1448                         /* TODO: head merge to next could be attempted here
1449                          * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1450                          * though it might not be worth of the additional hassle
1451                          *
1452                          * ...we can probably just fallback to what was done
1453                          * previously. We could try merging non-SACKed ones
1454                          * as well but it probably isn't going to buy off
1455                          * because later SACKs might again split them, and
1456                          * it would make skb timestamp tracking considerably
1457                          * harder problem.
1458                          */
1459                         goto fallback;
1460                 }
1461 
1462                 len = end_seq - TCP_SKB_CB(skb)->seq;
1463                 BUG_ON(len < 0);
1464                 BUG_ON(len > skb->len);
1465 
1466                 /* MSS boundaries should be honoured or else pcount will
1467                  * severely break even though it makes things bit trickier.
1468                  * Optimize common case to avoid most of the divides
1469                  */
1470                 mss = tcp_skb_mss(skb);
1471 
1472                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1473                  * drop this restriction as unnecessary
1474                  */
1475                 if (mss != tcp_skb_seglen(prev))
1476                         goto fallback;
1477 
1478                 if (len == mss) {
1479                         pcount = 1;
1480                 } else if (len < mss) {
1481                         goto noop;
1482                 } else {
1483                         pcount = len / mss;
1484                         len = pcount * mss;
1485                 }
1486         }
1487 
1488         /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1489         if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1490                 goto fallback;
1491 
1492         if (!skb_shift(prev, skb, len))
1493                 goto fallback;
1494         if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1495                 goto out;
1496 
1497         /* Hole filled allows collapsing with the next as well, this is very
1498          * useful when hole on every nth skb pattern happens
1499          */
1500         if (prev == tcp_write_queue_tail(sk))
1501                 goto out;
1502         skb = tcp_write_queue_next(sk, prev);
1503 
1504         if (!skb_can_shift(skb) ||
1505             (skb == tcp_send_head(sk)) ||
1506             ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1507             (mss != tcp_skb_seglen(skb)))
1508                 goto out;
1509 
1510         len = skb->len;
1511         if (skb_shift(prev, skb, len)) {
1512                 pcount += tcp_skb_pcount(skb);
1513                 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1514         }
1515 
1516 out:
1517         state->fack_count += pcount;
1518         return prev;
1519 
1520 noop:
1521         return skb;
1522 
1523 fallback:
1524         NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1525         return NULL;
1526 }
1527 
1528 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1529                                         struct tcp_sack_block *next_dup,
1530                                         struct tcp_sacktag_state *state,
1531                                         u32 start_seq, u32 end_seq,
1532                                         bool dup_sack_in)
1533 {
1534         struct tcp_sock *tp = tcp_sk(sk);
1535         struct sk_buff *tmp;
1536 
1537         tcp_for_write_queue_from(skb, sk) {
1538                 int in_sack = 0;
1539                 bool dup_sack = dup_sack_in;
1540 
1541                 if (skb == tcp_send_head(sk))
1542                         break;
1543 
1544                 /* queue is in-order => we can short-circuit the walk early */
1545                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1546                         break;
1547 
1548                 if (next_dup  &&
1549                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1550                         in_sack = tcp_match_skb_to_sack(sk, skb,
1551                                                         next_dup->start_seq,
1552                                                         next_dup->end_seq);
1553                         if (in_sack > 0)
1554                                 dup_sack = true;
1555                 }
1556 
1557                 /* skb reference here is a bit tricky to get right, since
1558                  * shifting can eat and free both this skb and the next,
1559                  * so not even _safe variant of the loop is enough.
1560                  */
1561                 if (in_sack <= 0) {
1562                         tmp = tcp_shift_skb_data(sk, skb, state,
1563                                                  start_seq, end_seq, dup_sack);
1564                         if (tmp) {
1565                                 if (tmp != skb) {
1566                                         skb = tmp;
1567                                         continue;
1568                                 }
1569 
1570                                 in_sack = 0;
1571                         } else {
1572                                 in_sack = tcp_match_skb_to_sack(sk, skb,
1573                                                                 start_seq,
1574                                                                 end_seq);
1575                         }
1576                 }
1577 
1578                 if (unlikely(in_sack < 0))
1579                         break;
1580 
1581                 if (in_sack) {
1582                         TCP_SKB_CB(skb)->sacked =
1583                                 tcp_sacktag_one(sk,
1584                                                 state,
1585                                                 TCP_SKB_CB(skb)->sacked,
1586                                                 TCP_SKB_CB(skb)->seq,
1587                                                 TCP_SKB_CB(skb)->end_seq,
1588                                                 dup_sack,
1589                                                 tcp_skb_pcount(skb),
1590                                                 &skb->skb_mstamp);
1591                         tcp_rate_skb_delivered(sk, skb, state->rate);
1592 
1593                         if (!before(TCP_SKB_CB(skb)->seq,
1594                                     tcp_highest_sack_seq(tp)))
1595                                 tcp_advance_highest_sack(sk, skb);
1596                 }
1597 
1598                 state->fack_count += tcp_skb_pcount(skb);
1599         }
1600         return skb;
1601 }
1602 
1603 /* Avoid all extra work that is being done by sacktag while walking in
1604  * a normal way
1605  */
1606 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1607                                         struct tcp_sacktag_state *state,
1608                                         u32 skip_to_seq)
1609 {
1610         tcp_for_write_queue_from(skb, sk) {
1611                 if (skb == tcp_send_head(sk))
1612                         break;
1613 
1614                 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1615                         break;
1616 
1617                 state->fack_count += tcp_skb_pcount(skb);
1618         }
1619         return skb;
1620 }
1621 
1622 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1623                                                 struct sock *sk,
1624                                                 struct tcp_sack_block *next_dup,
1625                                                 struct tcp_sacktag_state *state,
1626                                                 u32 skip_to_seq)
1627 {
1628         if (!next_dup)
1629                 return skb;
1630 
1631         if (before(next_dup->start_seq, skip_to_seq)) {
1632                 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1633                 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1634                                        next_dup->start_seq, next_dup->end_seq,
1635                                        1);
1636         }
1637 
1638         return skb;
1639 }
1640 
1641 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1642 {
1643         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1644 }
1645 
1646 static int
1647 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1648                         u32 prior_snd_una, struct tcp_sacktag_state *state)
1649 {
1650         struct tcp_sock *tp = tcp_sk(sk);
1651         const unsigned char *ptr = (skb_transport_header(ack_skb) +
1652                                     TCP_SKB_CB(ack_skb)->sacked);
1653         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1654         struct tcp_sack_block sp[TCP_NUM_SACKS];
1655         struct tcp_sack_block *cache;
1656         struct sk_buff *skb;
1657         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1658         int used_sacks;
1659         bool found_dup_sack = false;
1660         int i, j;
1661         int first_sack_index;
1662 
1663         state->flag = 0;
1664         state->reord = tp->packets_out;
1665 
1666         if (!tp->sacked_out) {
1667                 if (WARN_ON(tp->fackets_out))
1668                         tp->fackets_out = 0;
1669                 tcp_highest_sack_reset(sk);
1670         }
1671 
1672         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1673                                          num_sacks, prior_snd_una);
1674         if (found_dup_sack) {
1675                 state->flag |= FLAG_DSACKING_ACK;
1676                 tp->delivered++; /* A spurious retransmission is delivered */
1677         }
1678 
1679         /* Eliminate too old ACKs, but take into
1680          * account more or less fresh ones, they can
1681          * contain valid SACK info.
1682          */
1683         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1684                 return 0;
1685 
1686         if (!tp->packets_out)
1687                 goto out;
1688 
1689         used_sacks = 0;
1690         first_sack_index = 0;
1691         for (i = 0; i < num_sacks; i++) {
1692                 bool dup_sack = !i && found_dup_sack;
1693 
1694                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1695                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1696 
1697                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1698                                             sp[used_sacks].start_seq,
1699                                             sp[used_sacks].end_seq)) {
1700                         int mib_idx;
1701 
1702                         if (dup_sack) {
1703                                 if (!tp->undo_marker)
1704                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1705                                 else
1706                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1707                         } else {
1708                                 /* Don't count olds caused by ACK reordering */
1709                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1710                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1711                                         continue;
1712                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1713                         }
1714 
1715                         NET_INC_STATS(sock_net(sk), mib_idx);
1716                         if (i == 0)
1717                                 first_sack_index = -1;
1718                         continue;
1719                 }
1720 
1721                 /* Ignore very old stuff early */
1722                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1723                         continue;
1724 
1725                 used_sacks++;
1726         }
1727 
1728         /* order SACK blocks to allow in order walk of the retrans queue */
1729         for (i = used_sacks - 1; i > 0; i--) {
1730                 for (j = 0; j < i; j++) {
1731                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1732                                 swap(sp[j], sp[j + 1]);
1733 
1734                                 /* Track where the first SACK block goes to */
1735                                 if (j == first_sack_index)
1736                                         first_sack_index = j + 1;
1737                         }
1738                 }
1739         }
1740 
1741         skb = tcp_write_queue_head(sk);
1742         state->fack_count = 0;
1743         i = 0;
1744 
1745         if (!tp->sacked_out) {
1746                 /* It's already past, so skip checking against it */
1747                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1748         } else {
1749                 cache = tp->recv_sack_cache;
1750                 /* Skip empty blocks in at head of the cache */
1751                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1752                        !cache->end_seq)
1753                         cache++;
1754         }
1755 
1756         while (i < used_sacks) {
1757                 u32 start_seq = sp[i].start_seq;
1758                 u32 end_seq = sp[i].end_seq;
1759                 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1760                 struct tcp_sack_block *next_dup = NULL;
1761 
1762                 if (found_dup_sack && ((i + 1) == first_sack_index))
1763                         next_dup = &sp[i + 1];
1764 
1765                 /* Skip too early cached blocks */
1766                 while (tcp_sack_cache_ok(tp, cache) &&
1767                        !before(start_seq, cache->end_seq))
1768                         cache++;
1769 
1770                 /* Can skip some work by looking recv_sack_cache? */
1771                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1772                     after(end_seq, cache->start_seq)) {
1773 
1774                         /* Head todo? */
1775                         if (before(start_seq, cache->start_seq)) {
1776                                 skb = tcp_sacktag_skip(skb, sk, state,
1777                                                        start_seq);
1778                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1779                                                        state,
1780                                                        start_seq,
1781                                                        cache->start_seq,
1782                                                        dup_sack);
1783                         }
1784 
1785                         /* Rest of the block already fully processed? */
1786                         if (!after(end_seq, cache->end_seq))
1787                                 goto advance_sp;
1788 
1789                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1790                                                        state,
1791                                                        cache->end_seq);
1792 
1793                         /* ...tail remains todo... */
1794                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1795                                 /* ...but better entrypoint exists! */
1796                                 skb = tcp_highest_sack(sk);
1797                                 if (!skb)
1798                                         break;
1799                                 state->fack_count = tp->fackets_out;
1800                                 cache++;
1801                                 goto walk;
1802                         }
1803 
1804                         skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1805                         /* Check overlap against next cached too (past this one already) */
1806                         cache++;
1807                         continue;
1808                 }
1809 
1810                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1811                         skb = tcp_highest_sack(sk);
1812                         if (!skb)
1813                                 break;
1814                         state->fack_count = tp->fackets_out;
1815                 }
1816                 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1817 
1818 walk:
1819                 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1820                                        start_seq, end_seq, dup_sack);
1821 
1822 advance_sp:
1823                 i++;
1824         }
1825 
1826         /* Clear the head of the cache sack blocks so we can skip it next time */
1827         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1828                 tp->recv_sack_cache[i].start_seq = 0;
1829                 tp->recv_sack_cache[i].end_seq = 0;
1830         }
1831         for (j = 0; j < used_sacks; j++)
1832                 tp->recv_sack_cache[i++] = sp[j];
1833 
1834         if ((state->reord < tp->fackets_out) &&
1835             ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1836                 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
1837 
1838         tcp_verify_left_out(tp);
1839 out:
1840 
1841 #if FASTRETRANS_DEBUG > 0
1842         WARN_ON((int)tp->sacked_out < 0);
1843         WARN_ON((int)tp->lost_out < 0);
1844         WARN_ON((int)tp->retrans_out < 0);
1845         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1846 #endif
1847         return state->flag;
1848 }
1849 
1850 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1851  * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1852  */
1853 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1854 {
1855         u32 holes;
1856 
1857         holes = max(tp->lost_out, 1U);
1858         holes = min(holes, tp->packets_out);
1859 
1860         if ((tp->sacked_out + holes) > tp->packets_out) {
1861                 tp->sacked_out = tp->packets_out - holes;
1862                 return true;
1863         }
1864         return false;
1865 }
1866 
1867 /* If we receive more dupacks than we expected counting segments
1868  * in assumption of absent reordering, interpret this as reordering.
1869  * The only another reason could be bug in receiver TCP.
1870  */
1871 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1872 {
1873         struct tcp_sock *tp = tcp_sk(sk);
1874         if (tcp_limit_reno_sacked(tp))
1875                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1876 }
1877 
1878 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1879 
1880 static void tcp_add_reno_sack(struct sock *sk)
1881 {
1882         struct tcp_sock *tp = tcp_sk(sk);
1883         u32 prior_sacked = tp->sacked_out;
1884 
1885         tp->sacked_out++;
1886         tcp_check_reno_reordering(sk, 0);
1887         if (tp->sacked_out > prior_sacked)
1888                 tp->delivered++; /* Some out-of-order packet is delivered */
1889         tcp_verify_left_out(tp);
1890 }
1891 
1892 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1893 
1894 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1895 {
1896         struct tcp_sock *tp = tcp_sk(sk);
1897 
1898         if (acked > 0) {
1899                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1900                 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1901                 if (acked - 1 >= tp->sacked_out)
1902                         tp->sacked_out = 0;
1903                 else
1904                         tp->sacked_out -= acked - 1;
1905         }
1906         tcp_check_reno_reordering(sk, acked);
1907         tcp_verify_left_out(tp);
1908 }
1909 
1910 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1911 {
1912         tp->sacked_out = 0;
1913 }
1914 
1915 void tcp_clear_retrans(struct tcp_sock *tp)
1916 {
1917         tp->retrans_out = 0;
1918         tp->lost_out = 0;
1919         tp->undo_marker = 0;
1920         tp->undo_retrans = -1;
1921         tp->fackets_out = 0;
1922         tp->sacked_out = 0;
1923 }
1924 
1925 static inline void tcp_init_undo(struct tcp_sock *tp)
1926 {
1927         tp->undo_marker = tp->snd_una;
1928         /* Retransmission still in flight may cause DSACKs later. */
1929         tp->undo_retrans = tp->retrans_out ? : -1;
1930 }
1931 
1932 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1933  * and reset tags completely, otherwise preserve SACKs. If receiver
1934  * dropped its ofo queue, we will know this due to reneging detection.
1935  */
1936 void tcp_enter_loss(struct sock *sk)
1937 {
1938         const struct inet_connection_sock *icsk = inet_csk(sk);
1939         struct tcp_sock *tp = tcp_sk(sk);
1940         struct net *net = sock_net(sk);
1941         struct sk_buff *skb;
1942         bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1943         bool is_reneg;                  /* is receiver reneging on SACKs? */
1944         bool mark_lost;
1945 
1946         /* Reduce ssthresh if it has not yet been made inside this window. */
1947         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1948             !after(tp->high_seq, tp->snd_una) ||
1949             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1950                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1951                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1952                 tcp_ca_event(sk, CA_EVENT_LOSS);
1953                 tcp_init_undo(tp);
1954         }
1955         tp->snd_cwnd       = 1;
1956         tp->snd_cwnd_cnt   = 0;
1957         tp->snd_cwnd_stamp = tcp_time_stamp;
1958 
1959         tp->retrans_out = 0;
1960         tp->lost_out = 0;
1961 
1962         if (tcp_is_reno(tp))
1963                 tcp_reset_reno_sack(tp);
1964 
1965         skb = tcp_write_queue_head(sk);
1966         is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1967         if (is_reneg) {
1968                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1969                 tp->sacked_out = 0;
1970                 tp->fackets_out = 0;
1971         }
1972         tcp_clear_all_retrans_hints(tp);
1973 
1974         tcp_for_write_queue(skb, sk) {
1975                 if (skb == tcp_send_head(sk))
1976                         break;
1977 
1978                 mark_lost = (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
1979                              is_reneg);
1980                 if (mark_lost)
1981                         tcp_sum_lost(tp, skb);
1982                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1983                 if (mark_lost) {
1984                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1985                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1986                         tp->lost_out += tcp_skb_pcount(skb);
1987                 }
1988         }
1989         tcp_verify_left_out(tp);
1990 
1991         /* Timeout in disordered state after receiving substantial DUPACKs
1992          * suggests that the degree of reordering is over-estimated.
1993          */
1994         if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1995             tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
1996                 tp->reordering = min_t(unsigned int, tp->reordering,
1997                                        net->ipv4.sysctl_tcp_reordering);
1998         tcp_set_ca_state(sk, TCP_CA_Loss);
1999         tp->high_seq = tp->snd_nxt;
2000         tcp_ecn_queue_cwr(tp);
2001 
2002         /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2003          * loss recovery is underway except recurring timeout(s) on
2004          * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2005          *
2006          * In theory F-RTO can be used repeatedly during loss recovery.
2007          * In practice this interacts badly with broken middle-boxes that
2008          * falsely raise the receive window, which results in repeated
2009          * timeouts and stop-and-go behavior.
2010          */
2011         tp->frto = sysctl_tcp_frto &&
2012                    (new_recovery || icsk->icsk_retransmits) &&
2013                    !inet_csk(sk)->icsk_mtup.probe_size;
2014 }
2015 
2016 /* If ACK arrived pointing to a remembered SACK, it means that our
2017  * remembered SACKs do not reflect real state of receiver i.e.
2018  * receiver _host_ is heavily congested (or buggy).
2019  *
2020  * To avoid big spurious retransmission bursts due to transient SACK
2021  * scoreboard oddities that look like reneging, we give the receiver a
2022  * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2023  * restore sanity to the SACK scoreboard. If the apparent reneging
2024  * persists until this RTO then we'll clear the SACK scoreboard.
2025  */
2026 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2027 {
2028         if (flag & FLAG_SACK_RENEGING) {
2029                 struct tcp_sock *tp = tcp_sk(sk);
2030                 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2031                                           msecs_to_jiffies(10));
2032 
2033                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2034                                           delay, TCP_RTO_MAX);
2035                 return true;
2036         }
2037         return false;
2038 }
2039 
2040 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2041 {
2042         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2043 }
2044 
2045 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2046  * counter when SACK is enabled (without SACK, sacked_out is used for
2047  * that purpose).
2048  *
2049  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2050  * segments up to the highest received SACK block so far and holes in
2051  * between them.
2052  *
2053  * With reordering, holes may still be in flight, so RFC3517 recovery
2054  * uses pure sacked_out (total number of SACKed segments) even though
2055  * it violates the RFC that uses duplicate ACKs, often these are equal
2056  * but when e.g. out-of-window ACKs or packet duplication occurs,
2057  * they differ. Since neither occurs due to loss, TCP should really
2058  * ignore them.
2059  */
2060 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2061 {
2062         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2063 }
2064 
2065 /* Linux NewReno/SACK/FACK/ECN state machine.
2066  * --------------------------------------
2067  *
2068  * "Open"       Normal state, no dubious events, fast path.
2069  * "Disorder"   In all the respects it is "Open",
2070  *              but requires a bit more attention. It is entered when
2071  *              we see some SACKs or dupacks. It is split of "Open"
2072  *              mainly to move some processing from fast path to slow one.
2073  * "CWR"        CWND was reduced due to some Congestion Notification event.
2074  *              It can be ECN, ICMP source quench, local device congestion.
2075  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2076  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2077  *
2078  * tcp_fastretrans_alert() is entered:
2079  * - each incoming ACK, if state is not "Open"
2080  * - when arrived ACK is unusual, namely:
2081  *      * SACK
2082  *      * Duplicate ACK.
2083  *      * ECN ECE.
2084  *
2085  * Counting packets in flight is pretty simple.
2086  *
2087  *      in_flight = packets_out - left_out + retrans_out
2088  *
2089  *      packets_out is SND.NXT-SND.UNA counted in packets.
2090  *
2091  *      retrans_out is number of retransmitted segments.
2092  *
2093  *      left_out is number of segments left network, but not ACKed yet.
2094  *
2095  *              left_out = sacked_out + lost_out
2096  *
2097  *     sacked_out: Packets, which arrived to receiver out of order
2098  *                 and hence not ACKed. With SACKs this number is simply
2099  *                 amount of SACKed data. Even without SACKs
2100  *                 it is easy to give pretty reliable estimate of this number,
2101  *                 counting duplicate ACKs.
2102  *
2103  *       lost_out: Packets lost by network. TCP has no explicit
2104  *                 "loss notification" feedback from network (for now).
2105  *                 It means that this number can be only _guessed_.
2106  *                 Actually, it is the heuristics to predict lossage that
2107  *                 distinguishes different algorithms.
2108  *
2109  *      F.e. after RTO, when all the queue is considered as lost,
2110  *      lost_out = packets_out and in_flight = retrans_out.
2111  *
2112  *              Essentially, we have now a few algorithms detecting
2113  *              lost packets.
2114  *
2115  *              If the receiver supports SACK:
2116  *
2117  *              RFC6675/3517: It is the conventional algorithm. A packet is
2118  *              considered lost if the number of higher sequence packets
2119  *              SACKed is greater than or equal the DUPACK thoreshold
2120  *              (reordering). This is implemented in tcp_mark_head_lost and
2121  *              tcp_update_scoreboard.
2122  *
2123  *              RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2124  *              (2017-) that checks timing instead of counting DUPACKs.
2125  *              Essentially a packet is considered lost if it's not S/ACKed
2126  *              after RTT + reordering_window, where both metrics are
2127  *              dynamically measured and adjusted. This is implemented in
2128  *              tcp_rack_mark_lost.
2129  *
2130  *              FACK (Disabled by default. Subsumbed by RACK):
2131  *              It is the simplest heuristics. As soon as we decided
2132  *              that something is lost, we decide that _all_ not SACKed
2133  *              packets until the most forward SACK are lost. I.e.
2134  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2135  *              It is absolutely correct estimate, if network does not reorder
2136  *              packets. And it loses any connection to reality when reordering
2137  *              takes place. We use FACK by default until reordering
2138  *              is suspected on the path to this destination.
2139  *
2140  *              If the receiver does not support SACK:
2141  *
2142  *              NewReno (RFC6582): in Recovery 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  * Really tricky (and requiring careful tuning) part of algorithm
2149  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2150  * The first determines the moment _when_ we should reduce CWND and,
2151  * hence, slow down forward transmission. In fact, it determines the moment
2152  * when we decide that hole is caused by loss, rather than by a reorder.
2153  *
2154  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2155  * holes, caused by lost packets.
2156  *
2157  * And the most logically complicated part of algorithm is undo
2158  * heuristics. We detect false retransmits due to both too early
2159  * fast retransmit (reordering) and underestimated RTO, analyzing
2160  * timestamps and D-SACKs. When we detect that some segments were
2161  * retransmitted by mistake and CWND reduction was wrong, we undo
2162  * window reduction and abort recovery phase. This logic is hidden
2163  * inside several functions named tcp_try_undo_<something>.
2164  */
2165 
2166 /* This function decides, when we should leave Disordered state
2167  * and enter Recovery phase, reducing congestion window.
2168  *
2169  * Main question: may we further continue forward transmission
2170  * with the same cwnd?
2171  */
2172 static bool tcp_time_to_recover(struct sock *sk, int flag)
2173 {
2174         struct tcp_sock *tp = tcp_sk(sk);
2175 
2176         /* Trick#1: The loss is proven. */
2177         if (tp->lost_out)
2178                 return true;
2179 
2180         /* Not-A-Trick#2 : Classic rule... */
2181         if (tcp_dupack_heuristics(tp) > tp->reordering)
2182                 return true;
2183 
2184         return false;
2185 }
2186 
2187 /* Detect loss in event "A" above by marking head of queue up as lost.
2188  * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2189  * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2190  * has at least tp->reordering SACKed seqments above it; "packets" refers to
2191  * the maximum SACKed segments to pass before reaching this limit.
2192  */
2193 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2194 {
2195         struct tcp_sock *tp = tcp_sk(sk);
2196         struct sk_buff *skb;
2197         int cnt, oldcnt, lost;
2198         unsigned int mss;
2199         /* Use SACK to deduce losses of new sequences sent during recovery */
2200         const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2201 
2202         WARN_ON(packets > tp->packets_out);
2203         if (tp->lost_skb_hint) {
2204                 skb = tp->lost_skb_hint;
2205                 cnt = tp->lost_cnt_hint;
2206                 /* Head already handled? */
2207                 if (mark_head && skb != tcp_write_queue_head(sk))
2208                         return;
2209         } else {
2210                 skb = tcp_write_queue_head(sk);
2211                 cnt = 0;
2212         }
2213 
2214         tcp_for_write_queue_from(skb, sk) {
2215                 if (skb == tcp_send_head(sk))
2216                         break;
2217                 /* TODO: do this better */
2218                 /* this is not the most efficient way to do this... */
2219                 tp->lost_skb_hint = skb;
2220                 tp->lost_cnt_hint = cnt;
2221 
2222                 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2223                         break;
2224 
2225                 oldcnt = cnt;
2226                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2227                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2228                         cnt += tcp_skb_pcount(skb);
2229 
2230                 if (cnt > packets) {
2231                         if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2232                             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2233                             (oldcnt >= packets))
2234                                 break;
2235 
2236                         mss = tcp_skb_mss(skb);
2237                         /* If needed, chop off the prefix to mark as lost. */
2238                         lost = (packets - oldcnt) * mss;
2239                         if (lost < skb->len &&
2240                             tcp_fragment(sk, skb, lost, mss, GFP_ATOMIC) < 0)
2241                                 break;
2242                         cnt = packets;
2243                 }
2244 
2245                 tcp_skb_mark_lost(tp, skb);
2246 
2247                 if (mark_head)
2248                         break;
2249         }
2250         tcp_verify_left_out(tp);
2251 }
2252 
2253 /* Account newly detected lost packet(s) */
2254 
2255 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2256 {
2257         struct tcp_sock *tp = tcp_sk(sk);
2258 
2259         if (tcp_is_reno(tp)) {
2260                 tcp_mark_head_lost(sk, 1, 1);
2261         } else if (tcp_is_fack(tp)) {
2262                 int lost = tp->fackets_out - tp->reordering;
2263                 if (lost <= 0)
2264                         lost = 1;
2265                 tcp_mark_head_lost(sk, lost, 0);
2266         } else {
2267                 int sacked_upto = tp->sacked_out - tp->reordering;
2268                 if (sacked_upto >= 0)
2269                         tcp_mark_head_lost(sk, sacked_upto, 0);
2270                 else if (fast_rexmit)
2271                         tcp_mark_head_lost(sk, 1, 1);
2272         }
2273 }
2274 
2275 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2276 {
2277         return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2278                before(tp->rx_opt.rcv_tsecr, when);
2279 }
2280 
2281 /* skb is spurious retransmitted if the returned timestamp echo
2282  * reply is prior to the skb transmission time
2283  */
2284 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2285                                      const struct sk_buff *skb)
2286 {
2287         return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2288                tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2289 }
2290 
2291 /* Nothing was retransmitted or returned timestamp is less
2292  * than timestamp of the first retransmission.
2293  */
2294 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2295 {
2296         return !tp->retrans_stamp ||
2297                tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2298 }
2299 
2300 /* Undo procedures. */
2301 
2302 /* We can clear retrans_stamp when there are no retransmissions in the
2303  * window. It would seem that it is trivially available for us in
2304  * tp->retrans_out, however, that kind of assumptions doesn't consider
2305  * what will happen if errors occur when sending retransmission for the
2306  * second time. ...It could the that such segment has only
2307  * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2308  * the head skb is enough except for some reneging corner cases that
2309  * are not worth the effort.
2310  *
2311  * Main reason for all this complexity is the fact that connection dying
2312  * time now depends on the validity of the retrans_stamp, in particular,
2313  * that successive retransmissions of a segment must not advance
2314  * retrans_stamp under any conditions.
2315  */
2316 static bool tcp_any_retrans_done(const struct sock *sk)
2317 {
2318         const struct tcp_sock *tp = tcp_sk(sk);
2319         struct sk_buff *skb;
2320 
2321         if (tp->retrans_out)
2322                 return true;
2323 
2324         skb = tcp_write_queue_head(sk);
2325         if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2326                 return true;
2327 
2328         return false;
2329 }
2330 
2331 #if FASTRETRANS_DEBUG > 1
2332 static void DBGUNDO(struct sock *sk, const char *msg)
2333 {
2334         struct tcp_sock *tp = tcp_sk(sk);
2335         struct inet_sock *inet = inet_sk(sk);
2336 
2337         if (sk->sk_family == AF_INET) {
2338                 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2339                          msg,
2340                          &inet->inet_daddr, ntohs(inet->inet_dport),
2341                          tp->snd_cwnd, tcp_left_out(tp),
2342                          tp->snd_ssthresh, tp->prior_ssthresh,
2343                          tp->packets_out);
2344         }
2345 #if IS_ENABLED(CONFIG_IPV6)
2346         else if (sk->sk_family == AF_INET6) {
2347                 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2348                          msg,
2349                          &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2350                          tp->snd_cwnd, tcp_left_out(tp),
2351                          tp->snd_ssthresh, tp->prior_ssthresh,
2352                          tp->packets_out);
2353         }
2354 #endif
2355 }
2356 #else
2357 #define DBGUNDO(x...) do { } while (0)
2358 #endif
2359 
2360 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2361 {
2362         struct tcp_sock *tp = tcp_sk(sk);
2363 
2364         if (unmark_loss) {
2365                 struct sk_buff *skb;
2366 
2367                 tcp_for_write_queue(skb, sk) {
2368                         if (skb == tcp_send_head(sk))
2369                                 break;
2370                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2371                 }
2372                 tp->lost_out = 0;
2373                 tcp_clear_all_retrans_hints(tp);
2374         }
2375 
2376         if (tp->prior_ssthresh) {
2377                 const struct inet_connection_sock *icsk = inet_csk(sk);
2378 
2379                 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2380 
2381                 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2382                         tp->snd_ssthresh = tp->prior_ssthresh;
2383                         tcp_ecn_withdraw_cwr(tp);
2384                 }
2385         }
2386         tp->snd_cwnd_stamp = tcp_time_stamp;
2387         tp->undo_marker = 0;
2388 }
2389 
2390 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2391 {
2392         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2393 }
2394 
2395 /* People celebrate: "We love our President!" */
2396 static bool tcp_try_undo_recovery(struct sock *sk)
2397 {
2398         struct tcp_sock *tp = tcp_sk(sk);
2399 
2400         if (tcp_may_undo(tp)) {
2401                 int mib_idx;
2402 
2403                 /* Happy end! We did not retransmit anything
2404                  * or our original transmission succeeded.
2405                  */
2406                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2407                 tcp_undo_cwnd_reduction(sk, false);
2408                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2409                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2410                 else
2411                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2412 
2413                 NET_INC_STATS(sock_net(sk), mib_idx);
2414         }
2415         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2416                 /* Hold old state until something *above* high_seq
2417                  * is ACKed. For Reno it is MUST to prevent false
2418                  * fast retransmits (RFC2582). SACK TCP is safe. */
2419                 if (!tcp_any_retrans_done(sk))
2420                         tp->retrans_stamp = 0;
2421                 return true;
2422         }
2423         tcp_set_ca_state(sk, TCP_CA_Open);
2424         return false;
2425 }
2426 
2427 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2428 static bool tcp_try_undo_dsack(struct sock *sk)
2429 {
2430         struct tcp_sock *tp = tcp_sk(sk);
2431 
2432         if (tp->undo_marker && !tp->undo_retrans) {
2433                 DBGUNDO(sk, "D-SACK");
2434                 tcp_undo_cwnd_reduction(sk, false);
2435                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2436                 return true;
2437         }
2438         return false;
2439 }
2440 
2441 /* Undo during loss recovery after partial ACK or using F-RTO. */
2442 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2443 {
2444         struct tcp_sock *tp = tcp_sk(sk);
2445 
2446         if (frto_undo || tcp_may_undo(tp)) {
2447                 tcp_undo_cwnd_reduction(sk, true);
2448 
2449                 DBGUNDO(sk, "partial loss");
2450                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2451                 if (frto_undo)
2452                         NET_INC_STATS(sock_net(sk),
2453                                         LINUX_MIB_TCPSPURIOUSRTOS);
2454                 inet_csk(sk)->icsk_retransmits = 0;
2455                 if (frto_undo || tcp_is_sack(tp))
2456                         tcp_set_ca_state(sk, TCP_CA_Open);
2457                 return true;
2458         }
2459         return false;
2460 }
2461 
2462 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2463  * It computes the number of packets to send (sndcnt) based on packets newly
2464  * delivered:
2465  *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2466  *      cwnd reductions across a full RTT.
2467  *   2) Otherwise PRR uses packet conservation to send as much as delivered.
2468  *      But when the retransmits are acked without further losses, PRR
2469  *      slow starts cwnd up to ssthresh to speed up the recovery.
2470  */
2471 static void tcp_init_cwnd_reduction(struct sock *sk)
2472 {
2473         struct tcp_sock *tp = tcp_sk(sk);
2474 
2475         tp->high_seq = tp->snd_nxt;
2476         tp->tlp_high_seq = 0;
2477         tp->snd_cwnd_cnt = 0;
2478         tp->prior_cwnd = tp->snd_cwnd;
2479         tp->prr_delivered = 0;
2480         tp->prr_out = 0;
2481         tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2482         tcp_ecn_queue_cwr(tp);
2483 }
2484 
2485 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2486 {
2487         struct tcp_sock *tp = tcp_sk(sk);
2488         int sndcnt = 0;
2489         int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2490 
2491         if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2492                 return;
2493 
2494         tp->prr_delivered += newly_acked_sacked;
2495         if (delta < 0) {
2496                 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2497                                tp->prior_cwnd - 1;
2498                 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2499         } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2500                    !(flag & FLAG_LOST_RETRANS)) {
2501                 sndcnt = min_t(int, delta,
2502                                max_t(int, tp->prr_delivered - tp->prr_out,
2503                                      newly_acked_sacked) + 1);
2504         } else {
2505                 sndcnt = min(delta, newly_acked_sacked);
2506         }
2507         /* Force a fast retransmit upon entering fast recovery */
2508         sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2509         tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2510 }
2511 
2512 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2513 {
2514         struct tcp_sock *tp = tcp_sk(sk);
2515 
2516         if (inet_csk(sk)->icsk_ca_ops->cong_control)
2517                 return;
2518 
2519         /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2520         if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2521             (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2522                 tp->snd_cwnd = tp->snd_ssthresh;
2523                 tp->snd_cwnd_stamp = tcp_time_stamp;
2524         }
2525         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2526 }
2527 
2528 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2529 void tcp_enter_cwr(struct sock *sk)
2530 {
2531         struct tcp_sock *tp = tcp_sk(sk);
2532 
2533         tp->prior_ssthresh = 0;
2534         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2535                 tp->undo_marker = 0;
2536                 tcp_init_cwnd_reduction(sk);
2537                 tcp_set_ca_state(sk, TCP_CA_CWR);
2538         }
2539 }
2540 EXPORT_SYMBOL(tcp_enter_cwr);
2541 
2542 static void tcp_try_keep_open(struct sock *sk)
2543 {
2544         struct tcp_sock *tp = tcp_sk(sk);
2545         int state = TCP_CA_Open;
2546 
2547         if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2548                 state = TCP_CA_Disorder;
2549 
2550         if (inet_csk(sk)->icsk_ca_state != state) {
2551                 tcp_set_ca_state(sk, state);
2552                 tp->high_seq = tp->snd_nxt;
2553         }
2554 }
2555 
2556 static void tcp_try_to_open(struct sock *sk, int flag)
2557 {
2558         struct tcp_sock *tp = tcp_sk(sk);
2559 
2560         tcp_verify_left_out(tp);
2561 
2562         if (!tcp_any_retrans_done(sk))
2563                 tp->retrans_stamp = 0;
2564 
2565         if (flag & FLAG_ECE)
2566                 tcp_enter_cwr(sk);
2567 
2568         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2569                 tcp_try_keep_open(sk);
2570         }
2571 }
2572 
2573 static void tcp_mtup_probe_failed(struct sock *sk)
2574 {
2575         struct inet_connection_sock *icsk = inet_csk(sk);
2576 
2577         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2578         icsk->icsk_mtup.probe_size = 0;
2579         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2580 }
2581 
2582 static void tcp_mtup_probe_success(struct sock *sk)
2583 {
2584         struct tcp_sock *tp = tcp_sk(sk);
2585         struct inet_connection_sock *icsk = inet_csk(sk);
2586 
2587         /* FIXME: breaks with very large cwnd */
2588         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2589         tp->snd_cwnd = tp->snd_cwnd *
2590                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2591                        icsk->icsk_mtup.probe_size;
2592         tp->snd_cwnd_cnt = 0;
2593         tp->snd_cwnd_stamp = tcp_time_stamp;
2594         tp->snd_ssthresh = tcp_current_ssthresh(sk);
2595 
2596         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2597         icsk->icsk_mtup.probe_size = 0;
2598         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2599         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2600 }
2601 
2602 /* Do a simple retransmit without using the backoff mechanisms in
2603  * tcp_timer. This is used for path mtu discovery.
2604  * The socket is already locked here.
2605  */
2606 void tcp_simple_retransmit(struct sock *sk)
2607 {
2608         const struct inet_connection_sock *icsk = inet_csk(sk);
2609         struct tcp_sock *tp = tcp_sk(sk);
2610         struct sk_buff *skb;
2611         unsigned int mss = tcp_current_mss(sk);
2612         u32 prior_lost = tp->lost_out;
2613 
2614         tcp_for_write_queue(skb, sk) {
2615                 if (skb == tcp_send_head(sk))
2616                         break;
2617                 if (tcp_skb_seglen(skb) > mss &&
2618                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2619                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2620                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2621                                 tp->retrans_out -= tcp_skb_pcount(skb);
2622                         }
2623                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2624                 }
2625         }
2626 
2627         tcp_clear_retrans_hints_partial(tp);
2628 
2629         if (prior_lost == tp->lost_out)
2630                 return;
2631 
2632         if (tcp_is_reno(tp))
2633                 tcp_limit_reno_sacked(tp);
2634 
2635         tcp_verify_left_out(tp);
2636 
2637         /* Don't muck with the congestion window here.
2638          * Reason is that we do not increase amount of _data_
2639          * in network, but units changed and effective
2640          * cwnd/ssthresh really reduced now.
2641          */
2642         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2643                 tp->high_seq = tp->snd_nxt;
2644                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2645                 tp->prior_ssthresh = 0;
2646                 tp->undo_marker = 0;
2647                 tcp_set_ca_state(sk, TCP_CA_Loss);
2648         }
2649         tcp_xmit_retransmit_queue(sk);
2650 }
2651 EXPORT_SYMBOL(tcp_simple_retransmit);
2652 
2653 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2654 {
2655         struct tcp_sock *tp = tcp_sk(sk);
2656         int mib_idx;
2657 
2658         if (tcp_is_reno(tp))
2659                 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2660         else
2661                 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2662 
2663         NET_INC_STATS(sock_net(sk), mib_idx);
2664 
2665         tp->prior_ssthresh = 0;
2666         tcp_init_undo(tp);
2667 
2668         if (!tcp_in_cwnd_reduction(sk)) {
2669                 if (!ece_ack)
2670                         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2671                 tcp_init_cwnd_reduction(sk);
2672         }
2673         tcp_set_ca_state(sk, TCP_CA_Recovery);
2674 }
2675 
2676 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2677  * recovered or spurious. Otherwise retransmits more on partial ACKs.
2678  */
2679 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2680                              int *rexmit)
2681 {
2682         struct tcp_sock *tp = tcp_sk(sk);
2683         bool recovered = !before(tp->snd_una, tp->high_seq);
2684 
2685         if ((flag & FLAG_SND_UNA_ADVANCED) &&
2686             tcp_try_undo_loss(sk, false))
2687                 return;
2688 
2689         /* The ACK (s)acks some never-retransmitted data meaning not all
2690          * the data packets before the timeout were lost. Therefore we
2691          * undo the congestion window and state. This is essentially
2692          * the operation in F-RTO (RFC5682 section 3.1 step 3.b). Since
2693          * a retransmitted skb is permantly marked, we can apply such an
2694          * operation even if F-RTO was not used.
2695          */
2696         if ((flag & FLAG_ORIG_SACK_ACKED) &&
2697             tcp_try_undo_loss(sk, tp->undo_marker))
2698                 return;
2699 
2700         if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2701                 if (after(tp->snd_nxt, tp->high_seq)) {
2702                         if (flag & FLAG_DATA_SACKED || is_dupack)
2703                                 tp->frto = 0; /* Step 3.a. loss was real */
2704                 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2705                         tp->high_seq = tp->snd_nxt;
2706                         /* Step 2.b. Try send new data (but deferred until cwnd
2707                          * is updated in tcp_ack()). Otherwise fall back to
2708                          * the conventional recovery.
2709                          */
2710                         if (tcp_send_head(sk) &&
2711                             after(tcp_wnd_end(tp), tp->snd_nxt)) {
2712                                 *rexmit = REXMIT_NEW;
2713                                 return;
2714                         }
2715                         tp->frto = 0;
2716                 }
2717         }
2718 
2719         if (recovered) {
2720                 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2721                 tcp_try_undo_recovery(sk);
2722                 return;
2723         }
2724         if (tcp_is_reno(tp)) {
2725                 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2726                  * delivered. Lower inflight to clock out (re)tranmissions.
2727                  */
2728                 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2729                         tcp_add_reno_sack(sk);
2730                 else if (flag & FLAG_SND_UNA_ADVANCED)
2731                         tcp_reset_reno_sack(tp);
2732         }
2733         *rexmit = REXMIT_LOST;
2734 }
2735 
2736 /* Undo during fast recovery after partial ACK. */
2737 static bool tcp_try_undo_partial(struct sock *sk, const int acked)
2738 {
2739         struct tcp_sock *tp = tcp_sk(sk);
2740 
2741         if (tp->undo_marker && tcp_packet_delayed(tp)) {
2742                 /* Plain luck! Hole if filled with delayed
2743                  * packet, rather than with a retransmit.
2744                  */
2745                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2746 
2747                 /* We are getting evidence that the reordering degree is higher
2748                  * than we realized. If there are no retransmits out then we
2749                  * can undo. Otherwise we clock out new packets but do not
2750                  * mark more packets lost or retransmit more.
2751                  */
2752                 if (tp->retrans_out)
2753                         return true;
2754 
2755                 if (!tcp_any_retrans_done(sk))
2756                         tp->retrans_stamp = 0;
2757 
2758                 DBGUNDO(sk, "partial recovery");
2759                 tcp_undo_cwnd_reduction(sk, true);
2760                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2761                 tcp_try_keep_open(sk);
2762                 return true;
2763         }
2764         return false;
2765 }
2766 
2767 static void tcp_rack_identify_loss(struct sock *sk, int *ack_flag)
2768 {
2769         struct tcp_sock *tp = tcp_sk(sk);
2770 
2771         /* Use RACK to detect loss */
2772         if (sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION) {
2773                 u32 prior_retrans = tp->retrans_out;
2774 
2775                 tcp_rack_mark_lost(sk);
2776                 if (prior_retrans > tp->retrans_out)
2777                         *ack_flag |= FLAG_LOST_RETRANS;
2778         }
2779 }
2780 
2781 /* Process an event, which can update packets-in-flight not trivially.
2782  * Main goal of this function is to calculate new estimate for left_out,
2783  * taking into account both packets sitting in receiver's buffer and
2784  * packets lost by network.
2785  *
2786  * Besides that it updates the congestion state when packet loss or ECN
2787  * is detected. But it does not reduce the cwnd, it is done by the
2788  * congestion control later.
2789  *
2790  * It does _not_ decide what to send, it is made in function
2791  * tcp_xmit_retransmit_queue().
2792  */
2793 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2794                                   bool is_dupack, int *ack_flag, int *rexmit)
2795 {
2796         struct inet_connection_sock *icsk = inet_csk(sk);
2797         struct tcp_sock *tp = tcp_sk(sk);
2798         int fast_rexmit = 0, flag = *ack_flag;
2799         bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2800                                     (tcp_fackets_out(tp) > tp->reordering));
2801 
2802         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2803                 tp->sacked_out = 0;
2804         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2805                 tp->fackets_out = 0;
2806 
2807         /* Now state machine starts.
2808          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2809         if (flag & FLAG_ECE)
2810                 tp->prior_ssthresh = 0;
2811 
2812         /* B. In all the states check for reneging SACKs. */
2813         if (tcp_check_sack_reneging(sk, flag))
2814                 return;
2815 
2816         /* C. Check consistency of the current state. */
2817         tcp_verify_left_out(tp);
2818 
2819         /* D. Check state exit conditions. State can be terminated
2820          *    when high_seq is ACKed. */
2821         if (icsk->icsk_ca_state == TCP_CA_Open) {
2822                 WARN_ON(tp->retrans_out != 0);
2823                 tp->retrans_stamp = 0;
2824         } else if (!before(tp->snd_una, tp->high_seq)) {
2825                 switch (icsk->icsk_ca_state) {
2826                 case TCP_CA_CWR:
2827                         /* CWR is to be held something *above* high_seq
2828                          * is ACKed for CWR bit to reach receiver. */
2829                         if (tp->snd_una != tp->high_seq) {
2830                                 tcp_end_cwnd_reduction(sk);
2831                                 tcp_set_ca_state(sk, TCP_CA_Open);
2832                         }
2833                         break;
2834 
2835                 case TCP_CA_Recovery:
2836                         if (tcp_is_reno(tp))
2837                                 tcp_reset_reno_sack(tp);
2838                         if (tcp_try_undo_recovery(sk))
2839                                 return;
2840                         tcp_end_cwnd_reduction(sk);
2841                         break;
2842                 }
2843         }
2844 
2845         /* E. Process state. */
2846         switch (icsk->icsk_ca_state) {
2847         case TCP_CA_Recovery:
2848                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2849                         if (tcp_is_reno(tp) && is_dupack)
2850                                 tcp_add_reno_sack(sk);
2851                 } else {
2852                         if (tcp_try_undo_partial(sk, acked))
2853                                 return;
2854                         /* Partial ACK arrived. Force fast retransmit. */
2855                         do_lost = tcp_is_reno(tp) ||
2856                                   tcp_fackets_out(tp) > tp->reordering;
2857                 }
2858                 if (tcp_try_undo_dsack(sk)) {
2859                         tcp_try_keep_open(sk);
2860                         return;
2861                 }
2862                 tcp_rack_identify_loss(sk, ack_flag);
2863                 break;
2864         case TCP_CA_Loss:
2865                 tcp_process_loss(sk, flag, is_dupack, rexmit);
2866                 tcp_rack_identify_loss(sk, ack_flag);
2867                 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2868                       (*ack_flag & FLAG_LOST_RETRANS)))
2869                         return;
2870                 /* Change state if cwnd is undone or retransmits are lost */
2871         default:
2872                 if (tcp_is_reno(tp)) {
2873                         if (flag & FLAG_SND_UNA_ADVANCED)
2874                                 tcp_reset_reno_sack(tp);
2875                         if (is_dupack)
2876                                 tcp_add_reno_sack(sk);
2877                 }
2878 
2879                 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2880                         tcp_try_undo_dsack(sk);
2881 
2882                 tcp_rack_identify_loss(sk, ack_flag);
2883                 if (!tcp_time_to_recover(sk, flag)) {
2884                         tcp_try_to_open(sk, flag);
2885                         return;
2886                 }
2887 
2888                 /* MTU probe failure: don't reduce cwnd */
2889                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2890                     icsk->icsk_mtup.probe_size &&
2891                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2892                         tcp_mtup_probe_failed(sk);
2893                         /* Restores the reduction we did in tcp_mtup_probe() */
2894                         tp->snd_cwnd++;
2895                         tcp_simple_retransmit(sk);
2896                         return;
2897                 }
2898 
2899                 /* Otherwise enter Recovery state */
2900                 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2901                 fast_rexmit = 1;
2902         }
2903 
2904         if (do_lost)
2905                 tcp_update_scoreboard(sk, fast_rexmit);
2906         *rexmit = REXMIT_LOST;
2907 }
2908 
2909 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2910 {
2911         struct tcp_sock *tp = tcp_sk(sk);
2912         u32 wlen = sysctl_tcp_min_rtt_wlen * HZ;
2913 
2914         minmax_running_min(&tp->rtt_min, wlen, tcp_time_stamp,
2915                            rtt_us ? : jiffies_to_usecs(1));
2916 }
2917 
2918 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2919                                       long seq_rtt_us, long sack_rtt_us,
2920                                       long ca_rtt_us)
2921 {
2922         const struct tcp_sock *tp = tcp_sk(sk);
2923 
2924         /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2925          * broken middle-boxes or peers may corrupt TS-ECR fields. But
2926          * Karn's algorithm forbids taking RTT if some retransmitted data
2927          * is acked (RFC6298).
2928          */
2929         if (seq_rtt_us < 0)
2930                 seq_rtt_us = sack_rtt_us;
2931 
2932         /* RTTM Rule: A TSecr value received in a segment is used to
2933          * update the averaged RTT measurement only if the segment
2934          * acknowledges some new data, i.e., only if it advances the
2935          * left edge of the send window.
2936          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2937          */
2938         if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2939             flag & FLAG_ACKED)
2940                 seq_rtt_us = ca_rtt_us = jiffies_to_usecs(tcp_time_stamp -
2941                                                           tp->rx_opt.rcv_tsecr);
2942         if (seq_rtt_us < 0)
2943                 return false;
2944 
2945         /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2946          * always taken together with ACK, SACK, or TS-opts. Any negative
2947          * values will be skipped with the seq_rtt_us < 0 check above.
2948          */
2949         tcp_update_rtt_min(sk, ca_rtt_us);
2950         tcp_rtt_estimator(sk, seq_rtt_us);
2951         tcp_set_rto(sk);
2952 
2953         /* RFC6298: only reset backoff on valid RTT measurement. */
2954         inet_csk(sk)->icsk_backoff = 0;
2955         return true;
2956 }
2957 
2958 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2959 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2960 {
2961         long rtt_us = -1L;
2962 
2963         if (req && !req->num_retrans && tcp_rsk(req)->snt_synack.v64) {
2964                 struct skb_mstamp now;
2965 
2966                 skb_mstamp_get(&now);
2967                 rtt_us = skb_mstamp_us_delta(&now, &tcp_rsk(req)->snt_synack);
2968         }
2969 
2970         tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us);
2971 }
2972 
2973 
2974 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2975 {
2976         const struct inet_connection_sock *icsk = inet_csk(sk);
2977 
2978         icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2979         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2980 }
2981 
2982 /* Restart timer after forward progress on connection.
2983  * RFC2988 recommends to restart timer to now+rto.
2984  */
2985 void tcp_rearm_rto(struct sock *sk)
2986 {
2987         const struct inet_connection_sock *icsk = inet_csk(sk);
2988         struct tcp_sock *tp = tcp_sk(sk);
2989 
2990         /* If the retrans timer is currently being used by Fast Open
2991          * for SYN-ACK retrans purpose, stay put.
2992          */
2993         if (tp->fastopen_rsk)
2994                 return;
2995 
2996         if (!tp->packets_out) {
2997                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2998         } else {
2999                 u32 rto = inet_csk(sk)->icsk_rto;
3000                 /* Offset the time elapsed after installing regular RTO */
3001                 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3002                     icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3003                         struct sk_buff *skb = tcp_write_queue_head(sk);
3004                         const u32 rto_time_stamp =
3005                                 tcp_skb_timestamp(skb) + rto;
3006                         s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3007                         /* delta may not be positive if the socket is locked
3008                          * when the retrans timer fires and is rescheduled.
3009                          */
3010                         rto = max(delta, 1);
3011                 }
3012                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3013                                           TCP_RTO_MAX);
3014         }
3015 }
3016 
3017 /* If we get here, the whole TSO packet has not been acked. */
3018 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3019 {
3020         struct tcp_sock *tp = tcp_sk(sk);
3021         u32 packets_acked;
3022 
3023         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3024 
3025         packets_acked = tcp_skb_pcount(skb);
3026         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3027                 return 0;
3028         packets_acked -= tcp_skb_pcount(skb);
3029 
3030         if (packets_acked) {
3031                 BUG_ON(tcp_skb_pcount(skb) == 0);
3032                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3033         }
3034 
3035         return packets_acked;
3036 }
3037 
3038 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3039                            u32 prior_snd_una)
3040 {
3041         const struct skb_shared_info *shinfo;
3042 
3043         /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3044         if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3045                 return;
3046 
3047         shinfo = skb_shinfo(skb);
3048         if (!before(shinfo->tskey, prior_snd_una) &&
3049             before(shinfo->tskey, tcp_sk(sk)->snd_una))
3050                 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3051 }
3052 
3053 /* Remove acknowledged frames from the retransmission queue. If our packet
3054  * is before the ack sequence we can discard it as it's confirmed to have
3055  * arrived at the other end.
3056  */
3057 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3058                                u32 prior_snd_una, int *acked,
3059                                struct tcp_sacktag_state *sack)
3060 {
3061         const struct inet_connection_sock *icsk = inet_csk(sk);
3062         struct skb_mstamp first_ackt, last_ackt;
3063         struct tcp_sock *tp = tcp_sk(sk);
3064         struct skb_mstamp *now = &tp->tcp_mstamp;
3065         u32 prior_sacked = tp->sacked_out;
3066         u32 reord = tp->packets_out;
3067         bool fully_acked = true;
3068         long sack_rtt_us = -1L;
3069         long seq_rtt_us = -1L;
3070         long ca_rtt_us = -1L;
3071         struct sk_buff *skb;
3072         u32 pkts_acked = 0;
3073         u32 last_in_flight = 0;
3074         bool rtt_update;
3075         int flag = 0;
3076 
3077         first_ackt.v64 = 0;
3078 
3079         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3080                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3081                 u8 sacked = scb->sacked;
3082                 u32 acked_pcount;
3083 
3084                 tcp_ack_tstamp(sk, skb, prior_snd_una);
3085 
3086                 /* Determine how many packets and what bytes were acked, tso and else */
3087                 if (after(scb->end_seq, tp->snd_una)) {
3088                         if (tcp_skb_pcount(skb) == 1 ||
3089                             !after(tp->snd_una, scb->seq))
3090                                 break;
3091 
3092                         acked_pcount = tcp_tso_acked(sk, skb);
3093                         if (!acked_pcount)
3094                                 break;
3095                         fully_acked = false;
3096                 } else {
3097                         /* Speedup tcp_unlink_write_queue() and next loop */
3098                         prefetchw(skb->next);
3099                         acked_pcount = tcp_skb_pcount(skb);
3100                 }
3101 
3102                 if (unlikely(sacked & TCPCB_RETRANS)) {
3103                         if (sacked & TCPCB_SACKED_RETRANS)
3104                                 tp->retrans_out -= acked_pcount;
3105                         flag |= FLAG_RETRANS_DATA_ACKED;
3106                 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3107                         last_ackt = skb->skb_mstamp;
3108                         WARN_ON_ONCE(last_ackt.v64 == 0);
3109                         if (!first_ackt.v64)
3110                                 first_ackt = last_ackt;
3111 
3112                         last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3113                         reord = min(pkts_acked, reord);
3114                         if (!after(scb->end_seq, tp->high_seq))
3115                                 flag |= FLAG_ORIG_SACK_ACKED;
3116                 }
3117 
3118                 if (sacked & TCPCB_SACKED_ACKED) {
3119                         tp->sacked_out -= acked_pcount;
3120                 } else if (tcp_is_sack(tp)) {
3121                         tp->delivered += acked_pcount;
3122                         if (!tcp_skb_spurious_retrans(tp, skb))
3123                                 tcp_rack_advance(tp, sacked, scb->end_seq,
3124                                                  &skb->skb_mstamp);
3125                 }
3126                 if (sacked & TCPCB_LOST)
3127                         tp->lost_out -= acked_pcount;
3128 
3129                 tp->packets_out -= acked_pcount;
3130                 pkts_acked += acked_pcount;
3131                 tcp_rate_skb_delivered(sk, skb, sack->rate);
3132 
3133                 /* Initial outgoing SYN's get put onto the write_queue
3134                  * just like anything else we transmit.  It is not
3135                  * true data, and if we misinform our callers that
3136                  * this ACK acks real data, we will erroneously exit
3137                  * connection startup slow start one packet too
3138                  * quickly.  This is severely frowned upon behavior.
3139                  */
3140                 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3141                         flag |= FLAG_DATA_ACKED;
3142                 } else {
3143                         flag |= FLAG_SYN_ACKED;
3144                         tp->retrans_stamp = 0;
3145                 }
3146 
3147                 if (!fully_acked)
3148                         break;
3149 
3150                 tcp_unlink_write_queue(skb, sk);
3151                 sk_wmem_free_skb(sk, skb);
3152                 if (unlikely(skb == tp->retransmit_skb_hint))
3153                         tp->retransmit_skb_hint = NULL;
3154                 if (unlikely(skb == tp->lost_skb_hint))
3155                         tp->lost_skb_hint = NULL;
3156         }
3157 
3158         if (!skb)
3159                 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3160 
3161         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3162                 tp->snd_up = tp->snd_una;
3163 
3164         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3165                 flag |= FLAG_SACK_RENEGING;
3166 
3167         if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3168                 seq_rtt_us = skb_mstamp_us_delta(now, &first_ackt);
3169                 ca_rtt_us = skb_mstamp_us_delta(now, &last_ackt);
3170         }
3171         if (sack->first_sackt.v64) {
3172                 sack_rtt_us = skb_mstamp_us_delta(now, &sack->first_sackt);
3173                 ca_rtt_us = skb_mstamp_us_delta(now, &sack->last_sackt);
3174         }
3175         sack->rate->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet, or -1 */
3176         rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3177                                         ca_rtt_us);
3178 
3179         if (flag & FLAG_ACKED) {
3180                 tcp_rearm_rto(sk);
3181                 if (unlikely(icsk->icsk_mtup.probe_size &&
3182                              !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3183                         tcp_mtup_probe_success(sk);
3184                 }
3185 
3186                 if (tcp_is_reno(tp)) {
3187                         tcp_remove_reno_sacks(sk, pkts_acked);
3188                 } else {
3189                         int delta;
3190 
3191                         /* Non-retransmitted hole got filled? That's reordering */
3192                         if (reord < prior_fackets && reord <= tp->fackets_out)
3193                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3194 
3195                         delta = tcp_is_fack(tp) ? pkts_acked :
3196                                                   prior_sacked - tp->sacked_out;
3197                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3198                 }
3199 
3200                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3201 
3202         } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3203                    sack_rtt_us > skb_mstamp_us_delta(now, &skb->skb_mstamp)) {
3204                 /* Do not re-arm RTO if the sack RTT is measured from data sent
3205                  * after when the head was last (re)transmitted. Otherwise the
3206                  * timeout may continue to extend in loss recovery.
3207                  */
3208                 tcp_rearm_rto(sk);
3209         }
3210 
3211         if (icsk->icsk_ca_ops->pkts_acked) {
3212                 struct ack_sample sample = { .pkts_acked = pkts_acked,
3213                                              .rtt_us = ca_rtt_us,
3214                                              .in_flight = last_in_flight };
3215 
3216                 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3217         }
3218 
3219 #if FASTRETRANS_DEBUG > 0
3220         WARN_ON((int)tp->sacked_out < 0);
3221         WARN_ON((int)tp->lost_out < 0);
3222         WARN_ON((int)tp->retrans_out < 0);
3223         if (!tp->packets_out && tcp_is_sack(tp)) {
3224                 icsk = inet_csk(sk);
3225                 if (tp->lost_out) {
3226                         pr_debug("Leak l=%u %d\n",
3227                                  tp->lost_out, icsk->icsk_ca_state);
3228                         tp->lost_out = 0;
3229                 }
3230                 if (tp->sacked_out) {
3231                         pr_debug("Leak s=%u %d\n",
3232                                  tp->sacked_out, icsk->icsk_ca_state);
3233                         tp->sacked_out = 0;
3234                 }
3235                 if (tp->retrans_out) {
3236                         pr_debug("Leak r=%u %d\n",
3237                                  tp->retrans_out, icsk->icsk_ca_state);
3238                         tp->retrans_out = 0;
3239                 }
3240         }
3241 #endif
3242         *acked = pkts_acked;
3243         return flag;
3244 }
3245 
3246 static void tcp_ack_probe(struct sock *sk)
3247 {
3248         const struct tcp_sock *tp = tcp_sk(sk);
3249         struct inet_connection_sock *icsk = inet_csk(sk);
3250 
3251         /* Was it a usable window open? */
3252 
3253         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3254                 icsk->icsk_backoff = 0;
3255                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3256                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3257                  * This function is not for random using!
3258                  */
3259         } else {
3260                 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3261 
3262                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3263                                           when, TCP_RTO_MAX);
3264         }
3265 }
3266 
3267 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3268 {
3269         return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3270                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3271 }
3272 
3273 /* Decide wheather to run the increase function of congestion control. */
3274 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3275 {
3276         /* If reordering is high then always grow cwnd whenever data is
3277          * delivered regardless of its ordering. Otherwise stay conservative
3278          * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3279          * new SACK or ECE mark may first advance cwnd here and later reduce
3280          * cwnd in tcp_fastretrans_alert() based on more states.
3281          */
3282         if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3283                 return flag & FLAG_FORWARD_PROGRESS;
3284 
3285         return flag & FLAG_DATA_ACKED;
3286 }
3287 
3288 /* The "ultimate" congestion control function that aims to replace the rigid
3289  * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3290  * It's called toward the end of processing an ACK with precise rate
3291  * information. All transmission or retransmission are delayed afterwards.
3292  */
3293 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3294                              int flag, const struct rate_sample *rs)
3295 {
3296         const struct inet_connection_sock *icsk = inet_csk(sk);
3297 
3298         if (icsk->icsk_ca_ops->cong_control) {
3299                 icsk->icsk_ca_ops->cong_control(sk, rs);
3300                 return;
3301         }
3302 
3303         if (tcp_in_cwnd_reduction(sk)) {
3304                 /* Reduce cwnd if state mandates */
3305                 tcp_cwnd_reduction(sk, acked_sacked, flag);
3306         } else if (tcp_may_raise_cwnd(sk, flag)) {
3307                 /* Advance cwnd if state allows */
3308                 tcp_cong_avoid(sk, ack, acked_sacked);
3309         }
3310         tcp_update_pacing_rate(sk);
3311 }
3312 
3313 /* Check that window update is acceptable.
3314  * The function assumes that snd_una<=ack<=snd_next.
3315  */
3316 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3317                                         const u32 ack, const u32 ack_seq,
3318                                         const u32 nwin)
3319 {
3320         return  after(ack, tp->snd_una) ||
3321                 after(ack_seq, tp->snd_wl1) ||
3322                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3323 }
3324 
3325 /* If we update tp->snd_una, also update tp->bytes_acked */
3326 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3327 {
3328         u32 delta = ack - tp->snd_una;
3329 
3330         sock_owned_by_me((struct sock *)tp);
3331         tp->bytes_acked += delta;
3332         tp->snd_una = ack;
3333 }
3334 
3335 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3336 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3337 {
3338         u32 delta = seq - tp->rcv_nxt;
3339 
3340         sock_owned_by_me((struct sock *)tp);
3341         tp->bytes_received += delta;
3342         tp->rcv_nxt = seq;
3343 }
3344 
3345 /* Update our send window.
3346  *
3347  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3348  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3349  */
3350 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3351                                  u32 ack_seq)
3352 {
3353         struct tcp_sock *tp = tcp_sk(sk);
3354         int flag = 0;
3355         u32 nwin = ntohs(tcp_hdr(skb)->window);
3356 
3357         if (likely(!tcp_hdr(skb)->syn))
3358                 nwin <<= tp->rx_opt.snd_wscale;
3359 
3360         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3361                 flag |= FLAG_WIN_UPDATE;
3362                 tcp_update_wl(tp, ack_seq);
3363 
3364                 if (tp->snd_wnd != nwin) {
3365                         tp->snd_wnd = nwin;
3366 
3367                         /* Note, it is the only place, where
3368                          * fast path is recovered for sending TCP.
3369                          */
3370                         tp->pred_flags = 0;
3371                         tcp_fast_path_check(sk);
3372 
3373                         if (tcp_send_head(sk))
3374                                 tcp_slow_start_after_idle_check(sk);
3375 
3376                         if (nwin > tp->max_window) {
3377                                 tp->max_window = nwin;
3378                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3379                         }
3380                 }
3381         }
3382 
3383         tcp_snd_una_update(tp, ack);
3384 
3385         return flag;
3386 }
3387 
3388 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3389                                    u32 *last_oow_ack_time)
3390 {
3391         if (*last_oow_ack_time) {
3392                 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3393 
3394                 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3395                         NET_INC_STATS(net, mib_idx);
3396                         return true;    /* rate-limited: don't send yet! */
3397                 }
3398         }
3399 
3400         *last_oow_ack_time = tcp_time_stamp;
3401 
3402         return false;   /* not rate-limited: go ahead, send dupack now! */
3403 }
3404 
3405 /* Return true if we're currently rate-limiting out-of-window ACKs and
3406  * thus shouldn't send a dupack right now. We rate-limit dupacks in
3407  * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3408  * attacks that send repeated SYNs or ACKs for the same connection. To
3409  * do this, we do not send a duplicate SYNACK or ACK if the remote
3410  * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3411  */
3412 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3413                           int mib_idx, u32 *last_oow_ack_time)
3414 {
3415         /* Data packets without SYNs are not likely part of an ACK loop. */
3416         if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3417             !tcp_hdr(skb)->syn)
3418                 return false;
3419 
3420         return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3421 }
3422 
3423 /* RFC 5961 7 [ACK Throttling] */
3424 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3425 {
3426         /* unprotected vars, we dont care of overwrites */
3427         static u32 challenge_timestamp;
3428         static unsigned int challenge_count;
3429         struct tcp_sock *tp = tcp_sk(sk);
3430         u32 count, now;
3431 
3432         /* First check our per-socket dupack rate limit. */
3433         if (__tcp_oow_rate_limited(sock_net(sk),
3434                                    LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3435                                    &tp->last_oow_ack_time))
3436                 return;
3437 
3438         /* Then check host-wide RFC 5961 rate limit. */
3439         now = jiffies / HZ;
3440         if (now != challenge_timestamp) {
3441                 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3442 
3443                 challenge_timestamp = now;
3444                 WRITE_ONCE(challenge_count, half +
3445                            prandom_u32_max(sysctl_tcp_challenge_ack_limit));
3446         }
3447         count = READ_ONCE(challenge_count);
3448         if (count > 0) {
3449                 WRITE_ONCE(challenge_count, count - 1);
3450                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3451                 tcp_send_ack(sk);
3452         }
3453 }
3454 
3455 static void tcp_store_ts_recent(struct tcp_sock *tp)
3456 {
3457         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3458         tp->rx_opt.ts_recent_stamp = get_seconds();
3459 }
3460 
3461 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3462 {
3463         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3464                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3465                  * extra check below makes sure this can only happen
3466                  * for pure ACK frames.  -DaveM
3467                  *
3468                  * Not only, also it occurs for expired timestamps.
3469                  */
3470 
3471                 if (tcp_paws_check(&tp->rx_opt, 0))
3472                         tcp_store_ts_recent(tp);
3473         }
3474 }
3475 
3476 /* This routine deals with acks during a TLP episode.
3477  * We mark the end of a TLP episode on receiving TLP dupack or when
3478  * ack is after tlp_high_seq.
3479  * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3480  */
3481 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3482 {
3483         struct tcp_sock *tp = tcp_sk(sk);
3484 
3485         if (before(ack, tp->tlp_high_seq))
3486                 return;
3487 
3488         if (flag & FLAG_DSACKING_ACK) {
3489                 /* This DSACK means original and TLP probe arrived; no loss */
3490                 tp->tlp_high_seq = 0;
3491         } else if (after(ack, tp->tlp_high_seq)) {
3492                 /* ACK advances: there was a loss, so reduce cwnd. Reset
3493                  * tlp_high_seq in tcp_init_cwnd_reduction()
3494                  */
3495                 tcp_init_cwnd_reduction(sk);
3496                 tcp_set_ca_state(sk, TCP_CA_CWR);
3497                 tcp_end_cwnd_reduction(sk);
3498                 tcp_try_keep_open(sk);
3499                 NET_INC_STATS(sock_net(sk),
3500                                 LINUX_MIB_TCPLOSSPROBERECOVERY);
3501         } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3502                              FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3503                 /* Pure dupack: original and TLP probe arrived; no loss */
3504                 tp->tlp_high_seq = 0;
3505         }
3506 }
3507 
3508 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3509 {
3510         const struct inet_connection_sock *icsk = inet_csk(sk);
3511 
3512         if (icsk->icsk_ca_ops->in_ack_event)
3513                 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3514 }
3515 
3516 /* Congestion control has updated the cwnd already. So if we're in
3517  * loss recovery then now we do any new sends (for FRTO) or
3518  * retransmits (for CA_Loss or CA_recovery) that make sense.
3519  */
3520 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3521 {
3522         struct tcp_sock *tp = tcp_sk(sk);
3523 
3524         if (rexmit == REXMIT_NONE)
3525                 return;
3526 
3527         if (unlikely(rexmit == 2)) {
3528                 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3529                                           TCP_NAGLE_OFF);
3530                 if (after(tp->snd_nxt, tp->high_seq))
3531                         return;
3532                 tp->frto = 0;
3533         }
3534         tcp_xmit_retransmit_queue(sk);
3535 }
3536 
3537 /* This routine deals with incoming acks, but not outgoing ones. */
3538 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3539 {
3540         struct inet_connection_sock *icsk = inet_csk(sk);
3541         struct tcp_sock *tp = tcp_sk(sk);
3542         struct tcp_sacktag_state sack_state;
3543         struct rate_sample rs = { .prior_delivered = 0 };
3544         u32 prior_snd_una = tp->snd_una;
3545         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3546         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3547         bool is_dupack = false;
3548         u32 prior_fackets;
3549         int prior_packets = tp->packets_out;
3550         u32 delivered = tp->delivered;
3551         u32 lost = tp->lost;
3552         int acked = 0; /* Number of packets newly acked */
3553         int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3554 
3555         sack_state.first_sackt.v64 = 0;
3556         sack_state.rate = &rs;
3557 
3558         /* We very likely will need to access write queue head. */
3559         prefetchw(sk->sk_write_queue.next);
3560 
3561         /* If the ack is older than previous acks
3562          * then we can probably ignore it.
3563          */
3564         if (before(ack, prior_snd_una)) {
3565                 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3566                 if (before(ack, prior_snd_una - tp->max_window)) {
3567                         tcp_send_challenge_ack(sk, skb);
3568                         return -1;
3569                 }
3570                 goto old_ack;
3571         }
3572 
3573         /* If the ack includes data we haven't sent yet, discard
3574          * this segment (RFC793 Section 3.9).
3575          */
3576         if (after(ack, tp->snd_nxt))
3577                 goto invalid_ack;
3578 
3579         if (icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3580                 tcp_rearm_rto(sk);
3581 
3582         if (after(ack, prior_snd_una)) {
3583                 flag |= FLAG_SND_UNA_ADVANCED;
3584                 icsk->icsk_retransmits = 0;
3585         }
3586 
3587         prior_fackets = tp->fackets_out;
3588         rs.prior_in_flight = tcp_packets_in_flight(tp);
3589 
3590         /* ts_recent update must be made after we are sure that the packet
3591          * is in window.
3592          */
3593         if (flag & FLAG_UPDATE_TS_RECENT)
3594                 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3595 
3596         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3597                 /* Window is constant, pure forward advance.
3598                  * No more checks are required.
3599                  * Note, we use the fact that SND.UNA>=SND.WL2.
3600                  */
3601                 tcp_update_wl(tp, ack_seq);
3602                 tcp_snd_una_update(tp, ack);
3603                 flag |= FLAG_WIN_UPDATE;
3604 
3605                 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3606 
3607                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3608         } else {
3609                 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3610 
3611                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3612                         flag |= FLAG_DATA;
3613                 else
3614                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3615 
3616                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3617 
3618                 if (TCP_SKB_CB(skb)->sacked)
3619                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3620                                                         &sack_state);
3621 
3622                 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3623                         flag |= FLAG_ECE;
3624                         ack_ev_flags |= CA_ACK_ECE;
3625                 }
3626 
3627                 if (flag & FLAG_WIN_UPDATE)
3628                         ack_ev_flags |= CA_ACK_WIN_UPDATE;
3629 
3630                 tcp_in_ack_event(sk, ack_ev_flags);
3631         }
3632 
3633         /* We passed data and got it acked, remove any soft error
3634          * log. Something worked...
3635          */
3636         sk->sk_err_soft = 0;
3637         icsk->icsk_probes_out = 0;
3638         tp->rcv_tstamp = tcp_time_stamp;
3639         if (!prior_packets)
3640                 goto no_queue;
3641 
3642         /* See if we can take anything off of the retransmit queue. */
3643         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, &acked,
3644                                     &sack_state);
3645 
3646         if (tcp_ack_is_dubious(sk, flag)) {
3647                 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3648                 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3649         }
3650         if (tp->tlp_high_seq)
3651                 tcp_process_tlp_ack(sk, ack, flag);
3652 
3653         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3654                 sk_dst_confirm(sk);
3655 
3656         if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3657                 tcp_schedule_loss_probe(sk);
3658         delivered = tp->delivered - delivered;  /* freshly ACKed or SACKed */
3659         lost = tp->lost - lost;                 /* freshly marked lost */
3660         tcp_rate_gen(sk, delivered, lost, sack_state.rate);
3661         tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3662         tcp_xmit_recovery(sk, rexmit);
3663         return 1;
3664 
3665 no_queue:
3666         /* If data was DSACKed, see if we can undo a cwnd reduction. */
3667         if (flag & FLAG_DSACKING_ACK)
3668                 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3669         /* If this ack opens up a zero window, clear backoff.  It was
3670          * being used to time the probes, and is probably far higher than
3671          * it needs to be for normal retransmission.
3672          */
3673         if (tcp_send_head(sk))
3674                 tcp_ack_probe(sk);
3675 
3676         if (tp->tlp_high_seq)
3677                 tcp_process_tlp_ack(sk, ack, flag);
3678         return 1;
3679 
3680 invalid_ack:
3681         SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3682         return -1;
3683 
3684 old_ack:
3685         /* If data was SACKed, tag it and see if we should send more data.
3686          * If data was DSACKed, see if we can undo a cwnd reduction.
3687          */
3688         if (TCP_SKB_CB(skb)->sacked) {
3689                 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3690                                                 &sack_state);
3691                 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3692                 tcp_xmit_recovery(sk, rexmit);
3693         }
3694 
3695         SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3696         return 0;
3697 }
3698 
3699 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3700                                       bool syn, struct tcp_fastopen_cookie *foc,
3701                                       bool exp_opt)
3702 {
3703         /* Valid only in SYN or SYN-ACK with an even length.  */
3704         if (!foc || !syn || len < 0 || (len & 1))
3705                 return;
3706 
3707         if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3708             len <= TCP_FASTOPEN_COOKIE_MAX)
3709                 memcpy(foc->val, cookie, len);
3710         else if (len != 0)
3711                 len = -1;
3712         foc->len = len;
3713         foc->exp = exp_opt;
3714 }
3715 
3716 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3717  * But, this can also be called on packets in the established flow when
3718  * the fast version below fails.
3719  */
3720 void tcp_parse_options(const struct sk_buff *skb,
3721                        struct tcp_options_received *opt_rx, int estab,
3722                        struct tcp_fastopen_cookie *foc)
3723 {
3724         const unsigned char *ptr;
3725         const struct tcphdr *th = tcp_hdr(skb);
3726         int length = (th->doff * 4) - sizeof(struct tcphdr);
3727 
3728         ptr = (const unsigned char *)(th + 1);
3729         opt_rx->saw_tstamp = 0;
3730 
3731         while (length > 0) {
3732                 int opcode = *ptr++;
3733                 int opsize;
3734 
3735                 switch (opcode) {
3736                 case TCPOPT_EOL:
3737                         return;
3738                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3739                         length--;
3740                         continue;
3741                 default:
3742                         opsize = *ptr++;
3743                         if (opsize < 2) /* "silly options" */
3744                                 return;
3745                         if (opsize > length)
3746                                 return; /* don't parse partial options */
3747                         switch (opcode) {
3748                         case TCPOPT_MSS:
3749                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3750                                         u16 in_mss = get_unaligned_be16(ptr);
3751                                         if (in_mss) {
3752                                                 if (opt_rx->user_mss &&
3753                                                     opt_rx->user_mss < in_mss)
3754                                                         in_mss = opt_rx->user_mss;
3755                                                 opt_rx->mss_clamp = in_mss;
3756                                         }
3757                                 }
3758                                 break;
3759                         case TCPOPT_WINDOW:
3760                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3761                                     !estab && sysctl_tcp_window_scaling) {
3762                                         __u8 snd_wscale = *(__u8 *)ptr;
3763                                         opt_rx->wscale_ok = 1;
3764                                         if (snd_wscale > TCP_MAX_WSCALE) {
3765                                                 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3766                                                                      __func__,
3767                                                                      snd_wscale,
3768                                                                      TCP_MAX_WSCALE);
3769                                                 snd_wscale = TCP_MAX_WSCALE;
3770                                         }
3771                                         opt_rx->snd_wscale = snd_wscale;
3772                                 }
3773                                 break;
3774                         case TCPOPT_TIMESTAMP:
3775                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3776                                     ((estab && opt_rx->tstamp_ok) ||
3777                                      (!estab && sysctl_tcp_timestamps))) {
3778                                         opt_rx->saw_tstamp = 1;
3779                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3780                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3781                                 }
3782                                 break;
3783                         case TCPOPT_SACK_PERM:
3784                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3785                                     !estab && sysctl_tcp_sack) {
3786                                         opt_rx->sack_ok = TCP_SACK_SEEN;
3787                                         tcp_sack_reset(opt_rx);
3788                                 }
3789                                 break;
3790 
3791                         case TCPOPT_SACK:
3792                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3793                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3794                                    opt_rx->sack_ok) {
3795                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3796                                 }
3797                                 break;
3798 #ifdef CONFIG_TCP_MD5SIG
3799                         case TCPOPT_MD5SIG:
3800                                 /*
3801                                  * The MD5 Hash has already been
3802                                  * checked (see tcp_v{4,6}_do_rcv()).
3803                                  */
3804                                 break;
3805 #endif
3806                         case TCPOPT_FASTOPEN:
3807                                 tcp_parse_fastopen_option(
3808                                         opsize - TCPOLEN_FASTOPEN_BASE,
3809                                         ptr, th->syn, foc, false);
3810                                 break;
3811 
3812                         case TCPOPT_EXP:
3813                                 /* Fast Open option shares code 254 using a
3814                                  * 16 bits magic number.
3815                                  */
3816                                 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3817                                     get_unaligned_be16(ptr) ==
3818                                     TCPOPT_FASTOPEN_MAGIC)
3819                                         tcp_parse_fastopen_option(opsize -
3820                                                 TCPOLEN_EXP_FASTOPEN_BASE,
3821                                                 ptr + 2, th->syn, foc, true);
3822                                 break;
3823 
3824                         }
3825                         ptr += opsize-2;
3826                         length -= opsize;
3827                 }
3828         }
3829 }
3830 EXPORT_SYMBOL(tcp_parse_options);
3831 
3832 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3833 {
3834         const __be32 *ptr = (const __be32 *)(th + 1);
3835 
3836         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3837                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3838                 tp->rx_opt.saw_tstamp = 1;
3839                 ++ptr;
3840                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3841                 ++ptr;
3842                 if (*ptr)
3843                         tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3844                 else
3845                         tp->rx_opt.rcv_tsecr = 0;
3846                 return true;
3847         }
3848         return false;
3849 }
3850 
3851 /* Fast parse options. This hopes to only see timestamps.
3852  * If it is wrong it falls back on tcp_parse_options().
3853  */
3854 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3855                                    const struct tcphdr *th, struct tcp_sock *tp)
3856 {
3857         /* In the spirit of fast parsing, compare doff directly to constant
3858          * values.  Because equality is used, short doff can be ignored here.
3859          */
3860         if (th->doff == (sizeof(*th) / 4)) {
3861                 tp->rx_opt.saw_tstamp = 0;
3862                 return false;
3863         } else if (tp->rx_opt.tstamp_ok &&
3864                    th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3865                 if (tcp_parse_aligned_timestamp(tp, th))
3866                         return true;
3867         }
3868 
3869         tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3870         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3871                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3872 
3873         return true;
3874 }
3875 
3876 #ifdef CONFIG_TCP_MD5SIG
3877 /*
3878  * Parse MD5 Signature option
3879  */
3880 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3881 {
3882         int length = (th->doff << 2) - sizeof(*th);
3883         const u8 *ptr = (const u8 *)(th + 1);
3884 
3885         /* If the TCP option is too short, we can short cut */
3886         if (length < TCPOLEN_MD5SIG)
3887                 return NULL;
3888 
3889         while (length > 0) {
3890                 int opcode = *ptr++;
3891                 int opsize;
3892 
3893                 switch (opcode) {
3894                 case TCPOPT_EOL:
3895                         return NULL;
3896                 case TCPOPT_NOP:
3897                         length--;
3898                         continue;
3899                 default:
3900                         opsize = *ptr++;
3901                         if (opsize < 2 || opsize > length)
3902                                 return NULL;
3903                         if (opcode == TCPOPT_MD5SIG)
3904                                 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3905                 }
3906                 ptr += opsize - 2;
3907                 length -= opsize;
3908         }
3909         return NULL;
3910 }
3911 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3912 #endif
3913 
3914 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3915  *
3916  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3917  * it can pass through stack. So, the following predicate verifies that
3918  * this segment is not used for anything but congestion avoidance or
3919  * fast retransmit. Moreover, we even are able to eliminate most of such
3920  * second order effects, if we apply some small "replay" window (~RTO)
3921  * to timestamp space.
3922  *
3923  * All these measures still do not guarantee that we reject wrapped ACKs
3924  * on networks with high bandwidth, when sequence space is recycled fastly,
3925  * but it guarantees that such events will be very rare and do not affect
3926  * connection seriously. This doesn't look nice, but alas, PAWS is really
3927  * buggy extension.
3928  *
3929  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3930  * states that events when retransmit arrives after original data are rare.
3931  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3932  * the biggest problem on large power networks even with minor reordering.
3933  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3934  * up to bandwidth of 18Gigabit/sec. 8) ]
3935  */
3936 
3937 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3938 {
3939         const struct tcp_sock *tp = tcp_sk(sk);
3940         const struct tcphdr *th = tcp_hdr(skb);
3941         u32 seq = TCP_SKB_CB(skb)->seq;
3942         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3943 
3944         return (/* 1. Pure ACK with correct sequence number. */
3945                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3946 
3947                 /* 2. ... and duplicate ACK. */
3948                 ack == tp->snd_una &&
3949 
3950                 /* 3. ... and does not update window. */
3951                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3952 
3953                 /* 4. ... and sits in replay window. */
3954                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3955 }
3956 
3957 static inline bool tcp_paws_discard(const struct sock *sk,
3958                                    const struct sk_buff *skb)
3959 {
3960         const struct tcp_sock *tp = tcp_sk(sk);
3961 
3962         return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3963                !tcp_disordered_ack(sk, skb);
3964 }
3965 
3966 /* Check segment sequence number for validity.
3967  *
3968  * Segment controls are considered valid, if the segment
3969  * fits to the window after truncation to the window. Acceptability
3970  * of data (and SYN, FIN, of course) is checked separately.
3971  * See tcp_data_queue(), for example.
3972  *
3973  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3974  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3975  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3976  * (borrowed from freebsd)
3977  */
3978 
3979 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3980 {
3981         return  !before(end_seq, tp->rcv_wup) &&
3982                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3983 }
3984 
3985 /* When we get a reset we do this. */
3986 void tcp_reset(struct sock *sk)
3987 {
3988         /* We want the right error as BSD sees it (and indeed as we do). */
3989         switch (sk->sk_state) {
3990         case TCP_SYN_SENT:
3991                 sk->sk_err = ECONNREFUSED;
3992                 break;
3993         case TCP_CLOSE_WAIT:
3994                 sk->sk_err = EPIPE;
3995                 break;
3996         case TCP_CLOSE:
3997                 return;
3998         default:
3999                 sk->sk_err = ECONNRESET;
4000         }
4001         /* This barrier is coupled with smp_rmb() in tcp_poll() */
4002         smp_wmb();
4003 
4004         tcp_done(sk);
4005 
4006         if (!sock_flag(sk, SOCK_DEAD))
4007                 sk->sk_error_report(sk);
4008 }
4009 
4010 /*
4011  *      Process the FIN bit. This now behaves as it is supposed to work
4012  *      and the FIN takes effect when it is validly part of sequence
4013  *      space. Not before when we get holes.
4014  *
4015  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4016  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
4017  *      TIME-WAIT)
4018  *
4019  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
4020  *      close and we go into CLOSING (and later onto TIME-WAIT)
4021  *
4022  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4023  */
4024 void tcp_fin(struct sock *sk)
4025 {
4026         struct tcp_sock *tp = tcp_sk(sk);
4027 
4028         inet_csk_schedule_ack(sk);
4029 
4030         sk->sk_shutdown |= RCV_SHUTDOWN;
4031         sock_set_flag(sk, SOCK_DONE);
4032 
4033         switch (sk->sk_state) {
4034         case TCP_SYN_RECV:
4035         case TCP_ESTABLISHED:
4036                 /* Move to CLOSE_WAIT */
4037                 tcp_set_state(sk, TCP_CLOSE_WAIT);
4038                 inet_csk(sk)->icsk_ack.pingpong = 1;
4039                 break;
4040 
4041         case TCP_CLOSE_WAIT:
4042         case TCP_CLOSING:
4043                 /* Received a retransmission of the FIN, do
4044                  * nothing.
4045                  */
4046                 break;
4047         case TCP_LAST_ACK:
4048                 /* RFC793: Remain in the LAST-ACK state. */
4049                 break;
4050 
4051         case TCP_FIN_WAIT1:
4052                 /* This case occurs when a simultaneous close
4053                  * happens, we must ack the received FIN and
4054                  * enter the CLOSING state.
4055                  */
4056                 tcp_send_ack(sk);
4057                 tcp_set_state(sk, TCP_CLOSING);
4058                 break;
4059         case TCP_FIN_WAIT2:
4060                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4061                 tcp_send_ack(sk);
4062                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4063                 break;
4064         default:
4065                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4066                  * cases we should never reach this piece of code.
4067                  */
4068                 pr_err("%s: Impossible, sk->sk_state=%d\n",
4069                        __func__, sk->sk_state);
4070                 break;
4071         }
4072 
4073         /* It _is_ possible, that we have something out-of-order _after_ FIN.
4074          * Probably, we should reset in this case. For now drop them.
4075          */
4076         skb_rbtree_purge(&tp->out_of_order_queue);
4077         if (tcp_is_sack(tp))
4078                 tcp_sack_reset(&tp->rx_opt);
4079         sk_mem_reclaim(sk);
4080 
4081         if (!sock_flag(sk, SOCK_DEAD)) {
4082                 sk->sk_state_change(sk);
4083 
4084                 /* Do not send POLL_HUP for half duplex close. */
4085                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4086                     sk->sk_state == TCP_CLOSE)
4087                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4088                 else
4089                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4090         }
4091 }
4092 
4093 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4094                                   u32 end_seq)
4095 {
4096         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4097                 if (before(seq, sp->start_seq))
4098                         sp->start_seq = seq;
4099                 if (after(end_seq, sp->end_seq))
4100                         sp->end_seq = end_seq;
4101                 return true;
4102         }
4103         return false;
4104 }
4105 
4106 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4107 {
4108         struct tcp_sock *tp = tcp_sk(sk);
4109 
4110         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4111                 int mib_idx;
4112 
4113                 if (before(seq, tp->rcv_nxt))
4114                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4115                 else
4116                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4117 
4118                 NET_INC_STATS(sock_net(sk), mib_idx);
4119 
4120                 tp->rx_opt.dsack = 1;
4121                 tp->duplicate_sack[0].start_seq = seq;
4122                 tp->duplicate_sack[0].end_seq = end_seq;
4123         }
4124 }
4125 
4126 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4127 {
4128         struct tcp_sock *tp = tcp_sk(sk);
4129 
4130         if (!tp->rx_opt.dsack)
4131                 tcp_dsack_set(sk, seq, end_seq);
4132         else
4133                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4134 }
4135 
4136 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4137 {
4138         struct tcp_sock *tp = tcp_sk(sk);
4139 
4140         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4141             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4142                 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4143                 tcp_enter_quickack_mode(sk);
4144 
4145                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4146                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4147 
4148                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4149                                 end_seq = tp->rcv_nxt;
4150                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4151                 }
4152         }
4153 
4154         tcp_send_ack(sk);
4155 }
4156 
4157 /* These routines update the SACK block as out-of-order packets arrive or
4158  * in-order packets close up the sequence space.
4159  */
4160 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4161 {
4162         int this_sack;
4163         struct tcp_sack_block *sp = &tp->selective_acks[0];
4164         struct tcp_sack_block *swalk = sp + 1;
4165 
4166         /* See if the recent change to the first SACK eats into
4167          * or hits the sequence space of other SACK blocks, if so coalesce.
4168          */
4169         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4170                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4171                         int i;
4172 
4173                         /* Zap SWALK, by moving every further SACK up by one slot.
4174                          * Decrease num_sacks.
4175                          */
4176                         tp->rx_opt.num_sacks--;
4177                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4178                                 sp[i] = sp[i + 1];
4179                         continue;
4180                 }
4181                 this_sack++, swalk++;
4182         }
4183 }
4184 
4185 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4186 {
4187         struct tcp_sock *tp = tcp_sk(sk);
4188         struct tcp_sack_block *sp = &tp->selective_acks[0];
4189         int cur_sacks = tp->rx_opt.num_sacks;
4190         int this_sack;
4191 
4192         if (!cur_sacks)
4193                 goto new_sack;
4194 
4195         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4196                 if (tcp_sack_extend(sp, seq, end_seq)) {
4197                         /* Rotate this_sack to the first one. */
4198                         for (; this_sack > 0; this_sack--, sp--)
4199                                 swap(*sp, *(sp - 1));
4200                         if (cur_sacks > 1)
4201                                 tcp_sack_maybe_coalesce(tp);
4202                         return;
4203                 }
4204         }
4205 
4206         /* Could not find an adjacent existing SACK, build a new one,
4207          * put it at the front, and shift everyone else down.  We
4208          * always know there is at least one SACK present already here.
4209          *
4210          * If the sack array is full, forget about the last one.
4211          */
4212         if (this_sack >= TCP_NUM_SACKS) {
4213                 this_sack--;
4214                 tp->rx_opt.num_sacks--;
4215                 sp--;
4216         }
4217         for (; this_sack > 0; this_sack--, sp--)
4218                 *sp = *(sp - 1);
4219 
4220 new_sack:
4221         /* Build the new head SACK, and we're done. */
4222         sp->start_seq = seq;
4223         sp->end_seq = end_seq;
4224         tp->rx_opt.num_sacks++;
4225 }
4226 
4227 /* RCV.NXT advances, some SACKs should be eaten. */
4228 
4229 static void tcp_sack_remove(struct tcp_sock *tp)
4230 {
4231         struct tcp_sack_block *sp = &tp->selective_acks[0];
4232         int num_sacks = tp->rx_opt.num_sacks;
4233         int this_sack;
4234 
4235         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4236         if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4237                 tp->rx_opt.num_sacks = 0;
4238                 return;
4239         }
4240 
4241         for (this_sack = 0; this_sack < num_sacks;) {
4242                 /* Check if the start of the sack is covered by RCV.NXT. */
4243                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4244                         int i;
4245 
4246                         /* RCV.NXT must cover all the block! */
4247                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4248 
4249                         /* Zap this SACK, by moving forward any other SACKS. */
4250                         for (i = this_sack+1; i < num_sacks; i++)
4251                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4252                         num_sacks--;
4253                         continue;
4254                 }
4255                 this_sack++;
4256                 sp++;
4257         }
4258         tp->rx_opt.num_sacks = num_sacks;
4259 }
4260 
4261 /**
4262  * tcp_try_coalesce - try to merge skb to prior one
4263  * @sk: socket
4264  * @to: prior buffer
4265  * @from: buffer to add in queue
4266  * @fragstolen: pointer to boolean
4267  *
4268  * Before queueing skb @from after @to, try to merge them
4269  * to reduce overall memory use and queue lengths, if cost is small.
4270  * Packets in ofo or receive queues can stay a long time.
4271  * Better try to coalesce them right now to avoid future collapses.
4272  * Returns true if caller should free @from instead of queueing it
4273  */
4274 static bool tcp_try_coalesce(struct sock *sk,
4275                              struct sk_buff *to,
4276                              struct sk_buff *from,
4277                              bool *fragstolen)
4278 {
4279         int delta;
4280 
4281         *fragstolen = false;
4282 
4283         /* Its possible this segment overlaps with prior segment in queue */
4284         if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4285                 return false;
4286 
4287         if (!skb_try_coalesce(to, from, fragstolen, &delta))
4288                 return false;
4289 
4290         atomic_add(delta, &sk->sk_rmem_alloc);
4291         sk_mem_charge(sk, delta);
4292         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4293         TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4294         TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4295         TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4296         return true;
4297 }
4298 
4299 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4300 {
4301         sk_drops_add(sk, skb);
4302         __kfree_skb(skb);
4303 }
4304 
4305 /* This one checks to see if we can put data from the
4306  * out_of_order queue into the receive_queue.
4307  */
4308 static void tcp_ofo_queue(struct sock *sk)
4309 {
4310         struct tcp_sock *tp = tcp_sk(sk);
4311         __u32 dsack_high = tp->rcv_nxt;
4312         bool fin, fragstolen, eaten;
4313         struct sk_buff *skb, *tail;
4314         struct rb_node *p;
4315 
4316         p = rb_first(&tp->out_of_order_queue);
4317         while (p) {
4318                 skb = rb_entry(p, struct sk_buff, rbnode);
4319                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4320                         break;
4321 
4322                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4323                         __u32 dsack = dsack_high;
4324                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4325                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4326                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4327                 }
4328                 p = rb_next(p);
4329                 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4330 
4331                 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4332                         SOCK_DEBUG(sk, "ofo packet was already received\n");
4333                         tcp_drop(sk, skb);
4334                         continue;
4335                 }
4336                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4337                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4338                            TCP_SKB_CB(skb)->end_seq);
4339 
4340                 tail = skb_peek_tail(&sk->sk_receive_queue);
4341                 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4342                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4343                 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4344                 if (!eaten)
4345                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4346                 else
4347                         kfree_skb_partial(skb, fragstolen);
4348 
4349                 if (unlikely(fin)) {
4350                         tcp_fin(sk);
4351                         /* tcp_fin() purges tp->out_of_order_queue,
4352                          * so we must end this loop right now.
4353                          */
4354                         break;
4355                 }
4356         }
4357 }
4358 
4359 static bool tcp_prune_ofo_queue(struct sock *sk);
4360 static int tcp_prune_queue(struct sock *sk);
4361 
4362 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4363                                  unsigned int size)
4364 {
4365         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4366             !sk_rmem_schedule(sk, skb, size)) {
4367 
4368                 if (tcp_prune_queue(sk) < 0)
4369                         return -1;
4370 
4371                 while (!sk_rmem_schedule(sk, skb, size)) {
4372                         if (!tcp_prune_ofo_queue(sk))
4373                                 return -1;
4374                 }
4375         }
4376         return 0;
4377 }
4378 
4379 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4380 {
4381         struct tcp_sock *tp = tcp_sk(sk);
4382         struct rb_node **p, *q, *parent;
4383         struct sk_buff *skb1;
4384         u32 seq, end_seq;
4385         bool fragstolen;
4386 
4387         tcp_ecn_check_ce(tp, skb);
4388 
4389         if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4390                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4391                 tcp_drop(sk, skb);
4392                 return;
4393         }
4394 
4395         /* Disable header prediction. */
4396         tp->pred_flags = 0;
4397         inet_csk_schedule_ack(sk);
4398 
4399         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4400         seq = TCP_SKB_CB(skb)->seq;
4401         end_seq = TCP_SKB_CB(skb)->end_seq;
4402         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4403                    tp->rcv_nxt, seq, end_seq);
4404 
4405         p = &tp->out_of_order_queue.rb_node;
4406         if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4407                 /* Initial out of order segment, build 1 SACK. */
4408                 if (tcp_is_sack(tp)) {
4409                         tp->rx_opt.num_sacks = 1;
4410                         tp->selective_acks[0].start_seq = seq;
4411                         tp->selective_acks[0].end_seq = end_seq;
4412                 }
4413                 rb_link_node(&skb->rbnode, NULL, p);
4414                 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4415                 tp->ooo_last_skb = skb;
4416                 goto end;
4417         }
4418 
4419         /* In the typical case, we are adding an skb to the end of the list.
4420          * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4421          */
4422         if (tcp_try_coalesce(sk, tp->ooo_last_skb, skb, &fragstolen)) {
4423 coalesce_done:
4424                 tcp_grow_window(sk, skb);
4425                 kfree_skb_partial(skb, fragstolen);
4426                 skb = NULL;
4427                 goto add_sack;
4428         }
4429         /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4430         if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4431                 parent = &tp->ooo_last_skb->rbnode;
4432                 p = &parent->rb_right;
4433                 goto insert;
4434         }
4435 
4436         /* Find place to insert this segment. Handle overlaps on the way. */
4437         parent = NULL;
4438         while (*p) {
4439                 parent = *p;
4440                 skb1 = rb_entry(parent, struct sk_buff, rbnode);
4441                 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4442                         p = &parent->rb_left;
4443                         continue;
4444                 }
4445                 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4446                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4447                                 /* All the bits are present. Drop. */
4448                                 NET_INC_STATS(sock_net(sk),
4449                                               LINUX_MIB_TCPOFOMERGE);
4450                                 __kfree_skb(skb);
4451                                 skb = NULL;
4452                                 tcp_dsack_set(sk, seq, end_seq);
4453                                 goto add_sack;
4454                         }
4455                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4456                                 /* Partial overlap. */
4457                                 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4458                         } else {
4459                                 /* skb's seq == skb1's seq and skb covers skb1.
4460                                  * Replace skb1 with skb.
4461                                  */
4462                                 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4463                                                 &tp->out_of_order_queue);
4464                                 tcp_dsack_extend(sk,
4465                                                  TCP_SKB_CB(skb1)->seq,
4466                                                  TCP_SKB_CB(skb1)->end_seq);
4467                                 NET_INC_STATS(sock_net(sk),
4468                                               LINUX_MIB_TCPOFOMERGE);
4469                                 __kfree_skb(skb1);
4470                                 goto merge_right;
4471                         }
4472                 } else if (tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4473                         goto coalesce_done;
4474                 }
4475                 p = &parent->rb_right;
4476         }
4477 insert:
4478         /* Insert segment into RB tree. */
4479         rb_link_node(&skb->rbnode, parent, p);
4480         rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4481 
4482 merge_right:
4483         /* Remove other segments covered by skb. */
4484         while ((q = rb_next(&skb->rbnode)) != NULL) {
4485                 skb1 = rb_entry(q, struct sk_buff, rbnode);
4486 
4487                 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4488                         break;
4489                 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4490                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4491                                          end_seq);
4492                         break;
4493                 }
4494                 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4495                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4496                                  TCP_SKB_CB(skb1)->end_seq);
4497                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4498                 tcp_drop(sk, skb1);
4499         }
4500         /* If there is no skb after us, we are the last_skb ! */
4501         if (!q)
4502                 tp->ooo_last_skb = skb;
4503 
4504 add_sack:
4505         if (tcp_is_sack(tp))
4506                 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4507 end:
4508         if (skb) {
4509                 tcp_grow_window(sk, skb);
4510                 skb_condense(skb);
4511                 skb_set_owner_r(skb, sk);
4512         }
4513 }
4514 
4515 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4516                   bool *fragstolen)
4517 {
4518         int eaten;
4519         struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4520 
4521         __skb_pull(skb, hdrlen);
4522         eaten = (tail &&
4523                  tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4524         tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4525         if (!eaten) {
4526                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4527                 skb_set_owner_r(skb, sk);
4528         }
4529         return eaten;
4530 }
4531 
4532 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4533 {
4534         struct sk_buff *skb;
4535         int err = -ENOMEM;
4536         int data_len = 0;
4537         bool fragstolen;
4538 
4539         if (size == 0)
4540                 return 0;
4541 
4542         if (size > PAGE_SIZE) {
4543                 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4544 
4545                 data_len = npages << PAGE_SHIFT;
4546                 size = data_len + (size & ~PAGE_MASK);
4547         }
4548         skb = alloc_skb_with_frags(size - data_len, data_len,
4549                                    PAGE_ALLOC_COSTLY_ORDER,
4550                                    &err, sk->sk_allocation);
4551         if (!skb)
4552                 goto err;
4553 
4554         skb_put(skb, size - data_len);
4555         skb->data_len = data_len;
4556         skb->len = size;
4557 
4558         if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4559                 goto err_free;
4560 
4561         err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4562         if (err)
4563                 goto err_free;
4564 
4565         TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4566         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4567         TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4568 
4569         if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4570                 WARN_ON_ONCE(fragstolen); /* should not happen */
4571                 __kfree_skb(skb);
4572         }
4573         return size;
4574 
4575 err_free:
4576         kfree_skb(skb);
4577 err:
4578         return err;
4579 
4580 }
4581 
4582 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4583 {
4584         struct tcp_sock *tp = tcp_sk(sk);
4585         bool fragstolen = false;
4586         int eaten = -1;
4587 
4588         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4589                 __kfree_skb(skb);
4590                 return;
4591         }
4592         skb_dst_drop(skb);
4593         __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4594 
4595         tcp_ecn_accept_cwr(tp, skb);
4596 
4597         tp->rx_opt.dsack = 0;
4598 
4599         /*  Queue data for delivery to the user.
4600          *  Packets in sequence go to the receive queue.
4601          *  Out of sequence packets to the out_of_order_queue.
4602          */
4603         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4604                 if (tcp_receive_window(tp) == 0)
4605                         goto out_of_window;
4606 
4607                 /* Ok. In sequence. In window. */
4608                 if (tp->ucopy.task == current &&
4609                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4610                     sock_owned_by_user(sk) && !tp->urg_data) {
4611                         int chunk = min_t(unsigned int, skb->len,
4612                                           tp->ucopy.len);
4613 
4614                         __set_current_state(TASK_RUNNING);
4615 
4616                         if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4617                                 tp->ucopy.len -= chunk;
4618                                 tp->copied_seq += chunk;
4619                                 eaten = (chunk == skb->len);
4620                                 tcp_rcv_space_adjust(sk);
4621                         }
4622                 }
4623 
4624                 if (eaten <= 0) {
4625 queue_and_out:
4626                         if (eaten < 0) {
4627                                 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4628                                         sk_forced_mem_schedule(sk, skb->truesize);
4629                                 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4630                                         goto drop;
4631                         }
4632                         eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4633                 }
4634                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4635                 if (skb->len)
4636                         tcp_event_data_recv(sk, skb);
4637                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4638                         tcp_fin(sk);
4639 
4640                 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4641                         tcp_ofo_queue(sk);
4642 
4643                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4644                          * gap in queue is filled.
4645                          */
4646                         if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4647                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4648                 }
4649 
4650                 if (tp->rx_opt.num_sacks)
4651                         tcp_sack_remove(tp);
4652 
4653                 tcp_fast_path_check(sk);
4654 
4655                 if (eaten > 0)
4656                         kfree_skb_partial(skb, fragstolen);
4657                 if (!sock_flag(sk, SOCK_DEAD))
4658                         sk->sk_data_ready(sk);
4659                 return;
4660         }
4661 
4662         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4663                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4664                 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4665                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4666 
4667 out_of_window:
4668                 tcp_enter_quickack_mode(sk);
4669                 inet_csk_schedule_ack(sk);
4670 drop:
4671                 tcp_drop(sk, skb);
4672                 return;
4673         }
4674 
4675         /* Out of window. F.e. zero window probe. */
4676         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4677                 goto out_of_window;
4678 
4679         tcp_enter_quickack_mode(sk);
4680 
4681         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4682                 /* Partial packet, seq < rcv_next < end_seq */
4683                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4684                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4685                            TCP_SKB_CB(skb)->end_seq);
4686 
4687                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4688 
4689                 /* If window is closed, drop tail of packet. But after
4690                  * remembering D-SACK for its head made in previous line.
4691                  */
4692                 if (!tcp_receive_window(tp))
4693                         goto out_of_window;
4694                 goto queue_and_out;
4695         }
4696 
4697         tcp_data_queue_ofo(sk, skb);
4698 }
4699 
4700 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4701 {
4702         if (list)
4703                 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4704 
4705         return rb_entry_safe(rb_next(&skb->rbnode), struct sk_buff, rbnode);
4706 }
4707 
4708 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4709                                         struct sk_buff_head *list,
4710                                         struct rb_root *root)
4711 {
4712         struct sk_buff *next = tcp_skb_next(skb, list);
4713 
4714         if (list)
4715                 __skb_unlink(skb, list);
4716         else
4717                 rb_erase(&skb->rbnode, root);
4718 
4719         __kfree_skb(skb);
4720         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4721 
4722         return next;
4723 }
4724 
4725 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4726 static void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4727 {
4728         struct rb_node **p = &root->rb_node;
4729         struct rb_node *parent = NULL;
4730         struct sk_buff *skb1;
4731 
4732         while (*p) {
4733                 parent = *p;
4734                 skb1 = rb_entry(parent, struct sk_buff, rbnode);
4735                 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4736                         p = &parent->rb_left;
4737                 else
4738                         p = &parent->rb_right;
4739         }
4740         rb_link_node(&skb->rbnode, parent, p);
4741         rb_insert_color(&skb->rbnode, root);
4742 }
4743 
4744 /* Collapse contiguous sequence of skbs head..tail with
4745  * sequence numbers start..end.
4746  *
4747  * If tail is NULL, this means until the end of the queue.
4748  *
4749  * Segments with FIN/SYN are not collapsed (only because this
4750  * simplifies code)
4751  */
4752 static void
4753 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4754              struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4755 {
4756         struct sk_buff *skb = head, *n;
4757         struct sk_buff_head tmp;
4758         bool end_of_skbs;
4759 
4760         /* First, check that queue is collapsible and find
4761          * the point where collapsing can be useful.
4762          */
4763 restart:
4764         for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4765                 n = tcp_skb_next(skb, list);
4766 
4767                 /* No new bits? It is possible on ofo queue. */
4768                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4769                         skb = tcp_collapse_one(sk, skb, list, root);
4770                         if (!skb)
4771                                 break;
4772                         goto restart;
4773                 }
4774 
4775                 /* The first skb to collapse is:
4776                  * - not SYN/FIN and
4777                  * - bloated or contains data before "start" or
4778                  *   overlaps to the next one.
4779                  */
4780                 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4781                     (tcp_win_from_space(skb->truesize) > skb->len ||
4782                      before(TCP_SKB_CB(skb)->seq, start))) {
4783                         end_of_skbs = false;
4784                         break;
4785                 }
4786 
4787                 if (n && n != tail &&
4788                     TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4789                         end_of_skbs = false;
4790                         break;
4791                 }
4792 
4793                 /* Decided to skip this, advance start seq. */
4794                 start = TCP_SKB_CB(skb)->end_seq;
4795         }
4796         if (end_of_skbs ||
4797             (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4798                 return;
4799 
4800         __skb_queue_head_init(&tmp);
4801 
4802         while (before(start, end)) {
4803                 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4804                 struct sk_buff *nskb;
4805 
4806                 nskb = alloc_skb(copy, GFP_ATOMIC);
4807                 if (!nskb)
4808                         break;
4809 
4810                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4811                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4812                 if (list)
4813                         __skb_queue_before(list, skb, nskb);
4814                 else
4815                         __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4816                 skb_set_owner_r(nskb, sk);
4817 
4818                 /* Copy data, releasing collapsed skbs. */
4819                 while (copy > 0) {
4820                         int offset = start - TCP_SKB_CB(skb)->seq;
4821                         int size = TCP_SKB_CB(skb)->end_seq - start;
4822 
4823                         BUG_ON(offset < 0);
4824                         if (size > 0) {
4825                                 size = min(copy, size);
4826                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4827                                         BUG();
4828                                 TCP_SKB_CB(nskb)->end_seq += size;
4829                                 copy -= size;
4830                                 start += size;
4831                         }
4832                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4833                                 skb = tcp_collapse_one(sk, skb, list, root);
4834                                 if (!skb ||
4835                                     skb == tail ||
4836                                     (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4837                                         goto end;
4838                         }
4839                 }
4840         }
4841 end:
4842         skb_queue_walk_safe(&tmp, skb, n)
4843                 tcp_rbtree_insert(root, skb);
4844 }
4845 
4846 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4847  * and tcp_collapse() them until all the queue is collapsed.
4848  */
4849 static void tcp_collapse_ofo_queue(struct sock *sk)
4850 {
4851         struct tcp_sock *tp = tcp_sk(sk);
4852         struct sk_buff *skb, *head;
4853         struct rb_node *p;
4854         u32 start, end;
4855 
4856         p = rb_first(&tp->out_of_order_queue);
4857         skb = rb_entry_safe(p, struct sk_buff, rbnode);
4858 new_range:
4859         if (!skb) {
4860                 p = rb_last(&tp->out_of_order_queue);
4861                 /* Note: This is possible p is NULL here. We do not
4862                  * use rb_entry_safe(), as ooo_last_skb is valid only
4863                  * if rbtree is not empty.
4864                  */
4865                 tp->ooo_last_skb = rb_entry(p, struct sk_buff, rbnode);
4866                 return;
4867         }
4868         start = TCP_SKB_CB(skb)->seq;
4869         end = TCP_SKB_CB(skb)->end_seq;
4870 
4871         for (head = skb;;) {
4872                 skb = tcp_skb_next(skb, NULL);
4873 
4874                 /* Range is terminated when we see a gap or when
4875                  * we are at the queue end.
4876                  */
4877                 if (!skb ||
4878                     after(TCP_SKB_CB(skb)->seq, end) ||
4879                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4880                         tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4881                                      head, skb, start, end);
4882                         goto new_range;
4883                 }
4884 
4885                 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4886                         start = TCP_SKB_CB(skb)->seq;
4887                 if (after(TCP_SKB_CB(skb)->end_seq, end))
4888                         end = TCP_SKB_CB(skb)->end_seq;
4889         }
4890 }
4891 
4892 /*
4893  * Clean the out-of-order queue to make room.
4894  * We drop high sequences packets to :
4895  * 1) Let a chance for holes to be filled.
4896  * 2) not add too big latencies if thousands of packets sit there.
4897  *    (But if application shrinks SO_RCVBUF, we could still end up
4898  *     freeing whole queue here)
4899  *
4900  * Return true if queue has shrunk.
4901  */
4902 static bool tcp_prune_ofo_queue(struct sock *sk)
4903 {
4904         struct tcp_sock *tp = tcp_sk(sk);
4905         struct rb_node *node, *prev;
4906 
4907         if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4908                 return false;
4909 
4910         NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
4911         node = &tp->ooo_last_skb->rbnode;
4912         do {
4913                 prev = rb_prev(node);
4914                 rb_erase(node, &tp->out_of_order_queue);
4915                 tcp_drop(sk, rb_entry(node, struct sk_buff, rbnode));
4916                 sk_mem_reclaim(sk);
4917                 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
4918                     !tcp_under_memory_pressure(sk))
4919                         break;
4920                 node = prev;
4921         } while (node);
4922         tp->ooo_last_skb = rb_entry(prev, struct sk_buff, rbnode);
4923 
4924         /* Reset SACK state.  A conforming SACK implementation will
4925          * do the same at a timeout based retransmit.  When a connection
4926          * is in a sad state like this, we care only about integrity
4927          * of the connection not performance.
4928          */
4929         if (tp->rx_opt.sack_ok)
4930                 tcp_sack_reset(&tp->rx_opt);
4931         return true;
4932 }
4933 
4934 /* Reduce allocated memory if we can, trying to get
4935  * the socket within its memory limits again.
4936  *
4937  * Return less than zero if we should start dropping frames
4938  * until the socket owning process reads some of the data
4939  * to stabilize the situation.
4940  */
4941 static int tcp_prune_queue(struct sock *sk)
4942 {
4943         struct tcp_sock *tp = tcp_sk(sk);
4944 
4945         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4946 
4947         NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
4948 
4949         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4950                 tcp_clamp_window(sk);
4951         else if (tcp_under_memory_pressure(sk))
4952                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4953 
4954         tcp_collapse_ofo_queue(sk);
4955         if (!skb_queue_empty(&sk->sk_receive_queue))
4956                 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
4957                              skb_peek(&sk->sk_receive_queue),
4958                              NULL,
4959                              tp->copied_seq, tp->rcv_nxt);
4960         sk_mem_reclaim(sk);
4961 
4962         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4963                 return 0;
4964 
4965         /* Collapsing did not help, destructive actions follow.
4966          * This must not ever occur. */
4967 
4968         tcp_prune_ofo_queue(sk);
4969 
4970         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4971                 return 0;
4972 
4973         /* If we are really being abused, tell the caller to silently
4974          * drop receive data on the floor.  It will get retransmitted
4975          * and hopefully then we'll have sufficient space.
4976          */
4977         NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
4978 
4979         /* Massive buffer overcommit. */
4980         tp->pred_flags = 0;
4981         return -1;
4982 }
4983 
4984 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4985 {
4986         const struct tcp_sock *tp = tcp_sk(sk);
4987 
4988         /* If the user specified a specific send buffer setting, do
4989          * not modify it.
4990          */
4991         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4992                 return false;
4993 
4994         /* If we are under global TCP memory pressure, do not expand.  */
4995         if (tcp_under_memory_pressure(sk))
4996                 return false;
4997 
4998         /* If we are under soft global TCP memory pressure, do not expand.  */
4999         if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5000                 return false;
5001 
5002         /* If we filled the congestion window, do not expand.  */
5003         if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5004                 return false;
5005 
5006         return true;
5007 }
5008 
5009 /* When incoming ACK allowed to free some skb from write_queue,
5010  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5011  * on the exit from tcp input handler.
5012  *
5013  * PROBLEM: sndbuf expansion does not work well with largesend.
5014  */
5015 static void tcp_new_space(struct sock *sk)
5016 {
5017         struct tcp_sock *tp = tcp_sk(sk);
5018 
5019         if (tcp_should_expand_sndbuf(sk)) {
5020                 tcp_sndbuf_expand(sk);
5021                 tp->snd_cwnd_stamp = tcp_time_stamp;
5022         }
5023 
5024         sk->sk_write_space(sk);
5025 }
5026 
5027 static void tcp_check_space(struct sock *sk)
5028 {
5029         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5030                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5031                 /* pairs with tcp_poll() */
5032                 smp_mb();
5033                 if (sk->sk_socket &&
5034                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5035                         tcp_new_space(sk);
5036                         if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5037                                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5038                 }
5039         }
5040 }
5041 
5042 static inline void tcp_data_snd_check(struct sock *sk)
5043 {
5044         tcp_push_pending_frames(sk);
5045         tcp_check_space(sk);
5046 }
5047 
5048 /*
5049  * Check if sending an ack is needed.
5050  */
5051 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5052 {
5053         struct tcp_sock *tp = tcp_sk(sk);
5054 
5055             /* More than one full frame received... */
5056         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5057              /* ... and right edge of window advances far enough.
5058               * (tcp_recvmsg() will send ACK otherwise). Or...
5059               */
5060              __tcp_select_window(sk) >= tp->rcv_wnd) ||
5061             /* We ACK each frame or... */
5062             tcp_in_quickack_mode(sk) ||
5063             /* We have out of order data. */
5064             (ofo_possible && !RB_EMPTY_ROOT(&tp->out_of_order_queue))) {
5065                 /* Then ack it now */
5066                 tcp_send_ack(sk);
5067         } else {
5068                 /* Else, send delayed ack. */
5069                 tcp_send_delayed_ack(sk);
5070         }
5071 }
5072 
5073 static inline void tcp_ack_snd_check(struct sock *sk)
5074 {
5075         if (!inet_csk_ack_scheduled(sk)) {
5076                 /* We sent a data segment already. */
5077                 return;
5078         }
5079         __tcp_ack_snd_check(sk, 1);
5080 }
5081 
5082 /*
5083  *      This routine is only called when we have urgent data
5084  *      signaled. Its the 'slow' part of tcp_urg. It could be
5085  *      moved inline now as tcp_urg is only called from one
5086  *      place. We handle URGent data wrong. We have to - as
5087  *      BSD still doesn't use the correction from RFC961.
5088  *      For 1003.1g we should support a new option TCP_STDURG to permit
5089  *      either form (or just set the sysctl tcp_stdurg).
5090  */
5091 
5092 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5093 {
5094         struct tcp_sock *tp = tcp_sk(sk);
5095         u32 ptr = ntohs(th->urg_ptr);
5096 
5097         if (ptr && !sysctl_tcp_stdurg)
5098                 ptr--;
5099         ptr += ntohl(th->seq);
5100 
5101         /* Ignore urgent data that we've already seen and read. */
5102         if (after(tp->copied_seq, ptr))
5103                 return;
5104 
5105         /* Do not replay urg ptr.
5106          *
5107          * NOTE: interesting situation not covered by specs.
5108          * Misbehaving sender may send urg ptr, pointing to segment,
5109          * which we already have in ofo queue. We are not able to fetch
5110          * such data and will stay in TCP_URG_NOTYET until will be eaten
5111          * by recvmsg(). Seems, we are not obliged to handle such wicked
5112          * situations. But it is worth to think about possibility of some
5113          * DoSes using some hypothetical application level deadlock.
5114          */
5115         if (before(ptr, tp->rcv_nxt))
5116                 return;
5117 
5118         /* Do we already have a newer (or duplicate) urgent pointer? */
5119         if (tp->urg_data && !after(ptr, tp->urg_seq))
5120                 return;
5121 
5122         /* Tell the world about our new urgent pointer. */
5123         sk_send_sigurg(sk);
5124 
5125         /* We may be adding urgent data when the last byte read was
5126          * urgent. To do this requires some care. We cannot just ignore
5127          * tp->copied_seq since we would read the last urgent byte again
5128          * as data, nor can we alter copied_seq until this data arrives
5129          * or we break the semantics of SIOCATMARK (and thus sockatmark())
5130          *
5131          * NOTE. Double Dutch. Rendering to plain English: author of comment
5132          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
5133          * and expect that both A and B disappear from stream. This is _wrong_.
5134          * Though this happens in BSD with high probability, this is occasional.
5135          * Any application relying on this is buggy. Note also, that fix "works"
5136          * only in this artificial test. Insert some normal data between A and B and we will
5137          * decline of BSD again. Verdict: it is better to remove to trap
5138          * buggy users.
5139          */
5140         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5141             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5142                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5143                 tp->copied_seq++;
5144                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5145                         __skb_unlink(skb, &sk->sk_receive_queue);
5146                         __kfree_skb(skb);
5147                 }
5148         }
5149 
5150         tp->urg_data = TCP_URG_NOTYET;
5151         tp->urg_seq = ptr;
5152 
5153         /* Disable header prediction. */
5154         tp->pred_flags = 0;
5155 }
5156 
5157 /* This is the 'fast' part of urgent handling. */
5158 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5159 {
5160         struct tcp_sock *tp = tcp_sk(sk);
5161 
5162         /* Check if we get a new urgent pointer - normally not. */
5163         if (th->urg)
5164                 tcp_check_urg(sk, th);
5165 
5166         /* Do we wait for any urgent data? - normally not... */
5167         if (tp->urg_data == TCP_URG_NOTYET) {
5168                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5169                           th->syn;
5170 
5171                 /* Is the urgent pointer pointing into this packet? */
5172                 if (ptr < skb->len) {
5173                         u8 tmp;
5174                         if (skb_copy_bits(skb, ptr, &tmp, 1))
5175                                 BUG();
5176                         tp->urg_data = TCP_URG_VALID | tmp;
5177                         if (!sock_flag(sk, SOCK_DEAD))
5178                                 sk->sk_data_ready(sk);
5179                 }
5180         }
5181 }
5182 
5183 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5184 {
5185         struct tcp_sock *tp = tcp_sk(sk);
5186         int chunk = skb->len - hlen;
5187         int err;
5188 
5189         if (skb_csum_unnecessary(skb))
5190                 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5191         else
5192                 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5193 
5194         if (!err) {
5195                 tp->ucopy.len -= chunk;
5196                 tp->copied_seq += chunk;
5197                 tcp_rcv_space_adjust(sk);
5198         }
5199 
5200         return err;
5201 }
5202 
5203 /* Accept RST for rcv_nxt - 1 after a FIN.
5204  * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5205  * FIN is sent followed by a RST packet. The RST is sent with the same
5206  * sequence number as the FIN, and thus according to RFC 5961 a challenge
5207  * ACK should be sent. However, Mac OSX rate limits replies to challenge
5208  * ACKs on the closed socket. In addition middleboxes can drop either the
5209  * challenge ACK or a subsequent RST.
5210  */
5211 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5212 {
5213         struct tcp_sock *tp = tcp_sk(sk);
5214 
5215         return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5216                         (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5217                                                TCPF_CLOSING));
5218 }
5219 
5220 /* Does PAWS and seqno based validation of an incoming segment, flags will
5221  * play significant role here.
5222  */
5223 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5224                                   const struct tcphdr *th, int syn_inerr)
5225 {
5226         struct tcp_sock *tp = tcp_sk(sk);
5227         bool rst_seq_match = false;
5228 
5229         /* RFC1323: H1. Apply PAWS check first. */
5230         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5231             tcp_paws_discard(sk, skb)) {
5232                 if (!th->rst) {
5233                         NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5234                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5235                                                   LINUX_MIB_TCPACKSKIPPEDPAWS,
5236                                                   &tp->last_oow_ack_time))
5237                                 tcp_send_dupack(sk, skb);
5238                         goto discard;
5239                 }
5240                 /* Reset is accepted even if it did not pass PAWS. */
5241         }
5242 
5243         /* Step 1: check sequence number */
5244         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5245                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5246                  * (RST) segments are validated by checking their SEQ-fields."
5247                  * And page 69: "If an incoming segment is not acceptable,
5248                  * an acknowledgment should be sent in reply (unless the RST
5249                  * bit is set, if so drop the segment and return)".
5250                  */
5251                 if (!th->rst) {
5252                         if (th->syn)
5253                                 goto syn_challenge;
5254                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5255                                                   LINUX_MIB_TCPACKSKIPPEDSEQ,
5256                                                   &tp->last_oow_ack_time))
5257                                 tcp_send_dupack(sk, skb);
5258                 } else if (tcp_reset_check(sk, skb)) {
5259                         tcp_reset(sk);
5260                 }
5261                 goto discard;
5262         }
5263 
5264         /* Step 2: check RST bit */
5265         if (th->rst) {
5266                 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5267                  * FIN and SACK too if available):
5268                  * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5269                  * the right-most SACK block,
5270                  * then
5271                  *     RESET the connection
5272                  * else
5273                  *     Send a challenge ACK
5274                  */
5275                 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5276                     tcp_reset_check(sk, skb)) {
5277                         rst_seq_match = true;
5278                 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5279                         struct tcp_sack_block *sp = &tp->selective_acks[0];
5280                         int max_sack = sp[0].end_seq;
5281                         int this_sack;
5282 
5283                         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5284                              ++this_sack) {
5285                                 max_sack = after(sp[this_sack].end_seq,
5286                                                  max_sack) ?
5287                                         sp[this_sack].end_seq : max_sack;
5288                         }
5289 
5290                         if (TCP_SKB_CB(skb)->seq == max_sack)
5291                                 rst_seq_match = true;
5292                 }
5293 
5294                 if (rst_seq_match)
5295                         tcp_reset(sk);
5296                 else {
5297                         /* Disable TFO if RST is out-of-order
5298                          * and no data has been received
5299                          * for current active TFO socket
5300                          */
5301                         if (tp->syn_fastopen && !tp->data_segs_in &&
5302                             sk->sk_state == TCP_ESTABLISHED)
5303                                 tcp_fastopen_active_disable(sk);
5304                         tcp_send_challenge_ack(sk, skb);
5305                 }
5306                 goto discard;
5307         }
5308 
5309         /* step 3: check security and precedence [ignored] */
5310 
5311         /* step 4: Check for a SYN
5312          * RFC 5961 4.2 : Send a challenge ack
5313          */
5314         if (th->syn) {
5315 syn_challenge:
5316                 if (syn_inerr)
5317                         TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5318                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5319                 tcp_send_challenge_ack(sk, skb);
5320                 goto discard;
5321         }
5322 
5323         return true;
5324 
5325 discard:
5326         tcp_drop(sk, skb);
5327         return false;
5328 }
5329 
5330 /*
5331  *      TCP receive function for the ESTABLISHED state.
5332  *
5333  *      It is split into a fast path and a slow path. The fast path is
5334  *      disabled when:
5335  *      - A zero window was announced from us - zero window probing
5336  *        is only handled properly in the slow path.
5337  *      - Out of order segments arrived.
5338  *      - Urgent data is expected.
5339  *      - There is no buffer space left
5340  *      - Unexpected TCP flags/window values/header lengths are received
5341  *        (detected by checking the TCP header against pred_flags)
5342  *      - Data is sent in both directions. Fast path only supports pure senders
5343  *        or pure receivers (this means either the sequence number or the ack
5344  *        value must stay constant)
5345  *      - Unexpected TCP option.
5346  *
5347  *      When these conditions are not satisfied it drops into a standard
5348  *      receive procedure patterned after RFC793 to handle all cases.
5349  *      The first three cases are guaranteed by proper pred_flags setting,
5350  *      the rest is checked inline. Fast processing is turned on in
5351  *      tcp_data_queue when everything is OK.
5352  */
5353 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5354                          const struct tcphdr *th, unsigned int len)
5355 {
5356         struct tcp_sock *tp = tcp_sk(sk);
5357 
5358         skb_mstamp_get(&tp->tcp_mstamp);
5359         if (unlikely(!sk->sk_rx_dst))
5360                 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5361         /*
5362          *      Header prediction.
5363          *      The code loosely follows the one in the famous
5364          *      "30 instruction TCP receive" Van Jacobson mail.
5365          *
5366          *      Van's trick is to deposit buffers into socket queue
5367          *      on a device interrupt, to call tcp_recv function
5368          *      on the receive process context and checksum and copy
5369          *      the buffer to user space. smart...
5370          *
5371          *      Our current scheme is not silly either but we take the
5372          *      extra cost of the net_bh soft interrupt processing...
5373          *      We do checksum and copy also but from device to kernel.
5374          */
5375 
5376         tp->rx_opt.saw_tstamp = 0;
5377 
5378         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5379          *      if header_prediction is to be made
5380          *      'S' will always be tp->tcp_header_len >> 2
5381          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5382          *  turn it off (when there are holes in the receive
5383          *       space for instance)
5384          *      PSH flag is ignored.
5385          */
5386 
5387         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5388             TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5389             !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5390                 int tcp_header_len = tp->tcp_header_len;
5391 
5392                 /* Timestamp header prediction: tcp_header_len
5393                  * is automatically equal to th->doff*4 due to pred_flags
5394                  * match.
5395                  */
5396 
5397                 /* Check timestamp */
5398                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5399                         /* No? Slow path! */
5400                         if (!tcp_parse_aligned_timestamp(tp, th))
5401                                 goto slow_path;
5402 
5403                         /* If PAWS failed, check it more carefully in slow path */
5404                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5405                                 goto slow_path;
5406 
5407                         /* DO NOT update ts_recent here, if checksum fails
5408                          * and timestamp was corrupted part, it will result
5409                          * in a hung connection since we will drop all
5410                          * future packets due to the PAWS test.
5411                          */
5412                 }
5413 
5414                 if (len <= tcp_header_len) {
5415                         /* Bulk data transfer: sender */
5416                         if (len == tcp_header_len) {
5417                                 /* Predicted packet is in window by definition.
5418                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5419                                  * Hence, check seq<=rcv_wup reduces to:
5420                                  */
5421                                 if (tcp_header_len ==
5422                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5423                                     tp->rcv_nxt == tp->rcv_wup)
5424                                         tcp_store_ts_recent(tp);
5425 
5426                                 /* We know that such packets are checksummed
5427                                  * on entry.
5428                                  */
5429                                 tcp_ack(sk, skb, 0);
5430                                 __kfree_skb(skb);
5431                                 tcp_data_snd_check(sk);
5432                                 return;
5433                         } else { /* Header too small */
5434                                 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5435                                 goto discard;
5436                         }
5437                 } else {
5438                         int eaten = 0;
5439                         bool fragstolen = false;
5440 
5441                         if (tp->ucopy.task == current &&
5442                             tp->copied_seq == tp->rcv_nxt &&
5443                             len - tcp_header_len <= tp->ucopy.len &&
5444                             sock_owned_by_user(sk)) {
5445                                 __set_current_state(TASK_RUNNING);
5446 
5447                                 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5448                                         /* Predicted packet is in window by definition.
5449                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5450                                          * Hence, check seq<=rcv_wup reduces to:
5451                                          */
5452                                         if (tcp_header_len ==
5453                                             (sizeof(struct tcphdr) +
5454                                              TCPOLEN_TSTAMP_ALIGNED) &&
5455                                             tp->rcv_nxt == tp->rcv_wup)
5456                                                 tcp_store_ts_recent(tp);
5457 
5458                                         tcp_rcv_rtt_measure_ts(sk, skb);
5459 
5460                                         __skb_pull(skb, tcp_header_len);
5461                                         tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5462                                         NET_INC_STATS(sock_net(sk),
5463                                                         LINUX_MIB_TCPHPHITSTOUSER);
5464                                         eaten = 1;
5465                                 }
5466                         }
5467                         if (!eaten) {
5468                                 if (tcp_checksum_complete(skb))
5469                                         goto csum_error;
5470 
5471                                 if ((int)skb->truesize > sk->sk_forward_alloc)
5472                                         goto step5;
5473 
5474                                 /* Predicted packet is in window by definition.
5475                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5476                                  * Hence, check seq<=rcv_wup reduces to:
5477                                  */
5478                                 if (tcp_header_len ==
5479                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5480                                     tp->rcv_nxt == tp->rcv_wup)
5481                                         tcp_store_ts_recent(tp);
5482 
5483                                 tcp_rcv_rtt_measure_ts(sk, skb);
5484 
5485                                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5486 
5487                                 /* Bulk data transfer: receiver */
5488                                 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5489                                                       &fragstolen);
5490                         }
5491 
5492                         tcp_event_data_recv(sk, skb);
5493 
5494                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5495                                 /* Well, only one small jumplet in fast path... */
5496                                 tcp_ack(sk, skb, FLAG_DATA);
5497                                 tcp_data_snd_check(sk);
5498                                 if (!inet_csk_ack_scheduled(sk))
5499                                         goto no_ack;
5500                         }
5501 
5502                         __tcp_ack_snd_check(sk, 0);
5503 no_ack:
5504                         if (eaten)
5505                                 kfree_skb_partial(skb, fragstolen);
5506                         sk->sk_data_ready(sk);
5507                         return;
5508                 }
5509         }
5510 
5511 slow_path:
5512         if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5513                 goto csum_error;
5514 
5515         if (!th->ack && !th->rst && !th->syn)
5516                 goto discard;
5517 
5518         /*
5519          *      Standard slow path.
5520          */
5521 
5522         if (!tcp_validate_incoming(sk, skb, th, 1))
5523                 return;
5524 
5525 step5:
5526         if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5527                 goto discard;
5528 
5529         tcp_rcv_rtt_measure_ts(sk, skb);
5530 
5531         /* Process urgent data. */
5532         tcp_urg(sk, skb, th);
5533 
5534         /* step 7: process the segment text */
5535         tcp_data_queue(sk, skb);
5536 
5537         tcp_data_snd_check(sk);
5538         tcp_ack_snd_check(sk);
5539         return;
5540 
5541 csum_error:
5542         TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5543         TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5544 
5545 discard:
5546         tcp_drop(sk, skb);
5547 }
5548 EXPORT_SYMBOL(tcp_rcv_established);
5549 
5550 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5551 {
5552         struct tcp_sock *tp = tcp_sk(sk);
5553         struct inet_connection_sock *icsk = inet_csk(sk);
5554 
5555         tcp_set_state(sk, TCP_ESTABLISHED);
5556         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5557 
5558         if (skb) {
5559                 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5560                 security_inet_conn_established(sk, skb);
5561         }
5562 
5563         /* Make sure socket is routed, for correct metrics.  */
5564         icsk->icsk_af_ops->rebuild_header(sk);
5565 
5566         tcp_init_metrics(sk);
5567 
5568         tcp_init_congestion_control(sk);
5569 
5570         /* Prevent spurious tcp_cwnd_restart() on first data
5571          * packet.
5572          */
5573         tp->lsndtime = tcp_time_stamp;
5574 
5575         tcp_init_buffer_space(sk);
5576 
5577         if (sock_flag(sk, SOCK_KEEPOPEN))
5578                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5579 
5580         if (!tp->rx_opt.snd_wscale)
5581                 __tcp_fast_path_on(tp, tp->snd_wnd);
5582         else
5583                 tp->pred_flags = 0;
5584 
5585 }
5586 
5587 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5588                                     struct tcp_fastopen_cookie *cookie)
5589 {
5590         struct tcp_sock *tp = tcp_sk(sk);
5591         struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5592         u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5593         bool syn_drop = false;
5594 
5595         if (mss == tp->rx_opt.user_mss) {
5596                 struct tcp_options_received opt;
5597 
5598                 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5599                 tcp_clear_options(&opt);
5600                 opt.user_mss = opt.mss_clamp = 0;
5601                 tcp_parse_options(synack, &opt, 0, NULL);
5602                 mss = opt.mss_clamp;
5603         }
5604 
5605         if (!tp->syn_fastopen) {
5606                 /* Ignore an unsolicited cookie */
5607                 cookie->len = -1;
5608         } else if (tp->total_retrans) {
5609                 /* SYN timed out and the SYN-ACK neither has a cookie nor
5610                  * acknowledges data. Presumably the remote received only
5611                  * the retransmitted (regular) SYNs: either the original
5612                  * SYN-data or the corresponding SYN-ACK was dropped.
5613                  */
5614                 syn_drop = (cookie->len < 0 && data);
5615         } else if (cookie->len < 0 && !tp->syn_data) {
5616                 /* We requested a cookie but didn't get it. If we did not use
5617                  * the (old) exp opt format then try so next time (try_exp=1).
5618                  * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5619                  */
5620                 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5621         }
5622 
5623         tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5624 
5625         if (data) { /* Retransmit unacked data in SYN */
5626                 tcp_for_write_queue_from(data, sk) {
5627                         if (data == tcp_send_head(sk) ||
5628                             __tcp_retransmit_skb(sk, data, 1))
5629                                 break;
5630                 }
5631                 tcp_rearm_rto(sk);
5632                 NET_INC_STATS(sock_net(sk),
5633                                 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5634                 return true;
5635         }
5636         tp->syn_data_acked = tp->syn_data;
5637         if (tp->syn_data_acked)
5638                 NET_INC_STATS(sock_net(sk),
5639                                 LINUX_MIB_TCPFASTOPENACTIVE);
5640 
5641         tcp_fastopen_add_skb(sk, synack);
5642 
5643         return false;
5644 }
5645 
5646 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5647                                          const struct tcphdr *th)
5648 {
5649         struct inet_connection_sock *icsk = inet_csk(sk);
5650         struct tcp_sock *tp = tcp_sk(sk);
5651         struct tcp_fastopen_cookie foc = { .len = -1 };
5652         int saved_clamp = tp->rx_opt.mss_clamp;
5653         bool fastopen_fail;
5654 
5655         tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5656         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5657                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5658 
5659         if (th->ack) {
5660                 /* rfc793:
5661                  * "If the state is SYN-SENT then
5662                  *    first check the ACK bit
5663                  *      If the ACK bit is set
5664                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5665                  *        a reset (unless the RST bit is set, if so drop
5666                  *        the segment and return)"
5667                  */
5668                 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5669                     after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5670                         goto reset_and_undo;
5671 
5672                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5673                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5674                              tcp_time_stamp)) {
5675                         NET_INC_STATS(sock_net(sk),
5676                                         LINUX_MIB_PAWSACTIVEREJECTED);
5677                         goto reset_and_undo;
5678                 }
5679 
5680                 /* Now ACK is acceptable.
5681                  *
5682                  * "If the RST bit is set
5683                  *    If the ACK was acceptable then sig