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

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

Version: ~ [ linux-5.13-rc5 ] ~ [ linux-5.12.9 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.42 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.124 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.193 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.235 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.271 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.271 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ 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 = 1;
 83 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
 84 int sysctl_tcp_max_reordering __read_mostly = 300;
 85 EXPORT_SYMBOL(sysctl_tcp_reordering);
 86 int sysctl_tcp_dsack __read_mostly = 1;
 87 int sysctl_tcp_app_win __read_mostly = 31;
 88 int sysctl_tcp_adv_win_scale __read_mostly = 1;
 89 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
 90 
 91 /* rfc5961 challenge ack rate limiting */
 92 int sysctl_tcp_challenge_ack_limit = 100;
 93 
 94 int sysctl_tcp_stdurg __read_mostly;
 95 int sysctl_tcp_rfc1337 __read_mostly;
 96 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
 97 int sysctl_tcp_frto __read_mostly = 2;
 98 
 99 int sysctl_tcp_thin_dupack __read_mostly;
100 
101 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
102 int sysctl_tcp_early_retrans __read_mostly = 3;
103 int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
104 
105 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
106 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
107 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
108 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
109 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
110 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
111 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
112 #define FLAG_LOST_RETRANS       0x80 /* This ACK marks some retransmission lost */
113 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
114 #define FLAG_ORIG_SACK_ACKED    0x200 /* Never retransmitted data are (s)acked  */
115 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
116 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
117 #define FLAG_SACK_RENEGING      0x2000 /* snd_una advanced to a sacked seq */
118 #define FLAG_UPDATE_TS_RECENT   0x4000 /* tcp_replace_ts_recent() */
119 
120 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
121 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
122 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
123 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
124 
125 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
126 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
127 
128 /* Adapt the MSS value used to make delayed ack decision to the
129  * real world.
130  */
131 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
132 {
133         struct inet_connection_sock *icsk = inet_csk(sk);
134         const unsigned int lss = icsk->icsk_ack.last_seg_size;
135         unsigned int len;
136 
137         icsk->icsk_ack.last_seg_size = 0;
138 
139         /* skb->len may jitter because of SACKs, even if peer
140          * sends good full-sized frames.
141          */
142         len = skb_shinfo(skb)->gso_size ? : skb->len;
143         if (len >= icsk->icsk_ack.rcv_mss) {
144                 icsk->icsk_ack.rcv_mss = len;
145         } else {
146                 /* Otherwise, we make more careful check taking into account,
147                  * that SACKs block is variable.
148                  *
149                  * "len" is invariant segment length, including TCP header.
150                  */
151                 len += skb->data - skb_transport_header(skb);
152                 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
153                     /* If PSH is not set, packet should be
154                      * full sized, provided peer TCP is not badly broken.
155                      * This observation (if it is correct 8)) allows
156                      * to handle super-low mtu links fairly.
157                      */
158                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
159                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
160                         /* Subtract also invariant (if peer is RFC compliant),
161                          * tcp header plus fixed timestamp option length.
162                          * Resulting "len" is MSS free of SACK jitter.
163                          */
164                         len -= tcp_sk(sk)->tcp_header_len;
165                         icsk->icsk_ack.last_seg_size = len;
166                         if (len == lss) {
167                                 icsk->icsk_ack.rcv_mss = len;
168                                 return;
169                         }
170                 }
171                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
172                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
173                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
174         }
175 }
176 
177 static void tcp_incr_quickack(struct sock *sk)
178 {
179         struct inet_connection_sock *icsk = inet_csk(sk);
180         unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
181 
182         if (quickacks == 0)
183                 quickacks = 2;
184         if (quickacks > icsk->icsk_ack.quick)
185                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
186 }
187 
188 static void tcp_enter_quickack_mode(struct sock *sk)
189 {
190         struct inet_connection_sock *icsk = inet_csk(sk);
191         tcp_incr_quickack(sk);
192         icsk->icsk_ack.pingpong = 0;
193         icsk->icsk_ack.ato = TCP_ATO_MIN;
194 }
195 
196 /* Send ACKs quickly, if "quick" count is not exhausted
197  * and the session is not interactive.
198  */
199 
200 static bool tcp_in_quickack_mode(struct sock *sk)
201 {
202         const struct inet_connection_sock *icsk = inet_csk(sk);
203         const struct dst_entry *dst = __sk_dst_get(sk);
204 
205         return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
206                 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
207 }
208 
209 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
210 {
211         if (tp->ecn_flags & TCP_ECN_OK)
212                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
213 }
214 
215 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
216 {
217         if (tcp_hdr(skb)->cwr)
218                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
219 }
220 
221 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
222 {
223         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
224 }
225 
226 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
227 {
228         switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
229         case INET_ECN_NOT_ECT:
230                 /* Funny extension: if ECT is not set on a segment,
231                  * and we already seen ECT on a previous segment,
232                  * it is probably a retransmit.
233                  */
234                 if (tp->ecn_flags & TCP_ECN_SEEN)
235                         tcp_enter_quickack_mode((struct sock *)tp);
236                 break;
237         case INET_ECN_CE:
238                 if (tcp_ca_needs_ecn((struct sock *)tp))
239                         tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
240 
241                 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
242                         /* Better not delay acks, sender can have a very low cwnd */
243                         tcp_enter_quickack_mode((struct sock *)tp);
244                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
245                 }
246                 tp->ecn_flags |= TCP_ECN_SEEN;
247                 break;
248         default:
249                 if (tcp_ca_needs_ecn((struct sock *)tp))
250                         tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
251                 tp->ecn_flags |= TCP_ECN_SEEN;
252                 break;
253         }
254 }
255 
256 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
257 {
258         if (tp->ecn_flags & TCP_ECN_OK)
259                 __tcp_ecn_check_ce(tp, skb);
260 }
261 
262 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
263 {
264         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
265                 tp->ecn_flags &= ~TCP_ECN_OK;
266 }
267 
268 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
269 {
270         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
271                 tp->ecn_flags &= ~TCP_ECN_OK;
272 }
273 
274 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
275 {
276         if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
277                 return true;
278         return false;
279 }
280 
281 /* Buffer size and advertised window tuning.
282  *
283  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
284  */
285 
286 static void tcp_sndbuf_expand(struct sock *sk)
287 {
288         const struct tcp_sock *tp = tcp_sk(sk);
289         int sndmem, per_mss;
290         u32 nr_segs;
291 
292         /* Worst case is non GSO/TSO : each frame consumes one skb
293          * and skb->head is kmalloced using power of two area of memory
294          */
295         per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
296                   MAX_TCP_HEADER +
297                   SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
298 
299         per_mss = roundup_pow_of_two(per_mss) +
300                   SKB_DATA_ALIGN(sizeof(struct sk_buff));
301 
302         nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
303         nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
304 
305         /* Fast Recovery (RFC 5681 3.2) :
306          * Cubic needs 1.7 factor, rounded to 2 to include
307          * extra cushion (application might react slowly to POLLOUT)
308          */
309         sndmem = 2 * nr_segs * per_mss;
310 
311         if (sk->sk_sndbuf < sndmem)
312                 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
313 }
314 
315 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
316  *
317  * All tcp_full_space() is split to two parts: "network" buffer, allocated
318  * forward and advertised in receiver window (tp->rcv_wnd) and
319  * "application buffer", required to isolate scheduling/application
320  * latencies from network.
321  * window_clamp is maximal advertised window. It can be less than
322  * tcp_full_space(), in this case tcp_full_space() - window_clamp
323  * is reserved for "application" buffer. The less window_clamp is
324  * the smoother our behaviour from viewpoint of network, but the lower
325  * throughput and the higher sensitivity of the connection to losses. 8)
326  *
327  * rcv_ssthresh is more strict window_clamp used at "slow start"
328  * phase to predict further behaviour of this connection.
329  * It is used for two goals:
330  * - to enforce header prediction at sender, even when application
331  *   requires some significant "application buffer". It is check #1.
332  * - to prevent pruning of receive queue because of misprediction
333  *   of receiver window. Check #2.
334  *
335  * The scheme does not work when sender sends good segments opening
336  * window and then starts to feed us spaghetti. But it should work
337  * in common situations. Otherwise, we have to rely on queue collapsing.
338  */
339 
340 /* Slow part of check#2. */
341 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
342 {
343         struct tcp_sock *tp = tcp_sk(sk);
344         /* Optimize this! */
345         int truesize = tcp_win_from_space(skb->truesize) >> 1;
346         int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
347 
348         while (tp->rcv_ssthresh <= window) {
349                 if (truesize <= skb->len)
350                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
351 
352                 truesize >>= 1;
353                 window >>= 1;
354         }
355         return 0;
356 }
357 
358 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
359 {
360         struct tcp_sock *tp = tcp_sk(sk);
361 
362         /* Check #1 */
363         if (tp->rcv_ssthresh < tp->window_clamp &&
364             (int)tp->rcv_ssthresh < tcp_space(sk) &&
365             !tcp_under_memory_pressure(sk)) {
366                 int incr;
367 
368                 /* Check #2. Increase window, if skb with such overhead
369                  * will fit to rcvbuf in future.
370                  */
371                 if (tcp_win_from_space(skb->truesize) <= skb->len)
372                         incr = 2 * tp->advmss;
373                 else
374                         incr = __tcp_grow_window(sk, skb);
375 
376                 if (incr) {
377                         incr = max_t(int, incr, 2 * skb->len);
378                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
379                                                tp->window_clamp);
380                         inet_csk(sk)->icsk_ack.quick |= 1;
381                 }
382         }
383 }
384 
385 /* 3. Tuning rcvbuf, when connection enters established state. */
386 static void tcp_fixup_rcvbuf(struct sock *sk)
387 {
388         u32 mss = tcp_sk(sk)->advmss;
389         int rcvmem;
390 
391         rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
392                  tcp_default_init_rwnd(mss);
393 
394         /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
395          * Allow enough cushion so that sender is not limited by our window
396          */
397         if (sysctl_tcp_moderate_rcvbuf)
398                 rcvmem <<= 2;
399 
400         if (sk->sk_rcvbuf < rcvmem)
401                 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
402 }
403 
404 /* 4. Try to fixup all. It is made immediately after connection enters
405  *    established state.
406  */
407 void tcp_init_buffer_space(struct sock *sk)
408 {
409         struct tcp_sock *tp = tcp_sk(sk);
410         int maxwin;
411 
412         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
413                 tcp_fixup_rcvbuf(sk);
414         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
415                 tcp_sndbuf_expand(sk);
416 
417         tp->rcvq_space.space = tp->rcv_wnd;
418         tp->rcvq_space.time = tcp_time_stamp;
419         tp->rcvq_space.seq = tp->copied_seq;
420 
421         maxwin = tcp_full_space(sk);
422 
423         if (tp->window_clamp >= maxwin) {
424                 tp->window_clamp = maxwin;
425 
426                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
427                         tp->window_clamp = max(maxwin -
428                                                (maxwin >> sysctl_tcp_app_win),
429                                                4 * tp->advmss);
430         }
431 
432         /* Force reservation of one segment. */
433         if (sysctl_tcp_app_win &&
434             tp->window_clamp > 2 * tp->advmss &&
435             tp->window_clamp + tp->advmss > maxwin)
436                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
437 
438         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
439         tp->snd_cwnd_stamp = tcp_time_stamp;
440 }
441 
442 /* 5. Recalculate window clamp after socket hit its memory bounds. */
443 static void tcp_clamp_window(struct sock *sk)
444 {
445         struct tcp_sock *tp = tcp_sk(sk);
446         struct inet_connection_sock *icsk = inet_csk(sk);
447 
448         icsk->icsk_ack.quick = 0;
449 
450         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
451             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
452             !tcp_under_memory_pressure(sk) &&
453             sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
454                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
455                                     sysctl_tcp_rmem[2]);
456         }
457         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
458                 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
459 }
460 
461 /* Initialize RCV_MSS value.
462  * RCV_MSS is an our guess about MSS used by the peer.
463  * We haven't any direct information about the MSS.
464  * It's better to underestimate the RCV_MSS rather than overestimate.
465  * Overestimations make us ACKing less frequently than needed.
466  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
467  */
468 void tcp_initialize_rcv_mss(struct sock *sk)
469 {
470         const struct tcp_sock *tp = tcp_sk(sk);
471         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
472 
473         hint = min(hint, tp->rcv_wnd / 2);
474         hint = min(hint, TCP_MSS_DEFAULT);
475         hint = max(hint, TCP_MIN_MSS);
476 
477         inet_csk(sk)->icsk_ack.rcv_mss = hint;
478 }
479 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
480 
481 /* Receiver "autotuning" code.
482  *
483  * The algorithm for RTT estimation w/o timestamps is based on
484  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
485  * <http://public.lanl.gov/radiant/pubs.html#DRS>
486  *
487  * More detail on this code can be found at
488  * <http://staff.psc.edu/jheffner/>,
489  * though this reference is out of date.  A new paper
490  * is pending.
491  */
492 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
493 {
494         u32 new_sample = tp->rcv_rtt_est.rtt;
495         long m = sample;
496 
497         if (m == 0)
498                 m = 1;
499 
500         if (new_sample != 0) {
501                 /* If we sample in larger samples in the non-timestamp
502                  * case, we could grossly overestimate the RTT especially
503                  * with chatty applications or bulk transfer apps which
504                  * are stalled on filesystem I/O.
505                  *
506                  * Also, since we are only going for a minimum in the
507                  * non-timestamp case, we do not smooth things out
508                  * else with timestamps disabled convergence takes too
509                  * long.
510                  */
511                 if (!win_dep) {
512                         m -= (new_sample >> 3);
513                         new_sample += m;
514                 } else {
515                         m <<= 3;
516                         if (m < new_sample)
517                                 new_sample = m;
518                 }
519         } else {
520                 /* No previous measure. */
521                 new_sample = m << 3;
522         }
523 
524         if (tp->rcv_rtt_est.rtt != new_sample)
525                 tp->rcv_rtt_est.rtt = new_sample;
526 }
527 
528 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
529 {
530         if (tp->rcv_rtt_est.time == 0)
531                 goto new_measure;
532         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
533                 return;
534         tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
535 
536 new_measure:
537         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
538         tp->rcv_rtt_est.time = tcp_time_stamp;
539 }
540 
541 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
542                                           const struct sk_buff *skb)
543 {
544         struct tcp_sock *tp = tcp_sk(sk);
545         if (tp->rx_opt.rcv_tsecr &&
546             (TCP_SKB_CB(skb)->end_seq -
547              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
548                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
549 }
550 
551 /*
552  * This function should be called every time data is copied to user space.
553  * It calculates the appropriate TCP receive buffer space.
554  */
555 void tcp_rcv_space_adjust(struct sock *sk)
556 {
557         struct tcp_sock *tp = tcp_sk(sk);
558         int time;
559         int copied;
560 
561         time = tcp_time_stamp - tp->rcvq_space.time;
562         if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
563                 return;
564 
565         /* Number of bytes copied to user in last RTT */
566         copied = tp->copied_seq - tp->rcvq_space.seq;
567         if (copied <= tp->rcvq_space.space)
568                 goto new_measure;
569 
570         /* A bit of theory :
571          * copied = bytes received in previous RTT, our base window
572          * To cope with packet losses, we need a 2x factor
573          * To cope with slow start, and sender growing its cwin by 100 %
574          * every RTT, we need a 4x factor, because the ACK we are sending
575          * now is for the next RTT, not the current one :
576          * <prev RTT . ><current RTT .. ><next RTT .... >
577          */
578 
579         if (sysctl_tcp_moderate_rcvbuf &&
580             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
581                 int rcvwin, rcvmem, rcvbuf;
582 
583                 /* minimal window to cope with packet losses, assuming
584                  * steady state. Add some cushion because of small variations.
585                  */
586                 rcvwin = (copied << 1) + 16 * tp->advmss;
587 
588                 /* If rate increased by 25%,
589                  *      assume slow start, rcvwin = 3 * copied
590                  * If rate increased by 50%,
591                  *      assume sender can use 2x growth, rcvwin = 4 * copied
592                  */
593                 if (copied >=
594                     tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
595                         if (copied >=
596                             tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
597                                 rcvwin <<= 1;
598                         else
599                                 rcvwin += (rcvwin >> 1);
600                 }
601 
602                 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
603                 while (tcp_win_from_space(rcvmem) < tp->advmss)
604                         rcvmem += 128;
605 
606                 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
607                 if (rcvbuf > sk->sk_rcvbuf) {
608                         sk->sk_rcvbuf = rcvbuf;
609 
610                         /* Make the window clamp follow along.  */
611                         tp->window_clamp = rcvwin;
612                 }
613         }
614         tp->rcvq_space.space = copied;
615 
616 new_measure:
617         tp->rcvq_space.seq = tp->copied_seq;
618         tp->rcvq_space.time = tcp_time_stamp;
619 }
620 
621 /* There is something which you must keep in mind when you analyze the
622  * behavior of the tp->ato delayed ack timeout interval.  When a
623  * connection starts up, we want to ack as quickly as possible.  The
624  * problem is that "good" TCP's do slow start at the beginning of data
625  * transmission.  The means that until we send the first few ACK's the
626  * sender will sit on his end and only queue most of his data, because
627  * he can only send snd_cwnd unacked packets at any given time.  For
628  * each ACK we send, he increments snd_cwnd and transmits more of his
629  * queue.  -DaveM
630  */
631 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
632 {
633         struct tcp_sock *tp = tcp_sk(sk);
634         struct inet_connection_sock *icsk = inet_csk(sk);
635         u32 now;
636 
637         inet_csk_schedule_ack(sk);
638 
639         tcp_measure_rcv_mss(sk, skb);
640 
641         tcp_rcv_rtt_measure(tp);
642 
643         now = tcp_time_stamp;
644 
645         if (!icsk->icsk_ack.ato) {
646                 /* The _first_ data packet received, initialize
647                  * delayed ACK engine.
648                  */
649                 tcp_incr_quickack(sk);
650                 icsk->icsk_ack.ato = TCP_ATO_MIN;
651         } else {
652                 int m = now - icsk->icsk_ack.lrcvtime;
653 
654                 if (m <= TCP_ATO_MIN / 2) {
655                         /* The fastest case is the first. */
656                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
657                 } else if (m < icsk->icsk_ack.ato) {
658                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
659                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
660                                 icsk->icsk_ack.ato = icsk->icsk_rto;
661                 } else if (m > icsk->icsk_rto) {
662                         /* Too long gap. Apparently sender failed to
663                          * restart window, so that we send ACKs quickly.
664                          */
665                         tcp_incr_quickack(sk);
666                         sk_mem_reclaim(sk);
667                 }
668         }
669         icsk->icsk_ack.lrcvtime = now;
670 
671         tcp_ecn_check_ce(tp, skb);
672 
673         if (skb->len >= 128)
674                 tcp_grow_window(sk, skb);
675 }
676 
677 /* Called to compute a smoothed rtt estimate. The data fed to this
678  * routine either comes from timestamps, or from segments that were
679  * known _not_ to have been retransmitted [see Karn/Partridge
680  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
681  * piece by Van Jacobson.
682  * NOTE: the next three routines used to be one big routine.
683  * To save cycles in the RFC 1323 implementation it was better to break
684  * it up into three procedures. -- erics
685  */
686 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
687 {
688         struct tcp_sock *tp = tcp_sk(sk);
689         long m = mrtt_us; /* RTT */
690         u32 srtt = tp->srtt_us;
691 
692         /*      The following amusing code comes from Jacobson's
693          *      article in SIGCOMM '88.  Note that rtt and mdev
694          *      are scaled versions of rtt and mean deviation.
695          *      This is designed to be as fast as possible
696          *      m stands for "measurement".
697          *
698          *      On a 1990 paper the rto value is changed to:
699          *      RTO = rtt + 4 * mdev
700          *
701          * Funny. This algorithm seems to be very broken.
702          * These formulae increase RTO, when it should be decreased, increase
703          * too slowly, when it should be increased quickly, decrease too quickly
704          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
705          * does not matter how to _calculate_ it. Seems, it was trap
706          * that VJ failed to avoid. 8)
707          */
708         if (srtt != 0) {
709                 m -= (srtt >> 3);       /* m is now error in rtt est */
710                 srtt += m;              /* rtt = 7/8 rtt + 1/8 new */
711                 if (m < 0) {
712                         m = -m;         /* m is now abs(error) */
713                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
714                         /* This is similar to one of Eifel findings.
715                          * Eifel blocks mdev updates when rtt decreases.
716                          * This solution is a bit different: we use finer gain
717                          * for mdev in this case (alpha*beta).
718                          * Like Eifel it also prevents growth of rto,
719                          * but also it limits too fast rto decreases,
720                          * happening in pure Eifel.
721                          */
722                         if (m > 0)
723                                 m >>= 3;
724                 } else {
725                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
726                 }
727                 tp->mdev_us += m;               /* mdev = 3/4 mdev + 1/4 new */
728                 if (tp->mdev_us > tp->mdev_max_us) {
729                         tp->mdev_max_us = tp->mdev_us;
730                         if (tp->mdev_max_us > tp->rttvar_us)
731                                 tp->rttvar_us = tp->mdev_max_us;
732                 }
733                 if (after(tp->snd_una, tp->rtt_seq)) {
734                         if (tp->mdev_max_us < tp->rttvar_us)
735                                 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
736                         tp->rtt_seq = tp->snd_nxt;
737                         tp->mdev_max_us = tcp_rto_min_us(sk);
738                 }
739         } else {
740                 /* no previous measure. */
741                 srtt = m << 3;          /* take the measured time to be rtt */
742                 tp->mdev_us = m << 1;   /* make sure rto = 3*rtt */
743                 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
744                 tp->mdev_max_us = tp->rttvar_us;
745                 tp->rtt_seq = tp->snd_nxt;
746         }
747         tp->srtt_us = max(1U, srtt);
748 }
749 
750 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
751  * Note: TCP stack does not yet implement pacing.
752  * FQ packet scheduler can be used to implement cheap but effective
753  * TCP pacing, to smooth the burst on large writes when packets
754  * in flight is significantly lower than cwnd (or rwin)
755  */
756 int sysctl_tcp_pacing_ss_ratio __read_mostly = 200;
757 int sysctl_tcp_pacing_ca_ratio __read_mostly = 120;
758 
759 static void tcp_update_pacing_rate(struct sock *sk)
760 {
761         const struct tcp_sock *tp = tcp_sk(sk);
762         u64 rate;
763 
764         /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
765         rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
766 
767         /* current rate is (cwnd * mss) / srtt
768          * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
769          * In Congestion Avoidance phase, set it to 120 % the current rate.
770          *
771          * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
772          *       If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
773          *       end of slow start and should slow down.
774          */
775         if (tp->snd_cwnd < tp->snd_ssthresh / 2)
776                 rate *= sysctl_tcp_pacing_ss_ratio;
777         else
778                 rate *= sysctl_tcp_pacing_ca_ratio;
779 
780         rate *= max(tp->snd_cwnd, tp->packets_out);
781 
782         if (likely(tp->srtt_us))
783                 do_div(rate, tp->srtt_us);
784 
785         /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
786          * without any lock. We want to make sure compiler wont store
787          * intermediate values in this location.
788          */
789         ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
790                                                 sk->sk_max_pacing_rate);
791 }
792 
793 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
794  * routine referred to above.
795  */
796 static void tcp_set_rto(struct sock *sk)
797 {
798         const struct tcp_sock *tp = tcp_sk(sk);
799         /* Old crap is replaced with new one. 8)
800          *
801          * More seriously:
802          * 1. If rtt variance happened to be less 50msec, it is hallucination.
803          *    It cannot be less due to utterly erratic ACK generation made
804          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
805          *    to do with delayed acks, because at cwnd>2 true delack timeout
806          *    is invisible. Actually, Linux-2.4 also generates erratic
807          *    ACKs in some circumstances.
808          */
809         inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
810 
811         /* 2. Fixups made earlier cannot be right.
812          *    If we do not estimate RTO correctly without them,
813          *    all the algo is pure shit and should be replaced
814          *    with correct one. It is exactly, which we pretend to do.
815          */
816 
817         /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
818          * guarantees that rto is higher.
819          */
820         tcp_bound_rto(sk);
821 }
822 
823 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
824 {
825         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
826 
827         if (!cwnd)
828                 cwnd = TCP_INIT_CWND;
829         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
830 }
831 
832 /*
833  * Packet counting of FACK is based on in-order assumptions, therefore TCP
834  * disables it when reordering is detected
835  */
836 void tcp_disable_fack(struct tcp_sock *tp)
837 {
838         /* RFC3517 uses different metric in lost marker => reset on change */
839         if (tcp_is_fack(tp))
840                 tp->lost_skb_hint = NULL;
841         tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
842 }
843 
844 /* Take a notice that peer is sending D-SACKs */
845 static void tcp_dsack_seen(struct tcp_sock *tp)
846 {
847         tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
848 }
849 
850 static void tcp_update_reordering(struct sock *sk, const int metric,
851                                   const int ts)
852 {
853         struct tcp_sock *tp = tcp_sk(sk);
854         if (metric > tp->reordering) {
855                 int mib_idx;
856 
857                 tp->reordering = min(sysctl_tcp_max_reordering, metric);
858 
859                 /* This exciting event is worth to be remembered. 8) */
860                 if (ts)
861                         mib_idx = LINUX_MIB_TCPTSREORDER;
862                 else if (tcp_is_reno(tp))
863                         mib_idx = LINUX_MIB_TCPRENOREORDER;
864                 else if (tcp_is_fack(tp))
865                         mib_idx = LINUX_MIB_TCPFACKREORDER;
866                 else
867                         mib_idx = LINUX_MIB_TCPSACKREORDER;
868 
869                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
870 #if FASTRETRANS_DEBUG > 1
871                 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
872                          tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
873                          tp->reordering,
874                          tp->fackets_out,
875                          tp->sacked_out,
876                          tp->undo_marker ? tp->undo_retrans : 0);
877 #endif
878                 tcp_disable_fack(tp);
879         }
880 
881         if (metric > 0)
882                 tcp_disable_early_retrans(tp);
883 }
884 
885 /* This must be called before lost_out is incremented */
886 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
887 {
888         if (!tp->retransmit_skb_hint ||
889             before(TCP_SKB_CB(skb)->seq,
890                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
891                 tp->retransmit_skb_hint = skb;
892 
893         if (!tp->lost_out ||
894             after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
895                 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
896 }
897 
898 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
899 {
900         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
901                 tcp_verify_retransmit_hint(tp, skb);
902 
903                 tp->lost_out += tcp_skb_pcount(skb);
904                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
905         }
906 }
907 
908 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
909                                             struct sk_buff *skb)
910 {
911         tcp_verify_retransmit_hint(tp, skb);
912 
913         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
914                 tp->lost_out += tcp_skb_pcount(skb);
915                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
916         }
917 }
918 
919 /* This procedure tags the retransmission queue when SACKs arrive.
920  *
921  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
922  * Packets in queue with these bits set are counted in variables
923  * sacked_out, retrans_out and lost_out, correspondingly.
924  *
925  * Valid combinations are:
926  * Tag  InFlight        Description
927  * 0    1               - orig segment is in flight.
928  * S    0               - nothing flies, orig reached receiver.
929  * L    0               - nothing flies, orig lost by net.
930  * R    2               - both orig and retransmit are in flight.
931  * L|R  1               - orig is lost, retransmit is in flight.
932  * S|R  1               - orig reached receiver, retrans is still in flight.
933  * (L|S|R is logically valid, it could occur when L|R is sacked,
934  *  but it is equivalent to plain S and code short-curcuits it to S.
935  *  L|S is logically invalid, it would mean -1 packet in flight 8))
936  *
937  * These 6 states form finite state machine, controlled by the following events:
938  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
939  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
940  * 3. Loss detection event of two flavors:
941  *      A. Scoreboard estimator decided the packet is lost.
942  *         A'. Reno "three dupacks" marks head of queue lost.
943  *         A''. Its FACK modification, head until snd.fack is lost.
944  *      B. SACK arrives sacking SND.NXT at the moment, when the
945  *         segment was retransmitted.
946  * 4. D-SACK added new rule: D-SACK changes any tag to S.
947  *
948  * It is pleasant to note, that state diagram turns out to be commutative,
949  * so that we are allowed not to be bothered by order of our actions,
950  * when multiple events arrive simultaneously. (see the function below).
951  *
952  * Reordering detection.
953  * --------------------
954  * Reordering metric is maximal distance, which a packet can be displaced
955  * in packet stream. With SACKs we can estimate it:
956  *
957  * 1. SACK fills old hole and the corresponding segment was not
958  *    ever retransmitted -> reordering. Alas, we cannot use it
959  *    when segment was retransmitted.
960  * 2. The last flaw is solved with D-SACK. D-SACK arrives
961  *    for retransmitted and already SACKed segment -> reordering..
962  * Both of these heuristics are not used in Loss state, when we cannot
963  * account for retransmits accurately.
964  *
965  * SACK block validation.
966  * ----------------------
967  *
968  * SACK block range validation checks that the received SACK block fits to
969  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
970  * Note that SND.UNA is not included to the range though being valid because
971  * it means that the receiver is rather inconsistent with itself reporting
972  * SACK reneging when it should advance SND.UNA. Such SACK block this is
973  * perfectly valid, however, in light of RFC2018 which explicitly states
974  * that "SACK block MUST reflect the newest segment.  Even if the newest
975  * segment is going to be discarded ...", not that it looks very clever
976  * in case of head skb. Due to potentional receiver driven attacks, we
977  * choose to avoid immediate execution of a walk in write queue due to
978  * reneging and defer head skb's loss recovery to standard loss recovery
979  * procedure that will eventually trigger (nothing forbids us doing this).
980  *
981  * Implements also blockage to start_seq wrap-around. Problem lies in the
982  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
983  * there's no guarantee that it will be before snd_nxt (n). The problem
984  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
985  * wrap (s_w):
986  *
987  *         <- outs wnd ->                          <- wrapzone ->
988  *         u     e      n                         u_w   e_w  s n_w
989  *         |     |      |                          |     |   |  |
990  * |<------------+------+----- TCP seqno space --------------+---------->|
991  * ...-- <2^31 ->|                                           |<--------...
992  * ...---- >2^31 ------>|                                    |<--------...
993  *
994  * Current code wouldn't be vulnerable but it's better still to discard such
995  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
996  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
997  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
998  * equal to the ideal case (infinite seqno space without wrap caused issues).
999  *
1000  * With D-SACK the lower bound is extended to cover sequence space below
1001  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1002  * again, D-SACK block must not to go across snd_una (for the same reason as
1003  * for the normal SACK blocks, explained above). But there all simplicity
1004  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1005  * fully below undo_marker they do not affect behavior in anyway and can
1006  * therefore be safely ignored. In rare cases (which are more or less
1007  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1008  * fragmentation and packet reordering past skb's retransmission. To consider
1009  * them correctly, the acceptable range must be extended even more though
1010  * the exact amount is rather hard to quantify. However, tp->max_window can
1011  * be used as an exaggerated estimate.
1012  */
1013 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1014                                    u32 start_seq, u32 end_seq)
1015 {
1016         /* Too far in future, or reversed (interpretation is ambiguous) */
1017         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1018                 return false;
1019 
1020         /* Nasty start_seq wrap-around check (see comments above) */
1021         if (!before(start_seq, tp->snd_nxt))
1022                 return false;
1023 
1024         /* In outstanding window? ...This is valid exit for D-SACKs too.
1025          * start_seq == snd_una is non-sensical (see comments above)
1026          */
1027         if (after(start_seq, tp->snd_una))
1028                 return true;
1029 
1030         if (!is_dsack || !tp->undo_marker)
1031                 return false;
1032 
1033         /* ...Then it's D-SACK, and must reside below snd_una completely */
1034         if (after(end_seq, tp->snd_una))
1035                 return false;
1036 
1037         if (!before(start_seq, tp->undo_marker))
1038                 return true;
1039 
1040         /* Too old */
1041         if (!after(end_seq, tp->undo_marker))
1042                 return false;
1043 
1044         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1045          *   start_seq < undo_marker and end_seq >= undo_marker.
1046          */
1047         return !before(start_seq, end_seq - tp->max_window);
1048 }
1049 
1050 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1051  * Event "B". Later note: FACK people cheated me again 8), we have to account
1052  * for reordering! Ugly, but should help.
1053  *
1054  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1055  * less than what is now known to be received by the other end (derived from
1056  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1057  * retransmitted skbs to avoid some costly processing per ACKs.
1058  */
1059 static void tcp_mark_lost_retrans(struct sock *sk, int *flag)
1060 {
1061         const struct inet_connection_sock *icsk = inet_csk(sk);
1062         struct tcp_sock *tp = tcp_sk(sk);
1063         struct sk_buff *skb;
1064         int cnt = 0;
1065         u32 new_low_seq = tp->snd_nxt;
1066         u32 received_upto = tcp_highest_sack_seq(tp);
1067 
1068         if (!tcp_is_fack(tp) || !tp->retrans_out ||
1069             !after(received_upto, tp->lost_retrans_low) ||
1070             icsk->icsk_ca_state != TCP_CA_Recovery)
1071                 return;
1072 
1073         tcp_for_write_queue(skb, sk) {
1074                 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1075 
1076                 if (skb == tcp_send_head(sk))
1077                         break;
1078                 if (cnt == tp->retrans_out)
1079                         break;
1080                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1081                         continue;
1082 
1083                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1084                         continue;
1085 
1086                 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1087                  * constraint here (see above) but figuring out that at
1088                  * least tp->reordering SACK blocks reside between ack_seq
1089                  * and received_upto is not easy task to do cheaply with
1090                  * the available datastructures.
1091                  *
1092                  * Whether FACK should check here for tp->reordering segs
1093                  * in-between one could argue for either way (it would be
1094                  * rather simple to implement as we could count fack_count
1095                  * during the walk and do tp->fackets_out - fack_count).
1096                  */
1097                 if (after(received_upto, ack_seq)) {
1098                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1099                         tp->retrans_out -= tcp_skb_pcount(skb);
1100                         *flag |= FLAG_LOST_RETRANS;
1101                         tcp_skb_mark_lost_uncond_verify(tp, skb);
1102                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1103                 } else {
1104                         if (before(ack_seq, new_low_seq))
1105                                 new_low_seq = ack_seq;
1106                         cnt += tcp_skb_pcount(skb);
1107                 }
1108         }
1109 
1110         if (tp->retrans_out)
1111                 tp->lost_retrans_low = new_low_seq;
1112 }
1113 
1114 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1115                             struct tcp_sack_block_wire *sp, int num_sacks,
1116                             u32 prior_snd_una)
1117 {
1118         struct tcp_sock *tp = tcp_sk(sk);
1119         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1120         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1121         bool dup_sack = false;
1122 
1123         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1124                 dup_sack = true;
1125                 tcp_dsack_seen(tp);
1126                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1127         } else if (num_sacks > 1) {
1128                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1129                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1130 
1131                 if (!after(end_seq_0, end_seq_1) &&
1132                     !before(start_seq_0, start_seq_1)) {
1133                         dup_sack = true;
1134                         tcp_dsack_seen(tp);
1135                         NET_INC_STATS_BH(sock_net(sk),
1136                                         LINUX_MIB_TCPDSACKOFORECV);
1137                 }
1138         }
1139 
1140         /* D-SACK for already forgotten data... Do dumb counting. */
1141         if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1142             !after(end_seq_0, prior_snd_una) &&
1143             after(end_seq_0, tp->undo_marker))
1144                 tp->undo_retrans--;
1145 
1146         return dup_sack;
1147 }
1148 
1149 struct tcp_sacktag_state {
1150         int     reord;
1151         int     fack_count;
1152         /* Timestamps for earliest and latest never-retransmitted segment
1153          * that was SACKed. RTO needs the earliest RTT to stay conservative,
1154          * but congestion control should still get an accurate delay signal.
1155          */
1156         struct skb_mstamp first_sackt;
1157         struct skb_mstamp last_sackt;
1158         int     flag;
1159 };
1160 
1161 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1162  * the incoming SACK may not exactly match but we can find smaller MSS
1163  * aligned portion of it that matches. Therefore we might need to fragment
1164  * which may fail and creates some hassle (caller must handle error case
1165  * returns).
1166  *
1167  * FIXME: this could be merged to shift decision code
1168  */
1169 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1170                                   u32 start_seq, u32 end_seq)
1171 {
1172         int err;
1173         bool in_sack;
1174         unsigned int pkt_len;
1175         unsigned int mss;
1176 
1177         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1178                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1179 
1180         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1181             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1182                 mss = tcp_skb_mss(skb);
1183                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1184 
1185                 if (!in_sack) {
1186                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1187                         if (pkt_len < mss)
1188                                 pkt_len = mss;
1189                 } else {
1190                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1191                         if (pkt_len < mss)
1192                                 return -EINVAL;
1193                 }
1194 
1195                 /* Round if necessary so that SACKs cover only full MSSes
1196                  * and/or the remaining small portion (if present)
1197                  */
1198                 if (pkt_len > mss) {
1199                         unsigned int new_len = (pkt_len / mss) * mss;
1200                         if (!in_sack && new_len < pkt_len) {
1201                                 new_len += mss;
1202                                 if (new_len >= skb->len)
1203                                         return 0;
1204                         }
1205                         pkt_len = new_len;
1206                 }
1207                 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1208                 if (err < 0)
1209                         return err;
1210         }
1211 
1212         return in_sack;
1213 }
1214 
1215 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1216 static u8 tcp_sacktag_one(struct sock *sk,
1217                           struct tcp_sacktag_state *state, u8 sacked,
1218                           u32 start_seq, u32 end_seq,
1219                           int dup_sack, int pcount,
1220                           const struct skb_mstamp *xmit_time)
1221 {
1222         struct tcp_sock *tp = tcp_sk(sk);
1223         int fack_count = state->fack_count;
1224 
1225         /* Account D-SACK for retransmitted packet. */
1226         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1227                 if (tp->undo_marker && tp->undo_retrans > 0 &&
1228                     after(end_seq, tp->undo_marker))
1229                         tp->undo_retrans--;
1230                 if (sacked & TCPCB_SACKED_ACKED)
1231                         state->reord = min(fack_count, state->reord);
1232         }
1233 
1234         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1235         if (!after(end_seq, tp->snd_una))
1236                 return sacked;
1237 
1238         if (!(sacked & TCPCB_SACKED_ACKED)) {
1239                 if (sacked & TCPCB_SACKED_RETRANS) {
1240                         /* If the segment is not tagged as lost,
1241                          * we do not clear RETRANS, believing
1242                          * that retransmission is still in flight.
1243                          */
1244                         if (sacked & TCPCB_LOST) {
1245                                 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1246                                 tp->lost_out -= pcount;
1247                                 tp->retrans_out -= pcount;
1248                         }
1249                 } else {
1250                         if (!(sacked & TCPCB_RETRANS)) {
1251                                 /* New sack for not retransmitted frame,
1252                                  * which was in hole. It is reordering.
1253                                  */
1254                                 if (before(start_seq,
1255                                            tcp_highest_sack_seq(tp)))
1256                                         state->reord = min(fack_count,
1257                                                            state->reord);
1258                                 if (!after(end_seq, tp->high_seq))
1259                                         state->flag |= FLAG_ORIG_SACK_ACKED;
1260                                 if (state->first_sackt.v64 == 0)
1261                                         state->first_sackt = *xmit_time;
1262                                 state->last_sackt = *xmit_time;
1263                         }
1264 
1265                         if (sacked & TCPCB_LOST) {
1266                                 sacked &= ~TCPCB_LOST;
1267                                 tp->lost_out -= pcount;
1268                         }
1269                 }
1270 
1271                 sacked |= TCPCB_SACKED_ACKED;
1272                 state->flag |= FLAG_DATA_SACKED;
1273                 tp->sacked_out += pcount;
1274 
1275                 fack_count += pcount;
1276 
1277                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1278                 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1279                     before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1280                         tp->lost_cnt_hint += pcount;
1281 
1282                 if (fack_count > tp->fackets_out)
1283                         tp->fackets_out = fack_count;
1284         }
1285 
1286         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1287          * frames and clear it. undo_retrans is decreased above, L|R frames
1288          * are accounted above as well.
1289          */
1290         if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1291                 sacked &= ~TCPCB_SACKED_RETRANS;
1292                 tp->retrans_out -= pcount;
1293         }
1294 
1295         return sacked;
1296 }
1297 
1298 /* Shift newly-SACKed bytes from this skb to the immediately previous
1299  * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1300  */
1301 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1302                             struct tcp_sacktag_state *state,
1303                             unsigned int pcount, int shifted, int mss,
1304                             bool dup_sack)
1305 {
1306         struct tcp_sock *tp = tcp_sk(sk);
1307         struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1308         u32 start_seq = TCP_SKB_CB(skb)->seq;   /* start of newly-SACKed */
1309         u32 end_seq = start_seq + shifted;      /* end of newly-SACKed */
1310 
1311         BUG_ON(!pcount);
1312 
1313         /* Adjust counters and hints for the newly sacked sequence
1314          * range but discard the return value since prev is already
1315          * marked. We must tag the range first because the seq
1316          * advancement below implicitly advances
1317          * tcp_highest_sack_seq() when skb is highest_sack.
1318          */
1319         tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1320                         start_seq, end_seq, dup_sack, pcount,
1321                         &skb->skb_mstamp);
1322 
1323         if (skb == tp->lost_skb_hint)
1324                 tp->lost_cnt_hint += pcount;
1325 
1326         TCP_SKB_CB(prev)->end_seq += shifted;
1327         TCP_SKB_CB(skb)->seq += shifted;
1328 
1329         tcp_skb_pcount_add(prev, pcount);
1330         BUG_ON(tcp_skb_pcount(skb) < pcount);
1331         tcp_skb_pcount_add(skb, -pcount);
1332 
1333         /* When we're adding to gso_segs == 1, gso_size will be zero,
1334          * in theory this shouldn't be necessary but as long as DSACK
1335          * code can come after this skb later on it's better to keep
1336          * setting gso_size to something.
1337          */
1338         if (!TCP_SKB_CB(prev)->tcp_gso_size)
1339                 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1340 
1341         /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1342         if (tcp_skb_pcount(skb) <= 1)
1343                 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1344 
1345         /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1346         TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1347 
1348         if (skb->len > 0) {
1349                 BUG_ON(!tcp_skb_pcount(skb));
1350                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1351                 return false;
1352         }
1353 
1354         /* Whole SKB was eaten :-) */
1355 
1356         if (skb == tp->retransmit_skb_hint)
1357                 tp->retransmit_skb_hint = prev;
1358         if (skb == tp->lost_skb_hint) {
1359                 tp->lost_skb_hint = prev;
1360                 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1361         }
1362 
1363         TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1364         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1365                 TCP_SKB_CB(prev)->end_seq++;
1366 
1367         if (skb == tcp_highest_sack(sk))
1368                 tcp_advance_highest_sack(sk, skb);
1369 
1370         tcp_unlink_write_queue(skb, sk);
1371         sk_wmem_free_skb(sk, skb);
1372 
1373         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1374 
1375         return true;
1376 }
1377 
1378 /* I wish gso_size would have a bit more sane initialization than
1379  * something-or-zero which complicates things
1380  */
1381 static int tcp_skb_seglen(const struct sk_buff *skb)
1382 {
1383         return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1384 }
1385 
1386 /* Shifting pages past head area doesn't work */
1387 static int skb_can_shift(const struct sk_buff *skb)
1388 {
1389         return !skb_headlen(skb) && skb_is_nonlinear(skb);
1390 }
1391 
1392 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1393  * skb.
1394  */
1395 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1396                                           struct tcp_sacktag_state *state,
1397                                           u32 start_seq, u32 end_seq,
1398                                           bool dup_sack)
1399 {
1400         struct tcp_sock *tp = tcp_sk(sk);
1401         struct sk_buff *prev;
1402         int mss;
1403         int pcount = 0;
1404         int len;
1405         int in_sack;
1406 
1407         if (!sk_can_gso(sk))
1408                 goto fallback;
1409 
1410         /* Normally R but no L won't result in plain S */
1411         if (!dup_sack &&
1412             (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1413                 goto fallback;
1414         if (!skb_can_shift(skb))
1415                 goto fallback;
1416         /* This frame is about to be dropped (was ACKed). */
1417         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1418                 goto fallback;
1419 
1420         /* Can only happen with delayed DSACK + discard craziness */
1421         if (unlikely(skb == tcp_write_queue_head(sk)))
1422                 goto fallback;
1423         prev = tcp_write_queue_prev(sk, skb);
1424 
1425         if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1426                 goto fallback;
1427 
1428         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1429                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1430 
1431         if (in_sack) {
1432                 len = skb->len;
1433                 pcount = tcp_skb_pcount(skb);
1434                 mss = tcp_skb_seglen(skb);
1435 
1436                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1437                  * drop this restriction as unnecessary
1438                  */
1439                 if (mss != tcp_skb_seglen(prev))
1440                         goto fallback;
1441         } else {
1442                 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1443                         goto noop;
1444                 /* CHECKME: This is non-MSS split case only?, this will
1445                  * cause skipped skbs due to advancing loop btw, original
1446                  * has that feature too
1447                  */
1448                 if (tcp_skb_pcount(skb) <= 1)
1449                         goto noop;
1450 
1451                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1452                 if (!in_sack) {
1453                         /* TODO: head merge to next could be attempted here
1454                          * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1455                          * though it might not be worth of the additional hassle
1456                          *
1457                          * ...we can probably just fallback to what was done
1458                          * previously. We could try merging non-SACKed ones
1459                          * as well but it probably isn't going to buy off
1460                          * because later SACKs might again split them, and
1461                          * it would make skb timestamp tracking considerably
1462                          * harder problem.
1463                          */
1464                         goto fallback;
1465                 }
1466 
1467                 len = end_seq - TCP_SKB_CB(skb)->seq;
1468                 BUG_ON(len < 0);
1469                 BUG_ON(len > skb->len);
1470 
1471                 /* MSS boundaries should be honoured or else pcount will
1472                  * severely break even though it makes things bit trickier.
1473                  * Optimize common case to avoid most of the divides
1474                  */
1475                 mss = tcp_skb_mss(skb);
1476 
1477                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1478                  * drop this restriction as unnecessary
1479                  */
1480                 if (mss != tcp_skb_seglen(prev))
1481                         goto fallback;
1482 
1483                 if (len == mss) {
1484                         pcount = 1;
1485                 } else if (len < mss) {
1486                         goto noop;
1487                 } else {
1488                         pcount = len / mss;
1489                         len = pcount * mss;
1490                 }
1491         }
1492 
1493         /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1494         if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1495                 goto fallback;
1496 
1497         if (!skb_shift(prev, skb, len))
1498                 goto fallback;
1499         if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1500                 goto out;
1501 
1502         /* Hole filled allows collapsing with the next as well, this is very
1503          * useful when hole on every nth skb pattern happens
1504          */
1505         if (prev == tcp_write_queue_tail(sk))
1506                 goto out;
1507         skb = tcp_write_queue_next(sk, prev);
1508 
1509         if (!skb_can_shift(skb) ||
1510             (skb == tcp_send_head(sk)) ||
1511             ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1512             (mss != tcp_skb_seglen(skb)))
1513                 goto out;
1514 
1515         len = skb->len;
1516         if (skb_shift(prev, skb, len)) {
1517                 pcount += tcp_skb_pcount(skb);
1518                 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1519         }
1520 
1521 out:
1522         state->fack_count += pcount;
1523         return prev;
1524 
1525 noop:
1526         return skb;
1527 
1528 fallback:
1529         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1530         return NULL;
1531 }
1532 
1533 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1534                                         struct tcp_sack_block *next_dup,
1535                                         struct tcp_sacktag_state *state,
1536                                         u32 start_seq, u32 end_seq,
1537                                         bool dup_sack_in)
1538 {
1539         struct tcp_sock *tp = tcp_sk(sk);
1540         struct sk_buff *tmp;
1541 
1542         tcp_for_write_queue_from(skb, sk) {
1543                 int in_sack = 0;
1544                 bool dup_sack = dup_sack_in;
1545 
1546                 if (skb == tcp_send_head(sk))
1547                         break;
1548 
1549                 /* queue is in-order => we can short-circuit the walk early */
1550                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1551                         break;
1552 
1553                 if (next_dup  &&
1554                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1555                         in_sack = tcp_match_skb_to_sack(sk, skb,
1556                                                         next_dup->start_seq,
1557                                                         next_dup->end_seq);
1558                         if (in_sack > 0)
1559                                 dup_sack = true;
1560                 }
1561 
1562                 /* skb reference here is a bit tricky to get right, since
1563                  * shifting can eat and free both this skb and the next,
1564                  * so not even _safe variant of the loop is enough.
1565                  */
1566                 if (in_sack <= 0) {
1567                         tmp = tcp_shift_skb_data(sk, skb, state,
1568                                                  start_seq, end_seq, dup_sack);
1569                         if (tmp) {
1570                                 if (tmp != skb) {
1571                                         skb = tmp;
1572                                         continue;
1573                                 }
1574 
1575                                 in_sack = 0;
1576                         } else {
1577                                 in_sack = tcp_match_skb_to_sack(sk, skb,
1578                                                                 start_seq,
1579                                                                 end_seq);
1580                         }
1581                 }
1582 
1583                 if (unlikely(in_sack < 0))
1584                         break;
1585 
1586                 if (in_sack) {
1587                         TCP_SKB_CB(skb)->sacked =
1588                                 tcp_sacktag_one(sk,
1589                                                 state,
1590                                                 TCP_SKB_CB(skb)->sacked,
1591                                                 TCP_SKB_CB(skb)->seq,
1592                                                 TCP_SKB_CB(skb)->end_seq,
1593                                                 dup_sack,
1594                                                 tcp_skb_pcount(skb),
1595                                                 &skb->skb_mstamp);
1596 
1597                         if (!before(TCP_SKB_CB(skb)->seq,
1598                                     tcp_highest_sack_seq(tp)))
1599                                 tcp_advance_highest_sack(sk, skb);
1600                 }
1601 
1602                 state->fack_count += tcp_skb_pcount(skb);
1603         }
1604         return skb;
1605 }
1606 
1607 /* Avoid all extra work that is being done by sacktag while walking in
1608  * a normal way
1609  */
1610 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1611                                         struct tcp_sacktag_state *state,
1612                                         u32 skip_to_seq)
1613 {
1614         tcp_for_write_queue_from(skb, sk) {
1615                 if (skb == tcp_send_head(sk))
1616                         break;
1617 
1618                 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1619                         break;
1620 
1621                 state->fack_count += tcp_skb_pcount(skb);
1622         }
1623         return skb;
1624 }
1625 
1626 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1627                                                 struct sock *sk,
1628                                                 struct tcp_sack_block *next_dup,
1629                                                 struct tcp_sacktag_state *state,
1630                                                 u32 skip_to_seq)
1631 {
1632         if (!next_dup)
1633                 return skb;
1634 
1635         if (before(next_dup->start_seq, skip_to_seq)) {
1636                 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1637                 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1638                                        next_dup->start_seq, next_dup->end_seq,
1639                                        1);
1640         }
1641 
1642         return skb;
1643 }
1644 
1645 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1646 {
1647         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1648 }
1649 
1650 static int
1651 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1652                         u32 prior_snd_una, struct tcp_sacktag_state *state)
1653 {
1654         struct tcp_sock *tp = tcp_sk(sk);
1655         const unsigned char *ptr = (skb_transport_header(ack_skb) +
1656                                     TCP_SKB_CB(ack_skb)->sacked);
1657         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1658         struct tcp_sack_block sp[TCP_NUM_SACKS];
1659         struct tcp_sack_block *cache;
1660         struct sk_buff *skb;
1661         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1662         int used_sacks;
1663         bool found_dup_sack = false;
1664         int i, j;
1665         int first_sack_index;
1666 
1667         state->flag = 0;
1668         state->reord = tp->packets_out;
1669 
1670         if (!tp->sacked_out) {
1671                 if (WARN_ON(tp->fackets_out))
1672                         tp->fackets_out = 0;
1673                 tcp_highest_sack_reset(sk);
1674         }
1675 
1676         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1677                                          num_sacks, prior_snd_una);
1678         if (found_dup_sack)
1679                 state->flag |= FLAG_DSACKING_ACK;
1680 
1681         /* Eliminate too old ACKs, but take into
1682          * account more or less fresh ones, they can
1683          * contain valid SACK info.
1684          */
1685         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1686                 return 0;
1687 
1688         if (!tp->packets_out)
1689                 goto out;
1690 
1691         used_sacks = 0;
1692         first_sack_index = 0;
1693         for (i = 0; i < num_sacks; i++) {
1694                 bool dup_sack = !i && found_dup_sack;
1695 
1696                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1697                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1698 
1699                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1700                                             sp[used_sacks].start_seq,
1701                                             sp[used_sacks].end_seq)) {
1702                         int mib_idx;
1703 
1704                         if (dup_sack) {
1705                                 if (!tp->undo_marker)
1706                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1707                                 else
1708                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1709                         } else {
1710                                 /* Don't count olds caused by ACK reordering */
1711                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1712                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1713                                         continue;
1714                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1715                         }
1716 
1717                         NET_INC_STATS_BH(sock_net(sk), mib_idx);
1718                         if (i == 0)
1719                                 first_sack_index = -1;
1720                         continue;
1721                 }
1722 
1723                 /* Ignore very old stuff early */
1724                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1725                         continue;
1726 
1727                 used_sacks++;
1728         }
1729 
1730         /* order SACK blocks to allow in order walk of the retrans queue */
1731         for (i = used_sacks - 1; i > 0; i--) {
1732                 for (j = 0; j < i; j++) {
1733                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1734                                 swap(sp[j], sp[j + 1]);
1735 
1736                                 /* Track where the first SACK block goes to */
1737                                 if (j == first_sack_index)
1738                                         first_sack_index = j + 1;
1739                         }
1740                 }
1741         }
1742 
1743         skb = tcp_write_queue_head(sk);
1744         state->fack_count = 0;
1745         i = 0;
1746 
1747         if (!tp->sacked_out) {
1748                 /* It's already past, so skip checking against it */
1749                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1750         } else {
1751                 cache = tp->recv_sack_cache;
1752                 /* Skip empty blocks in at head of the cache */
1753                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1754                        !cache->end_seq)
1755                         cache++;
1756         }
1757 
1758         while (i < used_sacks) {
1759                 u32 start_seq = sp[i].start_seq;
1760                 u32 end_seq = sp[i].end_seq;
1761                 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1762                 struct tcp_sack_block *next_dup = NULL;
1763 
1764                 if (found_dup_sack && ((i + 1) == first_sack_index))
1765                         next_dup = &sp[i + 1];
1766 
1767                 /* Skip too early cached blocks */
1768                 while (tcp_sack_cache_ok(tp, cache) &&
1769                        !before(start_seq, cache->end_seq))
1770                         cache++;
1771 
1772                 /* Can skip some work by looking recv_sack_cache? */
1773                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1774                     after(end_seq, cache->start_seq)) {
1775 
1776                         /* Head todo? */
1777                         if (before(start_seq, cache->start_seq)) {
1778                                 skb = tcp_sacktag_skip(skb, sk, state,
1779                                                        start_seq);
1780                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1781                                                        state,
1782                                                        start_seq,
1783                                                        cache->start_seq,
1784                                                        dup_sack);
1785                         }
1786 
1787                         /* Rest of the block already fully processed? */
1788                         if (!after(end_seq, cache->end_seq))
1789                                 goto advance_sp;
1790 
1791                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1792                                                        state,
1793                                                        cache->end_seq);
1794 
1795                         /* ...tail remains todo... */
1796                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1797                                 /* ...but better entrypoint exists! */
1798                                 skb = tcp_highest_sack(sk);
1799                                 if (!skb)
1800                                         break;
1801                                 state->fack_count = tp->fackets_out;
1802                                 cache++;
1803                                 goto walk;
1804                         }
1805 
1806                         skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1807                         /* Check overlap against next cached too (past this one already) */
1808                         cache++;
1809                         continue;
1810                 }
1811 
1812                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1813                         skb = tcp_highest_sack(sk);
1814                         if (!skb)
1815                                 break;
1816                         state->fack_count = tp->fackets_out;
1817                 }
1818                 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1819 
1820 walk:
1821                 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1822                                        start_seq, end_seq, dup_sack);
1823 
1824 advance_sp:
1825                 i++;
1826         }
1827 
1828         /* Clear the head of the cache sack blocks so we can skip it next time */
1829         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1830                 tp->recv_sack_cache[i].start_seq = 0;
1831                 tp->recv_sack_cache[i].end_seq = 0;
1832         }
1833         for (j = 0; j < used_sacks; j++)
1834                 tp->recv_sack_cache[i++] = sp[j];
1835 
1836         if ((state->reord < tp->fackets_out) &&
1837             ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1838                 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
1839 
1840         tcp_mark_lost_retrans(sk, &state->flag);
1841         tcp_verify_left_out(tp);
1842 out:
1843 
1844 #if FASTRETRANS_DEBUG > 0
1845         WARN_ON((int)tp->sacked_out < 0);
1846         WARN_ON((int)tp->lost_out < 0);
1847         WARN_ON((int)tp->retrans_out < 0);
1848         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1849 #endif
1850         return state->flag;
1851 }
1852 
1853 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1854  * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1855  */
1856 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1857 {
1858         u32 holes;
1859 
1860         holes = max(tp->lost_out, 1U);
1861         holes = min(holes, tp->packets_out);
1862 
1863         if ((tp->sacked_out + holes) > tp->packets_out) {
1864                 tp->sacked_out = tp->packets_out - holes;
1865                 return true;
1866         }
1867         return false;
1868 }
1869 
1870 /* If we receive more dupacks than we expected counting segments
1871  * in assumption of absent reordering, interpret this as reordering.
1872  * The only another reason could be bug in receiver TCP.
1873  */
1874 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1875 {
1876         struct tcp_sock *tp = tcp_sk(sk);
1877         if (tcp_limit_reno_sacked(tp))
1878                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1879 }
1880 
1881 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1882 
1883 static void tcp_add_reno_sack(struct sock *sk)
1884 {
1885         struct tcp_sock *tp = tcp_sk(sk);
1886         tp->sacked_out++;
1887         tcp_check_reno_reordering(sk, 0);
1888         tcp_verify_left_out(tp);
1889 }
1890 
1891 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1892 
1893 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1894 {
1895         struct tcp_sock *tp = tcp_sk(sk);
1896 
1897         if (acked > 0) {
1898                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1899                 if (acked - 1 >= tp->sacked_out)
1900                         tp->sacked_out = 0;
1901                 else
1902                         tp->sacked_out -= acked - 1;
1903         }
1904         tcp_check_reno_reordering(sk, acked);
1905         tcp_verify_left_out(tp);
1906 }
1907 
1908 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1909 {
1910         tp->sacked_out = 0;
1911 }
1912 
1913 void tcp_clear_retrans(struct tcp_sock *tp)
1914 {
1915         tp->retrans_out = 0;
1916         tp->lost_out = 0;
1917         tp->undo_marker = 0;
1918         tp->undo_retrans = -1;
1919         tp->fackets_out = 0;
1920         tp->sacked_out = 0;
1921 }
1922 
1923 static inline void tcp_init_undo(struct tcp_sock *tp)
1924 {
1925         tp->undo_marker = tp->snd_una;
1926         /* Retransmission still in flight may cause DSACKs later. */
1927         tp->undo_retrans = tp->retrans_out ? : -1;
1928 }
1929 
1930 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1931  * and reset tags completely, otherwise preserve SACKs. If receiver
1932  * dropped its ofo queue, we will know this due to reneging detection.
1933  */
1934 void tcp_enter_loss(struct sock *sk)
1935 {
1936         const struct inet_connection_sock *icsk = inet_csk(sk);
1937         struct tcp_sock *tp = tcp_sk(sk);
1938         struct sk_buff *skb;
1939         bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1940         bool is_reneg;                  /* is receiver reneging on SACKs? */
1941 
1942         /* Reduce ssthresh if it has not yet been made inside this window. */
1943         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1944             !after(tp->high_seq, tp->snd_una) ||
1945             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1946                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1947                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1948                 tcp_ca_event(sk, CA_EVENT_LOSS);
1949                 tcp_init_undo(tp);
1950         }
1951         tp->snd_cwnd       = 1;
1952         tp->snd_cwnd_cnt   = 0;
1953         tp->snd_cwnd_stamp = tcp_time_stamp;
1954 
1955         tp->retrans_out = 0;
1956         tp->lost_out = 0;
1957 
1958         if (tcp_is_reno(tp))
1959                 tcp_reset_reno_sack(tp);
1960 
1961         skb = tcp_write_queue_head(sk);
1962         is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1963         if (is_reneg) {
1964                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1965                 tp->sacked_out = 0;
1966                 tp->fackets_out = 0;
1967         }
1968         tcp_clear_all_retrans_hints(tp);
1969 
1970         tcp_for_write_queue(skb, sk) {
1971                 if (skb == tcp_send_head(sk))
1972                         break;
1973 
1974                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1975                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
1976                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1977                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1978                         tp->lost_out += tcp_skb_pcount(skb);
1979                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1980                 }
1981         }
1982         tcp_verify_left_out(tp);
1983 
1984         /* Timeout in disordered state after receiving substantial DUPACKs
1985          * suggests that the degree of reordering is over-estimated.
1986          */
1987         if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1988             tp->sacked_out >= sysctl_tcp_reordering)
1989                 tp->reordering = min_t(unsigned int, tp->reordering,
1990                                        sysctl_tcp_reordering);
1991         tcp_set_ca_state(sk, TCP_CA_Loss);
1992         tp->high_seq = tp->snd_nxt;
1993         tcp_ecn_queue_cwr(tp);
1994 
1995         /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1996          * loss recovery is underway except recurring timeout(s) on
1997          * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1998          */
1999         tp->frto = sysctl_tcp_frto &&
2000                    (new_recovery || icsk->icsk_retransmits) &&
2001                    !inet_csk(sk)->icsk_mtup.probe_size;
2002 }
2003 
2004 /* If ACK arrived pointing to a remembered SACK, it means that our
2005  * remembered SACKs do not reflect real state of receiver i.e.
2006  * receiver _host_ is heavily congested (or buggy).
2007  *
2008  * To avoid big spurious retransmission bursts due to transient SACK
2009  * scoreboard oddities that look like reneging, we give the receiver a
2010  * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2011  * restore sanity to the SACK scoreboard. If the apparent reneging
2012  * persists until this RTO then we'll clear the SACK scoreboard.
2013  */
2014 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2015 {
2016         if (flag & FLAG_SACK_RENEGING) {
2017                 struct tcp_sock *tp = tcp_sk(sk);
2018                 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2019                                           msecs_to_jiffies(10));
2020 
2021                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2022                                           delay, TCP_RTO_MAX);
2023                 return true;
2024         }
2025         return false;
2026 }
2027 
2028 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2029 {
2030         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2031 }
2032 
2033 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2034  * counter when SACK is enabled (without SACK, sacked_out is used for
2035  * that purpose).
2036  *
2037  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2038  * segments up to the highest received SACK block so far and holes in
2039  * between them.
2040  *
2041  * With reordering, holes may still be in flight, so RFC3517 recovery
2042  * uses pure sacked_out (total number of SACKed segments) even though
2043  * it violates the RFC that uses duplicate ACKs, often these are equal
2044  * but when e.g. out-of-window ACKs or packet duplication occurs,
2045  * they differ. Since neither occurs due to loss, TCP should really
2046  * ignore them.
2047  */
2048 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2049 {
2050         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2051 }
2052 
2053 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2054 {
2055         struct tcp_sock *tp = tcp_sk(sk);
2056         unsigned long delay;
2057 
2058         /* Delay early retransmit and entering fast recovery for
2059          * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2060          * available, or RTO is scheduled to fire first.
2061          */
2062         if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2063             (flag & FLAG_ECE) || !tp->srtt_us)
2064                 return false;
2065 
2066         delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2067                     msecs_to_jiffies(2));
2068 
2069         if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2070                 return false;
2071 
2072         inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2073                                   TCP_RTO_MAX);
2074         return true;
2075 }
2076 
2077 /* Linux NewReno/SACK/FACK/ECN state machine.
2078  * --------------------------------------
2079  *
2080  * "Open"       Normal state, no dubious events, fast path.
2081  * "Disorder"   In all the respects it is "Open",
2082  *              but requires a bit more attention. It is entered when
2083  *              we see some SACKs or dupacks. It is split of "Open"
2084  *              mainly to move some processing from fast path to slow one.
2085  * "CWR"        CWND was reduced due to some Congestion Notification event.
2086  *              It can be ECN, ICMP source quench, local device congestion.
2087  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2088  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2089  *
2090  * tcp_fastretrans_alert() is entered:
2091  * - each incoming ACK, if state is not "Open"
2092  * - when arrived ACK is unusual, namely:
2093  *      * SACK
2094  *      * Duplicate ACK.
2095  *      * ECN ECE.
2096  *
2097  * Counting packets in flight is pretty simple.
2098  *
2099  *      in_flight = packets_out - left_out + retrans_out
2100  *
2101  *      packets_out is SND.NXT-SND.UNA counted in packets.
2102  *
2103  *      retrans_out is number of retransmitted segments.
2104  *
2105  *      left_out is number of segments left network, but not ACKed yet.
2106  *
2107  *              left_out = sacked_out + lost_out
2108  *
2109  *     sacked_out: Packets, which arrived to receiver out of order
2110  *                 and hence not ACKed. With SACKs this number is simply
2111  *                 amount of SACKed data. Even without SACKs
2112  *                 it is easy to give pretty reliable estimate of this number,
2113  *                 counting duplicate ACKs.
2114  *
2115  *       lost_out: Packets lost by network. TCP has no explicit
2116  *                 "loss notification" feedback from network (for now).
2117  *                 It means that this number can be only _guessed_.
2118  *                 Actually, it is the heuristics to predict lossage that
2119  *                 distinguishes different algorithms.
2120  *
2121  *      F.e. after RTO, when all the queue is considered as lost,
2122  *      lost_out = packets_out and in_flight = retrans_out.
2123  *
2124  *              Essentially, we have now two algorithms counting
2125  *              lost packets.
2126  *
2127  *              FACK: It is the simplest heuristics. As soon as we decided
2128  *              that something is lost, we decide that _all_ not SACKed
2129  *              packets until the most forward SACK are lost. I.e.
2130  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2131  *              It is absolutely correct estimate, if network does not reorder
2132  *              packets. And it loses any connection to reality when reordering
2133  *              takes place. We use FACK by default until reordering
2134  *              is suspected on the path to this destination.
2135  *
2136  *              NewReno: when Recovery is entered, we assume that one segment
2137  *              is lost (classic Reno). While we are in Recovery and
2138  *              a partial ACK arrives, we assume that one more packet
2139  *              is lost (NewReno). This heuristics are the same in NewReno
2140  *              and SACK.
2141  *
2142  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2143  *  deflation etc. CWND is real congestion window, never inflated, changes
2144  *  only according to classic VJ rules.
2145  *
2146  * Really tricky (and requiring careful tuning) part of algorithm
2147  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2148  * The first determines the moment _when_ we should reduce CWND and,
2149  * hence, slow down forward transmission. In fact, it determines the moment
2150  * when we decide that hole is caused by loss, rather than by a reorder.
2151  *
2152  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2153  * holes, caused by lost packets.
2154  *
2155  * And the most logically complicated part of algorithm is undo
2156  * heuristics. We detect false retransmits due to both too early
2157  * fast retransmit (reordering) and underestimated RTO, analyzing
2158  * timestamps and D-SACKs. When we detect that some segments were
2159  * retransmitted by mistake and CWND reduction was wrong, we undo
2160  * window reduction and abort recovery phase. This logic is hidden
2161  * inside several functions named tcp_try_undo_<something>.
2162  */
2163 
2164 /* This function decides, when we should leave Disordered state
2165  * and enter Recovery phase, reducing congestion window.
2166  *
2167  * Main question: may we further continue forward transmission
2168  * with the same cwnd?
2169  */
2170 static bool tcp_time_to_recover(struct sock *sk, int flag)
2171 {
2172         struct tcp_sock *tp = tcp_sk(sk);
2173         __u32 packets_out;
2174 
2175         /* Trick#1: The loss is proven. */
2176         if (tp->lost_out)
2177                 return true;
2178 
2179         /* Not-A-Trick#2 : Classic rule... */
2180         if (tcp_dupack_heuristics(tp) > tp->reordering)
2181                 return true;
2182 
2183         /* Trick#4: It is still not OK... But will it be useful to delay
2184          * recovery more?
2185          */
2186         packets_out = tp->packets_out;
2187         if (packets_out <= tp->reordering &&
2188             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2189             !tcp_may_send_now(sk)) {
2190                 /* We have nothing to send. This connection is limited
2191                  * either by receiver window or by application.
2192                  */
2193                 return true;
2194         }
2195 
2196         /* If a thin stream is detected, retransmit after first
2197          * received dupack. Employ only if SACK is supported in order
2198          * to avoid possible corner-case series of spurious retransmissions
2199          * Use only if there are no unsent data.
2200          */
2201         if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2202             tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2203             tcp_is_sack(tp) && !tcp_send_head(sk))
2204                 return true;
2205 
2206         /* Trick#6: TCP early retransmit, per RFC5827.  To avoid spurious
2207          * retransmissions due to small network reorderings, we implement
2208          * Mitigation A.3 in the RFC and delay the retransmission for a short
2209          * interval if appropriate.
2210          */
2211         if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2212             (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2213             !tcp_may_send_now(sk))
2214                 return !tcp_pause_early_retransmit(sk, flag);
2215 
2216         return false;
2217 }
2218 
2219 /* Detect loss in event "A" above by marking head of queue up as lost.
2220  * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2221  * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2222  * has at least tp->reordering SACKed seqments above it; "packets" refers to
2223  * the maximum SACKed segments to pass before reaching this limit.
2224  */
2225 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2226 {
2227         struct tcp_sock *tp = tcp_sk(sk);
2228         struct sk_buff *skb;
2229         int cnt, oldcnt;
2230         int err;
2231         unsigned int mss;
2232         /* Use SACK to deduce losses of new sequences sent during recovery */
2233         const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2234 
2235         WARN_ON(packets > tp->packets_out);
2236         if (tp->lost_skb_hint) {
2237                 skb = tp->lost_skb_hint;
2238                 cnt = tp->lost_cnt_hint;
2239                 /* Head already handled? */
2240                 if (mark_head && skb != tcp_write_queue_head(sk))
2241                         return;
2242         } else {
2243                 skb = tcp_write_queue_head(sk);
2244                 cnt = 0;
2245         }
2246 
2247         tcp_for_write_queue_from(skb, sk) {
2248                 if (skb == tcp_send_head(sk))
2249                         break;
2250                 /* TODO: do this better */
2251                 /* this is not the most efficient way to do this... */
2252                 tp->lost_skb_hint = skb;
2253                 tp->lost_cnt_hint = cnt;
2254 
2255                 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2256                         break;
2257 
2258                 oldcnt = cnt;
2259                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2260                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2261                         cnt += tcp_skb_pcount(skb);
2262 
2263                 if (cnt > packets) {
2264                         if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2265                             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2266                             (oldcnt >= packets))
2267                                 break;
2268 
2269                         mss = tcp_skb_mss(skb);
2270                         err = tcp_fragment(sk, skb, (packets - oldcnt) * mss,
2271                                            mss, GFP_ATOMIC);
2272                         if (err < 0)
2273                                 break;
2274                         cnt = packets;
2275                 }
2276 
2277                 tcp_skb_mark_lost(tp, skb);
2278 
2279                 if (mark_head)
2280                         break;
2281         }
2282         tcp_verify_left_out(tp);
2283 }
2284 
2285 /* Account newly detected lost packet(s) */
2286 
2287 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2288 {
2289         struct tcp_sock *tp = tcp_sk(sk);
2290 
2291         if (tcp_is_reno(tp)) {
2292                 tcp_mark_head_lost(sk, 1, 1);
2293         } else if (tcp_is_fack(tp)) {
2294                 int lost = tp->fackets_out - tp->reordering;
2295                 if (lost <= 0)
2296                         lost = 1;
2297                 tcp_mark_head_lost(sk, lost, 0);
2298         } else {
2299                 int sacked_upto = tp->sacked_out - tp->reordering;
2300                 if (sacked_upto >= 0)
2301                         tcp_mark_head_lost(sk, sacked_upto, 0);
2302                 else if (fast_rexmit)
2303                         tcp_mark_head_lost(sk, 1, 1);
2304         }
2305 }
2306 
2307 /* CWND moderation, preventing bursts due to too big ACKs
2308  * in dubious situations.
2309  */
2310 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2311 {
2312         tp->snd_cwnd = min(tp->snd_cwnd,
2313                            tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2314         tp->snd_cwnd_stamp = tcp_time_stamp;
2315 }
2316 
2317 /* Nothing was retransmitted or returned timestamp is less
2318  * than timestamp of the first retransmission.
2319  */
2320 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2321 {
2322         return !tp->retrans_stamp ||
2323                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2324                  before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2325 }
2326 
2327 /* Undo procedures. */
2328 
2329 /* We can clear retrans_stamp when there are no retransmissions in the
2330  * window. It would seem that it is trivially available for us in
2331  * tp->retrans_out, however, that kind of assumptions doesn't consider
2332  * what will happen if errors occur when sending retransmission for the
2333  * second time. ...It could the that such segment has only
2334  * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2335  * the head skb is enough except for some reneging corner cases that
2336  * are not worth the effort.
2337  *
2338  * Main reason for all this complexity is the fact that connection dying
2339  * time now depends on the validity of the retrans_stamp, in particular,
2340  * that successive retransmissions of a segment must not advance
2341  * retrans_stamp under any conditions.
2342  */
2343 static bool tcp_any_retrans_done(const struct sock *sk)
2344 {
2345         const struct tcp_sock *tp = tcp_sk(sk);
2346         struct sk_buff *skb;
2347 
2348         if (tp->retrans_out)
2349                 return true;
2350 
2351         skb = tcp_write_queue_head(sk);
2352         if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2353                 return true;
2354 
2355         return false;
2356 }
2357 
2358 #if FASTRETRANS_DEBUG > 1
2359 static void DBGUNDO(struct sock *sk, const char *msg)
2360 {
2361         struct tcp_sock *tp = tcp_sk(sk);
2362         struct inet_sock *inet = inet_sk(sk);
2363 
2364         if (sk->sk_family == AF_INET) {
2365                 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2366                          msg,
2367                          &inet->inet_daddr, ntohs(inet->inet_dport),
2368                          tp->snd_cwnd, tcp_left_out(tp),
2369                          tp->snd_ssthresh, tp->prior_ssthresh,
2370                          tp->packets_out);
2371         }
2372 #if IS_ENABLED(CONFIG_IPV6)
2373         else if (sk->sk_family == AF_INET6) {
2374                 struct ipv6_pinfo *np = inet6_sk(sk);
2375                 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2376                          msg,
2377                          &np->daddr, ntohs(inet->inet_dport),
2378                          tp->snd_cwnd, tcp_left_out(tp),
2379                          tp->snd_ssthresh, tp->prior_ssthresh,
2380                          tp->packets_out);
2381         }
2382 #endif
2383 }
2384 #else
2385 #define DBGUNDO(x...) do { } while (0)
2386 #endif
2387 
2388 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2389 {
2390         struct tcp_sock *tp = tcp_sk(sk);
2391 
2392         if (unmark_loss) {
2393                 struct sk_buff *skb;
2394 
2395                 tcp_for_write_queue(skb, sk) {
2396                         if (skb == tcp_send_head(sk))
2397                                 break;
2398                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2399                 }
2400                 tp->lost_out = 0;
2401                 tcp_clear_all_retrans_hints(tp);
2402         }
2403 
2404         if (tp->prior_ssthresh) {
2405                 const struct inet_connection_sock *icsk = inet_csk(sk);
2406 
2407                 if (icsk->icsk_ca_ops->undo_cwnd)
2408                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2409                 else
2410                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2411 
2412                 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2413                         tp->snd_ssthresh = tp->prior_ssthresh;
2414                         tcp_ecn_withdraw_cwr(tp);
2415                 }
2416         } else {
2417                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2418         }
2419         tp->snd_cwnd_stamp = tcp_time_stamp;
2420         tp->undo_marker = 0;
2421 }
2422 
2423 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2424 {
2425         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2426 }
2427 
2428 /* People celebrate: "We love our President!" */
2429 static bool tcp_try_undo_recovery(struct sock *sk)
2430 {
2431         struct tcp_sock *tp = tcp_sk(sk);
2432 
2433         if (tcp_may_undo(tp)) {
2434                 int mib_idx;
2435 
2436                 /* Happy end! We did not retransmit anything
2437                  * or our original transmission succeeded.
2438                  */
2439                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2440                 tcp_undo_cwnd_reduction(sk, false);
2441                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2442                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2443                 else
2444                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2445 
2446                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2447         }
2448         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2449                 /* Hold old state until something *above* high_seq
2450                  * is ACKed. For Reno it is MUST to prevent false
2451                  * fast retransmits (RFC2582). SACK TCP is safe. */
2452                 tcp_moderate_cwnd(tp);
2453                 if (!tcp_any_retrans_done(sk))
2454                         tp->retrans_stamp = 0;
2455                 return true;
2456         }
2457         tcp_set_ca_state(sk, TCP_CA_Open);
2458         return false;
2459 }
2460 
2461 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2462 static bool tcp_try_undo_dsack(struct sock *sk)
2463 {
2464         struct tcp_sock *tp = tcp_sk(sk);
2465 
2466         if (tp->undo_marker && !tp->undo_retrans) {
2467                 DBGUNDO(sk, "D-SACK");
2468                 tcp_undo_cwnd_reduction(sk, false);
2469                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2470                 return true;
2471         }
2472         return false;
2473 }
2474 
2475 /* Undo during loss recovery after partial ACK or using F-RTO. */
2476 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2477 {
2478         struct tcp_sock *tp = tcp_sk(sk);
2479 
2480         if (frto_undo || tcp_may_undo(tp)) {
2481                 tcp_undo_cwnd_reduction(sk, true);
2482 
2483                 DBGUNDO(sk, "partial loss");
2484                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2485                 if (frto_undo)
2486                         NET_INC_STATS_BH(sock_net(sk),
2487                                          LINUX_MIB_TCPSPURIOUSRTOS);
2488                 inet_csk(sk)->icsk_retransmits = 0;
2489                 if (frto_undo || tcp_is_sack(tp))
2490                         tcp_set_ca_state(sk, TCP_CA_Open);
2491                 return true;
2492         }
2493         return false;
2494 }
2495 
2496 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2497  * It computes the number of packets to send (sndcnt) based on packets newly
2498  * delivered:
2499  *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2500  *      cwnd reductions across a full RTT.
2501  *   2) Otherwise PRR uses packet conservation to send as much as delivered.
2502  *      But when the retransmits are acked without further losses, PRR
2503  *      slow starts cwnd up to ssthresh to speed up the recovery.
2504  */
2505 static void tcp_init_cwnd_reduction(struct sock *sk)
2506 {
2507         struct tcp_sock *tp = tcp_sk(sk);
2508 
2509         tp->high_seq = tp->snd_nxt;
2510         tp->tlp_high_seq = 0;
2511         tp->snd_cwnd_cnt = 0;
2512         tp->prior_cwnd = tp->snd_cwnd;
2513         tp->prr_delivered = 0;
2514         tp->prr_out = 0;
2515         tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2516         tcp_ecn_queue_cwr(tp);
2517 }
2518 
2519 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2520                                int fast_rexmit, int flag)
2521 {
2522         struct tcp_sock *tp = tcp_sk(sk);
2523         int sndcnt = 0;
2524         int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2525         int newly_acked_sacked = prior_unsacked -
2526                                  (tp->packets_out - tp->sacked_out);
2527 
2528         if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2529                 return;
2530 
2531         tp->prr_delivered += newly_acked_sacked;
2532         if (delta < 0) {
2533                 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2534                                tp->prior_cwnd - 1;
2535                 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2536         } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2537                    !(flag & FLAG_LOST_RETRANS)) {
2538                 sndcnt = min_t(int, delta,
2539                                max_t(int, tp->prr_delivered - tp->prr_out,
2540                                      newly_acked_sacked) + 1);
2541         } else {
2542                 sndcnt = min(delta, newly_acked_sacked);
2543         }
2544         sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2545         tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2546 }
2547 
2548 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2549 {
2550         struct tcp_sock *tp = tcp_sk(sk);
2551 
2552         /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2553         if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2554             (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2555                 tp->snd_cwnd = tp->snd_ssthresh;
2556                 tp->snd_cwnd_stamp = tcp_time_stamp;
2557         }
2558         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2559 }
2560 
2561 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2562 void tcp_enter_cwr(struct sock *sk)
2563 {
2564         struct tcp_sock *tp = tcp_sk(sk);
2565 
2566         tp->prior_ssthresh = 0;
2567         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2568                 tp->undo_marker = 0;
2569                 tcp_init_cwnd_reduction(sk);
2570                 tcp_set_ca_state(sk, TCP_CA_CWR);
2571         }
2572 }
2573 EXPORT_SYMBOL(tcp_enter_cwr);
2574 
2575 static void tcp_try_keep_open(struct sock *sk)
2576 {
2577         struct tcp_sock *tp = tcp_sk(sk);
2578         int state = TCP_CA_Open;
2579 
2580         if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2581                 state = TCP_CA_Disorder;
2582 
2583         if (inet_csk(sk)->icsk_ca_state != state) {
2584                 tcp_set_ca_state(sk, state);
2585                 tp->high_seq = tp->snd_nxt;
2586         }
2587 }
2588 
2589 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2590 {
2591         struct tcp_sock *tp = tcp_sk(sk);
2592 
2593         tcp_verify_left_out(tp);
2594 
2595         if (!tcp_any_retrans_done(sk))
2596                 tp->retrans_stamp = 0;
2597 
2598         if (flag & FLAG_ECE)
2599                 tcp_enter_cwr(sk);
2600 
2601         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2602                 tcp_try_keep_open(sk);
2603         } else {
2604                 tcp_cwnd_reduction(sk, prior_unsacked, 0, flag);
2605         }
2606 }
2607 
2608 static void tcp_mtup_probe_failed(struct sock *sk)
2609 {
2610         struct inet_connection_sock *icsk = inet_csk(sk);
2611 
2612         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2613         icsk->icsk_mtup.probe_size = 0;
2614         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2615 }
2616 
2617 static void tcp_mtup_probe_success(struct sock *sk)
2618 {
2619         struct tcp_sock *tp = tcp_sk(sk);
2620         struct inet_connection_sock *icsk = inet_csk(sk);
2621 
2622         /* FIXME: breaks with very large cwnd */
2623         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2624         tp->snd_cwnd = tp->snd_cwnd *
2625                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2626                        icsk->icsk_mtup.probe_size;
2627         tp->snd_cwnd_cnt = 0;
2628         tp->snd_cwnd_stamp = tcp_time_stamp;
2629         tp->snd_ssthresh = tcp_current_ssthresh(sk);
2630 
2631         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2632         icsk->icsk_mtup.probe_size = 0;
2633         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2634         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2635 }
2636 
2637 /* Do a simple retransmit without using the backoff mechanisms in
2638  * tcp_timer. This is used for path mtu discovery.
2639  * The socket is already locked here.
2640  */
2641 void tcp_simple_retransmit(struct sock *sk)
2642 {
2643         const struct inet_connection_sock *icsk = inet_csk(sk);
2644         struct tcp_sock *tp = tcp_sk(sk);
2645         struct sk_buff *skb;
2646         unsigned int mss = tcp_current_mss(sk);
2647         u32 prior_lost = tp->lost_out;
2648 
2649         tcp_for_write_queue(skb, sk) {
2650                 if (skb == tcp_send_head(sk))
2651                         break;
2652                 if (tcp_skb_seglen(skb) > mss &&
2653                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2654                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2655                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2656                                 tp->retrans_out -= tcp_skb_pcount(skb);
2657                         }
2658                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2659                 }
2660         }
2661 
2662         tcp_clear_retrans_hints_partial(tp);
2663 
2664         if (prior_lost == tp->lost_out)
2665                 return;
2666 
2667         if (tcp_is_reno(tp))
2668                 tcp_limit_reno_sacked(tp);
2669 
2670         tcp_verify_left_out(tp);
2671 
2672         /* Don't muck with the congestion window here.
2673          * Reason is that we do not increase amount of _data_
2674          * in network, but units changed and effective
2675          * cwnd/ssthresh really reduced now.
2676          */
2677         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2678                 tp->high_seq = tp->snd_nxt;
2679                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2680                 tp->prior_ssthresh = 0;
2681                 tp->undo_marker = 0;
2682                 tcp_set_ca_state(sk, TCP_CA_Loss);
2683         }
2684         tcp_xmit_retransmit_queue(sk);
2685 }
2686 EXPORT_SYMBOL(tcp_simple_retransmit);
2687 
2688 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2689 {
2690         struct tcp_sock *tp = tcp_sk(sk);
2691         int mib_idx;
2692 
2693         if (tcp_is_reno(tp))
2694                 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2695         else
2696                 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2697 
2698         NET_INC_STATS_BH(sock_net(sk), mib_idx);
2699 
2700         tp->prior_ssthresh = 0;
2701         tcp_init_undo(tp);
2702 
2703         if (!tcp_in_cwnd_reduction(sk)) {
2704                 if (!ece_ack)
2705                         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2706                 tcp_init_cwnd_reduction(sk);
2707         }
2708         tcp_set_ca_state(sk, TCP_CA_Recovery);
2709 }
2710 
2711 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2712  * recovered or spurious. Otherwise retransmits more on partial ACKs.
2713  */
2714 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2715 {
2716         struct tcp_sock *tp = tcp_sk(sk);
2717         bool recovered = !before(tp->snd_una, tp->high_seq);
2718 
2719         if ((flag & FLAG_SND_UNA_ADVANCED) &&
2720             tcp_try_undo_loss(sk, false))
2721                 return;
2722 
2723         if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2724                 /* Step 3.b. A timeout is spurious if not all data are
2725                  * lost, i.e., never-retransmitted data are (s)acked.
2726                  */
2727                 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2728                     tcp_try_undo_loss(sk, true))
2729                         return;
2730 
2731                 if (after(tp->snd_nxt, tp->high_seq)) {
2732                         if (flag & FLAG_DATA_SACKED || is_dupack)
2733                                 tp->frto = 0; /* Step 3.a. loss was real */
2734                 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2735                         tp->high_seq = tp->snd_nxt;
2736                         __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2737                                                   TCP_NAGLE_OFF);
2738                         if (after(tp->snd_nxt, tp->high_seq))
2739                                 return; /* Step 2.b */
2740                         tp->frto = 0;
2741                 }
2742         }
2743 
2744         if (recovered) {
2745                 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2746                 tcp_try_undo_recovery(sk);
2747                 return;
2748         }
2749         if (tcp_is_reno(tp)) {
2750                 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2751                  * delivered. Lower inflight to clock out (re)tranmissions.
2752                  */
2753                 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2754                         tcp_add_reno_sack(sk);
2755                 else if (flag & FLAG_SND_UNA_ADVANCED)
2756                         tcp_reset_reno_sack(tp);
2757         }
2758         tcp_xmit_retransmit_queue(sk);
2759 }
2760 
2761 /* Undo during fast recovery after partial ACK. */
2762 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2763                                  const int prior_unsacked, int flag)
2764 {
2765         struct tcp_sock *tp = tcp_sk(sk);
2766 
2767         if (tp->undo_marker && tcp_packet_delayed(tp)) {
2768                 /* Plain luck! Hole if filled with delayed
2769                  * packet, rather than with a retransmit.
2770                  */
2771                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2772 
2773                 /* We are getting evidence that the reordering degree is higher
2774                  * than we realized. If there are no retransmits out then we
2775                  * can undo. Otherwise we clock out new packets but do not
2776                  * mark more packets lost or retransmit more.
2777                  */
2778                 if (tp->retrans_out) {
2779                         tcp_cwnd_reduction(sk, prior_unsacked, 0, flag);
2780                         return true;
2781                 }
2782 
2783                 if (!tcp_any_retrans_done(sk))
2784                         tp->retrans_stamp = 0;
2785 
2786                 DBGUNDO(sk, "partial recovery");
2787                 tcp_undo_cwnd_reduction(sk, true);
2788                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2789                 tcp_try_keep_open(sk);
2790                 return true;
2791         }
2792         return false;
2793 }
2794 
2795 /* Process an event, which can update packets-in-flight not trivially.
2796  * Main goal of this function is to calculate new estimate for left_out,
2797  * taking into account both packets sitting in receiver's buffer and
2798  * packets lost by network.
2799  *
2800  * Besides that it does CWND reduction, when packet loss is detected
2801  * and changes state of machine.
2802  *
2803  * It does _not_ decide what to send, it is made in function
2804  * tcp_xmit_retransmit_queue().
2805  */
2806 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2807                                   const int prior_unsacked,
2808                                   bool is_dupack, int flag)
2809 {
2810         struct inet_connection_sock *icsk = inet_csk(sk);
2811         struct tcp_sock *tp = tcp_sk(sk);
2812         bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2813                                     (tcp_fackets_out(tp) > tp->reordering));
2814         int fast_rexmit = 0;
2815 
2816         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2817                 tp->sacked_out = 0;
2818         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2819                 tp->fackets_out = 0;
2820 
2821         /* Now state machine starts.
2822          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2823         if (flag & FLAG_ECE)
2824                 tp->prior_ssthresh = 0;
2825 
2826         /* B. In all the states check for reneging SACKs. */
2827         if (tcp_check_sack_reneging(sk, flag))
2828                 return;
2829 
2830         /* C. Check consistency of the current state. */
2831         tcp_verify_left_out(tp);
2832 
2833         /* D. Check state exit conditions. State can be terminated
2834          *    when high_seq is ACKed. */
2835         if (icsk->icsk_ca_state == TCP_CA_Open) {
2836                 WARN_ON(tp->retrans_out != 0);
2837                 tp->retrans_stamp = 0;
2838         } else if (!before(tp->snd_una, tp->high_seq)) {
2839                 switch (icsk->icsk_ca_state) {
2840                 case TCP_CA_CWR:
2841                         /* CWR is to be held something *above* high_seq
2842                          * is ACKed for CWR bit to reach receiver. */
2843                         if (tp->snd_una != tp->high_seq) {
2844                                 tcp_end_cwnd_reduction(sk);
2845                                 tcp_set_ca_state(sk, TCP_CA_Open);
2846                         }
2847                         break;
2848 
2849                 case TCP_CA_Recovery:
2850                         if (tcp_is_reno(tp))
2851                                 tcp_reset_reno_sack(tp);
2852                         if (tcp_try_undo_recovery(sk))
2853                                 return;
2854                         tcp_end_cwnd_reduction(sk);
2855                         break;
2856                 }
2857         }
2858 
2859         /* E. Process state. */
2860         switch (icsk->icsk_ca_state) {
2861         case TCP_CA_Recovery:
2862                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2863                         if (tcp_is_reno(tp) && is_dupack)
2864                                 tcp_add_reno_sack(sk);
2865                 } else {
2866                         if (tcp_try_undo_partial(sk, acked, prior_unsacked, flag))
2867                                 return;
2868                         /* Partial ACK arrived. Force fast retransmit. */
2869                         do_lost = tcp_is_reno(tp) ||
2870                                   tcp_fackets_out(tp) > tp->reordering;
2871                 }
2872                 if (tcp_try_undo_dsack(sk)) {
2873                         tcp_try_keep_open(sk);
2874                         return;
2875                 }
2876                 break;
2877         case TCP_CA_Loss:
2878                 tcp_process_loss(sk, flag, is_dupack);
2879                 if (icsk->icsk_ca_state != TCP_CA_Open &&
2880                     !(flag & FLAG_LOST_RETRANS))
2881                         return;
2882                 /* Change state if cwnd is undone or retransmits are lost */
2883         default:
2884                 if (tcp_is_reno(tp)) {
2885                         if (flag & FLAG_SND_UNA_ADVANCED)
2886                                 tcp_reset_reno_sack(tp);
2887                         if (is_dupack)
2888                                 tcp_add_reno_sack(sk);
2889                 }
2890 
2891                 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2892                         tcp_try_undo_dsack(sk);
2893 
2894                 if (!tcp_time_to_recover(sk, flag)) {
2895                         tcp_try_to_open(sk, flag, prior_unsacked);
2896                         return;
2897                 }
2898 
2899                 /* MTU probe failure: don't reduce cwnd */
2900                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2901                     icsk->icsk_mtup.probe_size &&
2902                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2903                         tcp_mtup_probe_failed(sk);
2904                         /* Restores the reduction we did in tcp_mtup_probe() */
2905                         tp->snd_cwnd++;
2906                         tcp_simple_retransmit(sk);
2907                         return;
2908                 }
2909 
2910                 /* Otherwise enter Recovery state */
2911                 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2912                 fast_rexmit = 1;
2913         }
2914 
2915         if (do_lost)
2916                 tcp_update_scoreboard(sk, fast_rexmit);
2917         tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit, flag);
2918         tcp_xmit_retransmit_queue(sk);
2919 }
2920 
2921 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2922                                       long seq_rtt_us, long sack_rtt_us)
2923 {
2924         const struct tcp_sock *tp = tcp_sk(sk);
2925 
2926         /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2927          * broken middle-boxes or peers may corrupt TS-ECR fields. But
2928          * Karn's algorithm forbids taking RTT if some retransmitted data
2929          * is acked (RFC6298).
2930          */
2931         if (flag & FLAG_RETRANS_DATA_ACKED)
2932                 seq_rtt_us = -1L;
2933 
2934         if (seq_rtt_us < 0)
2935                 seq_rtt_us = sack_rtt_us;
2936 
2937         /* RTTM Rule: A TSecr value received in a segment is used to
2938          * update the averaged RTT measurement only if the segment
2939          * acknowledges some new data, i.e., only if it advances the
2940          * left edge of the send window.
2941          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2942          */
2943         if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2944             flag & FLAG_ACKED)
2945                 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2946 
2947         if (seq_rtt_us < 0)
2948                 return false;
2949 
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 static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
2960 {
2961         struct tcp_sock *tp = tcp_sk(sk);
2962         long seq_rtt_us = -1L;
2963 
2964         if (synack_stamp && !tp->total_retrans)
2965                 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - synack_stamp);
2966 
2967         /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2968          * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2969          */
2970         if (!tp->srtt_us)
2971                 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt_us, -1L);
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_EARLY_RETRANS ||
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                         if (delta > 0)
3011                                 rto = delta;
3012                 }
3013                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3014                                           TCP_RTO_MAX);
3015         }
3016 }
3017 
3018 /* This function is called when the delayed ER timer fires. TCP enters
3019  * fast recovery and performs fast-retransmit.
3020  */
3021 void tcp_resume_early_retransmit(struct sock *sk)
3022 {
3023         struct tcp_sock *tp = tcp_sk(sk);
3024 
3025         tcp_rearm_rto(sk);
3026 
3027         /* Stop if ER is disabled after the delayed ER timer is scheduled */
3028         if (!tp->do_early_retrans)
3029                 return;
3030 
3031         tcp_enter_recovery(sk, false);
3032         tcp_update_scoreboard(sk, 1);
3033         tcp_xmit_retransmit_queue(sk);
3034 }
3035 
3036 /* If we get here, the whole TSO packet has not been acked. */
3037 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3038 {
3039         struct tcp_sock *tp = tcp_sk(sk);
3040         u32 packets_acked;
3041 
3042         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3043 
3044         packets_acked = tcp_skb_pcount(skb);
3045         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3046                 return 0;
3047         packets_acked -= tcp_skb_pcount(skb);
3048 
3049         if (packets_acked) {
3050                 BUG_ON(tcp_skb_pcount(skb) == 0);
3051                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3052         }
3053 
3054         return packets_acked;
3055 }
3056 
3057 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3058                            u32 prior_snd_una)
3059 {
3060         const struct skb_shared_info *shinfo;
3061 
3062         /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3063         if (likely(!(sk->sk_tsflags & SOF_TIMESTAMPING_TX_ACK)))
3064                 return;
3065 
3066         shinfo = skb_shinfo(skb);
3067         if ((shinfo->tx_flags & SKBTX_ACK_TSTAMP) &&
3068             between(shinfo->tskey, prior_snd_una, tcp_sk(sk)->snd_una - 1))
3069                 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3070 }
3071 
3072 /* Remove acknowledged frames from the retransmission queue. If our packet
3073  * is before the ack sequence we can discard it as it's confirmed to have
3074  * arrived at the other end.
3075  */
3076 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3077                                u32 prior_snd_una,
3078                                struct tcp_sacktag_state *sack)
3079 {
3080         const struct inet_connection_sock *icsk = inet_csk(sk);
3081         struct skb_mstamp first_ackt, last_ackt, now;
3082         struct tcp_sock *tp = tcp_sk(sk);
3083         u32 prior_sacked = tp->sacked_out;
3084         u32 reord = tp->packets_out;
3085         bool fully_acked = true;
3086         long sack_rtt_us = -1L;
3087         long seq_rtt_us = -1L;
3088         long ca_rtt_us = -1L;
3089         struct sk_buff *skb;
3090         u32 pkts_acked = 0;
3091         bool rtt_update;
3092         int flag = 0;
3093 
3094         first_ackt.v64 = 0;
3095 
3096         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3097                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3098                 u8 sacked = scb->sacked;
3099                 u32 acked_pcount;
3100 
3101                 tcp_ack_tstamp(sk, skb, prior_snd_una);
3102 
3103                 /* Determine how many packets and what bytes were acked, tso and else */
3104                 if (after(scb->end_seq, tp->snd_una)) {
3105                         if (tcp_skb_pcount(skb) == 1 ||
3106                             !after(tp->snd_una, scb->seq))
3107                                 break;
3108 
3109                         acked_pcount = tcp_tso_acked(sk, skb);
3110                         if (!acked_pcount)
3111                                 break;
3112 
3113                         fully_acked = false;
3114                 } else {
3115                         /* Speedup tcp_unlink_write_queue() and next loop */
3116                         prefetchw(skb->next);
3117                         acked_pcount = tcp_skb_pcount(skb);
3118                 }
3119 
3120                 if (unlikely(sacked & TCPCB_RETRANS)) {
3121                         if (sacked & TCPCB_SACKED_RETRANS)
3122                                 tp->retrans_out -= acked_pcount;
3123                         flag |= FLAG_RETRANS_DATA_ACKED;
3124                 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3125                         last_ackt = skb->skb_mstamp;
3126                         WARN_ON_ONCE(last_ackt.v64 == 0);
3127                         if (!first_ackt.v64)
3128                                 first_ackt = last_ackt;
3129 
3130                         reord = min(pkts_acked, reord);
3131                         if (!after(scb->end_seq, tp->high_seq))
3132                                 flag |= FLAG_ORIG_SACK_ACKED;
3133                 }
3134 
3135                 if (sacked & TCPCB_SACKED_ACKED)
3136                         tp->sacked_out -= acked_pcount;
3137                 if (sacked & TCPCB_LOST)
3138                         tp->lost_out -= acked_pcount;
3139 
3140                 tp->packets_out -= acked_pcount;
3141                 pkts_acked += acked_pcount;
3142 
3143                 /* Initial outgoing SYN's get put onto the write_queue
3144                  * just like anything else we transmit.  It is not
3145                  * true data, and if we misinform our callers that
3146                  * this ACK acks real data, we will erroneously exit
3147                  * connection startup slow start one packet too
3148                  * quickly.  This is severely frowned upon behavior.
3149                  */
3150                 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3151                         flag |= FLAG_DATA_ACKED;
3152                 } else {
3153                         flag |= FLAG_SYN_ACKED;
3154                         tp->retrans_stamp = 0;
3155                 }
3156 
3157                 if (!fully_acked)
3158                         break;
3159 
3160                 tcp_unlink_write_queue(skb, sk);
3161                 sk_wmem_free_skb(sk, skb);
3162                 if (unlikely(skb == tp->retransmit_skb_hint))
3163                         tp->retransmit_skb_hint = NULL;
3164                 if (unlikely(skb == tp->lost_skb_hint))
3165                         tp->lost_skb_hint = NULL;
3166         }
3167 
3168         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3169                 tp->snd_up = tp->snd_una;
3170 
3171         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3172                 flag |= FLAG_SACK_RENEGING;
3173 
3174         skb_mstamp_get(&now);
3175         if (likely(first_ackt.v64)) {
3176                 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3177                 ca_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3178         }
3179         if (sack->first_sackt.v64) {
3180                 sack_rtt_us = skb_mstamp_us_delta(&now, &sack->first_sackt);
3181                 ca_rtt_us = skb_mstamp_us_delta(&now, &sack->last_sackt);
3182         }
3183 
3184         rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us);
3185 
3186         if (flag & FLAG_ACKED) {
3187                 tcp_rearm_rto(sk);
3188                 if (unlikely(icsk->icsk_mtup.probe_size &&
3189                              !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3190                         tcp_mtup_probe_success(sk);
3191                 }
3192 
3193                 if (tcp_is_reno(tp)) {
3194                         tcp_remove_reno_sacks(sk, pkts_acked);
3195                 } else {
3196                         int delta;
3197 
3198                         /* Non-retransmitted hole got filled? That's reordering */
3199                         if (reord < prior_fackets)
3200                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3201 
3202                         delta = tcp_is_fack(tp) ? pkts_acked :
3203                                                   prior_sacked - tp->sacked_out;
3204                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3205                 }
3206 
3207                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3208 
3209         } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3210                    sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3211                 /* Do not re-arm RTO if the sack RTT is measured from data sent
3212                  * after when the head was last (re)transmitted. Otherwise the
3213                  * timeout may continue to extend in loss recovery.
3214                  */
3215                 tcp_rearm_rto(sk);
3216         }
3217 
3218         if (icsk->icsk_ca_ops->pkts_acked)
3219                 icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked, ca_rtt_us);
3220 
3221 #if FASTRETRANS_DEBUG > 0
3222         WARN_ON((int)tp->sacked_out < 0);
3223         WARN_ON((int)tp->lost_out < 0);
3224         WARN_ON((int)tp->retrans_out < 0);
3225         if (!tp->packets_out && tcp_is_sack(tp)) {
3226                 icsk = inet_csk(sk);
3227                 if (tp->lost_out) {
3228                         pr_debug("Leak l=%u %d\n",
3229                                  tp->lost_out, icsk->icsk_ca_state);
3230                         tp->lost_out = 0;
3231                 }
3232                 if (tp->sacked_out) {
3233                         pr_debug("Leak s=%u %d\n",
3234                                  tp->sacked_out, icsk->icsk_ca_state);
3235                         tp->sacked_out = 0;
3236                 }
3237                 if (tp->retrans_out) {
3238                         pr_debug("Leak r=%u %d\n",
3239                                  tp->retrans_out, icsk->icsk_ca_state);
3240                         tp->retrans_out = 0;
3241                 }
3242         }
3243 #endif
3244         return flag;
3245 }
3246 
3247 static void tcp_ack_probe(struct sock *sk)
3248 {
3249         const struct tcp_sock *tp = tcp_sk(sk);
3250         struct inet_connection_sock *icsk = inet_csk(sk);
3251 
3252         /* Was it a usable window open? */
3253 
3254         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3255                 icsk->icsk_backoff = 0;
3256                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3257                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3258                  * This function is not for random using!
3259                  */
3260         } else {
3261                 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3262 
3263                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3264                                           when, TCP_RTO_MAX);
3265         }
3266 }
3267 
3268 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3269 {
3270         return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3271                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3272 }
3273 
3274 /* Decide wheather to run the increase function of congestion control. */
3275 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3276 {
3277         if (tcp_in_cwnd_reduction(sk))
3278                 return false;
3279 
3280         /* If reordering is high then always grow cwnd whenever data is
3281          * delivered regardless of its ordering. Otherwise stay conservative
3282          * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3283          * new SACK or ECE mark may first advance cwnd here and later reduce
3284          * cwnd in tcp_fastretrans_alert() based on more states.
3285          */
3286         if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3287                 return flag & FLAG_FORWARD_PROGRESS;
3288 
3289         return flag & FLAG_DATA_ACKED;
3290 }
3291 
3292 /* Check that window update is acceptable.
3293  * The function assumes that snd_una<=ack<=snd_next.
3294  */
3295 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3296                                         const u32 ack, const u32 ack_seq,
3297                                         const u32 nwin)
3298 {
3299         return  after(ack, tp->snd_una) ||
3300                 after(ack_seq, tp->snd_wl1) ||
3301                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3302 }
3303 
3304 /* If we update tp->snd_una, also update tp->bytes_acked */
3305 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3306 {
3307         u32 delta = ack - tp->snd_una;
3308 
3309         u64_stats_update_begin(&tp->syncp);
3310         tp->bytes_acked += delta;
3311         u64_stats_update_end(&tp->syncp);
3312         tp->snd_una = ack;
3313 }
3314 
3315 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3316 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3317 {
3318         u32 delta = seq - tp->rcv_nxt;
3319 
3320         u64_stats_update_begin(&tp->syncp);
3321         tp->bytes_received += delta;
3322         u64_stats_update_end(&tp->syncp);
3323         tp->rcv_nxt = seq;
3324 }
3325 
3326 /* Update our send window.
3327  *
3328  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3329  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3330  */
3331 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3332                                  u32 ack_seq)
3333 {
3334         struct tcp_sock *tp = tcp_sk(sk);
3335         int flag = 0;
3336         u32 nwin = ntohs(tcp_hdr(skb)->window);
3337 
3338         if (likely(!tcp_hdr(skb)->syn))
3339                 nwin <<= tp->rx_opt.snd_wscale;
3340 
3341         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3342                 flag |= FLAG_WIN_UPDATE;
3343                 tcp_update_wl(tp, ack_seq);
3344 
3345                 if (tp->snd_wnd != nwin) {
3346                         tp->snd_wnd = nwin;
3347 
3348                         /* Note, it is the only place, where
3349                          * fast path is recovered for sending TCP.
3350                          */
3351                         tp->pred_flags = 0;
3352                         tcp_fast_path_check(sk);
3353 
3354                         if (tcp_send_head(sk))
3355                                 tcp_slow_start_after_idle_check(sk);
3356 
3357                         if (nwin > tp->max_window) {
3358                                 tp->max_window = nwin;
3359                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3360                         }
3361                 }
3362         }
3363 
3364         tcp_snd_una_update(tp, ack);
3365 
3366         return flag;
3367 }
3368 
3369 /* Return true if we're currently rate-limiting out-of-window ACKs and
3370  * thus shouldn't send a dupack right now. We rate-limit dupacks in
3371  * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3372  * attacks that send repeated SYNs or ACKs for the same connection. To
3373  * do this, we do not send a duplicate SYNACK or ACK if the remote
3374  * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3375  */
3376 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3377                           int mib_idx, u32 *last_oow_ack_time)
3378 {
3379         /* Data packets without SYNs are not likely part of an ACK loop. */
3380         if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3381             !tcp_hdr(skb)->syn)
3382                 goto not_rate_limited;
3383 
3384         if (*last_oow_ack_time) {
3385                 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3386 
3387                 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3388                         NET_INC_STATS_BH(net, mib_idx);
3389                         return true;    /* rate-limited: don't send yet! */
3390                 }
3391         }
3392 
3393         *last_oow_ack_time = tcp_time_stamp;
3394 
3395 not_rate_limited:
3396         return false;   /* not rate-limited: go ahead, send dupack now! */
3397 }
3398 
3399 /* RFC 5961 7 [ACK Throttling] */
3400 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3401 {
3402         /* unprotected vars, we dont care of overwrites */
3403         static u32 challenge_timestamp;
3404         static unsigned int challenge_count;
3405         struct tcp_sock *tp = tcp_sk(sk);
3406         u32 now;
3407 
3408         /* First check our per-socket dupack rate limit. */
3409         if (tcp_oow_rate_limited(sock_net(sk), skb,
3410                                  LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3411                                  &tp->last_oow_ack_time))
3412                 return;
3413 
3414         /* Then check the check host-wide RFC 5961 rate limit. */
3415         now = jiffies / HZ;
3416         if (now != challenge_timestamp) {
3417                 challenge_timestamp = now;
3418                 challenge_count = 0;
3419         }
3420         if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3421                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3422                 tcp_send_ack(sk);
3423         }
3424 }
3425 
3426 static void tcp_store_ts_recent(struct tcp_sock *tp)
3427 {
3428         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3429         tp->rx_opt.ts_recent_stamp = get_seconds();
3430 }
3431 
3432 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3433 {
3434         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3435                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3436                  * extra check below makes sure this can only happen
3437                  * for pure ACK frames.  -DaveM
3438                  *
3439                  * Not only, also it occurs for expired timestamps.
3440                  */
3441 
3442                 if (tcp_paws_check(&tp->rx_opt, 0))
3443                         tcp_store_ts_recent(tp);
3444         }
3445 }
3446 
3447 /* This routine deals with acks during a TLP episode.
3448  * We mark the end of a TLP episode on receiving TLP dupack or when
3449  * ack is after tlp_high_seq.
3450  * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3451  */
3452 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3453 {
3454         struct tcp_sock *tp = tcp_sk(sk);
3455 
3456         if (before(ack, tp->tlp_high_seq))
3457                 return;
3458 
3459         if (flag & FLAG_DSACKING_ACK) {
3460                 /* This DSACK means original and TLP probe arrived; no loss */
3461                 tp->tlp_high_seq = 0;
3462         } else if (after(ack, tp->tlp_high_seq)) {
3463                 /* ACK advances: there was a loss, so reduce cwnd. Reset
3464                  * tlp_high_seq in tcp_init_cwnd_reduction()
3465                  */
3466                 tcp_init_cwnd_reduction(sk);
3467                 tcp_set_ca_state(sk, TCP_CA_CWR);
3468                 tcp_end_cwnd_reduction(sk);
3469                 tcp_try_keep_open(sk);
3470                 NET_INC_STATS_BH(sock_net(sk),
3471                                  LINUX_MIB_TCPLOSSPROBERECOVERY);
3472         } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3473                              FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3474                 /* Pure dupack: original and TLP probe arrived; no loss */
3475                 tp->tlp_high_seq = 0;
3476         }
3477 }
3478 
3479 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3480 {
3481         const struct inet_connection_sock *icsk = inet_csk(sk);
3482 
3483         if (icsk->icsk_ca_ops->in_ack_event)
3484                 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3485 }
3486 
3487 /* This routine deals with incoming acks, but not outgoing ones. */
3488 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3489 {
3490         struct inet_connection_sock *icsk = inet_csk(sk);
3491         struct tcp_sock *tp = tcp_sk(sk);
3492         struct tcp_sacktag_state sack_state;
3493         u32 prior_snd_una = tp->snd_una;
3494         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3495         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3496         bool is_dupack = false;
3497         u32 prior_fackets;
3498         int prior_packets = tp->packets_out;
3499         const int prior_unsacked = tp->packets_out - tp->sacked_out;
3500         int acked = 0; /* Number of packets newly acked */
3501 
3502         sack_state.first_sackt.v64 = 0;
3503 
3504         /* We very likely will need to access write queue head. */
3505         prefetchw(sk->sk_write_queue.next);
3506 
3507         /* If the ack is older than previous acks
3508          * then we can probably ignore it.
3509          */
3510         if (before(ack, prior_snd_una)) {
3511                 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3512                 if (before(ack, prior_snd_una - tp->max_window)) {
3513                         tcp_send_challenge_ack(sk, skb);
3514                         return -1;
3515                 }
3516                 goto old_ack;
3517         }
3518 
3519         /* If the ack includes data we haven't sent yet, discard
3520          * this segment (RFC793 Section 3.9).
3521          */
3522         if (after(ack, tp->snd_nxt))
3523                 goto invalid_ack;
3524 
3525         if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3526             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3527                 tcp_rearm_rto(sk);
3528 
3529         if (after(ack, prior_snd_una)) {
3530                 flag |= FLAG_SND_UNA_ADVANCED;
3531                 icsk->icsk_retransmits = 0;
3532         }
3533 
3534         prior_fackets = tp->fackets_out;
3535 
3536         /* ts_recent update must be made after we are sure that the packet
3537          * is in window.
3538          */
3539         if (flag & FLAG_UPDATE_TS_RECENT)
3540                 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3541 
3542         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3543                 /* Window is constant, pure forward advance.
3544                  * No more checks are required.
3545                  * Note, we use the fact that SND.UNA>=SND.WL2.
3546                  */
3547                 tcp_update_wl(tp, ack_seq);
3548                 tcp_snd_una_update(tp, ack);
3549                 flag |= FLAG_WIN_UPDATE;
3550 
3551                 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3552 
3553                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3554         } else {
3555                 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3556 
3557                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3558                         flag |= FLAG_DATA;
3559                 else
3560                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3561 
3562                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3563 
3564                 if (TCP_SKB_CB(skb)->sacked)
3565                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3566                                                         &sack_state);
3567 
3568                 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3569                         flag |= FLAG_ECE;
3570                         ack_ev_flags |= CA_ACK_ECE;
3571                 }
3572 
3573                 if (flag & FLAG_WIN_UPDATE)
3574                         ack_ev_flags |= CA_ACK_WIN_UPDATE;
3575 
3576                 tcp_in_ack_event(sk, ack_ev_flags);
3577         }
3578 
3579         /* We passed data and got it acked, remove any soft error
3580          * log. Something worked...
3581          */
3582         sk->sk_err_soft = 0;
3583         icsk->icsk_probes_out = 0;
3584         tp->rcv_tstamp = tcp_time_stamp;
3585         if (!prior_packets)
3586                 goto no_queue;
3587 
3588         /* See if we can take anything off of the retransmit queue. */
3589         acked = tp->packets_out;
3590         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
3591                                     &sack_state);
3592         acked -= tp->packets_out;
3593 
3594         if (tcp_ack_is_dubious(sk, flag)) {
3595                 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3596                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3597                                       is_dupack, flag);
3598         }
3599         if (tp->tlp_high_seq)
3600                 tcp_process_tlp_ack(sk, ack, flag);
3601 
3602         /* Advance cwnd if state allows */
3603         if (tcp_may_raise_cwnd(sk, flag))
3604                 tcp_cong_avoid(sk, ack, acked);
3605 
3606         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3607                 struct dst_entry *dst = __sk_dst_get(sk);
3608                 if (dst)
3609                         dst_confirm(dst);
3610         }
3611 
3612         if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3613                 tcp_schedule_loss_probe(sk);
3614         tcp_update_pacing_rate(sk);
3615         return 1;
3616 
3617 no_queue:
3618         /* If data was DSACKed, see if we can undo a cwnd reduction. */
3619         if (flag & FLAG_DSACKING_ACK)
3620                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3621                                       is_dupack, flag);
3622         /* If this ack opens up a zero window, clear backoff.  It was
3623          * being used to time the probes, and is probably far higher than
3624          * it needs to be for normal retransmission.
3625          */
3626         if (tcp_send_head(sk))
3627                 tcp_ack_probe(sk);
3628 
3629         if (tp->tlp_high_seq)
3630                 tcp_process_tlp_ack(sk, ack, flag);
3631         return 1;
3632 
3633 invalid_ack:
3634         SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3635         return -1;
3636 
3637 old_ack:
3638         /* If data was SACKed, tag it and see if we should send more data.
3639          * If data was DSACKed, see if we can undo a cwnd reduction.
3640          */
3641         if (TCP_SKB_CB(skb)->sacked) {
3642                 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3643                                                 &sack_state);
3644                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3645                                       is_dupack, flag);
3646         }
3647 
3648         SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3649         return 0;
3650 }
3651 
3652 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3653                                       bool syn, struct tcp_fastopen_cookie *foc,
3654                                       bool exp_opt)
3655 {
3656         /* Valid only in SYN or SYN-ACK with an even length.  */
3657         if (!foc || !syn || len < 0 || (len & 1))
3658                 return;
3659 
3660         if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3661             len <= TCP_FASTOPEN_COOKIE_MAX)
3662                 memcpy(foc->val, cookie, len);
3663         else if (len != 0)
3664                 len = -1;
3665         foc->len = len;
3666         foc->exp = exp_opt;
3667 }
3668 
3669 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3670  * But, this can also be called on packets in the established flow when
3671  * the fast version below fails.
3672  */
3673 void tcp_parse_options(const struct sk_buff *skb,
3674                        struct tcp_options_received *opt_rx, int estab,
3675                        struct tcp_fastopen_cookie *foc)
3676 {
3677         const unsigned char *ptr;
3678         const struct tcphdr *th = tcp_hdr(skb);
3679         int length = (th->doff * 4) - sizeof(struct tcphdr);
3680 
3681         ptr = (const unsigned char *)(th + 1);
3682         opt_rx->saw_tstamp = 0;
3683 
3684         while (length > 0) {
3685                 int opcode = *ptr++;
3686                 int opsize;
3687 
3688                 switch (opcode) {
3689                 case TCPOPT_EOL:
3690                         return;
3691                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3692                         length--;
3693                         continue;
3694                 default:
3695                         opsize = *ptr++;
3696                         if (opsize < 2) /* "silly options" */
3697                                 return;
3698                         if (opsize > length)
3699                                 return; /* don't parse partial options */
3700                         switch (opcode) {
3701                         case TCPOPT_MSS:
3702                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3703                                         u16 in_mss = get_unaligned_be16(ptr);
3704                                         if (in_mss) {
3705                                                 if (opt_rx->user_mss &&
3706                                                     opt_rx->user_mss < in_mss)
3707                                                         in_mss = opt_rx->user_mss;
3708                                                 opt_rx->mss_clamp = in_mss;
3709                                         }
3710                                 }
3711                                 break;
3712                         case TCPOPT_WINDOW:
3713                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3714                                     !estab && sysctl_tcp_window_scaling) {
3715                                         __u8 snd_wscale = *(__u8 *)ptr;
3716                                         opt_rx->wscale_ok = 1;
3717                                         if (snd_wscale > 14) {
3718                                                 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3719                                                                      __func__,
3720                                                                      snd_wscale);
3721                                                 snd_wscale = 14;
3722                                         }
3723                                         opt_rx->snd_wscale = snd_wscale;
3724                                 }
3725                                 break;
3726                         case TCPOPT_TIMESTAMP:
3727                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3728                                     ((estab && opt_rx->tstamp_ok) ||
3729                                      (!estab && sysctl_tcp_timestamps))) {
3730                                         opt_rx->saw_tstamp = 1;
3731                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3732                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3733                                 }
3734                                 break;
3735                         case TCPOPT_SACK_PERM:
3736                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3737                                     !estab && sysctl_tcp_sack) {
3738                                         opt_rx->sack_ok = TCP_SACK_SEEN;
3739                                         tcp_sack_reset(opt_rx);
3740                                 }
3741                                 break;
3742 
3743                         case TCPOPT_SACK:
3744                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3745                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3746                                    opt_rx->sack_ok) {
3747                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3748                                 }
3749                                 break;
3750 #ifdef CONFIG_TCP_MD5SIG
3751                         case TCPOPT_MD5SIG:
3752                                 /*
3753                                  * The MD5 Hash has already been
3754                                  * checked (see tcp_v{4,6}_do_rcv()).
3755                                  */
3756                                 break;
3757 #endif
3758                         case TCPOPT_FASTOPEN:
3759                                 tcp_parse_fastopen_option(
3760                                         opsize - TCPOLEN_FASTOPEN_BASE,
3761                                         ptr, th->syn, foc, false);
3762                                 break;
3763 
3764                         case TCPOPT_EXP:
3765                                 /* Fast Open option shares code 254 using a
3766                                  * 16 bits magic number.
3767                                  */
3768                                 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3769                                     get_unaligned_be16(ptr) ==
3770                                     TCPOPT_FASTOPEN_MAGIC)
3771                                         tcp_parse_fastopen_option(opsize -
3772                                                 TCPOLEN_EXP_FASTOPEN_BASE,
3773                                                 ptr + 2, th->syn, foc, true);
3774                                 break;
3775 
3776                         }
3777                         ptr += opsize-2;
3778                         length -= opsize;
3779                 }
3780         }
3781 }
3782 EXPORT_SYMBOL(tcp_parse_options);
3783 
3784 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3785 {
3786         const __be32 *ptr = (const __be32 *)(th + 1);
3787 
3788         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3789                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3790                 tp->rx_opt.saw_tstamp = 1;
3791                 ++ptr;
3792                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3793                 ++ptr;
3794                 if (*ptr)
3795                         tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3796                 else
3797                         tp->rx_opt.rcv_tsecr = 0;
3798                 return true;
3799         }
3800         return false;
3801 }
3802 
3803 /* Fast parse options. This hopes to only see timestamps.
3804  * If it is wrong it falls back on tcp_parse_options().
3805  */
3806 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3807                                    const struct tcphdr *th, struct tcp_sock *tp)
3808 {
3809         /* In the spirit of fast parsing, compare doff directly to constant
3810          * values.  Because equality is used, short doff can be ignored here.
3811          */
3812         if (th->doff == (sizeof(*th) / 4)) {
3813                 tp->rx_opt.saw_tstamp = 0;
3814                 return false;
3815         } else if (tp->rx_opt.tstamp_ok &&
3816                    th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3817                 if (tcp_parse_aligned_timestamp(tp, th))
3818                         return true;
3819         }
3820 
3821         tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3822         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3823                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3824 
3825         return true;
3826 }
3827 
3828 #ifdef CONFIG_TCP_MD5SIG
3829 /*
3830  * Parse MD5 Signature option
3831  */
3832 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3833 {
3834         int length = (th->doff << 2) - sizeof(*th);
3835         const u8 *ptr = (const u8 *)(th + 1);
3836 
3837         /* If the TCP option is too short, we can short cut */
3838         if (length < TCPOLEN_MD5SIG)
3839                 return NULL;
3840 
3841         while (length > 0) {
3842                 int opcode = *ptr++;
3843                 int opsize;
3844 
3845                 switch (opcode) {
3846                 case TCPOPT_EOL:
3847                         return NULL;
3848                 case TCPOPT_NOP:
3849                         length--;
3850                         continue;
3851                 default:
3852                         opsize = *ptr++;
3853                         if (opsize < 2 || opsize > length)
3854                                 return NULL;
3855                         if (opcode == TCPOPT_MD5SIG)
3856                                 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3857                 }
3858                 ptr += opsize - 2;
3859                 length -= opsize;
3860         }
3861         return NULL;
3862 }
3863 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3864 #endif
3865 
3866 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3867  *
3868  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3869  * it can pass through stack. So, the following predicate verifies that
3870  * this segment is not used for anything but congestion avoidance or
3871  * fast retransmit. Moreover, we even are able to eliminate most of such
3872  * second order effects, if we apply some small "replay" window (~RTO)
3873  * to timestamp space.
3874  *
3875  * All these measures still do not guarantee that we reject wrapped ACKs
3876  * on networks with high bandwidth, when sequence space is recycled fastly,
3877  * but it guarantees that such events will be very rare and do not affect
3878  * connection seriously. This doesn't look nice, but alas, PAWS is really
3879  * buggy extension.
3880  *
3881  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3882  * states that events when retransmit arrives after original data are rare.
3883  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3884  * the biggest problem on large power networks even with minor reordering.
3885  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3886  * up to bandwidth of 18Gigabit/sec. 8) ]
3887  */
3888 
3889 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3890 {
3891         const struct tcp_sock *tp = tcp_sk(sk);
3892         const struct tcphdr *th = tcp_hdr(skb);
3893         u32 seq = TCP_SKB_CB(skb)->seq;
3894         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3895 
3896         return (/* 1. Pure ACK with correct sequence number. */
3897                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3898 
3899                 /* 2. ... and duplicate ACK. */
3900                 ack == tp->snd_una &&
3901 
3902                 /* 3. ... and does not update window. */
3903                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3904 
3905                 /* 4. ... and sits in replay window. */
3906                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3907 }
3908 
3909 static inline bool tcp_paws_discard(const struct sock *sk,
3910                                    const struct sk_buff *skb)
3911 {
3912         const struct tcp_sock *tp = tcp_sk(sk);
3913 
3914         return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3915                !tcp_disordered_ack(sk, skb);
3916 }
3917 
3918 /* Check segment sequence number for validity.
3919  *
3920  * Segment controls are considered valid, if the segment
3921  * fits to the window after truncation to the window. Acceptability
3922  * of data (and SYN, FIN, of course) is checked separately.
3923  * See tcp_data_queue(), for example.
3924  *
3925  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3926  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3927  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3928  * (borrowed from freebsd)
3929  */
3930 
3931 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3932 {
3933         return  !before(end_seq, tp->rcv_wup) &&
3934                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3935 }
3936 
3937 /* When we get a reset we do this. */
3938 void tcp_reset(struct sock *sk)
3939 {
3940         /* We want the right error as BSD sees it (and indeed as we do). */
3941         switch (sk->sk_state) {
3942         case TCP_SYN_SENT:
3943                 sk->sk_err = ECONNREFUSED;
3944                 break;
3945         case TCP_CLOSE_WAIT:
3946                 sk->sk_err = EPIPE;
3947                 break;
3948         case TCP_CLOSE:
3949                 return;
3950         default:
3951                 sk->sk_err = ECONNRESET;
3952         }
3953         /* This barrier is coupled with smp_rmb() in tcp_poll() */
3954         smp_wmb();
3955 
3956         if (!sock_flag(sk, SOCK_DEAD))
3957                 sk->sk_error_report(sk);
3958 
3959         tcp_done(sk);
3960 }
3961 
3962 /*
3963  *      Process the FIN bit. This now behaves as it is supposed to work
3964  *      and the FIN takes effect when it is validly part of sequence
3965  *      space. Not before when we get holes.
3966  *
3967  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3968  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3969  *      TIME-WAIT)
3970  *
3971  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3972  *      close and we go into CLOSING (and later onto TIME-WAIT)
3973  *
3974  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3975  */
3976 static void tcp_fin(struct sock *sk)
3977 {
3978         struct tcp_sock *tp = tcp_sk(sk);
3979 
3980         inet_csk_schedule_ack(sk);
3981 
3982         sk->sk_shutdown |= RCV_SHUTDOWN;
3983         sock_set_flag(sk, SOCK_DONE);
3984 
3985         switch (sk->sk_state) {
3986         case TCP_SYN_RECV:
3987         case TCP_ESTABLISHED:
3988                 /* Move to CLOSE_WAIT */
3989                 tcp_set_state(sk, TCP_CLOSE_WAIT);
3990                 inet_csk(sk)->icsk_ack.pingpong = 1;
3991                 break;
3992 
3993         case TCP_CLOSE_WAIT:
3994         case TCP_CLOSING:
3995                 /* Received a retransmission of the FIN, do
3996                  * nothing.
3997                  */
3998                 break;
3999         case TCP_LAST_ACK:
4000                 /* RFC793: Remain in the LAST-ACK state. */
4001                 break;
4002 
4003         case TCP_FIN_WAIT1:
4004                 /* This case occurs when a simultaneous close
4005                  * happens, we must ack the received FIN and
4006                  * enter the CLOSING state.
4007                  */
4008                 tcp_send_ack(sk);
4009                 tcp_set_state(sk, TCP_CLOSING);
4010                 break;
4011         case TCP_FIN_WAIT2:
4012                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4013                 tcp_send_ack(sk);
4014                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4015                 break;
4016         default:
4017                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4018                  * cases we should never reach this piece of code.
4019                  */
4020                 pr_err("%s: Impossible, sk->sk_state=%d\n",
4021                        __func__, sk->sk_state);
4022                 break;
4023         }
4024 
4025         /* It _is_ possible, that we have something out-of-order _after_ FIN.
4026          * Probably, we should reset in this case. For now drop them.
4027          */
4028         __skb_queue_purge(&tp->out_of_order_queue);
4029         if (tcp_is_sack(tp))
4030                 tcp_sack_reset(&tp->rx_opt);
4031         sk_mem_reclaim(sk);
4032 
4033         if (!sock_flag(sk, SOCK_DEAD)) {
4034                 sk->sk_state_change(sk);
4035 
4036                 /* Do not send POLL_HUP for half duplex close. */
4037                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4038                     sk->sk_state == TCP_CLOSE)
4039                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4040                 else
4041                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4042         }
4043 }
4044 
4045 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4046                                   u32 end_seq)
4047 {
4048         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4049                 if (before(seq, sp->start_seq))
4050                         sp->start_seq = seq;
4051                 if (after(end_seq, sp->end_seq))
4052                         sp->end_seq = end_seq;
4053                 return true;
4054         }
4055         return false;
4056 }
4057 
4058 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4059 {
4060         struct tcp_sock *tp = tcp_sk(sk);
4061 
4062         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4063                 int mib_idx;
4064 
4065                 if (before(seq, tp->rcv_nxt))
4066                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4067                 else
4068                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4069 
4070                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4071 
4072                 tp->rx_opt.dsack = 1;
4073                 tp->duplicate_sack[0].start_seq = seq;
4074                 tp->duplicate_sack[0].end_seq = end_seq;
4075         }
4076 }
4077 
4078 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4079 {
4080         struct tcp_sock *tp = tcp_sk(sk);
4081 
4082         if (!tp->rx_opt.dsack)
4083                 tcp_dsack_set(sk, seq, end_seq);
4084         else
4085                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4086 }
4087 
4088 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4089 {
4090         struct tcp_sock *tp = tcp_sk(sk);
4091 
4092         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4093             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4094                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4095                 tcp_enter_quickack_mode(sk);
4096 
4097                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4098                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4099 
4100                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4101                                 end_seq = tp->rcv_nxt;
4102                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4103                 }
4104         }
4105 
4106         tcp_send_ack(sk);
4107 }
4108 
4109 /* These routines update the SACK block as out-of-order packets arrive or
4110  * in-order packets close up the sequence space.
4111  */
4112 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4113 {
4114         int this_sack;
4115         struct tcp_sack_block *sp = &tp->selective_acks[0];
4116         struct tcp_sack_block *swalk = sp + 1;
4117 
4118         /* See if the recent change to the first SACK eats into
4119          * or hits the sequence space of other SACK blocks, if so coalesce.
4120          */
4121         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4122                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4123                         int i;
4124 
4125                         /* Zap SWALK, by moving every further SACK up by one slot.
4126                          * Decrease num_sacks.
4127                          */
4128                         tp->rx_opt.num_sacks--;
4129                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4130                                 sp[i] = sp[i + 1];
4131                         continue;
4132                 }
4133                 this_sack++, swalk++;
4134         }
4135 }
4136 
4137 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4138 {
4139         struct tcp_sock *tp = tcp_sk(sk);
4140         struct tcp_sack_block *sp = &tp->selective_acks[0];
4141         int cur_sacks = tp->rx_opt.num_sacks;
4142         int this_sack;
4143 
4144         if (!cur_sacks)
4145                 goto new_sack;
4146 
4147         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4148                 if (tcp_sack_extend(sp, seq, end_seq)) {
4149                         /* Rotate this_sack to the first one. */
4150                         for (; this_sack > 0; this_sack--, sp--)
4151                                 swap(*sp, *(sp - 1));
4152                         if (cur_sacks > 1)
4153                                 tcp_sack_maybe_coalesce(tp);
4154                         return;
4155                 }
4156         }
4157 
4158         /* Could not find an adjacent existing SACK, build a new one,
4159          * put it at the front, and shift everyone else down.  We
4160          * always know there is at least one SACK present already here.
4161          *
4162          * If the sack array is full, forget about the last one.
4163          */
4164         if (this_sack >= TCP_NUM_SACKS) {
4165                 this_sack--;
4166                 tp->rx_opt.num_sacks--;
4167                 sp--;
4168         }
4169         for (; this_sack > 0; this_sack--, sp--)
4170                 *sp = *(sp - 1);
4171 
4172 new_sack:
4173         /* Build the new head SACK, and we're done. */
4174         sp->start_seq = seq;
4175         sp->end_seq = end_seq;
4176         tp->rx_opt.num_sacks++;
4177 }
4178 
4179 /* RCV.NXT advances, some SACKs should be eaten. */
4180 
4181 static void tcp_sack_remove(struct tcp_sock *tp)
4182 {
4183         struct tcp_sack_block *sp = &tp->selective_acks[0];
4184         int num_sacks = tp->rx_opt.num_sacks;
4185         int this_sack;
4186 
4187         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4188         if (skb_queue_empty(&tp->out_of_order_queue)) {
4189                 tp->rx_opt.num_sacks = 0;
4190                 return;
4191         }
4192 
4193         for (this_sack = 0; this_sack < num_sacks;) {
4194                 /* Check if the start of the sack is covered by RCV.NXT. */
4195                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4196                         int i;
4197 
4198                         /* RCV.NXT must cover all the block! */
4199                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4200 
4201                         /* Zap this SACK, by moving forward any other SACKS. */
4202                         for (i = this_sack+1; i < num_sacks; i++)
4203                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4204                         num_sacks--;
4205                         continue;
4206                 }
4207                 this_sack++;
4208                 sp++;
4209         }
4210         tp->rx_opt.num_sacks = num_sacks;
4211 }
4212 
4213 /**
4214  * tcp_try_coalesce - try to merge skb to prior one
4215  * @sk: socket
4216  * @to: prior buffer
4217  * @from: buffer to add in queue
4218  * @fragstolen: pointer to boolean
4219  *
4220  * Before queueing skb @from after @to, try to merge them
4221  * to reduce overall memory use and queue lengths, if cost is small.
4222  * Packets in ofo or receive queues can stay a long time.
4223  * Better try to coalesce them right now to avoid future collapses.
4224  * Returns true if caller should free @from instead of queueing it
4225  */
4226 static bool tcp_try_coalesce(struct sock *sk,
4227                              struct sk_buff *to,
4228                              struct sk_buff *from,
4229                              bool *fragstolen)
4230 {
4231         int delta;
4232 
4233         *fragstolen = false;
4234 
4235         /* Its possible this segment overlaps with prior segment in queue */
4236         if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4237                 return false;
4238 
4239         if (!skb_try_coalesce(to, from, fragstolen, &delta))
4240                 return false;
4241 
4242         atomic_add(delta, &sk->sk_rmem_alloc);
4243         sk_mem_charge(sk, delta);
4244         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4245         TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4246         TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4247         TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4248         return true;
4249 }
4250 
4251 /* This one checks to see if we can put data from the
4252  * out_of_order queue into the receive_queue.
4253  */
4254 static void tcp_ofo_queue(struct sock *sk)
4255 {
4256         struct tcp_sock *tp = tcp_sk(sk);
4257         __u32 dsack_high = tp->rcv_nxt;
4258         struct sk_buff *skb, *tail;
4259         bool fragstolen, eaten;
4260 
4261         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4262                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4263                         break;
4264 
4265                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4266                         __u32 dsack = dsack_high;
4267                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4268                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4269                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4270                 }
4271 
4272                 __skb_unlink(skb, &tp->out_of_order_queue);
4273                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4274                         SOCK_DEBUG(sk, "ofo packet was already received\n");
4275                         __kfree_skb(skb);
4276                         continue;
4277                 }
4278                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4279                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4280                            TCP_SKB_CB(skb)->end_seq);
4281 
4282                 tail = skb_peek_tail(&sk->sk_receive_queue);
4283                 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4284                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4285                 if (!eaten)
4286                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4287                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4288                         tcp_fin(sk);
4289                 if (eaten)
4290                         kfree_skb_partial(skb, fragstolen);
4291         }
4292 }
4293 
4294 static bool tcp_prune_ofo_queue(struct sock *sk);
4295 static int tcp_prune_queue(struct sock *sk);
4296 
4297 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4298                                  unsigned int size)
4299 {
4300         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4301             !sk_rmem_schedule(sk, skb, size)) {
4302 
4303                 if (tcp_prune_queue(sk) < 0)
4304                         return -1;
4305 
4306                 if (!sk_rmem_schedule(sk, skb, size)) {
4307                         if (!tcp_prune_ofo_queue(sk))
4308                                 return -1;
4309 
4310                         if (!sk_rmem_schedule(sk, skb, size))
4311                                 return -1;
4312                 }
4313         }
4314         return 0;
4315 }
4316 
4317 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4318 {
4319         struct tcp_sock *tp = tcp_sk(sk);
4320         struct sk_buff *skb1;
4321         u32 seq, end_seq;
4322 
4323         tcp_ecn_check_ce(tp, skb);
4324 
4325         if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4326                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4327                 __kfree_skb(skb);
4328                 return;
4329         }
4330 
4331         /* Disable header prediction. */
4332         tp->pred_flags = 0;
4333         inet_csk_schedule_ack(sk);
4334 
4335         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4336         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4337                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4338 
4339         skb1 = skb_peek_tail(&tp->out_of_order_queue);
4340         if (!skb1) {
4341                 /* Initial out of order segment, build 1 SACK. */
4342                 if (tcp_is_sack(tp)) {
4343                         tp->rx_opt.num_sacks = 1;
4344                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4345                         tp->selective_acks[0].end_seq =
4346                                                 TCP_SKB_CB(skb)->end_seq;
4347                 }
4348                 __skb_queue_head(&tp->out_of_order_queue, skb);
4349                 goto end;
4350         }
4351 
4352         seq = TCP_SKB_CB(skb)->seq;
4353         end_seq = TCP_SKB_CB(skb)->end_seq;
4354 
4355         if (seq == TCP_SKB_CB(skb1)->end_seq) {
4356                 bool fragstolen;
4357 
4358                 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4359                         __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4360                 } else {
4361                         tcp_grow_window(sk, skb);
4362                         kfree_skb_partial(skb, fragstolen);
4363                         skb = NULL;
4364                 }
4365 
4366                 if (!tp->rx_opt.num_sacks ||
4367                     tp->selective_acks[0].end_seq != seq)
4368                         goto add_sack;
4369 
4370                 /* Common case: data arrive in order after hole. */
4371                 tp->selective_acks[0].end_seq = end_seq;
4372                 goto end;
4373         }
4374 
4375         /* Find place to insert this segment. */
4376         while (1) {
4377                 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4378                         break;
4379                 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4380                         skb1 = NULL;
4381                         break;
4382                 }
4383                 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4384         }
4385 
4386         /* Do skb overlap to previous one? */
4387         if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4388                 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4389                         /* All the bits are present. Drop. */
4390                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4391                         __kfree_skb(skb);
4392                         skb = NULL;
4393                         tcp_dsack_set(sk, seq, end_seq);
4394                         goto add_sack;
4395                 }
4396                 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4397                         /* Partial overlap. */
4398                         tcp_dsack_set(sk, seq,
4399                                       TCP_SKB_CB(skb1)->end_seq);
4400                 } else {
4401                         if (skb_queue_is_first(&tp->out_of_order_queue,
4402                                                skb1))
4403                                 skb1 = NULL;
4404                         else
4405                                 skb1 = skb_queue_prev(
4406                                         &tp->out_of_order_queue,
4407                                         skb1);
4408                 }
4409         }
4410         if (!skb1)
4411                 __skb_queue_head(&tp->out_of_order_queue, skb);
4412         else
4413                 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4414 
4415         /* And clean segments covered by new one as whole. */
4416         while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4417                 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4418 
4419                 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4420                         break;
4421                 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4422                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4423                                          end_seq);
4424                         break;
4425                 }
4426                 __skb_unlink(skb1, &tp->out_of_order_queue);
4427                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4428                                  TCP_SKB_CB(skb1)->end_seq);
4429                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4430                 __kfree_skb(skb1);
4431         }
4432 
4433 add_sack:
4434         if (tcp_is_sack(tp))
4435                 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4436 end:
4437         if (skb) {
4438                 tcp_grow_window(sk, skb);
4439                 skb_set_owner_r(skb, sk);
4440         }
4441 }
4442 
4443 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4444                   bool *fragstolen)
4445 {
4446         int eaten;
4447         struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4448 
4449         __skb_pull(skb, hdrlen);
4450         eaten = (tail &&
4451                  tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4452         tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4453         if (!eaten) {
4454                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4455                 skb_set_owner_r(skb, sk);
4456         }
4457         return eaten;
4458 }
4459 
4460 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4461 {
4462         struct sk_buff *skb;
4463         int err = -ENOMEM;
4464         int data_len = 0;
4465         bool fragstolen;
4466 
4467         if (size == 0)
4468                 return 0;
4469 
4470         if (size > PAGE_SIZE) {
4471                 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4472 
4473                 data_len = npages << PAGE_SHIFT;
4474                 size = data_len + (size & ~PAGE_MASK);
4475         }
4476         skb = alloc_skb_with_frags(size - data_len, data_len,
4477                                    PAGE_ALLOC_COSTLY_ORDER,
4478                                    &err, sk->sk_allocation);
4479         if (!skb)
4480                 goto err;
4481 
4482         skb_put(skb, size - data_len);
4483         skb->data_len = data_len;
4484         skb->len = size;
4485 
4486         if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4487                 goto err_free;
4488 
4489         err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4490         if (err)
4491                 goto err_free;
4492 
4493         TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4494         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4495         TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4496 
4497         if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4498                 WARN_ON_ONCE(fragstolen); /* should not happen */
4499                 __kfree_skb(skb);
4500         }
4501         return size;
4502 
4503 err_free:
4504         kfree_skb(skb);
4505 err:
4506         return err;
4507 
4508 }
4509 
4510 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4511 {
4512         struct tcp_sock *tp = tcp_sk(sk);
4513         int eaten = -1;
4514         bool fragstolen = false;
4515 
4516         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4517                 goto drop;
4518 
4519         skb_dst_drop(skb);
4520         __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4521 
4522         tcp_ecn_accept_cwr(tp, skb);
4523 
4524         tp->rx_opt.dsack = 0;
4525 
4526         /*  Queue data for delivery to the user.
4527          *  Packets in sequence go to the receive queue.
4528          *  Out of sequence packets to the out_of_order_queue.
4529          */
4530         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4531                 if (tcp_receive_window(tp) == 0)
4532                         goto out_of_window;
4533 
4534                 /* Ok. In sequence. In window. */
4535                 if (tp->ucopy.task == current &&
4536                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4537                     sock_owned_by_user(sk) && !tp->urg_data) {
4538                         int chunk = min_t(unsigned int, skb->len,
4539                                           tp->ucopy.len);
4540 
4541                         __set_current_state(TASK_RUNNING);
4542 
4543                         local_bh_enable();
4544                         if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4545                                 tp->ucopy.len -= chunk;
4546                                 tp->copied_seq += chunk;
4547                                 eaten = (chunk == skb->len);
4548                                 tcp_rcv_space_adjust(sk);
4549                         }
4550                         local_bh_disable();
4551                 }
4552 
4553                 if (eaten <= 0) {
4554 queue_and_out:
4555                         if (eaten < 0) {
4556                                 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4557                                         sk_forced_mem_schedule(sk, skb->truesize);
4558                                 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4559                                         goto drop;
4560                         }
4561                         eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4562                 }
4563                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4564                 if (skb->len)
4565                         tcp_event_data_recv(sk, skb);
4566                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4567                         tcp_fin(sk);
4568 
4569                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4570                         tcp_ofo_queue(sk);
4571 
4572                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4573                          * gap in queue is filled.
4574                          */
4575                         if (skb_queue_empty(&tp->out_of_order_queue))
4576                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4577                 }
4578 
4579                 if (tp->rx_opt.num_sacks)
4580                         tcp_sack_remove(tp);
4581 
4582                 tcp_fast_path_check(sk);
4583 
4584                 if (eaten > 0)
4585                         kfree_skb_partial(skb, fragstolen);
4586                 if (!sock_flag(sk, SOCK_DEAD))
4587                         sk->sk_data_ready(sk);
4588                 return;
4589         }
4590 
4591         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4592                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4593                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4594                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4595 
4596 out_of_window:
4597                 tcp_enter_quickack_mode(sk);
4598                 inet_csk_schedule_ack(sk);
4599 drop:
4600                 __kfree_skb(skb);
4601                 return;
4602         }
4603 
4604         /* Out of window. F.e. zero window probe. */
4605         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4606                 goto out_of_window;
4607 
4608         tcp_enter_quickack_mode(sk);
4609 
4610         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4611                 /* Partial packet, seq < rcv_next < end_seq */
4612                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4613                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4614                            TCP_SKB_CB(skb)->end_seq);
4615 
4616                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4617 
4618                 /* If window is closed, drop tail of packet. But after
4619                  * remembering D-SACK for its head made in previous line.
4620                  */
4621                 if (!tcp_receive_window(tp))
4622                         goto out_of_window;
4623                 goto queue_and_out;
4624         }
4625 
4626         tcp_data_queue_ofo(sk, skb);
4627 }
4628 
4629 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4630                                         struct sk_buff_head *list)
4631 {
4632         struct sk_buff *next = NULL;
4633 
4634         if (!skb_queue_is_last(list, skb))
4635                 next = skb_queue_next(list, skb);
4636 
4637         __skb_unlink(skb, list);
4638         __kfree_skb(skb);
4639         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4640 
4641         return next;
4642 }
4643 
4644 /* Collapse contiguous sequence of skbs head..tail with
4645  * sequence numbers start..end.
4646  *
4647  * If tail is NULL, this means until the end of the list.
4648  *
4649  * Segments with FIN/SYN are not collapsed (only because this
4650  * simplifies code)
4651  */
4652 static void
4653 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4654              struct sk_buff *head, struct sk_buff *tail,
4655              u32 start, u32 end)
4656 {
4657         struct sk_buff *skb, *n;
4658         bool end_of_skbs;
4659 
4660         /* First, check that queue is collapsible and find
4661          * the point where collapsing can be useful. */
4662         skb = head;
4663 restart:
4664         end_of_skbs = true;
4665         skb_queue_walk_from_safe(list, skb, n) {
4666                 if (skb == tail)
4667                         break;
4668                 /* No new bits? It is possible on ofo queue. */
4669                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4670                         skb = tcp_collapse_one(sk, skb, list);
4671                         if (!skb)
4672                                 break;
4673                         goto restart;
4674                 }
4675 
4676                 /* The first skb to collapse is:
4677                  * - not SYN/FIN and
4678                  * - bloated or contains data before "start" or
4679                  *   overlaps to the next one.
4680                  */
4681                 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4682                     (tcp_win_from_space(skb->truesize) > skb->len ||
4683                      before(TCP_SKB_CB(skb)->seq, start))) {
4684                         end_of_skbs = false;
4685                         break;
4686                 }
4687 
4688                 if (!skb_queue_is_last(list, skb)) {
4689                         struct sk_buff *next = skb_queue_next(list, skb);
4690                         if (next != tail &&
4691                             TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4692                                 end_of_skbs = false;
4693                                 break;
4694                         }
4695                 }
4696 
4697                 /* Decided to skip this, advance start seq. */
4698                 start = TCP_SKB_CB(skb)->end_seq;
4699         }
4700         if (end_of_skbs ||
4701             (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4702                 return;
4703 
4704         while (before(start, end)) {
4705                 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4706                 struct sk_buff *nskb;
4707 
4708                 nskb = alloc_skb(copy, GFP_ATOMIC);
4709                 if (!nskb)
4710                         return;
4711 
4712                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4713                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4714                 __skb_queue_before(list, skb, nskb);
4715                 skb_set_owner_r(nskb, sk);
4716 
4717                 /* Copy data, releasing collapsed skbs. */
4718                 while (copy > 0) {
4719                         int offset = start - TCP_SKB_CB(skb)->seq;
4720                         int size = TCP_SKB_CB(skb)->end_seq - start;
4721 
4722                         BUG_ON(offset < 0);
4723                         if (size > 0) {
4724                                 size = min(copy, size);
4725                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4726                                         BUG();
4727                                 TCP_SKB_CB(nskb)->end_seq += size;
4728                                 copy -= size;
4729                                 start += size;
4730                         }
4731                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4732                                 skb = tcp_collapse_one(sk, skb, list);
4733                                 if (!skb ||
4734                                     skb == tail ||
4735                                     (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4736                                         return;
4737                         }
4738                 }
4739         }
4740 }
4741 
4742 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4743  * and tcp_collapse() them until all the queue is collapsed.
4744  */
4745 static void tcp_collapse_ofo_queue(struct sock *sk)
4746 {
4747         struct tcp_sock *tp = tcp_sk(sk);
4748         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4749         struct sk_buff *head;
4750         u32 start, end;
4751 
4752         if (!skb)
4753                 return;
4754 
4755         start = TCP_SKB_CB(skb)->seq;
4756         end = TCP_SKB_CB(skb)->end_seq;
4757         head = skb;
4758 
4759         for (;;) {
4760                 struct sk_buff *next = NULL;
4761 
4762                 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4763                         next = skb_queue_next(&tp->out_of_order_queue, skb);
4764                 skb = next;
4765 
4766                 /* Segment is terminated when we see gap or when
4767                  * we are at the end of all the queue. */
4768                 if (!skb ||
4769                     after(TCP_SKB_CB(skb)->seq, end) ||
4770                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4771                         tcp_collapse(sk, &tp->out_of_order_queue,
4772                                      head, skb, start, end);
4773                         head = skb;
4774                         if (!skb)
4775                                 break;
4776                         /* Start new segment */
4777                         start = TCP_SKB_CB(skb)->seq;
4778                         end = TCP_SKB_CB(skb)->end_seq;
4779                 } else {
4780                         if (before(TCP_SKB_CB(skb)->seq, start))
4781                                 start = TCP_SKB_CB(skb)->seq;
4782                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4783                                 end = TCP_SKB_CB(skb)->end_seq;
4784                 }
4785         }
4786 }
4787 
4788 /*
4789  * Purge the out-of-order queue.
4790  * Return true if queue was pruned.
4791  */
4792 static bool tcp_prune_ofo_queue(struct sock *sk)
4793 {
4794         struct tcp_sock *tp = tcp_sk(sk);
4795         bool res = false;
4796 
4797         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4798                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4799                 __skb_queue_purge(&tp->out_of_order_queue);
4800 
4801                 /* Reset SACK state.  A conforming SACK implementation will
4802                  * do the same at a timeout based retransmit.  When a connection
4803                  * is in a sad state like this, we care only about integrity
4804                  * of the connection not performance.
4805                  */
4806                 if (tp->rx_opt.sack_ok)
4807                         tcp_sack_reset(&tp->rx_opt);
4808                 sk_mem_reclaim(sk);
4809                 res = true;
4810         }
4811         return res;
4812 }
4813 
4814 /* Reduce allocated memory if we can, trying to get
4815  * the socket within its memory limits again.
4816  *
4817  * Return less than zero if we should start dropping frames
4818  * until the socket owning process reads some of the data
4819  * to stabilize the situation.
4820  */
4821 static int tcp_prune_queue(struct sock *sk)
4822 {
4823         struct tcp_sock *tp = tcp_sk(sk);
4824 
4825         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4826 
4827         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4828 
4829         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4830                 tcp_clamp_window(sk);
4831         else if (tcp_under_memory_pressure(sk))
4832                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4833 
4834         tcp_collapse_ofo_queue(sk);
4835         if (!skb_queue_empty(&sk->sk_receive_queue))
4836                 tcp_collapse(sk, &sk->sk_receive_queue,
4837                              skb_peek(&sk->sk_receive_queue),
4838                              NULL,
4839                              tp->copied_seq, tp->rcv_nxt);
4840         sk_mem_reclaim(sk);
4841 
4842         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4843                 return 0;
4844 
4845         /* Collapsing did not help, destructive actions follow.
4846          * This must not ever occur. */
4847 
4848         tcp_prune_ofo_queue(sk);
4849 
4850         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4851                 return 0;
4852 
4853         /* If we are really being abused, tell the caller to silently
4854          * drop receive data on the floor.  It will get retransmitted
4855          * and hopefully then we'll have sufficient space.
4856          */
4857         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4858 
4859         /* Massive buffer overcommit. */
4860         tp->pred_flags = 0;
4861         return -1;
4862 }
4863 
4864 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4865 {
4866         const struct tcp_sock *tp = tcp_sk(sk);
4867 
4868         /* If the user specified a specific send buffer setting, do
4869          * not modify it.
4870          */
4871         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4872                 return false;
4873 
4874         /* If we are under global TCP memory pressure, do not expand.  */
4875         if (tcp_under_memory_pressure(sk))
4876                 return false;
4877 
4878         /* If we are under soft global TCP memory pressure, do not expand.  */
4879         if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4880                 return false;
4881 
4882         /* If we filled the congestion window, do not expand.  */
4883         if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4884                 return false;
4885 
4886         return true;
4887 }
4888 
4889 /* When incoming ACK allowed to free some skb from write_queue,
4890  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4891  * on the exit from tcp input handler.
4892  *
4893  * PROBLEM: sndbuf expansion does not work well with largesend.
4894  */
4895 static void tcp_new_space(struct sock *sk)
4896 {
4897         struct tcp_sock *tp = tcp_sk(sk);
4898 
4899         if (tcp_should_expand_sndbuf(sk)) {
4900                 tcp_sndbuf_expand(sk);
4901                 tp->snd_cwnd_stamp = tcp_time_stamp;
4902         }
4903 
4904         sk->sk_write_space(sk);
4905 }
4906 
4907 static void tcp_check_space(struct sock *sk)
4908 {
4909         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4910                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4911                 /* pairs with tcp_poll() */
4912                 smp_mb__after_atomic();
4913                 if (sk->sk_socket &&
4914                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4915                         tcp_new_space(sk);
4916         }
4917 }
4918 
4919 static inline void tcp_data_snd_check(struct sock *sk)
4920 {
4921         tcp_push_pending_frames(sk);
4922         tcp_check_space(sk);
4923 }
4924 
4925 /*
4926  * Check if sending an ack is needed.
4927  */
4928 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4929 {
4930         struct tcp_sock *tp = tcp_sk(sk);
4931 
4932             /* More than one full frame received... */
4933         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4934              /* ... and right edge of window advances far enough.
4935               * (tcp_recvmsg() will send ACK otherwise). Or...
4936               */
4937              __tcp_select_window(sk) >= tp->rcv_wnd) ||
4938             /* We ACK each frame or... */
4939             tcp_in_quickack_mode(sk) ||
4940             /* We have out of order data. */
4941             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4942                 /* Then ack it now */
4943                 tcp_send_ack(sk);
4944         } else {
4945                 /* Else, send delayed ack. */
4946                 tcp_send_delayed_ack(sk);
4947         }
4948 }
4949 
4950 static inline void tcp_ack_snd_check(struct sock *sk)
4951 {
4952         if (!inet_csk_ack_scheduled(sk)) {
4953                 /* We sent a data segment already. */
4954                 return;
4955         }
4956         __tcp_ack_snd_check(sk, 1);
4957 }
4958 
4959 /*
4960  *      This routine is only called when we have urgent data
4961  *      signaled. Its the 'slow' part of tcp_urg. It could be
4962  *      moved inline now as tcp_urg is only called from one
4963  *      place. We handle URGent data wrong. We have to - as
4964  *      BSD still doesn't use the correction from RFC961.
4965  *      For 1003.1g we should support a new option TCP_STDURG to permit
4966  *      either form (or just set the sysctl tcp_stdurg).
4967  */
4968 
4969 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4970 {
4971         struct tcp_sock *tp = tcp_sk(sk);
4972         u32 ptr = ntohs(th->urg_ptr);
4973 
4974         if (ptr && !sysctl_tcp_stdurg)
4975                 ptr--;
4976         ptr += ntohl(th->seq);
4977 
4978         /* Ignore urgent data that we've already seen and read. */
4979         if (after(tp->copied_seq, ptr))
4980                 return;
4981 
4982         /* Do not replay urg ptr.
4983          *
4984          * NOTE: interesting situation not covered by specs.
4985          * Misbehaving sender may send urg ptr, pointing to segment,
4986          * which we already have in ofo queue. We are not able to fetch
4987          * such data and will stay in TCP_URG_NOTYET until will be eaten
4988          * by recvmsg(). Seems, we are not obliged to handle such wicked
4989          * situations. But it is worth to think about possibility of some
4990          * DoSes using some hypothetical application level deadlock.
4991          */
4992         if (before(ptr, tp->rcv_nxt))
4993                 return;
4994 
4995         /* Do we already have a newer (or duplicate) urgent pointer? */
4996         if (tp->urg_data && !after(ptr, tp->urg_seq))
4997                 return;
4998 
4999         /* Tell the world about our new urgent pointer. */
5000         sk_send_sigurg(sk);
5001 
5002         /* We may be adding urgent data when the last byte read was
5003          * urgent. To do this requires some care. We cannot just ignore
5004          * tp->copied_seq since we would read the last urgent byte again
5005          * as data, nor can we alter copied_seq until this data arrives
5006          * or we break the semantics of SIOCATMARK (and thus sockatmark())
5007          *
5008          * NOTE. Double Dutch. Rendering to plain English: author of comment
5009          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
5010          * and expect that both A and B disappear from stream. This is _wrong_.
5011          * Though this happens in BSD with high probability, this is occasional.
5012          * Any application relying on this is buggy. Note also, that fix "works"
5013          * only in this artificial test. Insert some normal data between A and B and we will
5014          * decline of BSD again. Verdict: it is better to remove to trap
5015          * buggy users.
5016          */
5017         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5018             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5019                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5020                 tp->copied_seq++;
5021                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5022                         __skb_unlink(skb, &sk->sk_receive_queue);
5023                         __kfree_skb(skb);
5024                 }
5025         }
5026 
5027         tp->urg_data = TCP_URG_NOTYET;
5028         tp->urg_seq = ptr;
5029 
5030         /* Disable header prediction. */
5031         tp->pred_flags = 0;
5032 }
5033 
5034 /* This is the 'fast' part of urgent handling. */
5035 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5036 {
5037         struct tcp_sock *tp = tcp_sk(sk);
5038 
5039         /* Check if we get a new urgent pointer - normally not. */
5040         if (th->urg)
5041                 tcp_check_urg(sk, th);
5042 
5043         /* Do we wait for any urgent data? - normally not... */
5044         if (tp->urg_data == TCP_URG_NOTYET) {
5045                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5046                           th->syn;
5047 
5048                 /* Is the urgent pointer pointing into this packet? */
5049                 if (ptr < skb->len) {
5050                         u8 tmp;
5051                         if (skb_copy_bits(skb, ptr, &tmp, 1))
5052                                 BUG();
5053                         tp->urg_data = TCP_URG_VALID | tmp;
5054                         if (!sock_flag(sk, SOCK_DEAD))
5055                                 sk->sk_data_ready(sk);
5056                 }
5057         }
5058 }
5059 
5060 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5061 {
5062         struct tcp_sock *tp = tcp_sk(sk);
5063         int chunk = skb->len - hlen;
5064         int err;
5065 
5066         local_bh_enable();
5067         if (skb_csum_unnecessary(skb))
5068                 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5069         else
5070                 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5071 
5072         if (!err) {
5073                 tp->ucopy.len -= chunk;
5074                 tp->copied_seq += chunk;
5075                 tcp_rcv_space_adjust(sk);
5076         }
5077 
5078         local_bh_disable();
5079         return err;
5080 }
5081 
5082 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5083                                             struct sk_buff *skb)
5084 {
5085         __sum16 result;
5086 
5087         if (sock_owned_by_user(sk)) {
5088                 local_bh_enable();
5089                 result = __tcp_checksum_complete(skb);
5090                 local_bh_disable();
5091         } else {
5092                 result = __tcp_checksum_complete(skb);
5093         }
5094         return result;
5095 }
5096 
5097 static inline bool tcp_checksum_complete_user(struct sock *sk,
5098                                              struct sk_buff *skb)
5099 {
5100         return !skb_csum_unnecessary(skb) &&
5101                __tcp_checksum_complete_user(sk, skb);
5102 }
5103 
5104 /* Does PAWS and seqno based validation of an incoming segment, flags will
5105  * play significant role here.
5106  */
5107 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5108                                   const struct tcphdr *th, int syn_inerr)
5109 {
5110         struct tcp_sock *tp = tcp_sk(sk);
5111 
5112         /* RFC1323: H1. Apply PAWS check first. */
5113         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5114             tcp_paws_discard(sk, skb)) {
5115                 if (!th->rst) {
5116                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5117                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5118                                                   LINUX_MIB_TCPACKSKIPPEDPAWS,
5119                                                   &tp->last_oow_ack_time))
5120                                 tcp_send_dupack(sk, skb);
5121                         goto discard;
5122                 }
5123                 /* Reset is accepted even if it did not pass PAWS. */
5124         }
5125 
5126         /* Step 1: check sequence number */
5127         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5128                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5129                  * (RST) segments are validated by checking their SEQ-fields."
5130                  * And page 69: "If an incoming segment is not acceptable,
5131                  * an acknowledgment should be sent in reply (unless the RST
5132                  * bit is set, if so drop the segment and return)".
5133                  */
5134                 if (!th->rst) {
5135                         if (th->syn)
5136                                 goto syn_challenge;
5137                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5138                                                   LINUX_MIB_TCPACKSKIPPEDSEQ,
5139                                                   &tp->last_oow_ack_time))
5140                                 tcp_send_dupack(sk, skb);
5141                 }
5142                 goto discard;
5143         }
5144 
5145         /* Step 2: check RST bit */
5146         if (th->rst) {
5147                 /* RFC 5961 3.2 :
5148                  * If sequence number exactly matches RCV.NXT, then
5149                  *     RESET the connection
5150                  * else
5151                  *     Send a challenge ACK
5152                  */
5153                 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5154                         tcp_reset(sk);
5155                 else
5156                         tcp_send_challenge_ack(sk, skb);
5157                 goto discard;
5158         }
5159 
5160         /* step 3: check security and precedence [ignored] */
5161 
5162         /* step 4: Check for a SYN
5163          * RFC 5961 4.2 : Send a challenge ack
5164          */
5165         if (th->syn) {
5166 syn_challenge:
5167                 if (syn_inerr)
5168                         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5169                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5170                 tcp_send_challenge_ack(sk, skb);
5171                 goto discard;
5172         }
5173 
5174         return true;
5175 
5176 discard:
5177         __kfree_skb(skb);
5178         return false;
5179 }
5180 
5181 /*
5182  *      TCP receive function for the ESTABLISHED state.
5183  *
5184  *      It is split into a fast path and a slow path. The fast path is
5185  *      disabled when:
5186  *      - A zero window was announced from us - zero window probing
5187  *        is only handled properly in the slow path.
5188  *      - Out of order segments arrived.
5189  *      - Urgent data is expected.
5190  *      - There is no buffer space left
5191  *      - Unexpected TCP flags/window values/header lengths are received
5192  *        (detected by checking the TCP header against pred_flags)
5193  *      - Data is sent in both directions. Fast path only supports pure senders
5194  *        or pure receivers (this means either the sequence number or the ack
5195  *        value must stay constant)
5196  *      - Unexpected TCP option.
5197  *
5198  *      When these conditions are not satisfied it drops into a standard
5199  *      receive procedure patterned after RFC793 to handle all cases.
5200  *      The first three cases are guaranteed by proper pred_flags setting,
5201  *      the rest is checked inline. Fast processing is turned on in
5202  *      tcp_data_queue when everything is OK.
5203  */
5204 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5205                          const struct tcphdr *th, unsigned int len)
5206 {
5207         struct tcp_sock *tp = tcp_sk(sk);
5208 
5209         if (unlikely(!sk->sk_rx_dst))
5210                 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5211         /*
5212          *      Header prediction.
5213          *      The code loosely follows the one in the famous
5214          *      "30 instruction TCP receive" Van Jacobson mail.
5215          *
5216          *      Van's trick is to deposit buffers into socket queue
5217          *      on a device interrupt, to call tcp_recv function
5218          *      on the receive process context and checksum and copy
5219          *      the buffer to user space. smart...
5220          *
5221          *      Our current scheme is not silly either but we take the
5222          *      extra cost of the net_bh soft interrupt processing...
5223          *      We do checksum and copy also but from device to kernel.
5224          */
5225 
5226         tp->rx_opt.saw_tstamp = 0;
5227 
5228         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5229          *      if header_prediction is to be made
5230          *      'S' will always be tp->tcp_header_len >> 2
5231          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5232          *  turn it off (when there are holes in the receive
5233          *       space for instance)
5234          *      PSH flag is ignored.
5235          */
5236 
5237         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5238             TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5239             !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5240                 int tcp_header_len = tp->tcp_header_len;
5241 
5242                 /* Timestamp header prediction: tcp_header_len
5243                  * is automatically equal to th->doff*4 due to pred_flags
5244                  * match.
5245                  */
5246 
5247                 /* Check timestamp */
5248                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5249                         /* No? Slow path! */
5250                         if (!tcp_parse_aligned_timestamp(tp, th))
5251                                 goto slow_path;
5252 
5253                         /* If PAWS failed, check it more carefully in slow path */
5254                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5255                                 goto slow_path;
5256 
5257                         /* DO NOT update ts_recent here, if checksum fails
5258                          * and timestamp was corrupted part, it will result
5259                          * in a hung connection since we will drop all
5260                          * future packets due to the PAWS test.
5261                          */
5262                 }
5263 
5264                 if (len <= tcp_header_len) {
5265                         /* Bulk data transfer: sender */
5266                         if (len == tcp_header_len) {
5267                                 /* Predicted packet is in window by definition.
5268                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5269                                  * Hence, check seq<=rcv_wup reduces to:
5270                                  */
5271                                 if (tcp_header_len ==
5272                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5273                                     tp->rcv_nxt == tp->rcv_wup)
5274                                         tcp_store_ts_recent(tp);
5275 
5276                                 /* We know that such packets are checksummed
5277                                  * on entry.
5278                                  */
5279                                 tcp_ack(sk, skb, 0);
5280                                 __kfree_skb(skb);
5281                                 tcp_data_snd_check(sk);
5282                                 return;
5283                         } else { /* Header too small */
5284                                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5285                                 goto discard;
5286                         }
5287                 } else {
5288                         int eaten = 0;
5289                         bool fragstolen = false;
5290 
5291                         if (tp->ucopy.task == current &&
5292                             tp->copied_seq == tp->rcv_nxt &&
5293                             len - tcp_header_len <= tp->ucopy.len &&
5294                             sock_owned_by_user(sk)) {
5295                                 __set_current_state(TASK_RUNNING);
5296 
5297                                 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5298                                         /* Predicted packet is in window by definition.
5299                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5300                                          * Hence, check seq<=rcv_wup reduces to:
5301                                          */
5302                                         if (tcp_header_len ==
5303                                             (sizeof(struct tcphdr) +
5304                                              TCPOLEN_TSTAMP_ALIGNED) &&
5305                                             tp->rcv_nxt == tp->rcv_wup)
5306                                                 tcp_store_ts_recent(tp);
5307 
5308                                         tcp_rcv_rtt_measure_ts(sk, skb);
5309 
5310                                         __skb_pull(skb, tcp_header_len);
5311                                         tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5312                                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5313                                         eaten = 1;
5314                                 }
5315                         }
5316                         if (!eaten) {
5317                                 if (tcp_checksum_complete_user(sk, skb))
5318                                         goto csum_error;
5319 
5320                                 if ((int)skb->truesize > sk->sk_forward_alloc)
5321                                         goto step5;
5322 
5323                                 /* Predicted packet is in window by definition.
5324                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5325                                  * Hence, check seq<=rcv_wup reduces to:
5326                                  */
5327                                 if (tcp_header_len ==
5328                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5329                                     tp->rcv_nxt == tp->rcv_wup)
5330                                         tcp_store_ts_recent(tp);
5331 
5332                                 tcp_rcv_rtt_measure_ts(sk, skb);
5333 
5334                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5335 
5336                                 /* Bulk data transfer: receiver */
5337                                 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5338                                                       &fragstolen);
5339                         }
5340 
5341                         tcp_event_data_recv(sk, skb);
5342 
5343                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5344                                 /* Well, only one small jumplet in fast path... */
5345                                 tcp_ack(sk, skb, FLAG_DATA);
5346                                 tcp_data_snd_check(sk);
5347                                 if (!inet_csk_ack_scheduled(sk))
5348                                         goto no_ack;
5349                         }
5350 
5351                         __tcp_ack_snd_check(sk, 0);
5352 no_ack:
5353                         if (eaten)
5354                                 kfree_skb_partial(skb, fragstolen);
5355                         sk->sk_data_ready(sk);
5356                         return;
5357                 }
5358         }
5359 
5360 slow_path:
5361         if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5362                 goto csum_error;
5363 
5364         if (!th->ack && !th->rst && !th->syn)
5365                 goto discard;
5366 
5367         /*
5368          *      Standard slow path.
5369          */
5370 
5371         if (!tcp_validate_incoming(sk, skb, th, 1))
5372                 return;
5373 
5374 step5:
5375         if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5376                 goto discard;
5377 
5378         tcp_rcv_rtt_measure_ts(sk, skb);
5379 
5380         /* Process urgent data. */
5381         tcp_urg(sk, skb, th);
5382 
5383         /* step 7: process the segment text */
5384         tcp_data_queue(sk, skb);
5385 
5386         tcp_data_snd_check(sk);
5387         tcp_ack_snd_check(sk);
5388         return;
5389 
5390 csum_error:
5391         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5392         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5393 
5394 discard:
5395         __kfree_skb(skb);
5396 }
5397 EXPORT_SYMBOL(tcp_rcv_established);
5398 
5399 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5400 {
5401         struct tcp_sock *tp = tcp_sk(sk);
5402         struct inet_connection_sock *icsk = inet_csk(sk);
5403 
5404         tcp_set_state(sk, TCP_ESTABLISHED);
5405 
5406         if (skb) {
5407                 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5408                 security_inet_conn_established(sk, skb);
5409         }
5410 
5411         /* Make sure socket is routed, for correct metrics.  */
5412         icsk->icsk_af_ops->rebuild_header(sk);
5413 
5414         tcp_init_metrics(sk);
5415 
5416         tcp_init_congestion_control(sk);
5417 
5418         /* Prevent spurious tcp_cwnd_restart() on first data
5419          * packet.
5420          */
5421         tp->lsndtime = tcp_time_stamp;
5422 
5423         tcp_init_buffer_space(sk);
5424 
5425         if (sock_flag(sk, SOCK_KEEPOPEN))
5426                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5427 
5428         if (!tp->rx_opt.snd_wscale)
5429                 __tcp_fast_path_on(tp, tp->snd_wnd);
5430         else
5431                 tp->pred_flags = 0;
5432 
5433         if (!sock_flag(sk, SOCK_DEAD)) {
5434                 sk->sk_state_change(sk);
5435                 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5436         }
5437 }
5438 
5439 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5440                                     struct tcp_fastopen_cookie *cookie)
5441 {
5442         struct tcp_sock *tp = tcp_sk(sk);
5443         struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5444         u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5445         bool syn_drop = false;
5446 
5447         if (mss == tp->rx_opt.user_mss) {
5448                 struct tcp_options_received opt;
5449 
5450                 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5451                 tcp_clear_options(&opt);
5452                 opt.user_mss = opt.mss_clamp = 0;
5453                 tcp_parse_options(synack, &opt, 0, NULL);
5454                 mss = opt.mss_clamp;
5455         }
5456 
5457         if (!tp->syn_fastopen) {
5458                 /* Ignore an unsolicited cookie */
5459                 cookie->len = -1;
5460         } else if (tp->total_retrans) {
5461                 /* SYN timed out and the SYN-ACK neither has a cookie nor
5462                  * acknowledges data. Presumably the remote received only
5463                  * the retransmitted (regular) SYNs: either the original
5464                  * SYN-data or the corresponding SYN-ACK was dropped.
5465                  */
5466                 syn_drop = (cookie->len < 0 && data);
5467         } else if (cookie->len < 0 && !tp->syn_data) {
5468                 /* We requested a cookie but didn't get it. If we did not use
5469                  * the (old) exp opt format then try so next time (try_exp=1).
5470                  * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5471                  */
5472                 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5473         }
5474 
5475         tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5476 
5477         if (data) { /* Retransmit unacked data in SYN */
5478                 tcp_for_write_queue_from(data, sk) {
5479                         if (data == tcp_send_head(sk) ||
5480                             __tcp_retransmit_skb(sk, data))
5481                                 break;
5482                 }
5483                 tcp_rearm_rto(sk);
5484                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5485                 return true;
5486         }
5487         tp->syn_data_acked = tp->syn_data;
5488         if (tp->syn_data_acked)
5489                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5490         return false;
5491 }
5492 
5493 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5494                                          const struct tcphdr *th, unsigned int len)
5495 {
5496         struct inet_connection_sock *icsk = inet_csk(sk);
5497         struct tcp_sock *tp = tcp_sk(sk);
5498         struct tcp_fastopen_cookie foc = { .len = -1 };
5499         int saved_clamp = tp->rx_opt.mss_clamp;
5500 
5501         tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5502         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5503                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5504 
5505         if (th->ack) {
5506                 /* rfc793:
5507                  * "If the state is SYN-SENT then
5508                  *    first check the ACK bit
5509                  *      If the ACK bit is set
5510                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5511                  *        a reset (unless the RST bit is set, if so drop
5512                  *        the segment and return)"
5513                  */
5514                 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5515                     after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5516                         goto reset_and_undo;
5517 
5518                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5519                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5520                              tcp_time_stamp)) {
5521                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5522                         goto reset_and_undo;
5523                 }
5524 
5525                 /* Now ACK is acceptable.
5526                  *
5527                  * "If the RST bit is set
5528                  *    If the ACK was acceptable then signal the user "error:
5529                  *    connection reset", drop the segment, enter CLOSED state,
5530                  *    delete TCB, and return."
5531                  */
5532 
5533                 if (th->rst) {
5534                         tcp_reset(sk);
5535                         goto discard;
5536                 }
5537 
5538                 /* rfc793:
5539                  *   "fifth, if neither of the SYN or RST bits is set then
5540                  *    drop the segment and return."
5541                  *
5542                  *    See note below!
5543                  *                                        --ANK(990513)
5544                  */
5545                 if (!th->syn)
5546                         goto discard_and_undo;
5547 
5548                 /* rfc793:
5549                  *   "If the SYN bit is on ...
5550                  *    are acceptable then ...
5551                  *    (our SYN has been ACKed), change the connection
5552                  *    state to ESTABLISHED..."
5553                  */
5554 
5555                 tcp_ecn_rcv_synack(tp, th);
5556 
5557                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5558                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5559 
5560                 /* Ok.. it's good. Set up sequence numbers and
5561                  * move to established.
5562                  */
5563                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5564                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5565 
5566                 /* RFC1323: The window in SYN & SYN/ACK segments is
5567                  * never scaled.
5568                  */
5569                 tp->snd_wnd = ntohs(th->window);
5570 
5571                 if (!tp->rx_opt.wscale_ok) {
5572                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5573                         tp->window_clamp = min(tp->window_clamp, 65535U);
5574                 }
5575 
5576                 if (tp->rx_opt.saw_tstamp) {
5577                         tp->rx_opt.tstamp_ok       = 1;
5578                         tp->tcp_header_len =
5579                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5580                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5581                         tcp_store_ts_recent(tp);
5582                 } else {
5583                         tp->tcp_header_len = sizeof(struct tcphdr);
5584                 }
5585 
5586                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5587                         tcp_enable_fack(tp);
5588 
5589                 tcp_mtup_init(sk);
5590                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5591                 tcp_initialize_rcv_mss(sk);
5592 
5593                 /* Remember, tcp_poll() does not lock socket!
5594                  * Change state from SYN-SENT only after copied_seq
5595                  * is initialized. */
5596                 tp->copied_seq = tp->rcv_nxt;
5597 
5598                 smp_mb();
5599 
5600                 tcp_finish_connect(sk, skb);
5601 
5602                 if ((tp->syn_fastopen || tp->syn_data) &&
5603                     tcp_rcv_fastopen_synack(sk, skb, &foc))
5604                         return -1;
5605 
5606                 if (sk->sk_write_pending ||
5607                     icsk->icsk_accept_queue.rskq_defer_accept ||
5608                     icsk->icsk_ack.pingpong) {
5609                         /* Save one ACK. Data will be ready after
5610                          * several ticks, if write_pending is set.
5611                          *
5612                          * It may be deleted, but with this feature tcpdumps
5613                          * look so _wonderfully_ clever, that I was not able
5614                          * to stand against the temptation 8)     --ANK
5615                          */
5616                         inet_csk_schedule_ack(sk);
5617                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5618                         tcp_enter_quickack_mode(sk);
5619                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5620                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5621 
5622 discard:
5623                         __kfree_skb(skb);
5624                         return 0;
5625                 } else {
5626                         tcp_send_ack(sk);
5627                 }
5628                 return -1;
5629         }
5630 
5631         /* No ACK in the segment */
5632 
5633         if (th->rst) {
5634                 /* rfc793:
5635                  * "If the RST bit is set
5636                  *
5637                  *      Otherwise (no ACK) drop the segment and return."
5638                  */
5639 
5640                 goto discard_and_undo;
5641         }
5642 
5643         /* PAWS check. */
5644         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5645             tcp_paws_reject(&tp->rx_opt, 0))
5646                 goto discard_and_undo;
5647 
5648         if (th->syn) {
5649                 /* We see SYN without ACK. It is attempt of
5650                  * simultaneous connect with crossed SYNs.
5651                  * Particularly, it can be connect to self.
5652                  */
5653                 tcp_set_state(sk, TCP_SYN_RECV);
5654 
5655                 if (tp->rx_opt.saw_tstamp) {
5656                         tp->rx_opt.tstamp_ok = 1;
5657                         tcp_store_ts_recent(tp);
5658                         tp->tcp_header_len =
5659                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5660                 } else {
5661                         tp->tcp_header_len = sizeof(struct tcphdr);
5662                 }
5663 
5664                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5665                 tp->copied_seq = tp->rcv_nxt;
5666                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5667 
5668                 /* RFC1323: The window in SYN & SYN/ACK segments is
5669                  * never scaled.
5670                  */
5671                 tp->snd_wnd    = ntohs(th->window);
5672                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5673                 tp->max_window = tp->snd_wnd;
5674 
5675                 tcp_ecn_rcv_syn(tp, th);
5676 
5677                 tcp_mtup_init(sk);
5678                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5679                 tcp_initialize_rcv_mss(sk);
5680 
5681                 tcp_send_synack(sk);
5682 #if 0
5683                 /* Note, we could accept data and URG from this segment.
5684                  * There are no obstacles to make this (except that we must
5685                  * either change tcp_recvmsg() to prevent it from returning data
5686                  * before 3WHS completes per RFC793, or employ TCP Fast Open).
5687                  *
5688                  * However, if we ignore data in ACKless segments sometimes,
5689                  * we have no reasons to accept it sometimes.
5690                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5691                  * is not flawless. So, discard packet for sanity.
5692                  * Uncomment this return to process the data.
5693                  */
5694                 return -1;
5695 #else
5696                 goto discard;
5697 #endif
5698         }
5699         /* "fifth, if neither of the SYN or RST bits is set then
5700          * drop the segment and return."
5701          */
5702 
5703 discard_and_undo:
5704         tcp_clear_options(&tp->rx_opt);
5705         tp->rx_opt.mss_clamp = saved_clamp;
5706         goto discard;
5707 
5708 reset_and_undo:
5709         tcp_clear_options(&tp->rx_opt);
5710         tp->rx_opt.mss_clamp = saved_clamp;
5711         return 1;
5712 }
5713 
5714 /*
5715  *      This function implements the receiving procedure of RFC 793 for
5716  *      all states except ESTABLISHED and TIME_WAIT.
5717  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5718  *      address independent.
5719  */
5720 
5721 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5722                           const struct tcphdr *th, unsigned int len)
5723 {
5724         struct tcp_sock *tp = tcp_sk(sk);
5725         struct inet_connection_sock *icsk = inet_csk(sk);
5726         struct request_sock *req;
5727         int queued = 0;
5728         bool acceptable;
5729         u32 synack_stamp;
5730 
5731         tp->rx_opt.saw_tstamp = 0;
5732 
5733         switch (sk->sk_state) {
5734         case TCP_CLOSE:
5735                 goto discard;
5736 
5737         case TCP_LISTEN:
5738                 if (th->ack)
5739                         return 1;
5740 
5741                 if (th->rst)
5742                         goto discard;
5743 
5744                 if (th->syn) {
5745                         if (th->fin)
5746                                 goto discard;
5747                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5748                                 return 1;
5749 
5750                         /* Now we have several options: In theory there is
5751                          * nothing else in the frame. KA9Q has an option to
5752                          * send data with the syn, BSD accepts data with the
5753                          * syn up to the [to be] advertised window and
5754                          * Solaris 2.1 gives you a protocol error. For now
5755                          * we just ignore it, that fits the spec precisely
5756                          * and avoids incompatibilities. It would be nice in
5757                          * future to drop through and process the data.
5758                          *
5759                          * Now that TTCP is starting to be used we ought to
5760                          * queue this data.
5761                          * But, this leaves one open to an easy denial of
5762                          * service attack, and SYN cookies can't defend
5763                          * against this problem. So, we drop the data
5764                          * in the interest of security over speed unless
5765                          * it's still in use.
5766                          */
5767                         kfree_skb(skb);
5768                         return 0;
5769                 }
5770                 goto discard;
5771 
5772         case TCP_SYN_SENT:
5773                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5774                 if (queued >= 0)
5775                         return queued;
5776 
5777                 /* Do step6 onward by hand. */
5778                 tcp_urg(sk, skb, th);
5779                 __kfree_skb(skb);
5780                 tcp_data_snd_check(sk);
5781                 return 0;
5782         }
5783 
5784         req = tp->fastopen_rsk;
5785         if (req) {
5786                 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5787                     sk->sk_state != TCP_FIN_WAIT1);
5788 
5789                 if (!tcp_check_req(sk, skb, req, true))
5790                         goto discard;
5791         }
5792 
5793         if (!th->ack && !th->rst && !th->syn)
5794                 goto discard;
5795 
5796         if (!tcp_validate_incoming(sk, skb, th, 0))
5797                 return 0;
5798 
5799         /* step 5: check the ACK field */
5800         acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5801                                       FLAG_UPDATE_TS_RECENT) > 0;
5802 
5803         switch (sk->sk_state) {
5804         case TCP_SYN_RECV:
5805                 if (!acceptable)
5806                         return 1;
5807 
5808                 /* Once we leave TCP_SYN_RECV, we no longer need req
5809                  * so release it.
5810                  */
5811                 if (req) {
5812                         synack_stamp = tcp_rsk(req)->snt_synack;
5813                         tp->total_retrans = req->num_retrans;
5814                         reqsk_fastopen_remove(sk, req, false);
5815                 } else {
5816                         synack_stamp = tp->lsndtime;
5817                         /* Make sure socket is routed, for correct metrics. */
5818                         icsk->icsk_af_ops->rebuild_header(sk);
5819                         tcp_init_congestion_control(sk);
5820 
5821                         tcp_mtup_init(sk);
5822                         tp->copied_seq = tp->rcv_nxt;
5823                         tcp_init_buffer_space(sk);
5824                 }
5825                 smp_mb();
5826                 tcp_set_state(sk, TCP_ESTABLISHED);
5827                 sk->sk_state_change(sk);
5828 
5829                 /* Note, that this wakeup is only for marginal crossed SYN case.
5830                  * Passively open sockets are not waked up, because
5831                  * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5832                  */
5833                 if (sk->sk_socket)
5834                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5835 
5836                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5837                 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5838                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5839                 tcp_synack_rtt_meas(sk, synack_stamp);
5840 
5841                 if (tp->rx_opt.tstamp_ok)
5842                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5843 
5844                 if (req) {
5845                         /* Re-arm the timer because data may have been sent out.
5846                          * This is similar to the regular data transmission case
5847                          * when new data has just been ack'ed.
5848                          *
5849                          * (TFO) - we could try to be more aggressive and
5850                          * retransmitting any data sooner based on when they
5851                          * are sent out.
5852                          */
5853                         tcp_rearm_rto(sk);
5854                 } else
5855                         tcp_init_metrics(sk);
5856 
5857                 tcp_update_pacing_rate(sk);
5858 
5859                 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5860                 tp->lsndtime = tcp_time_stamp;
5861 
5862                 tcp_initialize_rcv_mss(sk);
5863                 tcp_fast_path_on(tp);
5864                 break;
5865 
5866         case TCP_FIN_WAIT1: {
5867                 struct dst_entry *dst;
5868                 int tmo;
5869 
5870                 /* If we enter the TCP_FIN_WAIT1 state and we are a
5871                  * Fast Open socket and this is the first acceptable
5872                  * ACK we have received, this would have acknowledged
5873                  * our SYNACK so stop the SYNACK timer.
5874                  */
5875                 if (req) {
5876                         /* Return RST if ack_seq is invalid.
5877                          * Note that RFC793 only says to generate a
5878                          * DUPACK for it but for TCP Fast Open it seems
5879                          * better to treat this case like TCP_SYN_RECV
5880                          * above.
5881                          */
5882                         if (!acceptable)
5883                                 return 1;
5884                         /* We no longer need the request sock. */
5885                         reqsk_fastopen_remove(sk, req, false);
5886                         tcp_rearm_rto(sk);
5887                 }
5888                 if (tp->snd_una != tp->write_seq)
5889                         break;
5890 
5891                 tcp_set_state(sk, TCP_FIN_WAIT2);
5892                 sk->sk_shutdown |= SEND_SHUTDOWN;
5893 
5894                 dst = __sk_dst_get(sk);
5895                 if (dst)
5896                         dst_confirm(dst);
5897 
5898                 if (!sock_flag(sk, SOCK_DEAD)) {
5899                         /* Wake up lingering close() */
5900                         sk->sk_state_change(sk);
5901                         break;
5902                 }
5903 
5904                 if (tp->linger2 < 0 ||
5905                     (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5906                      after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5907                         tcp_done(sk);
5908                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5909                         return 1;
5910                 }
5911 
5912                 tmo = tcp_fin_time(sk);
5913                 if (tmo > TCP_TIMEWAIT_LEN) {
5914                         inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5915                 } else if (th->fin || sock_owned_by_user(sk)) {
5916                         /* Bad case. We could lose such FIN otherwise.
5917                          * It is not a big problem, but it looks confusing
5918                          * and not so rare event. We still can lose it now,
5919                          * if it spins in bh_lock_sock(), but it is really
5920                          * marginal case.
5921                          */
5922                         inet_csk_reset_keepalive_timer(sk, tmo);
5923                 } else {
5924                         tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5925                         goto discard;
5926                 }
5927                 break;
5928         }
5929 
5930         case TCP_CLOSING:
5931                 if (tp->snd_una == tp->write_seq) {
5932                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5933                         goto discard;
5934                 }
5935                 break;
5936 
5937         case TCP_LAST_ACK:
5938                 if (tp->snd_una == tp->write_seq) {
5939                         tcp_update_metrics(sk);
5940                         tcp_done(sk);
5941                         goto discard;
5942                 }
5943                 break;
5944         }
5945 
5946         /* step 6: check the URG bit */
5947         tcp_urg(sk, skb, th);
5948 
5949         /* step 7: process the segment text */
5950         switch (sk->sk_state) {
5951         case TCP_CLOSE_WAIT:
5952         case TCP_CLOSING:
5953         case TCP_LAST_ACK:
5954                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5955                         break;
5956         case TCP_FIN_WAIT1:
5957         case TCP_FIN_WAIT2:
5958                 /* RFC 793 says to queue data in these states,
5959                  * RFC 1122 says we MUST send a reset.
5960                  * BSD 4.4 also does reset.
5961                  */
5962                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5963                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5964                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5965                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5966                                 tcp_reset(sk);
5967                                 return 1;
5968                         }
5969                 }
5970                 /* Fall through */
5971         case TCP_ESTABLISHED:
5972                 tcp_data_queue(sk, skb);
5973                 queued = 1;
5974                 break;
5975         }
5976 
5977         /* tcp_data could move socket to TIME-WAIT */
5978         if (sk->sk_state != TCP_CLOSE) {
5979                 tcp_data_snd_check(sk);
5980                 tcp_ack_snd_check(sk);
5981         }
5982 
5983         if (!queued) {
5984 discard:
5985                 __kfree_skb(skb);
5986         }
5987         return 0;
5988 }
5989 EXPORT_SYMBOL(tcp_rcv_state_process);
5990 
5991 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
5992 {
5993         struct inet_request_sock *ireq = inet_rsk(req);
5994 
5995         if (family == AF_INET)
5996                 net_dbg_ratelimited("drop open request from %pI4/%u\n",
5997                                     &ireq->ir_rmt_addr, port);
5998 #if IS_ENABLED(CONFIG_IPV6)
5999         else if (family == AF_INET6)
6000                 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6001                                     &ireq->ir_v6_rmt_addr, port);
6002 #endif
6003 }
6004 
6005 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6006  *
6007  * If we receive a SYN packet with these bits set, it means a
6008  * network is playing bad games with TOS bits. In order to
6009  * avoid possible false congestion notifications, we disable
6010  * TCP ECN negotiation.
6011  *
6012  * Exception: tcp_ca wants ECN. This is required for DCTCP
6013  * congestion control: Linux DCTCP asserts ECT on all packets,
6014  * including SYN, which is most optimal solution; however,
6015  * others, such as FreeBSD do not.
6016  */
6017 static void tcp_ecn_create_request(struct request_sock *req,
6018                                    const struct sk_buff *skb,
6019                                    const struct sock *listen_sk,
6020                                    const struct dst_entry *dst)
6021 {
6022         const struct tcphdr *th = tcp_hdr(skb);
6023         const struct net *net = sock_net(listen_sk);
6024         bool th_ecn = th->ece && th->cwr;
6025         bool ect, ecn_ok;
6026         u32 ecn_ok_dst;
6027 
6028         if (!th_ecn)
6029                 return;
6030 
6031         ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6032         ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6033         ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6034 
6035         if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6036             (ecn_ok_dst & DST_FEATURE_ECN_CA))
6037                 inet_rsk(req)->ecn_ok = 1;
6038 }
6039 
6040 static void tcp_openreq_init(struct request_sock *req,
6041                              const struct tcp_options_received *rx_opt,
6042                              struct sk_buff *skb, const struct sock *sk)
6043 {
6044         struct inet_request_sock *ireq = inet_rsk(req);
6045 
6046         req->rcv_wnd = 0;               /* So that tcp_send_synack() knows! */
6047         req->cookie_ts = 0;
6048         tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6049         tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6050         tcp_rsk(req)->snt_synack = tcp_time_stamp;
6051         tcp_rsk(req)->last_oow_ack_time = 0;
6052         req->mss = rx_opt->mss_clamp;
6053         req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6054         ireq->tstamp_ok = rx_opt->tstamp_ok;
6055         ireq->sack_ok = rx_opt->sack_ok;
6056         ireq->snd_wscale = rx_opt->snd_wscale;
6057         ireq->wscale_ok = rx_opt->wscale_ok;
6058         ireq->acked = 0;
6059         ireq->ecn_ok = 0;
6060         ireq->ir_rmt_port = tcp_hdr(skb)->source;
6061         ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6062         ireq->ir_mark = inet_request_mark(sk, skb);
6063 }
6064 
6065 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6066                                       struct sock *sk_listener)
6067 {
6068         struct request_sock *req = reqsk_alloc(ops, sk_listener);
6069 
6070         if (req) {
6071                 struct inet_request_sock *ireq = inet_rsk(req);
6072 
6073                 kmemcheck_annotate_bitfield(ireq, flags);
6074                 ireq->opt = NULL;
6075                 atomic64_set(&ireq->ir_cookie, 0);
6076                 ireq->ireq_state = TCP_NEW_SYN_RECV;
6077                 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6078                 ireq->ireq_family = sk_listener->sk_family;
6079         }
6080 
6081         return req;
6082 }
6083 EXPORT_SYMBOL(inet_reqsk_alloc);
6084 
6085 /*
6086  * Return true if a syncookie should be sent
6087  */
6088 static bool tcp_syn_flood_action(struct sock *sk,
6089                                  const struct sk_buff *skb,
6090                                  const char *proto)
6091 {
6092         const char *msg = "Dropping request";
6093         bool want_cookie = false;
6094         struct listen_sock *lopt;
6095 
6096 #ifdef CONFIG_SYN_COOKIES
6097         if (sysctl_tcp_syncookies) {
6098                 msg = "Sending cookies";
6099                 want_cookie = true;
6100                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6101         } else
6102 #endif
6103                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6104 
6105         lopt = inet_csk(sk)->icsk_accept_queue.listen_opt;
6106         if (!lopt->synflood_warned && sysctl_tcp_syncookies != 2) {
6107                 lopt->synflood_warned = 1;
6108                 pr_info("%s: Possible SYN flooding on port %d. %s.  Check SNMP counters.\n",
6109                         proto,