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

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

Version: ~ [ linux-5.18-rc6 ] ~ [ linux-5.17.6 ] ~ [ linux-5.16.20 ] ~ [ linux-5.15.38 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.114 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.192 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.241 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.277 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.312 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.302 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.9 ] ~ [ 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 = 1000;
 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_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
113 #define FLAG_ORIG_SACK_ACKED    0x200 /* Never retransmitted data are (s)acked  */
114 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
115 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
116 #define FLAG_SACK_RENEGING      0x2000 /* snd_una advanced to a sacked seq */
117 #define FLAG_UPDATE_TS_RECENT   0x4000 /* tcp_replace_ts_recent() */
118 
119 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
120 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
121 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
122 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
123 
124 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
125 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
126 
127 /* Adapt the MSS value used to make delayed ack decision to the
128  * real world.
129  */
130 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
131 {
132         struct inet_connection_sock *icsk = inet_csk(sk);
133         const unsigned int lss = icsk->icsk_ack.last_seg_size;
134         unsigned int len;
135 
136         icsk->icsk_ack.last_seg_size = 0;
137 
138         /* skb->len may jitter because of SACKs, even if peer
139          * sends good full-sized frames.
140          */
141         len = skb_shinfo(skb)->gso_size ? : skb->len;
142         if (len >= icsk->icsk_ack.rcv_mss) {
143                 icsk->icsk_ack.rcv_mss = len;
144         } else {
145                 /* Otherwise, we make more careful check taking into account,
146                  * that SACKs block is variable.
147                  *
148                  * "len" is invariant segment length, including TCP header.
149                  */
150                 len += skb->data - skb_transport_header(skb);
151                 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
152                     /* If PSH is not set, packet should be
153                      * full sized, provided peer TCP is not badly broken.
154                      * This observation (if it is correct 8)) allows
155                      * to handle super-low mtu links fairly.
156                      */
157                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
158                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
159                         /* Subtract also invariant (if peer is RFC compliant),
160                          * tcp header plus fixed timestamp option length.
161                          * Resulting "len" is MSS free of SACK jitter.
162                          */
163                         len -= tcp_sk(sk)->tcp_header_len;
164                         icsk->icsk_ack.last_seg_size = len;
165                         if (len == lss) {
166                                 icsk->icsk_ack.rcv_mss = len;
167                                 return;
168                         }
169                 }
170                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
171                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
172                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
173         }
174 }
175 
176 static void tcp_incr_quickack(struct sock *sk)
177 {
178         struct inet_connection_sock *icsk = inet_csk(sk);
179         unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
180 
181         if (quickacks == 0)
182                 quickacks = 2;
183         if (quickacks > icsk->icsk_ack.quick)
184                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
185 }
186 
187 static void tcp_enter_quickack_mode(struct sock *sk)
188 {
189         struct inet_connection_sock *icsk = inet_csk(sk);
190         tcp_incr_quickack(sk);
191         icsk->icsk_ack.pingpong = 0;
192         icsk->icsk_ack.ato = TCP_ATO_MIN;
193 }
194 
195 /* Send ACKs quickly, if "quick" count is not exhausted
196  * and the session is not interactive.
197  */
198 
199 static inline bool tcp_in_quickack_mode(const struct sock *sk)
200 {
201         const struct inet_connection_sock *icsk = inet_csk(sk);
202 
203         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
204 }
205 
206 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
207 {
208         if (tp->ecn_flags & TCP_ECN_OK)
209                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
210 }
211 
212 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
213 {
214         if (tcp_hdr(skb)->cwr)
215                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
216 }
217 
218 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
219 {
220         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
221 }
222 
223 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
224 {
225         switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
226         case INET_ECN_NOT_ECT:
227                 /* Funny extension: if ECT is not set on a segment,
228                  * and we already seen ECT on a previous segment,
229                  * it is probably a retransmit.
230                  */
231                 if (tp->ecn_flags & TCP_ECN_SEEN)
232                         tcp_enter_quickack_mode((struct sock *)tp);
233                 break;
234         case INET_ECN_CE:
235                 if (tcp_ca_needs_ecn((struct sock *)tp))
236                         tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
237 
238                 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
239                         /* Better not delay acks, sender can have a very low cwnd */
240                         tcp_enter_quickack_mode((struct sock *)tp);
241                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
242                 }
243                 tp->ecn_flags |= TCP_ECN_SEEN;
244                 break;
245         default:
246                 if (tcp_ca_needs_ecn((struct sock *)tp))
247                         tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
248                 tp->ecn_flags |= TCP_ECN_SEEN;
249                 break;
250         }
251 }
252 
253 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
254 {
255         if (tp->ecn_flags & TCP_ECN_OK)
256                 __tcp_ecn_check_ce(tp, skb);
257 }
258 
259 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
260 {
261         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
262                 tp->ecn_flags &= ~TCP_ECN_OK;
263 }
264 
265 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
266 {
267         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
268                 tp->ecn_flags &= ~TCP_ECN_OK;
269 }
270 
271 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
272 {
273         if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
274                 return true;
275         return false;
276 }
277 
278 /* Buffer size and advertised window tuning.
279  *
280  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
281  */
282 
283 static void tcp_sndbuf_expand(struct sock *sk)
284 {
285         const struct tcp_sock *tp = tcp_sk(sk);
286         int sndmem, per_mss;
287         u32 nr_segs;
288 
289         /* Worst case is non GSO/TSO : each frame consumes one skb
290          * and skb->head is kmalloced using power of two area of memory
291          */
292         per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
293                   MAX_TCP_HEADER +
294                   SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
295 
296         per_mss = roundup_pow_of_two(per_mss) +
297                   SKB_DATA_ALIGN(sizeof(struct sk_buff));
298 
299         nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
300         nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
301 
302         /* Fast Recovery (RFC 5681 3.2) :
303          * Cubic needs 1.7 factor, rounded to 2 to include
304          * extra cushion (application might react slowly to POLLOUT)
305          */
306         sndmem = 2 * nr_segs * per_mss;
307 
308         if (sk->sk_sndbuf < sndmem)
309                 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
310 }
311 
312 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
313  *
314  * All tcp_full_space() is split to two parts: "network" buffer, allocated
315  * forward and advertised in receiver window (tp->rcv_wnd) and
316  * "application buffer", required to isolate scheduling/application
317  * latencies from network.
318  * window_clamp is maximal advertised window. It can be less than
319  * tcp_full_space(), in this case tcp_full_space() - window_clamp
320  * is reserved for "application" buffer. The less window_clamp is
321  * the smoother our behaviour from viewpoint of network, but the lower
322  * throughput and the higher sensitivity of the connection to losses. 8)
323  *
324  * rcv_ssthresh is more strict window_clamp used at "slow start"
325  * phase to predict further behaviour of this connection.
326  * It is used for two goals:
327  * - to enforce header prediction at sender, even when application
328  *   requires some significant "application buffer". It is check #1.
329  * - to prevent pruning of receive queue because of misprediction
330  *   of receiver window. Check #2.
331  *
332  * The scheme does not work when sender sends good segments opening
333  * window and then starts to feed us spaghetti. But it should work
334  * in common situations. Otherwise, we have to rely on queue collapsing.
335  */
336 
337 /* Slow part of check#2. */
338 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
339 {
340         struct tcp_sock *tp = tcp_sk(sk);
341         /* Optimize this! */
342         int truesize = tcp_win_from_space(skb->truesize) >> 1;
343         int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
344 
345         while (tp->rcv_ssthresh <= window) {
346                 if (truesize <= skb->len)
347                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
348 
349                 truesize >>= 1;
350                 window >>= 1;
351         }
352         return 0;
353 }
354 
355 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
356 {
357         struct tcp_sock *tp = tcp_sk(sk);
358 
359         /* Check #1 */
360         if (tp->rcv_ssthresh < tp->window_clamp &&
361             (int)tp->rcv_ssthresh < tcp_space(sk) &&
362             !sk_under_memory_pressure(sk)) {
363                 int incr;
364 
365                 /* Check #2. Increase window, if skb with such overhead
366                  * will fit to rcvbuf in future.
367                  */
368                 if (tcp_win_from_space(skb->truesize) <= skb->len)
369                         incr = 2 * tp->advmss;
370                 else
371                         incr = __tcp_grow_window(sk, skb);
372 
373                 if (incr) {
374                         incr = max_t(int, incr, 2 * skb->len);
375                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
376                                                tp->window_clamp);
377                         inet_csk(sk)->icsk_ack.quick |= 1;
378                 }
379         }
380 }
381 
382 /* 3. Tuning rcvbuf, when connection enters established state. */
383 static void tcp_fixup_rcvbuf(struct sock *sk)
384 {
385         u32 mss = tcp_sk(sk)->advmss;
386         int rcvmem;
387 
388         rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
389                  tcp_default_init_rwnd(mss);
390 
391         /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
392          * Allow enough cushion so that sender is not limited by our window
393          */
394         if (sysctl_tcp_moderate_rcvbuf)
395                 rcvmem <<= 2;
396 
397         if (sk->sk_rcvbuf < rcvmem)
398                 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
399 }
400 
401 /* 4. Try to fixup all. It is made immediately after connection enters
402  *    established state.
403  */
404 void tcp_init_buffer_space(struct sock *sk)
405 {
406         struct tcp_sock *tp = tcp_sk(sk);
407         int maxwin;
408 
409         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
410                 tcp_fixup_rcvbuf(sk);
411         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
412                 tcp_sndbuf_expand(sk);
413 
414         tp->rcvq_space.space = tp->rcv_wnd;
415         tp->rcvq_space.time = tcp_time_stamp;
416         tp->rcvq_space.seq = tp->copied_seq;
417 
418         maxwin = tcp_full_space(sk);
419 
420         if (tp->window_clamp >= maxwin) {
421                 tp->window_clamp = maxwin;
422 
423                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
424                         tp->window_clamp = max(maxwin -
425                                                (maxwin >> sysctl_tcp_app_win),
426                                                4 * tp->advmss);
427         }
428 
429         /* Force reservation of one segment. */
430         if (sysctl_tcp_app_win &&
431             tp->window_clamp > 2 * tp->advmss &&
432             tp->window_clamp + tp->advmss > maxwin)
433                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
434 
435         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
436         tp->snd_cwnd_stamp = tcp_time_stamp;
437 }
438 
439 /* 5. Recalculate window clamp after socket hit its memory bounds. */
440 static void tcp_clamp_window(struct sock *sk)
441 {
442         struct tcp_sock *tp = tcp_sk(sk);
443         struct inet_connection_sock *icsk = inet_csk(sk);
444 
445         icsk->icsk_ack.quick = 0;
446 
447         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
448             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
449             !sk_under_memory_pressure(sk) &&
450             sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
451                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
452                                     sysctl_tcp_rmem[2]);
453         }
454         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
455                 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
456 }
457 
458 /* Initialize RCV_MSS value.
459  * RCV_MSS is an our guess about MSS used by the peer.
460  * We haven't any direct information about the MSS.
461  * It's better to underestimate the RCV_MSS rather than overestimate.
462  * Overestimations make us ACKing less frequently than needed.
463  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
464  */
465 void tcp_initialize_rcv_mss(struct sock *sk)
466 {
467         const struct tcp_sock *tp = tcp_sk(sk);
468         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
469 
470         hint = min(hint, tp->rcv_wnd / 2);
471         hint = min(hint, TCP_MSS_DEFAULT);
472         hint = max(hint, TCP_MIN_MSS);
473 
474         inet_csk(sk)->icsk_ack.rcv_mss = hint;
475 }
476 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
477 
478 /* Receiver "autotuning" code.
479  *
480  * The algorithm for RTT estimation w/o timestamps is based on
481  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
482  * <http://public.lanl.gov/radiant/pubs.html#DRS>
483  *
484  * More detail on this code can be found at
485  * <http://staff.psc.edu/jheffner/>,
486  * though this reference is out of date.  A new paper
487  * is pending.
488  */
489 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
490 {
491         u32 new_sample = tp->rcv_rtt_est.rtt;
492         long m = sample;
493 
494         if (m == 0)
495                 m = 1;
496 
497         if (new_sample != 0) {
498                 /* If we sample in larger samples in the non-timestamp
499                  * case, we could grossly overestimate the RTT especially
500                  * with chatty applications or bulk transfer apps which
501                  * are stalled on filesystem I/O.
502                  *
503                  * Also, since we are only going for a minimum in the
504                  * non-timestamp case, we do not smooth things out
505                  * else with timestamps disabled convergence takes too
506                  * long.
507                  */
508                 if (!win_dep) {
509                         m -= (new_sample >> 3);
510                         new_sample += m;
511                 } else {
512                         m <<= 3;
513                         if (m < new_sample)
514                                 new_sample = m;
515                 }
516         } else {
517                 /* No previous measure. */
518                 new_sample = m << 3;
519         }
520 
521         if (tp->rcv_rtt_est.rtt != new_sample)
522                 tp->rcv_rtt_est.rtt = new_sample;
523 }
524 
525 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
526 {
527         if (tp->rcv_rtt_est.time == 0)
528                 goto new_measure;
529         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
530                 return;
531         tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
532 
533 new_measure:
534         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
535         tp->rcv_rtt_est.time = tcp_time_stamp;
536 }
537 
538 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
539                                           const struct sk_buff *skb)
540 {
541         struct tcp_sock *tp = tcp_sk(sk);
542         if (tp->rx_opt.rcv_tsecr &&
543             (TCP_SKB_CB(skb)->end_seq -
544              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
545                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
546 }
547 
548 /*
549  * This function should be called every time data is copied to user space.
550  * It calculates the appropriate TCP receive buffer space.
551  */
552 void tcp_rcv_space_adjust(struct sock *sk)
553 {
554         struct tcp_sock *tp = tcp_sk(sk);
555         int time;
556         int copied;
557 
558         time = tcp_time_stamp - tp->rcvq_space.time;
559         if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
560                 return;
561 
562         /* Number of bytes copied to user in last RTT */
563         copied = tp->copied_seq - tp->rcvq_space.seq;
564         if (copied <= tp->rcvq_space.space)
565                 goto new_measure;
566 
567         /* A bit of theory :
568          * copied = bytes received in previous RTT, our base window
569          * To cope with packet losses, we need a 2x factor
570          * To cope with slow start, and sender growing its cwin by 100 %
571          * every RTT, we need a 4x factor, because the ACK we are sending
572          * now is for the next RTT, not the current one :
573          * <prev RTT . ><current RTT .. ><next RTT .... >
574          */
575 
576         if (sysctl_tcp_moderate_rcvbuf &&
577             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
578                 int rcvwin, rcvmem, rcvbuf;
579 
580                 /* minimal window to cope with packet losses, assuming
581                  * steady state. Add some cushion because of small variations.
582                  */
583                 rcvwin = (copied << 1) + 16 * tp->advmss;
584 
585                 /* If rate increased by 25%,
586                  *      assume slow start, rcvwin = 3 * copied
587                  * If rate increased by 50%,
588                  *      assume sender can use 2x growth, rcvwin = 4 * copied
589                  */
590                 if (copied >=
591                     tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
592                         if (copied >=
593                             tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
594                                 rcvwin <<= 1;
595                         else
596                                 rcvwin += (rcvwin >> 1);
597                 }
598 
599                 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
600                 while (tcp_win_from_space(rcvmem) < tp->advmss)
601                         rcvmem += 128;
602 
603                 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
604                 if (rcvbuf > sk->sk_rcvbuf) {
605                         sk->sk_rcvbuf = rcvbuf;
606 
607                         /* Make the window clamp follow along.  */
608                         tp->window_clamp = rcvwin;
609                 }
610         }
611         tp->rcvq_space.space = copied;
612 
613 new_measure:
614         tp->rcvq_space.seq = tp->copied_seq;
615         tp->rcvq_space.time = tcp_time_stamp;
616 }
617 
618 /* There is something which you must keep in mind when you analyze the
619  * behavior of the tp->ato delayed ack timeout interval.  When a
620  * connection starts up, we want to ack as quickly as possible.  The
621  * problem is that "good" TCP's do slow start at the beginning of data
622  * transmission.  The means that until we send the first few ACK's the
623  * sender will sit on his end and only queue most of his data, because
624  * he can only send snd_cwnd unacked packets at any given time.  For
625  * each ACK we send, he increments snd_cwnd and transmits more of his
626  * queue.  -DaveM
627  */
628 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
629 {
630         struct tcp_sock *tp = tcp_sk(sk);
631         struct inet_connection_sock *icsk = inet_csk(sk);
632         u32 now;
633 
634         inet_csk_schedule_ack(sk);
635 
636         tcp_measure_rcv_mss(sk, skb);
637 
638         tcp_rcv_rtt_measure(tp);
639 
640         now = tcp_time_stamp;
641 
642         if (!icsk->icsk_ack.ato) {
643                 /* The _first_ data packet received, initialize
644                  * delayed ACK engine.
645                  */
646                 tcp_incr_quickack(sk);
647                 icsk->icsk_ack.ato = TCP_ATO_MIN;
648         } else {
649                 int m = now - icsk->icsk_ack.lrcvtime;
650 
651                 if (m <= TCP_ATO_MIN / 2) {
652                         /* The fastest case is the first. */
653                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
654                 } else if (m < icsk->icsk_ack.ato) {
655                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
656                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
657                                 icsk->icsk_ack.ato = icsk->icsk_rto;
658                 } else if (m > icsk->icsk_rto) {
659                         /* Too long gap. Apparently sender failed to
660                          * restart window, so that we send ACKs quickly.
661                          */
662                         tcp_incr_quickack(sk);
663                         sk_mem_reclaim(sk);
664                 }
665         }
666         icsk->icsk_ack.lrcvtime = now;
667 
668         tcp_ecn_check_ce(tp, skb);
669 
670         if (skb->len >= 128)
671                 tcp_grow_window(sk, skb);
672 }
673 
674 /* Called to compute a smoothed rtt estimate. The data fed to this
675  * routine either comes from timestamps, or from segments that were
676  * known _not_ to have been retransmitted [see Karn/Partridge
677  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
678  * piece by Van Jacobson.
679  * NOTE: the next three routines used to be one big routine.
680  * To save cycles in the RFC 1323 implementation it was better to break
681  * it up into three procedures. -- erics
682  */
683 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
684 {
685         struct tcp_sock *tp = tcp_sk(sk);
686         long m = mrtt_us; /* RTT */
687         u32 srtt = tp->srtt_us;
688 
689         /*      The following amusing code comes from Jacobson's
690          *      article in SIGCOMM '88.  Note that rtt and mdev
691          *      are scaled versions of rtt and mean deviation.
692          *      This is designed to be as fast as possible
693          *      m stands for "measurement".
694          *
695          *      On a 1990 paper the rto value is changed to:
696          *      RTO = rtt + 4 * mdev
697          *
698          * Funny. This algorithm seems to be very broken.
699          * These formulae increase RTO, when it should be decreased, increase
700          * too slowly, when it should be increased quickly, decrease too quickly
701          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
702          * does not matter how to _calculate_ it. Seems, it was trap
703          * that VJ failed to avoid. 8)
704          */
705         if (srtt != 0) {
706                 m -= (srtt >> 3);       /* m is now error in rtt est */
707                 srtt += m;              /* rtt = 7/8 rtt + 1/8 new */
708                 if (m < 0) {
709                         m = -m;         /* m is now abs(error) */
710                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
711                         /* This is similar to one of Eifel findings.
712                          * Eifel blocks mdev updates when rtt decreases.
713                          * This solution is a bit different: we use finer gain
714                          * for mdev in this case (alpha*beta).
715                          * Like Eifel it also prevents growth of rto,
716                          * but also it limits too fast rto decreases,
717                          * happening in pure Eifel.
718                          */
719                         if (m > 0)
720                                 m >>= 3;
721                 } else {
722                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
723                 }
724                 tp->mdev_us += m;               /* mdev = 3/4 mdev + 1/4 new */
725                 if (tp->mdev_us > tp->mdev_max_us) {
726                         tp->mdev_max_us = tp->mdev_us;
727                         if (tp->mdev_max_us > tp->rttvar_us)
728                                 tp->rttvar_us = tp->mdev_max_us;
729                 }
730                 if (after(tp->snd_una, tp->rtt_seq)) {
731                         if (tp->mdev_max_us < tp->rttvar_us)
732                                 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
733                         tp->rtt_seq = tp->snd_nxt;
734                         tp->mdev_max_us = tcp_rto_min_us(sk);
735                 }
736         } else {
737                 /* no previous measure. */
738                 srtt = m << 3;          /* take the measured time to be rtt */
739                 tp->mdev_us = m << 1;   /* make sure rto = 3*rtt */
740                 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
741                 tp->mdev_max_us = tp->rttvar_us;
742                 tp->rtt_seq = tp->snd_nxt;
743         }
744         tp->srtt_us = max(1U, srtt);
745 }
746 
747 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
748  * Note: TCP stack does not yet implement pacing.
749  * FQ packet scheduler can be used to implement cheap but effective
750  * TCP pacing, to smooth the burst on large writes when packets
751  * in flight is significantly lower than cwnd (or rwin)
752  */
753 static void tcp_update_pacing_rate(struct sock *sk)
754 {
755         const struct tcp_sock *tp = tcp_sk(sk);
756         u64 rate;
757 
758         /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
759         rate = (u64)tp->mss_cache * 2 * (USEC_PER_SEC << 3);
760 
761         rate *= max(tp->snd_cwnd, tp->packets_out);
762 
763         if (likely(tp->srtt_us))
764                 do_div(rate, tp->srtt_us);
765 
766         /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
767          * without any lock. We want to make sure compiler wont store
768          * intermediate values in this location.
769          */
770         ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
771                                                 sk->sk_max_pacing_rate);
772 }
773 
774 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
775  * routine referred to above.
776  */
777 static void tcp_set_rto(struct sock *sk)
778 {
779         const struct tcp_sock *tp = tcp_sk(sk);
780         /* Old crap is replaced with new one. 8)
781          *
782          * More seriously:
783          * 1. If rtt variance happened to be less 50msec, it is hallucination.
784          *    It cannot be less due to utterly erratic ACK generation made
785          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
786          *    to do with delayed acks, because at cwnd>2 true delack timeout
787          *    is invisible. Actually, Linux-2.4 also generates erratic
788          *    ACKs in some circumstances.
789          */
790         inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
791 
792         /* 2. Fixups made earlier cannot be right.
793          *    If we do not estimate RTO correctly without them,
794          *    all the algo is pure shit and should be replaced
795          *    with correct one. It is exactly, which we pretend to do.
796          */
797 
798         /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
799          * guarantees that rto is higher.
800          */
801         tcp_bound_rto(sk);
802 }
803 
804 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
805 {
806         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
807 
808         if (!cwnd)
809                 cwnd = TCP_INIT_CWND;
810         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
811 }
812 
813 /*
814  * Packet counting of FACK is based on in-order assumptions, therefore TCP
815  * disables it when reordering is detected
816  */
817 void tcp_disable_fack(struct tcp_sock *tp)
818 {
819         /* RFC3517 uses different metric in lost marker => reset on change */
820         if (tcp_is_fack(tp))
821                 tp->lost_skb_hint = NULL;
822         tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
823 }
824 
825 /* Take a notice that peer is sending D-SACKs */
826 static void tcp_dsack_seen(struct tcp_sock *tp)
827 {
828         tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
829 }
830 
831 static void tcp_update_reordering(struct sock *sk, const int metric,
832                                   const int ts)
833 {
834         struct tcp_sock *tp = tcp_sk(sk);
835         if (metric > tp->reordering) {
836                 int mib_idx;
837 
838                 tp->reordering = min(sysctl_tcp_max_reordering, metric);
839 
840                 /* This exciting event is worth to be remembered. 8) */
841                 if (ts)
842                         mib_idx = LINUX_MIB_TCPTSREORDER;
843                 else if (tcp_is_reno(tp))
844                         mib_idx = LINUX_MIB_TCPRENOREORDER;
845                 else if (tcp_is_fack(tp))
846                         mib_idx = LINUX_MIB_TCPFACKREORDER;
847                 else
848                         mib_idx = LINUX_MIB_TCPSACKREORDER;
849 
850                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
851 #if FASTRETRANS_DEBUG > 1
852                 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
853                          tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
854                          tp->reordering,
855                          tp->fackets_out,
856                          tp->sacked_out,
857                          tp->undo_marker ? tp->undo_retrans : 0);
858 #endif
859                 tcp_disable_fack(tp);
860         }
861 
862         if (metric > 0)
863                 tcp_disable_early_retrans(tp);
864 }
865 
866 /* This must be called before lost_out is incremented */
867 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
868 {
869         if (!tp->retransmit_skb_hint ||
870             before(TCP_SKB_CB(skb)->seq,
871                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
872                 tp->retransmit_skb_hint = skb;
873 
874         if (!tp->lost_out ||
875             after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
876                 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
877 }
878 
879 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
880 {
881         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
882                 tcp_verify_retransmit_hint(tp, skb);
883 
884                 tp->lost_out += tcp_skb_pcount(skb);
885                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
886         }
887 }
888 
889 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
890                                             struct sk_buff *skb)
891 {
892         tcp_verify_retransmit_hint(tp, skb);
893 
894         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
895                 tp->lost_out += tcp_skb_pcount(skb);
896                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
897         }
898 }
899 
900 /* This procedure tags the retransmission queue when SACKs arrive.
901  *
902  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
903  * Packets in queue with these bits set are counted in variables
904  * sacked_out, retrans_out and lost_out, correspondingly.
905  *
906  * Valid combinations are:
907  * Tag  InFlight        Description
908  * 0    1               - orig segment is in flight.
909  * S    0               - nothing flies, orig reached receiver.
910  * L    0               - nothing flies, orig lost by net.
911  * R    2               - both orig and retransmit are in flight.
912  * L|R  1               - orig is lost, retransmit is in flight.
913  * S|R  1               - orig reached receiver, retrans is still in flight.
914  * (L|S|R is logically valid, it could occur when L|R is sacked,
915  *  but it is equivalent to plain S and code short-curcuits it to S.
916  *  L|S is logically invalid, it would mean -1 packet in flight 8))
917  *
918  * These 6 states form finite state machine, controlled by the following events:
919  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
920  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
921  * 3. Loss detection event of two flavors:
922  *      A. Scoreboard estimator decided the packet is lost.
923  *         A'. Reno "three dupacks" marks head of queue lost.
924  *         A''. Its FACK modification, head until snd.fack is lost.
925  *      B. SACK arrives sacking SND.NXT at the moment, when the
926  *         segment was retransmitted.
927  * 4. D-SACK added new rule: D-SACK changes any tag to S.
928  *
929  * It is pleasant to note, that state diagram turns out to be commutative,
930  * so that we are allowed not to be bothered by order of our actions,
931  * when multiple events arrive simultaneously. (see the function below).
932  *
933  * Reordering detection.
934  * --------------------
935  * Reordering metric is maximal distance, which a packet can be displaced
936  * in packet stream. With SACKs we can estimate it:
937  *
938  * 1. SACK fills old hole and the corresponding segment was not
939  *    ever retransmitted -> reordering. Alas, we cannot use it
940  *    when segment was retransmitted.
941  * 2. The last flaw is solved with D-SACK. D-SACK arrives
942  *    for retransmitted and already SACKed segment -> reordering..
943  * Both of these heuristics are not used in Loss state, when we cannot
944  * account for retransmits accurately.
945  *
946  * SACK block validation.
947  * ----------------------
948  *
949  * SACK block range validation checks that the received SACK block fits to
950  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
951  * Note that SND.UNA is not included to the range though being valid because
952  * it means that the receiver is rather inconsistent with itself reporting
953  * SACK reneging when it should advance SND.UNA. Such SACK block this is
954  * perfectly valid, however, in light of RFC2018 which explicitly states
955  * that "SACK block MUST reflect the newest segment.  Even if the newest
956  * segment is going to be discarded ...", not that it looks very clever
957  * in case of head skb. Due to potentional receiver driven attacks, we
958  * choose to avoid immediate execution of a walk in write queue due to
959  * reneging and defer head skb's loss recovery to standard loss recovery
960  * procedure that will eventually trigger (nothing forbids us doing this).
961  *
962  * Implements also blockage to start_seq wrap-around. Problem lies in the
963  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
964  * there's no guarantee that it will be before snd_nxt (n). The problem
965  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
966  * wrap (s_w):
967  *
968  *         <- outs wnd ->                          <- wrapzone ->
969  *         u     e      n                         u_w   e_w  s n_w
970  *         |     |      |                          |     |   |  |
971  * |<------------+------+----- TCP seqno space --------------+---------->|
972  * ...-- <2^31 ->|                                           |<--------...
973  * ...---- >2^31 ------>|                                    |<--------...
974  *
975  * Current code wouldn't be vulnerable but it's better still to discard such
976  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
977  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
978  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
979  * equal to the ideal case (infinite seqno space without wrap caused issues).
980  *
981  * With D-SACK the lower bound is extended to cover sequence space below
982  * SND.UNA down to undo_marker, which is the last point of interest. Yet
983  * again, D-SACK block must not to go across snd_una (for the same reason as
984  * for the normal SACK blocks, explained above). But there all simplicity
985  * ends, TCP might receive valid D-SACKs below that. As long as they reside
986  * fully below undo_marker they do not affect behavior in anyway and can
987  * therefore be safely ignored. In rare cases (which are more or less
988  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
989  * fragmentation and packet reordering past skb's retransmission. To consider
990  * them correctly, the acceptable range must be extended even more though
991  * the exact amount is rather hard to quantify. However, tp->max_window can
992  * be used as an exaggerated estimate.
993  */
994 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
995                                    u32 start_seq, u32 end_seq)
996 {
997         /* Too far in future, or reversed (interpretation is ambiguous) */
998         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
999                 return false;
1000 
1001         /* Nasty start_seq wrap-around check (see comments above) */
1002         if (!before(start_seq, tp->snd_nxt))
1003                 return false;
1004 
1005         /* In outstanding window? ...This is valid exit for D-SACKs too.
1006          * start_seq == snd_una is non-sensical (see comments above)
1007          */
1008         if (after(start_seq, tp->snd_una))
1009                 return true;
1010 
1011         if (!is_dsack || !tp->undo_marker)
1012                 return false;
1013 
1014         /* ...Then it's D-SACK, and must reside below snd_una completely */
1015         if (after(end_seq, tp->snd_una))
1016                 return false;
1017 
1018         if (!before(start_seq, tp->undo_marker))
1019                 return true;
1020 
1021         /* Too old */
1022         if (!after(end_seq, tp->undo_marker))
1023                 return false;
1024 
1025         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1026          *   start_seq < undo_marker and end_seq >= undo_marker.
1027          */
1028         return !before(start_seq, end_seq - tp->max_window);
1029 }
1030 
1031 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1032  * Event "B". Later note: FACK people cheated me again 8), we have to account
1033  * for reordering! Ugly, but should help.
1034  *
1035  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1036  * less than what is now known to be received by the other end (derived from
1037  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1038  * retransmitted skbs to avoid some costly processing per ACKs.
1039  */
1040 static void tcp_mark_lost_retrans(struct sock *sk)
1041 {
1042         const struct inet_connection_sock *icsk = inet_csk(sk);
1043         struct tcp_sock *tp = tcp_sk(sk);
1044         struct sk_buff *skb;
1045         int cnt = 0;
1046         u32 new_low_seq = tp->snd_nxt;
1047         u32 received_upto = tcp_highest_sack_seq(tp);
1048 
1049         if (!tcp_is_fack(tp) || !tp->retrans_out ||
1050             !after(received_upto, tp->lost_retrans_low) ||
1051             icsk->icsk_ca_state != TCP_CA_Recovery)
1052                 return;
1053 
1054         tcp_for_write_queue(skb, sk) {
1055                 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1056 
1057                 if (skb == tcp_send_head(sk))
1058                         break;
1059                 if (cnt == tp->retrans_out)
1060                         break;
1061                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1062                         continue;
1063 
1064                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1065                         continue;
1066 
1067                 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1068                  * constraint here (see above) but figuring out that at
1069                  * least tp->reordering SACK blocks reside between ack_seq
1070                  * and received_upto is not easy task to do cheaply with
1071                  * the available datastructures.
1072                  *
1073                  * Whether FACK should check here for tp->reordering segs
1074                  * in-between one could argue for either way (it would be
1075                  * rather simple to implement as we could count fack_count
1076                  * during the walk and do tp->fackets_out - fack_count).
1077                  */
1078                 if (after(received_upto, ack_seq)) {
1079                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1080                         tp->retrans_out -= tcp_skb_pcount(skb);
1081 
1082                         tcp_skb_mark_lost_uncond_verify(tp, skb);
1083                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1084                 } else {
1085                         if (before(ack_seq, new_low_seq))
1086                                 new_low_seq = ack_seq;
1087                         cnt += tcp_skb_pcount(skb);
1088                 }
1089         }
1090 
1091         if (tp->retrans_out)
1092                 tp->lost_retrans_low = new_low_seq;
1093 }
1094 
1095 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1096                             struct tcp_sack_block_wire *sp, int num_sacks,
1097                             u32 prior_snd_una)
1098 {
1099         struct tcp_sock *tp = tcp_sk(sk);
1100         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1101         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1102         bool dup_sack = false;
1103 
1104         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1105                 dup_sack = true;
1106                 tcp_dsack_seen(tp);
1107                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1108         } else if (num_sacks > 1) {
1109                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1110                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1111 
1112                 if (!after(end_seq_0, end_seq_1) &&
1113                     !before(start_seq_0, start_seq_1)) {
1114                         dup_sack = true;
1115                         tcp_dsack_seen(tp);
1116                         NET_INC_STATS_BH(sock_net(sk),
1117                                         LINUX_MIB_TCPDSACKOFORECV);
1118                 }
1119         }
1120 
1121         /* D-SACK for already forgotten data... Do dumb counting. */
1122         if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1123             !after(end_seq_0, prior_snd_una) &&
1124             after(end_seq_0, tp->undo_marker))
1125                 tp->undo_retrans--;
1126 
1127         return dup_sack;
1128 }
1129 
1130 struct tcp_sacktag_state {
1131         int     reord;
1132         int     fack_count;
1133         long    rtt_us; /* RTT measured by SACKing never-retransmitted data */
1134         int     flag;
1135 };
1136 
1137 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1138  * the incoming SACK may not exactly match but we can find smaller MSS
1139  * aligned portion of it that matches. Therefore we might need to fragment
1140  * which may fail and creates some hassle (caller must handle error case
1141  * returns).
1142  *
1143  * FIXME: this could be merged to shift decision code
1144  */
1145 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1146                                   u32 start_seq, u32 end_seq)
1147 {
1148         int err;
1149         bool in_sack;
1150         unsigned int pkt_len;
1151         unsigned int mss;
1152 
1153         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1154                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1155 
1156         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1157             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1158                 mss = tcp_skb_mss(skb);
1159                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1160 
1161                 if (!in_sack) {
1162                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1163                         if (pkt_len < mss)
1164                                 pkt_len = mss;
1165                 } else {
1166                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1167                         if (pkt_len < mss)
1168                                 return -EINVAL;
1169                 }
1170 
1171                 /* Round if necessary so that SACKs cover only full MSSes
1172                  * and/or the remaining small portion (if present)
1173                  */
1174                 if (pkt_len > mss) {
1175                         unsigned int new_len = (pkt_len / mss) * mss;
1176                         if (!in_sack && new_len < pkt_len)
1177                                 new_len += mss;
1178                         pkt_len = new_len;
1179                 }
1180 
1181                 if (pkt_len >= skb->len && !in_sack)
1182                         return 0;
1183 
1184                 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1185                 if (err < 0)
1186                         return err;
1187         }
1188 
1189         return in_sack;
1190 }
1191 
1192 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1193 static u8 tcp_sacktag_one(struct sock *sk,
1194                           struct tcp_sacktag_state *state, u8 sacked,
1195                           u32 start_seq, u32 end_seq,
1196                           int dup_sack, int pcount,
1197                           const struct skb_mstamp *xmit_time)
1198 {
1199         struct tcp_sock *tp = tcp_sk(sk);
1200         int fack_count = state->fack_count;
1201 
1202         /* Account D-SACK for retransmitted packet. */
1203         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1204                 if (tp->undo_marker && tp->undo_retrans > 0 &&
1205                     after(end_seq, tp->undo_marker))
1206                         tp->undo_retrans--;
1207                 if (sacked & TCPCB_SACKED_ACKED)
1208                         state->reord = min(fack_count, state->reord);
1209         }
1210 
1211         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1212         if (!after(end_seq, tp->snd_una))
1213                 return sacked;
1214 
1215         if (!(sacked & TCPCB_SACKED_ACKED)) {
1216                 if (sacked & TCPCB_SACKED_RETRANS) {
1217                         /* If the segment is not tagged as lost,
1218                          * we do not clear RETRANS, believing
1219                          * that retransmission is still in flight.
1220                          */
1221                         if (sacked & TCPCB_LOST) {
1222                                 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1223                                 tp->lost_out -= pcount;
1224                                 tp->retrans_out -= pcount;
1225                         }
1226                 } else {
1227                         if (!(sacked & TCPCB_RETRANS)) {
1228                                 /* New sack for not retransmitted frame,
1229                                  * which was in hole. It is reordering.
1230                                  */
1231                                 if (before(start_seq,
1232                                            tcp_highest_sack_seq(tp)))
1233                                         state->reord = min(fack_count,
1234                                                            state->reord);
1235                                 if (!after(end_seq, tp->high_seq))
1236                                         state->flag |= FLAG_ORIG_SACK_ACKED;
1237                                 /* Pick the earliest sequence sacked for RTT */
1238                                 if (state->rtt_us < 0) {
1239                                         struct skb_mstamp now;
1240 
1241                                         skb_mstamp_get(&now);
1242                                         state->rtt_us = skb_mstamp_us_delta(&now,
1243                                                                 xmit_time);
1244                                 }
1245                         }
1246 
1247                         if (sacked & TCPCB_LOST) {
1248                                 sacked &= ~TCPCB_LOST;
1249                                 tp->lost_out -= pcount;
1250                         }
1251                 }
1252 
1253                 sacked |= TCPCB_SACKED_ACKED;
1254                 state->flag |= FLAG_DATA_SACKED;
1255                 tp->sacked_out += pcount;
1256 
1257                 fack_count += pcount;
1258 
1259                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1260                 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1261                     before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1262                         tp->lost_cnt_hint += pcount;
1263 
1264                 if (fack_count > tp->fackets_out)
1265                         tp->fackets_out = fack_count;
1266         }
1267 
1268         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1269          * frames and clear it. undo_retrans is decreased above, L|R frames
1270          * are accounted above as well.
1271          */
1272         if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1273                 sacked &= ~TCPCB_SACKED_RETRANS;
1274                 tp->retrans_out -= pcount;
1275         }
1276 
1277         return sacked;
1278 }
1279 
1280 /* Shift newly-SACKed bytes from this skb to the immediately previous
1281  * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1282  */
1283 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1284                             struct tcp_sacktag_state *state,
1285                             unsigned int pcount, int shifted, int mss,
1286                             bool dup_sack)
1287 {
1288         struct tcp_sock *tp = tcp_sk(sk);
1289         struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1290         u32 start_seq = TCP_SKB_CB(skb)->seq;   /* start of newly-SACKed */
1291         u32 end_seq = start_seq + shifted;      /* end of newly-SACKed */
1292 
1293         BUG_ON(!pcount);
1294 
1295         /* Adjust counters and hints for the newly sacked sequence
1296          * range but discard the return value since prev is already
1297          * marked. We must tag the range first because the seq
1298          * advancement below implicitly advances
1299          * tcp_highest_sack_seq() when skb is highest_sack.
1300          */
1301         tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1302                         start_seq, end_seq, dup_sack, pcount,
1303                         &skb->skb_mstamp);
1304 
1305         if (skb == tp->lost_skb_hint)
1306                 tp->lost_cnt_hint += pcount;
1307 
1308         TCP_SKB_CB(prev)->end_seq += shifted;
1309         TCP_SKB_CB(skb)->seq += shifted;
1310 
1311         tcp_skb_pcount_add(prev, pcount);
1312         BUG_ON(tcp_skb_pcount(skb) < pcount);
1313         tcp_skb_pcount_add(skb, -pcount);
1314 
1315         /* When we're adding to gso_segs == 1, gso_size will be zero,
1316          * in theory this shouldn't be necessary but as long as DSACK
1317          * code can come after this skb later on it's better to keep
1318          * setting gso_size to something.
1319          */
1320         if (!skb_shinfo(prev)->gso_size) {
1321                 skb_shinfo(prev)->gso_size = mss;
1322                 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1323         }
1324 
1325         /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1326         if (tcp_skb_pcount(skb) <= 1) {
1327                 skb_shinfo(skb)->gso_size = 0;
1328                 skb_shinfo(skb)->gso_type = 0;
1329         }
1330 
1331         /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1332         TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1333 
1334         if (skb->len > 0) {
1335                 BUG_ON(!tcp_skb_pcount(skb));
1336                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1337                 return false;
1338         }
1339 
1340         /* Whole SKB was eaten :-) */
1341 
1342         if (skb == tp->retransmit_skb_hint)
1343                 tp->retransmit_skb_hint = prev;
1344         if (skb == tp->lost_skb_hint) {
1345                 tp->lost_skb_hint = prev;
1346                 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1347         }
1348 
1349         TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1350         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1351                 TCP_SKB_CB(prev)->end_seq++;
1352 
1353         if (skb == tcp_highest_sack(sk))
1354                 tcp_advance_highest_sack(sk, skb);
1355 
1356         tcp_unlink_write_queue(skb, sk);
1357         sk_wmem_free_skb(sk, skb);
1358 
1359         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1360 
1361         return true;
1362 }
1363 
1364 /* I wish gso_size would have a bit more sane initialization than
1365  * something-or-zero which complicates things
1366  */
1367 static int tcp_skb_seglen(const struct sk_buff *skb)
1368 {
1369         return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1370 }
1371 
1372 /* Shifting pages past head area doesn't work */
1373 static int skb_can_shift(const struct sk_buff *skb)
1374 {
1375         return !skb_headlen(skb) && skb_is_nonlinear(skb);
1376 }
1377 
1378 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1379  * skb.
1380  */
1381 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1382                                           struct tcp_sacktag_state *state,
1383                                           u32 start_seq, u32 end_seq,
1384                                           bool dup_sack)
1385 {
1386         struct tcp_sock *tp = tcp_sk(sk);
1387         struct sk_buff *prev;
1388         int mss;
1389         int pcount = 0;
1390         int len;
1391         int in_sack;
1392 
1393         if (!sk_can_gso(sk))
1394                 goto fallback;
1395 
1396         /* Normally R but no L won't result in plain S */
1397         if (!dup_sack &&
1398             (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1399                 goto fallback;
1400         if (!skb_can_shift(skb))
1401                 goto fallback;
1402         /* This frame is about to be dropped (was ACKed). */
1403         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1404                 goto fallback;
1405 
1406         /* Can only happen with delayed DSACK + discard craziness */
1407         if (unlikely(skb == tcp_write_queue_head(sk)))
1408                 goto fallback;
1409         prev = tcp_write_queue_prev(sk, skb);
1410 
1411         if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1412                 goto fallback;
1413 
1414         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1415                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1416 
1417         if (in_sack) {
1418                 len = skb->len;
1419                 pcount = tcp_skb_pcount(skb);
1420                 mss = tcp_skb_seglen(skb);
1421 
1422                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1423                  * drop this restriction as unnecessary
1424                  */
1425                 if (mss != tcp_skb_seglen(prev))
1426                         goto fallback;
1427         } else {
1428                 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1429                         goto noop;
1430                 /* CHECKME: This is non-MSS split case only?, this will
1431                  * cause skipped skbs due to advancing loop btw, original
1432                  * has that feature too
1433                  */
1434                 if (tcp_skb_pcount(skb) <= 1)
1435                         goto noop;
1436 
1437                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1438                 if (!in_sack) {
1439                         /* TODO: head merge to next could be attempted here
1440                          * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1441                          * though it might not be worth of the additional hassle
1442                          *
1443                          * ...we can probably just fallback to what was done
1444                          * previously. We could try merging non-SACKed ones
1445                          * as well but it probably isn't going to buy off
1446                          * because later SACKs might again split them, and
1447                          * it would make skb timestamp tracking considerably
1448                          * harder problem.
1449                          */
1450                         goto fallback;
1451                 }
1452 
1453                 len = end_seq - TCP_SKB_CB(skb)->seq;
1454                 BUG_ON(len < 0);
1455                 BUG_ON(len > skb->len);
1456 
1457                 /* MSS boundaries should be honoured or else pcount will
1458                  * severely break even though it makes things bit trickier.
1459                  * Optimize common case to avoid most of the divides
1460                  */
1461                 mss = tcp_skb_mss(skb);
1462 
1463                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1464                  * drop this restriction as unnecessary
1465                  */
1466                 if (mss != tcp_skb_seglen(prev))
1467                         goto fallback;
1468 
1469                 if (len == mss) {
1470                         pcount = 1;
1471                 } else if (len < mss) {
1472                         goto noop;
1473                 } else {
1474                         pcount = len / mss;
1475                         len = pcount * mss;
1476                 }
1477         }
1478 
1479         /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1480         if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1481                 goto fallback;
1482 
1483         if (!skb_shift(prev, skb, len))
1484                 goto fallback;
1485         if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1486                 goto out;
1487 
1488         /* Hole filled allows collapsing with the next as well, this is very
1489          * useful when hole on every nth skb pattern happens
1490          */
1491         if (prev == tcp_write_queue_tail(sk))
1492                 goto out;
1493         skb = tcp_write_queue_next(sk, prev);
1494 
1495         if (!skb_can_shift(skb) ||
1496             (skb == tcp_send_head(sk)) ||
1497             ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1498             (mss != tcp_skb_seglen(skb)))
1499                 goto out;
1500 
1501         len = skb->len;
1502         if (skb_shift(prev, skb, len)) {
1503                 pcount += tcp_skb_pcount(skb);
1504                 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1505         }
1506 
1507 out:
1508         state->fack_count += pcount;
1509         return prev;
1510 
1511 noop:
1512         return skb;
1513 
1514 fallback:
1515         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1516         return NULL;
1517 }
1518 
1519 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1520                                         struct tcp_sack_block *next_dup,
1521                                         struct tcp_sacktag_state *state,
1522                                         u32 start_seq, u32 end_seq,
1523                                         bool dup_sack_in)
1524 {
1525         struct tcp_sock *tp = tcp_sk(sk);
1526         struct sk_buff *tmp;
1527 
1528         tcp_for_write_queue_from(skb, sk) {
1529                 int in_sack = 0;
1530                 bool dup_sack = dup_sack_in;
1531 
1532                 if (skb == tcp_send_head(sk))
1533                         break;
1534 
1535                 /* queue is in-order => we can short-circuit the walk early */
1536                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1537                         break;
1538 
1539                 if (next_dup  &&
1540                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1541                         in_sack = tcp_match_skb_to_sack(sk, skb,
1542                                                         next_dup->start_seq,
1543                                                         next_dup->end_seq);
1544                         if (in_sack > 0)
1545                                 dup_sack = true;
1546                 }
1547 
1548                 /* skb reference here is a bit tricky to get right, since
1549                  * shifting can eat and free both this skb and the next,
1550                  * so not even _safe variant of the loop is enough.
1551                  */
1552                 if (in_sack <= 0) {
1553                         tmp = tcp_shift_skb_data(sk, skb, state,
1554                                                  start_seq, end_seq, dup_sack);
1555                         if (tmp) {
1556                                 if (tmp != skb) {
1557                                         skb = tmp;
1558                                         continue;
1559                                 }
1560 
1561                                 in_sack = 0;
1562                         } else {
1563                                 in_sack = tcp_match_skb_to_sack(sk, skb,
1564                                                                 start_seq,
1565                                                                 end_seq);
1566                         }
1567                 }
1568 
1569                 if (unlikely(in_sack < 0))
1570                         break;
1571 
1572                 if (in_sack) {
1573                         TCP_SKB_CB(skb)->sacked =
1574                                 tcp_sacktag_one(sk,
1575                                                 state,
1576                                                 TCP_SKB_CB(skb)->sacked,
1577                                                 TCP_SKB_CB(skb)->seq,
1578                                                 TCP_SKB_CB(skb)->end_seq,
1579                                                 dup_sack,
1580                                                 tcp_skb_pcount(skb),
1581                                                 &skb->skb_mstamp);
1582 
1583                         if (!before(TCP_SKB_CB(skb)->seq,
1584                                     tcp_highest_sack_seq(tp)))
1585                                 tcp_advance_highest_sack(sk, skb);
1586                 }
1587 
1588                 state->fack_count += tcp_skb_pcount(skb);
1589         }
1590         return skb;
1591 }
1592 
1593 /* Avoid all extra work that is being done by sacktag while walking in
1594  * a normal way
1595  */
1596 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1597                                         struct tcp_sacktag_state *state,
1598                                         u32 skip_to_seq)
1599 {
1600         tcp_for_write_queue_from(skb, sk) {
1601                 if (skb == tcp_send_head(sk))
1602                         break;
1603 
1604                 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1605                         break;
1606 
1607                 state->fack_count += tcp_skb_pcount(skb);
1608         }
1609         return skb;
1610 }
1611 
1612 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1613                                                 struct sock *sk,
1614                                                 struct tcp_sack_block *next_dup,
1615                                                 struct tcp_sacktag_state *state,
1616                                                 u32 skip_to_seq)
1617 {
1618         if (!next_dup)
1619                 return skb;
1620 
1621         if (before(next_dup->start_seq, skip_to_seq)) {
1622                 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1623                 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1624                                        next_dup->start_seq, next_dup->end_seq,
1625                                        1);
1626         }
1627 
1628         return skb;
1629 }
1630 
1631 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1632 {
1633         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1634 }
1635 
1636 static int
1637 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1638                         u32 prior_snd_una, long *sack_rtt_us)
1639 {
1640         struct tcp_sock *tp = tcp_sk(sk);
1641         const unsigned char *ptr = (skb_transport_header(ack_skb) +
1642                                     TCP_SKB_CB(ack_skb)->sacked);
1643         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1644         struct tcp_sack_block sp[TCP_NUM_SACKS];
1645         struct tcp_sack_block *cache;
1646         struct tcp_sacktag_state state;
1647         struct sk_buff *skb;
1648         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1649         int used_sacks;
1650         bool found_dup_sack = false;
1651         int i, j;
1652         int first_sack_index;
1653 
1654         state.flag = 0;
1655         state.reord = tp->packets_out;
1656         state.rtt_us = -1L;
1657 
1658         if (!tp->sacked_out) {
1659                 if (WARN_ON(tp->fackets_out))
1660                         tp->fackets_out = 0;
1661                 tcp_highest_sack_reset(sk);
1662         }
1663 
1664         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1665                                          num_sacks, prior_snd_una);
1666         if (found_dup_sack)
1667                 state.flag |= FLAG_DSACKING_ACK;
1668 
1669         /* Eliminate too old ACKs, but take into
1670          * account more or less fresh ones, they can
1671          * contain valid SACK info.
1672          */
1673         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1674                 return 0;
1675 
1676         if (!tp->packets_out)
1677                 goto out;
1678 
1679         used_sacks = 0;
1680         first_sack_index = 0;
1681         for (i = 0; i < num_sacks; i++) {
1682                 bool dup_sack = !i && found_dup_sack;
1683 
1684                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1685                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1686 
1687                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1688                                             sp[used_sacks].start_seq,
1689                                             sp[used_sacks].end_seq)) {
1690                         int mib_idx;
1691 
1692                         if (dup_sack) {
1693                                 if (!tp->undo_marker)
1694                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1695                                 else
1696                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1697                         } else {
1698                                 /* Don't count olds caused by ACK reordering */
1699                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1700                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1701                                         continue;
1702                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1703                         }
1704 
1705                         NET_INC_STATS_BH(sock_net(sk), mib_idx);
1706                         if (i == 0)
1707                                 first_sack_index = -1;
1708                         continue;
1709                 }
1710 
1711                 /* Ignore very old stuff early */
1712                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1713                         continue;
1714 
1715                 used_sacks++;
1716         }
1717 
1718         /* order SACK blocks to allow in order walk of the retrans queue */
1719         for (i = used_sacks - 1; i > 0; i--) {
1720                 for (j = 0; j < i; j++) {
1721                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1722                                 swap(sp[j], sp[j + 1]);
1723 
1724                                 /* Track where the first SACK block goes to */
1725                                 if (j == first_sack_index)
1726                                         first_sack_index = j + 1;
1727                         }
1728                 }
1729         }
1730 
1731         skb = tcp_write_queue_head(sk);
1732         state.fack_count = 0;
1733         i = 0;
1734 
1735         if (!tp->sacked_out) {
1736                 /* It's already past, so skip checking against it */
1737                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1738         } else {
1739                 cache = tp->recv_sack_cache;
1740                 /* Skip empty blocks in at head of the cache */
1741                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1742                        !cache->end_seq)
1743                         cache++;
1744         }
1745 
1746         while (i < used_sacks) {
1747                 u32 start_seq = sp[i].start_seq;
1748                 u32 end_seq = sp[i].end_seq;
1749                 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1750                 struct tcp_sack_block *next_dup = NULL;
1751 
1752                 if (found_dup_sack && ((i + 1) == first_sack_index))
1753                         next_dup = &sp[i + 1];
1754 
1755                 /* Skip too early cached blocks */
1756                 while (tcp_sack_cache_ok(tp, cache) &&
1757                        !before(start_seq, cache->end_seq))
1758                         cache++;
1759 
1760                 /* Can skip some work by looking recv_sack_cache? */
1761                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1762                     after(end_seq, cache->start_seq)) {
1763 
1764                         /* Head todo? */
1765                         if (before(start_seq, cache->start_seq)) {
1766                                 skb = tcp_sacktag_skip(skb, sk, &state,
1767                                                        start_seq);
1768                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1769                                                        &state,
1770                                                        start_seq,
1771                                                        cache->start_seq,
1772                                                        dup_sack);
1773                         }
1774 
1775                         /* Rest of the block already fully processed? */
1776                         if (!after(end_seq, cache->end_seq))
1777                                 goto advance_sp;
1778 
1779                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1780                                                        &state,
1781                                                        cache->end_seq);
1782 
1783                         /* ...tail remains todo... */
1784                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1785                                 /* ...but better entrypoint exists! */
1786                                 skb = tcp_highest_sack(sk);
1787                                 if (!skb)
1788                                         break;
1789                                 state.fack_count = tp->fackets_out;
1790                                 cache++;
1791                                 goto walk;
1792                         }
1793 
1794                         skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1795                         /* Check overlap against next cached too (past this one already) */
1796                         cache++;
1797                         continue;
1798                 }
1799 
1800                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1801                         skb = tcp_highest_sack(sk);
1802                         if (!skb)
1803                                 break;
1804                         state.fack_count = tp->fackets_out;
1805                 }
1806                 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1807 
1808 walk:
1809                 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1810                                        start_seq, end_seq, dup_sack);
1811 
1812 advance_sp:
1813                 i++;
1814         }
1815 
1816         /* Clear the head of the cache sack blocks so we can skip it next time */
1817         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1818                 tp->recv_sack_cache[i].start_seq = 0;
1819                 tp->recv_sack_cache[i].end_seq = 0;
1820         }
1821         for (j = 0; j < used_sacks; j++)
1822                 tp->recv_sack_cache[i++] = sp[j];
1823 
1824         if ((state.reord < tp->fackets_out) &&
1825             ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1826                 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1827 
1828         tcp_mark_lost_retrans(sk);
1829         tcp_verify_left_out(tp);
1830 out:
1831 
1832 #if FASTRETRANS_DEBUG > 0
1833         WARN_ON((int)tp->sacked_out < 0);
1834         WARN_ON((int)tp->lost_out < 0);
1835         WARN_ON((int)tp->retrans_out < 0);
1836         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1837 #endif
1838         *sack_rtt_us = state.rtt_us;
1839         return state.flag;
1840 }
1841 
1842 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1843  * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1844  */
1845 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1846 {
1847         u32 holes;
1848 
1849         holes = max(tp->lost_out, 1U);
1850         holes = min(holes, tp->packets_out);
1851 
1852         if ((tp->sacked_out + holes) > tp->packets_out) {
1853                 tp->sacked_out = tp->packets_out - holes;
1854                 return true;
1855         }
1856         return false;
1857 }
1858 
1859 /* If we receive more dupacks than we expected counting segments
1860  * in assumption of absent reordering, interpret this as reordering.
1861  * The only another reason could be bug in receiver TCP.
1862  */
1863 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1864 {
1865         struct tcp_sock *tp = tcp_sk(sk);
1866         if (tcp_limit_reno_sacked(tp))
1867                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1868 }
1869 
1870 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1871 
1872 static void tcp_add_reno_sack(struct sock *sk)
1873 {
1874         struct tcp_sock *tp = tcp_sk(sk);
1875         tp->sacked_out++;
1876         tcp_check_reno_reordering(sk, 0);
1877         tcp_verify_left_out(tp);
1878 }
1879 
1880 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1881 
1882 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1883 {
1884         struct tcp_sock *tp = tcp_sk(sk);
1885 
1886         if (acked > 0) {
1887                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1888                 if (acked - 1 >= tp->sacked_out)
1889                         tp->sacked_out = 0;
1890                 else
1891                         tp->sacked_out -= acked - 1;
1892         }
1893         tcp_check_reno_reordering(sk, acked);
1894         tcp_verify_left_out(tp);
1895 }
1896 
1897 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1898 {
1899         tp->sacked_out = 0;
1900 }
1901 
1902 void tcp_clear_retrans(struct tcp_sock *tp)
1903 {
1904         tp->retrans_out = 0;
1905         tp->lost_out = 0;
1906         tp->undo_marker = 0;
1907         tp->undo_retrans = -1;
1908         tp->fackets_out = 0;
1909         tp->sacked_out = 0;
1910 }
1911 
1912 static inline void tcp_init_undo(struct tcp_sock *tp)
1913 {
1914         tp->undo_marker = tp->snd_una;
1915         /* Retransmission still in flight may cause DSACKs later. */
1916         tp->undo_retrans = tp->retrans_out ? : -1;
1917 }
1918 
1919 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1920  * and reset tags completely, otherwise preserve SACKs. If receiver
1921  * dropped its ofo queue, we will know this due to reneging detection.
1922  */
1923 void tcp_enter_loss(struct sock *sk)
1924 {
1925         const struct inet_connection_sock *icsk = inet_csk(sk);
1926         struct tcp_sock *tp = tcp_sk(sk);
1927         struct sk_buff *skb;
1928         bool new_recovery = false;
1929         bool is_reneg;                  /* is receiver reneging on SACKs? */
1930 
1931         /* Reduce ssthresh if it has not yet been made inside this window. */
1932         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1933             !after(tp->high_seq, tp->snd_una) ||
1934             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1935                 new_recovery = true;
1936                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1937                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1938                 tcp_ca_event(sk, CA_EVENT_LOSS);
1939                 tcp_init_undo(tp);
1940         }
1941         tp->snd_cwnd       = 1;
1942         tp->snd_cwnd_cnt   = 0;
1943         tp->snd_cwnd_stamp = tcp_time_stamp;
1944 
1945         tp->retrans_out = 0;
1946         tp->lost_out = 0;
1947 
1948         if (tcp_is_reno(tp))
1949                 tcp_reset_reno_sack(tp);
1950 
1951         skb = tcp_write_queue_head(sk);
1952         is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1953         if (is_reneg) {
1954                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1955                 tp->sacked_out = 0;
1956                 tp->fackets_out = 0;
1957         }
1958         tcp_clear_all_retrans_hints(tp);
1959 
1960         tcp_for_write_queue(skb, sk) {
1961                 if (skb == tcp_send_head(sk))
1962                         break;
1963 
1964                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1965                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
1966                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1967                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1968                         tp->lost_out += tcp_skb_pcount(skb);
1969                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1970                 }
1971         }
1972         tcp_verify_left_out(tp);
1973 
1974         /* Timeout in disordered state after receiving substantial DUPACKs
1975          * suggests that the degree of reordering is over-estimated.
1976          */
1977         if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1978             tp->sacked_out >= sysctl_tcp_reordering)
1979                 tp->reordering = min_t(unsigned int, tp->reordering,
1980                                        sysctl_tcp_reordering);
1981         tcp_set_ca_state(sk, TCP_CA_Loss);
1982         tp->high_seq = tp->snd_nxt;
1983         tcp_ecn_queue_cwr(tp);
1984 
1985         /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1986          * loss recovery is underway except recurring timeout(s) on
1987          * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1988          */
1989         tp->frto = sysctl_tcp_frto &&
1990                    (new_recovery || icsk->icsk_retransmits) &&
1991                    !inet_csk(sk)->icsk_mtup.probe_size;
1992 }
1993 
1994 /* If ACK arrived pointing to a remembered SACK, it means that our
1995  * remembered SACKs do not reflect real state of receiver i.e.
1996  * receiver _host_ is heavily congested (or buggy).
1997  *
1998  * To avoid big spurious retransmission bursts due to transient SACK
1999  * scoreboard oddities that look like reneging, we give the receiver a
2000  * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2001  * restore sanity to the SACK scoreboard. If the apparent reneging
2002  * persists until this RTO then we'll clear the SACK scoreboard.
2003  */
2004 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2005 {
2006         if (flag & FLAG_SACK_RENEGING) {
2007                 struct tcp_sock *tp = tcp_sk(sk);
2008                 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2009                                           msecs_to_jiffies(10));
2010 
2011                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2012                                           delay, TCP_RTO_MAX);
2013                 return true;
2014         }
2015         return false;
2016 }
2017 
2018 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2019 {
2020         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2021 }
2022 
2023 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2024  * counter when SACK is enabled (without SACK, sacked_out is used for
2025  * that purpose).
2026  *
2027  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2028  * segments up to the highest received SACK block so far and holes in
2029  * between them.
2030  *
2031  * With reordering, holes may still be in flight, so RFC3517 recovery
2032  * uses pure sacked_out (total number of SACKed segments) even though
2033  * it violates the RFC that uses duplicate ACKs, often these are equal
2034  * but when e.g. out-of-window ACKs or packet duplication occurs,
2035  * they differ. Since neither occurs due to loss, TCP should really
2036  * ignore them.
2037  */
2038 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2039 {
2040         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2041 }
2042 
2043 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2044 {
2045         struct tcp_sock *tp = tcp_sk(sk);
2046         unsigned long delay;
2047 
2048         /* Delay early retransmit and entering fast recovery for
2049          * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2050          * available, or RTO is scheduled to fire first.
2051          */
2052         if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2053             (flag & FLAG_ECE) || !tp->srtt_us)
2054                 return false;
2055 
2056         delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2057                     msecs_to_jiffies(2));
2058 
2059         if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2060                 return false;
2061 
2062         inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2063                                   TCP_RTO_MAX);
2064         return true;
2065 }
2066 
2067 /* Linux NewReno/SACK/FACK/ECN state machine.
2068  * --------------------------------------
2069  *
2070  * "Open"       Normal state, no dubious events, fast path.
2071  * "Disorder"   In all the respects it is "Open",
2072  *              but requires a bit more attention. It is entered when
2073  *              we see some SACKs or dupacks. It is split of "Open"
2074  *              mainly to move some processing from fast path to slow one.
2075  * "CWR"        CWND was reduced due to some Congestion Notification event.
2076  *              It can be ECN, ICMP source quench, local device congestion.
2077  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2078  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2079  *
2080  * tcp_fastretrans_alert() is entered:
2081  * - each incoming ACK, if state is not "Open"
2082  * - when arrived ACK is unusual, namely:
2083  *      * SACK
2084  *      * Duplicate ACK.
2085  *      * ECN ECE.
2086  *
2087  * Counting packets in flight is pretty simple.
2088  *
2089  *      in_flight = packets_out - left_out + retrans_out
2090  *
2091  *      packets_out is SND.NXT-SND.UNA counted in packets.
2092  *
2093  *      retrans_out is number of retransmitted segments.
2094  *
2095  *      left_out is number of segments left network, but not ACKed yet.
2096  *
2097  *              left_out = sacked_out + lost_out
2098  *
2099  *     sacked_out: Packets, which arrived to receiver out of order
2100  *                 and hence not ACKed. With SACKs this number is simply
2101  *                 amount of SACKed data. Even without SACKs
2102  *                 it is easy to give pretty reliable estimate of this number,
2103  *                 counting duplicate ACKs.
2104  *
2105  *       lost_out: Packets lost by network. TCP has no explicit
2106  *                 "loss notification" feedback from network (for now).
2107  *                 It means that this number can be only _guessed_.
2108  *                 Actually, it is the heuristics to predict lossage that
2109  *                 distinguishes different algorithms.
2110  *
2111  *      F.e. after RTO, when all the queue is considered as lost,
2112  *      lost_out = packets_out and in_flight = retrans_out.
2113  *
2114  *              Essentially, we have now two algorithms counting
2115  *              lost packets.
2116  *
2117  *              FACK: It is the simplest heuristics. As soon as we decided
2118  *              that something is lost, we decide that _all_ not SACKed
2119  *              packets until the most forward SACK are lost. I.e.
2120  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2121  *              It is absolutely correct estimate, if network does not reorder
2122  *              packets. And it loses any connection to reality when reordering
2123  *              takes place. We use FACK by default until reordering
2124  *              is suspected on the path to this destination.
2125  *
2126  *              NewReno: when Recovery is entered, we assume that one segment
2127  *              is lost (classic Reno). While we are in Recovery and
2128  *              a partial ACK arrives, we assume that one more packet
2129  *              is lost (NewReno). This heuristics are the same in NewReno
2130  *              and SACK.
2131  *
2132  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2133  *  deflation etc. CWND is real congestion window, never inflated, changes
2134  *  only according to classic VJ rules.
2135  *
2136  * Really tricky (and requiring careful tuning) part of algorithm
2137  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2138  * The first determines the moment _when_ we should reduce CWND and,
2139  * hence, slow down forward transmission. In fact, it determines the moment
2140  * when we decide that hole is caused by loss, rather than by a reorder.
2141  *
2142  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2143  * holes, caused by lost packets.
2144  *
2145  * And the most logically complicated part of algorithm is undo
2146  * heuristics. We detect false retransmits due to both too early
2147  * fast retransmit (reordering) and underestimated RTO, analyzing
2148  * timestamps and D-SACKs. When we detect that some segments were
2149  * retransmitted by mistake and CWND reduction was wrong, we undo
2150  * window reduction and abort recovery phase. This logic is hidden
2151  * inside several functions named tcp_try_undo_<something>.
2152  */
2153 
2154 /* This function decides, when we should leave Disordered state
2155  * and enter Recovery phase, reducing congestion window.
2156  *
2157  * Main question: may we further continue forward transmission
2158  * with the same cwnd?
2159  */
2160 static bool tcp_time_to_recover(struct sock *sk, int flag)
2161 {
2162         struct tcp_sock *tp = tcp_sk(sk);
2163         __u32 packets_out;
2164 
2165         /* Trick#1: The loss is proven. */
2166         if (tp->lost_out)
2167                 return true;
2168 
2169         /* Not-A-Trick#2 : Classic rule... */
2170         if (tcp_dupack_heuristics(tp) > tp->reordering)
2171                 return true;
2172 
2173         /* Trick#4: It is still not OK... But will it be useful to delay
2174          * recovery more?
2175          */
2176         packets_out = tp->packets_out;
2177         if (packets_out <= tp->reordering &&
2178             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2179             !tcp_may_send_now(sk)) {
2180                 /* We have nothing to send. This connection is limited
2181                  * either by receiver window or by application.
2182                  */
2183                 return true;
2184         }
2185 
2186         /* If a thin stream is detected, retransmit after first
2187          * received dupack. Employ only if SACK is supported in order
2188          * to avoid possible corner-case series of spurious retransmissions
2189          * Use only if there are no unsent data.
2190          */
2191         if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2192             tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2193             tcp_is_sack(tp) && !tcp_send_head(sk))
2194                 return true;
2195 
2196         /* Trick#6: TCP early retransmit, per RFC5827.  To avoid spurious
2197          * retransmissions due to small network reorderings, we implement
2198          * Mitigation A.3 in the RFC and delay the retransmission for a short
2199          * interval if appropriate.
2200          */
2201         if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2202             (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2203             !tcp_may_send_now(sk))
2204                 return !tcp_pause_early_retransmit(sk, flag);
2205 
2206         return false;
2207 }
2208 
2209 /* Detect loss in event "A" above by marking head of queue up as lost.
2210  * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2211  * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2212  * has at least tp->reordering SACKed seqments above it; "packets" refers to
2213  * the maximum SACKed segments to pass before reaching this limit.
2214  */
2215 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2216 {
2217         struct tcp_sock *tp = tcp_sk(sk);
2218         struct sk_buff *skb;
2219         int cnt, oldcnt;
2220         int err;
2221         unsigned int mss;
2222         /* Use SACK to deduce losses of new sequences sent during recovery */
2223         const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2224 
2225         WARN_ON(packets > tp->packets_out);
2226         if (tp->lost_skb_hint) {
2227                 skb = tp->lost_skb_hint;
2228                 cnt = tp->lost_cnt_hint;
2229                 /* Head already handled? */
2230                 if (mark_head && skb != tcp_write_queue_head(sk))
2231                         return;
2232         } else {
2233                 skb = tcp_write_queue_head(sk);
2234                 cnt = 0;
2235         }
2236 
2237         tcp_for_write_queue_from(skb, sk) {
2238                 if (skb == tcp_send_head(sk))
2239                         break;
2240                 /* TODO: do this better */
2241                 /* this is not the most efficient way to do this... */
2242                 tp->lost_skb_hint = skb;
2243                 tp->lost_cnt_hint = cnt;
2244 
2245                 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2246                         break;
2247 
2248                 oldcnt = cnt;
2249                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2250                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2251                         cnt += tcp_skb_pcount(skb);
2252 
2253                 if (cnt > packets) {
2254                         if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2255                             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2256                             (oldcnt >= packets))
2257                                 break;
2258 
2259                         mss = skb_shinfo(skb)->gso_size;
2260                         err = tcp_fragment(sk, skb, (packets - oldcnt) * mss,
2261                                            mss, GFP_ATOMIC);
2262                         if (err < 0)
2263                                 break;
2264                         cnt = packets;
2265                 }
2266 
2267                 tcp_skb_mark_lost(tp, skb);
2268 
2269                 if (mark_head)
2270                         break;
2271         }
2272         tcp_verify_left_out(tp);
2273 }
2274 
2275 /* Account newly detected lost packet(s) */
2276 
2277 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2278 {
2279         struct tcp_sock *tp = tcp_sk(sk);
2280 
2281         if (tcp_is_reno(tp)) {
2282                 tcp_mark_head_lost(sk, 1, 1);
2283         } else if (tcp_is_fack(tp)) {
2284                 int lost = tp->fackets_out - tp->reordering;
2285                 if (lost <= 0)
2286                         lost = 1;
2287                 tcp_mark_head_lost(sk, lost, 0);
2288         } else {
2289                 int sacked_upto = tp->sacked_out - tp->reordering;
2290                 if (sacked_upto >= 0)
2291                         tcp_mark_head_lost(sk, sacked_upto, 0);
2292                 else if (fast_rexmit)
2293                         tcp_mark_head_lost(sk, 1, 1);
2294         }
2295 }
2296 
2297 /* CWND moderation, preventing bursts due to too big ACKs
2298  * in dubious situations.
2299  */
2300 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2301 {
2302         tp->snd_cwnd = min(tp->snd_cwnd,
2303                            tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2304         tp->snd_cwnd_stamp = tcp_time_stamp;
2305 }
2306 
2307 /* Nothing was retransmitted or returned timestamp is less
2308  * than timestamp of the first retransmission.
2309  */
2310 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2311 {
2312         return !tp->retrans_stamp ||
2313                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2314                  before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2315 }
2316 
2317 /* Undo procedures. */
2318 
2319 /* We can clear retrans_stamp when there are no retransmissions in the
2320  * window. It would seem that it is trivially available for us in
2321  * tp->retrans_out, however, that kind of assumptions doesn't consider
2322  * what will happen if errors occur when sending retransmission for the
2323  * second time. ...It could the that such segment has only
2324  * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2325  * the head skb is enough except for some reneging corner cases that
2326  * are not worth the effort.
2327  *
2328  * Main reason for all this complexity is the fact that connection dying
2329  * time now depends on the validity of the retrans_stamp, in particular,
2330  * that successive retransmissions of a segment must not advance
2331  * retrans_stamp under any conditions.
2332  */
2333 static bool tcp_any_retrans_done(const struct sock *sk)
2334 {
2335         const struct tcp_sock *tp = tcp_sk(sk);
2336         struct sk_buff *skb;
2337 
2338         if (tp->retrans_out)
2339                 return true;
2340 
2341         skb = tcp_write_queue_head(sk);
2342         if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2343                 return true;
2344 
2345         return false;
2346 }
2347 
2348 #if FASTRETRANS_DEBUG > 1
2349 static void DBGUNDO(struct sock *sk, const char *msg)
2350 {
2351         struct tcp_sock *tp = tcp_sk(sk);
2352         struct inet_sock *inet = inet_sk(sk);
2353 
2354         if (sk->sk_family == AF_INET) {
2355                 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2356                          msg,
2357                          &inet->inet_daddr, ntohs(inet->inet_dport),
2358                          tp->snd_cwnd, tcp_left_out(tp),
2359                          tp->snd_ssthresh, tp->prior_ssthresh,
2360                          tp->packets_out);
2361         }
2362 #if IS_ENABLED(CONFIG_IPV6)
2363         else if (sk->sk_family == AF_INET6) {
2364                 struct ipv6_pinfo *np = inet6_sk(sk);
2365                 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2366                          msg,
2367                          &np->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 #endif
2373 }
2374 #else
2375 #define DBGUNDO(x...) do { } while (0)
2376 #endif
2377 
2378 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2379 {
2380         struct tcp_sock *tp = tcp_sk(sk);
2381 
2382         if (unmark_loss) {
2383                 struct sk_buff *skb;
2384 
2385                 tcp_for_write_queue(skb, sk) {
2386                         if (skb == tcp_send_head(sk))
2387                                 break;
2388                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2389                 }
2390                 tp->lost_out = 0;
2391                 tcp_clear_all_retrans_hints(tp);
2392         }
2393 
2394         if (tp->prior_ssthresh) {
2395                 const struct inet_connection_sock *icsk = inet_csk(sk);
2396 
2397                 if (icsk->icsk_ca_ops->undo_cwnd)
2398                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2399                 else
2400                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2401 
2402                 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2403                         tp->snd_ssthresh = tp->prior_ssthresh;
2404                         tcp_ecn_withdraw_cwr(tp);
2405                 }
2406         } else {
2407                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2408         }
2409         tp->snd_cwnd_stamp = tcp_time_stamp;
2410         tp->undo_marker = 0;
2411 }
2412 
2413 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2414 {
2415         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2416 }
2417 
2418 /* People celebrate: "We love our President!" */
2419 static bool tcp_try_undo_recovery(struct sock *sk)
2420 {
2421         struct tcp_sock *tp = tcp_sk(sk);
2422 
2423         if (tcp_may_undo(tp)) {
2424                 int mib_idx;
2425 
2426                 /* Happy end! We did not retransmit anything
2427                  * or our original transmission succeeded.
2428                  */
2429                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2430                 tcp_undo_cwnd_reduction(sk, false);
2431                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2432                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2433                 else
2434                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2435 
2436                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2437         }
2438         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2439                 /* Hold old state until something *above* high_seq
2440                  * is ACKed. For Reno it is MUST to prevent false
2441                  * fast retransmits (RFC2582). SACK TCP is safe. */
2442                 tcp_moderate_cwnd(tp);
2443                 if (!tcp_any_retrans_done(sk))
2444                         tp->retrans_stamp = 0;
2445                 return true;
2446         }
2447         tcp_set_ca_state(sk, TCP_CA_Open);
2448         return false;
2449 }
2450 
2451 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2452 static bool tcp_try_undo_dsack(struct sock *sk)
2453 {
2454         struct tcp_sock *tp = tcp_sk(sk);
2455 
2456         if (tp->undo_marker && !tp->undo_retrans) {
2457                 DBGUNDO(sk, "D-SACK");
2458                 tcp_undo_cwnd_reduction(sk, false);
2459                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2460                 return true;
2461         }
2462         return false;
2463 }
2464 
2465 /* Undo during loss recovery after partial ACK or using F-RTO. */
2466 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2467 {
2468         struct tcp_sock *tp = tcp_sk(sk);
2469 
2470         if (frto_undo || tcp_may_undo(tp)) {
2471                 tcp_undo_cwnd_reduction(sk, true);
2472 
2473                 DBGUNDO(sk, "partial loss");
2474                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2475                 if (frto_undo)
2476                         NET_INC_STATS_BH(sock_net(sk),
2477                                          LINUX_MIB_TCPSPURIOUSRTOS);
2478                 inet_csk(sk)->icsk_retransmits = 0;
2479                 if (frto_undo || tcp_is_sack(tp))
2480                         tcp_set_ca_state(sk, TCP_CA_Open);
2481                 return true;
2482         }
2483         return false;
2484 }
2485 
2486 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2487  * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2488  * It computes the number of packets to send (sndcnt) based on packets newly
2489  * delivered:
2490  *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2491  *      cwnd reductions across a full RTT.
2492  *   2) If packets in flight is lower than ssthresh (such as due to excess
2493  *      losses and/or application stalls), do not perform any further cwnd
2494  *      reductions, but instead slow start up to ssthresh.
2495  */
2496 static void tcp_init_cwnd_reduction(struct sock *sk)
2497 {
2498         struct tcp_sock *tp = tcp_sk(sk);
2499 
2500         tp->high_seq = tp->snd_nxt;
2501         tp->tlp_high_seq = 0;
2502         tp->snd_cwnd_cnt = 0;
2503         tp->prior_cwnd = tp->snd_cwnd;
2504         tp->prr_delivered = 0;
2505         tp->prr_out = 0;
2506         tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2507         tcp_ecn_queue_cwr(tp);
2508 }
2509 
2510 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2511                                int fast_rexmit)
2512 {
2513         struct tcp_sock *tp = tcp_sk(sk);
2514         int sndcnt = 0;
2515         int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2516         int newly_acked_sacked = prior_unsacked -
2517                                  (tp->packets_out - tp->sacked_out);
2518 
2519         tp->prr_delivered += newly_acked_sacked;
2520         if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2521                 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2522                                tp->prior_cwnd - 1;
2523                 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2524         } else {
2525                 sndcnt = min_t(int, delta,
2526                                max_t(int, tp->prr_delivered - tp->prr_out,
2527                                      newly_acked_sacked) + 1);
2528         }
2529 
2530         sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2531         tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2532 }
2533 
2534 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2535 {
2536         struct tcp_sock *tp = tcp_sk(sk);
2537 
2538         /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2539         if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2540             (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2541                 tp->snd_cwnd = tp->snd_ssthresh;
2542                 tp->snd_cwnd_stamp = tcp_time_stamp;
2543         }
2544         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2545 }
2546 
2547 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2548 void tcp_enter_cwr(struct sock *sk)
2549 {
2550         struct tcp_sock *tp = tcp_sk(sk);
2551 
2552         tp->prior_ssthresh = 0;
2553         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2554                 tp->undo_marker = 0;
2555                 tcp_init_cwnd_reduction(sk);
2556                 tcp_set_ca_state(sk, TCP_CA_CWR);
2557         }
2558 }
2559 
2560 static void tcp_try_keep_open(struct sock *sk)
2561 {
2562         struct tcp_sock *tp = tcp_sk(sk);
2563         int state = TCP_CA_Open;
2564 
2565         if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2566                 state = TCP_CA_Disorder;
2567 
2568         if (inet_csk(sk)->icsk_ca_state != state) {
2569                 tcp_set_ca_state(sk, state);
2570                 tp->high_seq = tp->snd_nxt;
2571         }
2572 }
2573 
2574 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2575 {
2576         struct tcp_sock *tp = tcp_sk(sk);
2577 
2578         tcp_verify_left_out(tp);
2579 
2580         if (!tcp_any_retrans_done(sk))
2581                 tp->retrans_stamp = 0;
2582 
2583         if (flag & FLAG_ECE)
2584                 tcp_enter_cwr(sk);
2585 
2586         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2587                 tcp_try_keep_open(sk);
2588         } else {
2589                 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2590         }
2591 }
2592 
2593 static void tcp_mtup_probe_failed(struct sock *sk)
2594 {
2595         struct inet_connection_sock *icsk = inet_csk(sk);
2596 
2597         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2598         icsk->icsk_mtup.probe_size = 0;
2599 }
2600 
2601 static void tcp_mtup_probe_success(struct sock *sk)
2602 {
2603         struct tcp_sock *tp = tcp_sk(sk);
2604         struct inet_connection_sock *icsk = inet_csk(sk);
2605 
2606         /* FIXME: breaks with very large cwnd */
2607         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2608         tp->snd_cwnd = tp->snd_cwnd *
2609                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2610                        icsk->icsk_mtup.probe_size;
2611         tp->snd_cwnd_cnt = 0;
2612         tp->snd_cwnd_stamp = tcp_time_stamp;
2613         tp->snd_ssthresh = tcp_current_ssthresh(sk);
2614 
2615         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2616         icsk->icsk_mtup.probe_size = 0;
2617         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2618 }
2619 
2620 /* Do a simple retransmit without using the backoff mechanisms in
2621  * tcp_timer. This is used for path mtu discovery.
2622  * The socket is already locked here.
2623  */
2624 void tcp_simple_retransmit(struct sock *sk)
2625 {
2626         const struct inet_connection_sock *icsk = inet_csk(sk);
2627         struct tcp_sock *tp = tcp_sk(sk);
2628         struct sk_buff *skb;
2629         unsigned int mss = tcp_current_mss(sk);
2630         u32 prior_lost = tp->lost_out;
2631 
2632         tcp_for_write_queue(skb, sk) {
2633                 if (skb == tcp_send_head(sk))
2634                         break;
2635                 if (tcp_skb_seglen(skb) > mss &&
2636                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2637                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2638                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2639                                 tp->retrans_out -= tcp_skb_pcount(skb);
2640                         }
2641                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2642                 }
2643         }
2644 
2645         tcp_clear_retrans_hints_partial(tp);
2646 
2647         if (prior_lost == tp->lost_out)
2648                 return;
2649 
2650         if (tcp_is_reno(tp))
2651                 tcp_limit_reno_sacked(tp);
2652 
2653         tcp_verify_left_out(tp);
2654 
2655         /* Don't muck with the congestion window here.
2656          * Reason is that we do not increase amount of _data_
2657          * in network, but units changed and effective
2658          * cwnd/ssthresh really reduced now.
2659          */
2660         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2661                 tp->high_seq = tp->snd_nxt;
2662                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2663                 tp->prior_ssthresh = 0;
2664                 tp->undo_marker = 0;
2665                 tcp_set_ca_state(sk, TCP_CA_Loss);
2666         }
2667         tcp_xmit_retransmit_queue(sk);
2668 }
2669 EXPORT_SYMBOL(tcp_simple_retransmit);
2670 
2671 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2672 {
2673         struct tcp_sock *tp = tcp_sk(sk);
2674         int mib_idx;
2675 
2676         if (tcp_is_reno(tp))
2677                 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2678         else
2679                 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2680 
2681         NET_INC_STATS_BH(sock_net(sk), mib_idx);
2682 
2683         tp->prior_ssthresh = 0;
2684         tcp_init_undo(tp);
2685 
2686         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2687                 if (!ece_ack)
2688                         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2689                 tcp_init_cwnd_reduction(sk);
2690         }
2691         tcp_set_ca_state(sk, TCP_CA_Recovery);
2692 }
2693 
2694 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2695  * recovered or spurious. Otherwise retransmits more on partial ACKs.
2696  */
2697 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2698 {
2699         struct tcp_sock *tp = tcp_sk(sk);
2700         bool recovered = !before(tp->snd_una, tp->high_seq);
2701 
2702         if ((flag & FLAG_SND_UNA_ADVANCED) &&
2703             tcp_try_undo_loss(sk, false))
2704                 return;
2705 
2706         if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2707                 /* Step 3.b. A timeout is spurious if not all data are
2708                  * lost, i.e., never-retransmitted data are (s)acked.
2709                  */
2710                 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2711                     tcp_try_undo_loss(sk, true))
2712                         return;
2713 
2714                 if (after(tp->snd_nxt, tp->high_seq)) {
2715                         if (flag & FLAG_DATA_SACKED || is_dupack)
2716                                 tp->frto = 0; /* Step 3.a. loss was real */
2717                 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2718                         tp->high_seq = tp->snd_nxt;
2719                         __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2720                                                   TCP_NAGLE_OFF);
2721                         if (after(tp->snd_nxt, tp->high_seq))
2722                                 return; /* Step 2.b */
2723                         tp->frto = 0;
2724                 }
2725         }
2726 
2727         if (recovered) {
2728                 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2729                 tcp_try_undo_recovery(sk);
2730                 return;
2731         }
2732         if (tcp_is_reno(tp)) {
2733                 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2734                  * delivered. Lower inflight to clock out (re)tranmissions.
2735                  */
2736                 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2737                         tcp_add_reno_sack(sk);
2738                 else if (flag & FLAG_SND_UNA_ADVANCED)
2739                         tcp_reset_reno_sack(tp);
2740         }
2741         tcp_xmit_retransmit_queue(sk);
2742 }
2743 
2744 /* Undo during fast recovery after partial ACK. */
2745 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2746                                  const int prior_unsacked)
2747 {
2748         struct tcp_sock *tp = tcp_sk(sk);
2749 
2750         if (tp->undo_marker && tcp_packet_delayed(tp)) {
2751                 /* Plain luck! Hole if filled with delayed
2752                  * packet, rather than with a retransmit.
2753                  */
2754                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2755 
2756                 /* We are getting evidence that the reordering degree is higher
2757                  * than we realized. If there are no retransmits out then we
2758                  * can undo. Otherwise we clock out new packets but do not
2759                  * mark more packets lost or retransmit more.
2760                  */
2761                 if (tp->retrans_out) {
2762                         tcp_cwnd_reduction(sk, prior_unsacked, 0);
2763                         return true;
2764                 }
2765 
2766                 if (!tcp_any_retrans_done(sk))
2767                         tp->retrans_stamp = 0;
2768 
2769                 DBGUNDO(sk, "partial recovery");
2770                 tcp_undo_cwnd_reduction(sk, true);
2771                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2772                 tcp_try_keep_open(sk);
2773                 return true;
2774         }
2775         return false;
2776 }
2777 
2778 /* Process an event, which can update packets-in-flight not trivially.
2779  * Main goal of this function is to calculate new estimate for left_out,
2780  * taking into account both packets sitting in receiver's buffer and
2781  * packets lost by network.
2782  *
2783  * Besides that it does CWND reduction, when packet loss is detected
2784  * and changes state of machine.
2785  *
2786  * It does _not_ decide what to send, it is made in function
2787  * tcp_xmit_retransmit_queue().
2788  */
2789 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2790                                   const int prior_unsacked,
2791                                   bool is_dupack, int flag)
2792 {
2793         struct inet_connection_sock *icsk = inet_csk(sk);
2794         struct tcp_sock *tp = tcp_sk(sk);
2795         bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2796                                     (tcp_fackets_out(tp) > tp->reordering));
2797         int fast_rexmit = 0;
2798 
2799         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2800                 tp->sacked_out = 0;
2801         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2802                 tp->fackets_out = 0;
2803 
2804         /* Now state machine starts.
2805          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2806         if (flag & FLAG_ECE)
2807                 tp->prior_ssthresh = 0;
2808 
2809         /* B. In all the states check for reneging SACKs. */
2810         if (tcp_check_sack_reneging(sk, flag))
2811                 return;
2812 
2813         /* C. Check consistency of the current state. */
2814         tcp_verify_left_out(tp);
2815 
2816         /* D. Check state exit conditions. State can be terminated
2817          *    when high_seq is ACKed. */
2818         if (icsk->icsk_ca_state == TCP_CA_Open) {
2819                 WARN_ON(tp->retrans_out != 0);
2820                 tp->retrans_stamp = 0;
2821         } else if (!before(tp->snd_una, tp->high_seq)) {
2822                 switch (icsk->icsk_ca_state) {
2823                 case TCP_CA_CWR:
2824                         /* CWR is to be held something *above* high_seq
2825                          * is ACKed for CWR bit to reach receiver. */
2826                         if (tp->snd_una != tp->high_seq) {
2827                                 tcp_end_cwnd_reduction(sk);
2828                                 tcp_set_ca_state(sk, TCP_CA_Open);
2829                         }
2830                         break;
2831 
2832                 case TCP_CA_Recovery:
2833                         if (tcp_is_reno(tp))
2834                                 tcp_reset_reno_sack(tp);
2835                         if (tcp_try_undo_recovery(sk))
2836                                 return;
2837                         tcp_end_cwnd_reduction(sk);
2838                         break;
2839                 }
2840         }
2841 
2842         /* E. Process state. */
2843         switch (icsk->icsk_ca_state) {
2844         case TCP_CA_Recovery:
2845                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2846                         if (tcp_is_reno(tp) && is_dupack)
2847                                 tcp_add_reno_sack(sk);
2848                 } else {
2849                         if (tcp_try_undo_partial(sk, acked, prior_unsacked))
2850                                 return;
2851                         /* Partial ACK arrived. Force fast retransmit. */
2852                         do_lost = tcp_is_reno(tp) ||
2853                                   tcp_fackets_out(tp) > tp->reordering;
2854                 }
2855                 if (tcp_try_undo_dsack(sk)) {
2856                         tcp_try_keep_open(sk);
2857                         return;
2858                 }
2859                 break;
2860         case TCP_CA_Loss:
2861                 tcp_process_loss(sk, flag, is_dupack);
2862                 if (icsk->icsk_ca_state != TCP_CA_Open)
2863                         return;
2864                 /* Fall through to processing in Open state. */
2865         default:
2866                 if (tcp_is_reno(tp)) {
2867                         if (flag & FLAG_SND_UNA_ADVANCED)
2868                                 tcp_reset_reno_sack(tp);
2869                         if (is_dupack)
2870                                 tcp_add_reno_sack(sk);
2871                 }
2872 
2873                 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2874                         tcp_try_undo_dsack(sk);
2875 
2876                 if (!tcp_time_to_recover(sk, flag)) {
2877                         tcp_try_to_open(sk, flag, prior_unsacked);
2878                         return;
2879                 }
2880 
2881                 /* MTU probe failure: don't reduce cwnd */
2882                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2883                     icsk->icsk_mtup.probe_size &&
2884                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2885                         tcp_mtup_probe_failed(sk);
2886                         /* Restores the reduction we did in tcp_mtup_probe() */
2887                         tp->snd_cwnd++;
2888                         tcp_simple_retransmit(sk);
2889                         return;
2890                 }
2891 
2892                 /* Otherwise enter Recovery state */
2893                 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2894                 fast_rexmit = 1;
2895         }
2896 
2897         if (do_lost)
2898                 tcp_update_scoreboard(sk, fast_rexmit);
2899         tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
2900         tcp_xmit_retransmit_queue(sk);
2901 }
2902 
2903 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2904                                       long seq_rtt_us, long sack_rtt_us)
2905 {
2906         const struct tcp_sock *tp = tcp_sk(sk);
2907 
2908         /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2909          * broken middle-boxes or peers may corrupt TS-ECR fields. But
2910          * Karn's algorithm forbids taking RTT if some retransmitted data
2911          * is acked (RFC6298).
2912          */
2913         if (flag & FLAG_RETRANS_DATA_ACKED)
2914                 seq_rtt_us = -1L;
2915 
2916         if (seq_rtt_us < 0)
2917                 seq_rtt_us = sack_rtt_us;
2918 
2919         /* RTTM Rule: A TSecr value received in a segment is used to
2920          * update the averaged RTT measurement only if the segment
2921          * acknowledges some new data, i.e., only if it advances the
2922          * left edge of the send window.
2923          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2924          */
2925         if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2926             flag & FLAG_ACKED)
2927                 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2928 
2929         if (seq_rtt_us < 0)
2930                 return false;
2931 
2932         tcp_rtt_estimator(sk, seq_rtt_us);
2933         tcp_set_rto(sk);
2934 
2935         /* RFC6298: only reset backoff on valid RTT measurement. */
2936         inet_csk(sk)->icsk_backoff = 0;
2937         return true;
2938 }
2939 
2940 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2941 static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
2942 {
2943         struct tcp_sock *tp = tcp_sk(sk);
2944         long seq_rtt_us = -1L;
2945 
2946         if (synack_stamp && !tp->total_retrans)
2947                 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - synack_stamp);
2948 
2949         /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2950          * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2951          */
2952         if (!tp->srtt_us)
2953                 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt_us, -1L);
2954 }
2955 
2956 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2957 {
2958         const struct inet_connection_sock *icsk = inet_csk(sk);
2959 
2960         icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2961         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2962 }
2963 
2964 /* Restart timer after forward progress on connection.
2965  * RFC2988 recommends to restart timer to now+rto.
2966  */
2967 void tcp_rearm_rto(struct sock *sk)
2968 {
2969         const struct inet_connection_sock *icsk = inet_csk(sk);
2970         struct tcp_sock *tp = tcp_sk(sk);
2971 
2972         /* If the retrans timer is currently being used by Fast Open
2973          * for SYN-ACK retrans purpose, stay put.
2974          */
2975         if (tp->fastopen_rsk)
2976                 return;
2977 
2978         if (!tp->packets_out) {
2979                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2980         } else {
2981                 u32 rto = inet_csk(sk)->icsk_rto;
2982                 /* Offset the time elapsed after installing regular RTO */
2983                 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2984                     icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2985                         struct sk_buff *skb = tcp_write_queue_head(sk);
2986                         const u32 rto_time_stamp =
2987                                 tcp_skb_timestamp(skb) + rto;
2988                         s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2989                         /* delta may not be positive if the socket is locked
2990                          * when the retrans timer fires and is rescheduled.
2991                          */
2992                         rto = max(delta, 1);
2993                 }
2994                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2995                                           TCP_RTO_MAX);
2996         }
2997 }
2998 
2999 /* This function is called when the delayed ER timer fires. TCP enters
3000  * fast recovery and performs fast-retransmit.
3001  */
3002 void tcp_resume_early_retransmit(struct sock *sk)
3003 {
3004         struct tcp_sock *tp = tcp_sk(sk);
3005 
3006         tcp_rearm_rto(sk);
3007 
3008         /* Stop if ER is disabled after the delayed ER timer is scheduled */
3009         if (!tp->do_early_retrans)
3010                 return;
3011 
3012         tcp_enter_recovery(sk, false);
3013         tcp_update_scoreboard(sk, 1);
3014         tcp_xmit_retransmit_queue(sk);
3015 }
3016 
3017 /* If we get here, the whole TSO packet has not been acked. */
3018 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3019 {
3020         struct tcp_sock *tp = tcp_sk(sk);
3021         u32 packets_acked;
3022 
3023         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3024 
3025         packets_acked = tcp_skb_pcount(skb);
3026         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3027                 return 0;
3028         packets_acked -= tcp_skb_pcount(skb);
3029 
3030         if (packets_acked) {
3031                 BUG_ON(tcp_skb_pcount(skb) == 0);
3032                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3033         }
3034 
3035         return packets_acked;
3036 }
3037 
3038 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3039                            u32 prior_snd_una)
3040 {
3041         const struct skb_shared_info *shinfo;
3042 
3043         /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3044         if (likely(!(sk->sk_tsflags & SOF_TIMESTAMPING_TX_ACK)))
3045                 return;
3046 
3047         shinfo = skb_shinfo(skb);
3048         if ((shinfo->tx_flags & SKBTX_ACK_TSTAMP) &&
3049             between(shinfo->tskey, prior_snd_una, tcp_sk(sk)->snd_una - 1))
3050                 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3051 }
3052 
3053 /* Remove acknowledged frames from the retransmission queue. If our packet
3054  * is before the ack sequence we can discard it as it's confirmed to have
3055  * arrived at the other end.
3056  */
3057 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3058                                u32 prior_snd_una, long sack_rtt_us)
3059 {
3060         const struct inet_connection_sock *icsk = inet_csk(sk);
3061         struct skb_mstamp first_ackt, last_ackt, now;
3062         struct tcp_sock *tp = tcp_sk(sk);
3063         u32 prior_sacked = tp->sacked_out;
3064         u32 reord = tp->packets_out;
3065         bool fully_acked = true;
3066         long ca_seq_rtt_us = -1L;
3067         long seq_rtt_us = -1L;
3068         struct sk_buff *skb;
3069         u32 pkts_acked = 0;
3070         bool rtt_update;
3071         int flag = 0;
3072 
3073         first_ackt.v64 = 0;
3074 
3075         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3076                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3077                 u8 sacked = scb->sacked;
3078                 u32 acked_pcount;
3079 
3080                 tcp_ack_tstamp(sk, skb, prior_snd_una);
3081 
3082                 /* Determine how many packets and what bytes were acked, tso and else */
3083                 if (after(scb->end_seq, tp->snd_una)) {
3084                         if (tcp_skb_pcount(skb) == 1 ||
3085                             !after(tp->snd_una, scb->seq))
3086                                 break;
3087 
3088                         acked_pcount = tcp_tso_acked(sk, skb);
3089                         if (!acked_pcount)
3090                                 break;
3091 
3092                         fully_acked = false;
3093                 } else {
3094                         /* Speedup tcp_unlink_write_queue() and next loop */
3095                         prefetchw(skb->next);
3096                         acked_pcount = tcp_skb_pcount(skb);
3097                 }
3098 
3099                 if (unlikely(sacked & TCPCB_RETRANS)) {
3100                         if (sacked & TCPCB_SACKED_RETRANS)
3101                                 tp->retrans_out -= acked_pcount;
3102                         flag |= FLAG_RETRANS_DATA_ACKED;
3103                 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3104                         last_ackt = skb->skb_mstamp;
3105                         WARN_ON_ONCE(last_ackt.v64 == 0);
3106                         if (!first_ackt.v64)
3107                                 first_ackt = last_ackt;
3108 
3109                         reord = min(pkts_acked, reord);
3110                         if (!after(scb->end_seq, tp->high_seq))
3111                                 flag |= FLAG_ORIG_SACK_ACKED;
3112                 }
3113 
3114                 if (sacked & TCPCB_SACKED_ACKED)
3115                         tp->sacked_out -= acked_pcount;
3116                 if (sacked & TCPCB_LOST)
3117                         tp->lost_out -= acked_pcount;
3118 
3119                 tp->packets_out -= acked_pcount;
3120                 pkts_acked += acked_pcount;
3121 
3122                 /* Initial outgoing SYN's get put onto the write_queue
3123                  * just like anything else we transmit.  It is not
3124                  * true data, and if we misinform our callers that
3125                  * this ACK acks real data, we will erroneously exit
3126                  * connection startup slow start one packet too
3127                  * quickly.  This is severely frowned upon behavior.
3128                  */
3129                 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3130                         flag |= FLAG_DATA_ACKED;
3131                 } else {
3132                         flag |= FLAG_SYN_ACKED;
3133                         tp->retrans_stamp = 0;
3134                 }
3135 
3136                 if (!fully_acked)
3137                         break;
3138 
3139                 tcp_unlink_write_queue(skb, sk);
3140                 sk_wmem_free_skb(sk, skb);
3141                 if (unlikely(skb == tp->retransmit_skb_hint))
3142                         tp->retransmit_skb_hint = NULL;
3143                 if (unlikely(skb == tp->lost_skb_hint))
3144                         tp->lost_skb_hint = NULL;
3145         }
3146 
3147         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3148                 tp->snd_up = tp->snd_una;
3149 
3150         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3151                 flag |= FLAG_SACK_RENEGING;
3152 
3153         skb_mstamp_get(&now);
3154         if (likely(first_ackt.v64)) {
3155                 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3156                 ca_seq_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3157         }
3158 
3159         rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us);
3160 
3161         if (flag & FLAG_ACKED) {
3162                 const struct tcp_congestion_ops *ca_ops
3163                         = inet_csk(sk)->icsk_ca_ops;
3164 
3165                 tcp_rearm_rto(sk);
3166                 if (unlikely(icsk->icsk_mtup.probe_size &&
3167                              !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3168                         tcp_mtup_probe_success(sk);
3169                 }
3170 
3171                 if (tcp_is_reno(tp)) {
3172                         tcp_remove_reno_sacks(sk, pkts_acked);
3173                 } else {
3174                         int delta;
3175 
3176                         /* Non-retransmitted hole got filled? That's reordering */
3177                         if (reord < prior_fackets && reord <= tp->fackets_out)
3178                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3179 
3180                         delta = tcp_is_fack(tp) ? pkts_acked :
3181                                                   prior_sacked - tp->sacked_out;
3182                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3183                 }
3184 
3185                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3186 
3187                 if (ca_ops->pkts_acked) {
3188                         long rtt_us = min_t(ulong, ca_seq_rtt_us, sack_rtt_us);
3189                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3190                 }
3191 
3192         } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3193                    sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3194                 /* Do not re-arm RTO if the sack RTT is measured from data sent
3195                  * after when the head was last (re)transmitted. Otherwise the
3196                  * timeout may continue to extend in loss recovery.
3197                  */
3198                 tcp_rearm_rto(sk);
3199         }
3200 
3201 #if FASTRETRANS_DEBUG > 0
3202         WARN_ON((int)tp->sacked_out < 0);
3203         WARN_ON((int)tp->lost_out < 0);
3204         WARN_ON((int)tp->retrans_out < 0);
3205         if (!tp->packets_out && tcp_is_sack(tp)) {
3206                 icsk = inet_csk(sk);
3207                 if (tp->lost_out) {
3208                         pr_debug("Leak l=%u %d\n",
3209                                  tp->lost_out, icsk->icsk_ca_state);
3210                         tp->lost_out = 0;
3211                 }
3212                 if (tp->sacked_out) {
3213                         pr_debug("Leak s=%u %d\n",
3214                                  tp->sacked_out, icsk->icsk_ca_state);
3215                         tp->sacked_out = 0;
3216                 }
3217                 if (tp->retrans_out) {
3218                         pr_debug("Leak r=%u %d\n",
3219                                  tp->retrans_out, icsk->icsk_ca_state);
3220                         tp->retrans_out = 0;
3221                 }
3222         }
3223 #endif
3224         return flag;
3225 }
3226 
3227 static void tcp_ack_probe(struct sock *sk)
3228 {
3229         const struct tcp_sock *tp = tcp_sk(sk);
3230         struct inet_connection_sock *icsk = inet_csk(sk);
3231 
3232         /* Was it a usable window open? */
3233 
3234         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3235                 icsk->icsk_backoff = 0;
3236                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3237                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3238                  * This function is not for random using!
3239                  */
3240         } else {
3241                 unsigned long when = inet_csk_rto_backoff(icsk, TCP_RTO_MAX);
3242 
3243                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3244                                           when, TCP_RTO_MAX);
3245         }
3246 }
3247 
3248 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3249 {
3250         return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3251                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3252 }
3253 
3254 /* Decide wheather to run the increase function of congestion control. */
3255 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3256 {
3257         if (tcp_in_cwnd_reduction(sk))
3258                 return false;
3259 
3260         /* If reordering is high then always grow cwnd whenever data is
3261          * delivered regardless of its ordering. Otherwise stay conservative
3262          * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3263          * new SACK or ECE mark may first advance cwnd here and later reduce
3264          * cwnd in tcp_fastretrans_alert() based on more states.
3265          */
3266         if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3267                 return flag & FLAG_FORWARD_PROGRESS;
3268 
3269         return flag & FLAG_DATA_ACKED;
3270 }
3271 
3272 /* Check that window update is acceptable.
3273  * The function assumes that snd_una<=ack<=snd_next.
3274  */
3275 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3276                                         const u32 ack, const u32 ack_seq,
3277                                         const u32 nwin)
3278 {
3279         return  after(ack, tp->snd_una) ||
3280                 after(ack_seq, tp->snd_wl1) ||
3281                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3282 }
3283 
3284 /* If we update tp->snd_una, also update tp->bytes_acked */
3285 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3286 {
3287         u32 delta = ack - tp->snd_una;
3288 
3289         u64_stats_update_begin(&tp->syncp);
3290         tp->bytes_acked += delta;
3291         u64_stats_update_end(&tp->syncp);
3292         tp->snd_una = ack;
3293 }
3294 
3295 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3296 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3297 {
3298         u32 delta = seq - tp->rcv_nxt;
3299 
3300         u64_stats_update_begin(&tp->syncp);
3301         tp->bytes_received += delta;
3302         u64_stats_update_end(&tp->syncp);
3303         tp->rcv_nxt = seq;
3304 }
3305 
3306 /* Update our send window.
3307  *
3308  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3309  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3310  */
3311 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3312                                  u32 ack_seq)
3313 {
3314         struct tcp_sock *tp = tcp_sk(sk);
3315         int flag = 0;
3316         u32 nwin = ntohs(tcp_hdr(skb)->window);
3317 
3318         if (likely(!tcp_hdr(skb)->syn))
3319                 nwin <<= tp->rx_opt.snd_wscale;
3320 
3321         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3322                 flag |= FLAG_WIN_UPDATE;
3323                 tcp_update_wl(tp, ack_seq);
3324 
3325                 if (tp->snd_wnd != nwin) {
3326                         tp->snd_wnd = nwin;
3327 
3328                         /* Note, it is the only place, where
3329                          * fast path is recovered for sending TCP.
3330                          */
3331                         tp->pred_flags = 0;
3332                         tcp_fast_path_check(sk);
3333 
3334                         if (nwin > tp->max_window) {
3335                                 tp->max_window = nwin;
3336                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3337                         }
3338                 }
3339         }
3340 
3341         tcp_snd_una_update(tp, ack);
3342 
3343         return flag;
3344 }
3345 
3346 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3347                                    u32 *last_oow_ack_time)
3348 {
3349         if (*last_oow_ack_time) {
3350                 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3351 
3352                 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3353                         NET_INC_STATS(net, mib_idx);
3354                         return true;    /* rate-limited: don't send yet! */
3355                 }
3356         }
3357 
3358         *last_oow_ack_time = tcp_time_stamp;
3359 
3360         return false;   /* not rate-limited: go ahead, send dupack now! */
3361 }
3362 
3363 /* Return true if we're currently rate-limiting out-of-window ACKs and
3364  * thus shouldn't send a dupack right now. We rate-limit dupacks in
3365  * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3366  * attacks that send repeated SYNs or ACKs for the same connection. To
3367  * do this, we do not send a duplicate SYNACK or ACK if the remote
3368  * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3369  */
3370 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3371                           int mib_idx, u32 *last_oow_ack_time)
3372 {
3373         /* Data packets without SYNs are not likely part of an ACK loop. */
3374         if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3375             !tcp_hdr(skb)->syn)
3376                 return false;
3377 
3378         return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3379 }
3380 
3381 /* RFC 5961 7 [ACK Throttling] */
3382 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3383 {
3384         /* unprotected vars, we dont care of overwrites */
3385         static u32 challenge_timestamp;
3386         static unsigned int challenge_count;
3387         struct tcp_sock *tp = tcp_sk(sk);
3388         u32 count, now;
3389 
3390         /* First check our per-socket dupack rate limit. */
3391         if (__tcp_oow_rate_limited(sock_net(sk),
3392                                    LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3393                                    &tp->last_oow_ack_time))
3394                 return;
3395 
3396         /* Then check host-wide RFC 5961 rate limit. */
3397         now = jiffies / HZ;
3398         if (now != challenge_timestamp) {
3399                 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3400 
3401                 challenge_timestamp = now;
3402                 WRITE_ONCE(challenge_count, half +
3403                            prandom_u32_max(sysctl_tcp_challenge_ack_limit));
3404         }
3405         count = READ_ONCE(challenge_count);
3406         if (count > 0) {
3407                 WRITE_ONCE(challenge_count, count - 1);
3408                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3409                 tcp_send_ack(sk);
3410         }
3411 }
3412 
3413 static void tcp_store_ts_recent(struct tcp_sock *tp)
3414 {
3415         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3416         tp->rx_opt.ts_recent_stamp = get_seconds();
3417 }
3418 
3419 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3420 {
3421         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3422                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3423                  * extra check below makes sure this can only happen
3424                  * for pure ACK frames.  -DaveM
3425                  *
3426                  * Not only, also it occurs for expired timestamps.
3427                  */
3428 
3429                 if (tcp_paws_check(&tp->rx_opt, 0))
3430                         tcp_store_ts_recent(tp);
3431         }
3432 }
3433 
3434 /* This routine deals with acks during a TLP episode.
3435  * We mark the end of a TLP episode on receiving TLP dupack or when
3436  * ack is after tlp_high_seq.
3437  * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3438  */
3439 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3440 {
3441         struct tcp_sock *tp = tcp_sk(sk);
3442 
3443         if (before(ack, tp->tlp_high_seq))
3444                 return;
3445 
3446         if (flag & FLAG_DSACKING_ACK) {
3447                 /* This DSACK means original and TLP probe arrived; no loss */
3448                 tp->tlp_high_seq = 0;
3449         } else if (after(ack, tp->tlp_high_seq)) {
3450                 /* ACK advances: there was a loss, so reduce cwnd. Reset
3451                  * tlp_high_seq in tcp_init_cwnd_reduction()
3452                  */
3453                 tcp_init_cwnd_reduction(sk);
3454                 tcp_set_ca_state(sk, TCP_CA_CWR);
3455                 tcp_end_cwnd_reduction(sk);
3456                 tcp_try_keep_open(sk);
3457                 NET_INC_STATS_BH(sock_net(sk),
3458                                  LINUX_MIB_TCPLOSSPROBERECOVERY);
3459         } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3460                              FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3461                 /* Pure dupack: original and TLP probe arrived; no loss */
3462                 tp->tlp_high_seq = 0;
3463         }
3464 }
3465 
3466 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3467 {
3468         const struct inet_connection_sock *icsk = inet_csk(sk);
3469 
3470         if (icsk->icsk_ca_ops->in_ack_event)
3471                 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3472 }
3473 
3474 /* This routine deals with incoming acks, but not outgoing ones. */
3475 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3476 {
3477         struct inet_connection_sock *icsk = inet_csk(sk);
3478         struct tcp_sock *tp = tcp_sk(sk);
3479         u32 prior_snd_una = tp->snd_una;
3480         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3481         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3482         bool is_dupack = false;
3483         u32 prior_fackets;
3484         int prior_packets = tp->packets_out;
3485         const int prior_unsacked = tp->packets_out - tp->sacked_out;
3486         int acked = 0; /* Number of packets newly acked */
3487         long sack_rtt_us = -1L;
3488 
3489         /* We very likely will need to access write queue head. */
3490         prefetchw(sk->sk_write_queue.next);
3491 
3492         /* If the ack is older than previous acks
3493          * then we can probably ignore it.
3494          */
3495         if (before(ack, prior_snd_una)) {
3496                 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3497                 if (before(ack, prior_snd_una - tp->max_window)) {
3498                         tcp_send_challenge_ack(sk, skb);
3499                         return -1;
3500                 }
3501                 goto old_ack;
3502         }
3503 
3504         /* If the ack includes data we haven't sent yet, discard
3505          * this segment (RFC793 Section 3.9).
3506          */
3507         if (after(ack, tp->snd_nxt))
3508                 goto invalid_ack;
3509 
3510         if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3511             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3512                 tcp_rearm_rto(sk);
3513 
3514         if (after(ack, prior_snd_una)) {
3515                 flag |= FLAG_SND_UNA_ADVANCED;
3516                 icsk->icsk_retransmits = 0;
3517         }
3518 
3519         prior_fackets = tp->fackets_out;
3520 
3521         /* ts_recent update must be made after we are sure that the packet
3522          * is in window.
3523          */
3524         if (flag & FLAG_UPDATE_TS_RECENT)
3525                 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3526 
3527         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3528                 /* Window is constant, pure forward advance.
3529                  * No more checks are required.
3530                  * Note, we use the fact that SND.UNA>=SND.WL2.
3531                  */
3532                 tcp_update_wl(tp, ack_seq);
3533                 tcp_snd_una_update(tp, ack);
3534                 flag |= FLAG_WIN_UPDATE;
3535 
3536                 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3537 
3538                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3539         } else {
3540                 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3541 
3542                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3543                         flag |= FLAG_DATA;
3544                 else
3545                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3546 
3547                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3548 
3549                 if (TCP_SKB_CB(skb)->sacked)
3550                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3551                                                         &sack_rtt_us);
3552 
3553                 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3554                         flag |= FLAG_ECE;
3555                         ack_ev_flags |= CA_ACK_ECE;
3556                 }
3557 
3558                 if (flag & FLAG_WIN_UPDATE)
3559                         ack_ev_flags |= CA_ACK_WIN_UPDATE;
3560 
3561                 tcp_in_ack_event(sk, ack_ev_flags);
3562         }
3563 
3564         /* We passed data and got it acked, remove any soft error
3565          * log. Something worked...
3566          */
3567         sk->sk_err_soft = 0;
3568         icsk->icsk_probes_out = 0;
3569         tp->rcv_tstamp = tcp_time_stamp;
3570         if (!prior_packets)
3571                 goto no_queue;
3572 
3573         /* See if we can take anything off of the retransmit queue. */
3574         acked = tp->packets_out;
3575         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
3576                                     sack_rtt_us);
3577         acked -= tp->packets_out;
3578 
3579         /* Advance cwnd if state allows */
3580         if (tcp_may_raise_cwnd(sk, flag))
3581                 tcp_cong_avoid(sk, ack, acked);
3582 
3583         if (tcp_ack_is_dubious(sk, flag)) {
3584                 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3585                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3586                                       is_dupack, flag);
3587         }
3588         if (tp->tlp_high_seq)
3589                 tcp_process_tlp_ack(sk, ack, flag);
3590 
3591         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3592                 struct dst_entry *dst = __sk_dst_get(sk);
3593                 if (dst)
3594                         dst_confirm(dst);
3595         }
3596 
3597         if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3598                 tcp_schedule_loss_probe(sk);
3599         tcp_update_pacing_rate(sk);
3600         return 1;
3601 
3602 no_queue:
3603         /* If data was DSACKed, see if we can undo a cwnd reduction. */
3604         if (flag & FLAG_DSACKING_ACK)
3605                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3606                                       is_dupack, flag);
3607         /* If this ack opens up a zero window, clear backoff.  It was
3608          * being used to time the probes, and is probably far higher than
3609          * it needs to be for normal retransmission.
3610          */
3611         if (tcp_send_head(sk))
3612                 tcp_ack_probe(sk);
3613 
3614         if (tp->tlp_high_seq)
3615                 tcp_process_tlp_ack(sk, ack, flag);
3616         return 1;
3617 
3618 invalid_ack:
3619         SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3620         return -1;
3621 
3622 old_ack:
3623         /* If data was SACKed, tag it and see if we should send more data.
3624          * If data was DSACKed, see if we can undo a cwnd reduction.
3625          */
3626         if (TCP_SKB_CB(skb)->sacked) {
3627                 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3628                                                 &sack_rtt_us);
3629                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3630                                       is_dupack, flag);
3631         }
3632 
3633         SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3634         return 0;
3635 }
3636 
3637 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3638                                       bool syn, struct tcp_fastopen_cookie *foc,
3639                                       bool exp_opt)
3640 {
3641         /* Valid only in SYN or SYN-ACK with an even length.  */
3642         if (!foc || !syn || len < 0 || (len & 1))
3643                 return;
3644 
3645         if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3646             len <= TCP_FASTOPEN_COOKIE_MAX)
3647                 memcpy(foc->val, cookie, len);
3648         else if (len != 0)
3649                 len = -1;
3650         foc->len = len;
3651         foc->exp = exp_opt;
3652 }
3653 
3654 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3655  * But, this can also be called on packets in the established flow when
3656  * the fast version below fails.
3657  */
3658 void tcp_parse_options(const struct sk_buff *skb,
3659                        struct tcp_options_received *opt_rx, int estab,
3660                        struct tcp_fastopen_cookie *foc)
3661 {
3662         const unsigned char *ptr;
3663         const struct tcphdr *th = tcp_hdr(skb);
3664         int length = (th->doff * 4) - sizeof(struct tcphdr);
3665 
3666         ptr = (const unsigned char *)(th + 1);
3667         opt_rx->saw_tstamp = 0;
3668 
3669         while (length > 0) {
3670                 int opcode = *ptr++;
3671                 int opsize;
3672 
3673                 switch (opcode) {
3674                 case TCPOPT_EOL:
3675                         return;
3676                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3677                         length--;
3678                         continue;
3679                 default:
3680                         opsize = *ptr++;
3681                         if (opsize < 2) /* "silly options" */
3682                                 return;
3683                         if (opsize > length)
3684                                 return; /* don't parse partial options */
3685                         switch (opcode) {
3686                         case TCPOPT_MSS:
3687                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3688                                         u16 in_mss = get_unaligned_be16(ptr);
3689                                         if (in_mss) {
3690                                                 if (opt_rx->user_mss &&
3691                                                     opt_rx->user_mss < in_mss)
3692                                                         in_mss = opt_rx->user_mss;
3693                                                 opt_rx->mss_clamp = in_mss;
3694                                         }
3695                                 }
3696                                 break;
3697                         case TCPOPT_WINDOW:
3698                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3699                                     !estab && sysctl_tcp_window_scaling) {
3700                                         __u8 snd_wscale = *(__u8 *)ptr;
3701                                         opt_rx->wscale_ok = 1;
3702                                         if (snd_wscale > 14) {
3703                                                 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3704                                                                      __func__,
3705                                                                      snd_wscale);
3706                                                 snd_wscale = 14;
3707                                         }
3708                                         opt_rx->snd_wscale = snd_wscale;
3709                                 }
3710                                 break;
3711                         case TCPOPT_TIMESTAMP:
3712                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3713                                     ((estab && opt_rx->tstamp_ok) ||
3714                                      (!estab && sysctl_tcp_timestamps))) {
3715                                         opt_rx->saw_tstamp = 1;
3716                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3717                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3718                                 }
3719                                 break;
3720                         case TCPOPT_SACK_PERM:
3721                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3722                                     !estab && sysctl_tcp_sack) {
3723                                         opt_rx->sack_ok = TCP_SACK_SEEN;
3724                                         tcp_sack_reset(opt_rx);
3725                                 }
3726                                 break;
3727 
3728                         case TCPOPT_SACK:
3729                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3730                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3731                                    opt_rx->sack_ok) {
3732                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3733                                 }
3734                                 break;
3735 #ifdef CONFIG_TCP_MD5SIG
3736                         case TCPOPT_MD5SIG:
3737                                 /*
3738                                  * The MD5 Hash has already been
3739                                  * checked (see tcp_v{4,6}_do_rcv()).
3740                                  */
3741                                 break;
3742 #endif
3743                         case TCPOPT_FASTOPEN:
3744                                 tcp_parse_fastopen_option(
3745                                         opsize - TCPOLEN_FASTOPEN_BASE,
3746                                         ptr, th->syn, foc, false);
3747                                 break;
3748 
3749                         case TCPOPT_EXP:
3750                                 /* Fast Open option shares code 254 using a
3751                                  * 16 bits magic number.
3752                                  */
3753                                 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3754                                     get_unaligned_be16(ptr) ==
3755                                     TCPOPT_FASTOPEN_MAGIC)
3756                                         tcp_parse_fastopen_option(opsize -
3757                                                 TCPOLEN_EXP_FASTOPEN_BASE,
3758                                                 ptr + 2, th->syn, foc, true);
3759                                 break;
3760 
3761                         }
3762                         ptr += opsize-2;
3763                         length -= opsize;
3764                 }
3765         }
3766 }
3767 EXPORT_SYMBOL(tcp_parse_options);
3768 
3769 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3770 {
3771         const __be32 *ptr = (const __be32 *)(th + 1);
3772 
3773         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3774                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3775                 tp->rx_opt.saw_tstamp = 1;
3776                 ++ptr;
3777                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3778                 ++ptr;
3779                 if (*ptr)
3780                         tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3781                 else
3782                         tp->rx_opt.rcv_tsecr = 0;
3783                 return true;
3784         }
3785         return false;
3786 }
3787 
3788 /* Fast parse options. This hopes to only see timestamps.
3789  * If it is wrong it falls back on tcp_parse_options().
3790  */
3791 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3792                                    const struct tcphdr *th, struct tcp_sock *tp)
3793 {
3794         /* In the spirit of fast parsing, compare doff directly to constant
3795          * values.  Because equality is used, short doff can be ignored here.
3796          */
3797         if (th->doff == (sizeof(*th) / 4)) {
3798                 tp->rx_opt.saw_tstamp = 0;
3799                 return false;
3800         } else if (tp->rx_opt.tstamp_ok &&
3801                    th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3802                 if (tcp_parse_aligned_timestamp(tp, th))
3803                         return true;
3804         }
3805 
3806         tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3807         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3808                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3809 
3810         return true;
3811 }
3812 
3813 #ifdef CONFIG_TCP_MD5SIG
3814 /*
3815  * Parse MD5 Signature option
3816  */
3817 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3818 {
3819         int length = (th->doff << 2) - sizeof(*th);
3820         const u8 *ptr = (const u8 *)(th + 1);
3821 
3822         /* If not enough data remaining, we can short cut */
3823         while (length >= TCPOLEN_MD5SIG) {
3824                 int opcode = *ptr++;
3825                 int opsize;
3826 
3827                 switch (opcode) {
3828                 case TCPOPT_EOL:
3829                         return NULL;
3830                 case TCPOPT_NOP:
3831                         length--;
3832                         continue;
3833                 default:
3834                         opsize = *ptr++;
3835                         if (opsize < 2 || opsize > length)
3836                                 return NULL;
3837                         if (opcode == TCPOPT_MD5SIG)
3838                                 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3839                 }
3840                 ptr += opsize - 2;
3841                 length -= opsize;
3842         }
3843         return NULL;
3844 }
3845 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3846 #endif
3847 
3848 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3849  *
3850  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3851  * it can pass through stack. So, the following predicate verifies that
3852  * this segment is not used for anything but congestion avoidance or
3853  * fast retransmit. Moreover, we even are able to eliminate most of such
3854  * second order effects, if we apply some small "replay" window (~RTO)
3855  * to timestamp space.
3856  *
3857  * All these measures still do not guarantee that we reject wrapped ACKs
3858  * on networks with high bandwidth, when sequence space is recycled fastly,
3859  * but it guarantees that such events will be very rare and do not affect
3860  * connection seriously. This doesn't look nice, but alas, PAWS is really
3861  * buggy extension.
3862  *
3863  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3864  * states that events when retransmit arrives after original data are rare.
3865  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3866  * the biggest problem on large power networks even with minor reordering.
3867  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3868  * up to bandwidth of 18Gigabit/sec. 8) ]
3869  */
3870 
3871 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3872 {
3873         const struct tcp_sock *tp = tcp_sk(sk);
3874         const struct tcphdr *th = tcp_hdr(skb);
3875         u32 seq = TCP_SKB_CB(skb)->seq;
3876         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3877 
3878         return (/* 1. Pure ACK with correct sequence number. */
3879                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3880 
3881                 /* 2. ... and duplicate ACK. */
3882                 ack == tp->snd_una &&
3883 
3884                 /* 3. ... and does not update window. */
3885                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3886 
3887                 /* 4. ... and sits in replay window. */
3888                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3889 }
3890 
3891 static inline bool tcp_paws_discard(const struct sock *sk,
3892                                    const struct sk_buff *skb)
3893 {
3894         const struct tcp_sock *tp = tcp_sk(sk);
3895 
3896         return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3897                !tcp_disordered_ack(sk, skb);
3898 }
3899 
3900 /* Check segment sequence number for validity.
3901  *
3902  * Segment controls are considered valid, if the segment
3903  * fits to the window after truncation to the window. Acceptability
3904  * of data (and SYN, FIN, of course) is checked separately.
3905  * See tcp_data_queue(), for example.
3906  *
3907  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3908  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3909  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3910  * (borrowed from freebsd)
3911  */
3912 
3913 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3914 {
3915         return  !before(end_seq, tp->rcv_wup) &&
3916                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3917 }
3918 
3919 /* When we get a reset we do this. */
3920 void tcp_reset(struct sock *sk)
3921 {
3922         /* We want the right error as BSD sees it (and indeed as we do). */
3923         switch (sk->sk_state) {
3924         case TCP_SYN_SENT:
3925                 sk->sk_err = ECONNREFUSED;
3926                 break;
3927         case TCP_CLOSE_WAIT:
3928                 sk->sk_err = EPIPE;
3929                 break;
3930         case TCP_CLOSE:
3931                 return;
3932         default:
3933                 sk->sk_err = ECONNRESET;
3934         }
3935         /* This barrier is coupled with smp_rmb() in tcp_poll() */
3936         smp_wmb();
3937 
3938         if (!sock_flag(sk, SOCK_DEAD))
3939                 sk->sk_error_report(sk);
3940 
3941         tcp_done(sk);
3942 }
3943 
3944 /*
3945  *      Process the FIN bit. This now behaves as it is supposed to work
3946  *      and the FIN takes effect when it is validly part of sequence
3947  *      space. Not before when we get holes.
3948  *
3949  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3950  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3951  *      TIME-WAIT)
3952  *
3953  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3954  *      close and we go into CLOSING (and later onto TIME-WAIT)
3955  *
3956  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3957  */
3958 static void tcp_fin(struct sock *sk)
3959 {
3960         struct tcp_sock *tp = tcp_sk(sk);
3961         const struct dst_entry *dst;
3962 
3963         inet_csk_schedule_ack(sk);
3964 
3965         sk->sk_shutdown |= RCV_SHUTDOWN;
3966         sock_set_flag(sk, SOCK_DONE);
3967 
3968         switch (sk->sk_state) {
3969         case TCP_SYN_RECV:
3970         case TCP_ESTABLISHED:
3971                 /* Move to CLOSE_WAIT */
3972                 tcp_set_state(sk, TCP_CLOSE_WAIT);
3973                 dst = __sk_dst_get(sk);
3974                 if (!dst || !dst_metric(dst, RTAX_QUICKACK))
3975                         inet_csk(sk)->icsk_ack.pingpong = 1;
3976                 break;
3977 
3978         case TCP_CLOSE_WAIT:
3979         case TCP_CLOSING:
3980                 /* Received a retransmission of the FIN, do
3981                  * nothing.
3982                  */
3983                 break;
3984         case TCP_LAST_ACK:
3985                 /* RFC793: Remain in the LAST-ACK state. */
3986                 break;
3987 
3988         case TCP_FIN_WAIT1:
3989                 /* This case occurs when a simultaneous close
3990                  * happens, we must ack the received FIN and
3991                  * enter the CLOSING state.
3992                  */
3993                 tcp_send_ack(sk);
3994                 tcp_set_state(sk, TCP_CLOSING);
3995                 break;
3996         case TCP_FIN_WAIT2:
3997                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3998                 tcp_send_ack(sk);
3999                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4000                 break;
4001         default:
4002                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4003                  * cases we should never reach this piece of code.
4004                  */
4005                 pr_err("%s: Impossible, sk->sk_state=%d\n",
4006                        __func__, sk->sk_state);
4007                 break;
4008         }
4009 
4010         /* It _is_ possible, that we have something out-of-order _after_ FIN.
4011          * Probably, we should reset in this case. For now drop them.
4012          */
4013         __skb_queue_purge(&tp->out_of_order_queue);
4014         if (tcp_is_sack(tp))
4015                 tcp_sack_reset(&tp->rx_opt);
4016         sk_mem_reclaim(sk);
4017 
4018         if (!sock_flag(sk, SOCK_DEAD)) {
4019                 sk->sk_state_change(sk);
4020 
4021                 /* Do not send POLL_HUP for half duplex close. */
4022                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4023                     sk->sk_state == TCP_CLOSE)
4024                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4025                 else
4026                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4027         }
4028 }
4029 
4030 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4031                                   u32 end_seq)
4032 {
4033         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4034                 if (before(seq, sp->start_seq))
4035                         sp->start_seq = seq;
4036                 if (after(end_seq, sp->end_seq))
4037                         sp->end_seq = end_seq;
4038                 return true;
4039         }
4040         return false;
4041 }
4042 
4043 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4044 {
4045         struct tcp_sock *tp = tcp_sk(sk);
4046 
4047         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4048                 int mib_idx;
4049 
4050                 if (before(seq, tp->rcv_nxt))
4051                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4052                 else
4053                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4054 
4055                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4056 
4057                 tp->rx_opt.dsack = 1;
4058                 tp->duplicate_sack[0].start_seq = seq;
4059                 tp->duplicate_sack[0].end_seq = end_seq;
4060         }
4061 }
4062 
4063 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4064 {
4065         struct tcp_sock *tp = tcp_sk(sk);
4066 
4067         if (!tp->rx_opt.dsack)
4068                 tcp_dsack_set(sk, seq, end_seq);
4069         else
4070                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4071 }
4072 
4073 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4074 {
4075         struct tcp_sock *tp = tcp_sk(sk);
4076 
4077         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4078             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4079                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4080                 tcp_enter_quickack_mode(sk);
4081 
4082                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4083                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4084 
4085                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4086                                 end_seq = tp->rcv_nxt;
4087                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4088                 }
4089         }
4090 
4091         tcp_send_ack(sk);
4092 }
4093 
4094 /* These routines update the SACK block as out-of-order packets arrive or
4095  * in-order packets close up the sequence space.
4096  */
4097 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4098 {
4099         int this_sack;
4100         struct tcp_sack_block *sp = &tp->selective_acks[0];
4101         struct tcp_sack_block *swalk = sp + 1;
4102 
4103         /* See if the recent change to the first SACK eats into
4104          * or hits the sequence space of other SACK blocks, if so coalesce.
4105          */
4106         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4107                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4108                         int i;
4109 
4110                         /* Zap SWALK, by moving every further SACK up by one slot.
4111                          * Decrease num_sacks.
4112                          */
4113                         tp->rx_opt.num_sacks--;
4114                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4115                                 sp[i] = sp[i + 1];
4116                         continue;
4117                 }
4118                 this_sack++, swalk++;
4119         }
4120 }
4121 
4122 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4123 {
4124         struct tcp_sock *tp = tcp_sk(sk);
4125         struct tcp_sack_block *sp = &tp->selective_acks[0];
4126         int cur_sacks = tp->rx_opt.num_sacks;
4127         int this_sack;
4128 
4129         if (!cur_sacks)
4130                 goto new_sack;
4131 
4132         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4133                 if (tcp_sack_extend(sp, seq, end_seq)) {
4134                         /* Rotate this_sack to the first one. */
4135                         for (; this_sack > 0; this_sack--, sp--)
4136                                 swap(*sp, *(sp - 1));
4137                         if (cur_sacks > 1)
4138                                 tcp_sack_maybe_coalesce(tp);
4139                         return;
4140                 }
4141         }
4142 
4143         /* Could not find an adjacent existing SACK, build a new one,
4144          * put it at the front, and shift everyone else down.  We
4145          * always know there is at least one SACK present already here.
4146          *
4147          * If the sack array is full, forget about the last one.
4148          */
4149         if (this_sack >= TCP_NUM_SACKS) {
4150                 this_sack--;
4151                 tp->rx_opt.num_sacks--;
4152                 sp--;
4153         }
4154         for (; this_sack > 0; this_sack--, sp--)
4155                 *sp = *(sp - 1);
4156 
4157 new_sack:
4158         /* Build the new head SACK, and we're done. */
4159         sp->start_seq = seq;
4160         sp->end_seq = end_seq;
4161         tp->rx_opt.num_sacks++;
4162 }
4163 
4164 /* RCV.NXT advances, some SACKs should be eaten. */
4165 
4166 static void tcp_sack_remove(struct tcp_sock *tp)
4167 {
4168         struct tcp_sack_block *sp = &tp->selective_acks[0];
4169         int num_sacks = tp->rx_opt.num_sacks;
4170         int this_sack;
4171 
4172         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4173         if (skb_queue_empty(&tp->out_of_order_queue)) {
4174                 tp->rx_opt.num_sacks = 0;
4175                 return;
4176         }
4177 
4178         for (this_sack = 0; this_sack < num_sacks;) {
4179                 /* Check if the start of the sack is covered by RCV.NXT. */
4180                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4181                         int i;
4182 
4183                         /* RCV.NXT must cover all the block! */
4184                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4185 
4186                         /* Zap this SACK, by moving forward any other SACKS. */
4187                         for (i = this_sack+1; i < num_sacks; i++)
4188                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4189                         num_sacks--;
4190                         continue;
4191                 }
4192                 this_sack++;
4193                 sp++;
4194         }
4195         tp->rx_opt.num_sacks = num_sacks;
4196 }
4197 
4198 /**
4199  * tcp_try_coalesce - try to merge skb to prior one
4200  * @sk: socket
4201  * @to: prior buffer
4202  * @from: buffer to add in queue
4203  * @fragstolen: pointer to boolean
4204  *
4205  * Before queueing skb @from after @to, try to merge them
4206  * to reduce overall memory use and queue lengths, if cost is small.
4207  * Packets in ofo or receive queues can stay a long time.
4208  * Better try to coalesce them right now to avoid future collapses.
4209  * Returns true if caller should free @from instead of queueing it
4210  */
4211 static bool tcp_try_coalesce(struct sock *sk,
4212                              struct sk_buff *to,
4213                              struct sk_buff *from,
4214                              bool *fragstolen)
4215 {
4216         int delta;
4217 
4218         *fragstolen = false;
4219 
4220         /* Its possible this segment overlaps with prior segment in queue */
4221         if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4222                 return false;
4223 
4224         if (!skb_try_coalesce(to, from, fragstolen, &delta))
4225                 return false;
4226 
4227         atomic_add(delta, &sk->sk_rmem_alloc);
4228         sk_mem_charge(sk, delta);
4229         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4230         TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4231         TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4232         TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4233         return true;
4234 }
4235 
4236 /* This one checks to see if we can put data from the
4237  * out_of_order queue into the receive_queue.
4238  */
4239 static void tcp_ofo_queue(struct sock *sk)
4240 {
4241         struct tcp_sock *tp = tcp_sk(sk);
4242         __u32 dsack_high = tp->rcv_nxt;
4243         struct sk_buff *skb, *tail;
4244         bool fragstolen, eaten;
4245 
4246         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4247                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4248                         break;
4249 
4250                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4251                         __u32 dsack = dsack_high;
4252                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4253                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4254                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4255                 }
4256 
4257                 __skb_unlink(skb, &tp->out_of_order_queue);
4258                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4259                         SOCK_DEBUG(sk, "ofo packet was already received\n");
4260                         __kfree_skb(skb);
4261                         continue;
4262                 }
4263                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4264                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4265                            TCP_SKB_CB(skb)->end_seq);
4266 
4267                 tail = skb_peek_tail(&sk->sk_receive_queue);
4268                 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4269                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4270                 if (!eaten)
4271                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4272                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4273                         tcp_fin(sk);
4274                 if (eaten)
4275                         kfree_skb_partial(skb, fragstolen);
4276         }
4277 }
4278 
4279 static bool tcp_prune_ofo_queue(struct sock *sk);
4280 static int tcp_prune_queue(struct sock *sk);
4281 
4282 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4283                                  unsigned int size)
4284 {
4285         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4286             !sk_rmem_schedule(sk, skb, size)) {
4287 
4288                 if (tcp_prune_queue(sk) < 0)
4289                         return -1;
4290 
4291                 if (!sk_rmem_schedule(sk, skb, size)) {
4292                         if (!tcp_prune_ofo_queue(sk))
4293                                 return -1;
4294 
4295                         if (!sk_rmem_schedule(sk, skb, size))
4296                                 return -1;
4297                 }
4298         }
4299         return 0;
4300 }
4301 
4302 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4303 {
4304         struct tcp_sock *tp = tcp_sk(sk);
4305         struct sk_buff *skb1;
4306         u32 seq, end_seq;
4307 
4308         tcp_ecn_check_ce(tp, skb);
4309 
4310         if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4311                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4312                 __kfree_skb(skb);
4313                 return;
4314         }
4315 
4316         /* Disable header prediction. */
4317         tp->pred_flags = 0;
4318         inet_csk_schedule_ack(sk);
4319 
4320         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4321         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4322                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4323 
4324         skb1 = skb_peek_tail(&tp->out_of_order_queue);
4325         if (!skb1) {
4326                 /* Initial out of order segment, build 1 SACK. */
4327                 if (tcp_is_sack(tp)) {
4328                         tp->rx_opt.num_sacks = 1;
4329                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4330                         tp->selective_acks[0].end_seq =
4331                                                 TCP_SKB_CB(skb)->end_seq;
4332                 }
4333                 __skb_queue_head(&tp->out_of_order_queue, skb);
4334                 goto end;
4335         }
4336 
4337         seq = TCP_SKB_CB(skb)->seq;
4338         end_seq = TCP_SKB_CB(skb)->end_seq;
4339 
4340         if (seq == TCP_SKB_CB(skb1)->end_seq) {
4341                 bool fragstolen;
4342 
4343                 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4344                         __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4345                 } else {
4346                         tcp_grow_window(sk, skb);
4347                         kfree_skb_partial(skb, fragstolen);
4348                         skb = NULL;
4349                 }
4350 
4351                 if (!tp->rx_opt.num_sacks ||
4352                     tp->selective_acks[0].end_seq != seq)
4353                         goto add_sack;
4354 
4355                 /* Common case: data arrive in order after hole. */
4356                 tp->selective_acks[0].end_seq = end_seq;
4357                 goto end;
4358         }
4359 
4360         /* Find place to insert this segment. */
4361         while (1) {
4362                 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4363                         break;
4364                 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4365                         skb1 = NULL;
4366                         break;
4367                 }
4368                 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4369         }
4370 
4371         /* Do skb overlap to previous one? */
4372         if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4373                 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4374                         /* All the bits are present. Drop. */
4375                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4376                         __kfree_skb(skb);
4377                         skb = NULL;
4378                         tcp_dsack_set(sk, seq, end_seq);
4379                         goto add_sack;
4380                 }
4381                 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4382                         /* Partial overlap. */
4383                         tcp_dsack_set(sk, seq,
4384                                       TCP_SKB_CB(skb1)->end_seq);
4385                 } else {
4386                         if (skb_queue_is_first(&tp->out_of_order_queue,
4387                                                skb1))
4388                                 skb1 = NULL;
4389                         else
4390                                 skb1 = skb_queue_prev(
4391                                         &tp->out_of_order_queue,
4392                                         skb1);
4393                 }
4394         }
4395         if (!skb1)
4396                 __skb_queue_head(&tp->out_of_order_queue, skb);
4397         else
4398                 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4399 
4400         /* And clean segments covered by new one as whole. */
4401         while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4402                 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4403 
4404                 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4405                         break;
4406                 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4407                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4408                                          end_seq);
4409                         break;
4410                 }
4411                 __skb_unlink(skb1, &tp->out_of_order_queue);
4412                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4413                                  TCP_SKB_CB(skb1)->end_seq);
4414                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4415                 __kfree_skb(skb1);
4416         }
4417 
4418 add_sack:
4419         if (tcp_is_sack(tp))
4420                 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4421 end:
4422         if (skb) {
4423                 tcp_grow_window(sk, skb);
4424                 skb_set_owner_r(skb, sk);
4425         }
4426 }
4427 
4428 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4429                   bool *fragstolen)
4430 {
4431         int eaten;
4432         struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4433 
4434         __skb_pull(skb, hdrlen);
4435         eaten = (tail &&
4436                  tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4437         tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4438         if (!eaten) {
4439                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4440                 skb_set_owner_r(skb, sk);
4441         }
4442         return eaten;
4443 }
4444 
4445 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4446 {
4447         struct sk_buff *skb;
4448         int err = -ENOMEM;
4449         int data_len = 0;
4450         bool fragstolen;
4451 
4452         if (size == 0)
4453                 return 0;
4454 
4455         if (size > PAGE_SIZE) {
4456                 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4457 
4458                 data_len = npages << PAGE_SHIFT;
4459                 size = data_len + (size & ~PAGE_MASK);
4460         }
4461         skb = alloc_skb_with_frags(size - data_len, data_len,
4462                                    PAGE_ALLOC_COSTLY_ORDER,
4463                                    &err, sk->sk_allocation);
4464         if (!skb)
4465                 goto err;
4466 
4467         skb_put(skb, size - data_len);
4468         skb->data_len = data_len;
4469         skb->len = size;
4470 
4471         if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4472                 goto err_free;
4473 
4474         err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4475         if (err)
4476                 goto err_free;
4477 
4478         TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4479         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4480         TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4481 
4482         if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4483                 WARN_ON_ONCE(fragstolen); /* should not happen */
4484                 __kfree_skb(skb);
4485         }
4486         return size;
4487 
4488 err_free:
4489         kfree_skb(skb);
4490 err:
4491         return err;
4492 
4493 }
4494 
4495 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4496 {
4497         struct tcp_sock *tp = tcp_sk(sk);
4498         int eaten = -1;
4499         bool fragstolen = false;
4500 
4501         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4502                 goto drop;
4503 
4504         skb_dst_drop(skb);
4505         __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4506 
4507         tcp_ecn_accept_cwr(tp, skb);
4508 
4509         tp->rx_opt.dsack = 0;
4510 
4511         /*  Queue data for delivery to the user.
4512          *  Packets in sequence go to the receive queue.
4513          *  Out of sequence packets to the out_of_order_queue.
4514          */
4515         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4516                 if (tcp_receive_window(tp) == 0)
4517                         goto out_of_window;
4518 
4519                 /* Ok. In sequence. In window. */
4520                 if (tp->ucopy.task == current &&
4521                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4522                     sock_owned_by_user(sk) && !tp->urg_data) {
4523                         int chunk = min_t(unsigned int, skb->len,
4524                                           tp->ucopy.len);
4525 
4526                         __set_current_state(TASK_RUNNING);
4527 
4528                         local_bh_enable();
4529                         if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4530                                 tp->ucopy.len -= chunk;
4531                                 tp->copied_seq += chunk;
4532                                 eaten = (chunk == skb->len);
4533                                 tcp_rcv_space_adjust(sk);
4534                         }
4535                         local_bh_disable();
4536                 }
4537 
4538                 if (eaten <= 0) {
4539 queue_and_out:
4540                         if (eaten < 0 &&
4541                             tcp_try_rmem_schedule(sk, skb, skb->truesize))
4542                                 goto drop;
4543 
4544                         eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4545                 }
4546                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4547                 if (skb->len)
4548                         tcp_event_data_recv(sk, skb);
4549                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4550                         tcp_fin(sk);
4551 
4552                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4553                         tcp_ofo_queue(sk);
4554 
4555                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4556                          * gap in queue is filled.
4557                          */
4558                         if (skb_queue_empty(&tp->out_of_order_queue))
4559                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4560                 }
4561 
4562                 if (tp->rx_opt.num_sacks)
4563                         tcp_sack_remove(tp);
4564 
4565                 tcp_fast_path_check(sk);
4566 
4567                 if (eaten > 0)
4568                         kfree_skb_partial(skb, fragstolen);
4569                 if (!sock_flag(sk, SOCK_DEAD))
4570                         sk->sk_data_ready(sk);
4571                 return;
4572         }
4573 
4574         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4575                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4576                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4577                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4578 
4579 out_of_window:
4580                 tcp_enter_quickack_mode(sk);
4581                 inet_csk_schedule_ack(sk);
4582 drop:
4583                 __kfree_skb(skb);
4584                 return;
4585         }
4586 
4587         /* Out of window. F.e. zero window probe. */
4588         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4589                 goto out_of_window;
4590 
4591         tcp_enter_quickack_mode(sk);
4592 
4593         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4594                 /* Partial packet, seq < rcv_next < end_seq */
4595                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4596                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4597                            TCP_SKB_CB(skb)->end_seq);
4598 
4599                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4600 
4601                 /* If window is closed, drop tail of packet. But after
4602                  * remembering D-SACK for its head made in previous line.
4603                  */
4604                 if (!tcp_receive_window(tp))
4605                         goto out_of_window;
4606                 goto queue_and_out;
4607         }
4608 
4609         tcp_data_queue_ofo(sk, skb);
4610 }
4611 
4612 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4613                                         struct sk_buff_head *list)
4614 {
4615         struct sk_buff *next = NULL;
4616 
4617         if (!skb_queue_is_last(list, skb))
4618                 next = skb_queue_next(list, skb);
4619 
4620         __skb_unlink(skb, list);
4621         __kfree_skb(skb);
4622         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4623 
4624         return next;
4625 }
4626 
4627 /* Collapse contiguous sequence of skbs head..tail with
4628  * sequence numbers start..end.
4629  *
4630  * If tail is NULL, this means until the end of the list.
4631  *
4632  * Segments with FIN/SYN are not collapsed (only because this
4633  * simplifies code)
4634  */
4635 static void
4636 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4637              struct sk_buff *head, struct sk_buff *tail,
4638              u32 start, u32 end)
4639 {
4640         struct sk_buff *skb, *n;
4641         bool end_of_skbs;
4642 
4643         /* First, check that queue is collapsible and find
4644          * the point where collapsing can be useful. */
4645         skb = head;
4646 restart:
4647         end_of_skbs = true;
4648         skb_queue_walk_from_safe(list, skb, n) {
4649                 if (skb == tail)
4650                         break;
4651                 /* No new bits? It is possible on ofo queue. */
4652                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4653                         skb = tcp_collapse_one(sk, skb, list);
4654                         if (!skb)
4655                                 break;
4656                         goto restart;
4657                 }
4658 
4659                 /* The first skb to collapse is:
4660                  * - not SYN/FIN and
4661                  * - bloated or contains data before "start" or
4662                  *   overlaps to the next one.
4663                  */
4664                 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4665                     (tcp_win_from_space(skb->truesize) > skb->len ||
4666                      before(TCP_SKB_CB(skb)->seq, start))) {
4667                         end_of_skbs = false;
4668                         break;
4669                 }
4670 
4671                 if (!skb_queue_is_last(list, skb)) {
4672                         struct sk_buff *next = skb_queue_next(list, skb);
4673                         if (next != tail &&
4674                             TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4675                                 end_of_skbs = false;
4676                                 break;
4677                         }
4678                 }
4679 
4680                 /* Decided to skip this, advance start seq. */
4681                 start = TCP_SKB_CB(skb)->end_seq;
4682         }
4683         if (end_of_skbs ||
4684             (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4685                 return;
4686 
4687         while (before(start, end)) {
4688                 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4689                 struct sk_buff *nskb;
4690 
4691                 nskb = alloc_skb(copy, GFP_ATOMIC);
4692                 if (!nskb)
4693                         return;
4694 
4695                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4696                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4697                 __skb_queue_before(list, skb, nskb);
4698                 skb_set_owner_r(nskb, sk);
4699 
4700                 /* Copy data, releasing collapsed skbs. */
4701                 while (copy > 0) {
4702                         int offset = start - TCP_SKB_CB(skb)->seq;
4703                         int size = TCP_SKB_CB(skb)->end_seq - start;
4704 
4705                         BUG_ON(offset < 0);
4706                         if (size > 0) {
4707                                 size = min(copy, size);
4708                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4709                                         BUG();
4710                                 TCP_SKB_CB(nskb)->end_seq += size;
4711                                 copy -= size;
4712                                 start += size;
4713                         }
4714                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4715                                 skb = tcp_collapse_one(sk, skb, list);
4716                                 if (!skb ||
4717                                     skb == tail ||
4718                                     (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4719                                         return;
4720                         }
4721                 }
4722         }
4723 }
4724 
4725 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4726  * and tcp_collapse() them until all the queue is collapsed.
4727  */
4728 static void tcp_collapse_ofo_queue(struct sock *sk)
4729 {
4730         struct tcp_sock *tp = tcp_sk(sk);
4731         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4732         struct sk_buff *head;
4733         u32 start, end;
4734 
4735         if (!skb)
4736                 return;
4737 
4738         start = TCP_SKB_CB(skb)->seq;
4739         end = TCP_SKB_CB(skb)->end_seq;
4740         head = skb;
4741 
4742         for (;;) {
4743                 struct sk_buff *next = NULL;
4744 
4745                 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4746                         next = skb_queue_next(&tp->out_of_order_queue, skb);
4747                 skb = next;
4748 
4749                 /* Segment is terminated when we see gap or when
4750                  * we are at the end of all the queue. */
4751                 if (!skb ||
4752                     after(TCP_SKB_CB(skb)->seq, end) ||
4753                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4754                         tcp_collapse(sk, &tp->out_of_order_queue,
4755                                      head, skb, start, end);
4756                         head = skb;
4757                         if (!skb)
4758                                 break;
4759                         /* Start new segment */
4760                         start = TCP_SKB_CB(skb)->seq;
4761                         end = TCP_SKB_CB(skb)->end_seq;
4762                 } else {
4763                         if (before(TCP_SKB_CB(skb)->seq, start))
4764                                 start = TCP_SKB_CB(skb)->seq;
4765                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4766                                 end = TCP_SKB_CB(skb)->end_seq;
4767                 }
4768         }
4769 }
4770 
4771 /*
4772  * Purge the out-of-order queue.
4773  * Return true if queue was pruned.
4774  */
4775 static bool tcp_prune_ofo_queue(struct sock *sk)
4776 {
4777         struct tcp_sock *tp = tcp_sk(sk);
4778         bool res = false;
4779 
4780         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4781                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4782                 __skb_queue_purge(&tp->out_of_order_queue);
4783 
4784                 /* Reset SACK state.  A conforming SACK implementation will
4785                  * do the same at a timeout based retransmit.  When a connection
4786                  * is in a sad state like this, we care only about integrity
4787                  * of the connection not performance.
4788                  */
4789                 if (tp->rx_opt.sack_ok)
4790                         tcp_sack_reset(&tp->rx_opt);
4791                 sk_mem_reclaim(sk);
4792                 res = true;
4793         }
4794         return res;
4795 }
4796 
4797 /* Reduce allocated memory if we can, trying to get
4798  * the socket within its memory limits again.
4799  *
4800  * Return less than zero if we should start dropping frames
4801  * until the socket owning process reads some of the data
4802  * to stabilize the situation.
4803  */
4804 static int tcp_prune_queue(struct sock *sk)
4805 {
4806         struct tcp_sock *tp = tcp_sk(sk);
4807 
4808         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4809 
4810         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4811 
4812         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4813                 tcp_clamp_window(sk);
4814         else if (sk_under_memory_pressure(sk))
4815                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4816 
4817         tcp_collapse_ofo_queue(sk);
4818         if (!skb_queue_empty(&sk->sk_receive_queue))
4819                 tcp_collapse(sk, &sk->sk_receive_queue,
4820                              skb_peek(&sk->sk_receive_queue),
4821                              NULL,
4822                              tp->copied_seq, tp->rcv_nxt);
4823         sk_mem_reclaim(sk);
4824 
4825         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4826                 return 0;
4827 
4828         /* Collapsing did not help, destructive actions follow.
4829          * This must not ever occur. */
4830 
4831         tcp_prune_ofo_queue(sk);
4832 
4833         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4834                 return 0;
4835 
4836         /* If we are really being abused, tell the caller to silently
4837          * drop receive data on the floor.  It will get retransmitted
4838          * and hopefully then we'll have sufficient space.
4839          */
4840         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4841 
4842         /* Massive buffer overcommit. */
4843         tp->pred_flags = 0;
4844         return -1;
4845 }
4846 
4847 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4848 {
4849         const struct tcp_sock *tp = tcp_sk(sk);
4850 
4851         /* If the user specified a specific send buffer setting, do
4852          * not modify it.
4853          */
4854         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4855                 return false;
4856 
4857         /* If we are under global TCP memory pressure, do not expand.  */
4858         if (sk_under_memory_pressure(sk))
4859                 return false;
4860 
4861         /* If we are under soft global TCP memory pressure, do not expand.  */
4862         if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4863                 return false;
4864 
4865         /* If we filled the congestion window, do not expand.  */
4866         if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4867                 return false;
4868 
4869         return true;
4870 }
4871 
4872 /* When incoming ACK allowed to free some skb from write_queue,
4873  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4874  * on the exit from tcp input handler.
4875  *
4876  * PROBLEM: sndbuf expansion does not work well with largesend.
4877  */
4878 static void tcp_new_space(struct sock *sk)
4879 {
4880         struct tcp_sock *tp = tcp_sk(sk);
4881 
4882         if (tcp_should_expand_sndbuf(sk)) {
4883                 tcp_sndbuf_expand(sk);
4884                 tp->snd_cwnd_stamp = tcp_time_stamp;
4885         }
4886 
4887         sk->sk_write_space(sk);
4888 }
4889 
4890 static void tcp_check_space(struct sock *sk)
4891 {
4892         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4893                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4894                 /* pairs with tcp_poll() */
4895                 smp_mb();
4896                 if (sk->sk_socket &&
4897                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4898                         tcp_new_space(sk);
4899         }
4900 }
4901 
4902 static inline void tcp_data_snd_check(struct sock *sk)
4903 {
4904         tcp_push_pending_frames(sk);
4905         tcp_check_space(sk);
4906 }
4907 
4908 /*
4909  * Check if sending an ack is needed.
4910  */
4911 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4912 {
4913         struct tcp_sock *tp = tcp_sk(sk);
4914 
4915             /* More than one full frame received... */
4916         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4917              /* ... and right edge of window advances far enough.
4918               * (tcp_recvmsg() will send ACK otherwise). Or...
4919               */
4920              __tcp_select_window(sk) >= tp->rcv_wnd) ||
4921             /* We ACK each frame or... */
4922             tcp_in_quickack_mode(sk) ||
4923             /* We have out of order data. */
4924             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4925                 /* Then ack it now */
4926                 tcp_send_ack(sk);
4927         } else {
4928                 /* Else, send delayed ack. */
4929                 tcp_send_delayed_ack(sk);
4930         }
4931 }
4932 
4933 static inline void tcp_ack_snd_check(struct sock *sk)
4934 {
4935         if (!inet_csk_ack_scheduled(sk)) {
4936                 /* We sent a data segment already. */
4937                 return;
4938         }
4939         __tcp_ack_snd_check(sk, 1);
4940 }
4941 
4942 /*
4943  *      This routine is only called when we have urgent data
4944  *      signaled. Its the 'slow' part of tcp_urg. It could be
4945  *      moved inline now as tcp_urg is only called from one
4946  *      place. We handle URGent data wrong. We have to - as
4947  *      BSD still doesn't use the correction from RFC961.
4948  *      For 1003.1g we should support a new option TCP_STDURG to permit
4949  *      either form (or just set the sysctl tcp_stdurg).
4950  */
4951 
4952 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4953 {
4954         struct tcp_sock *tp = tcp_sk(sk);
4955         u32 ptr = ntohs(th->urg_ptr);
4956 
4957         if (ptr && !sysctl_tcp_stdurg)
4958                 ptr--;
4959         ptr += ntohl(th->seq);
4960 
4961         /* Ignore urgent data that we've already seen and read. */
4962         if (after(tp->copied_seq, ptr))
4963                 return;
4964 
4965         /* Do not replay urg ptr.
4966          *
4967          * NOTE: interesting situation not covered by specs.
4968          * Misbehaving sender may send urg ptr, pointing to segment,
4969          * which we already have in ofo queue. We are not able to fetch
4970          * such data and will stay in TCP_URG_NOTYET until will be eaten
4971          * by recvmsg(). Seems, we are not obliged to handle such wicked
4972          * situations. But it is worth to think about possibility of some
4973          * DoSes using some hypothetical application level deadlock.
4974          */
4975         if (before(ptr, tp->rcv_nxt))
4976                 return;
4977 
4978         /* Do we already have a newer (or duplicate) urgent pointer? */
4979         if (tp->urg_data && !after(ptr, tp->urg_seq))
4980                 return;
4981 
4982         /* Tell the world about our new urgent pointer. */
4983         sk_send_sigurg(sk);
4984 
4985         /* We may be adding urgent data when the last byte read was
4986          * urgent. To do this requires some care. We cannot just ignore
4987          * tp->copied_seq since we would read the last urgent byte again
4988          * as data, nor can we alter copied_seq until this data arrives
4989          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4990          *
4991          * NOTE. Double Dutch. Rendering to plain English: author of comment
4992          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4993          * and expect that both A and B disappear from stream. This is _wrong_.
4994          * Though this happens in BSD with high probability, this is occasional.
4995          * Any application relying on this is buggy. Note also, that fix "works"
4996          * only in this artificial test. Insert some normal data between A and B and we will
4997          * decline of BSD again. Verdict: it is better to remove to trap
4998          * buggy users.
4999          */
5000         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5001             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5002                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5003                 tp->copied_seq++;
5004                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5005                         __skb_unlink(skb, &sk->sk_receive_queue);
5006                         __kfree_skb(skb);
5007                 }
5008         }
5009 
5010         tp->urg_data = TCP_URG_NOTYET;
5011         tp->urg_seq = ptr;
5012 
5013         /* Disable header prediction. */
5014         tp->pred_flags = 0;
5015 }
5016 
5017 /* This is the 'fast' part of urgent handling. */
5018 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5019 {
5020         struct tcp_sock *tp = tcp_sk(sk);
5021 
5022         /* Check if we get a new urgent pointer - normally not. */
5023         if (th->urg)
5024                 tcp_check_urg(sk, th);
5025 
5026         /* Do we wait for any urgent data? - normally not... */
5027         if (tp->urg_data == TCP_URG_NOTYET) {
5028                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5029                           th->syn;
5030 
5031                 /* Is the urgent pointer pointing into this packet? */
5032                 if (ptr < skb->len) {
5033                         u8 tmp;
5034                         if (skb_copy_bits(skb, ptr, &tmp, 1))
5035                                 BUG();
5036                         tp->urg_data = TCP_URG_VALID | tmp;
5037                         if (!sock_flag(sk, SOCK_DEAD))
5038                                 sk->sk_data_ready(sk);
5039                 }
5040         }
5041 }
5042 
5043 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5044 {
5045         struct tcp_sock *tp = tcp_sk(sk);
5046         int chunk = skb->len - hlen;
5047         int err;
5048 
5049         local_bh_enable();
5050         if (skb_csum_unnecessary(skb))
5051                 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5052         else
5053                 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5054 
5055         if (!err) {
5056                 tp->ucopy.len -= chunk;
5057                 tp->copied_seq += chunk;
5058                 tcp_rcv_space_adjust(sk);
5059         }
5060 
5061         local_bh_disable();
5062         return err;
5063 }
5064 
5065 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5066                                             struct sk_buff *skb)
5067 {
5068         __sum16 result;
5069 
5070         if (sock_owned_by_user(sk)) {
5071                 local_bh_enable();
5072                 result = __tcp_checksum_complete(skb);
5073                 local_bh_disable();
5074         } else {
5075                 result = __tcp_checksum_complete(skb);
5076         }
5077         return result;
5078 }
5079 
5080 static inline bool tcp_checksum_complete_user(struct sock *sk,
5081                                              struct sk_buff *skb)
5082 {
5083         return !skb_csum_unnecessary(skb) &&
5084                __tcp_checksum_complete_user(sk, skb);
5085 }
5086 
5087 /* Does PAWS and seqno based validation of an incoming segment, flags will
5088  * play significant role here.
5089  */
5090 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5091                                   const struct tcphdr *th, int syn_inerr)
5092 {
5093         struct tcp_sock *tp = tcp_sk(sk);
5094 
5095         /* RFC1323: H1. Apply PAWS check first. */
5096         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5097             tcp_paws_discard(sk, skb)) {
5098                 if (!th->rst) {
5099                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5100                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5101                                                   LINUX_MIB_TCPACKSKIPPEDPAWS,
5102                                                   &tp->last_oow_ack_time))
5103                                 tcp_send_dupack(sk, skb);
5104                         goto discard;
5105                 }
5106                 /* Reset is accepted even if it did not pass PAWS. */
5107         }
5108 
5109         /* Step 1: check sequence number */
5110         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5111                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5112                  * (RST) segments are validated by checking their SEQ-fields."
5113                  * And page 69: "If an incoming segment is not acceptable,
5114                  * an acknowledgment should be sent in reply (unless the RST
5115                  * bit is set, if so drop the segment and return)".
5116                  */
5117                 if (!th->rst) {
5118                         if (th->syn)
5119                                 goto syn_challenge;
5120                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5121                                                   LINUX_MIB_TCPACKSKIPPEDSEQ,
5122                                                   &tp->last_oow_ack_time))
5123                                 tcp_send_dupack(sk, skb);
5124                 }
5125                 goto discard;
5126         }
5127 
5128         /* Step 2: check RST bit */
5129         if (th->rst) {
5130                 /* RFC 5961 3.2 :
5131                  * If sequence number exactly matches RCV.NXT, then
5132                  *     RESET the connection
5133                  * else
5134                  *     Send a challenge ACK
5135                  */
5136                 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5137                         tcp_reset(sk);
5138                 else
5139                         tcp_send_challenge_ack(sk, skb);
5140                 goto discard;
5141         }
5142 
5143         /* step 3: check security and precedence [ignored] */
5144 
5145         /* step 4: Check for a SYN
5146          * RFC 5961 4.2 : Send a challenge ack
5147          */
5148         if (th->syn) {
5149 syn_challenge:
5150                 if (syn_inerr)
5151                         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5152                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5153                 tcp_send_challenge_ack(sk, skb);
5154                 goto discard;
5155         }
5156 
5157         return true;
5158 
5159 discard:
5160         __kfree_skb(skb);
5161         return false;
5162 }
5163 
5164 /*
5165  *      TCP receive function for the ESTABLISHED state.
5166  *
5167  *      It is split into a fast path and a slow path. The fast path is
5168  *      disabled when:
5169  *      - A zero window was announced from us - zero window probing
5170  *        is only handled properly in the slow path.
5171  *      - Out of order segments arrived.
5172  *      - Urgent data is expected.
5173  *      - There is no buffer space left
5174  *      - Unexpected TCP flags/window values/header lengths are received
5175  *        (detected by checking the TCP header against pred_flags)
5176  *      - Data is sent in both directions. Fast path only supports pure senders
5177  *        or pure receivers (this means either the sequence number or the ack
5178  *        value must stay constant)
5179  *      - Unexpected TCP option.
5180  *
5181  *      When these conditions are not satisfied it drops into a standard
5182  *      receive procedure patterned after RFC793 to handle all cases.
5183  *      The first three cases are guaranteed by proper pred_flags setting,
5184  *      the rest is checked inline. Fast processing is turned on in
5185  *      tcp_data_queue when everything is OK.
5186  */
5187 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5188                          const struct tcphdr *th, unsigned int len)
5189 {
5190         struct tcp_sock *tp = tcp_sk(sk);
5191 
5192         if (unlikely(!sk->sk_rx_dst))
5193                 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5194         /*
5195          *      Header prediction.
5196          *      The code loosely follows the one in the famous
5197          *      "30 instruction TCP receive" Van Jacobson mail.
5198          *
5199          *      Van's trick is to deposit buffers into socket queue
5200          *      on a device interrupt, to call tcp_recv function
5201          *      on the receive process context and checksum and copy
5202          *      the buffer to user space. smart...
5203          *
5204          *      Our current scheme is not silly either but we take the
5205          *      extra cost of the net_bh soft interrupt processing...
5206          *      We do checksum and copy also but from device to kernel.
5207          */
5208 
5209         tp->rx_opt.saw_tstamp = 0;
5210 
5211         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5212          *      if header_prediction is to be made
5213          *      'S' will always be tp->tcp_header_len >> 2
5214          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5215          *  turn it off (when there are holes in the receive
5216          *       space for instance)
5217          *      PSH flag is ignored.
5218          */
5219 
5220         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5221             TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5222             !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5223                 int tcp_header_len = tp->tcp_header_len;
5224 
5225                 /* Timestamp header prediction: tcp_header_len
5226                  * is automatically equal to th->doff*4 due to pred_flags
5227                  * match.
5228                  */
5229 
5230                 /* Check timestamp */
5231                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5232                         /* No? Slow path! */
5233                         if (!tcp_parse_aligned_timestamp(tp, th))
5234                                 goto slow_path;
5235 
5236                         /* If PAWS failed, check it more carefully in slow path */
5237                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5238                                 goto slow_path;
5239 
5240                         /* DO NOT update ts_recent here, if checksum fails
5241                          * and timestamp was corrupted part, it will result
5242                          * in a hung connection since we will drop all
5243                          * future packets due to the PAWS test.
5244                          */
5245                 }
5246 
5247                 if (len <= tcp_header_len) {
5248                         /* Bulk data transfer: sender */
5249                         if (len == tcp_header_len) {
5250                                 /* Predicted packet is in window by definition.
5251                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5252                                  * Hence, check seq<=rcv_wup reduces to:
5253                                  */
5254                                 if (tcp_header_len ==
5255                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5256                                     tp->rcv_nxt == tp->rcv_wup)
5257                                         tcp_store_ts_recent(tp);
5258 
5259                                 /* We know that such packets are checksummed
5260                                  * on entry.
5261                                  */
5262                                 tcp_ack(sk, skb, 0);
5263                                 __kfree_skb(skb);
5264                                 tcp_data_snd_check(sk);
5265                                 return;
5266                         } else { /* Header too small */
5267                                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5268                                 goto discard;
5269                         }
5270                 } else {
5271                         int eaten = 0;
5272                         bool fragstolen = false;
5273 
5274                         if (tp->ucopy.task == current &&
5275                             tp->copied_seq == tp->rcv_nxt &&
5276                             len - tcp_header_len <= tp->ucopy.len &&
5277                             sock_owned_by_user(sk)) {
5278                                 __set_current_state(TASK_RUNNING);
5279 
5280                                 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5281                                         /* Predicted packet is in window by definition.
5282                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5283                                          * Hence, check seq<=rcv_wup reduces to:
5284                                          */
5285                                         if (tcp_header_len ==
5286                                             (sizeof(struct tcphdr) +
5287                                              TCPOLEN_TSTAMP_ALIGNED) &&
5288                                             tp->rcv_nxt == tp->rcv_wup)
5289                                                 tcp_store_ts_recent(tp);
5290 
5291                                         tcp_rcv_rtt_measure_ts(sk, skb);
5292 
5293                                         __skb_pull(skb, tcp_header_len);
5294                                         tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5295                                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5296                                         eaten = 1;
5297                                 }
5298                         }
5299                         if (!eaten) {
5300                                 if (tcp_checksum_complete_user(sk, skb))
5301                                         goto csum_error;
5302 
5303                                 if ((int)skb->truesize > sk->sk_forward_alloc)
5304                                         goto step5;
5305 
5306                                 /* Predicted packet is in window by definition.
5307                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5308                                  * Hence, check seq<=rcv_wup reduces to:
5309                                  */
5310                                 if (tcp_header_len ==
5311                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5312                                     tp->rcv_nxt == tp->rcv_wup)
5313                                         tcp_store_ts_recent(tp);
5314 
5315                                 tcp_rcv_rtt_measure_ts(sk, skb);
5316 
5317                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5318 
5319                                 /* Bulk data transfer: receiver */
5320                                 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5321                                                       &fragstolen);
5322                         }
5323 
5324                         tcp_event_data_recv(sk, skb);
5325 
5326                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5327                                 /* Well, only one small jumplet in fast path... */
5328                                 tcp_ack(sk, skb, FLAG_DATA);
5329                                 tcp_data_snd_check(sk);
5330                                 if (!inet_csk_ack_scheduled(sk))
5331                                         goto no_ack;
5332                         }
5333 
5334                         __tcp_ack_snd_check(sk, 0);
5335 no_ack:
5336                         if (eaten)
5337                                 kfree_skb_partial(skb, fragstolen);
5338                         sk->sk_data_ready(sk);
5339                         return;
5340                 }
5341         }
5342 
5343 slow_path:
5344         if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5345                 goto csum_error;
5346 
5347         if (!th->ack && !th->rst && !th->syn)
5348                 goto discard;
5349 
5350         /*
5351          *      Standard slow path.
5352          */
5353 
5354         if (!tcp_validate_incoming(sk, skb, th, 1))
5355                 return;
5356 
5357 step5:
5358         if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5359                 goto discard;
5360 
5361         tcp_rcv_rtt_measure_ts(sk, skb);
5362 
5363         /* Process urgent data. */
5364         tcp_urg(sk, skb, th);
5365 
5366         /* step 7: process the segment text */
5367         tcp_data_queue(sk, skb);
5368 
5369         tcp_data_snd_check(sk);
5370         tcp_ack_snd_check(sk);
5371         return;
5372 
5373 csum_error:
5374         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5375         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5376 
5377 discard:
5378         __kfree_skb(skb);
5379 }
5380 EXPORT_SYMBOL(tcp_rcv_established);
5381 
5382 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5383 {
5384         struct tcp_sock *tp = tcp_sk(sk);
5385         struct inet_connection_sock *icsk = inet_csk(sk);
5386 
5387         tcp_set_state(sk, TCP_ESTABLISHED);
5388         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5389 
5390         if (skb) {
5391                 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5392                 security_inet_conn_established(sk, skb);
5393         }
5394 
5395         /* Make sure socket is routed, for correct metrics.  */
5396         icsk->icsk_af_ops->rebuild_header(sk);
5397 
5398         tcp_init_metrics(sk);
5399 
5400         tcp_init_congestion_control(sk);
5401 
5402         /* Prevent spurious tcp_cwnd_restart() on first data
5403          * packet.
5404          */
5405         tp->lsndtime = tcp_time_stamp;
5406 
5407         tcp_init_buffer_space(sk);
5408 
5409         if (sock_flag(sk, SOCK_KEEPOPEN))
5410                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5411 
5412         if (!tp->rx_opt.snd_wscale)
5413                 __tcp_fast_path_on(tp, tp->snd_wnd);
5414         else
5415                 tp->pred_flags = 0;
5416 
5417 }
5418 
5419 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5420                                     struct tcp_fastopen_cookie *cookie)
5421 {
5422         struct tcp_sock *tp = tcp_sk(sk);
5423         struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5424         u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5425         bool syn_drop = false;
5426 
5427         if (mss == tp->rx_opt.user_mss) {
5428                 struct tcp_options_received opt;
5429 
5430                 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5431                 tcp_clear_options(&opt);
5432                 opt.user_mss = opt.mss_clamp = 0;
5433                 tcp_parse_options(synack, &opt, 0, NULL);
5434                 mss = opt.mss_clamp;
5435         }
5436 
5437         if (!tp->syn_fastopen) {
5438                 /* Ignore an unsolicited cookie */
5439                 cookie->len = -1;
5440         } else if (tp->total_retrans) {
5441                 /* SYN timed out and the SYN-ACK neither has a cookie nor
5442                  * acknowledges data. Presumably the remote received only
5443                  * the retransmitted (regular) SYNs: either the original
5444                  * SYN-data or the corresponding SYN-ACK was dropped.
5445                  */
5446                 syn_drop = (cookie->len < 0 && data);
5447         } else if (cookie->len < 0 && !tp->syn_data) {
5448                 /* We requested a cookie but didn't get it. If we did not use
5449                  * the (old) exp opt format then try so next time (try_exp=1).
5450                  * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5451                  */
5452                 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5453         }
5454 
5455         tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5456 
5457         if (data) { /* Retransmit unacked data in SYN */
5458                 tcp_for_write_queue_from(data, sk) {
5459                         if (data == tcp_send_head(sk) ||
5460                             __tcp_retransmit_skb(sk, data))
5461                                 break;
5462                 }
5463                 tcp_rearm_rto(sk);
5464                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5465                 return true;
5466         }
5467         tp->syn_data_acked = tp->syn_data;
5468         if (tp->syn_data_acked)
5469                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5470         return false;
5471 }
5472 
5473 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5474                                          const struct tcphdr *th, unsigned int len)
5475 {
5476         struct inet_connection_sock *icsk = inet_csk(sk);
5477         struct tcp_sock *tp = tcp_sk(sk);
5478         struct tcp_fastopen_cookie foc = { .len = -1 };
5479         int saved_clamp = tp->rx_opt.mss_clamp;
5480         bool fastopen_fail;
5481 
5482         tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5483         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5484                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5485 
5486         if (th->ack) {
5487                 /* rfc793:
5488                  * "If the state is SYN-SENT then
5489                  *    first check the ACK bit
5490                  *      If the ACK bit is set
5491                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5492                  *        a reset (unless the RST bit is set, if so drop
5493                  *        the segment and return)"
5494                  */
5495                 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5496                     after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5497                         goto reset_and_undo;
5498 
5499                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5500                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5501                              tcp_time_stamp)) {
5502                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5503                         goto reset_and_undo;
5504                 }
5505 
5506                 /* Now ACK is acceptable.
5507                  *
5508                  * "If the RST bit is set
5509                  *    If the ACK was acceptable then signal the user "error:
5510                  *    connection reset", drop the segment, enter CLOSED state,
5511                  *    delete TCB, and return."
5512                  */
5513 
5514                 if (th->rst) {
5515                         tcp_reset(sk);
5516                         goto discard;
5517                 }
5518 
5519                 /* rfc793:
5520                  *   "fifth, if neither of the SYN or RST bits is set then
5521                  *    drop the segment and return."
5522                  *
5523                  *    See note below!
5524                  *                                        --ANK(990513)
5525                  */
5526                 if (!th->syn)
5527                         goto discard_and_undo;
5528 
5529                 /* rfc793:
5530                  *   "If the SYN bit is on ...
5531                  *    are acceptable then ...
5532                  *    (our SYN has been ACKed), change the connection
5533                  *    state to ESTABLISHED..."
5534                  */
5535 
5536                 tcp_ecn_rcv_synack(tp, th);
5537 
5538                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5539                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5540 
5541                 /* Ok.. it's good. Set up sequence numbers and
5542                  * move to established.
5543                  */
5544                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5545                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5546 
5547                 /* RFC1323: The window in SYN & SYN/ACK segments is
5548                  * never scaled.
5549                  */
5550                 tp->snd_wnd = ntohs(th->window);
5551 
5552                 if (!tp->rx_opt.wscale_ok) {
5553                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5554                         tp->window_clamp = min(tp->window_clamp, 65535U);
5555                 }
5556 
5557                 if (tp->rx_opt.saw_tstamp) {
5558                         tp->rx_opt.tstamp_ok       = 1;
5559                         tp->tcp_header_len =
5560                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5561                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5562                         tcp_store_ts_recent(tp);
5563                 } else {
5564                         tp->tcp_header_len = sizeof(struct tcphdr);
5565                 }
5566 
5567                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5568                         tcp_enable_fack(tp);
5569 
5570                 tcp_mtup_init(sk);
5571                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5572                 tcp_initialize_rcv_mss(sk);
5573 
5574                 /* Remember, tcp_poll() does not lock socket!
5575                  * Change state from SYN-SENT only after copied_seq
5576                  * is initialized. */
5577                 tp->copied_seq = tp->rcv_nxt;
5578 
5579                 smp_mb();
5580 
5581                 tcp_finish_connect(sk, skb);
5582 
5583                 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5584                                 tcp_rcv_fastopen_synack(sk, skb, &foc);
5585 
5586                 if (!sock_flag(sk, SOCK_DEAD)) {
5587                         sk->sk_state_change(sk);
5588                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5589                 }
5590                 if (fastopen_fail)
5591                         return -1;
5592                 if (sk->sk_write_pending ||
5593                     icsk->icsk_accept_queue.rskq_defer_accept ||
5594                     icsk->icsk_ack.pingpong) {
5595                         /* Save one ACK. Data will be ready after
5596                          * several ticks, if write_pending is set.
5597                          *
5598                          * It may be deleted, but with this feature tcpdumps
5599                          * look so _wonderfully_ clever, that I was not able
5600                          * to stand against the temptation 8)     --ANK
5601                          */
5602                         inet_csk_schedule_ack(sk);
5603                         tcp_enter_quickack_mode(sk);
5604                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5605                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5606 
5607 discard:
5608                         __kfree_skb(skb);
5609                         return 0;
5610                 } else {
5611                         tcp_send_ack(sk);
5612                 }
5613                 return -1;
5614         }
5615 
5616         /* No ACK in the segment */
5617 
5618         if (th->rst) {
5619                 /* rfc793:
5620                  * "If the RST bit is set
5621                  *
5622                  *      Otherwise (no ACK) drop the segment and return."
5623                  */
5624 
5625                 goto discard_and_undo;
5626         }
5627 
5628         /* PAWS check. */
5629         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5630             tcp_paws_reject(&tp->rx_opt, 0))
5631                 goto discard_and_undo;
5632 
5633         if (th->syn) {
5634                 /* We see SYN without ACK. It is attempt of
5635                  * simultaneous connect with crossed SYNs.
5636                  * Particularly, it can be connect to self.
5637                  */
5638                 tcp_set_state(sk, TCP_SYN_RECV);
5639 
5640                 if (tp->rx_opt.saw_tstamp) {
5641                         tp->rx_opt.tstamp_ok = 1;
5642                         tcp_store_ts_recent(tp);
5643                         tp->tcp_header_len =
5644                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5645                 } else {
5646                         tp->tcp_header_len = sizeof(struct tcphdr);
5647                 }
5648 
5649                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5650                 tp->copied_seq = tp->rcv_nxt;
5651                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5652 
5653                 /* RFC1323: The window in SYN & SYN/ACK segments is
5654                  * never scaled.
5655                  */
5656                 tp->snd_wnd    = ntohs(th->window);
5657                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5658                 tp->max_window = tp->snd_wnd;
5659 
5660                 tcp_ecn_rcv_syn(tp, th);
5661 
5662                 tcp_mtup_init(sk);
5663                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5664                 tcp_initialize_rcv_mss(sk);
5665 
5666                 tcp_send_synack(sk);
5667 #if 0
5668                 /* Note, we could accept data and URG from this segment.
5669                  * There are no obstacles to make this (except that we must
5670                  * either change tcp_recvmsg() to prevent it from returning data
5671                  * before 3WHS completes per RFC793, or employ TCP Fast Open).
5672                  *
5673                  * However, if we ignore data in ACKless segments sometimes,
5674                  * we have no reasons to accept it sometimes.
5675                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5676                  * is not flawless. So, discard packet for sanity.
5677                  * Uncomment this return to process the data.
5678                  */
5679                 return -1;
5680 #else
5681                 goto discard;
5682 #endif
5683         }
5684         /* "fifth, if neither of the SYN or RST bits is set then
5685          * drop the segment and return."
5686          */
5687 
5688 discard_and_undo:
5689         tcp_clear_options(&tp->rx_opt);
5690         tp->rx_opt.mss_clamp = saved_clamp;
5691         goto discard;
5692 
5693 reset_and_undo:
5694         tcp_clear_options(&tp->rx_opt);
5695         tp->rx_opt.mss_clamp = saved_clamp;
5696         return 1;
5697 }
5698 
5699 /*
5700  *      This function implements the receiving procedure of RFC 793 for
5701  *      all states except ESTABLISHED and TIME_WAIT.
5702  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5703  *      address independent.
5704  */
5705 
5706 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5707                           const struct tcphdr *th, unsigned int len)
5708 {
5709         struct tcp_sock *tp = tcp_sk(sk);
5710         struct inet_connection_sock *icsk = inet_csk(sk);
5711         struct request_sock *req;
5712         int queued = 0;
5713         bool acceptable;
5714         u32 synack_stamp;
5715 
5716         tp->rx_opt.saw_tstamp = 0;
5717 
5718         switch (sk->sk_state) {
5719         case TCP_CLOSE:
5720                 goto discard;
5721 
5722         case TCP_LISTEN:
5723                 if (th->ack)
5724                         return 1;
5725 
5726                 if (th->rst)
5727                         goto discard;
5728 
5729                 if (th->syn) {
5730                         if (th->fin)
5731                                 goto discard;
5732                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5733                                 return 1;
5734 
5735                         /* Now we have several options: In theory there is
5736                          * nothing else in the frame. KA9Q has an option to
5737                          * send data with the syn, BSD accepts data with the
5738                          * syn up to the [to be] advertised window and
5739                          * Solaris 2.1 gives you a protocol error. For now
5740                          * we just ignore it, that fits the spec precisely
5741                          * and avoids incompatibilities. It would be nice in
5742                          * future to drop through and process the data.
5743                          *
5744                          * Now that TTCP is starting to be used we ought to
5745                          * queue this data.
5746                          * But, this leaves one open to an easy denial of
5747                          * service attack, and SYN cookies can't defend
5748                          * against this problem. So, we drop the data
5749                          * in the interest of security over speed unless
5750                          * it's still in use.
5751                          */
5752                         kfree_skb(skb);
5753                         return 0;
5754                 }
5755                 goto discard;
5756 
5757         case TCP_SYN_SENT:
5758                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5759                 if (queued >= 0)
5760                         return queued;
5761 
5762                 /* Do step6 onward by hand. */
5763                 tcp_urg(sk, skb, th);
5764                 __kfree_skb(skb);
5765                 tcp_data_snd_check(sk);
5766                 return 0;
5767         }
5768 
5769         req = tp->fastopen_rsk;
5770         if (req) {
5771                 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5772                     sk->sk_state != TCP_FIN_WAIT1);
5773 
5774                 if (!tcp_check_req(sk, skb, req, true))
5775                         goto discard;
5776         }
5777 
5778         if (!th->ack && !th->rst && !th->syn)
5779                 goto discard;
5780 
5781         if (!tcp_validate_incoming(sk, skb, th, 0))
5782                 return 0;
5783 
5784         /* step 5: check the ACK field */
5785         acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5786                                       FLAG_UPDATE_TS_RECENT) > 0;
5787 
5788         switch (sk->sk_state) {
5789         case TCP_SYN_RECV:
5790                 if (!acceptable)
5791                         return 1;
5792 
5793                 /* Once we leave TCP_SYN_RECV, we no longer need req
5794                  * so release it.
5795                  */
5796                 if (req) {
5797                         synack_stamp = tcp_rsk(req)->snt_synack;
5798                         tp->total_retrans = req->num_retrans;
5799                         reqsk_fastopen_remove(sk, req, false);
5800                 } else {
5801                         synack_stamp = tp->lsndtime;
5802                         /* Make sure socket is routed, for correct metrics. */
5803                         icsk->icsk_af_ops->rebuild_header(sk);
5804                         tcp_init_congestion_control(sk);
5805 
5806                         tcp_mtup_init(sk);
5807                         tp->copied_seq = tp->rcv_nxt;
5808                         tcp_init_buffer_space(sk);
5809                 }
5810                 smp_mb();
5811                 tcp_set_state(sk, TCP_ESTABLISHED);
5812                 sk->sk_state_change(sk);
5813 
5814                 /* Note, that this wakeup is only for marginal crossed SYN case.
5815                  * Passively open sockets are not waked up, because
5816                  * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5817                  */
5818                 if (sk->sk_socket)
5819                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5820 
5821                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5822                 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5823                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5824                 tcp_synack_rtt_meas(sk, synack_stamp);
5825 
5826                 if (tp->rx_opt.tstamp_ok)
5827                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5828 
5829                 if (req) {
5830                         /* Re-arm the timer because data may have been sent out.
5831                          * This is similar to the regular data transmission case
5832                          * when new data has just been ack'ed.
5833                          *
5834                          * (TFO) - we could try to be more aggressive and
5835                          * retransmitting any data sooner based on when they
5836                          * are sent out.
5837                          */
5838                         tcp_rearm_rto(sk);
5839                 } else
5840                         tcp_init_metrics(sk);
5841 
5842                 tcp_update_pacing_rate(sk);
5843 
5844                 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5845                 tp->lsndtime = tcp_time_stamp;
5846 
5847                 tcp_initialize_rcv_mss(sk);
5848                 tcp_fast_path_on(tp);
5849                 break;
5850 
5851         case TCP_FIN_WAIT1: {
5852                 struct dst_entry *dst;
5853                 int tmo;
5854 
5855                 /* If we enter the TCP_FIN_WAIT1 state and we are a
5856                  * Fast Open socket and this is the first acceptable
5857                  * ACK we have received, this would have acknowledged
5858                  * our SYNACK so stop the SYNACK timer.
5859                  */
5860                 if (req) {
5861                         /* Return RST if ack_seq is invalid.
5862                          * Note that RFC793 only says to generate a
5863                          * DUPACK for it but for TCP Fast Open it seems
5864                          * better to treat this case like TCP_SYN_RECV
5865                          * above.
5866                          */
5867                         if (!acceptable)
5868                                 return 1;
5869                         /* We no longer need the request sock. */
5870                         reqsk_fastopen_remove(sk, req, false);
5871                         tcp_rearm_rto(sk);
5872                 }
5873                 if (tp->snd_una != tp->write_seq)
5874                         break;
5875 
5876                 tcp_set_state(sk, TCP_FIN_WAIT2);
5877                 sk->sk_shutdown |= SEND_SHUTDOWN;
5878 
5879                 dst = __sk_dst_get(sk);
5880                 if (dst)
5881                         dst_confirm(dst);
5882 
5883                 if (!sock_flag(sk, SOCK_DEAD)) {
5884                         /* Wake up lingering close() */
5885                         sk->sk_state_change(sk);
5886                         break;
5887                 }
5888 
5889                 if (tp->linger2 < 0 ||
5890                     (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5891                      after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5892                         tcp_done(sk);
5893                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5894                         return 1;
5895                 }
5896 
5897                 tmo = tcp_fin_time(sk);
5898                 if (tmo > TCP_TIMEWAIT_LEN) {
5899                         inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5900                 } else if (th->fin || sock_owned_by_user(sk)) {
5901                         /* Bad case. We could lose such FIN otherwise.
5902                          * It is not a big problem, but it looks confusing
5903                          * and not so rare event. We still can lose it now,
5904                          * if it spins in bh_lock_sock(), but it is really
5905                          * marginal case.
5906                          */
5907                         inet_csk_reset_keepalive_timer(sk, tmo);
5908                 } else {
5909                         tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5910                         goto discard;
5911                 }
5912                 break;
5913         }
5914 
5915         case TCP_CLOSING:
5916                 if (tp->snd_una == tp->write_seq) {
5917                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5918                         goto discard;
5919                 }
5920                 break;
5921 
5922         case TCP_LAST_ACK:
5923                 if (tp->snd_una == tp->write_seq) {
5924                         tcp_update_metrics(sk);
5925                         tcp_done(sk);
5926                         goto discard;
5927                 }
5928                 break;
5929         }
5930 
5931         /* step 6: check the URG bit */
5932         tcp_urg(sk, skb, th);
5933 
5934         /* step 7: process the segment text */
5935         switch (sk->sk_state) {
5936         case TCP_CLOSE_WAIT:
5937         case TCP_CLOSING:
5938         case TCP_LAST_ACK:
5939                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5940                         break;
5941         case TCP_FIN_WAIT1:
5942         case TCP_FIN_WAIT2:
5943                 /* RFC 793 says to queue data in these states,
5944                  * RFC 1122 says we MUST send a reset.
5945                  * BSD 4.4 also does reset.
5946                  */
5947                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5948                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5949                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5950                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5951                                 tcp_reset(sk);
5952                                 return 1;
5953                         }
5954                 }
5955                 /* Fall through */
5956         case TCP_ESTABLISHED:
5957                 tcp_data_queue(sk, skb);
5958                 queued = 1;
5959                 break;
5960         }
5961 
5962         /* tcp_data could move socket to TIME-WAIT */
5963         if (sk->sk_state != TCP_CLOSE) {
5964                 tcp_data_snd_check(sk);
5965                 tcp_ack_snd_check(sk);
5966         }
5967 
5968         if (!queued) {
5969 discard:
5970                 __kfree_skb(skb);
5971         }
5972         return 0;
5973 }
5974 EXPORT_SYMBOL(tcp_rcv_state_process);
5975 
5976 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
5977 {
5978         struct inet_request_sock *ireq = inet_rsk(req);
5979 
5980         if (family == AF_INET)
5981                 net_dbg_ratelimited("drop open request from %pI4/%u\n",
5982                                     &ireq->ir_rmt_addr, port);
5983 #if IS_ENABLED(CONFIG_IPV6)
5984         else if (family == AF_INET6)
5985                 net_dbg_ratelimited("drop open request from %pI6/%u\n",
5986