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

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

Version: ~ [ linux-5.5-rc7 ] ~ [ linux-5.4.13 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.97 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.166 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.210 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.210 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.81 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
  3  *              operating system.  INET is implemented using the  BSD Socket
  4  *              interface as the means of communication with the user level.
  5  *
  6  *              Implementation of the Transmission Control Protocol(TCP).
  7  *
  8  * Authors:     Ross Biro
  9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
 10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
 11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
 12  *              Florian La Roche, <flla@stud.uni-sb.de>
 13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
 14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
 15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
 16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
 17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
 18  *              Jorge Cwik, <jorge@laser.satlink.net>
 19  */
 20 
 21 /*
 22  * Changes:
 23  *              Pedro Roque     :       Fast Retransmit/Recovery.
 24  *                                      Two receive queues.
 25  *                                      Retransmit queue handled by TCP.
 26  *                                      Better retransmit timer handling.
 27  *                                      New congestion avoidance.
 28  *                                      Header prediction.
 29  *                                      Variable renaming.
 30  *
 31  *              Eric            :       Fast Retransmit.
 32  *              Randy Scott     :       MSS option defines.
 33  *              Eric Schenk     :       Fixes to slow start algorithm.
 34  *              Eric Schenk     :       Yet another double ACK bug.
 35  *              Eric Schenk     :       Delayed ACK bug fixes.
 36  *              Eric Schenk     :       Floyd style fast retrans war avoidance.
 37  *              David S. Miller :       Don't allow zero congestion window.
 38  *              Eric Schenk     :       Fix retransmitter so that it sends
 39  *                                      next packet on ack of previous packet.
 40  *              Andi Kleen      :       Moved open_request checking here
 41  *                                      and process RSTs for open_requests.
 42  *              Andi Kleen      :       Better prune_queue, and other fixes.
 43  *              Andrey Savochkin:       Fix RTT measurements in the presence of
 44  *                                      timestamps.
 45  *              Andrey Savochkin:       Check sequence numbers correctly when
 46  *                                      removing SACKs due to in sequence incoming
 47  *                                      data segments.
 48  *              Andi Kleen:             Make sure we never ack data there is not
 49  *                                      enough room for. Also make this condition
 50  *                                      a fatal error if it might still happen.
 51  *              Andi Kleen:             Add tcp_measure_rcv_mss to make
 52  *                                      connections with MSS<min(MTU,ann. MSS)
 53  *                                      work without delayed acks.
 54  *              Andi Kleen:             Process packets with PSH set in the
 55  *                                      fast path.
 56  *              J Hadi Salim:           ECN support
 57  *              Andrei Gurtov,
 58  *              Pasi Sarolahti,
 59  *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
 60  *                                      engine. Lots of bugs are found.
 61  *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
 62  */
 63 
 64 #define pr_fmt(fmt) "TCP: " fmt
 65 
 66 #include <linux/mm.h>
 67 #include <linux/slab.h>
 68 #include <linux/module.h>
 69 #include <linux/sysctl.h>
 70 #include <linux/kernel.h>
 71 #include <net/dst.h>
 72 #include <net/tcp.h>
 73 #include <net/inet_common.h>
 74 #include <linux/ipsec.h>
 75 #include <asm/unaligned.h>
 76 #include <net/netdma.h>
 77 
 78 int sysctl_tcp_timestamps __read_mostly = 1;
 79 int sysctl_tcp_window_scaling __read_mostly = 1;
 80 int sysctl_tcp_sack __read_mostly = 1;
 81 int sysctl_tcp_fack __read_mostly = 1;
 82 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
 83 EXPORT_SYMBOL(sysctl_tcp_reordering);
 84 int sysctl_tcp_dsack __read_mostly = 1;
 85 int sysctl_tcp_app_win __read_mostly = 31;
 86 int sysctl_tcp_adv_win_scale __read_mostly = 1;
 87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
 88 
 89 /* rfc5961 challenge ack rate limiting */
 90 int sysctl_tcp_challenge_ack_limit = 1000;
 91 
 92 int sysctl_tcp_stdurg __read_mostly;
 93 int sysctl_tcp_rfc1337 __read_mostly;
 94 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
 95 int sysctl_tcp_frto __read_mostly = 2;
 96 
 97 int sysctl_tcp_thin_dupack __read_mostly;
 98 
 99 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
100 int sysctl_tcp_early_retrans __read_mostly = 3;
101 
102 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
103 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
104 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
105 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
106 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
107 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
108 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
109 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_ORIG_SACK_ACKED    0x200 /* Never retransmitted data are (s)acked  */
111 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
113 #define FLAG_SACK_RENEGING      0x2000 /* snd_una advanced to a sacked seq */
114 #define FLAG_UPDATE_TS_RECENT   0x4000 /* tcp_replace_ts_recent() */
115 
116 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
120 
121 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
123 
124 /* Adapt the MSS value used to make delayed ack decision to the
125  * real world.
126  */
127 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
128 {
129         struct inet_connection_sock *icsk = inet_csk(sk);
130         const unsigned int lss = icsk->icsk_ack.last_seg_size;
131         unsigned int len;
132 
133         icsk->icsk_ack.last_seg_size = 0;
134 
135         /* skb->len may jitter because of SACKs, even if peer
136          * sends good full-sized frames.
137          */
138         len = skb_shinfo(skb)->gso_size ? : skb->len;
139         if (len >= icsk->icsk_ack.rcv_mss) {
140                 icsk->icsk_ack.rcv_mss = len;
141         } else {
142                 /* Otherwise, we make more careful check taking into account,
143                  * that SACKs block is variable.
144                  *
145                  * "len" is invariant segment length, including TCP header.
146                  */
147                 len += skb->data - skb_transport_header(skb);
148                 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
149                     /* If PSH is not set, packet should be
150                      * full sized, provided peer TCP is not badly broken.
151                      * This observation (if it is correct 8)) allows
152                      * to handle super-low mtu links fairly.
153                      */
154                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
155                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
156                         /* Subtract also invariant (if peer is RFC compliant),
157                          * tcp header plus fixed timestamp option length.
158                          * Resulting "len" is MSS free of SACK jitter.
159                          */
160                         len -= tcp_sk(sk)->tcp_header_len;
161                         icsk->icsk_ack.last_seg_size = len;
162                         if (len == lss) {
163                                 icsk->icsk_ack.rcv_mss = len;
164                                 return;
165                         }
166                 }
167                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
168                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
169                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
170         }
171 }
172 
173 static void tcp_incr_quickack(struct sock *sk)
174 {
175         struct inet_connection_sock *icsk = inet_csk(sk);
176         unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
177 
178         if (quickacks == 0)
179                 quickacks = 2;
180         if (quickacks > icsk->icsk_ack.quick)
181                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
182 }
183 
184 static void tcp_enter_quickack_mode(struct sock *sk)
185 {
186         struct inet_connection_sock *icsk = inet_csk(sk);
187         tcp_incr_quickack(sk);
188         icsk->icsk_ack.pingpong = 0;
189         icsk->icsk_ack.ato = TCP_ATO_MIN;
190 }
191 
192 /* Send ACKs quickly, if "quick" count is not exhausted
193  * and the session is not interactive.
194  */
195 
196 static inline bool tcp_in_quickack_mode(const struct sock *sk)
197 {
198         const struct inet_connection_sock *icsk = inet_csk(sk);
199 
200         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
201 }
202 
203 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
204 {
205         if (tp->ecn_flags & TCP_ECN_OK)
206                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
207 }
208 
209 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
210 {
211         if (tcp_hdr(skb)->cwr)
212                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
213 }
214 
215 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
216 {
217         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
218 }
219 
220 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
221 {
222         if (!(tp->ecn_flags & TCP_ECN_OK))
223                 return;
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 (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
236                         /* Better not delay acks, sender can have a very low cwnd */
237                         tcp_enter_quickack_mode((struct sock *)tp);
238                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
239                 }
240                 /* fallinto */
241         default:
242                 tp->ecn_flags |= TCP_ECN_SEEN;
243         }
244 }
245 
246 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
247 {
248         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
249                 tp->ecn_flags &= ~TCP_ECN_OK;
250 }
251 
252 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
253 {
254         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
255                 tp->ecn_flags &= ~TCP_ECN_OK;
256 }
257 
258 static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
259 {
260         if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
261                 return true;
262         return false;
263 }
264 
265 /* Buffer size and advertised window tuning.
266  *
267  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
268  */
269 
270 static void tcp_fixup_sndbuf(struct sock *sk)
271 {
272         int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
273 
274         sndmem *= TCP_INIT_CWND;
275         if (sk->sk_sndbuf < sndmem)
276                 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
277 }
278 
279 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
280  *
281  * All tcp_full_space() is split to two parts: "network" buffer, allocated
282  * forward and advertised in receiver window (tp->rcv_wnd) and
283  * "application buffer", required to isolate scheduling/application
284  * latencies from network.
285  * window_clamp is maximal advertised window. It can be less than
286  * tcp_full_space(), in this case tcp_full_space() - window_clamp
287  * is reserved for "application" buffer. The less window_clamp is
288  * the smoother our behaviour from viewpoint of network, but the lower
289  * throughput and the higher sensitivity of the connection to losses. 8)
290  *
291  * rcv_ssthresh is more strict window_clamp used at "slow start"
292  * phase to predict further behaviour of this connection.
293  * It is used for two goals:
294  * - to enforce header prediction at sender, even when application
295  *   requires some significant "application buffer". It is check #1.
296  * - to prevent pruning of receive queue because of misprediction
297  *   of receiver window. Check #2.
298  *
299  * The scheme does not work when sender sends good segments opening
300  * window and then starts to feed us spaghetti. But it should work
301  * in common situations. Otherwise, we have to rely on queue collapsing.
302  */
303 
304 /* Slow part of check#2. */
305 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
306 {
307         struct tcp_sock *tp = tcp_sk(sk);
308         /* Optimize this! */
309         int truesize = tcp_win_from_space(skb->truesize) >> 1;
310         int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
311 
312         while (tp->rcv_ssthresh <= window) {
313                 if (truesize <= skb->len)
314                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
315 
316                 truesize >>= 1;
317                 window >>= 1;
318         }
319         return 0;
320 }
321 
322 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
323 {
324         struct tcp_sock *tp = tcp_sk(sk);
325 
326         /* Check #1 */
327         if (tp->rcv_ssthresh < tp->window_clamp &&
328             (int)tp->rcv_ssthresh < tcp_space(sk) &&
329             !sk_under_memory_pressure(sk)) {
330                 int incr;
331 
332                 /* Check #2. Increase window, if skb with such overhead
333                  * will fit to rcvbuf in future.
334                  */
335                 if (tcp_win_from_space(skb->truesize) <= skb->len)
336                         incr = 2 * tp->advmss;
337                 else
338                         incr = __tcp_grow_window(sk, skb);
339 
340                 if (incr) {
341                         incr = max_t(int, incr, 2 * skb->len);
342                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
343                                                tp->window_clamp);
344                         inet_csk(sk)->icsk_ack.quick |= 1;
345                 }
346         }
347 }
348 
349 /* 3. Tuning rcvbuf, when connection enters established state. */
350 static void tcp_fixup_rcvbuf(struct sock *sk)
351 {
352         u32 mss = tcp_sk(sk)->advmss;
353         int rcvmem;
354 
355         rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
356                  tcp_default_init_rwnd(mss);
357 
358         if (sk->sk_rcvbuf < rcvmem)
359                 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
360 }
361 
362 /* 4. Try to fixup all. It is made immediately after connection enters
363  *    established state.
364  */
365 void tcp_init_buffer_space(struct sock *sk)
366 {
367         struct tcp_sock *tp = tcp_sk(sk);
368         int maxwin;
369 
370         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
371                 tcp_fixup_rcvbuf(sk);
372         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
373                 tcp_fixup_sndbuf(sk);
374 
375         tp->rcvq_space.space = tp->rcv_wnd;
376 
377         maxwin = tcp_full_space(sk);
378 
379         if (tp->window_clamp >= maxwin) {
380                 tp->window_clamp = maxwin;
381 
382                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
383                         tp->window_clamp = max(maxwin -
384                                                (maxwin >> sysctl_tcp_app_win),
385                                                4 * tp->advmss);
386         }
387 
388         /* Force reservation of one segment. */
389         if (sysctl_tcp_app_win &&
390             tp->window_clamp > 2 * tp->advmss &&
391             tp->window_clamp + tp->advmss > maxwin)
392                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
393 
394         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
395         tp->snd_cwnd_stamp = tcp_time_stamp;
396 }
397 
398 /* 5. Recalculate window clamp after socket hit its memory bounds. */
399 static void tcp_clamp_window(struct sock *sk)
400 {
401         struct tcp_sock *tp = tcp_sk(sk);
402         struct inet_connection_sock *icsk = inet_csk(sk);
403 
404         icsk->icsk_ack.quick = 0;
405 
406         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
407             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
408             !sk_under_memory_pressure(sk) &&
409             sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
410                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
411                                     sysctl_tcp_rmem[2]);
412         }
413         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
414                 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
415 }
416 
417 /* Initialize RCV_MSS value.
418  * RCV_MSS is an our guess about MSS used by the peer.
419  * We haven't any direct information about the MSS.
420  * It's better to underestimate the RCV_MSS rather than overestimate.
421  * Overestimations make us ACKing less frequently than needed.
422  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
423  */
424 void tcp_initialize_rcv_mss(struct sock *sk)
425 {
426         const struct tcp_sock *tp = tcp_sk(sk);
427         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
428 
429         hint = min(hint, tp->rcv_wnd / 2);
430         hint = min(hint, TCP_MSS_DEFAULT);
431         hint = max(hint, TCP_MIN_MSS);
432 
433         inet_csk(sk)->icsk_ack.rcv_mss = hint;
434 }
435 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
436 
437 /* Receiver "autotuning" code.
438  *
439  * The algorithm for RTT estimation w/o timestamps is based on
440  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
441  * <http://public.lanl.gov/radiant/pubs.html#DRS>
442  *
443  * More detail on this code can be found at
444  * <http://staff.psc.edu/jheffner/>,
445  * though this reference is out of date.  A new paper
446  * is pending.
447  */
448 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
449 {
450         u32 new_sample = tp->rcv_rtt_est.rtt;
451         long m = sample;
452 
453         if (m == 0)
454                 m = 1;
455 
456         if (new_sample != 0) {
457                 /* If we sample in larger samples in the non-timestamp
458                  * case, we could grossly overestimate the RTT especially
459                  * with chatty applications or bulk transfer apps which
460                  * are stalled on filesystem I/O.
461                  *
462                  * Also, since we are only going for a minimum in the
463                  * non-timestamp case, we do not smooth things out
464                  * else with timestamps disabled convergence takes too
465                  * long.
466                  */
467                 if (!win_dep) {
468                         m -= (new_sample >> 3);
469                         new_sample += m;
470                 } else {
471                         m <<= 3;
472                         if (m < new_sample)
473                                 new_sample = m;
474                 }
475         } else {
476                 /* No previous measure. */
477                 new_sample = m << 3;
478         }
479 
480         if (tp->rcv_rtt_est.rtt != new_sample)
481                 tp->rcv_rtt_est.rtt = new_sample;
482 }
483 
484 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
485 {
486         if (tp->rcv_rtt_est.time == 0)
487                 goto new_measure;
488         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
489                 return;
490         tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
491 
492 new_measure:
493         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
494         tp->rcv_rtt_est.time = tcp_time_stamp;
495 }
496 
497 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
498                                           const struct sk_buff *skb)
499 {
500         struct tcp_sock *tp = tcp_sk(sk);
501         if (tp->rx_opt.rcv_tsecr &&
502             (TCP_SKB_CB(skb)->end_seq -
503              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
504                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
505 }
506 
507 /*
508  * This function should be called every time data is copied to user space.
509  * It calculates the appropriate TCP receive buffer space.
510  */
511 void tcp_rcv_space_adjust(struct sock *sk)
512 {
513         struct tcp_sock *tp = tcp_sk(sk);
514         int time;
515         int space;
516 
517         if (tp->rcvq_space.time == 0)
518                 goto new_measure;
519 
520         time = tcp_time_stamp - tp->rcvq_space.time;
521         if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
522                 return;
523 
524         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
525 
526         space = max(tp->rcvq_space.space, space);
527 
528         if (tp->rcvq_space.space != space) {
529                 int rcvmem;
530 
531                 tp->rcvq_space.space = space;
532 
533                 if (sysctl_tcp_moderate_rcvbuf &&
534                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
535                         int new_clamp = space;
536 
537                         /* Receive space grows, normalize in order to
538                          * take into account packet headers and sk_buff
539                          * structure overhead.
540                          */
541                         space /= tp->advmss;
542                         if (!space)
543                                 space = 1;
544                         rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
545                         while (tcp_win_from_space(rcvmem) < tp->advmss)
546                                 rcvmem += 128;
547                         space *= rcvmem;
548                         space = min(space, sysctl_tcp_rmem[2]);
549                         if (space > sk->sk_rcvbuf) {
550                                 sk->sk_rcvbuf = space;
551 
552                                 /* Make the window clamp follow along.  */
553                                 tp->window_clamp = new_clamp;
554                         }
555                 }
556         }
557 
558 new_measure:
559         tp->rcvq_space.seq = tp->copied_seq;
560         tp->rcvq_space.time = tcp_time_stamp;
561 }
562 
563 /* There is something which you must keep in mind when you analyze the
564  * behavior of the tp->ato delayed ack timeout interval.  When a
565  * connection starts up, we want to ack as quickly as possible.  The
566  * problem is that "good" TCP's do slow start at the beginning of data
567  * transmission.  The means that until we send the first few ACK's the
568  * sender will sit on his end and only queue most of his data, because
569  * he can only send snd_cwnd unacked packets at any given time.  For
570  * each ACK we send, he increments snd_cwnd and transmits more of his
571  * queue.  -DaveM
572  */
573 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
574 {
575         struct tcp_sock *tp = tcp_sk(sk);
576         struct inet_connection_sock *icsk = inet_csk(sk);
577         u32 now;
578 
579         inet_csk_schedule_ack(sk);
580 
581         tcp_measure_rcv_mss(sk, skb);
582 
583         tcp_rcv_rtt_measure(tp);
584 
585         now = tcp_time_stamp;
586 
587         if (!icsk->icsk_ack.ato) {
588                 /* The _first_ data packet received, initialize
589                  * delayed ACK engine.
590                  */
591                 tcp_incr_quickack(sk);
592                 icsk->icsk_ack.ato = TCP_ATO_MIN;
593         } else {
594                 int m = now - icsk->icsk_ack.lrcvtime;
595 
596                 if (m <= TCP_ATO_MIN / 2) {
597                         /* The fastest case is the first. */
598                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
599                 } else if (m < icsk->icsk_ack.ato) {
600                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
601                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
602                                 icsk->icsk_ack.ato = icsk->icsk_rto;
603                 } else if (m > icsk->icsk_rto) {
604                         /* Too long gap. Apparently sender failed to
605                          * restart window, so that we send ACKs quickly.
606                          */
607                         tcp_incr_quickack(sk);
608                         sk_mem_reclaim(sk);
609                 }
610         }
611         icsk->icsk_ack.lrcvtime = now;
612 
613         TCP_ECN_check_ce(tp, skb);
614 
615         if (skb->len >= 128)
616                 tcp_grow_window(sk, skb);
617 }
618 
619 /* Called to compute a smoothed rtt estimate. The data fed to this
620  * routine either comes from timestamps, or from segments that were
621  * known _not_ to have been retransmitted [see Karn/Partridge
622  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
623  * piece by Van Jacobson.
624  * NOTE: the next three routines used to be one big routine.
625  * To save cycles in the RFC 1323 implementation it was better to break
626  * it up into three procedures. -- erics
627  */
628 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
629 {
630         struct tcp_sock *tp = tcp_sk(sk);
631         long m = mrtt; /* RTT */
632 
633         /*      The following amusing code comes from Jacobson's
634          *      article in SIGCOMM '88.  Note that rtt and mdev
635          *      are scaled versions of rtt and mean deviation.
636          *      This is designed to be as fast as possible
637          *      m stands for "measurement".
638          *
639          *      On a 1990 paper the rto value is changed to:
640          *      RTO = rtt + 4 * mdev
641          *
642          * Funny. This algorithm seems to be very broken.
643          * These formulae increase RTO, when it should be decreased, increase
644          * too slowly, when it should be increased quickly, decrease too quickly
645          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
646          * does not matter how to _calculate_ it. Seems, it was trap
647          * that VJ failed to avoid. 8)
648          */
649         if (m == 0)
650                 m = 1;
651         if (tp->srtt != 0) {
652                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
653                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
654                 if (m < 0) {
655                         m = -m;         /* m is now abs(error) */
656                         m -= (tp->mdev >> 2);   /* similar update on mdev */
657                         /* This is similar to one of Eifel findings.
658                          * Eifel blocks mdev updates when rtt decreases.
659                          * This solution is a bit different: we use finer gain
660                          * for mdev in this case (alpha*beta).
661                          * Like Eifel it also prevents growth of rto,
662                          * but also it limits too fast rto decreases,
663                          * happening in pure Eifel.
664                          */
665                         if (m > 0)
666                                 m >>= 3;
667                 } else {
668                         m -= (tp->mdev >> 2);   /* similar update on mdev */
669                 }
670                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
671                 if (tp->mdev > tp->mdev_max) {
672                         tp->mdev_max = tp->mdev;
673                         if (tp->mdev_max > tp->rttvar)
674                                 tp->rttvar = tp->mdev_max;
675                 }
676                 if (after(tp->snd_una, tp->rtt_seq)) {
677                         if (tp->mdev_max < tp->rttvar)
678                                 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
679                         tp->rtt_seq = tp->snd_nxt;
680                         tp->mdev_max = tcp_rto_min(sk);
681                 }
682         } else {
683                 /* no previous measure. */
684                 tp->srtt = m << 3;      /* take the measured time to be rtt */
685                 tp->mdev = m << 1;      /* make sure rto = 3*rtt */
686                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
687                 tp->rtt_seq = tp->snd_nxt;
688         }
689 }
690 
691 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
692  * Note: TCP stack does not yet implement pacing.
693  * FQ packet scheduler can be used to implement cheap but effective
694  * TCP pacing, to smooth the burst on large writes when packets
695  * in flight is significantly lower than cwnd (or rwin)
696  */
697 static void tcp_update_pacing_rate(struct sock *sk)
698 {
699         const struct tcp_sock *tp = tcp_sk(sk);
700         u64 rate;
701 
702         /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
703         rate = (u64)tp->mss_cache * 2 * (HZ << 3);
704 
705         rate *= max(tp->snd_cwnd, tp->packets_out);
706 
707         /* Correction for small srtt : minimum srtt being 8 (1 jiffy << 3),
708          * be conservative and assume srtt = 1 (125 us instead of 1.25 ms)
709          * We probably need usec resolution in the future.
710          * Note: This also takes care of possible srtt=0 case,
711          * when tcp_rtt_estimator() was not yet called.
712          */
713         if (tp->srtt > 8 + 2)
714                 do_div(rate, tp->srtt);
715 
716         sk->sk_pacing_rate = min_t(u64, rate, ~0U);
717 }
718 
719 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
720  * routine referred to above.
721  */
722 void tcp_set_rto(struct sock *sk)
723 {
724         const struct tcp_sock *tp = tcp_sk(sk);
725         /* Old crap is replaced with new one. 8)
726          *
727          * More seriously:
728          * 1. If rtt variance happened to be less 50msec, it is hallucination.
729          *    It cannot be less due to utterly erratic ACK generation made
730          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
731          *    to do with delayed acks, because at cwnd>2 true delack timeout
732          *    is invisible. Actually, Linux-2.4 also generates erratic
733          *    ACKs in some circumstances.
734          */
735         inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
736 
737         /* 2. Fixups made earlier cannot be right.
738          *    If we do not estimate RTO correctly without them,
739          *    all the algo is pure shit and should be replaced
740          *    with correct one. It is exactly, which we pretend to do.
741          */
742 
743         /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
744          * guarantees that rto is higher.
745          */
746         tcp_bound_rto(sk);
747 }
748 
749 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
750 {
751         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
752 
753         if (!cwnd)
754                 cwnd = TCP_INIT_CWND;
755         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
756 }
757 
758 /*
759  * Packet counting of FACK is based on in-order assumptions, therefore TCP
760  * disables it when reordering is detected
761  */
762 void tcp_disable_fack(struct tcp_sock *tp)
763 {
764         /* RFC3517 uses different metric in lost marker => reset on change */
765         if (tcp_is_fack(tp))
766                 tp->lost_skb_hint = NULL;
767         tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
768 }
769 
770 /* Take a notice that peer is sending D-SACKs */
771 static void tcp_dsack_seen(struct tcp_sock *tp)
772 {
773         tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
774 }
775 
776 static void tcp_update_reordering(struct sock *sk, const int metric,
777                                   const int ts)
778 {
779         struct tcp_sock *tp = tcp_sk(sk);
780         if (metric > tp->reordering) {
781                 int mib_idx;
782 
783                 tp->reordering = min(TCP_MAX_REORDERING, metric);
784 
785                 /* This exciting event is worth to be remembered. 8) */
786                 if (ts)
787                         mib_idx = LINUX_MIB_TCPTSREORDER;
788                 else if (tcp_is_reno(tp))
789                         mib_idx = LINUX_MIB_TCPRENOREORDER;
790                 else if (tcp_is_fack(tp))
791                         mib_idx = LINUX_MIB_TCPFACKREORDER;
792                 else
793                         mib_idx = LINUX_MIB_TCPSACKREORDER;
794 
795                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
796 #if FASTRETRANS_DEBUG > 1
797                 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
798                          tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
799                          tp->reordering,
800                          tp->fackets_out,
801                          tp->sacked_out,
802                          tp->undo_marker ? tp->undo_retrans : 0);
803 #endif
804                 tcp_disable_fack(tp);
805         }
806 
807         if (metric > 0)
808                 tcp_disable_early_retrans(tp);
809 }
810 
811 /* This must be called before lost_out is incremented */
812 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
813 {
814         if ((tp->retransmit_skb_hint == NULL) ||
815             before(TCP_SKB_CB(skb)->seq,
816                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
817                 tp->retransmit_skb_hint = skb;
818 
819         if (!tp->lost_out ||
820             after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
821                 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
822 }
823 
824 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
825 {
826         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
827                 tcp_verify_retransmit_hint(tp, skb);
828 
829                 tp->lost_out += tcp_skb_pcount(skb);
830                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
831         }
832 }
833 
834 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
835                                             struct sk_buff *skb)
836 {
837         tcp_verify_retransmit_hint(tp, skb);
838 
839         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
840                 tp->lost_out += tcp_skb_pcount(skb);
841                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
842         }
843 }
844 
845 /* This procedure tags the retransmission queue when SACKs arrive.
846  *
847  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
848  * Packets in queue with these bits set are counted in variables
849  * sacked_out, retrans_out and lost_out, correspondingly.
850  *
851  * Valid combinations are:
852  * Tag  InFlight        Description
853  * 0    1               - orig segment is in flight.
854  * S    0               - nothing flies, orig reached receiver.
855  * L    0               - nothing flies, orig lost by net.
856  * R    2               - both orig and retransmit are in flight.
857  * L|R  1               - orig is lost, retransmit is in flight.
858  * S|R  1               - orig reached receiver, retrans is still in flight.
859  * (L|S|R is logically valid, it could occur when L|R is sacked,
860  *  but it is equivalent to plain S and code short-curcuits it to S.
861  *  L|S is logically invalid, it would mean -1 packet in flight 8))
862  *
863  * These 6 states form finite state machine, controlled by the following events:
864  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
865  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
866  * 3. Loss detection event of two flavors:
867  *      A. Scoreboard estimator decided the packet is lost.
868  *         A'. Reno "three dupacks" marks head of queue lost.
869  *         A''. Its FACK modification, head until snd.fack is lost.
870  *      B. SACK arrives sacking SND.NXT at the moment, when the
871  *         segment was retransmitted.
872  * 4. D-SACK added new rule: D-SACK changes any tag to S.
873  *
874  * It is pleasant to note, that state diagram turns out to be commutative,
875  * so that we are allowed not to be bothered by order of our actions,
876  * when multiple events arrive simultaneously. (see the function below).
877  *
878  * Reordering detection.
879  * --------------------
880  * Reordering metric is maximal distance, which a packet can be displaced
881  * in packet stream. With SACKs we can estimate it:
882  *
883  * 1. SACK fills old hole and the corresponding segment was not
884  *    ever retransmitted -> reordering. Alas, we cannot use it
885  *    when segment was retransmitted.
886  * 2. The last flaw is solved with D-SACK. D-SACK arrives
887  *    for retransmitted and already SACKed segment -> reordering..
888  * Both of these heuristics are not used in Loss state, when we cannot
889  * account for retransmits accurately.
890  *
891  * SACK block validation.
892  * ----------------------
893  *
894  * SACK block range validation checks that the received SACK block fits to
895  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
896  * Note that SND.UNA is not included to the range though being valid because
897  * it means that the receiver is rather inconsistent with itself reporting
898  * SACK reneging when it should advance SND.UNA. Such SACK block this is
899  * perfectly valid, however, in light of RFC2018 which explicitly states
900  * that "SACK block MUST reflect the newest segment.  Even if the newest
901  * segment is going to be discarded ...", not that it looks very clever
902  * in case of head skb. Due to potentional receiver driven attacks, we
903  * choose to avoid immediate execution of a walk in write queue due to
904  * reneging and defer head skb's loss recovery to standard loss recovery
905  * procedure that will eventually trigger (nothing forbids us doing this).
906  *
907  * Implements also blockage to start_seq wrap-around. Problem lies in the
908  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
909  * there's no guarantee that it will be before snd_nxt (n). The problem
910  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
911  * wrap (s_w):
912  *
913  *         <- outs wnd ->                          <- wrapzone ->
914  *         u     e      n                         u_w   e_w  s n_w
915  *         |     |      |                          |     |   |  |
916  * |<------------+------+----- TCP seqno space --------------+---------->|
917  * ...-- <2^31 ->|                                           |<--------...
918  * ...---- >2^31 ------>|                                    |<--------...
919  *
920  * Current code wouldn't be vulnerable but it's better still to discard such
921  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
922  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
923  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
924  * equal to the ideal case (infinite seqno space without wrap caused issues).
925  *
926  * With D-SACK the lower bound is extended to cover sequence space below
927  * SND.UNA down to undo_marker, which is the last point of interest. Yet
928  * again, D-SACK block must not to go across snd_una (for the same reason as
929  * for the normal SACK blocks, explained above). But there all simplicity
930  * ends, TCP might receive valid D-SACKs below that. As long as they reside
931  * fully below undo_marker they do not affect behavior in anyway and can
932  * therefore be safely ignored. In rare cases (which are more or less
933  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
934  * fragmentation and packet reordering past skb's retransmission. To consider
935  * them correctly, the acceptable range must be extended even more though
936  * the exact amount is rather hard to quantify. However, tp->max_window can
937  * be used as an exaggerated estimate.
938  */
939 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
940                                    u32 start_seq, u32 end_seq)
941 {
942         /* Too far in future, or reversed (interpretation is ambiguous) */
943         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
944                 return false;
945 
946         /* Nasty start_seq wrap-around check (see comments above) */
947         if (!before(start_seq, tp->snd_nxt))
948                 return false;
949 
950         /* In outstanding window? ...This is valid exit for D-SACKs too.
951          * start_seq == snd_una is non-sensical (see comments above)
952          */
953         if (after(start_seq, tp->snd_una))
954                 return true;
955 
956         if (!is_dsack || !tp->undo_marker)
957                 return false;
958 
959         /* ...Then it's D-SACK, and must reside below snd_una completely */
960         if (after(end_seq, tp->snd_una))
961                 return false;
962 
963         if (!before(start_seq, tp->undo_marker))
964                 return true;
965 
966         /* Too old */
967         if (!after(end_seq, tp->undo_marker))
968                 return false;
969 
970         /* Undo_marker boundary crossing (overestimates a lot). Known already:
971          *   start_seq < undo_marker and end_seq >= undo_marker.
972          */
973         return !before(start_seq, end_seq - tp->max_window);
974 }
975 
976 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
977  * Event "B". Later note: FACK people cheated me again 8), we have to account
978  * for reordering! Ugly, but should help.
979  *
980  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
981  * less than what is now known to be received by the other end (derived from
982  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
983  * retransmitted skbs to avoid some costly processing per ACKs.
984  */
985 static void tcp_mark_lost_retrans(struct sock *sk)
986 {
987         const struct inet_connection_sock *icsk = inet_csk(sk);
988         struct tcp_sock *tp = tcp_sk(sk);
989         struct sk_buff *skb;
990         int cnt = 0;
991         u32 new_low_seq = tp->snd_nxt;
992         u32 received_upto = tcp_highest_sack_seq(tp);
993 
994         if (!tcp_is_fack(tp) || !tp->retrans_out ||
995             !after(received_upto, tp->lost_retrans_low) ||
996             icsk->icsk_ca_state != TCP_CA_Recovery)
997                 return;
998 
999         tcp_for_write_queue(skb, sk) {
1000                 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1001 
1002                 if (skb == tcp_send_head(sk))
1003                         break;
1004                 if (cnt == tp->retrans_out)
1005                         break;
1006                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1007                         continue;
1008 
1009                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1010                         continue;
1011 
1012                 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1013                  * constraint here (see above) but figuring out that at
1014                  * least tp->reordering SACK blocks reside between ack_seq
1015                  * and received_upto is not easy task to do cheaply with
1016                  * the available datastructures.
1017                  *
1018                  * Whether FACK should check here for tp->reordering segs
1019                  * in-between one could argue for either way (it would be
1020                  * rather simple to implement as we could count fack_count
1021                  * during the walk and do tp->fackets_out - fack_count).
1022                  */
1023                 if (after(received_upto, ack_seq)) {
1024                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1025                         tp->retrans_out -= tcp_skb_pcount(skb);
1026 
1027                         tcp_skb_mark_lost_uncond_verify(tp, skb);
1028                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1029                 } else {
1030                         if (before(ack_seq, new_low_seq))
1031                                 new_low_seq = ack_seq;
1032                         cnt += tcp_skb_pcount(skb);
1033                 }
1034         }
1035 
1036         if (tp->retrans_out)
1037                 tp->lost_retrans_low = new_low_seq;
1038 }
1039 
1040 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1041                             struct tcp_sack_block_wire *sp, int num_sacks,
1042                             u32 prior_snd_una)
1043 {
1044         struct tcp_sock *tp = tcp_sk(sk);
1045         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1046         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1047         bool dup_sack = false;
1048 
1049         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1050                 dup_sack = true;
1051                 tcp_dsack_seen(tp);
1052                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1053         } else if (num_sacks > 1) {
1054                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1055                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1056 
1057                 if (!after(end_seq_0, end_seq_1) &&
1058                     !before(start_seq_0, start_seq_1)) {
1059                         dup_sack = true;
1060                         tcp_dsack_seen(tp);
1061                         NET_INC_STATS_BH(sock_net(sk),
1062                                         LINUX_MIB_TCPDSACKOFORECV);
1063                 }
1064         }
1065 
1066         /* D-SACK for already forgotten data... Do dumb counting. */
1067         if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1068             !after(end_seq_0, prior_snd_una) &&
1069             after(end_seq_0, tp->undo_marker))
1070                 tp->undo_retrans--;
1071 
1072         return dup_sack;
1073 }
1074 
1075 struct tcp_sacktag_state {
1076         int reord;
1077         int fack_count;
1078         int flag;
1079         s32 rtt; /* RTT measured by SACKing never-retransmitted data */
1080 };
1081 
1082 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1083  * the incoming SACK may not exactly match but we can find smaller MSS
1084  * aligned portion of it that matches. Therefore we might need to fragment
1085  * which may fail and creates some hassle (caller must handle error case
1086  * returns).
1087  *
1088  * FIXME: this could be merged to shift decision code
1089  */
1090 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1091                                   u32 start_seq, u32 end_seq)
1092 {
1093         int err;
1094         bool in_sack;
1095         unsigned int pkt_len;
1096         unsigned int mss;
1097 
1098         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1099                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1100 
1101         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1102             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1103                 mss = tcp_skb_mss(skb);
1104                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1105 
1106                 if (!in_sack) {
1107                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1108                         if (pkt_len < mss)
1109                                 pkt_len = mss;
1110                 } else {
1111                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1112                         if (pkt_len < mss)
1113                                 return -EINVAL;
1114                 }
1115 
1116                 /* Round if necessary so that SACKs cover only full MSSes
1117                  * and/or the remaining small portion (if present)
1118                  */
1119                 if (pkt_len > mss) {
1120                         unsigned int new_len = (pkt_len / mss) * mss;
1121                         if (!in_sack && new_len < pkt_len) {
1122                                 new_len += mss;
1123                                 if (new_len >= skb->len)
1124                                         return 0;
1125                         }
1126                         pkt_len = new_len;
1127                 }
1128                 err = tcp_fragment(sk, skb, pkt_len, mss);
1129                 if (err < 0)
1130                         return err;
1131         }
1132 
1133         return in_sack;
1134 }
1135 
1136 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1137 static u8 tcp_sacktag_one(struct sock *sk,
1138                           struct tcp_sacktag_state *state, u8 sacked,
1139                           u32 start_seq, u32 end_seq,
1140                           int dup_sack, int pcount, u32 xmit_time)
1141 {
1142         struct tcp_sock *tp = tcp_sk(sk);
1143         int fack_count = state->fack_count;
1144 
1145         /* Account D-SACK for retransmitted packet. */
1146         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1147                 if (tp->undo_marker && tp->undo_retrans > 0 &&
1148                     after(end_seq, tp->undo_marker))
1149                         tp->undo_retrans--;
1150                 if (sacked & TCPCB_SACKED_ACKED)
1151                         state->reord = min(fack_count, state->reord);
1152         }
1153 
1154         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1155         if (!after(end_seq, tp->snd_una))
1156                 return sacked;
1157 
1158         if (!(sacked & TCPCB_SACKED_ACKED)) {
1159                 if (sacked & TCPCB_SACKED_RETRANS) {
1160                         /* If the segment is not tagged as lost,
1161                          * we do not clear RETRANS, believing
1162                          * that retransmission is still in flight.
1163                          */
1164                         if (sacked & TCPCB_LOST) {
1165                                 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1166                                 tp->lost_out -= pcount;
1167                                 tp->retrans_out -= pcount;
1168                         }
1169                 } else {
1170                         if (!(sacked & TCPCB_RETRANS)) {
1171                                 /* New sack for not retransmitted frame,
1172                                  * which was in hole. It is reordering.
1173                                  */
1174                                 if (before(start_seq,
1175                                            tcp_highest_sack_seq(tp)))
1176                                         state->reord = min(fack_count,
1177                                                            state->reord);
1178                                 if (!after(end_seq, tp->high_seq))
1179                                         state->flag |= FLAG_ORIG_SACK_ACKED;
1180                                 /* Pick the earliest sequence sacked for RTT */
1181                                 if (state->rtt < 0)
1182                                         state->rtt = tcp_time_stamp - xmit_time;
1183                         }
1184 
1185                         if (sacked & TCPCB_LOST) {
1186                                 sacked &= ~TCPCB_LOST;
1187                                 tp->lost_out -= pcount;
1188                         }
1189                 }
1190 
1191                 sacked |= TCPCB_SACKED_ACKED;
1192                 state->flag |= FLAG_DATA_SACKED;
1193                 tp->sacked_out += pcount;
1194 
1195                 fack_count += pcount;
1196 
1197                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1198                 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1199                     before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1200                         tp->lost_cnt_hint += pcount;
1201 
1202                 if (fack_count > tp->fackets_out)
1203                         tp->fackets_out = fack_count;
1204         }
1205 
1206         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1207          * frames and clear it. undo_retrans is decreased above, L|R frames
1208          * are accounted above as well.
1209          */
1210         if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1211                 sacked &= ~TCPCB_SACKED_RETRANS;
1212                 tp->retrans_out -= pcount;
1213         }
1214 
1215         return sacked;
1216 }
1217 
1218 /* Shift newly-SACKed bytes from this skb to the immediately previous
1219  * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1220  */
1221 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1222                             struct tcp_sacktag_state *state,
1223                             unsigned int pcount, int shifted, int mss,
1224                             bool dup_sack)
1225 {
1226         struct tcp_sock *tp = tcp_sk(sk);
1227         struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1228         u32 start_seq = TCP_SKB_CB(skb)->seq;   /* start of newly-SACKed */
1229         u32 end_seq = start_seq + shifted;      /* end of newly-SACKed */
1230 
1231         BUG_ON(!pcount);
1232 
1233         /* Adjust counters and hints for the newly sacked sequence
1234          * range but discard the return value since prev is already
1235          * marked. We must tag the range first because the seq
1236          * advancement below implicitly advances
1237          * tcp_highest_sack_seq() when skb is highest_sack.
1238          */
1239         tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1240                         start_seq, end_seq, dup_sack, pcount,
1241                         TCP_SKB_CB(skb)->when);
1242 
1243         if (skb == tp->lost_skb_hint)
1244                 tp->lost_cnt_hint += pcount;
1245 
1246         TCP_SKB_CB(prev)->end_seq += shifted;
1247         TCP_SKB_CB(skb)->seq += shifted;
1248 
1249         skb_shinfo(prev)->gso_segs += pcount;
1250         BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1251         skb_shinfo(skb)->gso_segs -= pcount;
1252 
1253         /* When we're adding to gso_segs == 1, gso_size will be zero,
1254          * in theory this shouldn't be necessary but as long as DSACK
1255          * code can come after this skb later on it's better to keep
1256          * setting gso_size to something.
1257          */
1258         if (!skb_shinfo(prev)->gso_size) {
1259                 skb_shinfo(prev)->gso_size = mss;
1260                 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1261         }
1262 
1263         /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1264         if (skb_shinfo(skb)->gso_segs <= 1) {
1265                 skb_shinfo(skb)->gso_size = 0;
1266                 skb_shinfo(skb)->gso_type = 0;
1267         }
1268 
1269         /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1270         TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1271 
1272         if (skb->len > 0) {
1273                 BUG_ON(!tcp_skb_pcount(skb));
1274                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1275                 return false;
1276         }
1277 
1278         /* Whole SKB was eaten :-) */
1279 
1280         if (skb == tp->retransmit_skb_hint)
1281                 tp->retransmit_skb_hint = prev;
1282         if (skb == tp->lost_skb_hint) {
1283                 tp->lost_skb_hint = prev;
1284                 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1285         }
1286 
1287         TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1288         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1289                 TCP_SKB_CB(prev)->end_seq++;
1290 
1291         if (skb == tcp_highest_sack(sk))
1292                 tcp_advance_highest_sack(sk, skb);
1293 
1294         tcp_unlink_write_queue(skb, sk);
1295         sk_wmem_free_skb(sk, skb);
1296 
1297         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1298 
1299         return true;
1300 }
1301 
1302 /* I wish gso_size would have a bit more sane initialization than
1303  * something-or-zero which complicates things
1304  */
1305 static int tcp_skb_seglen(const struct sk_buff *skb)
1306 {
1307         return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1308 }
1309 
1310 /* Shifting pages past head area doesn't work */
1311 static int skb_can_shift(const struct sk_buff *skb)
1312 {
1313         return !skb_headlen(skb) && skb_is_nonlinear(skb);
1314 }
1315 
1316 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1317  * skb.
1318  */
1319 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1320                                           struct tcp_sacktag_state *state,
1321                                           u32 start_seq, u32 end_seq,
1322                                           bool dup_sack)
1323 {
1324         struct tcp_sock *tp = tcp_sk(sk);
1325         struct sk_buff *prev;
1326         int mss;
1327         int pcount = 0;
1328         int len;
1329         int in_sack;
1330 
1331         if (!sk_can_gso(sk))
1332                 goto fallback;
1333 
1334         /* Normally R but no L won't result in plain S */
1335         if (!dup_sack &&
1336             (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1337                 goto fallback;
1338         if (!skb_can_shift(skb))
1339                 goto fallback;
1340         /* This frame is about to be dropped (was ACKed). */
1341         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1342                 goto fallback;
1343 
1344         /* Can only happen with delayed DSACK + discard craziness */
1345         if (unlikely(skb == tcp_write_queue_head(sk)))
1346                 goto fallback;
1347         prev = tcp_write_queue_prev(sk, skb);
1348 
1349         if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1350                 goto fallback;
1351 
1352         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1353                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1354 
1355         if (in_sack) {
1356                 len = skb->len;
1357                 pcount = tcp_skb_pcount(skb);
1358                 mss = tcp_skb_seglen(skb);
1359 
1360                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1361                  * drop this restriction as unnecessary
1362                  */
1363                 if (mss != tcp_skb_seglen(prev))
1364                         goto fallback;
1365         } else {
1366                 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1367                         goto noop;
1368                 /* CHECKME: This is non-MSS split case only?, this will
1369                  * cause skipped skbs due to advancing loop btw, original
1370                  * has that feature too
1371                  */
1372                 if (tcp_skb_pcount(skb) <= 1)
1373                         goto noop;
1374 
1375                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1376                 if (!in_sack) {
1377                         /* TODO: head merge to next could be attempted here
1378                          * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1379                          * though it might not be worth of the additional hassle
1380                          *
1381                          * ...we can probably just fallback to what was done
1382                          * previously. We could try merging non-SACKed ones
1383                          * as well but it probably isn't going to buy off
1384                          * because later SACKs might again split them, and
1385                          * it would make skb timestamp tracking considerably
1386                          * harder problem.
1387                          */
1388                         goto fallback;
1389                 }
1390 
1391                 len = end_seq - TCP_SKB_CB(skb)->seq;
1392                 BUG_ON(len < 0);
1393                 BUG_ON(len > skb->len);
1394 
1395                 /* MSS boundaries should be honoured or else pcount will
1396                  * severely break even though it makes things bit trickier.
1397                  * Optimize common case to avoid most of the divides
1398                  */
1399                 mss = tcp_skb_mss(skb);
1400 
1401                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1402                  * drop this restriction as unnecessary
1403                  */
1404                 if (mss != tcp_skb_seglen(prev))
1405                         goto fallback;
1406 
1407                 if (len == mss) {
1408                         pcount = 1;
1409                 } else if (len < mss) {
1410                         goto noop;
1411                 } else {
1412                         pcount = len / mss;
1413                         len = pcount * mss;
1414                 }
1415         }
1416 
1417         /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1418         if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1419                 goto fallback;
1420 
1421         if (!skb_shift(prev, skb, len))
1422                 goto fallback;
1423         if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1424                 goto out;
1425 
1426         /* Hole filled allows collapsing with the next as well, this is very
1427          * useful when hole on every nth skb pattern happens
1428          */
1429         if (prev == tcp_write_queue_tail(sk))
1430                 goto out;
1431         skb = tcp_write_queue_next(sk, prev);
1432 
1433         if (!skb_can_shift(skb) ||
1434             (skb == tcp_send_head(sk)) ||
1435             ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1436             (mss != tcp_skb_seglen(skb)))
1437                 goto out;
1438 
1439         len = skb->len;
1440         if (skb_shift(prev, skb, len)) {
1441                 pcount += tcp_skb_pcount(skb);
1442                 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1443         }
1444 
1445 out:
1446         state->fack_count += pcount;
1447         return prev;
1448 
1449 noop:
1450         return skb;
1451 
1452 fallback:
1453         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1454         return NULL;
1455 }
1456 
1457 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1458                                         struct tcp_sack_block *next_dup,
1459                                         struct tcp_sacktag_state *state,
1460                                         u32 start_seq, u32 end_seq,
1461                                         bool dup_sack_in)
1462 {
1463         struct tcp_sock *tp = tcp_sk(sk);
1464         struct sk_buff *tmp;
1465 
1466         tcp_for_write_queue_from(skb, sk) {
1467                 int in_sack = 0;
1468                 bool dup_sack = dup_sack_in;
1469 
1470                 if (skb == tcp_send_head(sk))
1471                         break;
1472 
1473                 /* queue is in-order => we can short-circuit the walk early */
1474                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1475                         break;
1476 
1477                 if ((next_dup != NULL) &&
1478                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1479                         in_sack = tcp_match_skb_to_sack(sk, skb,
1480                                                         next_dup->start_seq,
1481                                                         next_dup->end_seq);
1482                         if (in_sack > 0)
1483                                 dup_sack = true;
1484                 }
1485 
1486                 /* skb reference here is a bit tricky to get right, since
1487                  * shifting can eat and free both this skb and the next,
1488                  * so not even _safe variant of the loop is enough.
1489                  */
1490                 if (in_sack <= 0) {
1491                         tmp = tcp_shift_skb_data(sk, skb, state,
1492                                                  start_seq, end_seq, dup_sack);
1493                         if (tmp != NULL) {
1494                                 if (tmp != skb) {
1495                                         skb = tmp;
1496                                         continue;
1497                                 }
1498 
1499                                 in_sack = 0;
1500                         } else {
1501                                 in_sack = tcp_match_skb_to_sack(sk, skb,
1502                                                                 start_seq,
1503                                                                 end_seq);
1504                         }
1505                 }
1506 
1507                 if (unlikely(in_sack < 0))
1508                         break;
1509 
1510                 if (in_sack) {
1511                         TCP_SKB_CB(skb)->sacked =
1512                                 tcp_sacktag_one(sk,
1513                                                 state,
1514                                                 TCP_SKB_CB(skb)->sacked,
1515                                                 TCP_SKB_CB(skb)->seq,
1516                                                 TCP_SKB_CB(skb)->end_seq,
1517                                                 dup_sack,
1518                                                 tcp_skb_pcount(skb),
1519                                                 TCP_SKB_CB(skb)->when);
1520 
1521                         if (!before(TCP_SKB_CB(skb)->seq,
1522                                     tcp_highest_sack_seq(tp)))
1523                                 tcp_advance_highest_sack(sk, skb);
1524                 }
1525 
1526                 state->fack_count += tcp_skb_pcount(skb);
1527         }
1528         return skb;
1529 }
1530 
1531 /* Avoid all extra work that is being done by sacktag while walking in
1532  * a normal way
1533  */
1534 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1535                                         struct tcp_sacktag_state *state,
1536                                         u32 skip_to_seq)
1537 {
1538         tcp_for_write_queue_from(skb, sk) {
1539                 if (skb == tcp_send_head(sk))
1540                         break;
1541 
1542                 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1543                         break;
1544 
1545                 state->fack_count += tcp_skb_pcount(skb);
1546         }
1547         return skb;
1548 }
1549 
1550 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1551                                                 struct sock *sk,
1552                                                 struct tcp_sack_block *next_dup,
1553                                                 struct tcp_sacktag_state *state,
1554                                                 u32 skip_to_seq)
1555 {
1556         if (next_dup == NULL)
1557                 return skb;
1558 
1559         if (before(next_dup->start_seq, skip_to_seq)) {
1560                 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1561                 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1562                                        next_dup->start_seq, next_dup->end_seq,
1563                                        1);
1564         }
1565 
1566         return skb;
1567 }
1568 
1569 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1570 {
1571         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1572 }
1573 
1574 static int
1575 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1576                         u32 prior_snd_una, s32 *sack_rtt)
1577 {
1578         struct tcp_sock *tp = tcp_sk(sk);
1579         const unsigned char *ptr = (skb_transport_header(ack_skb) +
1580                                     TCP_SKB_CB(ack_skb)->sacked);
1581         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1582         struct tcp_sack_block sp[TCP_NUM_SACKS];
1583         struct tcp_sack_block *cache;
1584         struct tcp_sacktag_state state;
1585         struct sk_buff *skb;
1586         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1587         int used_sacks;
1588         bool found_dup_sack = false;
1589         int i, j;
1590         int first_sack_index;
1591 
1592         state.flag = 0;
1593         state.reord = tp->packets_out;
1594         state.rtt = -1;
1595 
1596         if (!tp->sacked_out) {
1597                 if (WARN_ON(tp->fackets_out))
1598                         tp->fackets_out = 0;
1599                 tcp_highest_sack_reset(sk);
1600         }
1601 
1602         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1603                                          num_sacks, prior_snd_una);
1604         if (found_dup_sack)
1605                 state.flag |= FLAG_DSACKING_ACK;
1606 
1607         /* Eliminate too old ACKs, but take into
1608          * account more or less fresh ones, they can
1609          * contain valid SACK info.
1610          */
1611         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1612                 return 0;
1613 
1614         if (!tp->packets_out)
1615                 goto out;
1616 
1617         used_sacks = 0;
1618         first_sack_index = 0;
1619         for (i = 0; i < num_sacks; i++) {
1620                 bool dup_sack = !i && found_dup_sack;
1621 
1622                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1623                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1624 
1625                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1626                                             sp[used_sacks].start_seq,
1627                                             sp[used_sacks].end_seq)) {
1628                         int mib_idx;
1629 
1630                         if (dup_sack) {
1631                                 if (!tp->undo_marker)
1632                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1633                                 else
1634                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1635                         } else {
1636                                 /* Don't count olds caused by ACK reordering */
1637                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1638                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1639                                         continue;
1640                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1641                         }
1642 
1643                         NET_INC_STATS_BH(sock_net(sk), mib_idx);
1644                         if (i == 0)
1645                                 first_sack_index = -1;
1646                         continue;
1647                 }
1648 
1649                 /* Ignore very old stuff early */
1650                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1651                         continue;
1652 
1653                 used_sacks++;
1654         }
1655 
1656         /* order SACK blocks to allow in order walk of the retrans queue */
1657         for (i = used_sacks - 1; i > 0; i--) {
1658                 for (j = 0; j < i; j++) {
1659                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1660                                 swap(sp[j], sp[j + 1]);
1661 
1662                                 /* Track where the first SACK block goes to */
1663                                 if (j == first_sack_index)
1664                                         first_sack_index = j + 1;
1665                         }
1666                 }
1667         }
1668 
1669         skb = tcp_write_queue_head(sk);
1670         state.fack_count = 0;
1671         i = 0;
1672 
1673         if (!tp->sacked_out) {
1674                 /* It's already past, so skip checking against it */
1675                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1676         } else {
1677                 cache = tp->recv_sack_cache;
1678                 /* Skip empty blocks in at head of the cache */
1679                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1680                        !cache->end_seq)
1681                         cache++;
1682         }
1683 
1684         while (i < used_sacks) {
1685                 u32 start_seq = sp[i].start_seq;
1686                 u32 end_seq = sp[i].end_seq;
1687                 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1688                 struct tcp_sack_block *next_dup = NULL;
1689 
1690                 if (found_dup_sack && ((i + 1) == first_sack_index))
1691                         next_dup = &sp[i + 1];
1692 
1693                 /* Skip too early cached blocks */
1694                 while (tcp_sack_cache_ok(tp, cache) &&
1695                        !before(start_seq, cache->end_seq))
1696                         cache++;
1697 
1698                 /* Can skip some work by looking recv_sack_cache? */
1699                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1700                     after(end_seq, cache->start_seq)) {
1701 
1702                         /* Head todo? */
1703                         if (before(start_seq, cache->start_seq)) {
1704                                 skb = tcp_sacktag_skip(skb, sk, &state,
1705                                                        start_seq);
1706                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1707                                                        &state,
1708                                                        start_seq,
1709                                                        cache->start_seq,
1710                                                        dup_sack);
1711                         }
1712 
1713                         /* Rest of the block already fully processed? */
1714                         if (!after(end_seq, cache->end_seq))
1715                                 goto advance_sp;
1716 
1717                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1718                                                        &state,
1719                                                        cache->end_seq);
1720 
1721                         /* ...tail remains todo... */
1722                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1723                                 /* ...but better entrypoint exists! */
1724                                 skb = tcp_highest_sack(sk);
1725                                 if (skb == NULL)
1726                                         break;
1727                                 state.fack_count = tp->fackets_out;
1728                                 cache++;
1729                                 goto walk;
1730                         }
1731 
1732                         skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1733                         /* Check overlap against next cached too (past this one already) */
1734                         cache++;
1735                         continue;
1736                 }
1737 
1738                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1739                         skb = tcp_highest_sack(sk);
1740                         if (skb == NULL)
1741                                 break;
1742                         state.fack_count = tp->fackets_out;
1743                 }
1744                 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1745 
1746 walk:
1747                 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1748                                        start_seq, end_seq, dup_sack);
1749 
1750 advance_sp:
1751                 i++;
1752         }
1753 
1754         /* Clear the head of the cache sack blocks so we can skip it next time */
1755         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1756                 tp->recv_sack_cache[i].start_seq = 0;
1757                 tp->recv_sack_cache[i].end_seq = 0;
1758         }
1759         for (j = 0; j < used_sacks; j++)
1760                 tp->recv_sack_cache[i++] = sp[j];
1761 
1762         tcp_mark_lost_retrans(sk);
1763 
1764         tcp_verify_left_out(tp);
1765 
1766         if ((state.reord < tp->fackets_out) &&
1767             ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1768                 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1769 
1770 out:
1771 
1772 #if FASTRETRANS_DEBUG > 0
1773         WARN_ON((int)tp->sacked_out < 0);
1774         WARN_ON((int)tp->lost_out < 0);
1775         WARN_ON((int)tp->retrans_out < 0);
1776         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1777 #endif
1778         *sack_rtt = state.rtt;
1779         return state.flag;
1780 }
1781 
1782 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1783  * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1784  */
1785 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1786 {
1787         u32 holes;
1788 
1789         holes = max(tp->lost_out, 1U);
1790         holes = min(holes, tp->packets_out);
1791 
1792         if ((tp->sacked_out + holes) > tp->packets_out) {
1793                 tp->sacked_out = tp->packets_out - holes;
1794                 return true;
1795         }
1796         return false;
1797 }
1798 
1799 /* If we receive more dupacks than we expected counting segments
1800  * in assumption of absent reordering, interpret this as reordering.
1801  * The only another reason could be bug in receiver TCP.
1802  */
1803 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1804 {
1805         struct tcp_sock *tp = tcp_sk(sk);
1806         if (tcp_limit_reno_sacked(tp))
1807                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1808 }
1809 
1810 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1811 
1812 static void tcp_add_reno_sack(struct sock *sk)
1813 {
1814         struct tcp_sock *tp = tcp_sk(sk);
1815         tp->sacked_out++;
1816         tcp_check_reno_reordering(sk, 0);
1817         tcp_verify_left_out(tp);
1818 }
1819 
1820 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1821 
1822 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1823 {
1824         struct tcp_sock *tp = tcp_sk(sk);
1825 
1826         if (acked > 0) {
1827                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1828                 if (acked - 1 >= tp->sacked_out)
1829                         tp->sacked_out = 0;
1830                 else
1831                         tp->sacked_out -= acked - 1;
1832         }
1833         tcp_check_reno_reordering(sk, acked);
1834         tcp_verify_left_out(tp);
1835 }
1836 
1837 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1838 {
1839         tp->sacked_out = 0;
1840 }
1841 
1842 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1843 {
1844         tp->retrans_out = 0;
1845         tp->lost_out = 0;
1846 
1847         tp->undo_marker = 0;
1848         tp->undo_retrans = -1;
1849 }
1850 
1851 void tcp_clear_retrans(struct tcp_sock *tp)
1852 {
1853         tcp_clear_retrans_partial(tp);
1854 
1855         tp->fackets_out = 0;
1856         tp->sacked_out = 0;
1857 }
1858 
1859 /* Enter Loss state. If "how" is not zero, forget all SACK information
1860  * and reset tags completely, otherwise preserve SACKs. If receiver
1861  * dropped its ofo queue, we will know this due to reneging detection.
1862  */
1863 void tcp_enter_loss(struct sock *sk, int how)
1864 {
1865         const struct inet_connection_sock *icsk = inet_csk(sk);
1866         struct tcp_sock *tp = tcp_sk(sk);
1867         struct sk_buff *skb;
1868         bool new_recovery = false;
1869 
1870         /* Reduce ssthresh if it has not yet been made inside this window. */
1871         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1872             !after(tp->high_seq, tp->snd_una) ||
1873             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1874                 new_recovery = true;
1875                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1876                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1877                 tcp_ca_event(sk, CA_EVENT_LOSS);
1878         }
1879         tp->snd_cwnd       = 1;
1880         tp->snd_cwnd_cnt   = 0;
1881         tp->snd_cwnd_stamp = tcp_time_stamp;
1882 
1883         tcp_clear_retrans_partial(tp);
1884 
1885         if (tcp_is_reno(tp))
1886                 tcp_reset_reno_sack(tp);
1887 
1888         tp->undo_marker = tp->snd_una;
1889         if (how) {
1890                 tp->sacked_out = 0;
1891                 tp->fackets_out = 0;
1892         }
1893         tcp_clear_all_retrans_hints(tp);
1894 
1895         tcp_for_write_queue(skb, sk) {
1896                 if (skb == tcp_send_head(sk))
1897                         break;
1898 
1899                 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1900                         tp->undo_marker = 0;
1901                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1902                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1903                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1904                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1905                         tp->lost_out += tcp_skb_pcount(skb);
1906                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1907                 }
1908         }
1909         tcp_verify_left_out(tp);
1910 
1911         /* Timeout in disordered state after receiving substantial DUPACKs
1912          * suggests that the degree of reordering is over-estimated.
1913          */
1914         if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1915             tp->sacked_out >= sysctl_tcp_reordering)
1916                 tp->reordering = min_t(unsigned int, tp->reordering,
1917                                        sysctl_tcp_reordering);
1918         tcp_set_ca_state(sk, TCP_CA_Loss);
1919         tp->high_seq = tp->snd_nxt;
1920         TCP_ECN_queue_cwr(tp);
1921 
1922         /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1923          * loss recovery is underway except recurring timeout(s) on
1924          * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1925          */
1926         tp->frto = sysctl_tcp_frto &&
1927                    (new_recovery || icsk->icsk_retransmits) &&
1928                    !inet_csk(sk)->icsk_mtup.probe_size;
1929 }
1930 
1931 /* If ACK arrived pointing to a remembered SACK, it means that our
1932  * remembered SACKs do not reflect real state of receiver i.e.
1933  * receiver _host_ is heavily congested (or buggy).
1934  *
1935  * Do processing similar to RTO timeout.
1936  */
1937 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1938 {
1939         if (flag & FLAG_SACK_RENEGING) {
1940                 struct inet_connection_sock *icsk = inet_csk(sk);
1941                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1942 
1943                 tcp_enter_loss(sk, 1);
1944                 icsk->icsk_retransmits++;
1945                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1946                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1947                                           icsk->icsk_rto, TCP_RTO_MAX);
1948                 return true;
1949         }
1950         return false;
1951 }
1952 
1953 static inline int tcp_fackets_out(const struct tcp_sock *tp)
1954 {
1955         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
1956 }
1957 
1958 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1959  * counter when SACK is enabled (without SACK, sacked_out is used for
1960  * that purpose).
1961  *
1962  * Instead, with FACK TCP uses fackets_out that includes both SACKed
1963  * segments up to the highest received SACK block so far and holes in
1964  * between them.
1965  *
1966  * With reordering, holes may still be in flight, so RFC3517 recovery
1967  * uses pure sacked_out (total number of SACKed segments) even though
1968  * it violates the RFC that uses duplicate ACKs, often these are equal
1969  * but when e.g. out-of-window ACKs or packet duplication occurs,
1970  * they differ. Since neither occurs due to loss, TCP should really
1971  * ignore them.
1972  */
1973 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
1974 {
1975         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
1976 }
1977 
1978 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
1979 {
1980         struct tcp_sock *tp = tcp_sk(sk);
1981         unsigned long delay;
1982 
1983         /* Delay early retransmit and entering fast recovery for
1984          * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
1985          * available, or RTO is scheduled to fire first.
1986          */
1987         if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
1988             (flag & FLAG_ECE) || !tp->srtt)
1989                 return false;
1990 
1991         delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2));
1992         if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
1993                 return false;
1994 
1995         inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
1996                                   TCP_RTO_MAX);
1997         return true;
1998 }
1999 
2000 /* Linux NewReno/SACK/FACK/ECN state machine.
2001  * --------------------------------------
2002  *
2003  * "Open"       Normal state, no dubious events, fast path.
2004  * "Disorder"   In all the respects it is "Open",
2005  *              but requires a bit more attention. It is entered when
2006  *              we see some SACKs or dupacks. It is split of "Open"
2007  *              mainly to move some processing from fast path to slow one.
2008  * "CWR"        CWND was reduced due to some Congestion Notification event.
2009  *              It can be ECN, ICMP source quench, local device congestion.
2010  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2011  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2012  *
2013  * tcp_fastretrans_alert() is entered:
2014  * - each incoming ACK, if state is not "Open"
2015  * - when arrived ACK is unusual, namely:
2016  *      * SACK
2017  *      * Duplicate ACK.
2018  *      * ECN ECE.
2019  *
2020  * Counting packets in flight is pretty simple.
2021  *
2022  *      in_flight = packets_out - left_out + retrans_out
2023  *
2024  *      packets_out is SND.NXT-SND.UNA counted in packets.
2025  *
2026  *      retrans_out is number of retransmitted segments.
2027  *
2028  *      left_out is number of segments left network, but not ACKed yet.
2029  *
2030  *              left_out = sacked_out + lost_out
2031  *
2032  *     sacked_out: Packets, which arrived to receiver out of order
2033  *                 and hence not ACKed. With SACKs this number is simply
2034  *                 amount of SACKed data. Even without SACKs
2035  *                 it is easy to give pretty reliable estimate of this number,
2036  *                 counting duplicate ACKs.
2037  *
2038  *       lost_out: Packets lost by network. TCP has no explicit
2039  *                 "loss notification" feedback from network (for now).
2040  *                 It means that this number can be only _guessed_.
2041  *                 Actually, it is the heuristics to predict lossage that
2042  *                 distinguishes different algorithms.
2043  *
2044  *      F.e. after RTO, when all the queue is considered as lost,
2045  *      lost_out = packets_out and in_flight = retrans_out.
2046  *
2047  *              Essentially, we have now two algorithms counting
2048  *              lost packets.
2049  *
2050  *              FACK: It is the simplest heuristics. As soon as we decided
2051  *              that something is lost, we decide that _all_ not SACKed
2052  *              packets until the most forward SACK are lost. I.e.
2053  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2054  *              It is absolutely correct estimate, if network does not reorder
2055  *              packets. And it loses any connection to reality when reordering
2056  *              takes place. We use FACK by default until reordering
2057  *              is suspected on the path to this destination.
2058  *
2059  *              NewReno: when Recovery is entered, we assume that one segment
2060  *              is lost (classic Reno). While we are in Recovery and
2061  *              a partial ACK arrives, we assume that one more packet
2062  *              is lost (NewReno). This heuristics are the same in NewReno
2063  *              and SACK.
2064  *
2065  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2066  *  deflation etc. CWND is real congestion window, never inflated, changes
2067  *  only according to classic VJ rules.
2068  *
2069  * Really tricky (and requiring careful tuning) part of algorithm
2070  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2071  * The first determines the moment _when_ we should reduce CWND and,
2072  * hence, slow down forward transmission. In fact, it determines the moment
2073  * when we decide that hole is caused by loss, rather than by a reorder.
2074  *
2075  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2076  * holes, caused by lost packets.
2077  *
2078  * And the most logically complicated part of algorithm is undo
2079  * heuristics. We detect false retransmits due to both too early
2080  * fast retransmit (reordering) and underestimated RTO, analyzing
2081  * timestamps and D-SACKs. When we detect that some segments were
2082  * retransmitted by mistake and CWND reduction was wrong, we undo
2083  * window reduction and abort recovery phase. This logic is hidden
2084  * inside several functions named tcp_try_undo_<something>.
2085  */
2086 
2087 /* This function decides, when we should leave Disordered state
2088  * and enter Recovery phase, reducing congestion window.
2089  *
2090  * Main question: may we further continue forward transmission
2091  * with the same cwnd?
2092  */
2093 static bool tcp_time_to_recover(struct sock *sk, int flag)
2094 {
2095         struct tcp_sock *tp = tcp_sk(sk);
2096         __u32 packets_out;
2097 
2098         /* Trick#1: The loss is proven. */
2099         if (tp->lost_out)
2100                 return true;
2101 
2102         /* Not-A-Trick#2 : Classic rule... */
2103         if (tcp_dupack_heuristics(tp) > tp->reordering)
2104                 return true;
2105 
2106         /* Trick#4: It is still not OK... But will it be useful to delay
2107          * recovery more?
2108          */
2109         packets_out = tp->packets_out;
2110         if (packets_out <= tp->reordering &&
2111             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2112             !tcp_may_send_now(sk)) {
2113                 /* We have nothing to send. This connection is limited
2114                  * either by receiver window or by application.
2115                  */
2116                 return true;
2117         }
2118 
2119         /* If a thin stream is detected, retransmit after first
2120          * received dupack. Employ only if SACK is supported in order
2121          * to avoid possible corner-case series of spurious retransmissions
2122          * Use only if there are no unsent data.
2123          */
2124         if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2125             tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2126             tcp_is_sack(tp) && !tcp_send_head(sk))
2127                 return true;
2128 
2129         /* Trick#6: TCP early retransmit, per RFC5827.  To avoid spurious
2130          * retransmissions due to small network reorderings, we implement
2131          * Mitigation A.3 in the RFC and delay the retransmission for a short
2132          * interval if appropriate.
2133          */
2134         if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2135             (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2136             !tcp_may_send_now(sk))
2137                 return !tcp_pause_early_retransmit(sk, flag);
2138 
2139         return false;
2140 }
2141 
2142 /* Detect loss in event "A" above by marking head of queue up as lost.
2143  * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2144  * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2145  * has at least tp->reordering SACKed seqments above it; "packets" refers to
2146  * the maximum SACKed segments to pass before reaching this limit.
2147  */
2148 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2149 {
2150         struct tcp_sock *tp = tcp_sk(sk);
2151         struct sk_buff *skb;
2152         int cnt, oldcnt;
2153         int err;
2154         unsigned int mss;
2155         /* Use SACK to deduce losses of new sequences sent during recovery */
2156         const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2157 
2158         WARN_ON(packets > tp->packets_out);
2159         if (tp->lost_skb_hint) {
2160                 skb = tp->lost_skb_hint;
2161                 cnt = tp->lost_cnt_hint;
2162                 /* Head already handled? */
2163                 if (mark_head && skb != tcp_write_queue_head(sk))
2164                         return;
2165         } else {
2166                 skb = tcp_write_queue_head(sk);
2167                 cnt = 0;
2168         }
2169 
2170         tcp_for_write_queue_from(skb, sk) {
2171                 if (skb == tcp_send_head(sk))
2172                         break;
2173                 /* TODO: do this better */
2174                 /* this is not the most efficient way to do this... */
2175                 tp->lost_skb_hint = skb;
2176                 tp->lost_cnt_hint = cnt;
2177 
2178                 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2179                         break;
2180 
2181                 oldcnt = cnt;
2182                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2183                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2184                         cnt += tcp_skb_pcount(skb);
2185 
2186                 if (cnt > packets) {
2187                         if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2188                             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2189                             (oldcnt >= packets))
2190                                 break;
2191 
2192                         mss = skb_shinfo(skb)->gso_size;
2193                         err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2194                         if (err < 0)
2195                                 break;
2196                         cnt = packets;
2197                 }
2198 
2199                 tcp_skb_mark_lost(tp, skb);
2200 
2201                 if (mark_head)
2202                         break;
2203         }
2204         tcp_verify_left_out(tp);
2205 }
2206 
2207 /* Account newly detected lost packet(s) */
2208 
2209 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2210 {
2211         struct tcp_sock *tp = tcp_sk(sk);
2212 
2213         if (tcp_is_reno(tp)) {
2214                 tcp_mark_head_lost(sk, 1, 1);
2215         } else if (tcp_is_fack(tp)) {
2216                 int lost = tp->fackets_out - tp->reordering;
2217                 if (lost <= 0)
2218                         lost = 1;
2219                 tcp_mark_head_lost(sk, lost, 0);
2220         } else {
2221                 int sacked_upto = tp->sacked_out - tp->reordering;
2222                 if (sacked_upto >= 0)
2223                         tcp_mark_head_lost(sk, sacked_upto, 0);
2224                 else if (fast_rexmit)
2225                         tcp_mark_head_lost(sk, 1, 1);
2226         }
2227 }
2228 
2229 /* CWND moderation, preventing bursts due to too big ACKs
2230  * in dubious situations.
2231  */
2232 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2233 {
2234         tp->snd_cwnd = min(tp->snd_cwnd,
2235                            tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2236         tp->snd_cwnd_stamp = tcp_time_stamp;
2237 }
2238 
2239 /* Nothing was retransmitted or returned timestamp is less
2240  * than timestamp of the first retransmission.
2241  */
2242 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2243 {
2244         return !tp->retrans_stamp ||
2245                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2246                  before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2247 }
2248 
2249 /* Undo procedures. */
2250 
2251 #if FASTRETRANS_DEBUG > 1
2252 static void DBGUNDO(struct sock *sk, const char *msg)
2253 {
2254         struct tcp_sock *tp = tcp_sk(sk);
2255         struct inet_sock *inet = inet_sk(sk);
2256 
2257         if (sk->sk_family == AF_INET) {
2258                 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2259                          msg,
2260                          &inet->inet_daddr, ntohs(inet->inet_dport),
2261                          tp->snd_cwnd, tcp_left_out(tp),
2262                          tp->snd_ssthresh, tp->prior_ssthresh,
2263                          tp->packets_out);
2264         }
2265 #if IS_ENABLED(CONFIG_IPV6)
2266         else if (sk->sk_family == AF_INET6) {
2267                 struct ipv6_pinfo *np = inet6_sk(sk);
2268                 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2269                          msg,
2270                          &np->daddr, ntohs(inet->inet_dport),
2271                          tp->snd_cwnd, tcp_left_out(tp),
2272                          tp->snd_ssthresh, tp->prior_ssthresh,
2273                          tp->packets_out);
2274         }
2275 #endif
2276 }
2277 #else
2278 #define DBGUNDO(x...) do { } while (0)
2279 #endif
2280 
2281 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2282 {
2283         struct tcp_sock *tp = tcp_sk(sk);
2284 
2285         if (unmark_loss) {
2286                 struct sk_buff *skb;
2287 
2288                 tcp_for_write_queue(skb, sk) {
2289                         if (skb == tcp_send_head(sk))
2290                                 break;
2291                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2292                 }
2293                 tp->lost_out = 0;
2294                 tcp_clear_all_retrans_hints(tp);
2295         }
2296 
2297         if (tp->prior_ssthresh) {
2298                 const struct inet_connection_sock *icsk = inet_csk(sk);
2299 
2300                 if (icsk->icsk_ca_ops->undo_cwnd)
2301                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2302                 else
2303                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2304 
2305                 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2306                         tp->snd_ssthresh = tp->prior_ssthresh;
2307                         TCP_ECN_withdraw_cwr(tp);
2308                 }
2309         } else {
2310                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2311         }
2312         tp->snd_cwnd_stamp = tcp_time_stamp;
2313         tp->undo_marker = 0;
2314 }
2315 
2316 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2317 {
2318         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2319 }
2320 
2321 /* People celebrate: "We love our President!" */
2322 static bool tcp_try_undo_recovery(struct sock *sk)
2323 {
2324         struct tcp_sock *tp = tcp_sk(sk);
2325 
2326         if (tcp_may_undo(tp)) {
2327                 int mib_idx;
2328 
2329                 /* Happy end! We did not retransmit anything
2330                  * or our original transmission succeeded.
2331                  */
2332                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2333                 tcp_undo_cwnd_reduction(sk, false);
2334                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2335                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2336                 else
2337                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2338 
2339                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2340         }
2341         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2342                 /* Hold old state until something *above* high_seq
2343                  * is ACKed. For Reno it is MUST to prevent false
2344                  * fast retransmits (RFC2582). SACK TCP is safe. */
2345                 tcp_moderate_cwnd(tp);
2346                 return true;
2347         }
2348         tcp_set_ca_state(sk, TCP_CA_Open);
2349         return false;
2350 }
2351 
2352 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2353 static bool tcp_try_undo_dsack(struct sock *sk)
2354 {
2355         struct tcp_sock *tp = tcp_sk(sk);
2356 
2357         if (tp->undo_marker && !tp->undo_retrans) {
2358                 DBGUNDO(sk, "D-SACK");
2359                 tcp_undo_cwnd_reduction(sk, false);
2360                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2361                 return true;
2362         }
2363         return false;
2364 }
2365 
2366 /* We can clear retrans_stamp when there are no retransmissions in the
2367  * window. It would seem that it is trivially available for us in
2368  * tp->retrans_out, however, that kind of assumptions doesn't consider
2369  * what will happen if errors occur when sending retransmission for the
2370  * second time. ...It could the that such segment has only
2371  * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2372  * the head skb is enough except for some reneging corner cases that
2373  * are not worth the effort.
2374  *
2375  * Main reason for all this complexity is the fact that connection dying
2376  * time now depends on the validity of the retrans_stamp, in particular,
2377  * that successive retransmissions of a segment must not advance
2378  * retrans_stamp under any conditions.
2379  */
2380 static bool tcp_any_retrans_done(const struct sock *sk)
2381 {
2382         const struct tcp_sock *tp = tcp_sk(sk);
2383         struct sk_buff *skb;
2384 
2385         if (tp->retrans_out)
2386                 return true;
2387 
2388         skb = tcp_write_queue_head(sk);
2389         if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2390                 return true;
2391 
2392         return false;
2393 }
2394 
2395 /* Undo during loss recovery after partial ACK or using F-RTO. */
2396 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2397 {
2398         struct tcp_sock *tp = tcp_sk(sk);
2399 
2400         if (frto_undo || tcp_may_undo(tp)) {
2401                 tcp_undo_cwnd_reduction(sk, true);
2402 
2403                 DBGUNDO(sk, "partial loss");
2404                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2405                 if (frto_undo)
2406                         NET_INC_STATS_BH(sock_net(sk),
2407                                          LINUX_MIB_TCPSPURIOUSRTOS);
2408                 inet_csk(sk)->icsk_retransmits = 0;
2409                 if (frto_undo || tcp_is_sack(tp))
2410                         tcp_set_ca_state(sk, TCP_CA_Open);
2411                 return true;
2412         }
2413         return false;
2414 }
2415 
2416 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2417  * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2418  * It computes the number of packets to send (sndcnt) based on packets newly
2419  * delivered:
2420  *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2421  *      cwnd reductions across a full RTT.
2422  *   2) If packets in flight is lower than ssthresh (such as due to excess
2423  *      losses and/or application stalls), do not perform any further cwnd
2424  *      reductions, but instead slow start up to ssthresh.
2425  */
2426 static void tcp_init_cwnd_reduction(struct sock *sk, const bool set_ssthresh)
2427 {
2428         struct tcp_sock *tp = tcp_sk(sk);
2429 
2430         tp->high_seq = tp->snd_nxt;
2431         tp->tlp_high_seq = 0;
2432         tp->snd_cwnd_cnt = 0;
2433         tp->prior_cwnd = tp->snd_cwnd;
2434         tp->prr_delivered = 0;
2435         tp->prr_out = 0;
2436         if (set_ssthresh)
2437                 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2438         TCP_ECN_queue_cwr(tp);
2439 }
2440 
2441 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2442                                int fast_rexmit)
2443 {
2444         struct tcp_sock *tp = tcp_sk(sk);
2445         int sndcnt = 0;
2446         int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2447         int newly_acked_sacked = prior_unsacked -
2448                                  (tp->packets_out - tp->sacked_out);
2449 
2450         tp->prr_delivered += newly_acked_sacked;
2451         if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2452                 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2453                                tp->prior_cwnd - 1;
2454                 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2455         } else {
2456                 sndcnt = min_t(int, delta,
2457                                max_t(int, tp->prr_delivered - tp->prr_out,
2458                                      newly_acked_sacked) + 1);
2459         }
2460 
2461         sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2462         tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2463 }
2464 
2465 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2466 {
2467         struct tcp_sock *tp = tcp_sk(sk);
2468 
2469         /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2470         if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2471             (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2472                 tp->snd_cwnd = tp->snd_ssthresh;
2473                 tp->snd_cwnd_stamp = tcp_time_stamp;
2474         }
2475         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2476 }
2477 
2478 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2479 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
2480 {
2481         struct tcp_sock *tp = tcp_sk(sk);
2482 
2483         tp->prior_ssthresh = 0;
2484         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2485                 tp->undo_marker = 0;
2486                 tcp_init_cwnd_reduction(sk, set_ssthresh);
2487                 tcp_set_ca_state(sk, TCP_CA_CWR);
2488         }
2489 }
2490 
2491 static void tcp_try_keep_open(struct sock *sk)
2492 {
2493         struct tcp_sock *tp = tcp_sk(sk);
2494         int state = TCP_CA_Open;
2495 
2496         if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2497                 state = TCP_CA_Disorder;
2498 
2499         if (inet_csk(sk)->icsk_ca_state != state) {
2500                 tcp_set_ca_state(sk, state);
2501                 tp->high_seq = tp->snd_nxt;
2502         }
2503 }
2504 
2505 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2506 {
2507         struct tcp_sock *tp = tcp_sk(sk);
2508 
2509         tcp_verify_left_out(tp);
2510 
2511         if (!tcp_any_retrans_done(sk))
2512                 tp->retrans_stamp = 0;
2513 
2514         if (flag & FLAG_ECE)
2515                 tcp_enter_cwr(sk, 1);
2516 
2517         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2518                 tcp_try_keep_open(sk);
2519         } else {
2520                 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2521         }
2522 }
2523 
2524 static void tcp_mtup_probe_failed(struct sock *sk)
2525 {
2526         struct inet_connection_sock *icsk = inet_csk(sk);
2527 
2528         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2529         icsk->icsk_mtup.probe_size = 0;
2530 }
2531 
2532 static void tcp_mtup_probe_success(struct sock *sk)
2533 {
2534         struct tcp_sock *tp = tcp_sk(sk);
2535         struct inet_connection_sock *icsk = inet_csk(sk);
2536 
2537         /* FIXME: breaks with very large cwnd */
2538         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2539         tp->snd_cwnd = tp->snd_cwnd *
2540                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2541                        icsk->icsk_mtup.probe_size;
2542         tp->snd_cwnd_cnt = 0;
2543         tp->snd_cwnd_stamp = tcp_time_stamp;
2544         tp->snd_ssthresh = tcp_current_ssthresh(sk);
2545 
2546         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2547         icsk->icsk_mtup.probe_size = 0;
2548         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2549 }
2550 
2551 /* Do a simple retransmit without using the backoff mechanisms in
2552  * tcp_timer. This is used for path mtu discovery.
2553  * The socket is already locked here.
2554  */
2555 void tcp_simple_retransmit(struct sock *sk)
2556 {
2557         const struct inet_connection_sock *icsk = inet_csk(sk);
2558         struct tcp_sock *tp = tcp_sk(sk);
2559         struct sk_buff *skb;
2560         unsigned int mss = tcp_current_mss(sk);
2561         u32 prior_lost = tp->lost_out;
2562 
2563         tcp_for_write_queue(skb, sk) {
2564                 if (skb == tcp_send_head(sk))
2565                         break;
2566                 if (tcp_skb_seglen(skb) > mss &&
2567                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2568                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2569                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2570                                 tp->retrans_out -= tcp_skb_pcount(skb);
2571                         }
2572                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2573                 }
2574         }
2575 
2576         tcp_clear_retrans_hints_partial(tp);
2577 
2578         if (prior_lost == tp->lost_out)
2579                 return;
2580 
2581         if (tcp_is_reno(tp))
2582                 tcp_limit_reno_sacked(tp);
2583 
2584         tcp_verify_left_out(tp);
2585 
2586         /* Don't muck with the congestion window here.
2587          * Reason is that we do not increase amount of _data_
2588          * in network, but units changed and effective
2589          * cwnd/ssthresh really reduced now.
2590          */
2591         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2592                 tp->high_seq = tp->snd_nxt;
2593                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2594                 tp->prior_ssthresh = 0;
2595                 tp->undo_marker = 0;
2596                 tcp_set_ca_state(sk, TCP_CA_Loss);
2597         }
2598         tcp_xmit_retransmit_queue(sk);
2599 }
2600 EXPORT_SYMBOL(tcp_simple_retransmit);
2601 
2602 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2603 {
2604         struct tcp_sock *tp = tcp_sk(sk);
2605         int mib_idx;
2606 
2607         if (tcp_is_reno(tp))
2608                 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2609         else
2610                 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2611 
2612         NET_INC_STATS_BH(sock_net(sk), mib_idx);
2613 
2614         tp->prior_ssthresh = 0;
2615         tp->undo_marker = tp->snd_una;
2616         tp->undo_retrans = tp->retrans_out ? : -1;
2617 
2618         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2619                 if (!ece_ack)
2620                         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2621                 tcp_init_cwnd_reduction(sk, true);
2622         }
2623         tcp_set_ca_state(sk, TCP_CA_Recovery);
2624 }
2625 
2626 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2627  * recovered or spurious. Otherwise retransmits more on partial ACKs.
2628  */
2629 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2630 {
2631         struct tcp_sock *tp = tcp_sk(sk);
2632         bool recovered = !before(tp->snd_una, tp->high_seq);
2633 
2634         if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2635                 /* Step 3.b. A timeout is spurious if not all data are
2636                  * lost, i.e., never-retransmitted data are (s)acked.
2637                  */
2638                 if (tcp_try_undo_loss(sk, flag & FLAG_ORIG_SACK_ACKED))
2639                         return;
2640 
2641                 if (after(tp->snd_nxt, tp->high_seq) &&
2642                     (flag & FLAG_DATA_SACKED || is_dupack)) {
2643                         tp->frto = 0; /* Loss was real: 2nd part of step 3.a */
2644                 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2645                         tp->high_seq = tp->snd_nxt;
2646                         __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2647                                                   TCP_NAGLE_OFF);
2648                         if (after(tp->snd_nxt, tp->high_seq))
2649                                 return; /* Step 2.b */
2650                         tp->frto = 0;
2651                 }
2652         }
2653 
2654         if (recovered) {
2655                 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2656                 tcp_try_undo_recovery(sk);
2657                 return;
2658         }
2659         if (tcp_is_reno(tp)) {
2660                 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2661                  * delivered. Lower inflight to clock out (re)tranmissions.
2662                  */
2663                 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2664                         tcp_add_reno_sack(sk);
2665                 else if (flag & FLAG_SND_UNA_ADVANCED)
2666                         tcp_reset_reno_sack(tp);
2667         }
2668         if (tcp_try_undo_loss(sk, false))
2669                 return;
2670         tcp_xmit_retransmit_queue(sk);
2671 }
2672 
2673 /* Undo during fast recovery after partial ACK. */
2674 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2675                                  const int prior_unsacked)
2676 {
2677         struct tcp_sock *tp = tcp_sk(sk);
2678 
2679         if (tp->undo_marker && tcp_packet_delayed(tp)) {
2680                 /* Plain luck! Hole if filled with delayed
2681                  * packet, rather than with a retransmit.
2682                  */
2683                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2684 
2685                 /* We are getting evidence that the reordering degree is higher
2686                  * than we realized. If there are no retransmits out then we
2687                  * can undo. Otherwise we clock out new packets but do not
2688                  * mark more packets lost or retransmit more.
2689                  */
2690                 if (tp->retrans_out) {
2691                         tcp_cwnd_reduction(sk, prior_unsacked, 0);
2692                         return true;
2693                 }
2694 
2695                 if (!tcp_any_retrans_done(sk))
2696                         tp->retrans_stamp = 0;
2697 
2698                 DBGUNDO(sk, "partial recovery");
2699                 tcp_undo_cwnd_reduction(sk, true);
2700                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2701                 tcp_try_keep_open(sk);
2702                 return true;
2703         }
2704         return false;
2705 }
2706 
2707 /* Process an event, which can update packets-in-flight not trivially.
2708  * Main goal of this function is to calculate new estimate for left_out,
2709  * taking into account both packets sitting in receiver's buffer and
2710  * packets lost by network.
2711  *
2712  * Besides that it does CWND reduction, when packet loss is detected
2713  * and changes state of machine.
2714  *
2715  * It does _not_ decide what to send, it is made in function
2716  * tcp_xmit_retransmit_queue().
2717  */
2718 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2719                                   const int prior_unsacked,
2720                                   bool is_dupack, int flag)
2721 {
2722         struct inet_connection_sock *icsk = inet_csk(sk);
2723         struct tcp_sock *tp = tcp_sk(sk);
2724         bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2725                                     (tcp_fackets_out(tp) > tp->reordering));
2726         int fast_rexmit = 0;
2727 
2728         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2729                 tp->sacked_out = 0;
2730         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2731                 tp->fackets_out = 0;
2732 
2733         /* Now state machine starts.
2734          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2735         if (flag & FLAG_ECE)
2736                 tp->prior_ssthresh = 0;
2737 
2738         /* B. In all the states check for reneging SACKs. */
2739         if (tcp_check_sack_reneging(sk, flag))
2740                 return;
2741 
2742         /* C. Check consistency of the current state. */
2743         tcp_verify_left_out(tp);
2744 
2745         /* D. Check state exit conditions. State can be terminated
2746          *    when high_seq is ACKed. */
2747         if (icsk->icsk_ca_state == TCP_CA_Open) {
2748                 WARN_ON(tp->retrans_out != 0);
2749                 tp->retrans_stamp = 0;
2750         } else if (!before(tp->snd_una, tp->high_seq)) {
2751                 switch (icsk->icsk_ca_state) {
2752                 case TCP_CA_CWR:
2753                         /* CWR is to be held something *above* high_seq
2754                          * is ACKed for CWR bit to reach receiver. */
2755                         if (tp->snd_una != tp->high_seq) {
2756                                 tcp_end_cwnd_reduction(sk);
2757                                 tcp_set_ca_state(sk, TCP_CA_Open);
2758                         }
2759                         break;
2760 
2761                 case TCP_CA_Recovery:
2762                         if (tcp_is_reno(tp))
2763                                 tcp_reset_reno_sack(tp);
2764                         if (tcp_try_undo_recovery(sk))
2765                                 return;
2766                         tcp_end_cwnd_reduction(sk);
2767                         break;
2768                 }
2769         }
2770 
2771         /* E. Process state. */
2772         switch (icsk->icsk_ca_state) {
2773         case TCP_CA_Recovery:
2774                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2775                         if (tcp_is_reno(tp) && is_dupack)
2776                                 tcp_add_reno_sack(sk);
2777                 } else {
2778                         if (tcp_try_undo_partial(sk, acked, prior_unsacked))
2779                                 return;
2780                         /* Partial ACK arrived. Force fast retransmit. */
2781                         do_lost = tcp_is_reno(tp) ||
2782                                   tcp_fackets_out(tp) > tp->reordering;
2783                 }
2784                 if (tcp_try_undo_dsack(sk)) {
2785                         tcp_try_keep_open(sk);
2786                         return;
2787                 }
2788                 break;
2789         case TCP_CA_Loss:
2790                 tcp_process_loss(sk, flag, is_dupack);
2791                 if (icsk->icsk_ca_state != TCP_CA_Open)
2792                         return;
2793                 /* Fall through to processing in Open state. */
2794         default:
2795                 if (tcp_is_reno(tp)) {
2796                         if (flag & FLAG_SND_UNA_ADVANCED)
2797                                 tcp_reset_reno_sack(tp);
2798                         if (is_dupack)
2799                                 tcp_add_reno_sack(sk);
2800                 }
2801 
2802                 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2803                         tcp_try_undo_dsack(sk);
2804 
2805                 if (!tcp_time_to_recover(sk, flag)) {
2806                         tcp_try_to_open(sk, flag, prior_unsacked);
2807                         return;
2808                 }
2809 
2810                 /* MTU probe failure: don't reduce cwnd */
2811                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2812                     icsk->icsk_mtup.probe_size &&
2813                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2814                         tcp_mtup_probe_failed(sk);
2815                         /* Restores the reduction we did in tcp_mtup_probe() */
2816                         tp->snd_cwnd++;
2817                         tcp_simple_retransmit(sk);
2818                         return;
2819                 }
2820 
2821                 /* Otherwise enter Recovery state */
2822                 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2823                 fast_rexmit = 1;
2824         }
2825 
2826         if (do_lost)
2827                 tcp_update_scoreboard(sk, fast_rexmit);
2828         tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
2829         tcp_xmit_retransmit_queue(sk);
2830 }
2831 
2832 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2833                                       s32 seq_rtt, s32 sack_rtt)
2834 {
2835         const struct tcp_sock *tp = tcp_sk(sk);
2836 
2837         /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2838          * broken middle-boxes or peers may corrupt TS-ECR fields. But
2839          * Karn's algorithm forbids taking RTT if some retransmitted data
2840          * is acked (RFC6298).
2841          */
2842         if (flag & FLAG_RETRANS_DATA_ACKED)
2843                 seq_rtt = -1;
2844 
2845         if (seq_rtt < 0)
2846                 seq_rtt = sack_rtt;
2847 
2848         /* RTTM Rule: A TSecr value received in a segment is used to
2849          * update the averaged RTT measurement only if the segment
2850          * acknowledges some new data, i.e., only if it advances the
2851          * left edge of the send window.
2852          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2853          */
2854         if (seq_rtt < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2855             flag & FLAG_ACKED)
2856                 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2857 
2858         if (seq_rtt < 0)
2859                 return false;
2860 
2861         tcp_rtt_estimator(sk, seq_rtt);
2862         tcp_set_rto(sk);
2863 
2864         /* RFC6298: only reset backoff on valid RTT measurement. */
2865         inet_csk(sk)->icsk_backoff = 0;
2866         return true;
2867 }
2868 
2869 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2870 static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
2871 {
2872         struct tcp_sock *tp = tcp_sk(sk);
2873         s32 seq_rtt = -1;
2874 
2875         if (synack_stamp && !tp->total_retrans)
2876                 seq_rtt = tcp_time_stamp - synack_stamp;
2877 
2878         /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2879          * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2880          */
2881         if (!tp->srtt)
2882                 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt, -1);
2883 }
2884 
2885 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
2886 {
2887         const struct inet_connection_sock *icsk = inet_csk(sk);
2888         icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
2889         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2890 }
2891 
2892 /* Restart timer after forward progress on connection.
2893  * RFC2988 recommends to restart timer to now+rto.
2894  */
2895 void tcp_rearm_rto(struct sock *sk)
2896 {
2897         const struct inet_connection_sock *icsk = inet_csk(sk);
2898         struct tcp_sock *tp = tcp_sk(sk);
2899 
2900         /* If the retrans timer is currently being used by Fast Open
2901          * for SYN-ACK retrans purpose, stay put.
2902          */
2903         if (tp->fastopen_rsk)
2904                 return;
2905 
2906         if (!tp->packets_out) {
2907                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2908         } else {
2909                 u32 rto = inet_csk(sk)->icsk_rto;
2910                 /* Offset the time elapsed after installing regular RTO */
2911                 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2912                     icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2913                         struct sk_buff *skb = tcp_write_queue_head(sk);
2914                         const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
2915                         s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2916                         /* delta may not be positive if the socket is locked
2917                          * when the retrans timer fires and is rescheduled.
2918                          */
2919                         if (delta > 0)
2920                                 rto = delta;
2921                 }
2922                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2923                                           TCP_RTO_MAX);
2924         }
2925 }
2926 
2927 /* This function is called when the delayed ER timer fires. TCP enters
2928  * fast recovery and performs fast-retransmit.
2929  */
2930 void tcp_resume_early_retransmit(struct sock *sk)
2931 {
2932         struct tcp_sock *tp = tcp_sk(sk);
2933 
2934         tcp_rearm_rto(sk);
2935 
2936         /* Stop if ER is disabled after the delayed ER timer is scheduled */
2937         if (!tp->do_early_retrans)
2938                 return;
2939 
2940         tcp_enter_recovery(sk, false);
2941         tcp_update_scoreboard(sk, 1);
2942         tcp_xmit_retransmit_queue(sk);
2943 }
2944 
2945 /* If we get here, the whole TSO packet has not been acked. */
2946 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2947 {
2948         struct tcp_sock *tp = tcp_sk(sk);
2949         u32 packets_acked;
2950 
2951         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2952 
2953         packets_acked = tcp_skb_pcount(skb);
2954         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2955                 return 0;
2956         packets_acked -= tcp_skb_pcount(skb);
2957 
2958         if (packets_acked) {
2959                 BUG_ON(tcp_skb_pcount(skb) == 0);
2960                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2961         }
2962 
2963         return packets_acked;
2964 }
2965 
2966 /* Remove acknowledged frames from the retransmission queue. If our packet
2967  * is before the ack sequence we can discard it as it's confirmed to have
2968  * arrived at the other end.
2969  */
2970 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
2971                                u32 prior_snd_una, s32 sack_rtt)
2972 {
2973         struct tcp_sock *tp = tcp_sk(sk);
2974         const struct inet_connection_sock *icsk = inet_csk(sk);
2975         struct sk_buff *skb;
2976         u32 now = tcp_time_stamp;
2977         int fully_acked = true;
2978         int flag = 0;
2979         u32 pkts_acked = 0;
2980         u32 reord = tp->packets_out;
2981         u32 prior_sacked = tp->sacked_out;
2982         s32 seq_rtt = -1;
2983         s32 ca_seq_rtt = -1;
2984         ktime_t last_ackt = net_invalid_timestamp();
2985         bool rtt_update;
2986 
2987         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
2988                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2989                 u32 acked_pcount;
2990                 u8 sacked = scb->sacked;
2991 
2992                 /* Determine how many packets and what bytes were acked, tso and else */
2993                 if (after(scb->end_seq, tp->snd_una)) {
2994                         if (tcp_skb_pcount(skb) == 1 ||
2995                             !after(tp->snd_una, scb->seq))
2996                                 break;
2997 
2998                         acked_pcount = tcp_tso_acked(sk, skb);
2999                         if (!acked_pcount)
3000                                 break;
3001 
3002                         fully_acked = false;
3003                 } else {
3004                         acked_pcount = tcp_skb_pcount(skb);
3005                 }
3006 
3007                 if (sacked & TCPCB_RETRANS) {
3008                         if (sacked & TCPCB_SACKED_RETRANS)
3009                                 tp->retrans_out -= acked_pcount;
3010                         flag |= FLAG_RETRANS_DATA_ACKED;
3011                 } else {
3012                         ca_seq_rtt = now - scb->when;
3013                         last_ackt = skb->tstamp;
3014                         if (seq_rtt < 0) {
3015                                 seq_rtt = ca_seq_rtt;
3016                         }
3017                         if (!(sacked & TCPCB_SACKED_ACKED)) {
3018                                 reord = min(pkts_acked, reord);
3019                                 if (!after(scb->end_seq, tp->high_seq))
3020                                         flag |= FLAG_ORIG_SACK_ACKED;
3021                         }
3022                 }
3023 
3024                 if (sacked & TCPCB_SACKED_ACKED)
3025                         tp->sacked_out -= acked_pcount;
3026                 if (sacked & TCPCB_LOST)
3027                         tp->lost_out -= acked_pcount;
3028 
3029                 tp->packets_out -= acked_pcount;
3030                 pkts_acked += acked_pcount;
3031 
3032                 /* Initial outgoing SYN's get put onto the write_queue
3033                  * just like anything else we transmit.  It is not
3034                  * true data, and if we misinform our callers that
3035                  * this ACK acks real data, we will erroneously exit
3036                  * connection startup slow start one packet too
3037                  * quickly.  This is severely frowned upon behavior.
3038                  */
3039                 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3040                         flag |= FLAG_DATA_ACKED;
3041                 } else {
3042                         flag |= FLAG_SYN_ACKED;
3043                         tp->retrans_stamp = 0;
3044                 }
3045 
3046                 if (!fully_acked)
3047                         break;
3048 
3049                 tcp_unlink_write_queue(skb, sk);
3050                 sk_wmem_free_skb(sk, skb);
3051                 if (skb == tp->retransmit_skb_hint)
3052                         tp->retransmit_skb_hint = NULL;
3053                 if (skb == tp->lost_skb_hint)
3054                         tp->lost_skb_hint = NULL;
3055         }
3056 
3057         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3058                 tp->snd_up = tp->snd_una;
3059 
3060         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3061                 flag |= FLAG_SACK_RENEGING;
3062 
3063         rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt, sack_rtt);
3064 
3065         if (flag & FLAG_ACKED) {
3066                 const struct tcp_congestion_ops *ca_ops
3067                         = inet_csk(sk)->icsk_ca_ops;
3068 
3069                 tcp_rearm_rto(sk);
3070                 if (unlikely(icsk->icsk_mtup.probe_size &&
3071                              !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3072                         tcp_mtup_probe_success(sk);
3073                 }
3074 
3075                 if (tcp_is_reno(tp)) {
3076                         tcp_remove_reno_sacks(sk, pkts_acked);
3077                 } else {
3078                         int delta;
3079 
3080                         /* Non-retransmitted hole got filled? That's reordering */
3081                         if (reord < prior_fackets)
3082                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3083 
3084                         delta = tcp_is_fack(tp) ? pkts_acked :
3085                                                   prior_sacked - tp->sacked_out;
3086                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3087                 }
3088 
3089                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3090 
3091                 if (ca_ops->pkts_acked) {
3092                         s32 rtt_us = -1;
3093 
3094                         /* Is the ACK triggering packet unambiguous? */
3095                         if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3096                                 /* High resolution needed and available? */
3097                                 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3098                                     !ktime_equal(last_ackt,
3099                                                  net_invalid_timestamp()))
3100                                         rtt_us = ktime_us_delta(ktime_get_real(),
3101                                                                 last_ackt);
3102                                 else if (ca_seq_rtt >= 0)
3103                                         rtt_us = jiffies_to_usecs(ca_seq_rtt);
3104                         }
3105 
3106                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3107                 }
3108         } else if (skb && rtt_update && sack_rtt >= 0 &&
3109                    sack_rtt > (s32)(now - TCP_SKB_CB(skb)->when)) {
3110                 /* Do not re-arm RTO if the sack RTT is measured from data sent
3111                  * after when the head was last (re)transmitted. Otherwise the
3112                  * timeout may continue to extend in loss recovery.
3113                  */
3114                 tcp_rearm_rto(sk);
3115         }
3116 
3117 #if FASTRETRANS_DEBUG > 0
3118         WARN_ON((int)tp->sacked_out < 0);
3119         WARN_ON((int)tp->lost_out < 0);
3120         WARN_ON((int)tp->retrans_out < 0);
3121         if (!tp->packets_out && tcp_is_sack(tp)) {
3122                 icsk = inet_csk(sk);
3123                 if (tp->lost_out) {
3124                         pr_debug("Leak l=%u %d\n",
3125                                  tp->lost_out, icsk->icsk_ca_state);
3126                         tp->lost_out = 0;
3127                 }
3128                 if (tp->sacked_out) {
3129                         pr_debug("Leak s=%u %d\n",
3130                                  tp->sacked_out, icsk->icsk_ca_state);
3131                         tp->sacked_out = 0;
3132                 }
3133                 if (tp->retrans_out) {
3134                         pr_debug("Leak r=%u %d\n",
3135                                  tp->retrans_out, icsk->icsk_ca_state);
3136                         tp->retrans_out = 0;
3137                 }
3138         }
3139 #endif
3140         return flag;
3141 }
3142 
3143 static void tcp_ack_probe(struct sock *sk)
3144 {
3145         const struct tcp_sock *tp = tcp_sk(sk);
3146         struct inet_connection_sock *icsk = inet_csk(sk);
3147 
3148         /* Was it a usable window open? */
3149 
3150         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3151                 icsk->icsk_backoff = 0;
3152                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3153                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3154                  * This function is not for random using!
3155                  */
3156         } else {
3157                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3158                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3159                                           TCP_RTO_MAX);
3160         }
3161 }
3162 
3163 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3164 {
3165         return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3166                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3167 }
3168 
3169 /* Decide wheather to run the increase function of congestion control. */
3170 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3171 {
3172         if (tcp_in_cwnd_reduction(sk))
3173                 return false;
3174 
3175         /* If reordering is high then always grow cwnd whenever data is
3176          * delivered regardless of its ordering. Otherwise stay conservative
3177          * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3178          * new SACK or ECE mark may first advance cwnd here and later reduce
3179          * cwnd in tcp_fastretrans_alert() based on more states.
3180          */
3181         if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3182                 return flag & FLAG_FORWARD_PROGRESS;
3183 
3184         return flag & FLAG_DATA_ACKED;
3185 }
3186 
3187 /* Check that window update is acceptable.
3188  * The function assumes that snd_una<=ack<=snd_next.
3189  */
3190 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3191                                         const u32 ack, const u32 ack_seq,
3192                                         const u32 nwin)
3193 {
3194         return  after(ack, tp->snd_una) ||
3195                 after(ack_seq, tp->snd_wl1) ||
3196                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3197 }
3198 
3199 /* Update our send window.
3200  *
3201  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3202  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3203  */
3204 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3205                                  u32 ack_seq)
3206 {
3207         struct tcp_sock *tp = tcp_sk(sk);
3208         int flag = 0;
3209         u32 nwin = ntohs(tcp_hdr(skb)->window);
3210 
3211         if (likely(!tcp_hdr(skb)->syn))
3212                 nwin <<= tp->rx_opt.snd_wscale;
3213 
3214         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3215                 flag |= FLAG_WIN_UPDATE;
3216                 tcp_update_wl(tp, ack_seq);
3217 
3218                 if (tp->snd_wnd != nwin) {
3219                         tp->snd_wnd = nwin;
3220 
3221                         /* Note, it is the only place, where
3222                          * fast path is recovered for sending TCP.
3223                          */
3224                         tp->pred_flags = 0;
3225                         tcp_fast_path_check(sk);
3226 
3227                         if (nwin > tp->max_window) {
3228                                 tp->max_window = nwin;
3229                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3230                         }
3231                 }
3232         }
3233 
3234         tp->snd_una = ack;
3235 
3236         return flag;
3237 }
3238 
3239 /* RFC 5961 7 [ACK Throttling] */
3240 static void tcp_send_challenge_ack(struct sock *sk)
3241 {
3242         /* unprotected vars, we dont care of overwrites */
3243         static u32 challenge_timestamp;
3244         static unsigned int challenge_count;
3245         u32 count, now = jiffies / HZ;
3246 
3247         if (now != challenge_timestamp) {
3248                 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3249 
3250                 challenge_timestamp = now;
3251                 WRITE_ONCE(challenge_count, half +
3252                            prandom_u32_max(sysctl_tcp_challenge_ack_limit));
3253         }
3254         count = READ_ONCE(challenge_count);
3255         if (count > 0) {
3256                 WRITE_ONCE(challenge_count, count - 1);
3257                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3258                 tcp_send_ack(sk);
3259         }
3260 }
3261 
3262 static void tcp_store_ts_recent(struct tcp_sock *tp)
3263 {
3264         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3265         tp->rx_opt.ts_recent_stamp = get_seconds();
3266 }
3267 
3268 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3269 {
3270         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3271                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3272                  * extra check below makes sure this can only happen
3273                  * for pure ACK frames.  -DaveM
3274                  *
3275                  * Not only, also it occurs for expired timestamps.
3276                  */
3277 
3278                 if (tcp_paws_check(&tp->rx_opt, 0))
3279                         tcp_store_ts_recent(tp);
3280         }
3281 }
3282 
3283 /* This routine deals with acks during a TLP episode.
3284  * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3285  */
3286 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3287 {
3288         struct tcp_sock *tp = tcp_sk(sk);
3289         bool is_tlp_dupack = (ack == tp->tlp_high_seq) &&
3290                              !(flag & (FLAG_SND_UNA_ADVANCED |
3291                                        FLAG_NOT_DUP | FLAG_DATA_SACKED));
3292 
3293         /* Mark the end of TLP episode on receiving TLP dupack or when
3294          * ack is after tlp_high_seq.
3295          */
3296         if (is_tlp_dupack) {
3297                 tp->tlp_high_seq = 0;
3298                 return;
3299         }
3300 
3301         if (after(ack, tp->tlp_high_seq)) {
3302                 tp->tlp_high_seq = 0;
3303                 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3304                 if (!(flag & FLAG_DSACKING_ACK)) {
3305                         tcp_init_cwnd_reduction(sk, true);
3306                         tcp_set_ca_state(sk, TCP_CA_CWR);
3307                         tcp_end_cwnd_reduction(sk);
3308                         tcp_try_keep_open(sk);
3309                         NET_INC_STATS_BH(sock_net(sk),
3310                                          LINUX_MIB_TCPLOSSPROBERECOVERY);
3311                 }
3312         }
3313 }
3314 
3315 /* This routine deals with incoming acks, but not outgoing ones. */
3316 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3317 {
3318         struct inet_connection_sock *icsk = inet_csk(sk);
3319         struct tcp_sock *tp = tcp_sk(sk);
3320         u32 prior_snd_una = tp->snd_una;
3321         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3322         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3323         bool is_dupack = false;
3324         u32 prior_in_flight, prior_cwnd = tp->snd_cwnd, prior_rtt = tp->srtt;
3325         u32 prior_fackets;
3326         int prior_packets = tp->packets_out;
3327         const int prior_unsacked = tp->packets_out - tp->sacked_out;
3328         int acked = 0; /* Number of packets newly acked */
3329         s32 sack_rtt = -1;
3330 
3331         /* If the ack is older than previous acks
3332          * then we can probably ignore it.
3333          */
3334         if (before(ack, prior_snd_una)) {
3335                 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3336                 if (before(ack, prior_snd_una - tp->max_window)) {
3337                         tcp_send_challenge_ack(sk);
3338                         return -1;
3339                 }
3340                 goto old_ack;
3341         }
3342 
3343         /* If the ack includes data we haven't sent yet, discard
3344          * this segment (RFC793 Section 3.9).
3345          */
3346         if (after(ack, tp->snd_nxt))
3347                 goto invalid_ack;
3348 
3349         if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3350             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3351                 tcp_rearm_rto(sk);
3352 
3353         if (after(ack, prior_snd_una)) {
3354                 flag |= FLAG_SND_UNA_ADVANCED;
3355                 icsk->icsk_retransmits = 0;
3356         }
3357 
3358         prior_fackets = tp->fackets_out;
3359         prior_in_flight = tcp_packets_in_flight(tp);
3360 
3361         /* ts_recent update must be made after we are sure that the packet
3362          * is in window.
3363          */
3364         if (flag & FLAG_UPDATE_TS_RECENT)
3365                 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3366 
3367         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3368                 /* Window is constant, pure forward advance.
3369                  * No more checks are required.
3370                  * Note, we use the fact that SND.UNA>=SND.WL2.
3371                  */
3372                 tcp_update_wl(tp, ack_seq);
3373                 tp->snd_una = ack;
3374                 flag |= FLAG_WIN_UPDATE;
3375 
3376                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3377 
3378                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3379         } else {
3380                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3381                         flag |= FLAG_DATA;
3382                 else
3383                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3384 
3385                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3386 
3387                 if (TCP_SKB_CB(skb)->sacked)
3388                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3389                                                         &sack_rtt);
3390 
3391                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3392                         flag |= FLAG_ECE;
3393 
3394                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3395         }
3396 
3397         /* We passed data and got it acked, remove any soft error
3398          * log. Something worked...
3399          */
3400         sk->sk_err_soft = 0;
3401         icsk->icsk_probes_out = 0;
3402         tp->rcv_tstamp = tcp_time_stamp;
3403         if (!prior_packets)
3404                 goto no_queue;
3405 
3406         /* See if we can take anything off of the retransmit queue. */
3407         acked = tp->packets_out;
3408         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, sack_rtt);
3409         acked -= tp->packets_out;
3410 
3411         /* Advance cwnd if state allows */
3412         if (tcp_may_raise_cwnd(sk, flag))
3413                 tcp_cong_avoid(sk, ack, prior_in_flight);
3414 
3415         if (tcp_ack_is_dubious(sk, flag)) {
3416                 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3417                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3418                                       is_dupack, flag);
3419         }
3420         if (tp->tlp_high_seq)
3421                 tcp_process_tlp_ack(sk, ack, flag);
3422 
3423         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3424                 struct dst_entry *dst = __sk_dst_get(sk);
3425                 if (dst)
3426                         dst_confirm(dst);
3427         }
3428 
3429         if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3430                 tcp_schedule_loss_probe(sk);
3431         if (tp->srtt != prior_rtt || tp->snd_cwnd != prior_cwnd)
3432                 tcp_update_pacing_rate(sk);
3433         return 1;
3434 
3435 no_queue:
3436         /* If data was DSACKed, see if we can undo a cwnd reduction. */
3437         if (flag & FLAG_DSACKING_ACK)
3438                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3439                                       is_dupack, flag);
3440         /* If this ack opens up a zero window, clear backoff.  It was
3441          * being used to time the probes, and is probably far higher than
3442          * it needs to be for normal retransmission.
3443          */
3444         if (tcp_send_head(sk))
3445                 tcp_ack_probe(sk);
3446 
3447         if (tp->tlp_high_seq)
3448                 tcp_process_tlp_ack(sk, ack, flag);
3449         return 1;
3450 
3451 invalid_ack:
3452         SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3453         return -1;
3454 
3455 old_ack:
3456         /* If data was SACKed, tag it and see if we should send more data.
3457          * If data was DSACKed, see if we can undo a cwnd reduction.
3458          */
3459         if (TCP_SKB_CB(skb)->sacked) {
3460                 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3461                                                 &sack_rtt);
3462                 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3463                                       is_dupack, flag);
3464         }
3465 
3466         SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3467         return 0;
3468 }
3469 
3470 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3471  * But, this can also be called on packets in the established flow when
3472  * the fast version below fails.
3473  */
3474 void tcp_parse_options(const struct sk_buff *skb,
3475                        struct tcp_options_received *opt_rx, int estab,
3476                        struct tcp_fastopen_cookie *foc)
3477 {
3478         const unsigned char *ptr;
3479         const struct tcphdr *th = tcp_hdr(skb);
3480         int length = (th->doff * 4) - sizeof(struct tcphdr);
3481 
3482         ptr = (const unsigned char *)(th + 1);
3483         opt_rx->saw_tstamp = 0;
3484 
3485         while (length > 0) {
3486                 int opcode = *ptr++;
3487                 int opsize;
3488 
3489                 switch (opcode) {
3490                 case TCPOPT_EOL:
3491                         return;
3492                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3493                         length--;
3494                         continue;
3495                 default:
3496                         opsize = *ptr++;
3497                         if (opsize < 2) /* "silly options" */
3498                                 return;
3499                         if (opsize > length)
3500                                 return; /* don't parse partial options */
3501                         switch (opcode) {
3502                         case TCPOPT_MSS:
3503                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3504                                         u16 in_mss = get_unaligned_be16(ptr);
3505                                         if (in_mss) {
3506                                                 if (opt_rx->user_mss &&
3507                                                     opt_rx->user_mss < in_mss)
3508                                                         in_mss = opt_rx->user_mss;
3509                                                 opt_rx->mss_clamp = in_mss;
3510                                         }
3511                                 }
3512                                 break;
3513                         case TCPOPT_WINDOW:
3514                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3515                                     !estab && sysctl_tcp_window_scaling) {
3516                                         __u8 snd_wscale = *(__u8 *)ptr;
3517                                         opt_rx->wscale_ok = 1;
3518                                         if (snd_wscale > 14) {
3519                                                 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3520                                                                      __func__,
3521                                                                      snd_wscale);
3522                                                 snd_wscale = 14;
3523                                         }
3524                                         opt_rx->snd_wscale = snd_wscale;
3525                                 }
3526                                 break;
3527                         case TCPOPT_TIMESTAMP:
3528                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3529                                     ((estab && opt_rx->tstamp_ok) ||
3530                                      (!estab && sysctl_tcp_timestamps))) {
3531                                         opt_rx->saw_tstamp = 1;
3532                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3533                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3534                                 }
3535                                 break;
3536                         case TCPOPT_SACK_PERM:
3537                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3538                                     !estab && sysctl_tcp_sack) {
3539                                         opt_rx->sack_ok = TCP_SACK_SEEN;
3540                                         tcp_sack_reset(opt_rx);
3541                                 }
3542                                 break;
3543 
3544                         case TCPOPT_SACK:
3545                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3546                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3547                                    opt_rx->sack_ok) {
3548                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3549                                 }
3550                                 break;
3551 #ifdef CONFIG_TCP_MD5SIG
3552                         case TCPOPT_MD5SIG:
3553                                 /*
3554                                  * The MD5 Hash has already been
3555                                  * checked (see tcp_v{4,6}_do_rcv()).
3556                                  */
3557                                 break;
3558 #endif
3559                         case TCPOPT_EXP:
3560                                 /* Fast Open option shares code 254 using a
3561                                  * 16 bits magic number. It's valid only in
3562                                  * SYN or SYN-ACK with an even size.
3563                                  */
3564                                 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3565                                     get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3566                                     foc == NULL || !th->syn || (opsize & 1))
3567                                         break;
3568                                 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3569                                 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3570                                     foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3571                                         memcpy(foc->val, ptr + 2, foc->len);
3572                                 else if (foc->len != 0)
3573                                         foc->len = -1;
3574                                 break;
3575 
3576                         }
3577                         ptr += opsize-2;
3578                         length -= opsize;
3579                 }
3580         }
3581 }
3582 EXPORT_SYMBOL(tcp_parse_options);
3583 
3584 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3585 {
3586         const __be32 *ptr = (const __be32 *)(th + 1);
3587 
3588         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3589                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3590                 tp->rx_opt.saw_tstamp = 1;
3591                 ++ptr;
3592                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3593                 ++ptr;
3594                 if (*ptr)
3595                         tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3596                 else
3597                         tp->rx_opt.rcv_tsecr = 0;
3598                 return true;
3599         }
3600         return false;
3601 }
3602 
3603 /* Fast parse options. This hopes to only see timestamps.
3604  * If it is wrong it falls back on tcp_parse_options().
3605  */
3606 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3607                                    const struct tcphdr *th, struct tcp_sock *tp)
3608 {
3609         /* In the spirit of fast parsing, compare doff directly to constant
3610          * values.  Because equality is used, short doff can be ignored here.
3611          */
3612         if (th->doff == (sizeof(*th) / 4)) {
3613                 tp->rx_opt.saw_tstamp = 0;
3614                 return false;
3615         } else if (tp->rx_opt.tstamp_ok &&
3616                    th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3617                 if (tcp_parse_aligned_timestamp(tp, th))
3618                         return true;
3619         }
3620 
3621         tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3622         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3623                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3624 
3625         return true;
3626 }
3627 
3628 #ifdef CONFIG_TCP_MD5SIG
3629 /*
3630  * Parse MD5 Signature option
3631  */
3632 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3633 {
3634         int length = (th->doff << 2) - sizeof(*th);
3635         const u8 *ptr = (const u8 *)(th + 1);
3636 
3637         /* If the TCP option is too short, we can short cut */
3638         if (length < TCPOLEN_MD5SIG)
3639                 return NULL;
3640 
3641         while (length > 0) {
3642                 int opcode = *ptr++;
3643                 int opsize;
3644 
3645                 switch(opcode) {
3646                 case TCPOPT_EOL:
3647                         return NULL;
3648                 case TCPOPT_NOP:
3649                         length--;
3650                         continue;
3651                 default:
3652                         opsize = *ptr++;
3653                         if (opsize < 2 || opsize > length)
3654                                 return NULL;
3655                         if (opcode == TCPOPT_MD5SIG)
3656                                 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3657                 }
3658                 ptr += opsize - 2;
3659                 length -= opsize;
3660         }
3661         return NULL;
3662 }
3663 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3664 #endif
3665 
3666 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3667  *
3668  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3669  * it can pass through stack. So, the following predicate verifies that
3670  * this segment is not used for anything but congestion avoidance or
3671  * fast retransmit. Moreover, we even are able to eliminate most of such
3672  * second order effects, if we apply some small "replay" window (~RTO)
3673  * to timestamp space.
3674  *
3675  * All these measures still do not guarantee that we reject wrapped ACKs
3676  * on networks with high bandwidth, when sequence space is recycled fastly,
3677  * but it guarantees that such events will be very rare and do not affect
3678  * connection seriously. This doesn't look nice, but alas, PAWS is really
3679  * buggy extension.
3680  *
3681  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3682  * states that events when retransmit arrives after original data are rare.
3683  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3684  * the biggest problem on large power networks even with minor reordering.
3685  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3686  * up to bandwidth of 18Gigabit/sec. 8) ]
3687  */
3688 
3689 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3690 {
3691         const struct tcp_sock *tp = tcp_sk(sk);
3692         const struct tcphdr *th = tcp_hdr(skb);
3693         u32 seq = TCP_SKB_CB(skb)->seq;
3694         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3695 
3696         return (/* 1. Pure ACK with correct sequence number. */
3697                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3698 
3699                 /* 2. ... and duplicate ACK. */
3700                 ack == tp->snd_una &&
3701 
3702                 /* 3. ... and does not update window. */
3703                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3704 
3705                 /* 4. ... and sits in replay window. */
3706                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3707 }
3708 
3709 static inline bool tcp_paws_discard(const struct sock *sk,
3710                                    const struct sk_buff *skb)
3711 {
3712         const struct tcp_sock *tp = tcp_sk(sk);
3713 
3714         return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3715                !tcp_disordered_ack(sk, skb);
3716 }
3717 
3718 /* Check segment sequence number for validity.
3719  *
3720  * Segment controls are considered valid, if the segment
3721  * fits to the window after truncation to the window. Acceptability
3722  * of data (and SYN, FIN, of course) is checked separately.
3723  * See tcp_data_queue(), for example.
3724  *
3725  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3726  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3727  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3728  * (borrowed from freebsd)
3729  */
3730 
3731 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3732 {
3733         return  !before(end_seq, tp->rcv_wup) &&
3734                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3735 }
3736 
3737 /* When we get a reset we do this. */
3738 void tcp_reset(struct sock *sk)
3739 {
3740         /* We want the right error as BSD sees it (and indeed as we do). */
3741         switch (sk->sk_state) {
3742         case TCP_SYN_SENT:
3743                 sk->sk_err = ECONNREFUSED;
3744                 break;
3745         case TCP_CLOSE_WAIT:
3746                 sk->sk_err = EPIPE;
3747                 break;
3748         case TCP_CLOSE:
3749                 return;
3750         default:
3751                 sk->sk_err = ECONNRESET;
3752         }
3753         /* This barrier is coupled with smp_rmb() in tcp_poll() */
3754         smp_wmb();
3755 
3756         if (!sock_flag(sk, SOCK_DEAD))
3757                 sk->sk_error_report(sk);
3758 
3759         tcp_done(sk);
3760 }
3761 
3762 /*
3763  *      Process the FIN bit. This now behaves as it is supposed to work
3764  *      and the FIN takes effect when it is validly part of sequence
3765  *      space. Not before when we get holes.
3766  *
3767  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3768  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3769  *      TIME-WAIT)
3770  *
3771  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3772  *      close and we go into CLOSING (and later onto TIME-WAIT)
3773  *
3774  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3775  */
3776 static void tcp_fin(struct sock *sk)
3777 {
3778         struct tcp_sock *tp = tcp_sk(sk);
3779         const struct dst_entry *dst;
3780 
3781         inet_csk_schedule_ack(sk);
3782 
3783         sk->sk_shutdown |= RCV_SHUTDOWN;
3784         sock_set_flag(sk, SOCK_DONE);
3785 
3786         switch (sk->sk_state) {
3787         case TCP_SYN_RECV:
3788         case TCP_ESTABLISHED:
3789                 /* Move to CLOSE_WAIT */
3790                 tcp_set_state(sk, TCP_CLOSE_WAIT);
3791                 dst = __sk_dst_get(sk);
3792                 if (!dst || !dst_metric(dst, RTAX_QUICKACK))
3793                         inet_csk(sk)->icsk_ack.pingpong = 1;
3794                 break;
3795 
3796         case TCP_CLOSE_WAIT:
3797         case TCP_CLOSING:
3798                 /* Received a retransmission of the FIN, do
3799                  * nothing.
3800                  */
3801                 break;
3802         case TCP_LAST_ACK:
3803                 /* RFC793: Remain in the LAST-ACK state. */
3804                 break;
3805 
3806         case TCP_FIN_WAIT1:
3807                 /* This case occurs when a simultaneous close
3808                  * happens, we must ack the received FIN and
3809                  * enter the CLOSING state.
3810                  */
3811                 tcp_send_ack(sk);
3812                 tcp_set_state(sk, TCP_CLOSING);
3813                 break;
3814         case TCP_FIN_WAIT2:
3815                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3816                 tcp_send_ack(sk);
3817                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3818                 break;
3819         default:
3820                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3821                  * cases we should never reach this piece of code.
3822                  */
3823                 pr_err("%s: Impossible, sk->sk_state=%d\n",
3824                        __func__, sk->sk_state);
3825                 break;
3826         }
3827 
3828         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3829          * Probably, we should reset in this case. For now drop them.
3830          */
3831         __skb_queue_purge(&tp->out_of_order_queue);
3832         if (tcp_is_sack(tp))
3833                 tcp_sack_reset(&tp->rx_opt);
3834         sk_mem_reclaim(sk);
3835 
3836         if (!sock_flag(sk, SOCK_DEAD)) {
3837                 sk->sk_state_change(sk);
3838 
3839                 /* Do not send POLL_HUP for half duplex close. */
3840                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3841                     sk->sk_state == TCP_CLOSE)
3842                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3843                 else
3844                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3845         }
3846 }
3847 
3848 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3849                                   u32 end_seq)
3850 {
3851         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3852                 if (before(seq, sp->start_seq))
3853                         sp->start_seq = seq;
3854                 if (after(end_seq, sp->end_seq))
3855                         sp->end_seq = end_seq;
3856                 return true;
3857         }
3858         return false;
3859 }
3860 
3861 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3862 {
3863         struct tcp_sock *tp = tcp_sk(sk);
3864 
3865         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3866                 int mib_idx;
3867 
3868                 if (before(seq, tp->rcv_nxt))
3869                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3870                 else
3871                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3872 
3873                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3874 
3875                 tp->rx_opt.dsack = 1;
3876                 tp->duplicate_sack[0].start_seq = seq;
3877                 tp->duplicate_sack[0].end_seq = end_seq;
3878         }
3879 }
3880 
3881 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3882 {
3883         struct tcp_sock *tp = tcp_sk(sk);
3884 
3885         if (!tp->rx_opt.dsack)
3886                 tcp_dsack_set(sk, seq, end_seq);
3887         else
3888                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3889 }
3890 
3891 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
3892 {
3893         struct tcp_sock *tp = tcp_sk(sk);
3894 
3895         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3896             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3897                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3898                 tcp_enter_quickack_mode(sk);
3899 
3900                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3901                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3902 
3903                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3904                                 end_seq = tp->rcv_nxt;
3905                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
3906                 }
3907         }
3908 
3909         tcp_send_ack(sk);
3910 }
3911 
3912 /* These routines update the SACK block as out-of-order packets arrive or
3913  * in-order packets close up the sequence space.
3914  */
3915 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3916 {
3917         int this_sack;
3918         struct tcp_sack_block *sp = &tp->selective_acks[0];
3919         struct tcp_sack_block *swalk = sp + 1;
3920 
3921         /* See if the recent change to the first SACK eats into
3922          * or hits the sequence space of other SACK blocks, if so coalesce.
3923          */
3924         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3925                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3926                         int i;
3927 
3928                         /* Zap SWALK, by moving every further SACK up by one slot.
3929                          * Decrease num_sacks.
3930                          */
3931                         tp->rx_opt.num_sacks--;
3932                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3933                                 sp[i] = sp[i + 1];
3934                         continue;
3935                 }
3936                 this_sack++, swalk++;
3937         }
3938 }
3939 
3940 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3941 {
3942         struct tcp_sock *tp = tcp_sk(sk);
3943         struct tcp_sack_block *sp = &tp->selective_acks[0];
3944         int cur_sacks = tp->rx_opt.num_sacks;
3945         int this_sack;
3946 
3947         if (!cur_sacks)
3948                 goto new_sack;
3949 
3950         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
3951                 if (tcp_sack_extend(sp, seq, end_seq)) {
3952                         /* Rotate this_sack to the first one. */
3953                         for (; this_sack > 0; this_sack--, sp--)
3954                                 swap(*sp, *(sp - 1));
3955                         if (cur_sacks > 1)
3956                                 tcp_sack_maybe_coalesce(tp);
3957                         return;
3958                 }
3959         }
3960 
3961         /* Could not find an adjacent existing SACK, build a new one,
3962          * put it at the front, and shift everyone else down.  We
3963          * always know there is at least one SACK present already here.
3964          *
3965          * If the sack array is full, forget about the last one.
3966          */
3967         if (this_sack >= TCP_NUM_SACKS) {
3968                 this_sack--;
3969                 tp->rx_opt.num_sacks--;
3970                 sp--;
3971         }
3972         for (; this_sack > 0; this_sack--, sp--)
3973                 *sp = *(sp - 1);
3974 
3975 new_sack:
3976         /* Build the new head SACK, and we're done. */
3977         sp->start_seq = seq;
3978         sp->end_seq = end_seq;
3979         tp->rx_opt.num_sacks++;
3980 }
3981 
3982 /* RCV.NXT advances, some SACKs should be eaten. */
3983 
3984 static void tcp_sack_remove(struct tcp_sock *tp)
3985 {
3986         struct tcp_sack_block *sp = &tp->selective_acks[0];
3987         int num_sacks = tp->rx_opt.num_sacks;
3988         int this_sack;
3989 
3990         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3991         if (skb_queue_empty(&tp->out_of_order_queue)) {
3992                 tp->rx_opt.num_sacks = 0;
3993                 return;
3994         }
3995 
3996         for (this_sack = 0; this_sack < num_sacks;) {
3997                 /* Check if the start of the sack is covered by RCV.NXT. */
3998                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3999                         int i;
4000 
4001                         /* RCV.NXT must cover all the block! */
4002                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4003 
4004                         /* Zap this SACK, by moving forward any other SACKS. */
4005                         for (i=this_sack+1; i < num_sacks; i++)
4006                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4007                         num_sacks--;
4008                         continue;
4009                 }
4010                 this_sack++;
4011                 sp++;
4012         }
4013         tp->rx_opt.num_sacks = num_sacks;
4014 }
4015 
4016 /* This one checks to see if we can put data from the
4017  * out_of_order queue into the receive_queue.
4018  */
4019 static void tcp_ofo_queue(struct sock *sk)
4020 {
4021         struct tcp_sock *tp = tcp_sk(sk);
4022         __u32 dsack_high = tp->rcv_nxt;
4023         struct sk_buff *skb;
4024 
4025         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4026                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4027                         break;
4028 
4029                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4030                         __u32 dsack = dsack_high;
4031                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4032                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4033                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4034                 }
4035 
4036                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4037                         SOCK_DEBUG(sk, "ofo packet was already received\n");
4038                         __skb_unlink(skb, &tp->out_of_order_queue);
4039                         __kfree_skb(skb);
4040                         continue;
4041                 }
4042                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4043                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4044                            TCP_SKB_CB(skb)->end_seq);
4045 
4046                 __skb_unlink(skb, &tp->out_of_order_queue);
4047                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4048                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4049                 if (tcp_hdr(skb)->fin)
4050                         tcp_fin(sk);
4051         }
4052 }
4053 
4054 static bool tcp_prune_ofo_queue(struct sock *sk);
4055 static int tcp_prune_queue(struct sock *sk);
4056 
4057 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4058                                  unsigned int size)
4059 {
4060         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4061             !sk_rmem_schedule(sk, skb, size)) {
4062 
4063                 if (tcp_prune_queue(sk) < 0)
4064                         return -1;
4065 
4066                 if (!sk_rmem_schedule(sk, skb, size)) {
4067                         if (!tcp_prune_ofo_queue(sk))
4068                                 return -1;
4069 
4070                         if (!sk_rmem_schedule(sk, skb, size))
4071                                 return -1;
4072                 }
4073         }
4074         return 0;
4075 }
4076 
4077 /**
4078  * tcp_try_coalesce - try to merge skb to prior one
4079  * @sk: socket
4080  * @to: prior buffer
4081  * @from: buffer to add in queue
4082  * @fragstolen: pointer to boolean
4083  *
4084  * Before queueing skb @from after @to, try to merge them
4085  * to reduce overall memory use and queue lengths, if cost is small.
4086  * Packets in ofo or receive queues can stay a long time.
4087  * Better try to coalesce them right now to avoid future collapses.
4088  * Returns true if caller should free @from instead of queueing it
4089  */
4090 static bool tcp_try_coalesce(struct sock *sk,
4091                              struct sk_buff *to,
4092                              struct sk_buff *from,
4093                              bool *fragstolen)
4094 {
4095         int delta;
4096 
4097         *fragstolen = false;
4098 
4099         if (tcp_hdr(from)->fin)
4100                 return false;
4101 
4102         /* Its possible this segment overlaps with prior segment in queue */
4103         if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4104                 return false;
4105 
4106         if (!skb_try_coalesce(to, from, fragstolen, &delta))
4107                 return false;
4108 
4109         atomic_add(delta, &sk->sk_rmem_alloc);
4110         sk_mem_charge(sk, delta);
4111         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4112         TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4113         TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4114         return true;
4115 }
4116 
4117 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4118 {
4119         struct tcp_sock *tp = tcp_sk(sk);
4120         struct sk_buff *skb1;
4121         u32 seq, end_seq;
4122 
4123         TCP_ECN_check_ce(tp, skb);
4124 
4125         if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4126                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4127                 __kfree_skb(skb);
4128                 return;
4129         }
4130 
4131         /* Disable header prediction. */
4132         tp->pred_flags = 0;
4133         inet_csk_schedule_ack(sk);
4134 
4135         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4136         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4137                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4138 
4139         skb1 = skb_peek_tail(&tp->out_of_order_queue);
4140         if (!skb1) {
4141                 /* Initial out of order segment, build 1 SACK. */
4142                 if (tcp_is_sack(tp)) {
4143                         tp->rx_opt.num_sacks = 1;
4144                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4145                         tp->selective_acks[0].end_seq =
4146                                                 TCP_SKB_CB(skb)->end_seq;
4147                 }
4148                 __skb_queue_head(&tp->out_of_order_queue, skb);
4149                 goto end;
4150         }
4151 
4152         seq = TCP_SKB_CB(skb)->seq;
4153         end_seq = TCP_SKB_CB(skb)->end_seq;
4154 
4155         if (seq == TCP_SKB_CB(skb1)->end_seq) {
4156                 bool fragstolen;
4157 
4158                 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4159                         __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4160                 } else {
4161                         tcp_grow_window(sk, skb);
4162                         kfree_skb_partial(skb, fragstolen);
4163                         skb = NULL;
4164                 }
4165 
4166                 if (!tp->rx_opt.num_sacks ||
4167                     tp->selective_acks[0].end_seq != seq)
4168                         goto add_sack;
4169 
4170                 /* Common case: data arrive in order after hole. */
4171                 tp->selective_acks[0].end_seq = end_seq;
4172                 goto end;
4173         }
4174 
4175         /* Find place to insert this segment. */
4176         while (1) {
4177                 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4178                         break;
4179                 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4180                         skb1 = NULL;
4181                         break;
4182                 }
4183                 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4184         }
4185 
4186         /* Do skb overlap to previous one? */
4187         if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4188                 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4189                         /* All the bits are present. Drop. */
4190                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4191                         __kfree_skb(skb);
4192                         skb = NULL;
4193                         tcp_dsack_set(sk, seq, end_seq);
4194                         goto add_sack;
4195                 }
4196                 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4197                         /* Partial overlap. */
4198                         tcp_dsack_set(sk, seq,
4199                                       TCP_SKB_CB(skb1)->end_seq);
4200                 } else {
4201                         if (skb_queue_is_first(&tp->out_of_order_queue,
4202                                                skb1))
4203                                 skb1 = NULL;
4204                         else
4205                                 skb1 = skb_queue_prev(
4206                                         &tp->out_of_order_queue,
4207                                         skb1);
4208                 }
4209         }
4210         if (!skb1)
4211                 __skb_queue_head(&tp->out_of_order_queue, skb);
4212         else
4213                 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4214 
4215         /* And clean segments covered by new one as whole. */
4216         while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4217                 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4218 
4219                 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4220                         break;
4221                 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4222                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4223                                          end_seq);
4224                         break;
4225                 }
4226                 __skb_unlink(skb1, &tp->out_of_order_queue);
4227                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4228                                  TCP_SKB_CB(skb1)->end_seq);
4229                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4230                 __kfree_skb(skb1);
4231         }
4232 
4233 add_sack:
4234         if (tcp_is_sack(tp))
4235                 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4236 end:
4237         if (skb) {
4238                 tcp_grow_window(sk, skb);
4239                 skb_set_owner_r(skb, sk);
4240         }
4241 }
4242 
4243 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4244                   bool *fragstolen)
4245 {
4246         int eaten;
4247         struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4248 
4249         __skb_pull(skb, hdrlen);
4250         eaten = (tail &&
4251                  tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4252         tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4253         if (!eaten) {
4254                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4255                 skb_set_owner_r(skb, sk);
4256         }
4257         return eaten;
4258 }
4259 
4260 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4261 {
4262         struct sk_buff *skb = NULL;
4263         struct tcphdr *th;
4264         bool fragstolen;
4265 
4266         if (size == 0)
4267                 return 0;
4268 
4269         skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
4270         if (!skb)
4271                 goto err;
4272 
4273         if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th)))
4274                 goto err_free;
4275 
4276         th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4277         skb_reset_transport_header(skb);
4278         memset(th, 0, sizeof(*th));
4279 
4280         if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4281                 goto err_free;
4282 
4283         TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4284         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4285         TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4286 
4287         if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4288                 WARN_ON_ONCE(fragstolen); /* should not happen */
4289                 __kfree_skb(skb);
4290         }
4291         return size;
4292 
4293 err_free:
4294         kfree_skb(skb);
4295 err:
4296         return -ENOMEM;
4297 }
4298 
4299 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4300 {
4301         const struct tcphdr *th = tcp_hdr(skb);
4302         struct tcp_sock *tp = tcp_sk(sk);
4303         int eaten = -1;
4304         bool fragstolen = false;
4305 
4306         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4307                 goto drop;
4308 
4309         skb_dst_drop(skb);
4310         __skb_pull(skb, th->doff * 4);
4311 
4312         TCP_ECN_accept_cwr(tp, skb);
4313 
4314         tp->rx_opt.dsack = 0;
4315 
4316         /*  Queue data for delivery to the user.
4317          *  Packets in sequence go to the receive queue.
4318          *  Out of sequence packets to the out_of_order_queue.
4319          */
4320         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4321                 if (tcp_receive_window(tp) == 0)
4322                         goto out_of_window;
4323 
4324                 /* Ok. In sequence. In window. */
4325                 if (tp->ucopy.task == current &&
4326                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4327                     sock_owned_by_user(sk) && !tp->urg_data) {
4328                         int chunk = min_t(unsigned int, skb->len,
4329                                           tp->ucopy.len);
4330 
4331                         __set_current_state(TASK_RUNNING);
4332 
4333                         local_bh_enable();
4334                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4335                                 tp->ucopy.len -= chunk;
4336                                 tp->copied_seq += chunk;
4337                                 eaten = (chunk == skb->len);
4338                                 tcp_rcv_space_adjust(sk);
4339                         }
4340                         local_bh_disable();
4341                 }
4342 
4343                 if (eaten <= 0) {
4344 queue_and_out:
4345                         if (eaten < 0 &&
4346                             tcp_try_rmem_schedule(sk, skb, skb->truesize))
4347                                 goto drop;
4348 
4349                         eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4350                 }
4351                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4352                 if (skb->len)
4353                         tcp_event_data_recv(sk, skb);
4354                 if (th->fin)
4355                         tcp_fin(sk);
4356 
4357                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4358                         tcp_ofo_queue(sk);
4359 
4360                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4361                          * gap in queue is filled.
4362                          */
4363                         if (skb_queue_empty(&tp->out_of_order_queue))
4364                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4365                 }
4366 
4367                 if (tp->rx_opt.num_sacks)
4368                         tcp_sack_remove(tp);
4369 
4370                 tcp_fast_path_check(sk);
4371 
4372                 if (eaten > 0)
4373                         kfree_skb_partial(skb, fragstolen);
4374                 if (!sock_flag(sk, SOCK_DEAD))
4375                         sk->sk_data_ready(sk, 0);
4376                 return;
4377         }
4378 
4379         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4380                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4381                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4382                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4383 
4384 out_of_window:
4385                 tcp_enter_quickack_mode(sk);
4386                 inet_csk_schedule_ack(sk);
4387 drop:
4388                 __kfree_skb(skb);
4389                 return;
4390         }
4391 
4392         /* Out of window. F.e. zero window probe. */
4393         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4394                 goto out_of_window;
4395 
4396         tcp_enter_quickack_mode(sk);
4397 
4398         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4399                 /* Partial packet, seq < rcv_next < end_seq */
4400                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4401                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4402                            TCP_SKB_CB(skb)->end_seq);
4403 
4404                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4405 
4406                 /* If window is closed, drop tail of packet. But after
4407                  * remembering D-SACK for its head made in previous line.
4408                  */
4409                 if (!tcp_receive_window(tp))
4410                         goto out_of_window;
4411                 goto queue_and_out;
4412         }
4413 
4414         tcp_data_queue_ofo(sk, skb);
4415 }
4416 
4417 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4418                                         struct sk_buff_head *list)
4419 {
4420         struct sk_buff *next = NULL;
4421 
4422         if (!skb_queue_is_last(list, skb))
4423                 next = skb_queue_next(list, skb);
4424 
4425         __skb_unlink(skb, list);
4426         __kfree_skb(skb);
4427         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4428 
4429         return next;
4430 }
4431 
4432 /* Collapse contiguous sequence of skbs head..tail with
4433  * sequence numbers start..end.
4434  *
4435  * If tail is NULL, this means until the end of the list.
4436  *
4437  * Segments with FIN/SYN are not collapsed (only because this
4438  * simplifies code)
4439  */
4440 static void
4441 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4442              struct sk_buff *head, struct sk_buff *tail,
4443              u32 start, u32 end)
4444 {
4445         struct sk_buff *skb, *n;
4446         bool end_of_skbs;
4447 
4448         /* First, check that queue is collapsible and find
4449          * the point where collapsing can be useful. */
4450         skb = head;
4451 restart:
4452         end_of_skbs = true;
4453         skb_queue_walk_from_safe(list, skb, n) {
4454                 if (skb == tail)
4455                         break;
4456                 /* No new bits? It is possible on ofo queue. */
4457                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4458                         skb = tcp_collapse_one(sk, skb, list);
4459                         if (!skb)
4460                                 break;
4461                         goto restart;
4462                 }
4463 
4464                 /* The first skb to collapse is:
4465                  * - not SYN/FIN and
4466                  * - bloated or contains data before "start" or
4467                  *   overlaps to the next one.
4468                  */
4469                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4470                     (tcp_win_from_space(skb->truesize) > skb->len ||
4471                      before(TCP_SKB_CB(skb)->seq, start))) {
4472                         end_of_skbs = false;
4473                         break;
4474                 }
4475 
4476                 if (!skb_queue_is_last(list, skb)) {
4477                         struct sk_buff *next = skb_queue_next(list, skb);
4478                         if (next != tail &&
4479                             TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4480                                 end_of_skbs = false;
4481                                 break;
4482                         }
4483                 }
4484 
4485                 /* Decided to skip this, advance start seq. */
4486                 start = TCP_SKB_CB(skb)->end_seq;
4487         }
4488         if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4489                 return;
4490 
4491         while (before(start, end)) {
4492                 struct sk_buff *nskb;
4493                 unsigned int header = skb_headroom(skb);
4494                 int copy = SKB_MAX_ORDER(header, 0);
4495 
4496                 /* Too big header? This can happen with IPv6. */
4497                 if (copy < 0)
4498                         return;
4499                 if (end - start < copy)
4500                         copy = end - start;
4501                 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4502                 if (!nskb)
4503                         return;
4504 
4505                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4506                 skb_set_network_header(nskb, (skb_network_header(skb) -
4507                                               skb->head));
4508                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4509                                                 skb->head));
4510                 skb_reserve(nskb, header);
4511                 memcpy(nskb->head, skb->head, header);
4512                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4513                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4514                 __skb_queue_before(list, skb, nskb);
4515                 skb_set_owner_r(nskb, sk);
4516 
4517                 /* Copy data, releasing collapsed skbs. */
4518                 while (copy > 0) {
4519                         int offset = start - TCP_SKB_CB(skb)->seq;
4520                         int size = TCP_SKB_CB(skb)->end_seq - start;
4521 
4522                         BUG_ON(offset < 0);
4523                         if (size > 0) {
4524                                 size = min(copy, size);
4525                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4526                                         BUG();
4527                                 TCP_SKB_CB(nskb)->end_seq += size;
4528                                 copy -= size;
4529                                 start += size;
4530                         }
4531                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4532                                 skb = tcp_collapse_one(sk, skb, list);
4533                                 if (!skb ||
4534                                     skb == tail ||
4535                                     tcp_hdr(skb)->syn ||
4536                                     tcp_hdr(skb)->fin)
4537                                         return;
4538                         }
4539                 }
4540         }
4541 }
4542 
4543 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4544  * and tcp_collapse() them until all the queue is collapsed.
4545  */
4546 static void tcp_collapse_ofo_queue(struct sock *sk)
4547 {
4548         struct tcp_sock *tp = tcp_sk(sk);
4549         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4550         struct sk_buff *head;
4551         u32 start, end;
4552 
4553         if (skb == NULL)
4554                 return;
4555 
4556         start = TCP_SKB_CB(skb)->seq;
4557         end = TCP_SKB_CB(skb)->end_seq;
4558         head = skb;
4559 
4560         for (;;) {
4561                 struct sk_buff *next = NULL;
4562 
4563                 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4564                         next = skb_queue_next(&tp->out_of_order_queue, skb);
4565                 skb = next;
4566 
4567                 /* Segment is terminated when we see gap or when
4568                  * we are at the end of all the queue. */
4569                 if (!skb ||
4570                     after(TCP_SKB_CB(skb)->seq, end) ||
4571                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4572                         tcp_collapse(sk, &tp->out_of_order_queue,
4573                                      head, skb, start, end);
4574                         head = skb;
4575                         if (!skb)
4576                                 break;
4577                         /* Start new segment */
4578                         start = TCP_SKB_CB(skb)->seq;
4579                         end = TCP_SKB_CB(skb)->end_seq;
4580                 } else {
4581                         if (before(TCP_SKB_CB(skb)->seq, start))
4582                                 start = TCP_SKB_CB(skb)->seq;
4583                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4584                                 end = TCP_SKB_CB(skb)->end_seq;
4585                 }
4586         }
4587 }
4588 
4589 /*
4590  * Purge the out-of-order queue.
4591  * Return true if queue was pruned.
4592  */
4593 static bool tcp_prune_ofo_queue(struct sock *sk)
4594 {
4595         struct tcp_sock *tp = tcp_sk(sk);
4596         bool res = false;
4597 
4598         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4599                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4600                 __skb_queue_purge(&tp->out_of_order_queue);
4601 
4602                 /* Reset SACK state.  A conforming SACK implementation will
4603                  * do the same at a timeout based retransmit.  When a connection
4604                  * is in a sad state like this, we care only about integrity
4605                  * of the connection not performance.
4606                  */
4607                 if (tp->rx_opt.sack_ok)
4608                         tcp_sack_reset(&tp->rx_opt);
4609                 sk_mem_reclaim(sk);
4610                 res = true;
4611         }
4612         return res;
4613 }
4614 
4615 /* Reduce allocated memory if we can, trying to get
4616  * the socket within its memory limits again.
4617  *
4618  * Return less than zero if we should start dropping frames
4619  * until the socket owning process reads some of the data
4620  * to stabilize the situation.
4621  */
4622 static int tcp_prune_queue(struct sock *sk)
4623 {
4624         struct tcp_sock *tp = tcp_sk(sk);
4625 
4626         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4627 
4628         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4629 
4630         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4631                 tcp_clamp_window(sk);
4632         else if (sk_under_memory_pressure(sk))
4633                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4634 
4635         tcp_collapse_ofo_queue(sk);
4636         if (!skb_queue_empty(&sk->sk_receive_queue))
4637                 tcp_collapse(sk, &sk->sk_receive_queue,
4638                              skb_peek(&sk->sk_receive_queue),
4639                              NULL,
4640                              tp->copied_seq, tp->rcv_nxt);
4641         sk_mem_reclaim(sk);
4642 
4643         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4644                 return 0;
4645 
4646         /* Collapsing did not help, destructive actions follow.
4647          * This must not ever occur. */
4648 
4649         tcp_prune_ofo_queue(sk);
4650 
4651         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4652                 return 0;
4653 
4654         /* If we are really being abused, tell the caller to silently
4655          * drop receive data on the floor.  It will get retransmitted
4656          * and hopefully then we'll have sufficient space.
4657          */
4658         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4659 
4660         /* Massive buffer overcommit. */
4661         tp->pred_flags = 0;
4662         return -1;
4663 }
4664 
4665 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4666  * As additional protections, we do not touch cwnd in retransmission phases,
4667  * and if application hit its sndbuf limit recently.
4668  */
4669 void tcp_cwnd_application_limited(struct sock *sk)
4670 {
4671         struct tcp_sock *tp = tcp_sk(sk);
4672 
4673         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4674             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4675                 /* Limited by application or receiver window. */
4676                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4677                 u32 win_used = max(tp->snd_cwnd_used, init_win);
4678                 if (win_used < tp->snd_cwnd) {
4679                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
4680                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4681                 }
4682                 tp->snd_cwnd_used = 0;
4683         }
4684         tp->snd_cwnd_stamp = tcp_time_stamp;
4685 }
4686 
4687 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4688 {
4689         const struct tcp_sock *tp = tcp_sk(sk);
4690 
4691         /* If the user specified a specific send buffer setting, do
4692          * not modify it.
4693          */
4694         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4695                 return false;
4696 
4697         /* If we are under global TCP memory pressure, do not expand.  */
4698         if (sk_under_memory_pressure(sk))
4699                 return false;
4700 
4701         /* If we are under soft global TCP memory pressure, do not expand.  */
4702         if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4703                 return false;
4704 
4705         /* If we filled the congestion window, do not expand.  */
4706         if (tp->packets_out >= tp->snd_cwnd)
4707                 return false;
4708 
4709         return true;
4710 }
4711 
4712 /* When incoming ACK allowed to free some skb from write_queue,
4713  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4714  * on the exit from tcp input handler.
4715  *
4716  * PROBLEM: sndbuf expansion does not work well with largesend.
4717  */
4718 static void tcp_new_space(struct sock *sk)
4719 {
4720         struct tcp_sock *tp = tcp_sk(sk);
4721 
4722         if (tcp_should_expand_sndbuf(sk)) {
4723                 int sndmem = SKB_TRUESIZE(max_t(u32,
4724                                                 tp->rx_opt.mss_clamp,
4725                                                 tp->mss_cache) +
4726                                           MAX_TCP_HEADER);
4727                 int demanded = max_t(unsigned int, tp->snd_cwnd,
4728                                      tp->reordering + 1);
4729                 sndmem *= 2 * demanded;
4730                 if (sndmem > sk->sk_sndbuf)
4731                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4732                 tp->snd_cwnd_stamp = tcp_time_stamp;
4733         }
4734 
4735         sk->sk_write_space(sk);
4736 }
4737 
4738 static void tcp_check_space(struct sock *sk)
4739 {
4740         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4741                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4742                 if (sk->sk_socket &&
4743                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4744                         tcp_new_space(sk);
4745         }
4746 }
4747 
4748 static inline void tcp_data_snd_check(struct sock *sk)
4749 {
4750         tcp_push_pending_frames(sk);
4751         tcp_check_space(sk);
4752 }
4753 
4754 /*
4755  * Check if sending an ack is needed.
4756  */
4757 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4758 {
4759         struct tcp_sock *tp = tcp_sk(sk);
4760 
4761             /* More than one full frame received... */
4762         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4763              /* ... and right edge of window advances far enough.
4764               * (tcp_recvmsg() will send ACK otherwise). Or...
4765               */
4766              __tcp_select_window(sk) >= tp->rcv_wnd) ||
4767             /* We ACK each frame or... */
4768             tcp_in_quickack_mode(sk) ||
4769             /* We have out of order data. */
4770             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4771                 /* Then ack it now */
4772                 tcp_send_ack(sk);
4773         } else {
4774                 /* Else, send delayed ack. */
4775                 tcp_send_delayed_ack(sk);
4776         }
4777 }
4778 
4779 static inline void tcp_ack_snd_check(struct sock *sk)
4780 {
4781         if (!inet_csk_ack_scheduled(sk)) {
4782                 /* We sent a data segment already. */
4783                 return;
4784         }
4785         __tcp_ack_snd_check(sk, 1);
4786 }
4787 
4788 /*
4789  *      This routine is only called when we have urgent data
4790  *      signaled. Its the 'slow' part of tcp_urg. It could be
4791  *      moved inline now as tcp_urg is only called from one
4792  *      place. We handle URGent data wrong. We have to - as
4793  *      BSD still doesn't use the correction from RFC961.
4794  *      For 1003.1g we should support a new option TCP_STDURG to permit
4795  *      either form (or just set the sysctl tcp_stdurg).
4796  */
4797 
4798 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4799 {
4800         struct tcp_sock *tp = tcp_sk(sk);
4801         u32 ptr = ntohs(th->urg_ptr);
4802 
4803         if (ptr && !sysctl_tcp_stdurg)
4804                 ptr--;
4805         ptr += ntohl(th->seq);
4806 
4807         /* Ignore urgent data that we've already seen and read. */
4808         if (after(tp->copied_seq, ptr))
4809                 return;
4810 
4811         /* Do not replay urg ptr.
4812          *
4813          * NOTE: interesting situation not covered by specs.
4814          * Misbehaving sender may send urg ptr, pointing to segment,
4815          * which we already have in ofo queue. We are not able to fetch
4816          * such data and will stay in TCP_URG_NOTYET until will be eaten
4817          * by recvmsg(). Seems, we are not obliged to handle such wicked
4818          * situations. But it is worth to think about possibility of some
4819          * DoSes using some hypothetical application level deadlock.
4820          */
4821         if (before(ptr, tp->rcv_nxt))
4822                 return;
4823 
4824         /* Do we already have a newer (or duplicate) urgent pointer? */
4825         if (tp->urg_data && !after(ptr, tp->urg_seq))
4826                 return;
4827 
4828         /* Tell the world about our new urgent pointer. */
4829         sk_send_sigurg(sk);
4830 
4831         /* We may be adding urgent data when the last byte read was
4832          * urgent. To do this requires some care. We cannot just ignore
4833          * tp->copied_seq since we would read the last urgent byte again
4834          * as data, nor can we alter copied_seq until this data arrives
4835          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4836          *
4837          * NOTE. Double Dutch. Rendering to plain English: author of comment
4838          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4839          * and expect that both A and B disappear from stream. This is _wrong_.
4840          * Though this happens in BSD with high probability, this is occasional.
4841          * Any application relying on this is buggy. Note also, that fix "works"
4842          * only in this artificial test. Insert some normal data between A and B and we will
4843          * decline of BSD again. Verdict: it is better to remove to trap
4844          * buggy users.
4845          */
4846         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4847             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4848                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4849                 tp->copied_seq++;
4850                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4851                         __skb_unlink(skb, &sk->sk_receive_queue);
4852                         __kfree_skb(skb);
4853                 }
4854         }
4855 
4856         tp->urg_data = TCP_URG_NOTYET;
4857         tp->urg_seq = ptr;
4858 
4859         /* Disable header prediction. */
4860         tp->pred_flags = 0;
4861 }
4862 
4863 /* This is the 'fast' part of urgent handling. */
4864 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4865 {
4866         struct tcp_sock *tp = tcp_sk(sk);
4867 
4868         /* Check if we get a new urgent pointer - normally not. */
4869         if (th->urg)
4870                 tcp_check_urg(sk, th);
4871 
4872         /* Do we wait for any urgent data? - normally not... */
4873         if (tp->urg_data == TCP_URG_NOTYET) {
4874                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4875                           th->syn;
4876 
4877                 /* Is the urgent pointer pointing into this packet? */
4878                 if (ptr < skb->len) {
4879                         u8 tmp;
4880                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4881                                 BUG();
4882                         tp->urg_data = TCP_URG_VALID | tmp;
4883                         if (!sock_flag(sk, SOCK_DEAD))
4884                                 sk->sk_data_ready(sk, 0);
4885                 }
4886         }
4887 }
4888 
4889 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4890 {
4891         struct tcp_sock *tp = tcp_sk(sk);
4892         int chunk = skb->len - hlen;
4893         int err;
4894 
4895         local_bh_enable();
4896         if (skb_csum_unnecessary(skb))
4897                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4898         else
4899                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4900                                                        tp->ucopy.iov, chunk);
4901 
4902         if (!err) {
4903                 tp->ucopy.len -= chunk;
4904                 tp->copied_seq += chunk;
4905                 tcp_rcv_space_adjust(sk);
4906         }
4907 
4908         local_bh_disable();
4909         return err;
4910 }
4911 
4912 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4913                                             struct sk_buff *skb)
4914 {
4915         __sum16 result;
4916 
4917         if (sock_owned_by_user(sk)) {
4918                 local_bh_enable();
4919                 result = __tcp_checksum_complete(skb);
4920                 local_bh_disable();
4921         } else {
4922                 result = __tcp_checksum_complete(skb);
4923         }
4924         return result;
4925 }
4926 
4927 static inline bool tcp_checksum_complete_user(struct sock *sk,
4928                                              struct sk_buff *skb)
4929 {
4930         return !skb_csum_unnecessary(skb) &&
4931                __tcp_checksum_complete_user(sk, skb);
4932 }
4933 
4934 #ifdef CONFIG_NET_DMA
4935 static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4936                                   int hlen)
4937 {
4938         struct tcp_sock *tp = tcp_sk(sk);
4939         int chunk = skb->len - hlen;
4940         int dma_cookie;
4941         bool copied_early = false;
4942 
4943         if (tp->ucopy.wakeup)
4944                 return false;
4945 
4946         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4947                 tp->ucopy.dma_chan = net_dma_find_channel();
4948 
4949         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4950 
4951                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4952                                                          skb, hlen,
4953                                                          tp->ucopy.iov, chunk,
4954                                                          tp->ucopy.pinned_list);
4955 
4956                 if (dma_cookie < 0)
4957                         goto out;
4958 
4959                 tp->ucopy.dma_cookie = dma_cookie;
4960                 copied_early = true;
4961 
4962                 tp->ucopy.len -= chunk;
4963                 tp->copied_seq += chunk;
4964                 tcp_rcv_space_adjust(sk);
4965 
4966                 if ((tp->ucopy.len == 0) ||
4967                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4968                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4969                         tp->ucopy.wakeup = 1;
4970                         sk->sk_data_ready(sk, 0);
4971                 }
4972         } else if (chunk > 0) {
4973                 tp->ucopy.wakeup = 1;
4974                 sk->sk_data_ready(sk, 0);
4975         }
4976 out:
4977         return copied_early;
4978 }
4979 #endif /* CONFIG_NET_DMA */
4980 
4981 /* Does PAWS and seqno based validation of an incoming segment, flags will
4982  * play significant role here.
4983  */
4984 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
4985                                   const struct tcphdr *th, int syn_inerr)
4986 {
4987         struct tcp_sock *tp = tcp_sk(sk);
4988 
4989         /* RFC1323: H1. Apply PAWS check first. */
4990         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4991             tcp_paws_discard(sk, skb)) {
4992                 if (!th->rst) {
4993                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
4994                         tcp_send_dupack(sk, skb);
4995                         goto discard;
4996                 }
4997                 /* Reset is accepted even if it did not pass PAWS. */
4998         }
4999 
5000         /* Step 1: check sequence number */
5001         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5002                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5003                  * (RST) segments are validated by checking their SEQ-fields."
5004                  * And page 69: "If an incoming segment is not acceptable,
5005                  * an acknowledgment should be sent in reply (unless the RST
5006                  * bit is set, if so drop the segment and return)".
5007                  */
5008                 if (!th->rst) {
5009                         if (th->syn)
5010                                 goto syn_challenge;
5011                         tcp_send_dupack(sk, skb);
5012                 }
5013                 goto discard;
5014         }
5015 
5016         /* Step 2: check RST bit */
5017         if (th->rst) {
5018                 /* RFC 5961 3.2 :
5019                  * If sequence number exactly matches RCV.NXT, then
5020                  *     RESET the connection
5021                  * else
5022                  *     Send a challenge ACK
5023                  */
5024                 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5025                         tcp_reset(sk);
5026                 else
5027                         tcp_send_challenge_ack(sk);
5028                 goto discard;
5029         }
5030 
5031         /* step 3: check security and precedence [ignored] */
5032 
5033         /* step 4: Check for a SYN
5034          * RFC 5691 4.2 : Send a challenge ack
5035          */
5036         if (th->syn) {
5037 syn_challenge:
5038                 if (syn_inerr)
5039                         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5040                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5041                 tcp_send_challenge_ack(sk);
5042                 goto discard;
5043         }
5044 
5045         return true;
5046 
5047 discard:
5048         __kfree_skb(skb);
5049         return false;
5050 }
5051 
5052 /*
5053  *      TCP receive function for the ESTABLISHED state.
5054  *
5055  *      It is split into a fast path and a slow path. The fast path is
5056  *      disabled when:
5057  *      - A zero window was announced from us - zero window probing
5058  *        is only handled properly in the slow path.
5059  *      - Out of order segments arrived.
5060  *      - Urgent data is expected.
5061  *      - There is no buffer space left
5062  *      - Unexpected TCP flags/window values/header lengths are received
5063  *        (detected by checking the TCP header against pred_flags)
5064  *      - Data is sent in both directions. Fast path only supports pure senders
5065  *        or pure receivers (this means either the sequence number or the ack
5066  *        value must stay constant)
5067  *      - Unexpected TCP option.
5068  *
5069  *      When these conditions are not satisfied it drops into a standard
5070  *      receive procedure patterned after RFC793 to handle all cases.
5071  *      The first three cases are guaranteed by proper pred_flags setting,
5072  *      the rest is checked inline. Fast processing is turned on in
5073  *      tcp_data_queue when everything is OK.
5074  */
5075 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5076                          const struct tcphdr *th, unsigned int len)
5077 {
5078         struct tcp_sock *tp = tcp_sk(sk);
5079 
5080         if (unlikely(sk->sk_rx_dst == NULL))
5081                 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5082         /*
5083          *      Header prediction.
5084          *      The code loosely follows the one in the famous
5085          *      "30 instruction TCP receive" Van Jacobson mail.
5086          *
5087          *      Van's trick is to deposit buffers into socket queue
5088          *      on a device interrupt, to call tcp_recv function
5089          *      on the receive process context and checksum and copy
5090          *      the buffer to user space. smart...
5091          *
5092          *      Our current scheme is not silly either but we take the
5093          *      extra cost of the net_bh soft interrupt processing...
5094          *      We do checksum and copy also but from device to kernel.
5095          */
5096 
5097         tp->rx_opt.saw_tstamp = 0;
5098 
5099         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5100          *      if header_prediction is to be made
5101          *      'S' will always be tp->tcp_header_len >> 2
5102          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5103          *  turn it off (when there are holes in the receive
5104          *       space for instance)
5105          *      PSH flag is ignored.
5106          */
5107 
5108         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5109             TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5110             !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5111                 int tcp_header_len = tp->tcp_header_len;
5112 
5113                 /* Timestamp header prediction: tcp_header_len
5114                  * is automatically equal to th->doff*4 due to pred_flags
5115                  * match.
5116                  */
5117 
5118                 /* Check timestamp */
5119                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5120                         /* No? Slow path! */
5121                         if (!tcp_parse_aligned_timestamp(tp, th))
5122                                 goto slow_path;
5123 
5124                         /* If PAWS failed, check it more carefully in slow path */
5125                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5126                                 goto slow_path;
5127 
5128                         /* DO NOT update ts_recent here, if checksum fails
5129                          * and timestamp was corrupted part, it will result
5130                          * in a hung connection since we will drop all
5131                          * future packets due to the PAWS test.
5132                          */
5133                 }
5134 
5135                 if (len <= tcp_header_len) {
5136                         /* Bulk data transfer: sender */
5137                         if (len == tcp_header_len) {
5138                                 /* Predicted packet is in window by definition.
5139                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5140                                  * Hence, check seq<=rcv_wup reduces to:
5141                                  */
5142                                 if (tcp_header_len ==
5143                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5144                                     tp->rcv_nxt == tp->rcv_wup)
5145                                         tcp_store_ts_recent(tp);
5146 
5147                                 /* We know that such packets are checksummed
5148                                  * on entry.
5149                                  */
5150                                 tcp_ack(sk, skb, 0);
5151                                 __kfree_skb(skb);
5152                                 tcp_data_snd_check(sk);
5153                                 return;
5154                         } else { /* Header too small */
5155                                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5156                                 goto discard;
5157                         }
5158                 } else {
5159                         int eaten = 0;
5160                         int copied_early = 0;
5161                         bool fragstolen = false;
5162 
5163                         if (tp->copied_seq == tp->rcv_nxt &&
5164                             len - tcp_header_len <= tp->ucopy.len) {
5165 #ifdef CONFIG_NET_DMA
5166                                 if (tp->ucopy.task == current &&
5167                                     sock_owned_by_user(sk) &&
5168                                     tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5169                                         copied_early = 1;
5170                                         eaten = 1;
5171                                 }
5172 #endif
5173                                 if (tp->ucopy.task == current &&
5174                                     sock_owned_by_user(sk) && !copied_early) {
5175                                         __set_current_state(TASK_RUNNING);
5176 
5177                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5178                                                 eaten = 1;
5179                                 }
5180                                 if (eaten) {
5181                                         /* Predicted packet is in window by definition.
5182                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5183                                          * Hence, check seq<=rcv_wup reduces to:
5184                                          */
5185                                         if (tcp_header_len ==
5186                                             (sizeof(struct tcphdr) +
5187                                              TCPOLEN_TSTAMP_ALIGNED) &&
5188                                             tp->rcv_nxt == tp->rcv_wup)
5189                                                 tcp_store_ts_recent(tp);
5190 
5191                                         tcp_rcv_rtt_measure_ts(sk, skb);
5192 
5193                                         __skb_pull(skb, tcp_header_len);
5194                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5195                                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5196                                 }
5197                                 if (copied_early)
5198                                         tcp_cleanup_rbuf(sk, skb->len);
5199                         }
5200                         if (!eaten) {
5201                                 if (tcp_checksum_complete_user(sk, skb))
5202                                         goto csum_error;
5203 
5204                                 if ((int)skb->truesize > sk->sk_forward_alloc)
5205                                         goto step5;
5206 
5207                                 /* Predicted packet is in window by definition.
5208                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5209                                  * Hence, check seq<=rcv_wup reduces to:
5210                                  */
5211                                 if (tcp_header_len ==
5212                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5213                                     tp->rcv_nxt == tp->rcv_wup)
5214                                         tcp_store_ts_recent(tp);
5215 
5216                                 tcp_rcv_rtt_measure_ts(sk, skb);
5217 
5218                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5219 
5220                                 /* Bulk data transfer: receiver */
5221                                 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5222                                                       &fragstolen);
5223                         }
5224 
5225                         tcp_event_data_recv(sk, skb);
5226 
5227                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5228                                 /* Well, only one small jumplet in fast path... */
5229                                 tcp_ack(sk, skb, FLAG_DATA);
5230                                 tcp_data_snd_check(sk);
5231                                 if (!inet_csk_ack_scheduled(sk))
5232                                         goto no_ack;
5233                         }
5234 
5235                         if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5236                                 __tcp_ack_snd_check(sk, 0);
5237 no_ack:
5238 #ifdef CONFIG_NET_DMA
5239                         if (copied_early)
5240                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5241                         else
5242 #endif
5243                         if (eaten)
5244                                 kfree_skb_partial(skb, fragstolen);
5245                         sk->sk_data_ready(sk, 0);
5246                         return;
5247                 }
5248         }
5249 
5250 slow_path:
5251         if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5252                 goto csum_error;
5253 
5254         if (!th->ack && !th->rst)
5255                 goto discard;
5256 
5257         /*
5258          *      Standard slow path.
5259          */
5260 
5261         if (!tcp_validate_incoming(sk, skb, th, 1))
5262                 return;
5263 
5264 step5:
5265         if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5266                 goto discard;
5267 
5268         tcp_rcv_rtt_measure_ts(sk, skb);
5269 
5270         /* Process urgent data. */
5271         tcp_urg(sk, skb, th);
5272 
5273         /* step 7: process the segment text */
5274         tcp_data_queue(sk, skb);
5275 
5276         tcp_data_snd_check(sk);
5277         tcp_ack_snd_check(sk);
5278         return;
5279 
5280 csum_error:
5281         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5282         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5283 
5284 discard:
5285         __kfree_skb(skb);
5286 }
5287 EXPORT_SYMBOL(tcp_rcv_established);
5288 
5289 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5290 {
5291         struct tcp_sock *tp = tcp_sk(sk);
5292         struct inet_connection_sock *icsk = inet_csk(sk);
5293 
5294         tcp_set_state(sk, TCP_ESTABLISHED);
5295         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5296 
5297         if (skb != NULL) {
5298                 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5299                 security_inet_conn_established(sk, skb);
5300         }
5301 
5302         /* Make sure socket is routed, for correct metrics.  */
5303         icsk->icsk_af_ops->rebuild_header(sk);
5304 
5305         tcp_init_metrics(sk);
5306 
5307         tcp_init_congestion_control(sk);
5308 
5309         /* Prevent spurious tcp_cwnd_restart() on first data
5310          * packet.
5311          */
5312         tp->lsndtime = tcp_time_stamp;
5313 
5314         tcp_init_buffer_space(sk);
5315 
5316         if (sock_flag(sk, SOCK_KEEPOPEN))
5317                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5318 
5319         if (!tp->rx_opt.snd_wscale)
5320                 __tcp_fast_path_on(tp, tp->snd_wnd);
5321         else
5322                 tp->pred_flags = 0;
5323 
5324         if (!sock_flag(sk, SOCK_DEAD)) {
5325                 sk->sk_state_change(sk);
5326                 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5327         }
5328 }
5329 
5330 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5331                                     struct tcp_fastopen_cookie *cookie)
5332 {
5333         struct tcp_sock *tp = tcp_sk(sk);
5334         struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5335         u16 mss = tp->rx_opt.mss_clamp;
5336         bool syn_drop;
5337 
5338         if (mss == tp->rx_opt.user_mss) {
5339                 struct tcp_options_received opt;
5340 
5341                 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5342                 tcp_clear_options(&opt);
5343                 opt.user_mss = opt.mss_clamp = 0;
5344                 tcp_parse_options(synack, &opt, 0, NULL);
5345                 mss = opt.mss_clamp;
5346         }
5347 
5348         if (!tp->syn_fastopen)  /* Ignore an unsolicited cookie */
5349                 cookie->len = -1;
5350 
5351         /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5352          * the remote receives only the retransmitted (regular) SYNs: either
5353          * the original SYN-data or the corresponding SYN-ACK is lost.
5354          */
5355         syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
5356 
5357         tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
5358 
5359         if (data) { /* Retransmit unacked data in SYN */
5360                 tcp_for_write_queue_from(data, sk) {
5361                         if (data == tcp_send_head(sk) ||
5362                             __tcp_retransmit_skb(sk, data))
5363                                 break;
5364                 }
5365                 tcp_rearm_rto(sk);
5366                 return true;
5367         }
5368         tp->syn_data_acked = tp->syn_data;
5369         return false;
5370 }
5371 
5372 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5373                                          const struct tcphdr *th, unsigned int len)
5374 {
5375         struct inet_connection_sock *icsk = inet_csk(sk);
5376         struct tcp_sock *tp = tcp_sk(sk);
5377         struct tcp_fastopen_cookie foc = { .len = -1 };
5378         int saved_clamp = tp->rx_opt.mss_clamp;
5379 
5380         tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5381         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5382                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5383 
5384         if (th->ack) {
5385                 /* rfc793:
5386                  * "If the state is SYN-SENT then
5387                  *    first check the ACK bit
5388                  *      If the ACK bit is set
5389                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5390                  *        a reset (unless the RST bit is set, if so drop
5391                  *        the segment and return)"
5392                  */
5393                 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5394                     after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5395                         goto reset_and_undo;
5396 
5397                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5398                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5399                              tcp_time_stamp)) {
5400                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5401                         goto reset_and_undo;
5402                 }
5403 
5404                 /* Now ACK is acceptable.
5405                  *
5406                  * "If the RST bit is set
5407                  *    If the ACK was acceptable then signal the user "error:
5408                  *    connection reset", drop the segment, enter CLOSED state,
5409                  *    delete TCB, and return."
5410                  */
5411 
5412                 if (th->rst) {
5413                         tcp_reset(sk);
5414                         goto discard;
5415                 }
5416 
5417                 /* rfc793:
5418                  *   "fifth, if neither of the SYN or RST bits is set then
5419                  *    drop the segment and return."
5420                  *
5421                  *    See note below!
5422                  *                                        --ANK(990513)
5423                  */
5424                 if (!th->syn)
5425                         goto discard_and_undo;
5426 
5427                 /* rfc793:
5428                  *   "If the SYN bit is on ...
5429                  *    are acceptable then ...
5430                  *    (our SYN has been ACKed), change the connection
5431                  *    state to ESTABLISHED..."
5432                  */
5433 
5434                 TCP_ECN_rcv_synack(tp, th);
5435 
5436                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5437                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5438 
5439                 /* Ok.. it's good. Set up sequence numbers and
5440                  * move to established.
5441                  */
5442                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5443                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5444 
5445                 /* RFC1323: The window in SYN & SYN/ACK segments is
5446                  * never scaled.
5447                  */
5448                 tp->snd_wnd = ntohs(th->window);
5449 
5450                 if (!tp->rx_opt.wscale_ok) {
5451                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5452                         tp->window_clamp = min(tp->window_clamp, 65535U);
5453                 }
5454 
5455                 if (tp->rx_opt.saw_tstamp) {
5456                         tp->rx_opt.tstamp_ok       = 1;
5457                         tp->tcp_header_len =
5458                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5459                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5460                         tcp_store_ts_recent(tp);
5461                 } else {
5462                         tp->tcp_header_len = sizeof(struct tcphdr);
5463                 }
5464 
5465                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5466                         tcp_enable_fack(tp);
5467 
5468                 tcp_mtup_init(sk);
5469                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5470                 tcp_initialize_rcv_mss(sk);
5471 
5472                 /* Remember, tcp_poll() does not lock socket!
5473                  * Change state from SYN-SENT only after copied_seq
5474                  * is initialized. */
5475                 tp->copied_seq = tp->rcv_nxt;
5476 
5477                 smp_mb();
5478 
5479                 tcp_finish_connect(sk, skb);
5480 
5481                 if ((tp->syn_fastopen || tp->syn_data) &&
5482                     tcp_rcv_fastopen_synack(sk, skb, &foc))
5483                         return -1;
5484 
5485                 if (sk->sk_write_pending ||
5486                     icsk->icsk_accept_queue.rskq_defer_accept ||
5487                     icsk->icsk_ack.pingpong) {
5488                         /* Save one ACK. Data will be ready after
5489                          * several ticks, if write_pending is set.
5490                          *
5491                          * It may be deleted, but with this feature tcpdumps
5492                          * look so _wonderfully_ clever, that I was not able
5493                          * to stand against the temptation 8)     --ANK
5494                          */
5495                         inet_csk_schedule_ack(sk);
5496                         tcp_enter_quickack_mode(sk);
5497                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5498                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5499 
5500 discard:
5501                         __kfree_skb(skb);
5502                         return 0;
5503                 } else {
5504                         tcp_send_ack(sk);
5505                 }
5506                 return -1;
5507         }
5508 
5509         /* No ACK in the segment */
5510 
5511         if (th->rst) {
5512                 /* rfc793:
5513                  * "If the RST bit is set
5514                  *
5515                  *      Otherwise (no ACK) drop the segment and return."
5516                  */
5517 
5518                 goto discard_and_undo;
5519         }
5520 
5521         /* PAWS check. */
5522         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5523             tcp_paws_reject(&tp->rx_opt, 0))
5524                 goto discard_and_undo;
5525 
5526         if (th->syn) {
5527                 /* We see SYN without ACK. It is attempt of
5528                  * simultaneous connect with crossed SYNs.
5529                  * Particularly, it can be connect to self.
5530                  */
5531                 tcp_set_state(sk, TCP_SYN_RECV);
5532 
5533                 if (tp->rx_opt.saw_tstamp) {
5534                         tp->rx_opt.tstamp_ok = 1;
5535                         tcp_store_ts_recent(tp);
5536                         tp->tcp_header_len =
5537                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5538                 } else {
5539                         tp->tcp_header_len = sizeof(struct tcphdr);
5540                 }
5541 
5542                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5543                 tp->copied_seq = tp->rcv_nxt;
5544                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5545 
5546                 /* RFC1323: The window in SYN & SYN/ACK segments is
5547                  * never scaled.
5548                  */
5549                 tp->snd_wnd    = ntohs(th->window);
5550                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5551                 tp->max_window = tp->snd_wnd;
5552 
5553                 TCP_ECN_rcv_syn(tp, th);
5554 
5555                 tcp_mtup_init(sk);
5556                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5557                 tcp_initialize_rcv_mss(sk);
5558 
5559                 tcp_send_synack(sk);
5560 #if 0
5561                 /* Note, we could accept data and URG from this segment.
5562                  * There are no obstacles to make this (except that we must
5563                  * either change tcp_recvmsg() to prevent it from returning data
5564                  * before 3WHS completes per RFC793, or employ TCP Fast Open).
5565                  *
5566                  * However, if we ignore data in ACKless segments sometimes,
5567                  * we have no reasons to accept it sometimes.
5568                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5569                  * is not flawless. So, discard packet for sanity.
5570                  * Uncomment this return to process the data.
5571                  */
5572                 return -1;
5573 #else
5574                 goto discard;
5575 #endif
5576         }
5577         /* "fifth, if neither of the SYN or RST bits is set then
5578          * drop the segment and return."
5579          */
5580 
5581 discard_and_undo:
5582         tcp_clear_options(&tp->rx_opt);
5583         tp->rx_opt.mss_clamp = saved_clamp;
5584         goto discard;
5585 
5586 reset_and_undo:
5587         tcp_clear_options(&tp->rx_opt);
5588         tp->rx_opt.mss_clamp = saved_clamp;
5589         return 1;
5590 }
5591 
5592 /*
5593  *      This function implements the receiving procedure of RFC 793 for
5594  *      all states except ESTABLISHED and TIME_WAIT.
5595  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5596  *      address independent.
5597  */
5598 
5599 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5600                           const struct tcphdr *th, unsigned int len)
5601 {
5602         struct tcp_sock *tp = tcp_sk(sk);
5603         struct inet_connection_sock *icsk = inet_csk(sk);
5604         struct request_sock *req;
5605         int queued = 0;
5606         bool acceptable;
5607         u32 synack_stamp;
5608 
5609         tp->rx_opt.saw_tstamp = 0;
5610 
5611         switch (sk->sk_state) {
5612         case TCP_CLOSE:
5613                 goto discard;
5614 
5615         case TCP_LISTEN:
5616                 if (th->ack)
5617                         return 1;
5618 
5619                 if (th->rst)
5620                         goto discard;
5621 
5622                 if (th->syn) {
5623                         if (th->fin)
5624                                 goto discard;
5625                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5626                                 return 1;
5627 
5628                         /* Now we have several options: In theory there is
5629                          * nothing else in the frame. KA9Q has an option to
5630                          * send data with the syn, BSD accepts data with the
5631                          * syn up to the [to be] advertised window and
5632                          * Solaris 2.1 gives you a protocol error. For now
5633                          * we just ignore it, that fits the spec precisely
5634                          * and avoids incompatibilities. It would be nice in
5635                          * future to drop through and process the data.
5636                          *
5637                          * Now that TTCP is starting to be used we ought to
5638                          * queue this data.
5639                          * But, this leaves one open to an easy denial of
5640                          * service attack, and SYN cookies can't defend
5641                          * against this problem. So, we drop the data
5642                          * in the interest of security over speed unless
5643                          * it's still in use.
5644                          */
5645                         kfree_skb(skb);
5646                         return 0;
5647                 }
5648                 goto discard;
5649 
5650         case TCP_SYN_SENT:
5651                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5652                 if (queued >= 0)
5653                         return queued;
5654 
5655                 /* Do step6 onward by hand. */
5656                 tcp_urg(sk, skb, th);
5657                 __kfree_skb(skb);
5658                 tcp_data_snd_check(sk);
5659                 return 0;
5660         }
5661 
5662         req = tp->fastopen_rsk;
5663         if (req != NULL) {
5664                 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5665                     sk->sk_state != TCP_FIN_WAIT1);
5666 
5667                 if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
5668                         goto discard;
5669         }
5670 
5671         if (!th->ack && !th->rst)
5672                 goto discard;
5673 
5674         if (!tcp_validate_incoming(sk, skb, th, 0))
5675                 return 0;
5676 
5677         /* step 5: check the ACK field */
5678         acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5679                                       FLAG_UPDATE_TS_RECENT) > 0;
5680 
5681         switch (sk->sk_state) {
5682         case TCP_SYN_RECV:
5683                 if (!acceptable)
5684                         return 1;
5685 
5686                 /* Once we leave TCP_SYN_RECV, we no longer need req
5687                  * so release it.
5688                  */
5689                 if (req) {
5690                         synack_stamp = tcp_rsk(req)->snt_synack;
5691                         tp->total_retrans = req->num_retrans;
5692                         reqsk_fastopen_remove(sk, req, false);
5693                 } else {
5694                         synack_stamp = tp->lsndtime;
5695                         /* Make sure socket is routed, for correct metrics. */
5696                         icsk->icsk_af_ops->rebuild_header(sk);
5697                         tcp_init_congestion_control(sk);
5698 
5699                         tcp_mtup_init(sk);
5700                         tcp_init_buffer_space(sk);
5701                         tp->copied_seq = tp->rcv_nxt;
5702                 }
5703                 smp_mb();
5704                 tcp_set_state(sk, TCP_ESTABLISHED);
5705                 sk->sk_state_change(sk);
5706 
5707                 /* Note, that this wakeup is only for marginal crossed SYN case.
5708                  * Passively open sockets are not waked up, because
5709                  * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5710                  */
5711                 if (sk->sk_socket)
5712                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5713 
5714                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5715                 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5716                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5717                 tcp_synack_rtt_meas(sk, synack_stamp);
5718 
5719                 if (tp->rx_opt.tstamp_ok)
5720                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5721 
5722                 if (req) {
5723                         /* Re-arm the timer because data may have been sent out.
5724                          * This is similar to the regular data transmission case
5725                          * when new data has just been ack'ed.
5726                          *
5727                          * (TFO) - we could try to be more aggressive and
5728                          * retransmitting any data sooner based on when they
5729                          * are sent out.
5730                          */
5731                         tcp_rearm_rto(sk);
5732                 } else
5733                         tcp_init_metrics(sk);
5734 
5735                 tcp_update_pacing_rate(sk);
5736 
5737                 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5738                 tp->lsndtime = tcp_time_stamp;
5739 
5740                 tcp_initialize_rcv_mss(sk);
5741                 tcp_fast_path_on(tp);
5742                 break;
5743 
5744         case TCP_FIN_WAIT1: {
5745                 struct dst_entry *dst;
5746                 int tmo;
5747 
5748                 /* If we enter the TCP_FIN_WAIT1 state and we are a
5749                  * Fast Open socket and this is the first acceptable
5750                  * ACK we have received, this would have acknowledged
5751                  * our SYNACK so stop the SYNACK timer.
5752                  */
5753                 if (req != NULL) {
5754                         /* Return RST if ack_seq is invalid.
5755                          * Note that RFC793 only says to generate a
5756                          * DUPACK for it but for TCP Fast Open it seems
5757                          * better to treat this case like TCP_SYN_RECV
5758                          * above.
5759                          */
5760                         if (!acceptable)
5761                                 return 1;
5762                         /* We no longer need the request sock. */
5763                         reqsk_fastopen_remove(sk, req, false);
5764                         tcp_rearm_rto(sk);
5765                 }
5766                 if (tp->snd_una != tp->write_seq)
5767                         break;
5768 
5769                 tcp_set_state(sk, TCP_FIN_WAIT2);
5770                 sk->sk_shutdown |= SEND_SHUTDOWN;
5771 
5772                 dst = __sk_dst_get(sk);
5773                 if (dst)
5774                         dst_confirm(dst);
5775 
5776                 if (!sock_flag(sk, SOCK_DEAD)) {
5777                         /* Wake up lingering close() */
5778                         sk->sk_state_change(sk);
5779                         break;
5780                 }
5781 
5782                 if (tp->linger2 < 0 ||
5783                     (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5784                      after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5785                         tcp_done(sk);
5786                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5787                         return 1;
5788                 }
5789 
5790                 tmo = tcp_fin_time(sk);
5791                 if (tmo > TCP_TIMEWAIT_LEN) {
5792                         inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5793                 } else if (th->fin || sock_owned_by_user(sk)) {
5794                         /* Bad case. We could lose such FIN otherwise.
5795                          * It is not a big problem, but it looks confusing
5796                          * and not so rare event. We still can lose it now,
5797                          * if it spins in bh_lock_sock(), but it is really
5798                          * marginal case.
5799                          */
5800                         inet_csk_reset_keepalive_timer(sk, tmo);
5801                 } else {
5802                         tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5803                         goto discard;
5804                 }
5805                 break;
5806         }
5807 
5808         case TCP_CLOSING:
5809                 if (tp->snd_una == tp->write_seq) {
5810                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5811                         goto discard;
5812                 }
5813                 break;
5814 
5815         case TCP_LAST_ACK:
5816                 if (tp->snd_una == tp->write_seq) {
5817                         tcp_update_metrics(sk);
5818                         tcp_done(sk);
5819                         goto discard;
5820                 }
5821                 break;
5822         }
5823 
5824         /* step 6: check the URG bit */
5825         tcp_urg(sk, skb, th);
5826 
5827         /* step 7: process the segment text */
5828         switch (sk->sk_state) {
5829         case TCP_CLOSE_WAIT:
5830         case TCP_CLOSING:
5831         case TCP_LAST_ACK:
5832                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5833                         break;
5834         case TCP_FIN_WAIT1:
5835         case TCP_FIN_WAIT2:
5836                 /* RFC 793 says to queue data in these states,
5837                  * RFC 1122 says we MUST send a reset.
5838                  * BSD 4.4 also does reset.
5839                  */
5840                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5841                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5842                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5843                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5844                                 tcp_reset(sk);
5845                                 return 1;
5846                         }
5847                 }
5848                 /* Fall through */
5849         case TCP_ESTABLISHED:
5850                 tcp_data_queue(sk, skb);
5851                 queued = 1;
5852                 break;
5853         }
5854 
5855         /* tcp_data could move socket to TIME-WAIT */
5856         if (sk->sk_state != TCP_CLOSE) {
5857                 tcp_data_snd_check(sk);
5858                 tcp_ack_snd_check(sk);
5859         }
5860 
5861         if (!queued) {
5862 discard:
5863                 __kfree_skb(skb);
5864         }
5865         return 0;
5866 }
5867 EXPORT_SYMBOL(tcp_rcv_state_process);
5868 

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

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp