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

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

Version: ~ [ linux-5.6-rc3 ] ~ [ linux-5.5.6 ] ~ [ linux-5.4.22 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.106 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.171 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.214 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.214 ] ~ [ 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-3.16.82 ] ~ [ 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 <linux/reciprocal_div.h>
 72 #include <net/dst.h>
 73 #include <net/tcp.h>
 74 #include <net/inet_common.h>
 75 #include <linux/ipsec.h>
 76 #include <asm/unaligned.h>
 77 #include <net/netdma.h>
 78 
 79 int sysctl_tcp_timestamps __read_mostly = 1;
 80 int sysctl_tcp_window_scaling __read_mostly = 1;
 81 int sysctl_tcp_sack __read_mostly = 1;
 82 int sysctl_tcp_fack __read_mostly = 1;
 83 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
 84 EXPORT_SYMBOL(sysctl_tcp_reordering);
 85 int sysctl_tcp_dsack __read_mostly = 1;
 86 int sysctl_tcp_app_win __read_mostly = 31;
 87 int sysctl_tcp_adv_win_scale __read_mostly = 1;
 88 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
 89 
 90 /* rfc5961 challenge ack rate limiting */
 91 int sysctl_tcp_challenge_ack_limit = 1000;
 92 
 93 int sysctl_tcp_stdurg __read_mostly;
 94 int sysctl_tcp_rfc1337 __read_mostly;
 95 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
 96 int sysctl_tcp_frto __read_mostly = 2;
 97 
 98 int sysctl_tcp_thin_dupack __read_mostly;
 99 
100 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
101 int sysctl_tcp_early_retrans __read_mostly = 3;
102 
103 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
104 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
105 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
106 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
107 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
108 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
109 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
110 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
111 #define FLAG_ORIG_SACK_ACKED    0x200 /* Never retransmitted data are (s)acked  */
112 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
113 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
114 #define FLAG_SET_XMIT_TIMER     0x1000 /* Set TLP or RTO timer */
115 #define FLAG_SACK_RENEGING      0x2000 /* snd_una advanced to a sacked seq */
116 #define FLAG_UPDATE_TS_RECENT   0x4000 /* tcp_replace_ts_recent() */
117 
118 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
119 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
120 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
121 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
122 
123 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
124 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
125 
126 /* Adapt the MSS value used to make delayed ack decision to the
127  * real world.
128  */
129 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
130 {
131         struct inet_connection_sock *icsk = inet_csk(sk);
132         const unsigned int lss = icsk->icsk_ack.last_seg_size;
133         unsigned int len;
134 
135         icsk->icsk_ack.last_seg_size = 0;
136 
137         /* skb->len may jitter because of SACKs, even if peer
138          * sends good full-sized frames.
139          */
140         len = skb_shinfo(skb)->gso_size ? : skb->len;
141         if (len >= icsk->icsk_ack.rcv_mss) {
142                 icsk->icsk_ack.rcv_mss = len;
143         } else {
144                 /* Otherwise, we make more careful check taking into account,
145                  * that SACKs block is variable.
146                  *
147                  * "len" is invariant segment length, including TCP header.
148                  */
149                 len += skb->data - skb_transport_header(skb);
150                 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
151                     /* If PSH is not set, packet should be
152                      * full sized, provided peer TCP is not badly broken.
153                      * This observation (if it is correct 8)) allows
154                      * to handle super-low mtu links fairly.
155                      */
156                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
157                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
158                         /* Subtract also invariant (if peer is RFC compliant),
159                          * tcp header plus fixed timestamp option length.
160                          * Resulting "len" is MSS free of SACK jitter.
161                          */
162                         len -= tcp_sk(sk)->tcp_header_len;
163                         icsk->icsk_ack.last_seg_size = len;
164                         if (len == lss) {
165                                 icsk->icsk_ack.rcv_mss = len;
166                                 return;
167                         }
168                 }
169                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
170                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
171                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
172         }
173 }
174 
175 static void tcp_incr_quickack(struct sock *sk)
176 {
177         struct inet_connection_sock *icsk = inet_csk(sk);
178         unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
179 
180         if (quickacks == 0)
181                 quickacks = 2;
182         if (quickacks > icsk->icsk_ack.quick)
183                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
184 }
185 
186 static void tcp_enter_quickack_mode(struct sock *sk)
187 {
188         struct inet_connection_sock *icsk = inet_csk(sk);
189         tcp_incr_quickack(sk);
190         icsk->icsk_ack.pingpong = 0;
191         icsk->icsk_ack.ato = TCP_ATO_MIN;
192 }
193 
194 /* Send ACKs quickly, if "quick" count is not exhausted
195  * and the session is not interactive.
196  */
197 
198 static inline bool tcp_in_quickack_mode(const struct sock *sk)
199 {
200         const struct inet_connection_sock *icsk = inet_csk(sk);
201 
202         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
203 }
204 
205 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
206 {
207         if (tp->ecn_flags & TCP_ECN_OK)
208                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
209 }
210 
211 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
212 {
213         if (tcp_hdr(skb)->cwr)
214                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
215 }
216 
217 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
218 {
219         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
220 }
221 
222 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
223 {
224         if (!(tp->ecn_flags & TCP_ECN_OK))
225                 return;
226 
227         switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
228         case INET_ECN_NOT_ECT:
229                 /* Funny extension: if ECT is not set on a segment,
230                  * and we already seen ECT on a previous segment,
231                  * it is probably a retransmit.
232                  */
233                 if (tp->ecn_flags & TCP_ECN_SEEN)
234                         tcp_enter_quickack_mode((struct sock *)tp);
235                 break;
236         case INET_ECN_CE:
237                 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
238                         /* Better not delay acks, sender can have a very low cwnd */
239                         tcp_enter_quickack_mode((struct sock *)tp);
240                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
241                 }
242                 /* fallinto */
243         default:
244                 tp->ecn_flags |= TCP_ECN_SEEN;
245         }
246 }
247 
248 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
249 {
250         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
251                 tp->ecn_flags &= ~TCP_ECN_OK;
252 }
253 
254 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
255 {
256         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
257                 tp->ecn_flags &= ~TCP_ECN_OK;
258 }
259 
260 static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
261 {
262         if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
263                 return true;
264         return false;
265 }
266 
267 /* Buffer size and advertised window tuning.
268  *
269  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
270  */
271 
272 static void tcp_fixup_sndbuf(struct sock *sk)
273 {
274         int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
275 
276         sndmem *= TCP_INIT_CWND;
277         if (sk->sk_sndbuf < sndmem)
278                 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
279 }
280 
281 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
282  *
283  * All tcp_full_space() is split to two parts: "network" buffer, allocated
284  * forward and advertised in receiver window (tp->rcv_wnd) and
285  * "application buffer", required to isolate scheduling/application
286  * latencies from network.
287  * window_clamp is maximal advertised window. It can be less than
288  * tcp_full_space(), in this case tcp_full_space() - window_clamp
289  * is reserved for "application" buffer. The less window_clamp is
290  * the smoother our behaviour from viewpoint of network, but the lower
291  * throughput and the higher sensitivity of the connection to losses. 8)
292  *
293  * rcv_ssthresh is more strict window_clamp used at "slow start"
294  * phase to predict further behaviour of this connection.
295  * It is used for two goals:
296  * - to enforce header prediction at sender, even when application
297  *   requires some significant "application buffer". It is check #1.
298  * - to prevent pruning of receive queue because of misprediction
299  *   of receiver window. Check #2.
300  *
301  * The scheme does not work when sender sends good segments opening
302  * window and then starts to feed us spaghetti. But it should work
303  * in common situations. Otherwise, we have to rely on queue collapsing.
304  */
305 
306 /* Slow part of check#2. */
307 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
308 {
309         struct tcp_sock *tp = tcp_sk(sk);
310         /* Optimize this! */
311         int truesize = tcp_win_from_space(skb->truesize) >> 1;
312         int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
313 
314         while (tp->rcv_ssthresh <= window) {
315                 if (truesize <= skb->len)
316                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
317 
318                 truesize >>= 1;
319                 window >>= 1;
320         }
321         return 0;
322 }
323 
324 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
325 {
326         struct tcp_sock *tp = tcp_sk(sk);
327 
328         /* Check #1 */
329         if (tp->rcv_ssthresh < tp->window_clamp &&
330             (int)tp->rcv_ssthresh < tcp_space(sk) &&
331             !sk_under_memory_pressure(sk)) {
332                 int incr;
333 
334                 /* Check #2. Increase window, if skb with such overhead
335                  * will fit to rcvbuf in future.
336                  */
337                 if (tcp_win_from_space(skb->truesize) <= skb->len)
338                         incr = 2 * tp->advmss;
339                 else
340                         incr = __tcp_grow_window(sk, skb);
341 
342                 if (incr) {
343                         incr = max_t(int, incr, 2 * skb->len);
344                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
345                                                tp->window_clamp);
346                         inet_csk(sk)->icsk_ack.quick |= 1;
347                 }
348         }
349 }
350 
351 /* 3. Tuning rcvbuf, when connection enters established state. */
352 
353 static void tcp_fixup_rcvbuf(struct sock *sk)
354 {
355         u32 mss = tcp_sk(sk)->advmss;
356         u32 icwnd = TCP_DEFAULT_INIT_RCVWND;
357         int rcvmem;
358 
359         /* Limit to 10 segments if mss <= 1460,
360          * or 14600/mss segments, with a minimum of two segments.
361          */
362         if (mss > 1460)
363                 icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
364 
365         rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
366         while (tcp_win_from_space(rcvmem) < mss)
367                 rcvmem += 128;
368 
369         rcvmem *= icwnd;
370 
371         if (sk->sk_rcvbuf < rcvmem)
372                 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
373 }
374 
375 /* 4. Try to fixup all. It is made immediately after connection enters
376  *    established state.
377  */
378 void tcp_init_buffer_space(struct sock *sk)
379 {
380         struct tcp_sock *tp = tcp_sk(sk);
381         int maxwin;
382 
383         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
384                 tcp_fixup_rcvbuf(sk);
385         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
386                 tcp_fixup_sndbuf(sk);
387 
388         tp->rcvq_space.space = tp->rcv_wnd;
389 
390         maxwin = tcp_full_space(sk);
391 
392         if (tp->window_clamp >= maxwin) {
393                 tp->window_clamp = maxwin;
394 
395                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
396                         tp->window_clamp = max(maxwin -
397                                                (maxwin >> sysctl_tcp_app_win),
398                                                4 * tp->advmss);
399         }
400 
401         /* Force reservation of one segment. */
402         if (sysctl_tcp_app_win &&
403             tp->window_clamp > 2 * tp->advmss &&
404             tp->window_clamp + tp->advmss > maxwin)
405                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
406 
407         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
408         tp->snd_cwnd_stamp = tcp_time_stamp;
409 }
410 
411 /* 5. Recalculate window clamp after socket hit its memory bounds. */
412 static void tcp_clamp_window(struct sock *sk)
413 {
414         struct tcp_sock *tp = tcp_sk(sk);
415         struct inet_connection_sock *icsk = inet_csk(sk);
416 
417         icsk->icsk_ack.quick = 0;
418 
419         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
420             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
421             !sk_under_memory_pressure(sk) &&
422             sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
423                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
424                                     sysctl_tcp_rmem[2]);
425         }
426         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
427                 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
428 }
429 
430 /* Initialize RCV_MSS value.
431  * RCV_MSS is an our guess about MSS used by the peer.
432  * We haven't any direct information about the MSS.
433  * It's better to underestimate the RCV_MSS rather than overestimate.
434  * Overestimations make us ACKing less frequently than needed.
435  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
436  */
437 void tcp_initialize_rcv_mss(struct sock *sk)
438 {
439         const struct tcp_sock *tp = tcp_sk(sk);
440         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
441 
442         hint = min(hint, tp->rcv_wnd / 2);
443         hint = min(hint, TCP_MSS_DEFAULT);
444         hint = max(hint, TCP_MIN_MSS);
445 
446         inet_csk(sk)->icsk_ack.rcv_mss = hint;
447 }
448 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
449 
450 /* Receiver "autotuning" code.
451  *
452  * The algorithm for RTT estimation w/o timestamps is based on
453  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
454  * <http://public.lanl.gov/radiant/pubs.html#DRS>
455  *
456  * More detail on this code can be found at
457  * <http://staff.psc.edu/jheffner/>,
458  * though this reference is out of date.  A new paper
459  * is pending.
460  */
461 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
462 {
463         u32 new_sample = tp->rcv_rtt_est.rtt;
464         long m = sample;
465 
466         if (m == 0)
467                 m = 1;
468 
469         if (new_sample != 0) {
470                 /* If we sample in larger samples in the non-timestamp
471                  * case, we could grossly overestimate the RTT especially
472                  * with chatty applications or bulk transfer apps which
473                  * are stalled on filesystem I/O.
474                  *
475                  * Also, since we are only going for a minimum in the
476                  * non-timestamp case, we do not smooth things out
477                  * else with timestamps disabled convergence takes too
478                  * long.
479                  */
480                 if (!win_dep) {
481                         m -= (new_sample >> 3);
482                         new_sample += m;
483                 } else {
484                         m <<= 3;
485                         if (m < new_sample)
486                                 new_sample = m;
487                 }
488         } else {
489                 /* No previous measure. */
490                 new_sample = m << 3;
491         }
492 
493         if (tp->rcv_rtt_est.rtt != new_sample)
494                 tp->rcv_rtt_est.rtt = new_sample;
495 }
496 
497 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
498 {
499         if (tp->rcv_rtt_est.time == 0)
500                 goto new_measure;
501         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
502                 return;
503         tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
504 
505 new_measure:
506         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
507         tp->rcv_rtt_est.time = tcp_time_stamp;
508 }
509 
510 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
511                                           const struct sk_buff *skb)
512 {
513         struct tcp_sock *tp = tcp_sk(sk);
514         if (tp->rx_opt.rcv_tsecr &&
515             (TCP_SKB_CB(skb)->end_seq -
516              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
517                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
518 }
519 
520 /*
521  * This function should be called every time data is copied to user space.
522  * It calculates the appropriate TCP receive buffer space.
523  */
524 void tcp_rcv_space_adjust(struct sock *sk)
525 {
526         struct tcp_sock *tp = tcp_sk(sk);
527         int time;
528         int space;
529 
530         if (tp->rcvq_space.time == 0)
531                 goto new_measure;
532 
533         time = tcp_time_stamp - tp->rcvq_space.time;
534         if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
535                 return;
536 
537         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
538 
539         space = max(tp->rcvq_space.space, space);
540 
541         if (tp->rcvq_space.space != space) {
542                 int rcvmem;
543 
544                 tp->rcvq_space.space = space;
545 
546                 if (sysctl_tcp_moderate_rcvbuf &&
547                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
548                         int new_clamp = space;
549 
550                         /* Receive space grows, normalize in order to
551                          * take into account packet headers and sk_buff
552                          * structure overhead.
553                          */
554                         space /= tp->advmss;
555                         if (!space)
556                                 space = 1;
557                         rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
558                         while (tcp_win_from_space(rcvmem) < tp->advmss)
559                                 rcvmem += 128;
560                         space *= rcvmem;
561                         space = min(space, sysctl_tcp_rmem[2]);
562                         if (space > sk->sk_rcvbuf) {
563                                 sk->sk_rcvbuf = space;
564 
565                                 /* Make the window clamp follow along.  */
566                                 tp->window_clamp = new_clamp;
567                         }
568                 }
569         }
570 
571 new_measure:
572         tp->rcvq_space.seq = tp->copied_seq;
573         tp->rcvq_space.time = tcp_time_stamp;
574 }
575 
576 /* There is something which you must keep in mind when you analyze the
577  * behavior of the tp->ato delayed ack timeout interval.  When a
578  * connection starts up, we want to ack as quickly as possible.  The
579  * problem is that "good" TCP's do slow start at the beginning of data
580  * transmission.  The means that until we send the first few ACK's the
581  * sender will sit on his end and only queue most of his data, because
582  * he can only send snd_cwnd unacked packets at any given time.  For
583  * each ACK we send, he increments snd_cwnd and transmits more of his
584  * queue.  -DaveM
585  */
586 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
587 {
588         struct tcp_sock *tp = tcp_sk(sk);
589         struct inet_connection_sock *icsk = inet_csk(sk);
590         u32 now;
591 
592         inet_csk_schedule_ack(sk);
593 
594         tcp_measure_rcv_mss(sk, skb);
595 
596         tcp_rcv_rtt_measure(tp);
597 
598         now = tcp_time_stamp;
599 
600         if (!icsk->icsk_ack.ato) {
601                 /* The _first_ data packet received, initialize
602                  * delayed ACK engine.
603                  */
604                 tcp_incr_quickack(sk);
605                 icsk->icsk_ack.ato = TCP_ATO_MIN;
606         } else {
607                 int m = now - icsk->icsk_ack.lrcvtime;
608 
609                 if (m <= TCP_ATO_MIN / 2) {
610                         /* The fastest case is the first. */
611                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
612                 } else if (m < icsk->icsk_ack.ato) {
613                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
614                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
615                                 icsk->icsk_ack.ato = icsk->icsk_rto;
616                 } else if (m > icsk->icsk_rto) {
617                         /* Too long gap. Apparently sender failed to
618                          * restart window, so that we send ACKs quickly.
619                          */
620                         tcp_incr_quickack(sk);
621                         sk_mem_reclaim(sk);
622                 }
623         }
624         icsk->icsk_ack.lrcvtime = now;
625 
626         TCP_ECN_check_ce(tp, skb);
627 
628         if (skb->len >= 128)
629                 tcp_grow_window(sk, skb);
630 }
631 
632 /* Called to compute a smoothed rtt estimate. The data fed to this
633  * routine either comes from timestamps, or from segments that were
634  * known _not_ to have been retransmitted [see Karn/Partridge
635  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
636  * piece by Van Jacobson.
637  * NOTE: the next three routines used to be one big routine.
638  * To save cycles in the RFC 1323 implementation it was better to break
639  * it up into three procedures. -- erics
640  */
641 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
642 {
643         struct tcp_sock *tp = tcp_sk(sk);
644         long m = mrtt; /* RTT */
645 
646         /*      The following amusing code comes from Jacobson's
647          *      article in SIGCOMM '88.  Note that rtt and mdev
648          *      are scaled versions of rtt and mean deviation.
649          *      This is designed to be as fast as possible
650          *      m stands for "measurement".
651          *
652          *      On a 1990 paper the rto value is changed to:
653          *      RTO = rtt + 4 * mdev
654          *
655          * Funny. This algorithm seems to be very broken.
656          * These formulae increase RTO, when it should be decreased, increase
657          * too slowly, when it should be increased quickly, decrease too quickly
658          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
659          * does not matter how to _calculate_ it. Seems, it was trap
660          * that VJ failed to avoid. 8)
661          */
662         if (m == 0)
663                 m = 1;
664         if (tp->srtt != 0) {
665                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
666                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
667                 if (m < 0) {
668                         m = -m;         /* m is now abs(error) */
669                         m -= (tp->mdev >> 2);   /* similar update on mdev */
670                         /* This is similar to one of Eifel findings.
671                          * Eifel blocks mdev updates when rtt decreases.
672                          * This solution is a bit different: we use finer gain
673                          * for mdev in this case (alpha*beta).
674                          * Like Eifel it also prevents growth of rto,
675                          * but also it limits too fast rto decreases,
676                          * happening in pure Eifel.
677                          */
678                         if (m > 0)
679                                 m >>= 3;
680                 } else {
681                         m -= (tp->mdev >> 2);   /* similar update on mdev */
682                 }
683                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
684                 if (tp->mdev > tp->mdev_max) {
685                         tp->mdev_max = tp->mdev;
686                         if (tp->mdev_max > tp->rttvar)
687                                 tp->rttvar = tp->mdev_max;
688                 }
689                 if (after(tp->snd_una, tp->rtt_seq)) {
690                         if (tp->mdev_max < tp->rttvar)
691                                 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
692                         tp->rtt_seq = tp->snd_nxt;
693                         tp->mdev_max = tcp_rto_min(sk);
694                 }
695         } else {
696                 /* no previous measure. */
697                 tp->srtt = m << 3;      /* take the measured time to be rtt */
698                 tp->mdev = m << 1;      /* make sure rto = 3*rtt */
699                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
700                 tp->rtt_seq = tp->snd_nxt;
701         }
702 }
703 
704 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
705  * Note: TCP stack does not yet implement pacing.
706  * FQ packet scheduler can be used to implement cheap but effective
707  * TCP pacing, to smooth the burst on large writes when packets
708  * in flight is significantly lower than cwnd (or rwin)
709  */
710 static void tcp_update_pacing_rate(struct sock *sk)
711 {
712         const struct tcp_sock *tp = tcp_sk(sk);
713         u64 rate;
714 
715         /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
716         rate = (u64)tp->mss_cache * 2 * (HZ << 3);
717 
718         rate *= max(tp->snd_cwnd, tp->packets_out);
719 
720         /* Correction for small srtt : minimum srtt being 8 (1 jiffy << 3),
721          * be conservative and assume srtt = 1 (125 us instead of 1.25 ms)
722          * We probably need usec resolution in the future.
723          * Note: This also takes care of possible srtt=0 case,
724          * when tcp_rtt_estimator() was not yet called.
725          */
726         if (tp->srtt > 8 + 2)
727                 do_div(rate, tp->srtt);
728 
729         sk->sk_pacing_rate = min_t(u64, rate, ~0U);
730 }
731 
732 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
733  * routine referred to above.
734  */
735 void tcp_set_rto(struct sock *sk)
736 {
737         const struct tcp_sock *tp = tcp_sk(sk);
738         /* Old crap is replaced with new one. 8)
739          *
740          * More seriously:
741          * 1. If rtt variance happened to be less 50msec, it is hallucination.
742          *    It cannot be less due to utterly erratic ACK generation made
743          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
744          *    to do with delayed acks, because at cwnd>2 true delack timeout
745          *    is invisible. Actually, Linux-2.4 also generates erratic
746          *    ACKs in some circumstances.
747          */
748         inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
749 
750         /* 2. Fixups made earlier cannot be right.
751          *    If we do not estimate RTO correctly without them,
752          *    all the algo is pure shit and should be replaced
753          *    with correct one. It is exactly, which we pretend to do.
754          */
755 
756         /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
757          * guarantees that rto is higher.
758          */
759         tcp_bound_rto(sk);
760 }
761 
762 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
763 {
764         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
765 
766         if (!cwnd)
767                 cwnd = TCP_INIT_CWND;
768         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
769 }
770 
771 /*
772  * Packet counting of FACK is based on in-order assumptions, therefore TCP
773  * disables it when reordering is detected
774  */
775 void tcp_disable_fack(struct tcp_sock *tp)
776 {
777         /* RFC3517 uses different metric in lost marker => reset on change */
778         if (tcp_is_fack(tp))
779                 tp->lost_skb_hint = NULL;
780         tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
781 }
782 
783 /* Take a notice that peer is sending D-SACKs */
784 static void tcp_dsack_seen(struct tcp_sock *tp)
785 {
786         tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
787 }
788 
789 static void tcp_update_reordering(struct sock *sk, const int metric,
790                                   const int ts)
791 {
792         struct tcp_sock *tp = tcp_sk(sk);
793         if (metric > tp->reordering) {
794                 int mib_idx;
795 
796                 tp->reordering = min(TCP_MAX_REORDERING, metric);
797 
798                 /* This exciting event is worth to be remembered. 8) */
799                 if (ts)
800                         mib_idx = LINUX_MIB_TCPTSREORDER;
801                 else if (tcp_is_reno(tp))
802                         mib_idx = LINUX_MIB_TCPRENOREORDER;
803                 else if (tcp_is_fack(tp))
804                         mib_idx = LINUX_MIB_TCPFACKREORDER;
805                 else
806                         mib_idx = LINUX_MIB_TCPSACKREORDER;
807 
808                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
809 #if FASTRETRANS_DEBUG > 1
810                 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
811                          tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
812                          tp->reordering,
813                          tp->fackets_out,
814                          tp->sacked_out,
815                          tp->undo_marker ? tp->undo_retrans : 0);
816 #endif
817                 tcp_disable_fack(tp);
818         }
819 
820         if (metric > 0)
821                 tcp_disable_early_retrans(tp);
822 }
823 
824 /* This must be called before lost_out is incremented */
825 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
826 {
827         if ((tp->retransmit_skb_hint == NULL) ||
828             before(TCP_SKB_CB(skb)->seq,
829                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
830                 tp->retransmit_skb_hint = skb;
831 
832         if (!tp->lost_out ||
833             after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
834                 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
835 }
836 
837 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
838 {
839         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
840                 tcp_verify_retransmit_hint(tp, skb);
841 
842                 tp->lost_out += tcp_skb_pcount(skb);
843                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
844         }
845 }
846 
847 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
848                                             struct sk_buff *skb)
849 {
850         tcp_verify_retransmit_hint(tp, skb);
851 
852         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
853                 tp->lost_out += tcp_skb_pcount(skb);
854                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
855         }
856 }
857 
858 /* This procedure tags the retransmission queue when SACKs arrive.
859  *
860  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
861  * Packets in queue with these bits set are counted in variables
862  * sacked_out, retrans_out and lost_out, correspondingly.
863  *
864  * Valid combinations are:
865  * Tag  InFlight        Description
866  * 0    1               - orig segment is in flight.
867  * S    0               - nothing flies, orig reached receiver.
868  * L    0               - nothing flies, orig lost by net.
869  * R    2               - both orig and retransmit are in flight.
870  * L|R  1               - orig is lost, retransmit is in flight.
871  * S|R  1               - orig reached receiver, retrans is still in flight.
872  * (L|S|R is logically valid, it could occur when L|R is sacked,
873  *  but it is equivalent to plain S and code short-curcuits it to S.
874  *  L|S is logically invalid, it would mean -1 packet in flight 8))
875  *
876  * These 6 states form finite state machine, controlled by the following events:
877  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
878  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
879  * 3. Loss detection event of two flavors:
880  *      A. Scoreboard estimator decided the packet is lost.
881  *         A'. Reno "three dupacks" marks head of queue lost.
882  *         A''. Its FACK modification, head until snd.fack is lost.
883  *      B. SACK arrives sacking SND.NXT at the moment, when the
884  *         segment was retransmitted.
885  * 4. D-SACK added new rule: D-SACK changes any tag to S.
886  *
887  * It is pleasant to note, that state diagram turns out to be commutative,
888  * so that we are allowed not to be bothered by order of our actions,
889  * when multiple events arrive simultaneously. (see the function below).
890  *
891  * Reordering detection.
892  * --------------------
893  * Reordering metric is maximal distance, which a packet can be displaced
894  * in packet stream. With SACKs we can estimate it:
895  *
896  * 1. SACK fills old hole and the corresponding segment was not
897  *    ever retransmitted -> reordering. Alas, we cannot use it
898  *    when segment was retransmitted.
899  * 2. The last flaw is solved with D-SACK. D-SACK arrives
900  *    for retransmitted and already SACKed segment -> reordering..
901  * Both of these heuristics are not used in Loss state, when we cannot
902  * account for retransmits accurately.
903  *
904  * SACK block validation.
905  * ----------------------
906  *
907  * SACK block range validation checks that the received SACK block fits to
908  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
909  * Note that SND.UNA is not included to the range though being valid because
910  * it means that the receiver is rather inconsistent with itself reporting
911  * SACK reneging when it should advance SND.UNA. Such SACK block this is
912  * perfectly valid, however, in light of RFC2018 which explicitly states
913  * that "SACK block MUST reflect the newest segment.  Even if the newest
914  * segment is going to be discarded ...", not that it looks very clever
915  * in case of head skb. Due to potentional receiver driven attacks, we
916  * choose to avoid immediate execution of a walk in write queue due to
917  * reneging and defer head skb's loss recovery to standard loss recovery
918  * procedure that will eventually trigger (nothing forbids us doing this).
919  *
920  * Implements also blockage to start_seq wrap-around. Problem lies in the
921  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
922  * there's no guarantee that it will be before snd_nxt (n). The problem
923  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
924  * wrap (s_w):
925  *
926  *         <- outs wnd ->                          <- wrapzone ->
927  *         u     e      n                         u_w   e_w  s n_w
928  *         |     |      |                          |     |   |  |
929  * |<------------+------+----- TCP seqno space --------------+---------->|
930  * ...-- <2^31 ->|                                           |<--------...
931  * ...---- >2^31 ------>|                                    |<--------...
932  *
933  * Current code wouldn't be vulnerable but it's better still to discard such
934  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
935  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
936  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
937  * equal to the ideal case (infinite seqno space without wrap caused issues).
938  *
939  * With D-SACK the lower bound is extended to cover sequence space below
940  * SND.UNA down to undo_marker, which is the last point of interest. Yet
941  * again, D-SACK block must not to go across snd_una (for the same reason as
942  * for the normal SACK blocks, explained above). But there all simplicity
943  * ends, TCP might receive valid D-SACKs below that. As long as they reside
944  * fully below undo_marker they do not affect behavior in anyway and can
945  * therefore be safely ignored. In rare cases (which are more or less
946  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
947  * fragmentation and packet reordering past skb's retransmission. To consider
948  * them correctly, the acceptable range must be extended even more though
949  * the exact amount is rather hard to quantify. However, tp->max_window can
950  * be used as an exaggerated estimate.
951  */
952 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
953                                    u32 start_seq, u32 end_seq)
954 {
955         /* Too far in future, or reversed (interpretation is ambiguous) */
956         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
957                 return false;
958 
959         /* Nasty start_seq wrap-around check (see comments above) */
960         if (!before(start_seq, tp->snd_nxt))
961                 return false;
962 
963         /* In outstanding window? ...This is valid exit for D-SACKs too.
964          * start_seq == snd_una is non-sensical (see comments above)
965          */
966         if (after(start_seq, tp->snd_una))
967                 return true;
968 
969         if (!is_dsack || !tp->undo_marker)
970                 return false;
971 
972         /* ...Then it's D-SACK, and must reside below snd_una completely */
973         if (after(end_seq, tp->snd_una))
974                 return false;
975 
976         if (!before(start_seq, tp->undo_marker))
977                 return true;
978 
979         /* Too old */
980         if (!after(end_seq, tp->undo_marker))
981                 return false;
982 
983         /* Undo_marker boundary crossing (overestimates a lot). Known already:
984          *   start_seq < undo_marker and end_seq >= undo_marker.
985          */
986         return !before(start_seq, end_seq - tp->max_window);
987 }
988 
989 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
990  * Event "B". Later note: FACK people cheated me again 8), we have to account
991  * for reordering! Ugly, but should help.
992  *
993  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
994  * less than what is now known to be received by the other end (derived from
995  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
996  * retransmitted skbs to avoid some costly processing per ACKs.
997  */
998 static void tcp_mark_lost_retrans(struct sock *sk)
999 {
1000         const struct inet_connection_sock *icsk = inet_csk(sk);
1001         struct tcp_sock *tp = tcp_sk(sk);
1002         struct sk_buff *skb;
1003         int cnt = 0;
1004         u32 new_low_seq = tp->snd_nxt;
1005         u32 received_upto = tcp_highest_sack_seq(tp);
1006 
1007         if (!tcp_is_fack(tp) || !tp->retrans_out ||
1008             !after(received_upto, tp->lost_retrans_low) ||
1009             icsk->icsk_ca_state != TCP_CA_Recovery)
1010                 return;
1011 
1012         tcp_for_write_queue(skb, sk) {
1013                 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1014 
1015                 if (skb == tcp_send_head(sk))
1016                         break;
1017                 if (cnt == tp->retrans_out)
1018                         break;
1019                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1020                         continue;
1021 
1022                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1023                         continue;
1024 
1025                 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1026                  * constraint here (see above) but figuring out that at
1027                  * least tp->reordering SACK blocks reside between ack_seq
1028                  * and received_upto is not easy task to do cheaply with
1029                  * the available datastructures.
1030                  *
1031                  * Whether FACK should check here for tp->reordering segs
1032                  * in-between one could argue for either way (it would be
1033                  * rather simple to implement as we could count fack_count
1034                  * during the walk and do tp->fackets_out - fack_count).
1035                  */
1036                 if (after(received_upto, ack_seq)) {
1037                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1038                         tp->retrans_out -= tcp_skb_pcount(skb);
1039 
1040                         tcp_skb_mark_lost_uncond_verify(tp, skb);
1041                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1042                 } else {
1043                         if (before(ack_seq, new_low_seq))
1044                                 new_low_seq = ack_seq;
1045                         cnt += tcp_skb_pcount(skb);
1046                 }
1047         }
1048 
1049         if (tp->retrans_out)
1050                 tp->lost_retrans_low = new_low_seq;
1051 }
1052 
1053 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1054                             struct tcp_sack_block_wire *sp, int num_sacks,
1055                             u32 prior_snd_una)
1056 {
1057         struct tcp_sock *tp = tcp_sk(sk);
1058         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1059         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1060         bool dup_sack = false;
1061 
1062         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1063                 dup_sack = true;
1064                 tcp_dsack_seen(tp);
1065                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1066         } else if (num_sacks > 1) {
1067                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1068                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1069 
1070                 if (!after(end_seq_0, end_seq_1) &&
1071                     !before(start_seq_0, start_seq_1)) {
1072                         dup_sack = true;
1073                         tcp_dsack_seen(tp);
1074                         NET_INC_STATS_BH(sock_net(sk),
1075                                         LINUX_MIB_TCPDSACKOFORECV);
1076                 }
1077         }
1078 
1079         /* D-SACK for already forgotten data... Do dumb counting. */
1080         if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1081             !after(end_seq_0, prior_snd_una) &&
1082             after(end_seq_0, tp->undo_marker))
1083                 tp->undo_retrans--;
1084 
1085         return dup_sack;
1086 }
1087 
1088 struct tcp_sacktag_state {
1089         int reord;
1090         int fack_count;
1091         int flag;
1092 };
1093 
1094 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1095  * the incoming SACK may not exactly match but we can find smaller MSS
1096  * aligned portion of it that matches. Therefore we might need to fragment
1097  * which may fail and creates some hassle (caller must handle error case
1098  * returns).
1099  *
1100  * FIXME: this could be merged to shift decision code
1101  */
1102 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1103                                   u32 start_seq, u32 end_seq)
1104 {
1105         int err;
1106         bool in_sack;
1107         unsigned int pkt_len;
1108         unsigned int mss;
1109 
1110         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1111                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1112 
1113         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1114             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1115                 mss = tcp_skb_mss(skb);
1116                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1117 
1118                 if (!in_sack) {
1119                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1120                         if (pkt_len < mss)
1121                                 pkt_len = mss;
1122                 } else {
1123                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1124                         if (pkt_len < mss)
1125                                 return -EINVAL;
1126                 }
1127 
1128                 /* Round if necessary so that SACKs cover only full MSSes
1129                  * and/or the remaining small portion (if present)
1130                  */
1131                 if (pkt_len > mss) {
1132                         unsigned int new_len = (pkt_len / mss) * mss;
1133                         if (!in_sack && new_len < pkt_len) {
1134                                 new_len += mss;
1135                                 if (new_len >= skb->len)
1136                                         return 0;
1137                         }
1138                         pkt_len = new_len;
1139                 }
1140                 err = tcp_fragment(sk, skb, pkt_len, mss);
1141                 if (err < 0)
1142                         return err;
1143         }
1144 
1145         return in_sack;
1146 }
1147 
1148 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1149 static u8 tcp_sacktag_one(struct sock *sk,
1150                           struct tcp_sacktag_state *state, u8 sacked,
1151                           u32 start_seq, u32 end_seq,
1152                           bool dup_sack, int pcount)
1153 {
1154         struct tcp_sock *tp = tcp_sk(sk);
1155         int fack_count = state->fack_count;
1156 
1157         /* Account D-SACK for retransmitted packet. */
1158         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1159                 if (tp->undo_marker && tp->undo_retrans > 0 &&
1160                     after(end_seq, tp->undo_marker))
1161                         tp->undo_retrans--;
1162                 if (sacked & TCPCB_SACKED_ACKED)
1163                         state->reord = min(fack_count, state->reord);
1164         }
1165 
1166         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1167         if (!after(end_seq, tp->snd_una))
1168                 return sacked;
1169 
1170         if (!(sacked & TCPCB_SACKED_ACKED)) {
1171                 if (sacked & TCPCB_SACKED_RETRANS) {
1172                         /* If the segment is not tagged as lost,
1173                          * we do not clear RETRANS, believing
1174                          * that retransmission is still in flight.
1175                          */
1176                         if (sacked & TCPCB_LOST) {
1177                                 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1178                                 tp->lost_out -= pcount;
1179                                 tp->retrans_out -= pcount;
1180                         }
1181                 } else {
1182                         if (!(sacked & TCPCB_RETRANS)) {
1183                                 /* New sack for not retransmitted frame,
1184                                  * which was in hole. It is reordering.
1185                                  */
1186                                 if (before(start_seq,
1187                                            tcp_highest_sack_seq(tp)))
1188                                         state->reord = min(fack_count,
1189                                                            state->reord);
1190                                 if (!after(end_seq, tp->high_seq))
1191                                         state->flag |= FLAG_ORIG_SACK_ACKED;
1192                         }
1193 
1194                         if (sacked & TCPCB_LOST) {
1195                                 sacked &= ~TCPCB_LOST;
1196                                 tp->lost_out -= pcount;
1197                         }
1198                 }
1199 
1200                 sacked |= TCPCB_SACKED_ACKED;
1201                 state->flag |= FLAG_DATA_SACKED;
1202                 tp->sacked_out += pcount;
1203 
1204                 fack_count += pcount;
1205 
1206                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1207                 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1208                     before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1209                         tp->lost_cnt_hint += pcount;
1210 
1211                 if (fack_count > tp->fackets_out)
1212                         tp->fackets_out = fack_count;
1213         }
1214 
1215         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1216          * frames and clear it. undo_retrans is decreased above, L|R frames
1217          * are accounted above as well.
1218          */
1219         if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1220                 sacked &= ~TCPCB_SACKED_RETRANS;
1221                 tp->retrans_out -= pcount;
1222         }
1223 
1224         return sacked;
1225 }
1226 
1227 /* Shift newly-SACKed bytes from this skb to the immediately previous
1228  * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1229  */
1230 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1231                             struct tcp_sacktag_state *state,
1232                             unsigned int pcount, int shifted, int mss,
1233                             bool dup_sack)
1234 {
1235         struct tcp_sock *tp = tcp_sk(sk);
1236         struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1237         u32 start_seq = TCP_SKB_CB(skb)->seq;   /* start of newly-SACKed */
1238         u32 end_seq = start_seq + shifted;      /* end of newly-SACKed */
1239 
1240         BUG_ON(!pcount);
1241 
1242         /* Adjust counters and hints for the newly sacked sequence
1243          * range but discard the return value since prev is already
1244          * marked. We must tag the range first because the seq
1245          * advancement below implicitly advances
1246          * tcp_highest_sack_seq() when skb is highest_sack.
1247          */
1248         tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1249                         start_seq, end_seq, dup_sack, pcount);
1250 
1251         if (skb == tp->lost_skb_hint)
1252                 tp->lost_cnt_hint += pcount;
1253 
1254         TCP_SKB_CB(prev)->end_seq += shifted;
1255         TCP_SKB_CB(skb)->seq += shifted;
1256 
1257         skb_shinfo(prev)->gso_segs += pcount;
1258         BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1259         skb_shinfo(skb)->gso_segs -= pcount;
1260 
1261         /* When we're adding to gso_segs == 1, gso_size will be zero,
1262          * in theory this shouldn't be necessary but as long as DSACK
1263          * code can come after this skb later on it's better to keep
1264          * setting gso_size to something.
1265          */
1266         if (!skb_shinfo(prev)->gso_size) {
1267                 skb_shinfo(prev)->gso_size = mss;
1268                 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1269         }
1270 
1271         /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1272         if (skb_shinfo(skb)->gso_segs <= 1) {
1273                 skb_shinfo(skb)->gso_size = 0;
1274                 skb_shinfo(skb)->gso_type = 0;
1275         }
1276 
1277         /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1278         TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1279 
1280         if (skb->len > 0) {
1281                 BUG_ON(!tcp_skb_pcount(skb));
1282                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1283                 return false;
1284         }
1285 
1286         /* Whole SKB was eaten :-) */
1287 
1288         if (skb == tp->retransmit_skb_hint)
1289                 tp->retransmit_skb_hint = prev;
1290         if (skb == tp->scoreboard_skb_hint)
1291                 tp->scoreboard_skb_hint = prev;
1292         if (skb == tp->lost_skb_hint) {
1293                 tp->lost_skb_hint = prev;
1294                 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1295         }
1296 
1297         TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1298         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1299                 TCP_SKB_CB(prev)->end_seq++;
1300 
1301         if (skb == tcp_highest_sack(sk))
1302                 tcp_advance_highest_sack(sk, skb);
1303 
1304         tcp_unlink_write_queue(skb, sk);
1305         sk_wmem_free_skb(sk, skb);
1306 
1307         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1308 
1309         return true;
1310 }
1311 
1312 /* I wish gso_size would have a bit more sane initialization than
1313  * something-or-zero which complicates things
1314  */
1315 static int tcp_skb_seglen(const struct sk_buff *skb)
1316 {
1317         return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1318 }
1319 
1320 /* Shifting pages past head area doesn't work */
1321 static int skb_can_shift(const struct sk_buff *skb)
1322 {
1323         return !skb_headlen(skb) && skb_is_nonlinear(skb);
1324 }
1325 
1326 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1327  * skb.
1328  */
1329 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1330                                           struct tcp_sacktag_state *state,
1331                                           u32 start_seq, u32 end_seq,
1332                                           bool dup_sack)
1333 {
1334         struct tcp_sock *tp = tcp_sk(sk);
1335         struct sk_buff *prev;
1336         int mss;
1337         int pcount = 0;
1338         int len;
1339         int in_sack;
1340 
1341         if (!sk_can_gso(sk))
1342                 goto fallback;
1343 
1344         /* Normally R but no L won't result in plain S */
1345         if (!dup_sack &&
1346             (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1347                 goto fallback;
1348         if (!skb_can_shift(skb))
1349                 goto fallback;
1350         /* This frame is about to be dropped (was ACKed). */
1351         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1352                 goto fallback;
1353 
1354         /* Can only happen with delayed DSACK + discard craziness */
1355         if (unlikely(skb == tcp_write_queue_head(sk)))
1356                 goto fallback;
1357         prev = tcp_write_queue_prev(sk, skb);
1358 
1359         if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1360                 goto fallback;
1361 
1362         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1363                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1364 
1365         if (in_sack) {
1366                 len = skb->len;
1367                 pcount = tcp_skb_pcount(skb);
1368                 mss = tcp_skb_seglen(skb);
1369 
1370                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1371                  * drop this restriction as unnecessary
1372                  */
1373                 if (mss != tcp_skb_seglen(prev))
1374                         goto fallback;
1375         } else {
1376                 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1377                         goto noop;
1378                 /* CHECKME: This is non-MSS split case only?, this will
1379                  * cause skipped skbs due to advancing loop btw, original
1380                  * has that feature too
1381                  */
1382                 if (tcp_skb_pcount(skb) <= 1)
1383                         goto noop;
1384 
1385                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1386                 if (!in_sack) {
1387                         /* TODO: head merge to next could be attempted here
1388                          * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1389                          * though it might not be worth of the additional hassle
1390                          *
1391                          * ...we can probably just fallback to what was done
1392                          * previously. We could try merging non-SACKed ones
1393                          * as well but it probably isn't going to buy off
1394                          * because later SACKs might again split them, and
1395                          * it would make skb timestamp tracking considerably
1396                          * harder problem.
1397                          */
1398                         goto fallback;
1399                 }
1400 
1401                 len = end_seq - TCP_SKB_CB(skb)->seq;
1402                 BUG_ON(len < 0);
1403                 BUG_ON(len > skb->len);
1404 
1405                 /* MSS boundaries should be honoured or else pcount will
1406                  * severely break even though it makes things bit trickier.
1407                  * Optimize common case to avoid most of the divides
1408                  */
1409                 mss = tcp_skb_mss(skb);
1410 
1411                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1412                  * drop this restriction as unnecessary
1413                  */
1414                 if (mss != tcp_skb_seglen(prev))
1415                         goto fallback;
1416 
1417                 if (len == mss) {
1418                         pcount = 1;
1419                 } else if (len < mss) {
1420                         goto noop;
1421                 } else {
1422                         pcount = len / mss;
1423                         len = pcount * mss;
1424                 }
1425         }
1426 
1427         /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1428         if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1429                 goto fallback;
1430 
1431         if (!skb_shift(prev, skb, len))
1432                 goto fallback;
1433         if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1434                 goto out;
1435 
1436         /* Hole filled allows collapsing with the next as well, this is very
1437          * useful when hole on every nth skb pattern happens
1438          */
1439         if (prev == tcp_write_queue_tail(sk))
1440                 goto out;
1441         skb = tcp_write_queue_next(sk, prev);
1442 
1443         if (!skb_can_shift(skb) ||
1444             (skb == tcp_send_head(sk)) ||
1445             ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1446             (mss != tcp_skb_seglen(skb)))
1447                 goto out;
1448 
1449         len = skb->len;
1450         if (skb_shift(prev, skb, len)) {
1451                 pcount += tcp_skb_pcount(skb);
1452                 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1453         }
1454 
1455 out:
1456         state->fack_count += pcount;
1457         return prev;
1458 
1459 noop:
1460         return skb;
1461 
1462 fallback:
1463         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1464         return NULL;
1465 }
1466 
1467 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1468                                         struct tcp_sack_block *next_dup,
1469                                         struct tcp_sacktag_state *state,
1470                                         u32 start_seq, u32 end_seq,
1471                                         bool dup_sack_in)
1472 {
1473         struct tcp_sock *tp = tcp_sk(sk);
1474         struct sk_buff *tmp;
1475 
1476         tcp_for_write_queue_from(skb, sk) {
1477                 int in_sack = 0;
1478                 bool dup_sack = dup_sack_in;
1479 
1480                 if (skb == tcp_send_head(sk))
1481                         break;
1482 
1483                 /* queue is in-order => we can short-circuit the walk early */
1484                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1485                         break;
1486 
1487                 if ((next_dup != NULL) &&
1488                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1489                         in_sack = tcp_match_skb_to_sack(sk, skb,
1490                                                         next_dup->start_seq,
1491                                                         next_dup->end_seq);
1492                         if (in_sack > 0)
1493                                 dup_sack = true;
1494                 }
1495 
1496                 /* skb reference here is a bit tricky to get right, since
1497                  * shifting can eat and free both this skb and the next,
1498                  * so not even _safe variant of the loop is enough.
1499                  */
1500                 if (in_sack <= 0) {
1501                         tmp = tcp_shift_skb_data(sk, skb, state,
1502                                                  start_seq, end_seq, dup_sack);
1503                         if (tmp != NULL) {
1504                                 if (tmp != skb) {
1505                                         skb = tmp;
1506                                         continue;
1507                                 }
1508 
1509                                 in_sack = 0;
1510                         } else {
1511                                 in_sack = tcp_match_skb_to_sack(sk, skb,
1512                                                                 start_seq,
1513                                                                 end_seq);
1514                         }
1515                 }
1516 
1517                 if (unlikely(in_sack < 0))
1518                         break;
1519 
1520                 if (in_sack) {
1521                         TCP_SKB_CB(skb)->sacked =
1522                                 tcp_sacktag_one(sk,
1523                                                 state,
1524                                                 TCP_SKB_CB(skb)->sacked,
1525                                                 TCP_SKB_CB(skb)->seq,
1526                                                 TCP_SKB_CB(skb)->end_seq,
1527                                                 dup_sack,
1528                                                 tcp_skb_pcount(skb));
1529 
1530                         if (!before(TCP_SKB_CB(skb)->seq,
1531                                     tcp_highest_sack_seq(tp)))
1532                                 tcp_advance_highest_sack(sk, skb);
1533                 }
1534 
1535                 state->fack_count += tcp_skb_pcount(skb);
1536         }
1537         return skb;
1538 }
1539 
1540 /* Avoid all extra work that is being done by sacktag while walking in
1541  * a normal way
1542  */
1543 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1544                                         struct tcp_sacktag_state *state,
1545                                         u32 skip_to_seq)
1546 {
1547         tcp_for_write_queue_from(skb, sk) {
1548                 if (skb == tcp_send_head(sk))
1549                         break;
1550 
1551                 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1552                         break;
1553 
1554                 state->fack_count += tcp_skb_pcount(skb);
1555         }
1556         return skb;
1557 }
1558 
1559 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1560                                                 struct sock *sk,
1561                                                 struct tcp_sack_block *next_dup,
1562                                                 struct tcp_sacktag_state *state,
1563                                                 u32 skip_to_seq)
1564 {
1565         if (next_dup == NULL)
1566                 return skb;
1567 
1568         if (before(next_dup->start_seq, skip_to_seq)) {
1569                 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1570                 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1571                                        next_dup->start_seq, next_dup->end_seq,
1572                                        1);
1573         }
1574 
1575         return skb;
1576 }
1577 
1578 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1579 {
1580         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1581 }
1582 
1583 static int
1584 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1585                         u32 prior_snd_una)
1586 {
1587         struct tcp_sock *tp = tcp_sk(sk);
1588         const unsigned char *ptr = (skb_transport_header(ack_skb) +
1589                                     TCP_SKB_CB(ack_skb)->sacked);
1590         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1591         struct tcp_sack_block sp[TCP_NUM_SACKS];
1592         struct tcp_sack_block *cache;
1593         struct tcp_sacktag_state state;
1594         struct sk_buff *skb;
1595         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1596         int used_sacks;
1597         bool found_dup_sack = false;
1598         int i, j;
1599         int first_sack_index;
1600 
1601         state.flag = 0;
1602         state.reord = tp->packets_out;
1603 
1604         if (!tp->sacked_out) {
1605                 if (WARN_ON(tp->fackets_out))
1606                         tp->fackets_out = 0;
1607                 tcp_highest_sack_reset(sk);
1608         }
1609 
1610         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1611                                          num_sacks, prior_snd_una);
1612         if (found_dup_sack)
1613                 state.flag |= FLAG_DSACKING_ACK;
1614 
1615         /* Eliminate too old ACKs, but take into
1616          * account more or less fresh ones, they can
1617          * contain valid SACK info.
1618          */
1619         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1620                 return 0;
1621 
1622         if (!tp->packets_out)
1623                 goto out;
1624 
1625         used_sacks = 0;
1626         first_sack_index = 0;
1627         for (i = 0; i < num_sacks; i++) {
1628                 bool dup_sack = !i && found_dup_sack;
1629 
1630                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1631                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1632 
1633                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1634                                             sp[used_sacks].start_seq,
1635                                             sp[used_sacks].end_seq)) {
1636                         int mib_idx;
1637 
1638                         if (dup_sack) {
1639                                 if (!tp->undo_marker)
1640                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1641                                 else
1642                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1643                         } else {
1644                                 /* Don't count olds caused by ACK reordering */
1645                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1646                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1647                                         continue;
1648                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1649                         }
1650 
1651                         NET_INC_STATS_BH(sock_net(sk), mib_idx);
1652                         if (i == 0)
1653                                 first_sack_index = -1;
1654                         continue;
1655                 }
1656 
1657                 /* Ignore very old stuff early */
1658                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1659                         continue;
1660 
1661                 used_sacks++;
1662         }
1663 
1664         /* order SACK blocks to allow in order walk of the retrans queue */
1665         for (i = used_sacks - 1; i > 0; i--) {
1666                 for (j = 0; j < i; j++) {
1667                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1668                                 swap(sp[j], sp[j + 1]);
1669 
1670                                 /* Track where the first SACK block goes to */
1671                                 if (j == first_sack_index)
1672                                         first_sack_index = j + 1;
1673                         }
1674                 }
1675         }
1676 
1677         skb = tcp_write_queue_head(sk);
1678         state.fack_count = 0;
1679         i = 0;
1680 
1681         if (!tp->sacked_out) {
1682                 /* It's already past, so skip checking against it */
1683                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1684         } else {
1685                 cache = tp->recv_sack_cache;
1686                 /* Skip empty blocks in at head of the cache */
1687                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1688                        !cache->end_seq)
1689                         cache++;
1690         }
1691 
1692         while (i < used_sacks) {
1693                 u32 start_seq = sp[i].start_seq;
1694                 u32 end_seq = sp[i].end_seq;
1695                 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1696                 struct tcp_sack_block *next_dup = NULL;
1697 
1698                 if (found_dup_sack && ((i + 1) == first_sack_index))
1699                         next_dup = &sp[i + 1];
1700 
1701                 /* Skip too early cached blocks */
1702                 while (tcp_sack_cache_ok(tp, cache) &&
1703                        !before(start_seq, cache->end_seq))
1704                         cache++;
1705 
1706                 /* Can skip some work by looking recv_sack_cache? */
1707                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1708                     after(end_seq, cache->start_seq)) {
1709 
1710                         /* Head todo? */
1711                         if (before(start_seq, cache->start_seq)) {
1712                                 skb = tcp_sacktag_skip(skb, sk, &state,
1713                                                        start_seq);
1714                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1715                                                        &state,
1716                                                        start_seq,
1717                                                        cache->start_seq,
1718                                                        dup_sack);
1719                         }
1720 
1721                         /* Rest of the block already fully processed? */
1722                         if (!after(end_seq, cache->end_seq))
1723                                 goto advance_sp;
1724 
1725                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1726                                                        &state,
1727                                                        cache->end_seq);
1728 
1729                         /* ...tail remains todo... */
1730                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1731                                 /* ...but better entrypoint exists! */
1732                                 skb = tcp_highest_sack(sk);
1733                                 if (skb == NULL)
1734                                         break;
1735                                 state.fack_count = tp->fackets_out;
1736                                 cache++;
1737                                 goto walk;
1738                         }
1739 
1740                         skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1741                         /* Check overlap against next cached too (past this one already) */
1742                         cache++;
1743                         continue;
1744                 }
1745 
1746                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1747                         skb = tcp_highest_sack(sk);
1748                         if (skb == NULL)
1749                                 break;
1750                         state.fack_count = tp->fackets_out;
1751                 }
1752                 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1753 
1754 walk:
1755                 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1756                                        start_seq, end_seq, dup_sack);
1757 
1758 advance_sp:
1759                 i++;
1760         }
1761 
1762         /* Clear the head of the cache sack blocks so we can skip it next time */
1763         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1764                 tp->recv_sack_cache[i].start_seq = 0;
1765                 tp->recv_sack_cache[i].end_seq = 0;
1766         }
1767         for (j = 0; j < used_sacks; j++)
1768                 tp->recv_sack_cache[i++] = sp[j];
1769 
1770         tcp_mark_lost_retrans(sk);
1771 
1772         tcp_verify_left_out(tp);
1773 
1774         if ((state.reord < tp->fackets_out) &&
1775             ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1776                 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1777 
1778 out:
1779 
1780 #if FASTRETRANS_DEBUG > 0
1781         WARN_ON((int)tp->sacked_out < 0);
1782         WARN_ON((int)tp->lost_out < 0);
1783         WARN_ON((int)tp->retrans_out < 0);
1784         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1785 #endif
1786         return state.flag;
1787 }
1788 
1789 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1790  * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1791  */
1792 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1793 {
1794         u32 holes;
1795 
1796         holes = max(tp->lost_out, 1U);
1797         holes = min(holes, tp->packets_out);
1798 
1799         if ((tp->sacked_out + holes) > tp->packets_out) {
1800                 tp->sacked_out = tp->packets_out - holes;
1801                 return true;
1802         }
1803         return false;
1804 }
1805 
1806 /* If we receive more dupacks than we expected counting segments
1807  * in assumption of absent reordering, interpret this as reordering.
1808  * The only another reason could be bug in receiver TCP.
1809  */
1810 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1811 {
1812         struct tcp_sock *tp = tcp_sk(sk);
1813         if (tcp_limit_reno_sacked(tp))
1814                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1815 }
1816 
1817 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1818 
1819 static void tcp_add_reno_sack(struct sock *sk)
1820 {
1821         struct tcp_sock *tp = tcp_sk(sk);
1822         tp->sacked_out++;
1823         tcp_check_reno_reordering(sk, 0);
1824         tcp_verify_left_out(tp);
1825 }
1826 
1827 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1828 
1829 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1830 {
1831         struct tcp_sock *tp = tcp_sk(sk);
1832 
1833         if (acked > 0) {
1834                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1835                 if (acked - 1 >= tp->sacked_out)
1836                         tp->sacked_out = 0;
1837                 else
1838                         tp->sacked_out -= acked - 1;
1839         }
1840         tcp_check_reno_reordering(sk, acked);
1841         tcp_verify_left_out(tp);
1842 }
1843 
1844 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1845 {
1846         tp->sacked_out = 0;
1847 }
1848 
1849 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1850 {
1851         tp->retrans_out = 0;
1852         tp->lost_out = 0;
1853 
1854         tp->undo_marker = 0;
1855         tp->undo_retrans = -1;
1856 }
1857 
1858 void tcp_clear_retrans(struct tcp_sock *tp)
1859 {
1860         tcp_clear_retrans_partial(tp);
1861 
1862         tp->fackets_out = 0;
1863         tp->sacked_out = 0;
1864 }
1865 
1866 /* Enter Loss state. If "how" is not zero, forget all SACK information
1867  * and reset tags completely, otherwise preserve SACKs. If receiver
1868  * dropped its ofo queue, we will know this due to reneging detection.
1869  */
1870 void tcp_enter_loss(struct sock *sk, int how)
1871 {
1872         const struct inet_connection_sock *icsk = inet_csk(sk);
1873         struct tcp_sock *tp = tcp_sk(sk);
1874         struct sk_buff *skb;
1875         bool new_recovery = false;
1876 
1877         /* Reduce ssthresh if it has not yet been made inside this window. */
1878         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1879             !after(tp->high_seq, tp->snd_una) ||
1880             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1881                 new_recovery = true;
1882                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1883                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1884                 tcp_ca_event(sk, CA_EVENT_LOSS);
1885         }
1886         tp->snd_cwnd       = 1;
1887         tp->snd_cwnd_cnt   = 0;
1888         tp->snd_cwnd_stamp = tcp_time_stamp;
1889 
1890         tcp_clear_retrans_partial(tp);
1891 
1892         if (tcp_is_reno(tp))
1893                 tcp_reset_reno_sack(tp);
1894 
1895         tp->undo_marker = tp->snd_una;
1896         if (how) {
1897                 tp->sacked_out = 0;
1898                 tp->fackets_out = 0;
1899         }
1900         tcp_clear_all_retrans_hints(tp);
1901 
1902         tcp_for_write_queue(skb, sk) {
1903                 if (skb == tcp_send_head(sk))
1904                         break;
1905 
1906                 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1907                         tp->undo_marker = 0;
1908                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1909                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1910                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1911                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1912                         tp->lost_out += tcp_skb_pcount(skb);
1913                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1914                 }
1915         }
1916         tcp_verify_left_out(tp);
1917 
1918         tp->reordering = min_t(unsigned int, tp->reordering,
1919                                sysctl_tcp_reordering);
1920         tcp_set_ca_state(sk, TCP_CA_Loss);
1921         tp->high_seq = tp->snd_nxt;
1922         TCP_ECN_queue_cwr(tp);
1923 
1924         /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1925          * loss recovery is underway except recurring timeout(s) on
1926          * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1927          */
1928         tp->frto = sysctl_tcp_frto &&
1929                    (new_recovery || icsk->icsk_retransmits) &&
1930                    !inet_csk(sk)->icsk_mtup.probe_size;
1931 }
1932 
1933 /* If ACK arrived pointing to a remembered SACK, it means that our
1934  * remembered SACKs do not reflect real state of receiver i.e.
1935  * receiver _host_ is heavily congested (or buggy).
1936  *
1937  * Do processing similar to RTO timeout.
1938  */
1939 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1940 {
1941         if (flag & FLAG_SACK_RENEGING) {
1942                 struct inet_connection_sock *icsk = inet_csk(sk);
1943                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1944 
1945                 tcp_enter_loss(sk, 1);
1946                 icsk->icsk_retransmits++;
1947                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1948                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1949                                           icsk->icsk_rto, TCP_RTO_MAX);
1950                 return true;
1951         }
1952         return false;
1953 }
1954 
1955 static inline int tcp_fackets_out(const struct tcp_sock *tp)
1956 {
1957         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
1958 }
1959 
1960 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1961  * counter when SACK is enabled (without SACK, sacked_out is used for
1962  * that purpose).
1963  *
1964  * Instead, with FACK TCP uses fackets_out that includes both SACKed
1965  * segments up to the highest received SACK block so far and holes in
1966  * between them.
1967  *
1968  * With reordering, holes may still be in flight, so RFC3517 recovery
1969  * uses pure sacked_out (total number of SACKed segments) even though
1970  * it violates the RFC that uses duplicate ACKs, often these are equal
1971  * but when e.g. out-of-window ACKs or packet duplication occurs,
1972  * they differ. Since neither occurs due to loss, TCP should really
1973  * ignore them.
1974  */
1975 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
1976 {
1977         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
1978 }
1979 
1980 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
1981 {
1982         struct tcp_sock *tp = tcp_sk(sk);
1983         unsigned long delay;
1984 
1985         /* Delay early retransmit and entering fast recovery for
1986          * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
1987          * available, or RTO is scheduled to fire first.
1988          */
1989         if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
1990             (flag & FLAG_ECE) || !tp->srtt)
1991                 return false;
1992 
1993         delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2));
1994         if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
1995                 return false;
1996 
1997         inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
1998                                   TCP_RTO_MAX);
1999         return true;
2000 }
2001 
2002 static inline int tcp_skb_timedout(const struct sock *sk,
2003                                    const struct sk_buff *skb)
2004 {
2005         return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2006 }
2007 
2008 static inline int tcp_head_timedout(const struct sock *sk)
2009 {
2010         const struct tcp_sock *tp = tcp_sk(sk);
2011 
2012         return tp->packets_out &&
2013                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2014 }
2015 
2016 /* Linux NewReno/SACK/FACK/ECN state machine.
2017  * --------------------------------------
2018  *
2019  * "Open"       Normal state, no dubious events, fast path.
2020  * "Disorder"   In all the respects it is "Open",
2021  *              but requires a bit more attention. It is entered when
2022  *              we see some SACKs or dupacks. It is split of "Open"
2023  *              mainly to move some processing from fast path to slow one.
2024  * "CWR"        CWND was reduced due to some Congestion Notification event.
2025  *              It can be ECN, ICMP source quench, local device congestion.
2026  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2027  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2028  *
2029  * tcp_fastretrans_alert() is entered:
2030  * - each incoming ACK, if state is not "Open"
2031  * - when arrived ACK is unusual, namely:
2032  *      * SACK
2033  *      * Duplicate ACK.
2034  *      * ECN ECE.
2035  *
2036  * Counting packets in flight is pretty simple.
2037  *
2038  *      in_flight = packets_out - left_out + retrans_out
2039  *
2040  *      packets_out is SND.NXT-SND.UNA counted in packets.
2041  *
2042  *      retrans_out is number of retransmitted segments.
2043  *
2044  *      left_out is number of segments left network, but not ACKed yet.
2045  *
2046  *              left_out = sacked_out + lost_out
2047  *
2048  *     sacked_out: Packets, which arrived to receiver out of order
2049  *                 and hence not ACKed. With SACKs this number is simply
2050  *                 amount of SACKed data. Even without SACKs
2051  *                 it is easy to give pretty reliable estimate of this number,
2052  *                 counting duplicate ACKs.
2053  *
2054  *       lost_out: Packets lost by network. TCP has no explicit
2055  *                 "loss notification" feedback from network (for now).
2056  *                 It means that this number can be only _guessed_.
2057  *                 Actually, it is the heuristics to predict lossage that
2058  *                 distinguishes different algorithms.
2059  *
2060  *      F.e. after RTO, when all the queue is considered as lost,
2061  *      lost_out = packets_out and in_flight = retrans_out.
2062  *
2063  *              Essentially, we have now two algorithms counting
2064  *              lost packets.
2065  *
2066  *              FACK: It is the simplest heuristics. As soon as we decided
2067  *              that something is lost, we decide that _all_ not SACKed
2068  *              packets until the most forward SACK are lost. I.e.
2069  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2070  *              It is absolutely correct estimate, if network does not reorder
2071  *              packets. And it loses any connection to reality when reordering
2072  *              takes place. We use FACK by default until reordering
2073  *              is suspected on the path to this destination.
2074  *
2075  *              NewReno: when Recovery is entered, we assume that one segment
2076  *              is lost (classic Reno). While we are in Recovery and
2077  *              a partial ACK arrives, we assume that one more packet
2078  *              is lost (NewReno). This heuristics are the same in NewReno
2079  *              and SACK.
2080  *
2081  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2082  *  deflation etc. CWND is real congestion window, never inflated, changes
2083  *  only according to classic VJ rules.
2084  *
2085  * Really tricky (and requiring careful tuning) part of algorithm
2086  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2087  * The first determines the moment _when_ we should reduce CWND and,
2088  * hence, slow down forward transmission. In fact, it determines the moment
2089  * when we decide that hole is caused by loss, rather than by a reorder.
2090  *
2091  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2092  * holes, caused by lost packets.
2093  *
2094  * And the most logically complicated part of algorithm is undo
2095  * heuristics. We detect false retransmits due to both too early
2096  * fast retransmit (reordering) and underestimated RTO, analyzing
2097  * timestamps and D-SACKs. When we detect that some segments were
2098  * retransmitted by mistake and CWND reduction was wrong, we undo
2099  * window reduction and abort recovery phase. This logic is hidden
2100  * inside several functions named tcp_try_undo_<something>.
2101  */
2102 
2103 /* This function decides, when we should leave Disordered state
2104  * and enter Recovery phase, reducing congestion window.
2105  *
2106  * Main question: may we further continue forward transmission
2107  * with the same cwnd?
2108  */
2109 static bool tcp_time_to_recover(struct sock *sk, int flag)
2110 {
2111         struct tcp_sock *tp = tcp_sk(sk);
2112         __u32 packets_out;
2113 
2114         /* Trick#1: The loss is proven. */
2115         if (tp->lost_out)
2116                 return true;
2117 
2118         /* Not-A-Trick#2 : Classic rule... */
2119         if (tcp_dupack_heuristics(tp) > tp->reordering)
2120                 return true;
2121 
2122         /* Trick#3 : when we use RFC2988 timer restart, fast
2123          * retransmit can be triggered by timeout of queue head.
2124          */
2125         if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2126                 return true;
2127 
2128         /* Trick#4: It is still not OK... But will it be useful to delay
2129          * recovery more?
2130          */
2131         packets_out = tp->packets_out;
2132         if (packets_out <= tp->reordering &&
2133             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2134             !tcp_may_send_now(sk)) {
2135                 /* We have nothing to send. This connection is limited
2136                  * either by receiver window or by application.
2137                  */
2138                 return true;
2139         }
2140 
2141         /* If a thin stream is detected, retransmit after first
2142          * received dupack. Employ only if SACK is supported in order
2143          * to avoid possible corner-case series of spurious retransmissions
2144          * Use only if there are no unsent data.
2145          */
2146         if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2147             tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2148             tcp_is_sack(tp) && !tcp_send_head(sk))
2149                 return true;
2150 
2151         /* Trick#6: TCP early retransmit, per RFC5827.  To avoid spurious
2152          * retransmissions due to small network reorderings, we implement
2153          * Mitigation A.3 in the RFC and delay the retransmission for a short
2154          * interval if appropriate.
2155          */
2156         if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2157             (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2158             !tcp_may_send_now(sk))
2159                 return !tcp_pause_early_retransmit(sk, flag);
2160 
2161         return false;
2162 }
2163 
2164 /* New heuristics: it is possible only after we switched to restart timer
2165  * each time when something is ACKed. Hence, we can detect timed out packets
2166  * during fast retransmit without falling to slow start.
2167  *
2168  * Usefulness of this as is very questionable, since we should know which of
2169  * the segments is the next to timeout which is relatively expensive to find
2170  * in general case unless we add some data structure just for that. The
2171  * current approach certainly won't find the right one too often and when it
2172  * finally does find _something_ it usually marks large part of the window
2173  * right away (because a retransmission with a larger timestamp blocks the
2174  * loop from advancing). -ij
2175  */
2176 static void tcp_timeout_skbs(struct sock *sk)
2177 {
2178         struct tcp_sock *tp = tcp_sk(sk);
2179         struct sk_buff *skb;
2180 
2181         if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2182                 return;
2183 
2184         skb = tp->scoreboard_skb_hint;
2185         if (tp->scoreboard_skb_hint == NULL)
2186                 skb = tcp_write_queue_head(sk);
2187 
2188         tcp_for_write_queue_from(skb, sk) {
2189                 if (skb == tcp_send_head(sk))
2190                         break;
2191                 if (!tcp_skb_timedout(sk, skb))
2192                         break;
2193 
2194                 tcp_skb_mark_lost(tp, skb);
2195         }
2196 
2197         tp->scoreboard_skb_hint = skb;
2198 
2199         tcp_verify_left_out(tp);
2200 }
2201 
2202 /* Detect loss in event "A" above by marking head of queue up as lost.
2203  * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2204  * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2205  * has at least tp->reordering SACKed seqments above it; "packets" refers to
2206  * the maximum SACKed segments to pass before reaching this limit.
2207  */
2208 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2209 {
2210         struct tcp_sock *tp = tcp_sk(sk);
2211         struct sk_buff *skb;
2212         int cnt, oldcnt;
2213         int err;
2214         unsigned int mss;
2215         /* Use SACK to deduce losses of new sequences sent during recovery */
2216         const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2217 
2218         WARN_ON(packets > tp->packets_out);
2219         if (tp->lost_skb_hint) {
2220                 skb = tp->lost_skb_hint;
2221                 cnt = tp->lost_cnt_hint;
2222                 /* Head already handled? */
2223                 if (mark_head && skb != tcp_write_queue_head(sk))
2224                         return;
2225         } else {
2226                 skb = tcp_write_queue_head(sk);
2227                 cnt = 0;
2228         }
2229 
2230         tcp_for_write_queue_from(skb, sk) {
2231                 if (skb == tcp_send_head(sk))
2232                         break;
2233                 /* TODO: do this better */
2234                 /* this is not the most efficient way to do this... */
2235                 tp->lost_skb_hint = skb;
2236                 tp->lost_cnt_hint = cnt;
2237 
2238                 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2239                         break;
2240 
2241                 oldcnt = cnt;
2242                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2243                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2244                         cnt += tcp_skb_pcount(skb);
2245 
2246                 if (cnt > packets) {
2247                         if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2248                             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2249                             (oldcnt >= packets))
2250                                 break;
2251 
2252                         mss = skb_shinfo(skb)->gso_size;
2253                         err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2254                         if (err < 0)
2255                                 break;
2256                         cnt = packets;
2257                 }
2258 
2259                 tcp_skb_mark_lost(tp, skb);
2260 
2261                 if (mark_head)
2262                         break;
2263         }
2264         tcp_verify_left_out(tp);
2265 }
2266 
2267 /* Account newly detected lost packet(s) */
2268 
2269 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2270 {
2271         struct tcp_sock *tp = tcp_sk(sk);
2272 
2273         if (tcp_is_reno(tp)) {
2274                 tcp_mark_head_lost(sk, 1, 1);
2275         } else if (tcp_is_fack(tp)) {
2276                 int lost = tp->fackets_out - tp->reordering;
2277                 if (lost <= 0)
2278                         lost = 1;
2279                 tcp_mark_head_lost(sk, lost, 0);
2280         } else {
2281                 int sacked_upto = tp->sacked_out - tp->reordering;
2282                 if (sacked_upto >= 0)
2283                         tcp_mark_head_lost(sk, sacked_upto, 0);
2284                 else if (fast_rexmit)
2285                         tcp_mark_head_lost(sk, 1, 1);
2286         }
2287 
2288         tcp_timeout_skbs(sk);
2289 }
2290 
2291 /* CWND moderation, preventing bursts due to too big ACKs
2292  * in dubious situations.
2293  */
2294 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2295 {
2296         tp->snd_cwnd = min(tp->snd_cwnd,
2297                            tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2298         tp->snd_cwnd_stamp = tcp_time_stamp;
2299 }
2300 
2301 /* Nothing was retransmitted or returned timestamp is less
2302  * than timestamp of the first retransmission.
2303  */
2304 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2305 {
2306         return !tp->retrans_stamp ||
2307                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2308                  before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2309 }
2310 
2311 /* Undo procedures. */
2312 
2313 #if FASTRETRANS_DEBUG > 1
2314 static void DBGUNDO(struct sock *sk, const char *msg)
2315 {
2316         struct tcp_sock *tp = tcp_sk(sk);
2317         struct inet_sock *inet = inet_sk(sk);
2318 
2319         if (sk->sk_family == AF_INET) {
2320                 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2321                          msg,
2322                          &inet->inet_daddr, ntohs(inet->inet_dport),
2323                          tp->snd_cwnd, tcp_left_out(tp),
2324                          tp->snd_ssthresh, tp->prior_ssthresh,
2325                          tp->packets_out);
2326         }
2327 #if IS_ENABLED(CONFIG_IPV6)
2328         else if (sk->sk_family == AF_INET6) {
2329                 struct ipv6_pinfo *np = inet6_sk(sk);
2330                 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2331                          msg,
2332                          &np->daddr, ntohs(inet->inet_dport),
2333                          tp->snd_cwnd, tcp_left_out(tp),
2334                          tp->snd_ssthresh, tp->prior_ssthresh,
2335                          tp->packets_out);
2336         }
2337 #endif
2338 }
2339 #else
2340 #define DBGUNDO(x...) do { } while (0)
2341 #endif
2342 
2343 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2344 {
2345         struct tcp_sock *tp = tcp_sk(sk);
2346 
2347         if (tp->prior_ssthresh) {
2348                 const struct inet_connection_sock *icsk = inet_csk(sk);
2349 
2350                 if (icsk->icsk_ca_ops->undo_cwnd)
2351                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2352                 else
2353                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2354 
2355                 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2356                         tp->snd_ssthresh = tp->prior_ssthresh;
2357                         TCP_ECN_withdraw_cwr(tp);
2358                 }
2359         } else {
2360                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2361         }
2362         tp->snd_cwnd_stamp = tcp_time_stamp;
2363 }
2364 
2365 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2366 {
2367         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2368 }
2369 
2370 /* People celebrate: "We love our President!" */
2371 static bool tcp_try_undo_recovery(struct sock *sk)
2372 {
2373         struct tcp_sock *tp = tcp_sk(sk);
2374 
2375         if (tcp_may_undo(tp)) {
2376                 int mib_idx;
2377 
2378                 /* Happy end! We did not retransmit anything
2379                  * or our original transmission succeeded.
2380                  */
2381                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2382                 tcp_undo_cwr(sk, true);
2383                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2384                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2385                 else
2386                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2387 
2388                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2389                 tp->undo_marker = 0;
2390         }
2391         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2392                 /* Hold old state until something *above* high_seq
2393                  * is ACKed. For Reno it is MUST to prevent false
2394                  * fast retransmits (RFC2582). SACK TCP is safe. */
2395                 tcp_moderate_cwnd(tp);
2396                 return true;
2397         }
2398         tcp_set_ca_state(sk, TCP_CA_Open);
2399         return false;
2400 }
2401 
2402 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2403 static void tcp_try_undo_dsack(struct sock *sk)
2404 {
2405         struct tcp_sock *tp = tcp_sk(sk);
2406 
2407         if (tp->undo_marker && !tp->undo_retrans) {
2408                 DBGUNDO(sk, "D-SACK");
2409                 tcp_undo_cwr(sk, true);
2410                 tp->undo_marker = 0;
2411                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2412         }
2413 }
2414 
2415 /* We can clear retrans_stamp when there are no retransmissions in the
2416  * window. It would seem that it is trivially available for us in
2417  * tp->retrans_out, however, that kind of assumptions doesn't consider
2418  * what will happen if errors occur when sending retransmission for the
2419  * second time. ...It could the that such segment has only
2420  * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2421  * the head skb is enough except for some reneging corner cases that
2422  * are not worth the effort.
2423  *
2424  * Main reason for all this complexity is the fact that connection dying
2425  * time now depends on the validity of the retrans_stamp, in particular,
2426  * that successive retransmissions of a segment must not advance
2427  * retrans_stamp under any conditions.
2428  */
2429 static bool tcp_any_retrans_done(const struct sock *sk)
2430 {
2431         const struct tcp_sock *tp = tcp_sk(sk);
2432         struct sk_buff *skb;
2433 
2434         if (tp->retrans_out)
2435                 return true;
2436 
2437         skb = tcp_write_queue_head(sk);
2438         if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2439                 return true;
2440 
2441         return false;
2442 }
2443 
2444 /* Undo during fast recovery after partial ACK. */
2445 
2446 static int tcp_try_undo_partial(struct sock *sk, int acked)
2447 {
2448         struct tcp_sock *tp = tcp_sk(sk);
2449         /* Partial ACK arrived. Force Hoe's retransmit. */
2450         int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2451 
2452         if (tcp_may_undo(tp)) {
2453                 /* Plain luck! Hole if filled with delayed
2454                  * packet, rather than with a retransmit.
2455                  */
2456                 if (!tcp_any_retrans_done(sk))
2457                         tp->retrans_stamp = 0;
2458 
2459                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2460 
2461                 DBGUNDO(sk, "Hoe");
2462                 tcp_undo_cwr(sk, false);
2463                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2464 
2465                 /* So... Do not make Hoe's retransmit yet.
2466                  * If the first packet was delayed, the rest
2467                  * ones are most probably delayed as well.
2468                  */
2469                 failed = 0;
2470         }
2471         return failed;
2472 }
2473 
2474 /* Undo during loss recovery after partial ACK or using F-RTO. */
2475 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2476 {
2477         struct tcp_sock *tp = tcp_sk(sk);
2478 
2479         if (frto_undo || tcp_may_undo(tp)) {
2480                 struct sk_buff *skb;
2481                 tcp_for_write_queue(skb, sk) {
2482                         if (skb == tcp_send_head(sk))
2483                                 break;
2484                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2485                 }
2486 
2487                 tcp_clear_all_retrans_hints(tp);
2488 
2489                 DBGUNDO(sk, "partial loss");
2490                 tp->lost_out = 0;
2491                 tcp_undo_cwr(sk, true);
2492                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2493                 if (frto_undo)
2494                         NET_INC_STATS_BH(sock_net(sk),
2495                                          LINUX_MIB_TCPSPURIOUSRTOS);
2496                 inet_csk(sk)->icsk_retransmits = 0;
2497                 tp->undo_marker = 0;
2498                 if (frto_undo || tcp_is_sack(tp))
2499                         tcp_set_ca_state(sk, TCP_CA_Open);
2500                 return true;
2501         }
2502         return false;
2503 }
2504 
2505 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2506  * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2507  * It computes the number of packets to send (sndcnt) based on packets newly
2508  * delivered:
2509  *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2510  *      cwnd reductions across a full RTT.
2511  *   2) If packets in flight is lower than ssthresh (such as due to excess
2512  *      losses and/or application stalls), do not perform any further cwnd
2513  *      reductions, but instead slow start up to ssthresh.
2514  */
2515 static void tcp_init_cwnd_reduction(struct sock *sk, const bool set_ssthresh)
2516 {
2517         struct tcp_sock *tp = tcp_sk(sk);
2518 
2519         tp->high_seq = tp->snd_nxt;
2520         tp->tlp_high_seq = 0;
2521         tp->snd_cwnd_cnt = 0;
2522         tp->prior_cwnd = tp->snd_cwnd;
2523         tp->prr_delivered = 0;
2524         tp->prr_out = 0;
2525         if (set_ssthresh)
2526                 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2527         TCP_ECN_queue_cwr(tp);
2528 }
2529 
2530 static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked,
2531                                int fast_rexmit)
2532 {
2533         struct tcp_sock *tp = tcp_sk(sk);
2534         int sndcnt = 0;
2535         int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2536 
2537         tp->prr_delivered += newly_acked_sacked;
2538         if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2539                 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2540                                tp->prior_cwnd - 1;
2541                 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2542         } else {
2543                 sndcnt = min_t(int, delta,
2544                                max_t(int, tp->prr_delivered - tp->prr_out,
2545                                      newly_acked_sacked) + 1);
2546         }
2547 
2548         sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2549         tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2550 }
2551 
2552 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2553 {
2554         struct tcp_sock *tp = tcp_sk(sk);
2555 
2556         /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2557         if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2558             (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2559                 tp->snd_cwnd = tp->snd_ssthresh;
2560                 tp->snd_cwnd_stamp = tcp_time_stamp;
2561         }
2562         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2563 }
2564 
2565 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2566 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
2567 {
2568         struct tcp_sock *tp = tcp_sk(sk);
2569 
2570         tp->prior_ssthresh = 0;
2571         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2572                 tp->undo_marker = 0;
2573                 tcp_init_cwnd_reduction(sk, set_ssthresh);
2574                 tcp_set_ca_state(sk, TCP_CA_CWR);
2575         }
2576 }
2577 
2578 static void tcp_try_keep_open(struct sock *sk)
2579 {
2580         struct tcp_sock *tp = tcp_sk(sk);
2581         int state = TCP_CA_Open;
2582 
2583         if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2584                 state = TCP_CA_Disorder;
2585 
2586         if (inet_csk(sk)->icsk_ca_state != state) {
2587                 tcp_set_ca_state(sk, state);
2588                 tp->high_seq = tp->snd_nxt;
2589         }
2590 }
2591 
2592 static void tcp_try_to_open(struct sock *sk, int flag, int newly_acked_sacked)
2593 {
2594         struct tcp_sock *tp = tcp_sk(sk);
2595 
2596         tcp_verify_left_out(tp);
2597 
2598         if (!tcp_any_retrans_done(sk))
2599                 tp->retrans_stamp = 0;
2600 
2601         if (flag & FLAG_ECE)
2602                 tcp_enter_cwr(sk, 1);
2603 
2604         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2605                 tcp_try_keep_open(sk);
2606                 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2607                         tcp_moderate_cwnd(tp);
2608         } else {
2609                 tcp_cwnd_reduction(sk, newly_acked_sacked, 0);
2610         }
2611 }
2612 
2613 static void tcp_mtup_probe_failed(struct sock *sk)
2614 {
2615         struct inet_connection_sock *icsk = inet_csk(sk);
2616 
2617         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2618         icsk->icsk_mtup.probe_size = 0;
2619 }
2620 
2621 static void tcp_mtup_probe_success(struct sock *sk)
2622 {
2623         struct tcp_sock *tp = tcp_sk(sk);
2624         struct inet_connection_sock *icsk = inet_csk(sk);
2625 
2626         /* FIXME: breaks with very large cwnd */
2627         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2628         tp->snd_cwnd = tp->snd_cwnd *
2629                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2630                        icsk->icsk_mtup.probe_size;
2631         tp->snd_cwnd_cnt = 0;
2632         tp->snd_cwnd_stamp = tcp_time_stamp;
2633         tp->snd_ssthresh = tcp_current_ssthresh(sk);
2634 
2635         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2636         icsk->icsk_mtup.probe_size = 0;
2637         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2638 }
2639 
2640 /* Do a simple retransmit without using the backoff mechanisms in
2641  * tcp_timer. This is used for path mtu discovery.
2642  * The socket is already locked here.
2643  */
2644 void tcp_simple_retransmit(struct sock *sk)
2645 {
2646         const struct inet_connection_sock *icsk = inet_csk(sk);
2647         struct tcp_sock *tp = tcp_sk(sk);
2648         struct sk_buff *skb;
2649         unsigned int mss = tcp_current_mss(sk);
2650         u32 prior_lost = tp->lost_out;
2651 
2652         tcp_for_write_queue(skb, sk) {
2653                 if (skb == tcp_send_head(sk))
2654                         break;
2655                 if (tcp_skb_seglen(skb) > mss &&
2656                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2657                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2658                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2659                                 tp->retrans_out -= tcp_skb_pcount(skb);
2660                         }
2661                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2662                 }
2663         }
2664 
2665         tcp_clear_retrans_hints_partial(tp);
2666 
2667         if (prior_lost == tp->lost_out)
2668                 return;
2669 
2670         if (tcp_is_reno(tp))
2671                 tcp_limit_reno_sacked(tp);
2672 
2673         tcp_verify_left_out(tp);
2674 
2675         /* Don't muck with the congestion window here.
2676          * Reason is that we do not increase amount of _data_
2677          * in network, but units changed and effective
2678          * cwnd/ssthresh really reduced now.
2679          */
2680         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2681                 tp->high_seq = tp->snd_nxt;
2682                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2683                 tp->prior_ssthresh = 0;
2684                 tp->undo_marker = 0;
2685                 tcp_set_ca_state(sk, TCP_CA_Loss);
2686         }
2687         tcp_xmit_retransmit_queue(sk);
2688 }
2689 EXPORT_SYMBOL(tcp_simple_retransmit);
2690 
2691 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2692 {
2693         struct tcp_sock *tp = tcp_sk(sk);
2694         int mib_idx;
2695 
2696         if (tcp_is_reno(tp))
2697                 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2698         else
2699                 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2700 
2701         NET_INC_STATS_BH(sock_net(sk), mib_idx);
2702 
2703         tp->prior_ssthresh = 0;
2704         tp->undo_marker = tp->snd_una;
2705         tp->undo_retrans = tp->retrans_out ? : -1;
2706 
2707         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2708                 if (!ece_ack)
2709                         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2710                 tcp_init_cwnd_reduction(sk, true);
2711         }
2712         tcp_set_ca_state(sk, TCP_CA_Recovery);
2713 }
2714 
2715 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2716  * recovered or spurious. Otherwise retransmits more on partial ACKs.
2717  */
2718 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2719 {
2720         struct inet_connection_sock *icsk = inet_csk(sk);
2721         struct tcp_sock *tp = tcp_sk(sk);
2722         bool recovered = !before(tp->snd_una, tp->high_seq);
2723 
2724         if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2725                 /* Step 3.b. A timeout is spurious if not all data are
2726                  * lost, i.e., never-retransmitted data are (s)acked.
2727                  */
2728                 if (tcp_try_undo_loss(sk, flag & FLAG_ORIG_SACK_ACKED))
2729                         return;
2730 
2731                 if (after(tp->snd_nxt, tp->high_seq) &&
2732                     (flag & FLAG_DATA_SACKED || is_dupack)) {
2733                         tp->frto = 0; /* Loss was real: 2nd part of step 3.a */
2734                 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2735                         tp->high_seq = tp->snd_nxt;
2736                         __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2737                                                   TCP_NAGLE_OFF);
2738                         if (after(tp->snd_nxt, tp->high_seq))
2739                                 return; /* Step 2.b */
2740                         tp->frto = 0;
2741                 }
2742         }
2743 
2744         if (recovered) {
2745                 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2746                 icsk->icsk_retransmits = 0;
2747                 tcp_try_undo_recovery(sk);
2748                 return;
2749         }
2750         if (flag & FLAG_DATA_ACKED)
2751                 icsk->icsk_retransmits = 0;
2752         if (tcp_is_reno(tp)) {
2753                 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2754                  * delivered. Lower inflight to clock out (re)tranmissions.
2755                  */
2756                 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2757                         tcp_add_reno_sack(sk);
2758                 else if (flag & FLAG_SND_UNA_ADVANCED)
2759                         tcp_reset_reno_sack(tp);
2760         }
2761         if (tcp_try_undo_loss(sk, false))
2762                 return;
2763         tcp_xmit_retransmit_queue(sk);
2764 }
2765 
2766 /* Process an event, which can update packets-in-flight not trivially.
2767  * Main goal of this function is to calculate new estimate for left_out,
2768  * taking into account both packets sitting in receiver's buffer and
2769  * packets lost by network.
2770  *
2771  * Besides that it does CWND reduction, when packet loss is detected
2772  * and changes state of machine.
2773  *
2774  * It does _not_ decide what to send, it is made in function
2775  * tcp_xmit_retransmit_queue().
2776  */
2777 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
2778                                   int prior_sacked, int prior_packets,
2779                                   bool is_dupack, int flag)
2780 {
2781         struct inet_connection_sock *icsk = inet_csk(sk);
2782         struct tcp_sock *tp = tcp_sk(sk);
2783         int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2784                                     (tcp_fackets_out(tp) > tp->reordering));
2785         int newly_acked_sacked = 0;
2786         int fast_rexmit = 0;
2787 
2788         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2789                 tp->sacked_out = 0;
2790         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2791                 tp->fackets_out = 0;
2792 
2793         /* Now state machine starts.
2794          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2795         if (flag & FLAG_ECE)
2796                 tp->prior_ssthresh = 0;
2797 
2798         /* B. In all the states check for reneging SACKs. */
2799         if (tcp_check_sack_reneging(sk, flag))
2800                 return;
2801 
2802         /* C. Check consistency of the current state. */
2803         tcp_verify_left_out(tp);
2804 
2805         /* D. Check state exit conditions. State can be terminated
2806          *    when high_seq is ACKed. */
2807         if (icsk->icsk_ca_state == TCP_CA_Open) {
2808                 WARN_ON(tp->retrans_out != 0);
2809                 tp->retrans_stamp = 0;
2810         } else if (!before(tp->snd_una, tp->high_seq)) {
2811                 switch (icsk->icsk_ca_state) {
2812                 case TCP_CA_CWR:
2813                         /* CWR is to be held something *above* high_seq
2814                          * is ACKed for CWR bit to reach receiver. */
2815                         if (tp->snd_una != tp->high_seq) {
2816                                 tcp_end_cwnd_reduction(sk);
2817                                 tcp_set_ca_state(sk, TCP_CA_Open);
2818                         }
2819                         break;
2820 
2821                 case TCP_CA_Recovery:
2822                         if (tcp_is_reno(tp))
2823                                 tcp_reset_reno_sack(tp);
2824                         if (tcp_try_undo_recovery(sk))
2825                                 return;
2826                         tcp_end_cwnd_reduction(sk);
2827                         break;
2828                 }
2829         }
2830 
2831         /* E. Process state. */
2832         switch (icsk->icsk_ca_state) {
2833         case TCP_CA_Recovery:
2834                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2835                         if (tcp_is_reno(tp) && is_dupack)
2836                                 tcp_add_reno_sack(sk);
2837                 } else
2838                         do_lost = tcp_try_undo_partial(sk, pkts_acked);
2839                 newly_acked_sacked = prior_packets - tp->packets_out +
2840                                      tp->sacked_out - prior_sacked;
2841                 break;
2842         case TCP_CA_Loss:
2843                 tcp_process_loss(sk, flag, is_dupack);
2844                 if (icsk->icsk_ca_state != TCP_CA_Open)
2845                         return;
2846                 /* Fall through to processing in Open state. */
2847         default:
2848                 if (tcp_is_reno(tp)) {
2849                         if (flag & FLAG_SND_UNA_ADVANCED)
2850                                 tcp_reset_reno_sack(tp);
2851                         if (is_dupack)
2852                                 tcp_add_reno_sack(sk);
2853                 }
2854                 newly_acked_sacked = prior_packets - tp->packets_out +
2855                                      tp->sacked_out - prior_sacked;
2856 
2857                 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2858                         tcp_try_undo_dsack(sk);
2859 
2860                 if (!tcp_time_to_recover(sk, flag)) {
2861                         tcp_try_to_open(sk, flag, newly_acked_sacked);
2862                         return;
2863                 }
2864 
2865                 /* MTU probe failure: don't reduce cwnd */
2866                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2867                     icsk->icsk_mtup.probe_size &&
2868                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2869                         tcp_mtup_probe_failed(sk);
2870                         /* Restores the reduction we did in tcp_mtup_probe() */
2871                         tp->snd_cwnd++;
2872                         tcp_simple_retransmit(sk);
2873                         return;
2874                 }
2875 
2876                 /* Otherwise enter Recovery state */
2877                 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2878                 fast_rexmit = 1;
2879         }
2880 
2881         if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
2882                 tcp_update_scoreboard(sk, fast_rexmit);
2883         tcp_cwnd_reduction(sk, newly_acked_sacked, fast_rexmit);
2884         tcp_xmit_retransmit_queue(sk);
2885 }
2886 
2887 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
2888 {
2889         tcp_rtt_estimator(sk, seq_rtt);
2890         tcp_set_rto(sk);
2891         inet_csk(sk)->icsk_backoff = 0;
2892 }
2893 EXPORT_SYMBOL(tcp_valid_rtt_meas);
2894 
2895 /* Read draft-ietf-tcplw-high-performance before mucking
2896  * with this code. (Supersedes RFC1323)
2897  */
2898 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2899 {
2900         /* RTTM Rule: A TSecr value received in a segment is used to
2901          * update the averaged RTT measurement only if the segment
2902          * acknowledges some new data, i.e., only if it advances the
2903          * left edge of the send window.
2904          *
2905          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2906          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2907          *
2908          * Changed: reset backoff as soon as we see the first valid sample.
2909          * If we do not, we get strongly overestimated rto. With timestamps
2910          * samples are accepted even from very old segments: f.e., when rtt=1
2911          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2912          * answer arrives rto becomes 120 seconds! If at least one of segments
2913          * in window is lost... Voila.                          --ANK (010210)
2914          */
2915         struct tcp_sock *tp = tcp_sk(sk);
2916 
2917         tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2918 }
2919 
2920 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2921 {
2922         /* We don't have a timestamp. Can only use
2923          * packets that are not retransmitted to determine
2924          * rtt estimates. Also, we must not reset the
2925          * backoff for rto until we get a non-retransmitted
2926          * packet. This allows us to deal with a situation
2927          * where the network delay has increased suddenly.
2928          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2929          */
2930 
2931         if (flag & FLAG_RETRANS_DATA_ACKED)
2932                 return;
2933 
2934         tcp_valid_rtt_meas(sk, seq_rtt);
2935 }
2936 
2937 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2938                                       const s32 seq_rtt)
2939 {
2940         const struct tcp_sock *tp = tcp_sk(sk);
2941         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2942         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2943                 tcp_ack_saw_tstamp(sk, flag);
2944         else if (seq_rtt >= 0)
2945                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2946 }
2947 
2948 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
2949 {
2950         const struct inet_connection_sock *icsk = inet_csk(sk);
2951         icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
2952         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2953 }
2954 
2955 /* Restart timer after forward progress on connection.
2956  * RFC2988 recommends to restart timer to now+rto.
2957  */
2958 void tcp_rearm_rto(struct sock *sk)
2959 {
2960         const struct inet_connection_sock *icsk = inet_csk(sk);
2961         struct tcp_sock *tp = tcp_sk(sk);
2962 
2963         /* If the retrans timer is currently being used by Fast Open
2964          * for SYN-ACK retrans purpose, stay put.
2965          */
2966         if (tp->fastopen_rsk)
2967                 return;
2968 
2969         if (!tp->packets_out) {
2970                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2971         } else {
2972                 u32 rto = inet_csk(sk)->icsk_rto;
2973                 /* Offset the time elapsed after installing regular RTO */
2974                 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2975                     icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2976                         s32 delta = tcp_rto_delta(sk);
2977                         /* delta may not be positive if the socket is locked
2978                          * when the retrans timer fires and is rescheduled.
2979                          */
2980                         rto = max_t(int, delta, 1);
2981                 }
2982                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2983                                           TCP_RTO_MAX);
2984         }
2985 }
2986 
2987 /* This function is called when the delayed ER timer fires. TCP enters
2988  * fast recovery and performs fast-retransmit.
2989  */
2990 void tcp_resume_early_retransmit(struct sock *sk)
2991 {
2992         struct tcp_sock *tp = tcp_sk(sk);
2993 
2994         tcp_rearm_rto(sk);
2995 
2996         /* Stop if ER is disabled after the delayed ER timer is scheduled */
2997         if (!tp->do_early_retrans)
2998                 return;
2999 
3000         tcp_enter_recovery(sk, false);
3001         tcp_update_scoreboard(sk, 1);
3002         tcp_xmit_retransmit_queue(sk);
3003 }
3004 
3005 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3006 static void tcp_set_xmit_timer(struct sock *sk)
3007 {
3008         if (!tcp_schedule_loss_probe(sk))
3009                 tcp_rearm_rto(sk);
3010 }
3011 
3012 /* If we get here, the whole TSO packet has not been acked. */
3013 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3014 {
3015         struct tcp_sock *tp = tcp_sk(sk);
3016         u32 packets_acked;
3017 
3018         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3019 
3020         packets_acked = tcp_skb_pcount(skb);
3021         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3022                 return 0;
3023         packets_acked -= tcp_skb_pcount(skb);
3024 
3025         if (packets_acked) {
3026                 BUG_ON(tcp_skb_pcount(skb) == 0);
3027                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3028         }
3029 
3030         return packets_acked;
3031 }
3032 
3033 /* Remove acknowledged frames from the retransmission queue. If our packet
3034  * is before the ack sequence we can discard it as it's confirmed to have
3035  * arrived at the other end.
3036  */
3037 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3038                                u32 prior_snd_una)
3039 {
3040         struct tcp_sock *tp = tcp_sk(sk);
3041         const struct inet_connection_sock *icsk = inet_csk(sk);
3042         struct sk_buff *skb;
3043         u32 now = tcp_time_stamp;
3044         int fully_acked = true;
3045         int flag = 0;
3046         u32 pkts_acked = 0;
3047         u32 reord = tp->packets_out;
3048         u32 prior_sacked = tp->sacked_out;
3049         s32 seq_rtt = -1;
3050         s32 ca_seq_rtt = -1;
3051         ktime_t last_ackt = net_invalid_timestamp();
3052 
3053         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3054                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3055                 u32 acked_pcount;
3056                 u8 sacked = scb->sacked;
3057 
3058                 /* Determine how many packets and what bytes were acked, tso and else */
3059                 if (after(scb->end_seq, tp->snd_una)) {
3060                         if (tcp_skb_pcount(skb) == 1 ||
3061                             !after(tp->snd_una, scb->seq))
3062                                 break;
3063 
3064                         acked_pcount = tcp_tso_acked(sk, skb);
3065                         if (!acked_pcount)
3066                                 break;
3067 
3068                         fully_acked = false;
3069                 } else {
3070                         acked_pcount = tcp_skb_pcount(skb);
3071                 }
3072 
3073                 if (sacked & TCPCB_RETRANS) {
3074                         if (sacked & TCPCB_SACKED_RETRANS)
3075                                 tp->retrans_out -= acked_pcount;
3076                         flag |= FLAG_RETRANS_DATA_ACKED;
3077                         ca_seq_rtt = -1;
3078                         seq_rtt = -1;
3079                 } else {
3080                         ca_seq_rtt = now - scb->when;
3081                         last_ackt = skb->tstamp;
3082                         if (seq_rtt < 0) {
3083                                 seq_rtt = ca_seq_rtt;
3084                         }
3085                         if (!(sacked & TCPCB_SACKED_ACKED)) {
3086                                 reord = min(pkts_acked, reord);
3087                                 if (!after(scb->end_seq, tp->high_seq))
3088                                         flag |= FLAG_ORIG_SACK_ACKED;
3089                         }
3090                 }
3091 
3092                 if (sacked & TCPCB_SACKED_ACKED)
3093                         tp->sacked_out -= acked_pcount;
3094                 if (sacked & TCPCB_LOST)
3095                         tp->lost_out -= acked_pcount;
3096 
3097                 tp->packets_out -= acked_pcount;
3098                 pkts_acked += acked_pcount;
3099 
3100                 /* Initial outgoing SYN's get put onto the write_queue
3101                  * just like anything else we transmit.  It is not
3102                  * true data, and if we misinform our callers that
3103                  * this ACK acks real data, we will erroneously exit
3104                  * connection startup slow start one packet too
3105                  * quickly.  This is severely frowned upon behavior.
3106                  */
3107                 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3108                         flag |= FLAG_DATA_ACKED;
3109                 } else {
3110                         flag |= FLAG_SYN_ACKED;
3111                         tp->retrans_stamp = 0;
3112                 }
3113 
3114                 if (!fully_acked)
3115                         break;
3116 
3117                 tcp_unlink_write_queue(skb, sk);
3118                 sk_wmem_free_skb(sk, skb);
3119                 tp->scoreboard_skb_hint = NULL;
3120                 if (skb == tp->retransmit_skb_hint)
3121                         tp->retransmit_skb_hint = NULL;
3122                 if (skb == tp->lost_skb_hint)
3123                         tp->lost_skb_hint = NULL;
3124         }
3125 
3126         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3127                 tp->snd_up = tp->snd_una;
3128 
3129         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3130                 flag |= FLAG_SACK_RENEGING;
3131 
3132         if (flag & FLAG_ACKED) {
3133                 const struct tcp_congestion_ops *ca_ops
3134                         = inet_csk(sk)->icsk_ca_ops;
3135 
3136                 if (unlikely(icsk->icsk_mtup.probe_size &&
3137                              !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3138                         tcp_mtup_probe_success(sk);
3139                 }
3140 
3141                 tcp_ack_update_rtt(sk, flag, seq_rtt);
3142                 flag |= FLAG_SET_XMIT_TIMER;  /* set TLP or RTO timer */
3143 
3144                 if (tcp_is_reno(tp)) {
3145                         tcp_remove_reno_sacks(sk, pkts_acked);
3146                 } else {
3147                         int delta;
3148 
3149                         /* Non-retransmitted hole got filled? That's reordering */
3150                         if (reord < prior_fackets)
3151                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3152 
3153                         delta = tcp_is_fack(tp) ? pkts_acked :
3154                                                   prior_sacked - tp->sacked_out;
3155                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3156                 }
3157 
3158                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3159 
3160                 if (ca_ops->pkts_acked) {
3161                         s32 rtt_us = -1;
3162 
3163                         /* Is the ACK triggering packet unambiguous? */
3164                         if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3165                                 /* High resolution needed and available? */
3166                                 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3167                                     !ktime_equal(last_ackt,
3168                                                  net_invalid_timestamp()))
3169                                         rtt_us = ktime_us_delta(ktime_get_real(),
3170                                                                 last_ackt);
3171                                 else if (ca_seq_rtt >= 0)
3172                                         rtt_us = jiffies_to_usecs(ca_seq_rtt);
3173                         }
3174 
3175                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3176                 }
3177         }
3178 
3179 #if FASTRETRANS_DEBUG > 0
3180         WARN_ON((int)tp->sacked_out < 0);
3181         WARN_ON((int)tp->lost_out < 0);
3182         WARN_ON((int)tp->retrans_out < 0);
3183         if (!tp->packets_out && tcp_is_sack(tp)) {
3184                 icsk = inet_csk(sk);
3185                 if (tp->lost_out) {
3186                         pr_debug("Leak l=%u %d\n",
3187                                  tp->lost_out, icsk->icsk_ca_state);
3188                         tp->lost_out = 0;
3189                 }
3190                 if (tp->sacked_out) {
3191                         pr_debug("Leak s=%u %d\n",
3192                                  tp->sacked_out, icsk->icsk_ca_state);
3193                         tp->sacked_out = 0;
3194                 }
3195                 if (tp->retrans_out) {
3196                         pr_debug("Leak r=%u %d\n",
3197                                  tp->retrans_out, icsk->icsk_ca_state);
3198                         tp->retrans_out = 0;
3199                 }
3200         }
3201 #endif
3202         return flag;
3203 }
3204 
3205 static void tcp_ack_probe(struct sock *sk)
3206 {
3207         const struct tcp_sock *tp = tcp_sk(sk);
3208         struct inet_connection_sock *icsk = inet_csk(sk);
3209 
3210         /* Was it a usable window open? */
3211 
3212         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3213                 icsk->icsk_backoff = 0;
3214                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3215                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3216                  * This function is not for random using!
3217                  */
3218         } else {
3219                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3220                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3221                                           TCP_RTO_MAX);
3222         }
3223 }
3224 
3225 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3226 {
3227         return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3228                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3229 }
3230 
3231 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3232 {
3233         const struct tcp_sock *tp = tcp_sk(sk);
3234         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3235                 !tcp_in_cwnd_reduction(sk);
3236 }
3237 
3238 /* Check that window update is acceptable.
3239  * The function assumes that snd_una<=ack<=snd_next.
3240  */
3241 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3242                                         const u32 ack, const u32 ack_seq,
3243                                         const u32 nwin)
3244 {
3245         return  after(ack, tp->snd_una) ||
3246                 after(ack_seq, tp->snd_wl1) ||
3247                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3248 }
3249 
3250 /* Update our send window.
3251  *
3252  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3253  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3254  */
3255 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3256                                  u32 ack_seq)
3257 {
3258         struct tcp_sock *tp = tcp_sk(sk);
3259         int flag = 0;
3260         u32 nwin = ntohs(tcp_hdr(skb)->window);
3261 
3262         if (likely(!tcp_hdr(skb)->syn))
3263                 nwin <<= tp->rx_opt.snd_wscale;
3264 
3265         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3266                 flag |= FLAG_WIN_UPDATE;
3267                 tcp_update_wl(tp, ack_seq);
3268 
3269                 if (tp->snd_wnd != nwin) {
3270                         tp->snd_wnd = nwin;
3271 
3272                         /* Note, it is the only place, where
3273                          * fast path is recovered for sending TCP.
3274                          */
3275                         tp->pred_flags = 0;
3276                         tcp_fast_path_check(sk);
3277 
3278                         if (nwin > tp->max_window) {
3279                                 tp->max_window = nwin;
3280                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3281                         }
3282                 }
3283         }
3284 
3285         tp->snd_una = ack;
3286 
3287         return flag;
3288 }
3289 
3290 /* RFC 5961 7 [ACK Throttling] */
3291 static void tcp_send_challenge_ack(struct sock *sk)
3292 {
3293         /* unprotected vars, we dont care of overwrites */
3294         static u32 challenge_timestamp;
3295         static unsigned int challenge_count;
3296         u32 now = jiffies / HZ;
3297         u32 count;
3298 
3299         if (now != challenge_timestamp) {
3300                 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3301 
3302                 challenge_timestamp = now;
3303                 ACCESS_ONCE(challenge_count) = half +
3304                                   reciprocal_divide(prandom_u32(),
3305                                         sysctl_tcp_challenge_ack_limit);
3306         }
3307         count = ACCESS_ONCE(challenge_count);
3308         if (count > 0) {
3309                 ACCESS_ONCE(challenge_count) = count - 1;
3310                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3311                 tcp_send_ack(sk);
3312         }
3313 }
3314 
3315 static void tcp_store_ts_recent(struct tcp_sock *tp)
3316 {
3317         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3318         tp->rx_opt.ts_recent_stamp = get_seconds();
3319 }
3320 
3321 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3322 {
3323         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3324                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3325                  * extra check below makes sure this can only happen
3326                  * for pure ACK frames.  -DaveM
3327                  *
3328                  * Not only, also it occurs for expired timestamps.
3329                  */
3330 
3331                 if (tcp_paws_check(&tp->rx_opt, 0))
3332                         tcp_store_ts_recent(tp);
3333         }
3334 }
3335 
3336 /* This routine deals with acks during a TLP episode.
3337  * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3338  */
3339 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3340 {
3341         struct tcp_sock *tp = tcp_sk(sk);
3342         bool is_tlp_dupack = (ack == tp->tlp_high_seq) &&
3343                              !(flag & (FLAG_SND_UNA_ADVANCED |
3344                                        FLAG_NOT_DUP | FLAG_DATA_SACKED));
3345 
3346         /* Mark the end of TLP episode on receiving TLP dupack or when
3347          * ack is after tlp_high_seq.
3348          */
3349         if (is_tlp_dupack) {
3350                 tp->tlp_high_seq = 0;
3351                 return;
3352         }
3353 
3354         if (after(ack, tp->tlp_high_seq)) {
3355                 tp->tlp_high_seq = 0;
3356                 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3357                 if (!(flag & FLAG_DSACKING_ACK)) {
3358                         tcp_init_cwnd_reduction(sk, true);
3359                         tcp_set_ca_state(sk, TCP_CA_CWR);
3360                         tcp_end_cwnd_reduction(sk);
3361                         tcp_try_keep_open(sk);
3362                         NET_INC_STATS_BH(sock_net(sk),
3363                                          LINUX_MIB_TCPLOSSPROBERECOVERY);
3364                 }
3365         }
3366 }
3367 
3368 /* This routine deals with incoming acks, but not outgoing ones. */
3369 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3370 {
3371         struct inet_connection_sock *icsk = inet_csk(sk);
3372         struct tcp_sock *tp = tcp_sk(sk);
3373         u32 prior_snd_una = tp->snd_una;
3374         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3375         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3376         bool is_dupack = false;
3377         u32 prior_in_flight, prior_cwnd = tp->snd_cwnd, prior_rtt = tp->srtt;
3378         u32 prior_fackets;
3379         int prior_packets = tp->packets_out;
3380         int prior_sacked = tp->sacked_out;
3381         int pkts_acked = 0;
3382         int previous_packets_out = 0;
3383 
3384         /* If the ack is older than previous acks
3385          * then we can probably ignore it.
3386          */
3387         if (before(ack, prior_snd_una)) {
3388                 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3389                 if (before(ack, prior_snd_una - tp->max_window)) {
3390                         tcp_send_challenge_ack(sk);
3391                         return -1;
3392                 }
3393                 goto old_ack;
3394         }
3395 
3396         /* If the ack includes data we haven't sent yet, discard
3397          * this segment (RFC793 Section 3.9).
3398          */
3399         if (after(ack, tp->snd_nxt))
3400                 goto invalid_ack;
3401 
3402         if (after(ack, prior_snd_una))
3403                 flag |= FLAG_SND_UNA_ADVANCED;
3404 
3405         prior_fackets = tp->fackets_out;
3406         prior_in_flight = tcp_packets_in_flight(tp);
3407 
3408         /* ts_recent update must be made after we are sure that the packet
3409          * is in window.
3410          */
3411         if (flag & FLAG_UPDATE_TS_RECENT)
3412                 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3413 
3414         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3415                 /* Window is constant, pure forward advance.
3416                  * No more checks are required.
3417                  * Note, we use the fact that SND.UNA>=SND.WL2.
3418                  */
3419                 tcp_update_wl(tp, ack_seq);
3420                 tp->snd_una = ack;
3421                 flag |= FLAG_WIN_UPDATE;
3422 
3423                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3424 
3425                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3426         } else {
3427                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3428                         flag |= FLAG_DATA;
3429                 else
3430                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3431 
3432                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3433 
3434                 if (TCP_SKB_CB(skb)->sacked)
3435                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3436 
3437                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3438                         flag |= FLAG_ECE;
3439 
3440                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3441         }
3442 
3443         /* We passed data and got it acked, remove any soft error
3444          * log. Something worked...
3445          */
3446         sk->sk_err_soft = 0;
3447         icsk->icsk_probes_out = 0;
3448         tp->rcv_tstamp = tcp_time_stamp;
3449         if (!prior_packets)
3450                 goto no_queue;
3451 
3452         /* See if we can take anything off of the retransmit queue. */
3453         previous_packets_out = tp->packets_out;
3454         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3455 
3456         pkts_acked = previous_packets_out - tp->packets_out;
3457 
3458         if (tp->tlp_high_seq)
3459                 tcp_process_tlp_ack(sk, ack, flag);
3460         /* If needed, reset TLP/RTO timer; RACK may later override this. */
3461         if (flag & FLAG_SET_XMIT_TIMER)
3462                 tcp_set_xmit_timer(sk);
3463 
3464         if (tcp_ack_is_dubious(sk, flag)) {
3465                 /* Advance CWND, if state allows this. */
3466                 if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag))
3467                         tcp_cong_avoid(sk, ack, prior_in_flight);
3468                 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3469                 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3470                                       prior_packets, is_dupack, flag);
3471         } else {
3472                 if (flag & FLAG_DATA_ACKED)
3473                         tcp_cong_avoid(sk, ack, prior_in_flight);
3474         }
3475 
3476         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3477                 struct dst_entry *dst = __sk_dst_get(sk);
3478                 if (dst)
3479                         dst_confirm(dst);
3480         }
3481 
3482         if (tp->srtt != prior_rtt || tp->snd_cwnd != prior_cwnd)
3483                 tcp_update_pacing_rate(sk);
3484         return 1;
3485 
3486 no_queue:
3487         /* If data was DSACKed, see if we can undo a cwnd reduction. */
3488         if (flag & FLAG_DSACKING_ACK)
3489                 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3490                                       prior_packets, is_dupack, flag);
3491         /* If this ack opens up a zero window, clear backoff.  It was
3492          * being used to time the probes, and is probably far higher than
3493          * it needs to be for normal retransmission.
3494          */
3495         if (tcp_send_head(sk))
3496                 tcp_ack_probe(sk);
3497 
3498         if (tp->tlp_high_seq)
3499                 tcp_process_tlp_ack(sk, ack, flag);
3500         return 1;
3501 
3502 invalid_ack:
3503         SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3504         return -1;
3505 
3506 old_ack:
3507         /* If data was SACKed, tag it and see if we should send more data.
3508          * If data was DSACKed, see if we can undo a cwnd reduction.
3509          */
3510         if (TCP_SKB_CB(skb)->sacked) {
3511                 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3512                 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3513                                       prior_packets, is_dupack, flag);
3514         }
3515 
3516         SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3517         return 0;
3518 }
3519 
3520 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3521  * But, this can also be called on packets in the established flow when
3522  * the fast version below fails.
3523  */
3524 void tcp_parse_options(const struct sk_buff *skb,
3525                        struct tcp_options_received *opt_rx, int estab,
3526                        struct tcp_fastopen_cookie *foc)
3527 {
3528         const unsigned char *ptr;
3529         const struct tcphdr *th = tcp_hdr(skb);
3530         int length = (th->doff * 4) - sizeof(struct tcphdr);
3531 
3532         ptr = (const unsigned char *)(th + 1);
3533         opt_rx->saw_tstamp = 0;
3534 
3535         while (length > 0) {
3536                 int opcode = *ptr++;
3537                 int opsize;
3538 
3539                 switch (opcode) {
3540                 case TCPOPT_EOL:
3541                         return;
3542                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3543                         length--;
3544                         continue;
3545                 default:
3546                         opsize = *ptr++;
3547                         if (opsize < 2) /* "silly options" */
3548                                 return;
3549                         if (opsize > length)
3550                                 return; /* don't parse partial options */
3551                         switch (opcode) {
3552                         case TCPOPT_MSS:
3553                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3554                                         u16 in_mss = get_unaligned_be16(ptr);
3555                                         if (in_mss) {
3556                                                 if (opt_rx->user_mss &&
3557                                                     opt_rx->user_mss < in_mss)
3558                                                         in_mss = opt_rx->user_mss;
3559                                                 opt_rx->mss_clamp = in_mss;
3560                                         }
3561                                 }
3562                                 break;
3563                         case TCPOPT_WINDOW:
3564                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3565                                     !estab && sysctl_tcp_window_scaling) {
3566                                         __u8 snd_wscale = *(__u8 *)ptr;
3567                                         opt_rx->wscale_ok = 1;
3568                                         if (snd_wscale > 14) {
3569                                                 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3570                                                                      __func__,
3571                                                                      snd_wscale);
3572                                                 snd_wscale = 14;
3573                                         }
3574                                         opt_rx->snd_wscale = snd_wscale;
3575                                 }
3576                                 break;
3577                         case TCPOPT_TIMESTAMP:
3578                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3579                                     ((estab && opt_rx->tstamp_ok) ||
3580                                      (!estab && sysctl_tcp_timestamps))) {
3581                                         opt_rx->saw_tstamp = 1;
3582                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3583                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3584                                 }
3585                                 break;
3586                         case TCPOPT_SACK_PERM:
3587                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3588                                     !estab && sysctl_tcp_sack) {
3589                                         opt_rx->sack_ok = TCP_SACK_SEEN;
3590                                         tcp_sack_reset(opt_rx);
3591                                 }
3592                                 break;
3593 
3594                         case TCPOPT_SACK:
3595                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3596                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3597                                    opt_rx->sack_ok) {
3598                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3599                                 }
3600                                 break;
3601 #ifdef CONFIG_TCP_MD5SIG
3602                         case TCPOPT_MD5SIG:
3603                                 /*
3604                                  * The MD5 Hash has already been
3605                                  * checked (see tcp_v{4,6}_do_rcv()).
3606                                  */
3607                                 break;
3608 #endif
3609                         case TCPOPT_EXP:
3610                                 /* Fast Open option shares code 254 using a
3611                                  * 16 bits magic number. It's valid only in
3612                                  * SYN or SYN-ACK with an even size.
3613                                  */
3614                                 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3615                                     get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3616                                     foc == NULL || !th->syn || (opsize & 1))
3617                                         break;
3618                                 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3619                                 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3620                                     foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3621                                         memcpy(foc->val, ptr + 2, foc->len);
3622                                 else if (foc->len != 0)
3623                                         foc->len = -1;
3624                                 break;
3625 
3626                         }
3627                         ptr += opsize-2;
3628                         length -= opsize;
3629                 }
3630         }
3631 }
3632 EXPORT_SYMBOL(tcp_parse_options);
3633 
3634 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3635 {
3636         const __be32 *ptr = (const __be32 *)(th + 1);
3637 
3638         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3639                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3640                 tp->rx_opt.saw_tstamp = 1;
3641                 ++ptr;
3642                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3643                 ++ptr;
3644                 if (*ptr)
3645                         tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3646                 else
3647                         tp->rx_opt.rcv_tsecr = 0;
3648                 return true;
3649         }
3650         return false;
3651 }
3652 
3653 /* Fast parse options. This hopes to only see timestamps.
3654  * If it is wrong it falls back on tcp_parse_options().
3655  */
3656 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3657                                    const struct tcphdr *th, struct tcp_sock *tp)
3658 {
3659         /* In the spirit of fast parsing, compare doff directly to constant
3660          * values.  Because equality is used, short doff can be ignored here.
3661          */
3662         if (th->doff == (sizeof(*th) / 4)) {
3663                 tp->rx_opt.saw_tstamp = 0;
3664                 return false;
3665         } else if (tp->rx_opt.tstamp_ok &&
3666                    th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3667                 if (tcp_parse_aligned_timestamp(tp, th))
3668                         return true;
3669         }
3670 
3671         tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3672         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3673                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3674 
3675         return true;
3676 }
3677 
3678 #ifdef CONFIG_TCP_MD5SIG
3679 /*
3680  * Parse MD5 Signature option
3681  */
3682 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3683 {
3684         int length = (th->doff << 2) - sizeof(*th);
3685         const u8 *ptr = (const u8 *)(th + 1);
3686 
3687         /* If the TCP option is too short, we can short cut */
3688         if (length < TCPOLEN_MD5SIG)
3689                 return NULL;
3690 
3691         while (length > 0) {
3692                 int opcode = *ptr++;
3693                 int opsize;
3694 
3695                 switch(opcode) {
3696                 case TCPOPT_EOL:
3697                         return NULL;
3698                 case TCPOPT_NOP:
3699                         length--;
3700                         continue;
3701                 default:
3702                         opsize = *ptr++;
3703                         if (opsize < 2 || opsize > length)
3704                                 return NULL;
3705                         if (opcode == TCPOPT_MD5SIG)
3706                                 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3707                 }
3708                 ptr += opsize - 2;
3709                 length -= opsize;
3710         }
3711         return NULL;
3712 }
3713 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3714 #endif
3715 
3716 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3717  *
3718  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3719  * it can pass through stack. So, the following predicate verifies that
3720  * this segment is not used for anything but congestion avoidance or
3721  * fast retransmit. Moreover, we even are able to eliminate most of such
3722  * second order effects, if we apply some small "replay" window (~RTO)
3723  * to timestamp space.
3724  *
3725  * All these measures still do not guarantee that we reject wrapped ACKs
3726  * on networks with high bandwidth, when sequence space is recycled fastly,
3727  * but it guarantees that such events will be very rare and do not affect
3728  * connection seriously. This doesn't look nice, but alas, PAWS is really
3729  * buggy extension.
3730  *
3731  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3732  * states that events when retransmit arrives after original data are rare.
3733  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3734  * the biggest problem on large power networks even with minor reordering.
3735  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3736  * up to bandwidth of 18Gigabit/sec. 8) ]
3737  */
3738 
3739 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3740 {
3741         const struct tcp_sock *tp = tcp_sk(sk);
3742         const struct tcphdr *th = tcp_hdr(skb);
3743         u32 seq = TCP_SKB_CB(skb)->seq;
3744         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3745 
3746         return (/* 1. Pure ACK with correct sequence number. */
3747                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3748 
3749                 /* 2. ... and duplicate ACK. */
3750                 ack == tp->snd_una &&
3751 
3752                 /* 3. ... and does not update window. */
3753                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3754 
3755                 /* 4. ... and sits in replay window. */
3756                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3757 }
3758 
3759 static inline bool tcp_paws_discard(const struct sock *sk,
3760                                    const struct sk_buff *skb)
3761 {
3762         const struct tcp_sock *tp = tcp_sk(sk);
3763 
3764         return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3765                !tcp_disordered_ack(sk, skb);
3766 }
3767 
3768 /* Check segment sequence number for validity.
3769  *
3770  * Segment controls are considered valid, if the segment
3771  * fits to the window after truncation to the window. Acceptability
3772  * of data (and SYN, FIN, of course) is checked separately.
3773  * See tcp_data_queue(), for example.
3774  *
3775  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3776  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3777  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3778  * (borrowed from freebsd)
3779  */
3780 
3781 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3782 {
3783         return  !before(end_seq, tp->rcv_wup) &&
3784                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3785 }
3786 
3787 /* When we get a reset we do this. */
3788 void tcp_reset(struct sock *sk)
3789 {
3790         /* We want the right error as BSD sees it (and indeed as we do). */
3791         switch (sk->sk_state) {
3792         case TCP_SYN_SENT:
3793                 sk->sk_err = ECONNREFUSED;
3794                 break;
3795         case TCP_CLOSE_WAIT:
3796                 sk->sk_err = EPIPE;
3797                 break;
3798         case TCP_CLOSE:
3799                 return;
3800         default:
3801                 sk->sk_err = ECONNRESET;
3802         }
3803         /* This barrier is coupled with smp_rmb() in tcp_poll() */
3804         smp_wmb();
3805 
3806         if (!sock_flag(sk, SOCK_DEAD))
3807                 sk->sk_error_report(sk);
3808 
3809         tcp_done(sk);
3810 }
3811 
3812 /*
3813  *      Process the FIN bit. This now behaves as it is supposed to work
3814  *      and the FIN takes effect when it is validly part of sequence
3815  *      space. Not before when we get holes.
3816  *
3817  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3818  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3819  *      TIME-WAIT)
3820  *
3821  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3822  *      close and we go into CLOSING (and later onto TIME-WAIT)
3823  *
3824  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3825  */
3826 static void tcp_fin(struct sock *sk)
3827 {
3828         struct tcp_sock *tp = tcp_sk(sk);
3829 
3830         inet_csk_schedule_ack(sk);
3831 
3832         sk->sk_shutdown |= RCV_SHUTDOWN;
3833         sock_set_flag(sk, SOCK_DONE);
3834 
3835         switch (sk->sk_state) {
3836         case TCP_SYN_RECV:
3837         case TCP_ESTABLISHED:
3838                 /* Move to CLOSE_WAIT */
3839                 tcp_set_state(sk, TCP_CLOSE_WAIT);
3840                 inet_csk(sk)->icsk_ack.pingpong = 1;
3841                 break;
3842 
3843         case TCP_CLOSE_WAIT:
3844         case TCP_CLOSING:
3845                 /* Received a retransmission of the FIN, do
3846                  * nothing.
3847                  */
3848                 break;
3849         case TCP_LAST_ACK:
3850                 /* RFC793: Remain in the LAST-ACK state. */
3851                 break;
3852 
3853         case TCP_FIN_WAIT1:
3854                 /* This case occurs when a simultaneous close
3855                  * happens, we must ack the received FIN and
3856                  * enter the CLOSING state.
3857                  */
3858                 tcp_send_ack(sk);
3859                 tcp_set_state(sk, TCP_CLOSING);
3860                 break;
3861         case TCP_FIN_WAIT2:
3862                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3863                 tcp_send_ack(sk);
3864                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3865                 break;
3866         default:
3867                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3868                  * cases we should never reach this piece of code.
3869                  */
3870                 pr_err("%s: Impossible, sk->sk_state=%d\n",
3871                        __func__, sk->sk_state);
3872                 break;
3873         }
3874 
3875         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3876          * Probably, we should reset in this case. For now drop them.
3877          */
3878         __skb_queue_purge(&tp->out_of_order_queue);
3879         if (tcp_is_sack(tp))
3880                 tcp_sack_reset(&tp->rx_opt);
3881         sk_mem_reclaim(sk);
3882 
3883         if (!sock_flag(sk, SOCK_DEAD)) {
3884                 sk->sk_state_change(sk);
3885 
3886                 /* Do not send POLL_HUP for half duplex close. */
3887                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3888                     sk->sk_state == TCP_CLOSE)
3889                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3890                 else
3891                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3892         }
3893 }
3894 
3895 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3896                                   u32 end_seq)
3897 {
3898         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3899                 if (before(seq, sp->start_seq))
3900                         sp->start_seq = seq;
3901                 if (after(end_seq, sp->end_seq))
3902                         sp->end_seq = end_seq;
3903                 return true;
3904         }
3905         return false;
3906 }
3907 
3908 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3909 {
3910         struct tcp_sock *tp = tcp_sk(sk);
3911 
3912         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3913                 int mib_idx;
3914 
3915                 if (before(seq, tp->rcv_nxt))
3916                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3917                 else
3918                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3919 
3920                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3921 
3922                 tp->rx_opt.dsack = 1;
3923                 tp->duplicate_sack[0].start_seq = seq;
3924                 tp->duplicate_sack[0].end_seq = end_seq;
3925         }
3926 }
3927 
3928 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3929 {
3930         struct tcp_sock *tp = tcp_sk(sk);
3931 
3932         if (!tp->rx_opt.dsack)
3933                 tcp_dsack_set(sk, seq, end_seq);
3934         else
3935                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3936 }
3937 
3938 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
3939 {
3940         struct tcp_sock *tp = tcp_sk(sk);
3941 
3942         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3943             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3944                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3945                 tcp_enter_quickack_mode(sk);
3946 
3947                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3948                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3949 
3950                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3951                                 end_seq = tp->rcv_nxt;
3952                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
3953                 }
3954         }
3955 
3956         tcp_send_ack(sk);
3957 }
3958 
3959 /* These routines update the SACK block as out-of-order packets arrive or
3960  * in-order packets close up the sequence space.
3961  */
3962 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3963 {
3964         int this_sack;
3965         struct tcp_sack_block *sp = &tp->selective_acks[0];
3966         struct tcp_sack_block *swalk = sp + 1;
3967 
3968         /* See if the recent change to the first SACK eats into
3969          * or hits the sequence space of other SACK blocks, if so coalesce.
3970          */
3971         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3972                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3973                         int i;
3974 
3975                         /* Zap SWALK, by moving every further SACK up by one slot.
3976                          * Decrease num_sacks.
3977                          */
3978                         tp->rx_opt.num_sacks--;
3979                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3980                                 sp[i] = sp[i + 1];
3981                         continue;
3982                 }
3983                 this_sack++, swalk++;
3984         }
3985 }
3986 
3987 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3988 {
3989         struct tcp_sock *tp = tcp_sk(sk);
3990         struct tcp_sack_block *sp = &tp->selective_acks[0];
3991         int cur_sacks = tp->rx_opt.num_sacks;
3992         int this_sack;
3993 
3994         if (!cur_sacks)
3995                 goto new_sack;
3996 
3997         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
3998                 if (tcp_sack_extend(sp, seq, end_seq)) {
3999                         /* Rotate this_sack to the first one. */
4000                         for (; this_sack > 0; this_sack--, sp--)
4001                                 swap(*sp, *(sp - 1));
4002                         if (cur_sacks > 1)
4003                                 tcp_sack_maybe_coalesce(tp);
4004                         return;
4005                 }
4006         }
4007 
4008         /* Could not find an adjacent existing SACK, build a new one,
4009          * put it at the front, and shift everyone else down.  We
4010          * always know there is at least one SACK present already here.
4011          *
4012          * If the sack array is full, forget about the last one.
4013          */
4014         if (this_sack >= TCP_NUM_SACKS) {
4015                 this_sack--;
4016                 tp->rx_opt.num_sacks--;
4017                 sp--;
4018         }
4019         for (; this_sack > 0; this_sack--, sp--)
4020                 *sp = *(sp - 1);
4021 
4022 new_sack:
4023         /* Build the new head SACK, and we're done. */
4024         sp->start_seq = seq;
4025         sp->end_seq = end_seq;
4026         tp->rx_opt.num_sacks++;
4027 }
4028 
4029 /* RCV.NXT advances, some SACKs should be eaten. */
4030 
4031 static void tcp_sack_remove(struct tcp_sock *tp)
4032 {
4033         struct tcp_sack_block *sp = &tp->selective_acks[0];
4034         int num_sacks = tp->rx_opt.num_sacks;
4035         int this_sack;
4036 
4037         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4038         if (skb_queue_empty(&tp->out_of_order_queue)) {
4039                 tp->rx_opt.num_sacks = 0;
4040                 return;
4041         }
4042 
4043         for (this_sack = 0; this_sack < num_sacks;) {
4044                 /* Check if the start of the sack is covered by RCV.NXT. */
4045                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4046                         int i;
4047 
4048                         /* RCV.NXT must cover all the block! */
4049                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4050 
4051                         /* Zap this SACK, by moving forward any other SACKS. */
4052                         for (i=this_sack+1; i < num_sacks; i++)
4053                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4054                         num_sacks--;
4055                         continue;
4056                 }
4057                 this_sack++;
4058                 sp++;
4059         }
4060         tp->rx_opt.num_sacks = num_sacks;
4061 }
4062 
4063 /* This one checks to see if we can put data from the
4064  * out_of_order queue into the receive_queue.
4065  */
4066 static void tcp_ofo_queue(struct sock *sk)
4067 {
4068         struct tcp_sock *tp = tcp_sk(sk);
4069         __u32 dsack_high = tp->rcv_nxt;
4070         struct sk_buff *skb;
4071 
4072         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4073                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4074                         break;
4075 
4076                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4077                         __u32 dsack = dsack_high;
4078                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4079                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4080                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4081                 }
4082 
4083                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4084                         SOCK_DEBUG(sk, "ofo packet was already received\n");
4085                         __skb_unlink(skb, &tp->out_of_order_queue);
4086                         __kfree_skb(skb);
4087                         continue;
4088                 }
4089                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4090                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4091                            TCP_SKB_CB(skb)->end_seq);
4092 
4093                 __skb_unlink(skb, &tp->out_of_order_queue);
4094                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4095                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4096                 if (tcp_hdr(skb)->fin)
4097                         tcp_fin(sk);
4098         }
4099 }
4100 
4101 static bool tcp_prune_ofo_queue(struct sock *sk);
4102 static int tcp_prune_queue(struct sock *sk);
4103 
4104 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4105                                  unsigned int size)
4106 {
4107         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4108             !sk_rmem_schedule(sk, skb, size)) {
4109 
4110                 if (tcp_prune_queue(sk) < 0)
4111                         return -1;
4112 
4113                 if (!sk_rmem_schedule(sk, skb, size)) {
4114                         if (!tcp_prune_ofo_queue(sk))
4115                                 return -1;
4116 
4117                         if (!sk_rmem_schedule(sk, skb, size))
4118                                 return -1;
4119                 }
4120         }
4121         return 0;
4122 }
4123 
4124 /**
4125  * tcp_try_coalesce - try to merge skb to prior one
4126  * @sk: socket
4127  * @to: prior buffer
4128  * @from: buffer to add in queue
4129  * @fragstolen: pointer to boolean
4130  *
4131  * Before queueing skb @from after @to, try to merge them
4132  * to reduce overall memory use and queue lengths, if cost is small.
4133  * Packets in ofo or receive queues can stay a long time.
4134  * Better try to coalesce them right now to avoid future collapses.
4135  * Returns true if caller should free @from instead of queueing it
4136  */
4137 static bool tcp_try_coalesce(struct sock *sk,
4138                              struct sk_buff *to,
4139                              struct sk_buff *from,
4140                              bool *fragstolen)
4141 {
4142         int delta;
4143 
4144         *fragstolen = false;
4145 
4146         if (tcp_hdr(from)->fin)
4147                 return false;
4148 
4149         /* Its possible this segment overlaps with prior segment in queue */
4150         if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4151                 return false;
4152 
4153         if (!skb_try_coalesce(to, from, fragstolen, &delta))
4154                 return false;
4155 
4156         atomic_add(delta, &sk->sk_rmem_alloc);
4157         sk_mem_charge(sk, delta);
4158         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4159         TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4160         TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4161         return true;
4162 }
4163 
4164 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4165 {
4166         struct tcp_sock *tp = tcp_sk(sk);
4167         struct sk_buff *skb1;
4168         u32 seq, end_seq;
4169 
4170         TCP_ECN_check_ce(tp, skb);
4171 
4172         if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4173                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4174                 __kfree_skb(skb);
4175                 return;
4176         }
4177 
4178         /* Disable header prediction. */
4179         tp->pred_flags = 0;
4180         inet_csk_schedule_ack(sk);
4181 
4182         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4183         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4184                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4185 
4186         skb1 = skb_peek_tail(&tp->out_of_order_queue);
4187         if (!skb1) {
4188                 /* Initial out of order segment, build 1 SACK. */
4189                 if (tcp_is_sack(tp)) {
4190                         tp->rx_opt.num_sacks = 1;
4191                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4192                         tp->selective_acks[0].end_seq =
4193                                                 TCP_SKB_CB(skb)->end_seq;
4194                 }
4195                 __skb_queue_head(&tp->out_of_order_queue, skb);
4196                 goto end;
4197         }
4198 
4199         seq = TCP_SKB_CB(skb)->seq;
4200         end_seq = TCP_SKB_CB(skb)->end_seq;
4201 
4202         if (seq == TCP_SKB_CB(skb1)->end_seq) {
4203                 bool fragstolen;
4204 
4205                 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4206                         __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4207                 } else {
4208                         kfree_skb_partial(skb, fragstolen);
4209                         skb = NULL;
4210                 }
4211 
4212                 if (!tp->rx_opt.num_sacks ||
4213                     tp->selective_acks[0].end_seq != seq)
4214                         goto add_sack;
4215 
4216                 /* Common case: data arrive in order after hole. */
4217                 tp->selective_acks[0].end_seq = end_seq;
4218                 goto end;
4219         }
4220 
4221         /* Find place to insert this segment. */
4222         while (1) {
4223                 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4224                         break;
4225                 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4226                         skb1 = NULL;
4227                         break;
4228                 }
4229                 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4230         }
4231 
4232         /* Do skb overlap to previous one? */
4233         if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4234                 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4235                         /* All the bits are present. Drop. */
4236                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4237                         __kfree_skb(skb);
4238                         skb = NULL;
4239                         tcp_dsack_set(sk, seq, end_seq);
4240                         goto add_sack;
4241                 }
4242                 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4243                         /* Partial overlap. */
4244                         tcp_dsack_set(sk, seq,
4245                                       TCP_SKB_CB(skb1)->end_seq);
4246                 } else {
4247                         if (skb_queue_is_first(&tp->out_of_order_queue,
4248                                                skb1))
4249                                 skb1 = NULL;
4250                         else
4251                                 skb1 = skb_queue_prev(
4252                                         &tp->out_of_order_queue,
4253                                         skb1);
4254                 }
4255         }
4256         if (!skb1)
4257                 __skb_queue_head(&tp->out_of_order_queue, skb);
4258         else
4259                 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4260 
4261         /* And clean segments covered by new one as whole. */
4262         while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4263                 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4264 
4265                 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4266                         break;
4267                 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4268                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4269                                          end_seq);
4270                         break;
4271                 }
4272                 __skb_unlink(skb1, &tp->out_of_order_queue);
4273                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4274                                  TCP_SKB_CB(skb1)->end_seq);
4275                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4276                 __kfree_skb(skb1);
4277         }
4278 
4279 add_sack:
4280         if (tcp_is_sack(tp))
4281                 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4282 end:
4283         if (skb)
4284                 skb_set_owner_r(skb, sk);
4285 }
4286 
4287 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4288                   bool *fragstolen)
4289 {
4290         int eaten;
4291         struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4292 
4293         __skb_pull(skb, hdrlen);
4294         eaten = (tail &&
4295                  tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4296         tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4297         if (!eaten) {
4298                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4299                 skb_set_owner_r(skb, sk);
4300         }
4301         return eaten;
4302 }
4303 
4304 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4305 {
4306         struct sk_buff *skb = NULL;
4307         struct tcphdr *th;
4308         bool fragstolen;
4309 
4310         if (size == 0)
4311                 return 0;
4312 
4313         skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
4314         if (!skb)
4315                 goto err;
4316 
4317         if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th)))
4318                 goto err_free;
4319 
4320         th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4321         skb_reset_transport_header(skb);
4322         memset(th, 0, sizeof(*th));
4323 
4324         if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4325                 goto err_free;
4326 
4327         TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4328         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4329         TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4330 
4331         if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4332                 WARN_ON_ONCE(fragstolen); /* should not happen */
4333                 __kfree_skb(skb);
4334         }
4335         return size;
4336 
4337 err_free:
4338         kfree_skb(skb);
4339 err:
4340         return -ENOMEM;
4341 }
4342 
4343 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4344 {
4345         const struct tcphdr *th = tcp_hdr(skb);
4346         struct tcp_sock *tp = tcp_sk(sk);
4347         int eaten = -1;
4348         bool fragstolen = false;
4349 
4350         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4351                 goto drop;
4352 
4353         skb_dst_drop(skb);
4354         __skb_pull(skb, th->doff * 4);
4355 
4356         TCP_ECN_accept_cwr(tp, skb);
4357 
4358         tp->rx_opt.dsack = 0;
4359 
4360         /*  Queue data for delivery to the user.
4361          *  Packets in sequence go to the receive queue.
4362          *  Out of sequence packets to the out_of_order_queue.
4363          */
4364         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4365                 if (tcp_receive_window(tp) == 0)
4366                         goto out_of_window;
4367 
4368                 /* Ok. In sequence. In window. */
4369                 if (tp->ucopy.task == current &&
4370                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4371                     sock_owned_by_user(sk) && !tp->urg_data) {
4372                         int chunk = min_t(unsigned int, skb->len,
4373                                           tp->ucopy.len);
4374 
4375                         __set_current_state(TASK_RUNNING);
4376 
4377                         local_bh_enable();
4378                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4379                                 tp->ucopy.len -= chunk;
4380                                 tp->copied_seq += chunk;
4381                                 eaten = (chunk == skb->len);
4382                                 tcp_rcv_space_adjust(sk);
4383                         }
4384                         local_bh_disable();
4385                 }
4386 
4387                 if (eaten <= 0) {
4388 queue_and_out:
4389                         if (eaten < 0 &&
4390                             tcp_try_rmem_schedule(sk, skb, skb->truesize))
4391                                 goto drop;
4392 
4393                         eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4394                 }
4395                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4396                 if (skb->len)
4397                         tcp_event_data_recv(sk, skb);
4398                 if (th->fin)
4399                         tcp_fin(sk);
4400 
4401                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4402                         tcp_ofo_queue(sk);
4403 
4404                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4405                          * gap in queue is filled.
4406                          */
4407                         if (skb_queue_empty(&tp->out_of_order_queue))
4408                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4409                 }
4410 
4411                 if (tp->rx_opt.num_sacks)
4412                         tcp_sack_remove(tp);
4413 
4414                 tcp_fast_path_check(sk);
4415 
4416                 if (eaten > 0)
4417                         kfree_skb_partial(skb, fragstolen);
4418                 if (!sock_flag(sk, SOCK_DEAD))
4419                         sk->sk_data_ready(sk, 0);
4420                 return;
4421         }
4422 
4423         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4424                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4425                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4426                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4427 
4428 out_of_window:
4429                 tcp_enter_quickack_mode(sk);
4430                 inet_csk_schedule_ack(sk);
4431 drop:
4432                 __kfree_skb(skb);
4433                 return;
4434         }
4435 
4436         /* Out of window. F.e. zero window probe. */
4437         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4438                 goto out_of_window;
4439 
4440         tcp_enter_quickack_mode(sk);
4441 
4442         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4443                 /* Partial packet, seq < rcv_next < end_seq */
4444                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4445                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4446                            TCP_SKB_CB(skb)->end_seq);
4447 
4448                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4449 
4450                 /* If window is closed, drop tail of packet. But after
4451                  * remembering D-SACK for its head made in previous line.
4452                  */
4453                 if (!tcp_receive_window(tp))
4454                         goto out_of_window;
4455                 goto queue_and_out;
4456         }
4457 
4458         tcp_data_queue_ofo(sk, skb);
4459 }
4460 
4461 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4462                                         struct sk_buff_head *list)
4463 {
4464         struct sk_buff *next = NULL;
4465 
4466         if (!skb_queue_is_last(list, skb))
4467                 next = skb_queue_next(list, skb);
4468 
4469         __skb_unlink(skb, list);
4470         __kfree_skb(skb);
4471         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4472 
4473         return next;
4474 }
4475 
4476 /* Collapse contiguous sequence of skbs head..tail with
4477  * sequence numbers start..end.
4478  *
4479  * If tail is NULL, this means until the end of the list.
4480  *
4481  * Segments with FIN/SYN are not collapsed (only because this
4482  * simplifies code)
4483  */
4484 static void
4485 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4486              struct sk_buff *head, struct sk_buff *tail,
4487              u32 start, u32 end)
4488 {
4489         struct sk_buff *skb, *n;
4490         bool end_of_skbs;
4491 
4492         /* First, check that queue is collapsible and find
4493          * the point where collapsing can be useful. */
4494         skb = head;
4495 restart:
4496         end_of_skbs = true;
4497         skb_queue_walk_from_safe(list, skb, n) {
4498                 if (skb == tail)
4499                         break;
4500                 /* No new bits? It is possible on ofo queue. */
4501                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4502                         skb = tcp_collapse_one(sk, skb, list);
4503                         if (!skb)
4504                                 break;
4505                         goto restart;
4506                 }
4507 
4508                 /* The first skb to collapse is:
4509                  * - not SYN/FIN and
4510                  * - bloated or contains data before "start" or
4511                  *   overlaps to the next one.
4512                  */
4513                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4514                     (tcp_win_from_space(skb->truesize) > skb->len ||
4515                      before(TCP_SKB_CB(skb)->seq, start))) {
4516                         end_of_skbs = false;
4517                         break;
4518                 }
4519 
4520                 if (!skb_queue_is_last(list, skb)) {
4521                         struct sk_buff *next = skb_queue_next(list, skb);
4522                         if (next != tail &&
4523                             TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4524                                 end_of_skbs = false;
4525                                 break;
4526                         }
4527                 }
4528 
4529                 /* Decided to skip this, advance start seq. */
4530                 start = TCP_SKB_CB(skb)->end_seq;
4531         }
4532         if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4533                 return;
4534 
4535         while (before(start, end)) {
4536                 struct sk_buff *nskb;
4537                 unsigned int header = skb_headroom(skb);
4538                 int copy = SKB_MAX_ORDER(header, 0);
4539 
4540                 /* Too big header? This can happen with IPv6. */
4541                 if (copy < 0)
4542                         return;
4543                 if (end - start < copy)
4544                         copy = end - start;
4545                 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4546                 if (!nskb)
4547                         return;
4548 
4549                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4550                 skb_set_network_header(nskb, (skb_network_header(skb) -
4551                                               skb->head));
4552                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4553                                                 skb->head));
4554                 skb_reserve(nskb, header);
4555                 memcpy(nskb->head, skb->head, header);
4556                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4557                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4558                 __skb_queue_before(list, skb, nskb);
4559                 skb_set_owner_r(nskb, sk);
4560 
4561                 /* Copy data, releasing collapsed skbs. */
4562                 while (copy > 0) {
4563                         int offset = start - TCP_SKB_CB(skb)->seq;
4564                         int size = TCP_SKB_CB(skb)->end_seq - start;
4565 
4566                         BUG_ON(offset < 0);
4567                         if (size > 0) {
4568                                 size = min(copy, size);
4569                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4570                                         BUG();
4571                                 TCP_SKB_CB(nskb)->end_seq += size;
4572                                 copy -= size;
4573                                 start += size;
4574                         }
4575                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4576                                 skb = tcp_collapse_one(sk, skb, list);
4577                                 if (!skb ||
4578                                     skb == tail ||
4579                                     tcp_hdr(skb)->syn ||
4580                                     tcp_hdr(skb)->fin)
4581                                         return;
4582                         }
4583                 }
4584         }
4585 }
4586 
4587 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4588  * and tcp_collapse() them until all the queue is collapsed.
4589  */
4590 static void tcp_collapse_ofo_queue(struct sock *sk)
4591 {
4592         struct tcp_sock *tp = tcp_sk(sk);
4593         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4594         struct sk_buff *head;
4595         u32 start, end;
4596 
4597         if (skb == NULL)
4598                 return;
4599 
4600         start = TCP_SKB_CB(skb)->seq;
4601         end = TCP_SKB_CB(skb)->end_seq;
4602         head = skb;
4603 
4604         for (;;) {
4605                 struct sk_buff *next = NULL;
4606 
4607                 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4608                         next = skb_queue_next(&tp->out_of_order_queue, skb);
4609                 skb = next;
4610 
4611                 /* Segment is terminated when we see gap or when
4612                  * we are at the end of all the queue. */
4613                 if (!skb ||
4614                     after(TCP_SKB_CB(skb)->seq, end) ||
4615                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4616                         tcp_collapse(sk, &tp->out_of_order_queue,
4617                                      head, skb, start, end);
4618                         head = skb;
4619                         if (!skb)
4620                                 break;
4621                         /* Start new segment */
4622                         start = TCP_SKB_CB(skb)->seq;
4623                         end = TCP_SKB_CB(skb)->end_seq;
4624                 } else {
4625                         if (before(TCP_SKB_CB(skb)->seq, start))
4626                                 start = TCP_SKB_CB(skb)->seq;
4627                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4628                                 end = TCP_SKB_CB(skb)->end_seq;
4629                 }
4630         }
4631 }
4632 
4633 /*
4634  * Purge the out-of-order queue.
4635  * Return true if queue was pruned.
4636  */
4637 static bool tcp_prune_ofo_queue(struct sock *sk)
4638 {
4639         struct tcp_sock *tp = tcp_sk(sk);
4640         bool res = false;
4641 
4642         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4643                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4644                 __skb_queue_purge(&tp->out_of_order_queue);
4645 
4646                 /* Reset SACK state.  A conforming SACK implementation will
4647                  * do the same at a timeout based retransmit.  When a connection
4648                  * is in a sad state like this, we care only about integrity
4649                  * of the connection not performance.
4650                  */
4651                 if (tp->rx_opt.sack_ok)
4652                         tcp_sack_reset(&tp->rx_opt);
4653                 sk_mem_reclaim(sk);
4654                 res = true;
4655         }
4656         return res;
4657 }
4658 
4659 /* Reduce allocated memory if we can, trying to get
4660  * the socket within its memory limits again.
4661  *
4662  * Return less than zero if we should start dropping frames
4663  * until the socket owning process reads some of the data
4664  * to stabilize the situation.
4665  */
4666 static int tcp_prune_queue(struct sock *sk)
4667 {
4668         struct tcp_sock *tp = tcp_sk(sk);
4669 
4670         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4671 
4672         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4673 
4674         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4675                 tcp_clamp_window(sk);
4676         else if (sk_under_memory_pressure(sk))
4677                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4678 
4679         tcp_collapse_ofo_queue(sk);
4680         if (!skb_queue_empty(&sk->sk_receive_queue))
4681                 tcp_collapse(sk, &sk->sk_receive_queue,
4682                              skb_peek(&sk->sk_receive_queue),
4683                              NULL,
4684                              tp->copied_seq, tp->rcv_nxt);
4685         sk_mem_reclaim(sk);
4686 
4687         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4688                 return 0;
4689 
4690         /* Collapsing did not help, destructive actions follow.
4691          * This must not ever occur. */
4692 
4693         tcp_prune_ofo_queue(sk);
4694 
4695         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4696                 return 0;
4697 
4698         /* If we are really being abused, tell the caller to silently
4699          * drop receive data on the floor.  It will get retransmitted
4700          * and hopefully then we'll have sufficient space.
4701          */
4702         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4703 
4704         /* Massive buffer overcommit. */
4705         tp->pred_flags = 0;
4706         return -1;
4707 }
4708 
4709 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4710  * As additional protections, we do not touch cwnd in retransmission phases,
4711  * and if application hit its sndbuf limit recently.
4712  */
4713 void tcp_cwnd_application_limited(struct sock *sk)
4714 {
4715         struct tcp_sock *tp = tcp_sk(sk);
4716 
4717         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4718             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4719                 /* Limited by application or receiver window. */
4720                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4721                 u32 win_used = max(tp->snd_cwnd_used, init_win);
4722                 if (win_used < tp->snd_cwnd) {
4723                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
4724                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4725                 }
4726                 tp->snd_cwnd_used = 0;
4727         }
4728         tp->snd_cwnd_stamp = tcp_time_stamp;
4729 }
4730 
4731 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4732 {
4733         const struct tcp_sock *tp = tcp_sk(sk);
4734 
4735         /* If the user specified a specific send buffer setting, do
4736          * not modify it.
4737          */
4738         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4739                 return false;
4740 
4741         /* If we are under global TCP memory pressure, do not expand.  */
4742         if (sk_under_memory_pressure(sk))
4743                 return false;
4744 
4745         /* If we are under soft global TCP memory pressure, do not expand.  */
4746         if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4747                 return false;
4748 
4749         /* If we filled the congestion window, do not expand.  */
4750         if (tp->packets_out >= tp->snd_cwnd)
4751                 return false;
4752 
4753         return true;
4754 }
4755 
4756 /* When incoming ACK allowed to free some skb from write_queue,
4757  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4758  * on the exit from tcp input handler.
4759  *
4760  * PROBLEM: sndbuf expansion does not work well with largesend.
4761  */
4762 static void tcp_new_space(struct sock *sk)
4763 {
4764         struct tcp_sock *tp = tcp_sk(sk);
4765 
4766         if (tcp_should_expand_sndbuf(sk)) {
4767                 int sndmem = SKB_TRUESIZE(max_t(u32,
4768                                                 tp->rx_opt.mss_clamp,
4769                                                 tp->mss_cache) +
4770                                           MAX_TCP_HEADER);
4771                 int demanded = max_t(unsigned int, tp->snd_cwnd,
4772                                      tp->reordering + 1);
4773                 sndmem *= 2 * demanded;
4774                 if (sndmem > sk->sk_sndbuf)
4775                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4776                 tp->snd_cwnd_stamp = tcp_time_stamp;
4777         }
4778 
4779         sk->sk_write_space(sk);
4780 }
4781 
4782 static void tcp_check_space(struct sock *sk)
4783 {
4784         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4785                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4786                 if (sk->sk_socket &&
4787                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4788                         tcp_new_space(sk);
4789         }
4790 }
4791 
4792 static inline void tcp_data_snd_check(struct sock *sk)
4793 {
4794         tcp_push_pending_frames(sk);
4795         tcp_check_space(sk);
4796 }
4797 
4798 /*
4799  * Check if sending an ack is needed.
4800  */
4801 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4802 {
4803         struct tcp_sock *tp = tcp_sk(sk);
4804 
4805             /* More than one full frame received... */
4806         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4807              /* ... and right edge of window advances far enough.
4808               * (tcp_recvmsg() will send ACK otherwise). Or...
4809               */
4810              __tcp_select_window(sk) >= tp->rcv_wnd) ||
4811             /* We ACK each frame or... */
4812             tcp_in_quickack_mode(sk) ||
4813             /* We have out of order data. */
4814             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4815                 /* Then ack it now */
4816                 tcp_send_ack(sk);
4817         } else {
4818                 /* Else, send delayed ack. */
4819                 tcp_send_delayed_ack(sk);
4820         }
4821 }
4822 
4823 static inline void tcp_ack_snd_check(struct sock *sk)
4824 {
4825         if (!inet_csk_ack_scheduled(sk)) {
4826                 /* We sent a data segment already. */
4827                 return;
4828         }
4829         __tcp_ack_snd_check(sk, 1);
4830 }
4831 
4832 /*
4833  *      This routine is only called when we have urgent data
4834  *      signaled. Its the 'slow' part of tcp_urg. It could be
4835  *      moved inline now as tcp_urg is only called from one
4836  *      place. We handle URGent data wrong. We have to - as
4837  *      BSD still doesn't use the correction from RFC961.
4838  *      For 1003.1g we should support a new option TCP_STDURG to permit
4839  *      either form (or just set the sysctl tcp_stdurg).
4840  */
4841 
4842 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4843 {
4844         struct tcp_sock *tp = tcp_sk(sk);
4845         u32 ptr = ntohs(th->urg_ptr);
4846 
4847         if (ptr && !sysctl_tcp_stdurg)
4848                 ptr--;
4849         ptr += ntohl(th->seq);
4850 
4851         /* Ignore urgent data that we've already seen and read. */
4852         if (after(tp->copied_seq, ptr))
4853                 return;
4854 
4855         /* Do not replay urg ptr.
4856          *
4857          * NOTE: interesting situation not covered by specs.
4858          * Misbehaving sender may send urg ptr, pointing to segment,
4859          * which we already have in ofo queue. We are not able to fetch
4860          * such data and will stay in TCP_URG_NOTYET until will be eaten
4861          * by recvmsg(). Seems, we are not obliged to handle such wicked
4862          * situations. But it is worth to think about possibility of some
4863          * DoSes using some hypothetical application level deadlock.
4864          */
4865         if (before(ptr, tp->rcv_nxt))
4866                 return;
4867 
4868         /* Do we already have a newer (or duplicate) urgent pointer? */
4869         if (tp->urg_data && !after(ptr, tp->urg_seq))
4870                 return;
4871 
4872         /* Tell the world about our new urgent pointer. */
4873         sk_send_sigurg(sk);
4874 
4875         /* We may be adding urgent data when the last byte read was
4876          * urgent. To do this requires some care. We cannot just ignore
4877          * tp->copied_seq since we would read the last urgent byte again
4878          * as data, nor can we alter copied_seq until this data arrives
4879          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4880          *
4881          * NOTE. Double Dutch. Rendering to plain English: author of comment
4882          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4883          * and expect that both A and B disappear from stream. This is _wrong_.
4884          * Though this happens in BSD with high probability, this is occasional.
4885          * Any application relying on this is buggy. Note also, that fix "works"
4886          * only in this artificial test. Insert some normal data between A and B and we will
4887          * decline of BSD again. Verdict: it is better to remove to trap
4888          * buggy users.
4889          */
4890         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4891             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4892                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4893                 tp->copied_seq++;
4894                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4895                         __skb_unlink(skb, &sk->sk_receive_queue);
4896                         __kfree_skb(skb);
4897                 }
4898         }
4899 
4900         tp->urg_data = TCP_URG_NOTYET;
4901         tp->urg_seq = ptr;
4902 
4903         /* Disable header prediction. */
4904         tp->pred_flags = 0;
4905 }
4906 
4907 /* This is the 'fast' part of urgent handling. */
4908 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4909 {
4910         struct tcp_sock *tp = tcp_sk(sk);
4911 
4912         /* Check if we get a new urgent pointer - normally not. */
4913         if (th->urg)
4914                 tcp_check_urg(sk, th);
4915 
4916         /* Do we wait for any urgent data? - normally not... */
4917         if (tp->urg_data == TCP_URG_NOTYET) {
4918                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4919                           th->syn;
4920 
4921                 /* Is the urgent pointer pointing into this packet? */
4922                 if (ptr < skb->len) {
4923                         u8 tmp;
4924                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4925                                 BUG();
4926                         tp->urg_data = TCP_URG_VALID | tmp;
4927                         if (!sock_flag(sk, SOCK_DEAD))
4928                                 sk->sk_data_ready(sk, 0);
4929                 }
4930         }
4931 }
4932 
4933 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4934 {
4935         struct tcp_sock *tp = tcp_sk(sk);
4936         int chunk = skb->len - hlen;
4937         int err;
4938 
4939         local_bh_enable();
4940         if (skb_csum_unnecessary(skb))
4941                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4942         else
4943                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4944                                                        tp->ucopy.iov);
4945 
4946         if (!err) {
4947                 tp->ucopy.len -= chunk;
4948                 tp->copied_seq += chunk;
4949                 tcp_rcv_space_adjust(sk);
4950         }
4951 
4952         local_bh_disable();
4953         return err;
4954 }
4955 
4956 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4957                                             struct sk_buff *skb)
4958 {
4959         __sum16 result;
4960 
4961         if (sock_owned_by_user(sk)) {
4962                 local_bh_enable();
4963                 result = __tcp_checksum_complete(skb);
4964                 local_bh_disable();
4965         } else {
4966                 result = __tcp_checksum_complete(skb);
4967         }
4968         return result;
4969 }
4970 
4971 static inline bool tcp_checksum_complete_user(struct sock *sk,
4972                                              struct sk_buff *skb)
4973 {
4974         return !skb_csum_unnecessary(skb) &&
4975                __tcp_checksum_complete_user(sk, skb);
4976 }
4977 
4978 #ifdef CONFIG_NET_DMA
4979 static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4980                                   int hlen)
4981 {
4982         struct tcp_sock *tp = tcp_sk(sk);
4983         int chunk = skb->len - hlen;
4984         int dma_cookie;
4985         bool copied_early = false;
4986 
4987         if (tp->ucopy.wakeup)
4988                 return false;
4989 
4990         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4991                 tp->ucopy.dma_chan = net_dma_find_channel();
4992 
4993         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4994 
4995                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4996                                                          skb, hlen,
4997                                                          tp->ucopy.iov, chunk,
4998                                                          tp->ucopy.pinned_list);
4999 
5000                 if (dma_cookie < 0)
5001                         goto out;
5002 
5003                 tp->ucopy.dma_cookie = dma_cookie;
5004                 copied_early = true;
5005 
5006                 tp->ucopy.len -= chunk;
5007                 tp->copied_seq += chunk;
5008                 tcp_rcv_space_adjust(sk);
5009 
5010                 if ((tp->ucopy.len == 0) ||
5011                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5012                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5013                         tp->ucopy.wakeup = 1;
5014                         sk->sk_data_ready(sk, 0);
5015                 }
5016         } else if (chunk > 0) {
5017                 tp->ucopy.wakeup = 1;
5018                 sk->sk_data_ready(sk, 0);
5019         }
5020 out:
5021         return copied_early;
5022 }
5023 #endif /* CONFIG_NET_DMA */
5024 
5025 /* Does PAWS and seqno based validation of an incoming segment, flags will
5026  * play significant role here.
5027  */
5028 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5029                                   const struct tcphdr *th, int syn_inerr)
5030 {
5031         struct tcp_sock *tp = tcp_sk(sk);
5032 
5033         /* RFC1323: H1. Apply PAWS check first. */
5034         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5035             tcp_paws_discard(sk, skb)) {
5036                 if (!th->rst) {
5037                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5038                         tcp_send_dupack(sk, skb);
5039                         goto discard;
5040                 }
5041                 /* Reset is accepted even if it did not pass PAWS. */
5042         }
5043 
5044         /* Step 1: check sequence number */
5045         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5046                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5047                  * (RST) segments are validated by checking their SEQ-fields."
5048                  * And page 69: "If an incoming segment is not acceptable,
5049                  * an acknowledgment should be sent in reply (unless the RST
5050                  * bit is set, if so drop the segment and return)".
5051                  */
5052                 if (!th->rst) {
5053                         if (th->syn)
5054                                 goto syn_challenge;
5055                         tcp_send_dupack(sk, skb);
5056                 }
5057                 goto discard;
5058         }
5059 
5060         /* Step 2: check RST bit */
5061         if (th->rst) {
5062                 /* RFC 5961 3.2 :
5063                  * If sequence number exactly matches RCV.NXT, then
5064                  *     RESET the connection
5065                  * else
5066                  *     Send a challenge ACK
5067                  */
5068                 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5069                         tcp_reset(sk);
5070                 else
5071                         tcp_send_challenge_ack(sk);
5072                 goto discard;
5073         }
5074 
5075         /* step 3: check security and precedence [ignored] */
5076 
5077         /* step 4: Check for a SYN
5078          * RFC 5691 4.2 : Send a challenge ack
5079          */
5080         if (th->syn) {
5081 syn_challenge:
5082                 if (syn_inerr)
5083                         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5084                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5085                 tcp_send_challenge_ack(sk);
5086                 goto discard;
5087         }
5088 
5089         return true;
5090 
5091 discard:
5092         __kfree_skb(skb);
5093         return false;
5094 }
5095 
5096 /*
5097  *      TCP receive function for the ESTABLISHED state.
5098  *
5099  *      It is split into a fast path and a slow path. The fast path is
5100  *      disabled when:
5101  *      - A zero window was announced from us - zero window probing
5102  *        is only handled properly in the slow path.
5103  *      - Out of order segments arrived.
5104  *      - Urgent data is expected.
5105  *      - There is no buffer space left
5106  *      - Unexpected TCP flags/window values/header lengths are received
5107  *        (detected by checking the TCP header against pred_flags)
5108  *      - Data is sent in both directions. Fast path only supports pure senders
5109  *        or pure receivers (this means either the sequence number or the ack
5110  *        value must stay constant)
5111  *      - Unexpected TCP option.
5112  *
5113  *      When these conditions are not satisfied it drops into a standard
5114  *      receive procedure patterned after RFC793 to handle all cases.
5115  *      The first three cases are guaranteed by proper pred_flags setting,
5116  *      the rest is checked inline. Fast processing is turned on in
5117  *      tcp_data_queue when everything is OK.
5118  */
5119 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5120                         const struct tcphdr *th, unsigned int len)
5121 {
5122         struct tcp_sock *tp = tcp_sk(sk);
5123 
5124         if (unlikely(sk->sk_rx_dst == NULL))
5125                 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5126         /*
5127          *      Header prediction.
5128          *      The code loosely follows the one in the famous
5129          *      "30 instruction TCP receive" Van Jacobson mail.
5130          *
5131          *      Van's trick is to deposit buffers into socket queue
5132          *      on a device interrupt, to call tcp_recv function
5133          *      on the receive process context and checksum and copy
5134          *      the buffer to user space. smart...
5135          *
5136          *      Our current scheme is not silly either but we take the
5137          *      extra cost of the net_bh soft interrupt processing...
5138          *      We do checksum and copy also but from device to kernel.
5139          */
5140 
5141         tp->rx_opt.saw_tstamp = 0;
5142 
5143         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5144          *      if header_prediction is to be made
5145          *      'S' will always be tp->tcp_header_len >> 2
5146          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5147          *  turn it off (when there are holes in the receive
5148          *       space for instance)
5149          *      PSH flag is ignored.
5150          */
5151 
5152         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5153             TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5154             !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5155                 int tcp_header_len = tp->tcp_header_len;
5156 
5157                 /* Timestamp header prediction: tcp_header_len
5158                  * is automatically equal to th->doff*4 due to pred_flags
5159                  * match.
5160                  */
5161 
5162                 /* Check timestamp */
5163                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5164                         /* No? Slow path! */
5165                         if (!tcp_parse_aligned_timestamp(tp, th))
5166                                 goto slow_path;
5167 
5168                         /* If PAWS failed, check it more carefully in slow path */
5169                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5170                                 goto slow_path;
5171 
5172                         /* DO NOT update ts_recent here, if checksum fails
5173                          * and timestamp was corrupted part, it will result
5174                          * in a hung connection since we will drop all
5175                          * future packets due to the PAWS test.
5176                          */
5177                 }
5178 
5179                 if (len <= tcp_header_len) {
5180                         /* Bulk data transfer: sender */
5181                         if (len == tcp_header_len) {
5182                                 /* Predicted packet is in window by definition.
5183                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5184                                  * Hence, check seq<=rcv_wup reduces to:
5185                                  */
5186                                 if (tcp_header_len ==
5187                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5188                                     tp->rcv_nxt == tp->rcv_wup)
5189                                         tcp_store_ts_recent(tp);
5190 
5191                                 /* We know that such packets are checksummed
5192                                  * on entry.
5193                                  */
5194                                 tcp_ack(sk, skb, 0);
5195                                 __kfree_skb(skb);
5196                                 tcp_data_snd_check(sk);
5197                                 return 0;
5198                         } else { /* Header too small */
5199                                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5200                                 goto discard;
5201                         }
5202                 } else {
5203                         int eaten = 0;
5204                         int copied_early = 0;
5205                         bool fragstolen = false;
5206 
5207                         if (tp->copied_seq == tp->rcv_nxt &&
5208                             len - tcp_header_len <= tp->ucopy.len) {
5209 #ifdef CONFIG_NET_DMA
5210                                 if (tp->ucopy.task == current &&
5211                                     sock_owned_by_user(sk) &&
5212                                     tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5213                                         copied_early = 1;
5214                                         eaten = 1;
5215                                 }
5216 #endif
5217                                 if (tp->ucopy.task == current &&
5218                                     sock_owned_by_user(sk) && !copied_early) {
5219                                         __set_current_state(TASK_RUNNING);
5220 
5221                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5222                                                 eaten = 1;
5223                                 }
5224                                 if (eaten) {
5225                                         /* Predicted packet is in window by definition.
5226                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5227                                          * Hence, check seq<=rcv_wup reduces to:
5228                                          */
5229                                         if (tcp_header_len ==
5230                                             (sizeof(struct tcphdr) +
5231                                              TCPOLEN_TSTAMP_ALIGNED) &&
5232                                             tp->rcv_nxt == tp->rcv_wup)
5233                                                 tcp_store_ts_recent(tp);
5234 
5235                                         tcp_rcv_rtt_measure_ts(sk, skb);
5236 
5237                                         __skb_pull(skb, tcp_header_len);
5238                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5239                                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5240                                 }
5241                                 if (copied_early)
5242                                         tcp_cleanup_rbuf(sk, skb->len);
5243                         }
5244                         if (!eaten) {
5245                                 if (tcp_checksum_complete_user(sk, skb))
5246                                         goto csum_error;
5247 
5248                                 if ((int)skb->truesize > sk->sk_forward_alloc)
5249                                         goto step5;
5250 
5251                                 /* Predicted packet is in window by definition.
5252                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5253                                  * Hence, check seq<=rcv_wup reduces to:
5254                                  */
5255                                 if (tcp_header_len ==
5256                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5257                                     tp->rcv_nxt == tp->rcv_wup)
5258                                         tcp_store_ts_recent(tp);
5259 
5260                                 tcp_rcv_rtt_measure_ts(sk, skb);
5261 
5262                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5263 
5264                                 /* Bulk data transfer: receiver */
5265                                 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5266                                                       &fragstolen);
5267                         }
5268 
5269                         tcp_event_data_recv(sk, skb);
5270 
5271                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5272                                 /* Well, only one small jumplet in fast path... */
5273                                 tcp_ack(sk, skb, FLAG_DATA);
5274                                 tcp_data_snd_check(sk);
5275                                 if (!inet_csk_ack_scheduled(sk))
5276                                         goto no_ack;
5277                         }
5278 
5279                         if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5280                                 __tcp_ack_snd_check(sk, 0);
5281 no_ack:
5282 #ifdef CONFIG_NET_DMA
5283                         if (copied_early)
5284                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5285                         else
5286 #endif
5287                         if (eaten)
5288                                 kfree_skb_partial(skb, fragstolen);
5289                         sk->sk_data_ready(sk, 0);
5290                         return 0;
5291                 }
5292         }
5293 
5294 slow_path:
5295         if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5296                 goto csum_error;
5297 
5298         if (!th->ack && !th->rst)
5299                 goto discard;
5300 
5301         /*
5302          *      Standard slow path.
5303          */
5304 
5305         if (!tcp_validate_incoming(sk, skb, th, 1))
5306                 return 0;
5307 
5308 step5:
5309         if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5310                 goto discard;
5311 
5312         tcp_rcv_rtt_measure_ts(sk, skb);
5313 
5314         /* Process urgent data. */
5315         tcp_urg(sk, skb, th);
5316 
5317         /* step 7: process the segment text */
5318         tcp_data_queue(sk, skb);
5319 
5320         tcp_data_snd_check(sk);
5321         tcp_ack_snd_check(sk);
5322         return 0;
5323 
5324 csum_error:
5325         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5326         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5327 
5328 discard:
5329         __kfree_skb(skb);
5330         return 0;
5331 }
5332 EXPORT_SYMBOL(tcp_rcv_established);
5333 
5334 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5335 {
5336         struct tcp_sock *tp = tcp_sk(sk);
5337         struct inet_connection_sock *icsk = inet_csk(sk);
5338 
5339         tcp_set_state(sk, TCP_ESTABLISHED);
5340         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5341 
5342         if (skb != NULL) {
5343                 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5344                 security_inet_conn_established(sk, skb);
5345         }
5346 
5347         /* Make sure socket is routed, for correct metrics.  */
5348         icsk->icsk_af_ops->rebuild_header(sk);
5349 
5350         tcp_init_metrics(sk);
5351 
5352         tcp_init_congestion_control(sk);
5353 
5354         /* Prevent spurious tcp_cwnd_restart() on first data
5355          * packet.
5356          */
5357         tp->lsndtime = tcp_time_stamp;
5358 
5359         tcp_init_buffer_space(sk);
5360 
5361         if (sock_flag(sk, SOCK_KEEPOPEN))
5362                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5363 
5364         if (!tp->rx_opt.snd_wscale)
5365                 __tcp_fast_path_on(tp, tp->snd_wnd);
5366         else
5367                 tp->pred_flags = 0;
5368 
5369         if (!sock_flag(sk, SOCK_DEAD)) {
5370                 sk->sk_state_change(sk);
5371                 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5372         }
5373 }
5374 
5375 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5376                                     struct tcp_fastopen_cookie *cookie)
5377 {
5378         struct tcp_sock *tp = tcp_sk(sk);
5379         struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5380         u16 mss = tp->rx_opt.mss_clamp;
5381         bool syn_drop;
5382 
5383         if (mss == tp->rx_opt.user_mss) {
5384                 struct tcp_options_received opt;
5385 
5386                 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5387                 tcp_clear_options(&opt);
5388                 opt.user_mss = opt.mss_clamp = 0;
5389                 tcp_parse_options(synack, &opt, 0, NULL);
5390                 mss = opt.mss_clamp;
5391         }
5392 
5393         if (!tp->syn_fastopen)  /* Ignore an unsolicited cookie */
5394                 cookie->len = -1;
5395 
5396         /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5397          * the remote receives only the retransmitted (regular) SYNs: either
5398          * the original SYN-data or the corresponding SYN-ACK is lost.
5399          */
5400         syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
5401 
5402         tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
5403 
5404         if (data) { /* Retransmit unacked data in SYN */
5405                 tcp_for_write_queue_from(data, sk) {
5406                         if (data == tcp_send_head(sk) ||
5407                             __tcp_retransmit_skb(sk, data))
5408                                 break;
5409                 }
5410                 tcp_rearm_rto(sk);
5411                 return true;
5412         }
5413         tp->syn_data_acked = tp->syn_data;
5414         return false;
5415 }
5416 
5417 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5418                                          const struct tcphdr *th, unsigned int len)
5419 {
5420         struct inet_connection_sock *icsk = inet_csk(sk);
5421         struct tcp_sock *tp = tcp_sk(sk);
5422         struct tcp_fastopen_cookie foc = { .len = -1 };
5423         int saved_clamp = tp->rx_opt.mss_clamp;
5424 
5425         tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5426         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5427                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5428 
5429         if (th->ack) {
5430                 /* rfc793:
5431                  * "If the state is SYN-SENT then
5432                  *    first check the ACK bit
5433                  *      If the ACK bit is set
5434                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5435                  *        a reset (unless the RST bit is set, if so drop
5436                  *        the segment and return)"
5437                  */
5438                 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5439                     after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5440                         goto reset_and_undo;
5441 
5442                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5443                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5444                              tcp_time_stamp)) {
5445                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5446                         goto reset_and_undo;
5447                 }
5448 
5449                 /* Now ACK is acceptable.
5450                  *
5451                  * "If the RST bit is set
5452                  *    If the ACK was acceptable then signal the user "error:
5453                  *    connection reset", drop the segment, enter CLOSED state,
5454                  *    delete TCB, and return."
5455                  */
5456 
5457                 if (th->rst) {
5458                         tcp_reset(sk);
5459                         goto discard;
5460                 }
5461 
5462                 /* rfc793:
5463                  *   "fifth, if neither of the SYN or RST bits is set then
5464                  *    drop the segment and return."
5465                  *
5466                  *    See note below!
5467                  *                                        --ANK(990513)
5468                  */
5469                 if (!th->syn)
5470                         goto discard_and_undo;
5471 
5472                 /* rfc793:
5473                  *   "If the SYN bit is on ...
5474                  *    are acceptable then ...
5475                  *    (our SYN has been ACKed), change the connection
5476                  *    state to ESTABLISHED..."
5477                  */
5478 
5479                 TCP_ECN_rcv_synack(tp, th);
5480 
5481                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5482                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5483 
5484                 /* Ok.. it's good. Set up sequence numbers and
5485                  * move to established.
5486                  */
5487                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5488                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5489 
5490                 /* RFC1323: The window in SYN & SYN/ACK segments is
5491                  * never scaled.
5492                  */
5493                 tp->snd_wnd = ntohs(th->window);
5494 
5495                 if (!tp->rx_opt.wscale_ok) {
5496                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5497                         tp->window_clamp = min(tp->window_clamp, 65535U);
5498                 }
5499 
5500                 if (tp->rx_opt.saw_tstamp) {
5501                         tp->rx_opt.tstamp_ok       = 1;
5502                         tp->tcp_header_len =
5503                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5504                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5505                         tcp_store_ts_recent(tp);
5506                 } else {
5507                         tp->tcp_header_len = sizeof(struct tcphdr);
5508                 }
5509 
5510                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5511                         tcp_enable_fack(tp);
5512 
5513                 tcp_mtup_init(sk);
5514                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5515                 tcp_initialize_rcv_mss(sk);
5516 
5517                 /* Remember, tcp_poll() does not lock socket!
5518                  * Change state from SYN-SENT only after copied_seq
5519                  * is initialized. */
5520                 tp->copied_seq = tp->rcv_nxt;
5521 
5522                 smp_mb();
5523 
5524                 tcp_finish_connect(sk, skb);
5525 
5526                 if ((tp->syn_fastopen || tp->syn_data) &&
5527                     tcp_rcv_fastopen_synack(sk, skb, &foc))
5528                         return -1;
5529 
5530                 if (sk->sk_write_pending ||
5531                     icsk->icsk_accept_queue.rskq_defer_accept ||
5532                     icsk->icsk_ack.pingpong) {
5533                         /* Save one ACK. Data will be ready after
5534                          * several ticks, if write_pending is set.
5535                          *
5536                          * It may be deleted, but with this feature tcpdumps
5537                          * look so _wonderfully_ clever, that I was not able
5538                          * to stand against the temptation 8)     --ANK
5539                          */
5540                         inet_csk_schedule_ack(sk);
5541                         tcp_enter_quickack_mode(sk);
5542                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5543                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5544 
5545 discard:
5546                         __kfree_skb(skb);
5547                         return 0;
5548                 } else {
5549                         tcp_send_ack(sk);
5550                 }
5551                 return -1;
5552         }
5553 
5554         /* No ACK in the segment */
5555 
5556         if (th->rst) {
5557                 /* rfc793:
5558                  * "If the RST bit is set
5559                  *
5560                  *      Otherwise (no ACK) drop the segment and return."
5561                  */
5562 
5563                 goto discard_and_undo;
5564         }
5565 
5566         /* PAWS check. */
5567         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5568             tcp_paws_reject(&tp->rx_opt, 0))
5569                 goto discard_and_undo;
5570 
5571         if (th->syn) {
5572                 /* We see SYN without ACK. It is attempt of
5573                  * simultaneous connect with crossed SYNs.
5574                  * Particularly, it can be connect to self.
5575                  */
5576                 tcp_set_state(sk, TCP_SYN_RECV);
5577 
5578                 if (tp->rx_opt.saw_tstamp) {
5579                         tp->rx_opt.tstamp_ok = 1;
5580                         tcp_store_ts_recent(tp);
5581                         tp->tcp_header_len =
5582                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5583                 } else {
5584                         tp->tcp_header_len = sizeof(struct tcphdr);
5585                 }
5586 
5587                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5588                 tp->copied_seq = tp->rcv_nxt;
5589                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5590 
5591                 /* RFC1323: The window in SYN & SYN/ACK segments is
5592                  * never scaled.
5593                  */
5594                 tp->snd_wnd    = ntohs(th->window);
5595                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5596                 tp->max_window = tp->snd_wnd;
5597 
5598                 TCP_ECN_rcv_syn(tp, th);
5599 
5600                 tcp_mtup_init(sk);
5601                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5602                 tcp_initialize_rcv_mss(sk);
5603 
5604                 tcp_send_synack(sk);
5605 #if 0
5606                 /* Note, we could accept data and URG from this segment.
5607                  * There are no obstacles to make this (except that we must
5608                  * either change tcp_recvmsg() to prevent it from returning data
5609                  * before 3WHS completes per RFC793, or employ TCP Fast Open).
5610                  *
5611                  * However, if we ignore data in ACKless segments sometimes,
5612                  * we have no reasons to accept it sometimes.
5613                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5614                  * is not flawless. So, discard packet for sanity.
5615                  * Uncomment this return to process the data.
5616                  */
5617                 return -1;
5618 #else
5619                 goto discard;
5620 #endif
5621         }
5622         /* "fifth, if neither of the SYN or RST bits is set then
5623          * drop the segment and return."
5624          */
5625 
5626 discard_and_undo:
5627         tcp_clear_options(&tp->rx_opt);
5628         tp->rx_opt.mss_clamp = saved_clamp;
5629         goto discard;
5630 
5631 reset_and_undo:
5632         tcp_clear_options(&tp->rx_opt);
5633         tp->rx_opt.mss_clamp = saved_clamp;
5634         return 1;
5635 }
5636 
5637 /*
5638  *      This function implements the receiving procedure of RFC 793 for
5639  *      all states except ESTABLISHED and TIME_WAIT.
5640  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5641  *      address independent.
5642  */
5643 
5644 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5645                           const struct tcphdr *th, unsigned int len)
5646 {
5647         struct tcp_sock *tp = tcp_sk(sk);
5648         struct inet_connection_sock *icsk = inet_csk(sk);
5649         struct request_sock *req;
5650         int queued = 0;
5651 
5652         tp->rx_opt.saw_tstamp = 0;
5653 
5654         switch (sk->sk_state) {
5655         case TCP_CLOSE:
5656                 goto discard;
5657 
5658         case TCP_LISTEN:
5659                 if (th->ack)
5660                         return 1;
5661 
5662                 if (th->rst)
5663                         goto discard;
5664 
5665                 if (th->syn) {
5666                         if (th->fin)
5667                                 goto discard;
5668                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5669                                 return 1;
5670 
5671                         /* Now we have several options: In theory there is
5672                          * nothing else in the frame. KA9Q has an option to
5673                          * send data with the syn, BSD accepts data with the
5674                          * syn up to the [to be] advertised window and
5675                          * Solaris 2.1 gives you a protocol error. For now
5676                          * we just ignore it, that fits the spec precisely
5677                          * and avoids incompatibilities. It would be nice in
5678                          * future to drop through and process the data.
5679                          *
5680                          * Now that TTCP is starting to be used we ought to
5681                          * queue this data.
5682                          * But, this leaves one open to an easy denial of
5683                          * service attack, and SYN cookies can't defend
5684                          * against this problem. So, we drop the data
5685                          * in the interest of security over speed unless
5686                          * it's still in use.
5687                          */
5688                         kfree_skb(skb);
5689                         return 0;
5690                 }
5691                 goto discard;
5692 
5693         case TCP_SYN_SENT:
5694                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5695                 if (queued >= 0)
5696                         return queued;
5697 
5698                 /* Do step6 onward by hand. */
5699                 tcp_urg(sk, skb, th);
5700                 __kfree_skb(skb);
5701                 tcp_data_snd_check(sk);
5702                 return 0;
5703         }
5704 
5705         req = tp->fastopen_rsk;
5706         if (req != NULL) {
5707                 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5708                     sk->sk_state != TCP_FIN_WAIT1);
5709 
5710                 if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
5711                         goto discard;
5712         }
5713 
5714         if (!th->ack && !th->rst)
5715                 goto discard;
5716 
5717         if (!tcp_validate_incoming(sk, skb, th, 0))
5718                 return 0;
5719 
5720         /* step 5: check the ACK field */
5721         if (true) {
5722                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5723                                                   FLAG_UPDATE_TS_RECENT) > 0;
5724 
5725                 switch (sk->sk_state) {
5726                 case TCP_SYN_RECV:
5727                         if (acceptable) {
5728                                 /* Once we leave TCP_SYN_RECV, we no longer
5729                                  * need req so release it.
5730                                  */
5731                                 if (req) {
5732                                         tcp_synack_rtt_meas(sk, req);
5733                                         tp->total_retrans = req->num_retrans;
5734 
5735                                         reqsk_fastopen_remove(sk, req, false);
5736                                 } else {
5737                                         /* Make sure socket is routed, for
5738                                          * correct metrics.
5739                                          */
5740                                         icsk->icsk_af_ops->rebuild_header(sk);
5741                                         tcp_init_congestion_control(sk);
5742 
5743                                         tcp_mtup_init(sk);
5744                                         tcp_init_buffer_space(sk);
5745                                         tp->copied_seq = tp->rcv_nxt;
5746                                 }
5747                                 smp_mb();
5748                                 tcp_set_state(sk, TCP_ESTABLISHED);
5749                                 sk->sk_state_change(sk);
5750 
5751                                 /* Note, that this wakeup is only for marginal
5752                                  * crossed SYN case. Passively open sockets
5753                                  * are not waked up, because sk->sk_sleep ==
5754                                  * NULL and sk->sk_socket == NULL.
5755                                  */
5756                                 if (sk->sk_socket)
5757                                         sk_wake_async(sk,
5758                                                       SOCK_WAKE_IO, POLL_OUT);
5759 
5760                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5761                                 tp->snd_wnd = ntohs(th->window) <<
5762                                               tp->rx_opt.snd_wscale;
5763                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5764 
5765                                 if (tp->rx_opt.tstamp_ok)
5766                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5767 
5768                                 if (req) {
5769                                         /* Re-arm the timer because data may
5770                                          * have been sent out. This is similar
5771                                          * to the regular data transmission case
5772                                          * when new data has just been ack'ed.
5773                                          *
5774                                          * (TFO) - we could try to be more
5775                                          * aggressive and retranmitting any data
5776                                          * sooner based on when they were sent
5777                                          * out.
5778                                          */
5779                                         tcp_rearm_rto(sk);
5780                                 } else
5781                                         tcp_init_metrics(sk);
5782 
5783                                 tcp_update_pacing_rate(sk);
5784 
5785                                 /* Prevent spurious tcp_cwnd_restart() on
5786                                  * first data packet.
5787                                  */
5788                                 tp->lsndtime = tcp_time_stamp;
5789 
5790                                 tcp_initialize_rcv_mss(sk);
5791                                 tcp_fast_path_on(tp);
5792                         } else {
5793                                 return 1;
5794                         }
5795                         break;
5796 
5797                 case TCP_FIN_WAIT1:
5798                         /* If we enter the TCP_FIN_WAIT1 state and we are a
5799                          * Fast Open socket and this is the first acceptable
5800                          * ACK we have received, this would have acknowledged
5801                          * our SYNACK so stop the SYNACK timer.
5802                          */
5803                         if (req != NULL) {
5804                                 /* Return RST if ack_seq is invalid.
5805                                  * Note that RFC793 only says to generate a
5806                                  * DUPACK for it but for TCP Fast Open it seems
5807                                  * better to treat this case like TCP_SYN_RECV
5808                                  * above.
5809                                  */
5810                                 if (!acceptable)
5811                                         return 1;
5812                                 /* We no longer need the request sock. */
5813                                 reqsk_fastopen_remove(sk, req, false);
5814                                 tcp_rearm_rto(sk);
5815                         }
5816                         if (tp->snd_una == tp->write_seq) {
5817                                 struct dst_entry *dst;
5818 
5819                                 tcp_set_state(sk, TCP_FIN_WAIT2);
5820                                 sk->sk_shutdown |= SEND_SHUTDOWN;
5821 
5822                                 dst = __sk_dst_get(sk);
5823                                 if (dst)
5824                                         dst_confirm(dst);
5825 
5826                                 if (!sock_flag(sk, SOCK_DEAD))
5827                                         /* Wake up lingering close() */
5828                                         sk->sk_state_change(sk);
5829                                 else {
5830                                         int tmo;
5831 
5832                                         if (tp->linger2 < 0 ||
5833                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5834                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5835                                                 tcp_done(sk);
5836                                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5837                                                 return 1;
5838                                         }
5839 
5840                                         tmo = tcp_fin_time(sk);
5841                                         if (tmo > TCP_TIMEWAIT_LEN) {
5842                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5843                                         } else if (th->fin || sock_owned_by_user(sk)) {
5844                                                 /* Bad case. We could lose such FIN otherwise.
5845                                                  * It is not a big problem, but it looks confusing
5846                                                  * and not so rare event. We still can lose it now,
5847                                                  * if it spins in bh_lock_sock(), but it is really
5848                                                  * marginal case.
5849                                                  */
5850                                                 inet_csk_reset_keepalive_timer(sk, tmo);
5851                                         } else {
5852                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5853                                                 goto discard;
5854                                         }
5855                                 }
5856                         }
5857                         break;
5858 
5859                 case TCP_CLOSING:
5860                         if (tp->snd_una == tp->write_seq) {
5861                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5862                                 goto discard;
5863                         }
5864                         break;
5865 
5866                 case TCP_LAST_ACK:
5867                         if (tp->snd_una == tp->write_seq) {
5868                                 tcp_update_metrics(sk);
5869                                 tcp_done(sk);
5870                                 goto discard;
5871                         }
5872                         break;
5873                 }
5874         }
5875 
5876         /* step 6: check the URG bit */
5877         tcp_urg(sk, skb, th);
5878 
5879         /* step 7: process the segment text */
5880         switch (sk->sk_state) {
5881         case TCP_CLOSE_WAIT:
5882         case TCP_CLOSING:
5883         case TCP_LAST_ACK:
5884                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5885                         break;
5886         case TCP_FIN_WAIT1:
5887         case TCP_FIN_WAIT2:
5888                 /* RFC 793 says to queue data in these states,
5889                  * RFC 1122 says we MUST send a reset.
5890                  * BSD 4.4 also does reset.
5891                  */
5892                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5893                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5894                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5895                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5896                                 tcp_reset(sk);
5897                                 return 1;
5898                         }
5899                 }
5900                 /* Fall through */
5901         case TCP_ESTABLISHED:
5902                 tcp_data_queue(sk, skb);
5903                 queued = 1;
5904                 break;
5905         }
5906 
5907         /* tcp_data could move socket to TIME-WAIT */
5908         if (sk->sk_state != TCP_CLOSE) {
5909                 tcp_data_snd_check(sk);
5910                 tcp_ack_snd_check(sk);
5911         }
5912 
5913         if (!queued) {
5914 discard:
5915                 __kfree_skb(skb);
5916         }
5917         return 0;
5918 }
5919 EXPORT_SYMBOL(tcp_rcv_state_process);
5920 

~ [ 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