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

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

Version: ~ [ linux-5.17-rc1 ] ~ [ linux-5.16.2 ] ~ [ linux-5.15.16 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.93 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.173 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.225 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.262 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.297 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.299 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 // SPDX-License-Identifier: GPL-2.0
  2 /*
  3  * INET         An implementation of the TCP/IP protocol suite for the LINUX
  4  *              operating system.  INET is implemented using the  BSD Socket
  5  *              interface as the means of communication with the user level.
  6  *
  7  *              Implementation of the Transmission Control Protocol(TCP).
  8  *
  9  * Authors:     Ross Biro
 10  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
 11  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
 12  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
 13  *              Florian La Roche, <flla@stud.uni-sb.de>
 14  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
 15  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
 16  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
 17  *              Matthew Dillon, <dillon@apollo.west.oic.com>
 18  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
 19  *              Jorge Cwik, <jorge@laser.satlink.net>
 20  */
 21 
 22 /*
 23  * Changes:
 24  *              Pedro Roque     :       Fast Retransmit/Recovery.
 25  *                                      Two receive queues.
 26  *                                      Retransmit queue handled by TCP.
 27  *                                      Better retransmit timer handling.
 28  *                                      New congestion avoidance.
 29  *                                      Header prediction.
 30  *                                      Variable renaming.
 31  *
 32  *              Eric            :       Fast Retransmit.
 33  *              Randy Scott     :       MSS option defines.
 34  *              Eric Schenk     :       Fixes to slow start algorithm.
 35  *              Eric Schenk     :       Yet another double ACK bug.
 36  *              Eric Schenk     :       Delayed ACK bug fixes.
 37  *              Eric Schenk     :       Floyd style fast retrans war avoidance.
 38  *              David S. Miller :       Don't allow zero congestion window.
 39  *              Eric Schenk     :       Fix retransmitter so that it sends
 40  *                                      next packet on ack of previous packet.
 41  *              Andi Kleen      :       Moved open_request checking here
 42  *                                      and process RSTs for open_requests.
 43  *              Andi Kleen      :       Better prune_queue, and other fixes.
 44  *              Andrey Savochkin:       Fix RTT measurements in the presence of
 45  *                                      timestamps.
 46  *              Andrey Savochkin:       Check sequence numbers correctly when
 47  *                                      removing SACKs due to in sequence incoming
 48  *                                      data segments.
 49  *              Andi Kleen:             Make sure we never ack data there is not
 50  *                                      enough room for. Also make this condition
 51  *                                      a fatal error if it might still happen.
 52  *              Andi Kleen:             Add tcp_measure_rcv_mss to make
 53  *                                      connections with MSS<min(MTU,ann. MSS)
 54  *                                      work without delayed acks.
 55  *              Andi Kleen:             Process packets with PSH set in the
 56  *                                      fast path.
 57  *              J Hadi Salim:           ECN support
 58  *              Andrei Gurtov,
 59  *              Pasi Sarolahti,
 60  *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
 61  *                                      engine. Lots of bugs are found.
 62  *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
 63  */
 64 
 65 #define pr_fmt(fmt) "TCP: " fmt
 66 
 67 #include <linux/mm.h>
 68 #include <linux/slab.h>
 69 #include <linux/module.h>
 70 #include <linux/sysctl.h>
 71 #include <linux/kernel.h>
 72 #include <linux/prefetch.h>
 73 #include <net/dst.h>
 74 #include <net/tcp.h>
 75 #include <net/inet_common.h>
 76 #include <linux/ipsec.h>
 77 #include <asm/unaligned.h>
 78 #include <linux/errqueue.h>
 79 #include <trace/events/tcp.h>
 80 #include <linux/static_key.h>
 81 #include <net/busy_poll.h>
 82 
 83 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
 84 
 85 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
 86 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
 87 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
 88 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
 89 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
 90 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
 91 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
 92 #define FLAG_LOST_RETRANS       0x80 /* This ACK marks some retransmission lost */
 93 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
 94 #define FLAG_ORIG_SACK_ACKED    0x200 /* Never retransmitted data are (s)acked  */
 95 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
 96 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
 97 #define FLAG_SET_XMIT_TIMER     0x1000 /* Set TLP or RTO timer */
 98 #define FLAG_SACK_RENEGING      0x2000 /* snd_una advanced to a sacked seq */
 99 #define FLAG_UPDATE_TS_RECENT   0x4000 /* tcp_replace_ts_recent() */
100 #define FLAG_NO_CHALLENGE_ACK   0x8000 /* do not call tcp_send_challenge_ack()  */
101 #define FLAG_ACK_MAYBE_DELAYED  0x10000 /* Likely a delayed ACK */
102 
103 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
104 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
105 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
106 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
107 
108 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
109 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
110 
111 #define REXMIT_NONE     0 /* no loss recovery to do */
112 #define REXMIT_LOST     1 /* retransmit packets marked lost */
113 #define REXMIT_NEW      2 /* FRTO-style transmit of unsent/new packets */
114 
115 #if IS_ENABLED(CONFIG_TLS_DEVICE)
116 static DEFINE_STATIC_KEY_FALSE(clean_acked_data_enabled);
117 
118 void clean_acked_data_enable(struct inet_connection_sock *icsk,
119                              void (*cad)(struct sock *sk, u32 ack_seq))
120 {
121         icsk->icsk_clean_acked = cad;
122         static_branch_inc(&clean_acked_data_enabled);
123 }
124 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
125 
126 void clean_acked_data_disable(struct inet_connection_sock *icsk)
127 {
128         static_branch_dec(&clean_acked_data_enabled);
129         icsk->icsk_clean_acked = NULL;
130 }
131 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
132 #endif
133 
134 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
135                              unsigned int len)
136 {
137         static bool __once __read_mostly;
138 
139         if (!__once) {
140                 struct net_device *dev;
141 
142                 __once = true;
143 
144                 rcu_read_lock();
145                 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
146                 if (!dev || len >= dev->mtu)
147                         pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
148                                 dev ? dev->name : "Unknown driver");
149                 rcu_read_unlock();
150         }
151 }
152 
153 /* Adapt the MSS value used to make delayed ack decision to the
154  * real world.
155  */
156 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
157 {
158         struct inet_connection_sock *icsk = inet_csk(sk);
159         const unsigned int lss = icsk->icsk_ack.last_seg_size;
160         unsigned int len;
161 
162         icsk->icsk_ack.last_seg_size = 0;
163 
164         /* skb->len may jitter because of SACKs, even if peer
165          * sends good full-sized frames.
166          */
167         len = skb_shinfo(skb)->gso_size ? : skb->len;
168         if (len >= icsk->icsk_ack.rcv_mss) {
169                 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
170                                                tcp_sk(sk)->advmss);
171                 /* Account for possibly-removed options */
172                 if (unlikely(len > icsk->icsk_ack.rcv_mss +
173                                    MAX_TCP_OPTION_SPACE))
174                         tcp_gro_dev_warn(sk, skb, len);
175         } else {
176                 /* Otherwise, we make more careful check taking into account,
177                  * that SACKs block is variable.
178                  *
179                  * "len" is invariant segment length, including TCP header.
180                  */
181                 len += skb->data - skb_transport_header(skb);
182                 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
183                     /* If PSH is not set, packet should be
184                      * full sized, provided peer TCP is not badly broken.
185                      * This observation (if it is correct 8)) allows
186                      * to handle super-low mtu links fairly.
187                      */
188                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
189                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
190                         /* Subtract also invariant (if peer is RFC compliant),
191                          * tcp header plus fixed timestamp option length.
192                          * Resulting "len" is MSS free of SACK jitter.
193                          */
194                         len -= tcp_sk(sk)->tcp_header_len;
195                         icsk->icsk_ack.last_seg_size = len;
196                         if (len == lss) {
197                                 icsk->icsk_ack.rcv_mss = len;
198                                 return;
199                         }
200                 }
201                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
202                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
203                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
204         }
205 }
206 
207 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
208 {
209         struct inet_connection_sock *icsk = inet_csk(sk);
210         unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
211 
212         if (quickacks == 0)
213                 quickacks = 2;
214         quickacks = min(quickacks, max_quickacks);
215         if (quickacks > icsk->icsk_ack.quick)
216                 icsk->icsk_ack.quick = quickacks;
217 }
218 
219 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
220 {
221         struct inet_connection_sock *icsk = inet_csk(sk);
222 
223         tcp_incr_quickack(sk, max_quickacks);
224         inet_csk_exit_pingpong_mode(sk);
225         icsk->icsk_ack.ato = TCP_ATO_MIN;
226 }
227 EXPORT_SYMBOL(tcp_enter_quickack_mode);
228 
229 /* Send ACKs quickly, if "quick" count is not exhausted
230  * and the session is not interactive.
231  */
232 
233 static bool tcp_in_quickack_mode(struct sock *sk)
234 {
235         const struct inet_connection_sock *icsk = inet_csk(sk);
236         const struct dst_entry *dst = __sk_dst_get(sk);
237 
238         return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
239                 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
240 }
241 
242 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
243 {
244         if (tp->ecn_flags & TCP_ECN_OK)
245                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
246 }
247 
248 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
249 {
250         if (tcp_hdr(skb)->cwr) {
251                 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
252 
253                 /* If the sender is telling us it has entered CWR, then its
254                  * cwnd may be very low (even just 1 packet), so we should ACK
255                  * immediately.
256                  */
257                 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
258         }
259 }
260 
261 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
262 {
263         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
264 }
265 
266 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
267 {
268         struct tcp_sock *tp = tcp_sk(sk);
269 
270         switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
271         case INET_ECN_NOT_ECT:
272                 /* Funny extension: if ECT is not set on a segment,
273                  * and we already seen ECT on a previous segment,
274                  * it is probably a retransmit.
275                  */
276                 if (tp->ecn_flags & TCP_ECN_SEEN)
277                         tcp_enter_quickack_mode(sk, 2);
278                 break;
279         case INET_ECN_CE:
280                 if (tcp_ca_needs_ecn(sk))
281                         tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
282 
283                 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
284                         /* Better not delay acks, sender can have a very low cwnd */
285                         tcp_enter_quickack_mode(sk, 2);
286                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
287                 }
288                 tp->ecn_flags |= TCP_ECN_SEEN;
289                 break;
290         default:
291                 if (tcp_ca_needs_ecn(sk))
292                         tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
293                 tp->ecn_flags |= TCP_ECN_SEEN;
294                 break;
295         }
296 }
297 
298 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
299 {
300         if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
301                 __tcp_ecn_check_ce(sk, skb);
302 }
303 
304 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
305 {
306         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
307                 tp->ecn_flags &= ~TCP_ECN_OK;
308 }
309 
310 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
311 {
312         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
313                 tp->ecn_flags &= ~TCP_ECN_OK;
314 }
315 
316 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
317 {
318         if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
319                 return true;
320         return false;
321 }
322 
323 /* Buffer size and advertised window tuning.
324  *
325  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
326  */
327 
328 static void tcp_sndbuf_expand(struct sock *sk)
329 {
330         const struct tcp_sock *tp = tcp_sk(sk);
331         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
332         int sndmem, per_mss;
333         u32 nr_segs;
334 
335         /* Worst case is non GSO/TSO : each frame consumes one skb
336          * and skb->head is kmalloced using power of two area of memory
337          */
338         per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
339                   MAX_TCP_HEADER +
340                   SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
341 
342         per_mss = roundup_pow_of_two(per_mss) +
343                   SKB_DATA_ALIGN(sizeof(struct sk_buff));
344 
345         nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
346         nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
347 
348         /* Fast Recovery (RFC 5681 3.2) :
349          * Cubic needs 1.7 factor, rounded to 2 to include
350          * extra cushion (application might react slowly to EPOLLOUT)
351          */
352         sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
353         sndmem *= nr_segs * per_mss;
354 
355         if (sk->sk_sndbuf < sndmem)
356                 sk->sk_sndbuf = min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]);
357 }
358 
359 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
360  *
361  * All tcp_full_space() is split to two parts: "network" buffer, allocated
362  * forward and advertised in receiver window (tp->rcv_wnd) and
363  * "application buffer", required to isolate scheduling/application
364  * latencies from network.
365  * window_clamp is maximal advertised window. It can be less than
366  * tcp_full_space(), in this case tcp_full_space() - window_clamp
367  * is reserved for "application" buffer. The less window_clamp is
368  * the smoother our behaviour from viewpoint of network, but the lower
369  * throughput and the higher sensitivity of the connection to losses. 8)
370  *
371  * rcv_ssthresh is more strict window_clamp used at "slow start"
372  * phase to predict further behaviour of this connection.
373  * It is used for two goals:
374  * - to enforce header prediction at sender, even when application
375  *   requires some significant "application buffer". It is check #1.
376  * - to prevent pruning of receive queue because of misprediction
377  *   of receiver window. Check #2.
378  *
379  * The scheme does not work when sender sends good segments opening
380  * window and then starts to feed us spaghetti. But it should work
381  * in common situations. Otherwise, we have to rely on queue collapsing.
382  */
383 
384 /* Slow part of check#2. */
385 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
386 {
387         struct tcp_sock *tp = tcp_sk(sk);
388         /* Optimize this! */
389         int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
390         int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
391 
392         while (tp->rcv_ssthresh <= window) {
393                 if (truesize <= skb->len)
394                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
395 
396                 truesize >>= 1;
397                 window >>= 1;
398         }
399         return 0;
400 }
401 
402 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
403 {
404         struct tcp_sock *tp = tcp_sk(sk);
405         int room;
406 
407         room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
408 
409         /* Check #1 */
410         if (room > 0 && !tcp_under_memory_pressure(sk)) {
411                 int incr;
412 
413                 /* Check #2. Increase window, if skb with such overhead
414                  * will fit to rcvbuf in future.
415                  */
416                 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
417                         incr = 2 * tp->advmss;
418                 else
419                         incr = __tcp_grow_window(sk, skb);
420 
421                 if (incr) {
422                         incr = max_t(int, incr, 2 * skb->len);
423                         tp->rcv_ssthresh += min(room, incr);
424                         inet_csk(sk)->icsk_ack.quick |= 1;
425                 }
426         }
427 }
428 
429 /* 3. Try to fixup all. It is made immediately after connection enters
430  *    established state.
431  */
432 void tcp_init_buffer_space(struct sock *sk)
433 {
434         int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
435         struct tcp_sock *tp = tcp_sk(sk);
436         int maxwin;
437 
438         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
439                 tcp_sndbuf_expand(sk);
440 
441         tp->rcvq_space.space = min_t(u32, tp->rcv_wnd, TCP_INIT_CWND * tp->advmss);
442         tcp_mstamp_refresh(tp);
443         tp->rcvq_space.time = tp->tcp_mstamp;
444         tp->rcvq_space.seq = tp->copied_seq;
445 
446         maxwin = tcp_full_space(sk);
447 
448         if (tp->window_clamp >= maxwin) {
449                 tp->window_clamp = maxwin;
450 
451                 if (tcp_app_win && maxwin > 4 * tp->advmss)
452                         tp->window_clamp = max(maxwin -
453                                                (maxwin >> tcp_app_win),
454                                                4 * tp->advmss);
455         }
456 
457         /* Force reservation of one segment. */
458         if (tcp_app_win &&
459             tp->window_clamp > 2 * tp->advmss &&
460             tp->window_clamp + tp->advmss > maxwin)
461                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
462 
463         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
464         tp->snd_cwnd_stamp = tcp_jiffies32;
465 }
466 
467 /* 4. Recalculate window clamp after socket hit its memory bounds. */
468 static void tcp_clamp_window(struct sock *sk)
469 {
470         struct tcp_sock *tp = tcp_sk(sk);
471         struct inet_connection_sock *icsk = inet_csk(sk);
472         struct net *net = sock_net(sk);
473 
474         icsk->icsk_ack.quick = 0;
475 
476         if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
477             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
478             !tcp_under_memory_pressure(sk) &&
479             sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
480                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
481                                     net->ipv4.sysctl_tcp_rmem[2]);
482         }
483         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
484                 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
485 }
486 
487 /* Initialize RCV_MSS value.
488  * RCV_MSS is an our guess about MSS used by the peer.
489  * We haven't any direct information about the MSS.
490  * It's better to underestimate the RCV_MSS rather than overestimate.
491  * Overestimations make us ACKing less frequently than needed.
492  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
493  */
494 void tcp_initialize_rcv_mss(struct sock *sk)
495 {
496         const struct tcp_sock *tp = tcp_sk(sk);
497         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
498 
499         hint = min(hint, tp->rcv_wnd / 2);
500         hint = min(hint, TCP_MSS_DEFAULT);
501         hint = max(hint, TCP_MIN_MSS);
502 
503         inet_csk(sk)->icsk_ack.rcv_mss = hint;
504 }
505 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
506 
507 /* Receiver "autotuning" code.
508  *
509  * The algorithm for RTT estimation w/o timestamps is based on
510  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
511  * <http://public.lanl.gov/radiant/pubs.html#DRS>
512  *
513  * More detail on this code can be found at
514  * <http://staff.psc.edu/jheffner/>,
515  * though this reference is out of date.  A new paper
516  * is pending.
517  */
518 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
519 {
520         u32 new_sample = tp->rcv_rtt_est.rtt_us;
521         long m = sample;
522 
523         if (new_sample != 0) {
524                 /* If we sample in larger samples in the non-timestamp
525                  * case, we could grossly overestimate the RTT especially
526                  * with chatty applications or bulk transfer apps which
527                  * are stalled on filesystem I/O.
528                  *
529                  * Also, since we are only going for a minimum in the
530                  * non-timestamp case, we do not smooth things out
531                  * else with timestamps disabled convergence takes too
532                  * long.
533                  */
534                 if (!win_dep) {
535                         m -= (new_sample >> 3);
536                         new_sample += m;
537                 } else {
538                         m <<= 3;
539                         if (m < new_sample)
540                                 new_sample = m;
541                 }
542         } else {
543                 /* No previous measure. */
544                 new_sample = m << 3;
545         }
546 
547         tp->rcv_rtt_est.rtt_us = new_sample;
548 }
549 
550 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
551 {
552         u32 delta_us;
553 
554         if (tp->rcv_rtt_est.time == 0)
555                 goto new_measure;
556         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
557                 return;
558         delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
559         if (!delta_us)
560                 delta_us = 1;
561         tcp_rcv_rtt_update(tp, delta_us, 1);
562 
563 new_measure:
564         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
565         tp->rcv_rtt_est.time = tp->tcp_mstamp;
566 }
567 
568 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
569                                           const struct sk_buff *skb)
570 {
571         struct tcp_sock *tp = tcp_sk(sk);
572 
573         if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
574                 return;
575         tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
576 
577         if (TCP_SKB_CB(skb)->end_seq -
578             TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
579                 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
580                 u32 delta_us;
581 
582                 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
583                         if (!delta)
584                                 delta = 1;
585                         delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
586                         tcp_rcv_rtt_update(tp, delta_us, 0);
587                 }
588         }
589 }
590 
591 /*
592  * This function should be called every time data is copied to user space.
593  * It calculates the appropriate TCP receive buffer space.
594  */
595 void tcp_rcv_space_adjust(struct sock *sk)
596 {
597         struct tcp_sock *tp = tcp_sk(sk);
598         u32 copied;
599         int time;
600 
601         trace_tcp_rcv_space_adjust(sk);
602 
603         tcp_mstamp_refresh(tp);
604         time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
605         if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
606                 return;
607 
608         /* Number of bytes copied to user in last RTT */
609         copied = tp->copied_seq - tp->rcvq_space.seq;
610         if (copied <= tp->rcvq_space.space)
611                 goto new_measure;
612 
613         /* A bit of theory :
614          * copied = bytes received in previous RTT, our base window
615          * To cope with packet losses, we need a 2x factor
616          * To cope with slow start, and sender growing its cwin by 100 %
617          * every RTT, we need a 4x factor, because the ACK we are sending
618          * now is for the next RTT, not the current one :
619          * <prev RTT . ><current RTT .. ><next RTT .... >
620          */
621 
622         if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
623             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
624                 int rcvmem, rcvbuf;
625                 u64 rcvwin, grow;
626 
627                 /* minimal window to cope with packet losses, assuming
628                  * steady state. Add some cushion because of small variations.
629                  */
630                 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
631 
632                 /* Accommodate for sender rate increase (eg. slow start) */
633                 grow = rcvwin * (copied - tp->rcvq_space.space);
634                 do_div(grow, tp->rcvq_space.space);
635                 rcvwin += (grow << 1);
636 
637                 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
638                 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
639                         rcvmem += 128;
640 
641                 do_div(rcvwin, tp->advmss);
642                 rcvbuf = min_t(u64, rcvwin * rcvmem,
643                                sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
644                 if (rcvbuf > sk->sk_rcvbuf) {
645                         sk->sk_rcvbuf = rcvbuf;
646 
647                         /* Make the window clamp follow along.  */
648                         tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
649                 }
650         }
651         tp->rcvq_space.space = copied;
652 
653 new_measure:
654         tp->rcvq_space.seq = tp->copied_seq;
655         tp->rcvq_space.time = tp->tcp_mstamp;
656 }
657 
658 /* There is something which you must keep in mind when you analyze the
659  * behavior of the tp->ato delayed ack timeout interval.  When a
660  * connection starts up, we want to ack as quickly as possible.  The
661  * problem is that "good" TCP's do slow start at the beginning of data
662  * transmission.  The means that until we send the first few ACK's the
663  * sender will sit on his end and only queue most of his data, because
664  * he can only send snd_cwnd unacked packets at any given time.  For
665  * each ACK we send, he increments snd_cwnd and transmits more of his
666  * queue.  -DaveM
667  */
668 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
669 {
670         struct tcp_sock *tp = tcp_sk(sk);
671         struct inet_connection_sock *icsk = inet_csk(sk);
672         u32 now;
673 
674         inet_csk_schedule_ack(sk);
675 
676         tcp_measure_rcv_mss(sk, skb);
677 
678         tcp_rcv_rtt_measure(tp);
679 
680         now = tcp_jiffies32;
681 
682         if (!icsk->icsk_ack.ato) {
683                 /* The _first_ data packet received, initialize
684                  * delayed ACK engine.
685                  */
686                 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
687                 icsk->icsk_ack.ato = TCP_ATO_MIN;
688         } else {
689                 int m = now - icsk->icsk_ack.lrcvtime;
690 
691                 if (m <= TCP_ATO_MIN / 2) {
692                         /* The fastest case is the first. */
693                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
694                 } else if (m < icsk->icsk_ack.ato) {
695                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
696                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
697                                 icsk->icsk_ack.ato = icsk->icsk_rto;
698                 } else if (m > icsk->icsk_rto) {
699                         /* Too long gap. Apparently sender failed to
700                          * restart window, so that we send ACKs quickly.
701                          */
702                         tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
703                         sk_mem_reclaim(sk);
704                 }
705         }
706         icsk->icsk_ack.lrcvtime = now;
707 
708         tcp_ecn_check_ce(sk, skb);
709 
710         if (skb->len >= 128)
711                 tcp_grow_window(sk, skb);
712 }
713 
714 /* Called to compute a smoothed rtt estimate. The data fed to this
715  * routine either comes from timestamps, or from segments that were
716  * known _not_ to have been retransmitted [see Karn/Partridge
717  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
718  * piece by Van Jacobson.
719  * NOTE: the next three routines used to be one big routine.
720  * To save cycles in the RFC 1323 implementation it was better to break
721  * it up into three procedures. -- erics
722  */
723 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
724 {
725         struct tcp_sock *tp = tcp_sk(sk);
726         long m = mrtt_us; /* RTT */
727         u32 srtt = tp->srtt_us;
728 
729         /*      The following amusing code comes from Jacobson's
730          *      article in SIGCOMM '88.  Note that rtt and mdev
731          *      are scaled versions of rtt and mean deviation.
732          *      This is designed to be as fast as possible
733          *      m stands for "measurement".
734          *
735          *      On a 1990 paper the rto value is changed to:
736          *      RTO = rtt + 4 * mdev
737          *
738          * Funny. This algorithm seems to be very broken.
739          * These formulae increase RTO, when it should be decreased, increase
740          * too slowly, when it should be increased quickly, decrease too quickly
741          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
742          * does not matter how to _calculate_ it. Seems, it was trap
743          * that VJ failed to avoid. 8)
744          */
745         if (srtt != 0) {
746                 m -= (srtt >> 3);       /* m is now error in rtt est */
747                 srtt += m;              /* rtt = 7/8 rtt + 1/8 new */
748                 if (m < 0) {
749                         m = -m;         /* m is now abs(error) */
750                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
751                         /* This is similar to one of Eifel findings.
752                          * Eifel blocks mdev updates when rtt decreases.
753                          * This solution is a bit different: we use finer gain
754                          * for mdev in this case (alpha*beta).
755                          * Like Eifel it also prevents growth of rto,
756                          * but also it limits too fast rto decreases,
757                          * happening in pure Eifel.
758                          */
759                         if (m > 0)
760                                 m >>= 3;
761                 } else {
762                         m -= (tp->mdev_us >> 2);   /* similar update on mdev */
763                 }
764                 tp->mdev_us += m;               /* mdev = 3/4 mdev + 1/4 new */
765                 if (tp->mdev_us > tp->mdev_max_us) {
766                         tp->mdev_max_us = tp->mdev_us;
767                         if (tp->mdev_max_us > tp->rttvar_us)
768                                 tp->rttvar_us = tp->mdev_max_us;
769                 }
770                 if (after(tp->snd_una, tp->rtt_seq)) {
771                         if (tp->mdev_max_us < tp->rttvar_us)
772                                 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
773                         tp->rtt_seq = tp->snd_nxt;
774                         tp->mdev_max_us = tcp_rto_min_us(sk);
775                 }
776         } else {
777                 /* no previous measure. */
778                 srtt = m << 3;          /* take the measured time to be rtt */
779                 tp->mdev_us = m << 1;   /* make sure rto = 3*rtt */
780                 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
781                 tp->mdev_max_us = tp->rttvar_us;
782                 tp->rtt_seq = tp->snd_nxt;
783         }
784         tp->srtt_us = max(1U, srtt);
785 }
786 
787 static void tcp_update_pacing_rate(struct sock *sk)
788 {
789         const struct tcp_sock *tp = tcp_sk(sk);
790         u64 rate;
791 
792         /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
793         rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
794 
795         /* current rate is (cwnd * mss) / srtt
796          * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
797          * In Congestion Avoidance phase, set it to 120 % the current rate.
798          *
799          * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
800          *       If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
801          *       end of slow start and should slow down.
802          */
803         if (tp->snd_cwnd < tp->snd_ssthresh / 2)
804                 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
805         else
806                 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
807 
808         rate *= max(tp->snd_cwnd, tp->packets_out);
809 
810         if (likely(tp->srtt_us))
811                 do_div(rate, tp->srtt_us);
812 
813         /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
814          * without any lock. We want to make sure compiler wont store
815          * intermediate values in this location.
816          */
817         WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
818                                              sk->sk_max_pacing_rate));
819 }
820 
821 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
822  * routine referred to above.
823  */
824 static void tcp_set_rto(struct sock *sk)
825 {
826         const struct tcp_sock *tp = tcp_sk(sk);
827         /* Old crap is replaced with new one. 8)
828          *
829          * More seriously:
830          * 1. If rtt variance happened to be less 50msec, it is hallucination.
831          *    It cannot be less due to utterly erratic ACK generation made
832          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
833          *    to do with delayed acks, because at cwnd>2 true delack timeout
834          *    is invisible. Actually, Linux-2.4 also generates erratic
835          *    ACKs in some circumstances.
836          */
837         inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
838 
839         /* 2. Fixups made earlier cannot be right.
840          *    If we do not estimate RTO correctly without them,
841          *    all the algo is pure shit and should be replaced
842          *    with correct one. It is exactly, which we pretend to do.
843          */
844 
845         /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
846          * guarantees that rto is higher.
847          */
848         tcp_bound_rto(sk);
849 }
850 
851 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
852 {
853         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
854 
855         if (!cwnd)
856                 cwnd = TCP_INIT_CWND;
857         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
858 }
859 
860 /* Take a notice that peer is sending D-SACKs */
861 static void tcp_dsack_seen(struct tcp_sock *tp)
862 {
863         tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
864         tp->rack.dsack_seen = 1;
865         tp->dsack_dups++;
866 }
867 
868 /* It's reordering when higher sequence was delivered (i.e. sacked) before
869  * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
870  * distance is approximated in full-mss packet distance ("reordering").
871  */
872 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
873                                       const int ts)
874 {
875         struct tcp_sock *tp = tcp_sk(sk);
876         const u32 mss = tp->mss_cache;
877         u32 fack, metric;
878 
879         fack = tcp_highest_sack_seq(tp);
880         if (!before(low_seq, fack))
881                 return;
882 
883         metric = fack - low_seq;
884         if ((metric > tp->reordering * mss) && mss) {
885 #if FASTRETRANS_DEBUG > 1
886                 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
887                          tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
888                          tp->reordering,
889                          0,
890                          tp->sacked_out,
891                          tp->undo_marker ? tp->undo_retrans : 0);
892 #endif
893                 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
894                                        sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
895         }
896 
897         /* This exciting event is worth to be remembered. 8) */
898         tp->reord_seen++;
899         NET_INC_STATS(sock_net(sk),
900                       ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
901 }
902 
903 /* This must be called before lost_out is incremented */
904 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
905 {
906         if (!tp->retransmit_skb_hint ||
907             before(TCP_SKB_CB(skb)->seq,
908                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
909                 tp->retransmit_skb_hint = skb;
910 }
911 
912 /* Sum the number of packets on the wire we have marked as lost.
913  * There are two cases we care about here:
914  * a) Packet hasn't been marked lost (nor retransmitted),
915  *    and this is the first loss.
916  * b) Packet has been marked both lost and retransmitted,
917  *    and this means we think it was lost again.
918  */
919 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
920 {
921         __u8 sacked = TCP_SKB_CB(skb)->sacked;
922 
923         if (!(sacked & TCPCB_LOST) ||
924             ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
925                 tp->lost += tcp_skb_pcount(skb);
926 }
927 
928 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
929 {
930         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
931                 tcp_verify_retransmit_hint(tp, skb);
932 
933                 tp->lost_out += tcp_skb_pcount(skb);
934                 tcp_sum_lost(tp, skb);
935                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
936         }
937 }
938 
939 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
940 {
941         tcp_verify_retransmit_hint(tp, skb);
942 
943         tcp_sum_lost(tp, skb);
944         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
945                 tp->lost_out += tcp_skb_pcount(skb);
946                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
947         }
948 }
949 
950 /* This procedure tags the retransmission queue when SACKs arrive.
951  *
952  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
953  * Packets in queue with these bits set are counted in variables
954  * sacked_out, retrans_out and lost_out, correspondingly.
955  *
956  * Valid combinations are:
957  * Tag  InFlight        Description
958  * 0    1               - orig segment is in flight.
959  * S    0               - nothing flies, orig reached receiver.
960  * L    0               - nothing flies, orig lost by net.
961  * R    2               - both orig and retransmit are in flight.
962  * L|R  1               - orig is lost, retransmit is in flight.
963  * S|R  1               - orig reached receiver, retrans is still in flight.
964  * (L|S|R is logically valid, it could occur when L|R is sacked,
965  *  but it is equivalent to plain S and code short-curcuits it to S.
966  *  L|S is logically invalid, it would mean -1 packet in flight 8))
967  *
968  * These 6 states form finite state machine, controlled by the following events:
969  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
970  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
971  * 3. Loss detection event of two flavors:
972  *      A. Scoreboard estimator decided the packet is lost.
973  *         A'. Reno "three dupacks" marks head of queue lost.
974  *      B. SACK arrives sacking SND.NXT at the moment, when the
975  *         segment was retransmitted.
976  * 4. D-SACK added new rule: D-SACK changes any tag to S.
977  *
978  * It is pleasant to note, that state diagram turns out to be commutative,
979  * so that we are allowed not to be bothered by order of our actions,
980  * when multiple events arrive simultaneously. (see the function below).
981  *
982  * Reordering detection.
983  * --------------------
984  * Reordering metric is maximal distance, which a packet can be displaced
985  * in packet stream. With SACKs we can estimate it:
986  *
987  * 1. SACK fills old hole and the corresponding segment was not
988  *    ever retransmitted -> reordering. Alas, we cannot use it
989  *    when segment was retransmitted.
990  * 2. The last flaw is solved with D-SACK. D-SACK arrives
991  *    for retransmitted and already SACKed segment -> reordering..
992  * Both of these heuristics are not used in Loss state, when we cannot
993  * account for retransmits accurately.
994  *
995  * SACK block validation.
996  * ----------------------
997  *
998  * SACK block range validation checks that the received SACK block fits to
999  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1000  * Note that SND.UNA is not included to the range though being valid because
1001  * it means that the receiver is rather inconsistent with itself reporting
1002  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1003  * perfectly valid, however, in light of RFC2018 which explicitly states
1004  * that "SACK block MUST reflect the newest segment.  Even if the newest
1005  * segment is going to be discarded ...", not that it looks very clever
1006  * in case of head skb. Due to potentional receiver driven attacks, we
1007  * choose to avoid immediate execution of a walk in write queue due to
1008  * reneging and defer head skb's loss recovery to standard loss recovery
1009  * procedure that will eventually trigger (nothing forbids us doing this).
1010  *
1011  * Implements also blockage to start_seq wrap-around. Problem lies in the
1012  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1013  * there's no guarantee that it will be before snd_nxt (n). The problem
1014  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1015  * wrap (s_w):
1016  *
1017  *         <- outs wnd ->                          <- wrapzone ->
1018  *         u     e      n                         u_w   e_w  s n_w
1019  *         |     |      |                          |     |   |  |
1020  * |<------------+------+----- TCP seqno space --------------+---------->|
1021  * ...-- <2^31 ->|                                           |<--------...
1022  * ...---- >2^31 ------>|                                    |<--------...
1023  *
1024  * Current code wouldn't be vulnerable but it's better still to discard such
1025  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1026  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1027  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1028  * equal to the ideal case (infinite seqno space without wrap caused issues).
1029  *
1030  * With D-SACK the lower bound is extended to cover sequence space below
1031  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1032  * again, D-SACK block must not to go across snd_una (for the same reason as
1033  * for the normal SACK blocks, explained above). But there all simplicity
1034  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1035  * fully below undo_marker they do not affect behavior in anyway and can
1036  * therefore be safely ignored. In rare cases (which are more or less
1037  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1038  * fragmentation and packet reordering past skb's retransmission. To consider
1039  * them correctly, the acceptable range must be extended even more though
1040  * the exact amount is rather hard to quantify. However, tp->max_window can
1041  * be used as an exaggerated estimate.
1042  */
1043 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1044                                    u32 start_seq, u32 end_seq)
1045 {
1046         /* Too far in future, or reversed (interpretation is ambiguous) */
1047         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1048                 return false;
1049 
1050         /* Nasty start_seq wrap-around check (see comments above) */
1051         if (!before(start_seq, tp->snd_nxt))
1052                 return false;
1053 
1054         /* In outstanding window? ...This is valid exit for D-SACKs too.
1055          * start_seq == snd_una is non-sensical (see comments above)
1056          */
1057         if (after(start_seq, tp->snd_una))
1058                 return true;
1059 
1060         if (!is_dsack || !tp->undo_marker)
1061                 return false;
1062 
1063         /* ...Then it's D-SACK, and must reside below snd_una completely */
1064         if (after(end_seq, tp->snd_una))
1065                 return false;
1066 
1067         if (!before(start_seq, tp->undo_marker))
1068                 return true;
1069 
1070         /* Too old */
1071         if (!after(end_seq, tp->undo_marker))
1072                 return false;
1073 
1074         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1075          *   start_seq < undo_marker and end_seq >= undo_marker.
1076          */
1077         return !before(start_seq, end_seq - tp->max_window);
1078 }
1079 
1080 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1081                             struct tcp_sack_block_wire *sp, int num_sacks,
1082                             u32 prior_snd_una)
1083 {
1084         struct tcp_sock *tp = tcp_sk(sk);
1085         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1086         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1087         bool dup_sack = false;
1088 
1089         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1090                 dup_sack = true;
1091                 tcp_dsack_seen(tp);
1092                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1093         } else if (num_sacks > 1) {
1094                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1095                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1096 
1097                 if (!after(end_seq_0, end_seq_1) &&
1098                     !before(start_seq_0, start_seq_1)) {
1099                         dup_sack = true;
1100                         tcp_dsack_seen(tp);
1101                         NET_INC_STATS(sock_net(sk),
1102                                         LINUX_MIB_TCPDSACKOFORECV);
1103                 }
1104         }
1105 
1106         /* D-SACK for already forgotten data... Do dumb counting. */
1107         if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1108             !after(end_seq_0, prior_snd_una) &&
1109             after(end_seq_0, tp->undo_marker))
1110                 tp->undo_retrans--;
1111 
1112         return dup_sack;
1113 }
1114 
1115 struct tcp_sacktag_state {
1116         u32     reord;
1117         /* Timestamps for earliest and latest never-retransmitted segment
1118          * that was SACKed. RTO needs the earliest RTT to stay conservative,
1119          * but congestion control should still get an accurate delay signal.
1120          */
1121         u64     first_sackt;
1122         u64     last_sackt;
1123         struct rate_sample *rate;
1124         int     flag;
1125         unsigned int mss_now;
1126 };
1127 
1128 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1129  * the incoming SACK may not exactly match but we can find smaller MSS
1130  * aligned portion of it that matches. Therefore we might need to fragment
1131  * which may fail and creates some hassle (caller must handle error case
1132  * returns).
1133  *
1134  * FIXME: this could be merged to shift decision code
1135  */
1136 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1137                                   u32 start_seq, u32 end_seq)
1138 {
1139         int err;
1140         bool in_sack;
1141         unsigned int pkt_len;
1142         unsigned int mss;
1143 
1144         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1145                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1146 
1147         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1148             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1149                 mss = tcp_skb_mss(skb);
1150                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1151 
1152                 if (!in_sack) {
1153                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1154                         if (pkt_len < mss)
1155                                 pkt_len = mss;
1156                 } else {
1157                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1158                         if (pkt_len < mss)
1159                                 return -EINVAL;
1160                 }
1161 
1162                 /* Round if necessary so that SACKs cover only full MSSes
1163                  * and/or the remaining small portion (if present)
1164                  */
1165                 if (pkt_len > mss) {
1166                         unsigned int new_len = (pkt_len / mss) * mss;
1167                         if (!in_sack && new_len < pkt_len)
1168                                 new_len += mss;
1169                         pkt_len = new_len;
1170                 }
1171 
1172                 if (pkt_len >= skb->len && !in_sack)
1173                         return 0;
1174 
1175                 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1176                                    pkt_len, mss, GFP_ATOMIC);
1177                 if (err < 0)
1178                         return err;
1179         }
1180 
1181         return in_sack;
1182 }
1183 
1184 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1185 static u8 tcp_sacktag_one(struct sock *sk,
1186                           struct tcp_sacktag_state *state, u8 sacked,
1187                           u32 start_seq, u32 end_seq,
1188                           int dup_sack, int pcount,
1189                           u64 xmit_time)
1190 {
1191         struct tcp_sock *tp = tcp_sk(sk);
1192 
1193         /* Account D-SACK for retransmitted packet. */
1194         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1195                 if (tp->undo_marker && tp->undo_retrans > 0 &&
1196                     after(end_seq, tp->undo_marker))
1197                         tp->undo_retrans--;
1198                 if ((sacked & TCPCB_SACKED_ACKED) &&
1199                     before(start_seq, state->reord))
1200                                 state->reord = start_seq;
1201         }
1202 
1203         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1204         if (!after(end_seq, tp->snd_una))
1205                 return sacked;
1206 
1207         if (!(sacked & TCPCB_SACKED_ACKED)) {
1208                 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1209 
1210                 if (sacked & TCPCB_SACKED_RETRANS) {
1211                         /* If the segment is not tagged as lost,
1212                          * we do not clear RETRANS, believing
1213                          * that retransmission is still in flight.
1214                          */
1215                         if (sacked & TCPCB_LOST) {
1216                                 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1217                                 tp->lost_out -= pcount;
1218                                 tp->retrans_out -= pcount;
1219                         }
1220                 } else {
1221                         if (!(sacked & TCPCB_RETRANS)) {
1222                                 /* New sack for not retransmitted frame,
1223                                  * which was in hole. It is reordering.
1224                                  */
1225                                 if (before(start_seq,
1226                                            tcp_highest_sack_seq(tp)) &&
1227                                     before(start_seq, state->reord))
1228                                         state->reord = start_seq;
1229 
1230                                 if (!after(end_seq, tp->high_seq))
1231                                         state->flag |= FLAG_ORIG_SACK_ACKED;
1232                                 if (state->first_sackt == 0)
1233                                         state->first_sackt = xmit_time;
1234                                 state->last_sackt = xmit_time;
1235                         }
1236 
1237                         if (sacked & TCPCB_LOST) {
1238                                 sacked &= ~TCPCB_LOST;
1239                                 tp->lost_out -= pcount;
1240                         }
1241                 }
1242 
1243                 sacked |= TCPCB_SACKED_ACKED;
1244                 state->flag |= FLAG_DATA_SACKED;
1245                 tp->sacked_out += pcount;
1246                 tp->delivered += pcount;  /* Out-of-order packets delivered */
1247 
1248                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1249                 if (tp->lost_skb_hint &&
1250                     before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1251                         tp->lost_cnt_hint += pcount;
1252         }
1253 
1254         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1255          * frames and clear it. undo_retrans is decreased above, L|R frames
1256          * are accounted above as well.
1257          */
1258         if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1259                 sacked &= ~TCPCB_SACKED_RETRANS;
1260                 tp->retrans_out -= pcount;
1261         }
1262 
1263         return sacked;
1264 }
1265 
1266 /* Shift newly-SACKed bytes from this skb to the immediately previous
1267  * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1268  */
1269 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1270                             struct sk_buff *skb,
1271                             struct tcp_sacktag_state *state,
1272                             unsigned int pcount, int shifted, int mss,
1273                             bool dup_sack)
1274 {
1275         struct tcp_sock *tp = tcp_sk(sk);
1276         u32 start_seq = TCP_SKB_CB(skb)->seq;   /* start of newly-SACKed */
1277         u32 end_seq = start_seq + shifted;      /* end of newly-SACKed */
1278 
1279         BUG_ON(!pcount);
1280 
1281         /* Adjust counters and hints for the newly sacked sequence
1282          * range but discard the return value since prev is already
1283          * marked. We must tag the range first because the seq
1284          * advancement below implicitly advances
1285          * tcp_highest_sack_seq() when skb is highest_sack.
1286          */
1287         tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1288                         start_seq, end_seq, dup_sack, pcount,
1289                         tcp_skb_timestamp_us(skb));
1290         tcp_rate_skb_delivered(sk, skb, state->rate);
1291 
1292         if (skb == tp->lost_skb_hint)
1293                 tp->lost_cnt_hint += pcount;
1294 
1295         TCP_SKB_CB(prev)->end_seq += shifted;
1296         TCP_SKB_CB(skb)->seq += shifted;
1297 
1298         tcp_skb_pcount_add(prev, pcount);
1299         WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1300         tcp_skb_pcount_add(skb, -pcount);
1301 
1302         /* When we're adding to gso_segs == 1, gso_size will be zero,
1303          * in theory this shouldn't be necessary but as long as DSACK
1304          * code can come after this skb later on it's better to keep
1305          * setting gso_size to something.
1306          */
1307         if (!TCP_SKB_CB(prev)->tcp_gso_size)
1308                 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1309 
1310         /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1311         if (tcp_skb_pcount(skb) <= 1)
1312                 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1313 
1314         /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1315         TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1316 
1317         if (skb->len > 0) {
1318                 BUG_ON(!tcp_skb_pcount(skb));
1319                 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1320                 return false;
1321         }
1322 
1323         /* Whole SKB was eaten :-) */
1324 
1325         if (skb == tp->retransmit_skb_hint)
1326                 tp->retransmit_skb_hint = prev;
1327         if (skb == tp->lost_skb_hint) {
1328                 tp->lost_skb_hint = prev;
1329                 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1330         }
1331 
1332         TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1333         TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1334         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1335                 TCP_SKB_CB(prev)->end_seq++;
1336 
1337         if (skb == tcp_highest_sack(sk))
1338                 tcp_advance_highest_sack(sk, skb);
1339 
1340         tcp_skb_collapse_tstamp(prev, skb);
1341         if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1342                 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1343 
1344         tcp_rtx_queue_unlink_and_free(skb, sk);
1345 
1346         NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1347 
1348         return true;
1349 }
1350 
1351 /* I wish gso_size would have a bit more sane initialization than
1352  * something-or-zero which complicates things
1353  */
1354 static int tcp_skb_seglen(const struct sk_buff *skb)
1355 {
1356         return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1357 }
1358 
1359 /* Shifting pages past head area doesn't work */
1360 static int skb_can_shift(const struct sk_buff *skb)
1361 {
1362         return !skb_headlen(skb) && skb_is_nonlinear(skb);
1363 }
1364 
1365 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1366                   int pcount, int shiftlen)
1367 {
1368         /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1369          * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1370          * to make sure not storing more than 65535 * 8 bytes per skb,
1371          * even if current MSS is bigger.
1372          */
1373         if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1374                 return 0;
1375         if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1376                 return 0;
1377         return skb_shift(to, from, shiftlen);
1378 }
1379 
1380 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1381  * skb.
1382  */
1383 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1384                                           struct tcp_sacktag_state *state,
1385                                           u32 start_seq, u32 end_seq,
1386                                           bool dup_sack)
1387 {
1388         struct tcp_sock *tp = tcp_sk(sk);
1389         struct sk_buff *prev;
1390         int mss;
1391         int pcount = 0;
1392         int len;
1393         int in_sack;
1394 
1395         /* Normally R but no L won't result in plain S */
1396         if (!dup_sack &&
1397             (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1398                 goto fallback;
1399         if (!skb_can_shift(skb))
1400                 goto fallback;
1401         /* This frame is about to be dropped (was ACKed). */
1402         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1403                 goto fallback;
1404 
1405         /* Can only happen with delayed DSACK + discard craziness */
1406         prev = skb_rb_prev(skb);
1407         if (!prev)
1408                 goto fallback;
1409 
1410         if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1411                 goto fallback;
1412 
1413         if (!tcp_skb_can_collapse_to(prev))
1414                 goto fallback;
1415 
1416         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1417                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1418 
1419         if (in_sack) {
1420                 len = skb->len;
1421                 pcount = tcp_skb_pcount(skb);
1422                 mss = tcp_skb_seglen(skb);
1423 
1424                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1425                  * drop this restriction as unnecessary
1426                  */
1427                 if (mss != tcp_skb_seglen(prev))
1428                         goto fallback;
1429         } else {
1430                 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1431                         goto noop;
1432                 /* CHECKME: This is non-MSS split case only?, this will
1433                  * cause skipped skbs due to advancing loop btw, original
1434                  * has that feature too
1435                  */
1436                 if (tcp_skb_pcount(skb) <= 1)
1437                         goto noop;
1438 
1439                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1440                 if (!in_sack) {
1441                         /* TODO: head merge to next could be attempted here
1442                          * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1443                          * though it might not be worth of the additional hassle
1444                          *
1445                          * ...we can probably just fallback to what was done
1446                          * previously. We could try merging non-SACKed ones
1447                          * as well but it probably isn't going to buy off
1448                          * because later SACKs might again split them, and
1449                          * it would make skb timestamp tracking considerably
1450                          * harder problem.
1451                          */
1452                         goto fallback;
1453                 }
1454 
1455                 len = end_seq - TCP_SKB_CB(skb)->seq;
1456                 BUG_ON(len < 0);
1457                 BUG_ON(len > skb->len);
1458 
1459                 /* MSS boundaries should be honoured or else pcount will
1460                  * severely break even though it makes things bit trickier.
1461                  * Optimize common case to avoid most of the divides
1462                  */
1463                 mss = tcp_skb_mss(skb);
1464 
1465                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1466                  * drop this restriction as unnecessary
1467                  */
1468                 if (mss != tcp_skb_seglen(prev))
1469                         goto fallback;
1470 
1471                 if (len == mss) {
1472                         pcount = 1;
1473                 } else if (len < mss) {
1474                         goto noop;
1475                 } else {
1476                         pcount = len / mss;
1477                         len = pcount * mss;
1478                 }
1479         }
1480 
1481         /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1482         if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1483                 goto fallback;
1484 
1485         if (!tcp_skb_shift(prev, skb, pcount, len))
1486                 goto fallback;
1487         if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1488                 goto out;
1489 
1490         /* Hole filled allows collapsing with the next as well, this is very
1491          * useful when hole on every nth skb pattern happens
1492          */
1493         skb = skb_rb_next(prev);
1494         if (!skb)
1495                 goto out;
1496 
1497         if (!skb_can_shift(skb) ||
1498             ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1499             (mss != tcp_skb_seglen(skb)))
1500                 goto out;
1501 
1502         len = skb->len;
1503         pcount = tcp_skb_pcount(skb);
1504         if (tcp_skb_shift(prev, skb, pcount, len))
1505                 tcp_shifted_skb(sk, prev, skb, state, pcount,
1506                                 len, mss, 0);
1507 
1508 out:
1509         return prev;
1510 
1511 noop:
1512         return skb;
1513 
1514 fallback:
1515         NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1516         return NULL;
1517 }
1518 
1519 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1520                                         struct tcp_sack_block *next_dup,
1521                                         struct tcp_sacktag_state *state,
1522                                         u32 start_seq, u32 end_seq,
1523                                         bool dup_sack_in)
1524 {
1525         struct tcp_sock *tp = tcp_sk(sk);
1526         struct sk_buff *tmp;
1527 
1528         skb_rbtree_walk_from(skb) {
1529                 int in_sack = 0;
1530                 bool dup_sack = dup_sack_in;
1531 
1532                 /* queue is in-order => we can short-circuit the walk early */
1533                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1534                         break;
1535 
1536                 if (next_dup  &&
1537                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1538                         in_sack = tcp_match_skb_to_sack(sk, skb,
1539                                                         next_dup->start_seq,
1540                                                         next_dup->end_seq);
1541                         if (in_sack > 0)
1542                                 dup_sack = true;
1543                 }
1544 
1545                 /* skb reference here is a bit tricky to get right, since
1546                  * shifting can eat and free both this skb and the next,
1547                  * so not even _safe variant of the loop is enough.
1548                  */
1549                 if (in_sack <= 0) {
1550                         tmp = tcp_shift_skb_data(sk, skb, state,
1551                                                  start_seq, end_seq, dup_sack);
1552                         if (tmp) {
1553                                 if (tmp != skb) {
1554                                         skb = tmp;
1555                                         continue;
1556                                 }
1557 
1558                                 in_sack = 0;
1559                         } else {
1560                                 in_sack = tcp_match_skb_to_sack(sk, skb,
1561                                                                 start_seq,
1562                                                                 end_seq);
1563                         }
1564                 }
1565 
1566                 if (unlikely(in_sack < 0))
1567                         break;
1568 
1569                 if (in_sack) {
1570                         TCP_SKB_CB(skb)->sacked =
1571                                 tcp_sacktag_one(sk,
1572                                                 state,
1573                                                 TCP_SKB_CB(skb)->sacked,
1574                                                 TCP_SKB_CB(skb)->seq,
1575                                                 TCP_SKB_CB(skb)->end_seq,
1576                                                 dup_sack,
1577                                                 tcp_skb_pcount(skb),
1578                                                 tcp_skb_timestamp_us(skb));
1579                         tcp_rate_skb_delivered(sk, skb, state->rate);
1580                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1581                                 list_del_init(&skb->tcp_tsorted_anchor);
1582 
1583                         if (!before(TCP_SKB_CB(skb)->seq,
1584                                     tcp_highest_sack_seq(tp)))
1585                                 tcp_advance_highest_sack(sk, skb);
1586                 }
1587         }
1588         return skb;
1589 }
1590 
1591 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1592 {
1593         struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1594         struct sk_buff *skb;
1595 
1596         while (*p) {
1597                 parent = *p;
1598                 skb = rb_to_skb(parent);
1599                 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1600                         p = &parent->rb_left;
1601                         continue;
1602                 }
1603                 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1604                         p = &parent->rb_right;
1605                         continue;
1606                 }
1607                 return skb;
1608         }
1609         return NULL;
1610 }
1611 
1612 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1613                                         u32 skip_to_seq)
1614 {
1615         if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1616                 return skb;
1617 
1618         return tcp_sacktag_bsearch(sk, skip_to_seq);
1619 }
1620 
1621 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1622                                                 struct sock *sk,
1623                                                 struct tcp_sack_block *next_dup,
1624                                                 struct tcp_sacktag_state *state,
1625                                                 u32 skip_to_seq)
1626 {
1627         if (!next_dup)
1628                 return skb;
1629 
1630         if (before(next_dup->start_seq, skip_to_seq)) {
1631                 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1632                 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1633                                        next_dup->start_seq, next_dup->end_seq,
1634                                        1);
1635         }
1636 
1637         return skb;
1638 }
1639 
1640 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1641 {
1642         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1643 }
1644 
1645 static int
1646 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1647                         u32 prior_snd_una, struct tcp_sacktag_state *state)
1648 {
1649         struct tcp_sock *tp = tcp_sk(sk);
1650         const unsigned char *ptr = (skb_transport_header(ack_skb) +
1651                                     TCP_SKB_CB(ack_skb)->sacked);
1652         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1653         struct tcp_sack_block sp[TCP_NUM_SACKS];
1654         struct tcp_sack_block *cache;
1655         struct sk_buff *skb;
1656         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1657         int used_sacks;
1658         bool found_dup_sack = false;
1659         int i, j;
1660         int first_sack_index;
1661 
1662         state->flag = 0;
1663         state->reord = tp->snd_nxt;
1664 
1665         if (!tp->sacked_out)
1666                 tcp_highest_sack_reset(sk);
1667 
1668         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1669                                          num_sacks, prior_snd_una);
1670         if (found_dup_sack) {
1671                 state->flag |= FLAG_DSACKING_ACK;
1672                 tp->delivered++; /* A spurious retransmission is delivered */
1673         }
1674 
1675         /* Eliminate too old ACKs, but take into
1676          * account more or less fresh ones, they can
1677          * contain valid SACK info.
1678          */
1679         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1680                 return 0;
1681 
1682         if (!tp->packets_out)
1683                 goto out;
1684 
1685         used_sacks = 0;
1686         first_sack_index = 0;
1687         for (i = 0; i < num_sacks; i++) {
1688                 bool dup_sack = !i && found_dup_sack;
1689 
1690                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1691                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1692 
1693                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1694                                             sp[used_sacks].start_seq,
1695                                             sp[used_sacks].end_seq)) {
1696                         int mib_idx;
1697 
1698                         if (dup_sack) {
1699                                 if (!tp->undo_marker)
1700                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1701                                 else
1702                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1703                         } else {
1704                                 /* Don't count olds caused by ACK reordering */
1705                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1706                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1707                                         continue;
1708                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1709                         }
1710 
1711                         NET_INC_STATS(sock_net(sk), mib_idx);
1712                         if (i == 0)
1713                                 first_sack_index = -1;
1714                         continue;
1715                 }
1716 
1717                 /* Ignore very old stuff early */
1718                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1719                         continue;
1720 
1721                 used_sacks++;
1722         }
1723 
1724         /* order SACK blocks to allow in order walk of the retrans queue */
1725         for (i = used_sacks - 1; i > 0; i--) {
1726                 for (j = 0; j < i; j++) {
1727                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1728                                 swap(sp[j], sp[j + 1]);
1729 
1730                                 /* Track where the first SACK block goes to */
1731                                 if (j == first_sack_index)
1732                                         first_sack_index = j + 1;
1733                         }
1734                 }
1735         }
1736 
1737         state->mss_now = tcp_current_mss(sk);
1738         skb = NULL;
1739         i = 0;
1740 
1741         if (!tp->sacked_out) {
1742                 /* It's already past, so skip checking against it */
1743                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1744         } else {
1745                 cache = tp->recv_sack_cache;
1746                 /* Skip empty blocks in at head of the cache */
1747                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1748                        !cache->end_seq)
1749                         cache++;
1750         }
1751 
1752         while (i < used_sacks) {
1753                 u32 start_seq = sp[i].start_seq;
1754                 u32 end_seq = sp[i].end_seq;
1755                 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1756                 struct tcp_sack_block *next_dup = NULL;
1757 
1758                 if (found_dup_sack && ((i + 1) == first_sack_index))
1759                         next_dup = &sp[i + 1];
1760 
1761                 /* Skip too early cached blocks */
1762                 while (tcp_sack_cache_ok(tp, cache) &&
1763                        !before(start_seq, cache->end_seq))
1764                         cache++;
1765 
1766                 /* Can skip some work by looking recv_sack_cache? */
1767                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1768                     after(end_seq, cache->start_seq)) {
1769 
1770                         /* Head todo? */
1771                         if (before(start_seq, cache->start_seq)) {
1772                                 skb = tcp_sacktag_skip(skb, sk, start_seq);
1773                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1774                                                        state,
1775                                                        start_seq,
1776                                                        cache->start_seq,
1777                                                        dup_sack);
1778                         }
1779 
1780                         /* Rest of the block already fully processed? */
1781                         if (!after(end_seq, cache->end_seq))
1782                                 goto advance_sp;
1783 
1784                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1785                                                        state,
1786                                                        cache->end_seq);
1787 
1788                         /* ...tail remains todo... */
1789                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1790                                 /* ...but better entrypoint exists! */
1791                                 skb = tcp_highest_sack(sk);
1792                                 if (!skb)
1793                                         break;
1794                                 cache++;
1795                                 goto walk;
1796                         }
1797 
1798                         skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1799                         /* Check overlap against next cached too (past this one already) */
1800                         cache++;
1801                         continue;
1802                 }
1803 
1804                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1805                         skb = tcp_highest_sack(sk);
1806                         if (!skb)
1807                                 break;
1808                 }
1809                 skb = tcp_sacktag_skip(skb, sk, start_seq);
1810 
1811 walk:
1812                 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1813                                        start_seq, end_seq, dup_sack);
1814 
1815 advance_sp:
1816                 i++;
1817         }
1818 
1819         /* Clear the head of the cache sack blocks so we can skip it next time */
1820         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1821                 tp->recv_sack_cache[i].start_seq = 0;
1822                 tp->recv_sack_cache[i].end_seq = 0;
1823         }
1824         for (j = 0; j < used_sacks; j++)
1825                 tp->recv_sack_cache[i++] = sp[j];
1826 
1827         if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1828                 tcp_check_sack_reordering(sk, state->reord, 0);
1829 
1830         tcp_verify_left_out(tp);
1831 out:
1832 
1833 #if FASTRETRANS_DEBUG > 0
1834         WARN_ON((int)tp->sacked_out < 0);
1835         WARN_ON((int)tp->lost_out < 0);
1836         WARN_ON((int)tp->retrans_out < 0);
1837         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1838 #endif
1839         return state->flag;
1840 }
1841 
1842 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1843  * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1844  */
1845 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1846 {
1847         u32 holes;
1848 
1849         holes = max(tp->lost_out, 1U);
1850         holes = min(holes, tp->packets_out);
1851 
1852         if ((tp->sacked_out + holes) > tp->packets_out) {
1853                 tp->sacked_out = tp->packets_out - holes;
1854                 return true;
1855         }
1856         return false;
1857 }
1858 
1859 /* If we receive more dupacks than we expected counting segments
1860  * in assumption of absent reordering, interpret this as reordering.
1861  * The only another reason could be bug in receiver TCP.
1862  */
1863 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1864 {
1865         struct tcp_sock *tp = tcp_sk(sk);
1866 
1867         if (!tcp_limit_reno_sacked(tp))
1868                 return;
1869 
1870         tp->reordering = min_t(u32, tp->packets_out + addend,
1871                                sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1872         tp->reord_seen++;
1873         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1874 }
1875 
1876 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1877 
1878 static void tcp_add_reno_sack(struct sock *sk, int num_dupack)
1879 {
1880         if (num_dupack) {
1881                 struct tcp_sock *tp = tcp_sk(sk);
1882                 u32 prior_sacked = tp->sacked_out;
1883                 s32 delivered;
1884 
1885                 tp->sacked_out += num_dupack;
1886                 tcp_check_reno_reordering(sk, 0);
1887                 delivered = tp->sacked_out - prior_sacked;
1888                 if (delivered > 0)
1889                         tp->delivered += delivered;
1890                 tcp_verify_left_out(tp);
1891         }
1892 }
1893 
1894 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1895 
1896 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1897 {
1898         struct tcp_sock *tp = tcp_sk(sk);
1899 
1900         if (acked > 0) {
1901                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1902                 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1903                 if (acked - 1 >= tp->sacked_out)
1904                         tp->sacked_out = 0;
1905                 else
1906                         tp->sacked_out -= acked - 1;
1907         }
1908         tcp_check_reno_reordering(sk, acked);
1909         tcp_verify_left_out(tp);
1910 }
1911 
1912 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1913 {
1914         tp->sacked_out = 0;
1915 }
1916 
1917 void tcp_clear_retrans(struct tcp_sock *tp)
1918 {
1919         tp->retrans_out = 0;
1920         tp->lost_out = 0;
1921         tp->undo_marker = 0;
1922         tp->undo_retrans = -1;
1923         tp->sacked_out = 0;
1924 }
1925 
1926 static inline void tcp_init_undo(struct tcp_sock *tp)
1927 {
1928         tp->undo_marker = tp->snd_una;
1929         /* Retransmission still in flight may cause DSACKs later. */
1930         tp->undo_retrans = tp->retrans_out ? : -1;
1931 }
1932 
1933 static bool tcp_is_rack(const struct sock *sk)
1934 {
1935         return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
1936 }
1937 
1938 /* If we detect SACK reneging, forget all SACK information
1939  * and reset tags completely, otherwise preserve SACKs. If receiver
1940  * dropped its ofo queue, we will know this due to reneging detection.
1941  */
1942 static void tcp_timeout_mark_lost(struct sock *sk)
1943 {
1944         struct tcp_sock *tp = tcp_sk(sk);
1945         struct sk_buff *skb, *head;
1946         bool is_reneg;                  /* is receiver reneging on SACKs? */
1947 
1948         head = tcp_rtx_queue_head(sk);
1949         is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
1950         if (is_reneg) {
1951                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1952                 tp->sacked_out = 0;
1953                 /* Mark SACK reneging until we recover from this loss event. */
1954                 tp->is_sack_reneg = 1;
1955         } else if (tcp_is_reno(tp)) {
1956                 tcp_reset_reno_sack(tp);
1957         }
1958 
1959         skb = head;
1960         skb_rbtree_walk_from(skb) {
1961                 if (is_reneg)
1962                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1963                 else if (tcp_is_rack(sk) && skb != head &&
1964                          tcp_rack_skb_timeout(tp, skb, 0) > 0)
1965                         continue; /* Don't mark recently sent ones lost yet */
1966                 tcp_mark_skb_lost(sk, skb);
1967         }
1968         tcp_verify_left_out(tp);
1969         tcp_clear_all_retrans_hints(tp);
1970 }
1971 
1972 /* Enter Loss state. */
1973 void tcp_enter_loss(struct sock *sk)
1974 {
1975         const struct inet_connection_sock *icsk = inet_csk(sk);
1976         struct tcp_sock *tp = tcp_sk(sk);
1977         struct net *net = sock_net(sk);
1978         bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1979 
1980         tcp_timeout_mark_lost(sk);
1981 
1982         /* Reduce ssthresh if it has not yet been made inside this window. */
1983         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1984             !after(tp->high_seq, tp->snd_una) ||
1985             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1986                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1987                 tp->prior_cwnd = tp->snd_cwnd;
1988                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1989                 tcp_ca_event(sk, CA_EVENT_LOSS);
1990                 tcp_init_undo(tp);
1991         }
1992         tp->snd_cwnd       = tcp_packets_in_flight(tp) + 1;
1993         tp->snd_cwnd_cnt   = 0;
1994         tp->snd_cwnd_stamp = tcp_jiffies32;
1995 
1996         /* Timeout in disordered state after receiving substantial DUPACKs
1997          * suggests that the degree of reordering is over-estimated.
1998          */
1999         if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2000             tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
2001                 tp->reordering = min_t(unsigned int, tp->reordering,
2002                                        net->ipv4.sysctl_tcp_reordering);
2003         tcp_set_ca_state(sk, TCP_CA_Loss);
2004         tp->high_seq = tp->snd_nxt;
2005         tcp_ecn_queue_cwr(tp);
2006 
2007         /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2008          * loss recovery is underway except recurring timeout(s) on
2009          * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2010          */
2011         tp->frto = net->ipv4.sysctl_tcp_frto &&
2012                    (new_recovery || icsk->icsk_retransmits) &&
2013                    !inet_csk(sk)->icsk_mtup.probe_size;
2014 }
2015 
2016 /* If ACK arrived pointing to a remembered SACK, it means that our
2017  * remembered SACKs do not reflect real state of receiver i.e.
2018  * receiver _host_ is heavily congested (or buggy).
2019  *
2020  * To avoid big spurious retransmission bursts due to transient SACK
2021  * scoreboard oddities that look like reneging, we give the receiver a
2022  * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2023  * restore sanity to the SACK scoreboard. If the apparent reneging
2024  * persists until this RTO then we'll clear the SACK scoreboard.
2025  */
2026 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2027 {
2028         if (flag & FLAG_SACK_RENEGING) {
2029                 struct tcp_sock *tp = tcp_sk(sk);
2030                 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2031                                           msecs_to_jiffies(10));
2032 
2033                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2034                                           delay, TCP_RTO_MAX);
2035                 return true;
2036         }
2037         return false;
2038 }
2039 
2040 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2041  * counter when SACK is enabled (without SACK, sacked_out is used for
2042  * that purpose).
2043  *
2044  * With reordering, holes may still be in flight, so RFC3517 recovery
2045  * uses pure sacked_out (total number of SACKed segments) even though
2046  * it violates the RFC that uses duplicate ACKs, often these are equal
2047  * but when e.g. out-of-window ACKs or packet duplication occurs,
2048  * they differ. Since neither occurs due to loss, TCP should really
2049  * ignore them.
2050  */
2051 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2052 {
2053         return tp->sacked_out + 1;
2054 }
2055 
2056 /* Linux NewReno/SACK/ECN state machine.
2057  * --------------------------------------
2058  *
2059  * "Open"       Normal state, no dubious events, fast path.
2060  * "Disorder"   In all the respects it is "Open",
2061  *              but requires a bit more attention. It is entered when
2062  *              we see some SACKs or dupacks. It is split of "Open"
2063  *              mainly to move some processing from fast path to slow one.
2064  * "CWR"        CWND was reduced due to some Congestion Notification event.
2065  *              It can be ECN, ICMP source quench, local device congestion.
2066  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2067  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2068  *
2069  * tcp_fastretrans_alert() is entered:
2070  * - each incoming ACK, if state is not "Open"
2071  * - when arrived ACK is unusual, namely:
2072  *      * SACK
2073  *      * Duplicate ACK.
2074  *      * ECN ECE.
2075  *
2076  * Counting packets in flight is pretty simple.
2077  *
2078  *      in_flight = packets_out - left_out + retrans_out
2079  *
2080  *      packets_out is SND.NXT-SND.UNA counted in packets.
2081  *
2082  *      retrans_out is number of retransmitted segments.
2083  *
2084  *      left_out is number of segments left network, but not ACKed yet.
2085  *
2086  *              left_out = sacked_out + lost_out
2087  *
2088  *     sacked_out: Packets, which arrived to receiver out of order
2089  *                 and hence not ACKed. With SACKs this number is simply
2090  *                 amount of SACKed data. Even without SACKs
2091  *                 it is easy to give pretty reliable estimate of this number,
2092  *                 counting duplicate ACKs.
2093  *
2094  *       lost_out: Packets lost by network. TCP has no explicit
2095  *                 "loss notification" feedback from network (for now).
2096  *                 It means that this number can be only _guessed_.
2097  *                 Actually, it is the heuristics to predict lossage that
2098  *                 distinguishes different algorithms.
2099  *
2100  *      F.e. after RTO, when all the queue is considered as lost,
2101  *      lost_out = packets_out and in_flight = retrans_out.
2102  *
2103  *              Essentially, we have now a few algorithms detecting
2104  *              lost packets.
2105  *
2106  *              If the receiver supports SACK:
2107  *
2108  *              RFC6675/3517: It is the conventional algorithm. A packet is
2109  *              considered lost if the number of higher sequence packets
2110  *              SACKed is greater than or equal the DUPACK thoreshold
2111  *              (reordering). This is implemented in tcp_mark_head_lost and
2112  *              tcp_update_scoreboard.
2113  *
2114  *              RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2115  *              (2017-) that checks timing instead of counting DUPACKs.
2116  *              Essentially a packet is considered lost if it's not S/ACKed
2117  *              after RTT + reordering_window, where both metrics are
2118  *              dynamically measured and adjusted. This is implemented in
2119  *              tcp_rack_mark_lost.
2120  *
2121  *              If the receiver does not support SACK:
2122  *
2123  *              NewReno (RFC6582): in Recovery we assume that one segment
2124  *              is lost (classic Reno). While we are in Recovery and
2125  *              a partial ACK arrives, we assume that one more packet
2126  *              is lost (NewReno). This heuristics are the same in NewReno
2127  *              and SACK.
2128  *
2129  * Really tricky (and requiring careful tuning) part of algorithm
2130  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2131  * The first determines the moment _when_ we should reduce CWND and,
2132  * hence, slow down forward transmission. In fact, it determines the moment
2133  * when we decide that hole is caused by loss, rather than by a reorder.
2134  *
2135  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2136  * holes, caused by lost packets.
2137  *
2138  * And the most logically complicated part of algorithm is undo
2139  * heuristics. We detect false retransmits due to both too early
2140  * fast retransmit (reordering) and underestimated RTO, analyzing
2141  * timestamps and D-SACKs. When we detect that some segments were
2142  * retransmitted by mistake and CWND reduction was wrong, we undo
2143  * window reduction and abort recovery phase. This logic is hidden
2144  * inside several functions named tcp_try_undo_<something>.
2145  */
2146 
2147 /* This function decides, when we should leave Disordered state
2148  * and enter Recovery phase, reducing congestion window.
2149  *
2150  * Main question: may we further continue forward transmission
2151  * with the same cwnd?
2152  */
2153 static bool tcp_time_to_recover(struct sock *sk, int flag)
2154 {
2155         struct tcp_sock *tp = tcp_sk(sk);
2156 
2157         /* Trick#1: The loss is proven. */
2158         if (tp->lost_out)
2159                 return true;
2160 
2161         /* Not-A-Trick#2 : Classic rule... */
2162         if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2163                 return true;
2164 
2165         return false;
2166 }
2167 
2168 /* Detect loss in event "A" above by marking head of queue up as lost.
2169  * For non-SACK(Reno) senders, the first "packets" number of segments
2170  * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2171  * has at least tp->reordering SACKed seqments above it; "packets" refers to
2172  * the maximum SACKed segments to pass before reaching this limit.
2173  */
2174 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2175 {
2176         struct tcp_sock *tp = tcp_sk(sk);
2177         struct sk_buff *skb;
2178         int cnt, oldcnt, lost;
2179         unsigned int mss;
2180         /* Use SACK to deduce losses of new sequences sent during recovery */
2181         const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2182 
2183         WARN_ON(packets > tp->packets_out);
2184         skb = tp->lost_skb_hint;
2185         if (skb) {
2186                 /* Head already handled? */
2187                 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2188                         return;
2189                 cnt = tp->lost_cnt_hint;
2190         } else {
2191                 skb = tcp_rtx_queue_head(sk);
2192                 cnt = 0;
2193         }
2194 
2195         skb_rbtree_walk_from(skb) {
2196                 /* TODO: do this better */
2197                 /* this is not the most efficient way to do this... */
2198                 tp->lost_skb_hint = skb;
2199                 tp->lost_cnt_hint = cnt;
2200 
2201                 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2202                         break;
2203 
2204                 oldcnt = cnt;
2205                 if (tcp_is_reno(tp) ||
2206                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2207                         cnt += tcp_skb_pcount(skb);
2208 
2209                 if (cnt > packets) {
2210                         if (tcp_is_sack(tp) ||
2211                             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2212                             (oldcnt >= packets))
2213                                 break;
2214 
2215                         mss = tcp_skb_mss(skb);
2216                         /* If needed, chop off the prefix to mark as lost. */
2217                         lost = (packets - oldcnt) * mss;
2218                         if (lost < skb->len &&
2219                             tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2220                                          lost, mss, GFP_ATOMIC) < 0)
2221                                 break;
2222                         cnt = packets;
2223                 }
2224 
2225                 tcp_skb_mark_lost(tp, skb);
2226 
2227                 if (mark_head)
2228                         break;
2229         }
2230         tcp_verify_left_out(tp);
2231 }
2232 
2233 /* Account newly detected lost packet(s) */
2234 
2235 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2236 {
2237         struct tcp_sock *tp = tcp_sk(sk);
2238 
2239         if (tcp_is_sack(tp)) {
2240                 int sacked_upto = tp->sacked_out - tp->reordering;
2241                 if (sacked_upto >= 0)
2242                         tcp_mark_head_lost(sk, sacked_upto, 0);
2243                 else if (fast_rexmit)
2244                         tcp_mark_head_lost(sk, 1, 1);
2245         }
2246 }
2247 
2248 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2249 {
2250         return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2251                before(tp->rx_opt.rcv_tsecr, when);
2252 }
2253 
2254 /* skb is spurious retransmitted if the returned timestamp echo
2255  * reply is prior to the skb transmission time
2256  */
2257 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2258                                      const struct sk_buff *skb)
2259 {
2260         return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2261                tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2262 }
2263 
2264 /* Nothing was retransmitted or returned timestamp is less
2265  * than timestamp of the first retransmission.
2266  */
2267 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2268 {
2269         return !tp->retrans_stamp ||
2270                tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2271 }
2272 
2273 /* Undo procedures. */
2274 
2275 /* We can clear retrans_stamp when there are no retransmissions in the
2276  * window. It would seem that it is trivially available for us in
2277  * tp->retrans_out, however, that kind of assumptions doesn't consider
2278  * what will happen if errors occur when sending retransmission for the
2279  * second time. ...It could the that such segment has only
2280  * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2281  * the head skb is enough except for some reneging corner cases that
2282  * are not worth the effort.
2283  *
2284  * Main reason for all this complexity is the fact that connection dying
2285  * time now depends on the validity of the retrans_stamp, in particular,
2286  * that successive retransmissions of a segment must not advance
2287  * retrans_stamp under any conditions.
2288  */
2289 static bool tcp_any_retrans_done(const struct sock *sk)
2290 {
2291         const struct tcp_sock *tp = tcp_sk(sk);
2292         struct sk_buff *skb;
2293 
2294         if (tp->retrans_out)
2295                 return true;
2296 
2297         skb = tcp_rtx_queue_head(sk);
2298         if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2299                 return true;
2300 
2301         return false;
2302 }
2303 
2304 static void DBGUNDO(struct sock *sk, const char *msg)
2305 {
2306 #if FASTRETRANS_DEBUG > 1
2307         struct tcp_sock *tp = tcp_sk(sk);
2308         struct inet_sock *inet = inet_sk(sk);
2309 
2310         if (sk->sk_family == AF_INET) {
2311                 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2312                          msg,
2313                          &inet->inet_daddr, ntohs(inet->inet_dport),
2314                          tp->snd_cwnd, tcp_left_out(tp),
2315                          tp->snd_ssthresh, tp->prior_ssthresh,
2316                          tp->packets_out);
2317         }
2318 #if IS_ENABLED(CONFIG_IPV6)
2319         else if (sk->sk_family == AF_INET6) {
2320                 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2321                          msg,
2322                          &sk->sk_v6_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 #endif
2328 #endif
2329 }
2330 
2331 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2332 {
2333         struct tcp_sock *tp = tcp_sk(sk);
2334 
2335         if (unmark_loss) {
2336                 struct sk_buff *skb;
2337 
2338                 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2339                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2340                 }
2341                 tp->lost_out = 0;
2342                 tcp_clear_all_retrans_hints(tp);
2343         }
2344 
2345         if (tp->prior_ssthresh) {
2346                 const struct inet_connection_sock *icsk = inet_csk(sk);
2347 
2348                 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2349 
2350                 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2351                         tp->snd_ssthresh = tp->prior_ssthresh;
2352                         tcp_ecn_withdraw_cwr(tp);
2353                 }
2354         }
2355         tp->snd_cwnd_stamp = tcp_jiffies32;
2356         tp->undo_marker = 0;
2357         tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2358 }
2359 
2360 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2361 {
2362         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2363 }
2364 
2365 /* People celebrate: "We love our President!" */
2366 static bool tcp_try_undo_recovery(struct sock *sk)
2367 {
2368         struct tcp_sock *tp = tcp_sk(sk);
2369 
2370         if (tcp_may_undo(tp)) {
2371                 int mib_idx;
2372 
2373                 /* Happy end! We did not retransmit anything
2374                  * or our original transmission succeeded.
2375                  */
2376                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2377                 tcp_undo_cwnd_reduction(sk, false);
2378                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2379                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2380                 else
2381                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2382 
2383                 NET_INC_STATS(sock_net(sk), mib_idx);
2384         } else if (tp->rack.reo_wnd_persist) {
2385                 tp->rack.reo_wnd_persist--;
2386         }
2387         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2388                 /* Hold old state until something *above* high_seq
2389                  * is ACKed. For Reno it is MUST to prevent false
2390                  * fast retransmits (RFC2582). SACK TCP is safe. */
2391                 if (!tcp_any_retrans_done(sk))
2392                         tp->retrans_stamp = 0;
2393                 return true;
2394         }
2395         tcp_set_ca_state(sk, TCP_CA_Open);
2396         tp->is_sack_reneg = 0;
2397         return false;
2398 }
2399 
2400 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2401 static bool tcp_try_undo_dsack(struct sock *sk)
2402 {
2403         struct tcp_sock *tp = tcp_sk(sk);
2404 
2405         if (tp->undo_marker && !tp->undo_retrans) {
2406                 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2407                                                tp->rack.reo_wnd_persist + 1);
2408                 DBGUNDO(sk, "D-SACK");
2409                 tcp_undo_cwnd_reduction(sk, false);
2410                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2411                 return true;
2412         }
2413         return false;
2414 }
2415 
2416 /* Undo during loss recovery after partial ACK or using F-RTO. */
2417 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2418 {
2419         struct tcp_sock *tp = tcp_sk(sk);
2420 
2421         if (frto_undo || tcp_may_undo(tp)) {
2422                 tcp_undo_cwnd_reduction(sk, true);
2423 
2424                 DBGUNDO(sk, "partial loss");
2425                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2426                 if (frto_undo)
2427                         NET_INC_STATS(sock_net(sk),
2428                                         LINUX_MIB_TCPSPURIOUSRTOS);
2429                 inet_csk(sk)->icsk_retransmits = 0;
2430                 if (frto_undo || tcp_is_sack(tp)) {
2431                         tcp_set_ca_state(sk, TCP_CA_Open);
2432                         tp->is_sack_reneg = 0;
2433                 }
2434                 return true;
2435         }
2436         return false;
2437 }
2438 
2439 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2440  * It computes the number of packets to send (sndcnt) based on packets newly
2441  * delivered:
2442  *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2443  *      cwnd reductions across a full RTT.
2444  *   2) Otherwise PRR uses packet conservation to send as much as delivered.
2445  *      But when the retransmits are acked without further losses, PRR
2446  *      slow starts cwnd up to ssthresh to speed up the recovery.
2447  */
2448 static void tcp_init_cwnd_reduction(struct sock *sk)
2449 {
2450         struct tcp_sock *tp = tcp_sk(sk);
2451 
2452         tp->high_seq = tp->snd_nxt;
2453         tp->tlp_high_seq = 0;
2454         tp->snd_cwnd_cnt = 0;
2455         tp->prior_cwnd = tp->snd_cwnd;
2456         tp->prr_delivered = 0;
2457         tp->prr_out = 0;
2458         tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2459         tcp_ecn_queue_cwr(tp);
2460 }
2461 
2462 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2463 {
2464         struct tcp_sock *tp = tcp_sk(sk);
2465         int sndcnt = 0;
2466         int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2467 
2468         if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2469                 return;
2470 
2471         tp->prr_delivered += newly_acked_sacked;
2472         if (delta < 0) {
2473                 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2474                                tp->prior_cwnd - 1;
2475                 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2476         } else if ((flag & (FLAG_RETRANS_DATA_ACKED | FLAG_LOST_RETRANS)) ==
2477                    FLAG_RETRANS_DATA_ACKED) {
2478                 sndcnt = min_t(int, delta,
2479                                max_t(int, tp->prr_delivered - tp->prr_out,
2480                                      newly_acked_sacked) + 1);
2481         } else {
2482                 sndcnt = min(delta, newly_acked_sacked);
2483         }
2484         /* Force a fast retransmit upon entering fast recovery */
2485         sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2486         tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2487 }
2488 
2489 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2490 {
2491         struct tcp_sock *tp = tcp_sk(sk);
2492 
2493         if (inet_csk(sk)->icsk_ca_ops->cong_control)
2494                 return;
2495 
2496         /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2497         if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2498             (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2499                 tp->snd_cwnd = tp->snd_ssthresh;
2500                 tp->snd_cwnd_stamp = tcp_jiffies32;
2501         }
2502         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2503 }
2504 
2505 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2506 void tcp_enter_cwr(struct sock *sk)
2507 {
2508         struct tcp_sock *tp = tcp_sk(sk);
2509 
2510         tp->prior_ssthresh = 0;
2511         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2512                 tp->undo_marker = 0;
2513                 tcp_init_cwnd_reduction(sk);
2514                 tcp_set_ca_state(sk, TCP_CA_CWR);
2515         }
2516 }
2517 EXPORT_SYMBOL(tcp_enter_cwr);
2518 
2519 static void tcp_try_keep_open(struct sock *sk)
2520 {
2521         struct tcp_sock *tp = tcp_sk(sk);
2522         int state = TCP_CA_Open;
2523 
2524         if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2525                 state = TCP_CA_Disorder;
2526 
2527         if (inet_csk(sk)->icsk_ca_state != state) {
2528                 tcp_set_ca_state(sk, state);
2529                 tp->high_seq = tp->snd_nxt;
2530         }
2531 }
2532 
2533 static void tcp_try_to_open(struct sock *sk, int flag)
2534 {
2535         struct tcp_sock *tp = tcp_sk(sk);
2536 
2537         tcp_verify_left_out(tp);
2538 
2539         if (!tcp_any_retrans_done(sk))
2540                 tp->retrans_stamp = 0;
2541 
2542         if (flag & FLAG_ECE)
2543                 tcp_enter_cwr(sk);
2544 
2545         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2546                 tcp_try_keep_open(sk);
2547         }
2548 }
2549 
2550 static void tcp_mtup_probe_failed(struct sock *sk)
2551 {
2552         struct inet_connection_sock *icsk = inet_csk(sk);
2553 
2554         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2555         icsk->icsk_mtup.probe_size = 0;
2556         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2557 }
2558 
2559 static void tcp_mtup_probe_success(struct sock *sk)
2560 {
2561         struct tcp_sock *tp = tcp_sk(sk);
2562         struct inet_connection_sock *icsk = inet_csk(sk);
2563 
2564         /* FIXME: breaks with very large cwnd */
2565         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2566         tp->snd_cwnd = tp->snd_cwnd *
2567                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2568                        icsk->icsk_mtup.probe_size;
2569         tp->snd_cwnd_cnt = 0;
2570         tp->snd_cwnd_stamp = tcp_jiffies32;
2571         tp->snd_ssthresh = tcp_current_ssthresh(sk);
2572 
2573         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2574         icsk->icsk_mtup.probe_size = 0;
2575         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2576         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2577 }
2578 
2579 /* Do a simple retransmit without using the backoff mechanisms in
2580  * tcp_timer. This is used for path mtu discovery.
2581  * The socket is already locked here.
2582  */
2583 void tcp_simple_retransmit(struct sock *sk)
2584 {
2585         const struct inet_connection_sock *icsk = inet_csk(sk);
2586         struct tcp_sock *tp = tcp_sk(sk);
2587         struct sk_buff *skb;
2588         unsigned int mss = tcp_current_mss(sk);
2589 
2590         skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2591                 if (tcp_skb_seglen(skb) > mss &&
2592                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2593                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2594                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2595                                 tp->retrans_out -= tcp_skb_pcount(skb);
2596                         }
2597                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2598                 }
2599         }
2600 
2601         tcp_clear_retrans_hints_partial(tp);
2602 
2603         if (!tp->lost_out)
2604                 return;
2605 
2606         if (tcp_is_reno(tp))
2607                 tcp_limit_reno_sacked(tp);
2608 
2609         tcp_verify_left_out(tp);
2610 
2611         /* Don't muck with the congestion window here.
2612          * Reason is that we do not increase amount of _data_
2613          * in network, but units changed and effective
2614          * cwnd/ssthresh really reduced now.
2615          */
2616         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2617                 tp->high_seq = tp->snd_nxt;
2618                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2619                 tp->prior_ssthresh = 0;
2620                 tp->undo_marker = 0;
2621                 tcp_set_ca_state(sk, TCP_CA_Loss);
2622         }
2623         tcp_xmit_retransmit_queue(sk);
2624 }
2625 EXPORT_SYMBOL(tcp_simple_retransmit);
2626 
2627 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2628 {
2629         struct tcp_sock *tp = tcp_sk(sk);
2630         int mib_idx;
2631 
2632         if (tcp_is_reno(tp))
2633                 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2634         else
2635                 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2636 
2637         NET_INC_STATS(sock_net(sk), mib_idx);
2638 
2639         tp->prior_ssthresh = 0;
2640         tcp_init_undo(tp);
2641 
2642         if (!tcp_in_cwnd_reduction(sk)) {
2643                 if (!ece_ack)
2644                         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2645                 tcp_init_cwnd_reduction(sk);
2646         }
2647         tcp_set_ca_state(sk, TCP_CA_Recovery);
2648 }
2649 
2650 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2651  * recovered or spurious. Otherwise retransmits more on partial ACKs.
2652  */
2653 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2654                              int *rexmit)
2655 {
2656         struct tcp_sock *tp = tcp_sk(sk);
2657         bool recovered = !before(tp->snd_una, tp->high_seq);
2658 
2659         if ((flag & FLAG_SND_UNA_ADVANCED) &&
2660             tcp_try_undo_loss(sk, false))
2661                 return;
2662 
2663         if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2664                 /* Step 3.b. A timeout is spurious if not all data are
2665                  * lost, i.e., never-retransmitted data are (s)acked.
2666                  */
2667                 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2668                     tcp_try_undo_loss(sk, true))
2669                         return;
2670 
2671                 if (after(tp->snd_nxt, tp->high_seq)) {
2672                         if (flag & FLAG_DATA_SACKED || num_dupack)
2673                                 tp->frto = 0; /* Step 3.a. loss was real */
2674                 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2675                         tp->high_seq = tp->snd_nxt;
2676                         /* Step 2.b. Try send new data (but deferred until cwnd
2677                          * is updated in tcp_ack()). Otherwise fall back to
2678                          * the conventional recovery.
2679                          */
2680                         if (!tcp_write_queue_empty(sk) &&
2681                             after(tcp_wnd_end(tp), tp->snd_nxt)) {
2682                                 *rexmit = REXMIT_NEW;
2683                                 return;
2684                         }
2685                         tp->frto = 0;
2686                 }
2687         }
2688 
2689         if (recovered) {
2690                 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2691                 tcp_try_undo_recovery(sk);
2692                 return;
2693         }
2694         if (tcp_is_reno(tp)) {
2695                 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2696                  * delivered. Lower inflight to clock out (re)tranmissions.
2697                  */
2698                 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2699                         tcp_add_reno_sack(sk, num_dupack);
2700                 else if (flag & FLAG_SND_UNA_ADVANCED)
2701                         tcp_reset_reno_sack(tp);
2702         }
2703         *rexmit = REXMIT_LOST;
2704 }
2705 
2706 /* Undo during fast recovery after partial ACK. */
2707 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2708 {
2709         struct tcp_sock *tp = tcp_sk(sk);
2710 
2711         if (tp->undo_marker && tcp_packet_delayed(tp)) {
2712                 /* Plain luck! Hole if filled with delayed
2713                  * packet, rather than with a retransmit. Check reordering.
2714                  */
2715                 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2716 
2717                 /* We are getting evidence that the reordering degree is higher
2718                  * than we realized. If there are no retransmits out then we
2719                  * can undo. Otherwise we clock out new packets but do not
2720                  * mark more packets lost or retransmit more.
2721                  */
2722                 if (tp->retrans_out)
2723                         return true;
2724 
2725                 if (!tcp_any_retrans_done(sk))
2726                         tp->retrans_stamp = 0;
2727 
2728                 DBGUNDO(sk, "partial recovery");
2729                 tcp_undo_cwnd_reduction(sk, true);
2730                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2731                 tcp_try_keep_open(sk);
2732                 return true;
2733         }
2734         return false;
2735 }
2736 
2737 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2738 {
2739         struct tcp_sock *tp = tcp_sk(sk);
2740 
2741         if (tcp_rtx_queue_empty(sk))
2742                 return;
2743 
2744         if (unlikely(tcp_is_reno(tp))) {
2745                 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2746         } else if (tcp_is_rack(sk)) {
2747                 u32 prior_retrans = tp->retrans_out;
2748 
2749                 tcp_rack_mark_lost(sk);
2750                 if (prior_retrans > tp->retrans_out)
2751                         *ack_flag |= FLAG_LOST_RETRANS;
2752         }
2753 }
2754 
2755 static bool tcp_force_fast_retransmit(struct sock *sk)
2756 {
2757         struct tcp_sock *tp = tcp_sk(sk);
2758 
2759         return after(tcp_highest_sack_seq(tp),
2760                      tp->snd_una + tp->reordering * tp->mss_cache);
2761 }
2762 
2763 /* Process an event, which can update packets-in-flight not trivially.
2764  * Main goal of this function is to calculate new estimate for left_out,
2765  * taking into account both packets sitting in receiver's buffer and
2766  * packets lost by network.
2767  *
2768  * Besides that it updates the congestion state when packet loss or ECN
2769  * is detected. But it does not reduce the cwnd, it is done by the
2770  * congestion control later.
2771  *
2772  * It does _not_ decide what to send, it is made in function
2773  * tcp_xmit_retransmit_queue().
2774  */
2775 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2776                                   int num_dupack, int *ack_flag, int *rexmit)
2777 {
2778         struct inet_connection_sock *icsk = inet_csk(sk);
2779         struct tcp_sock *tp = tcp_sk(sk);
2780         int fast_rexmit = 0, flag = *ack_flag;
2781         bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2782                                       tcp_force_fast_retransmit(sk));
2783 
2784         if (!tp->packets_out && tp->sacked_out)
2785                 tp->sacked_out = 0;
2786 
2787         /* Now state machine starts.
2788          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2789         if (flag & FLAG_ECE)
2790                 tp->prior_ssthresh = 0;
2791 
2792         /* B. In all the states check for reneging SACKs. */
2793         if (tcp_check_sack_reneging(sk, flag))
2794                 return;
2795 
2796         /* C. Check consistency of the current state. */
2797         tcp_verify_left_out(tp);
2798 
2799         /* D. Check state exit conditions. State can be terminated
2800          *    when high_seq is ACKed. */
2801         if (icsk->icsk_ca_state == TCP_CA_Open) {
2802                 WARN_ON(tp->retrans_out != 0);
2803                 tp->retrans_stamp = 0;
2804         } else if (!before(tp->snd_una, tp->high_seq)) {
2805                 switch (icsk->icsk_ca_state) {
2806                 case TCP_CA_CWR:
2807                         /* CWR is to be held something *above* high_seq
2808                          * is ACKed for CWR bit to reach receiver. */
2809                         if (tp->snd_una != tp->high_seq) {
2810                                 tcp_end_cwnd_reduction(sk);
2811                                 tcp_set_ca_state(sk, TCP_CA_Open);
2812                         }
2813                         break;
2814 
2815                 case TCP_CA_Recovery:
2816                         if (tcp_is_reno(tp))
2817                                 tcp_reset_reno_sack(tp);
2818                         if (tcp_try_undo_recovery(sk))
2819                                 return;
2820                         tcp_end_cwnd_reduction(sk);
2821                         break;
2822                 }
2823         }
2824 
2825         /* E. Process state. */
2826         switch (icsk->icsk_ca_state) {
2827         case TCP_CA_Recovery:
2828                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2829                         if (tcp_is_reno(tp))
2830                                 tcp_add_reno_sack(sk, num_dupack);
2831                 } else {
2832                         if (tcp_try_undo_partial(sk, prior_snd_una))
2833                                 return;
2834                         /* Partial ACK arrived. Force fast retransmit. */
2835                         do_lost = tcp_is_reno(tp) ||
2836                                   tcp_force_fast_retransmit(sk);
2837                 }
2838                 if (tcp_try_undo_dsack(sk)) {
2839                         tcp_try_keep_open(sk);
2840                         return;
2841                 }
2842                 tcp_identify_packet_loss(sk, ack_flag);
2843                 break;
2844         case TCP_CA_Loss:
2845                 tcp_process_loss(sk, flag, num_dupack, rexmit);
2846                 tcp_identify_packet_loss(sk, ack_flag);
2847                 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2848                       (*ack_flag & FLAG_LOST_RETRANS)))
2849                         return;
2850                 /* Change state if cwnd is undone or retransmits are lost */
2851                 /* fall through */
2852         default:
2853                 if (tcp_is_reno(tp)) {
2854                         if (flag & FLAG_SND_UNA_ADVANCED)
2855                                 tcp_reset_reno_sack(tp);
2856                         tcp_add_reno_sack(sk, num_dupack);
2857                 }
2858 
2859                 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2860                         tcp_try_undo_dsack(sk);
2861 
2862                 tcp_identify_packet_loss(sk, ack_flag);
2863                 if (!tcp_time_to_recover(sk, flag)) {
2864                         tcp_try_to_open(sk, flag);
2865                         return;
2866                 }
2867 
2868                 /* MTU probe failure: don't reduce cwnd */
2869                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2870                     icsk->icsk_mtup.probe_size &&
2871                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2872                         tcp_mtup_probe_failed(sk);
2873                         /* Restores the reduction we did in tcp_mtup_probe() */
2874                         tp->snd_cwnd++;
2875                         tcp_simple_retransmit(sk);
2876                         return;
2877                 }
2878 
2879                 /* Otherwise enter Recovery state */
2880                 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2881                 fast_rexmit = 1;
2882         }
2883 
2884         if (!tcp_is_rack(sk) && do_lost)
2885                 tcp_update_scoreboard(sk, fast_rexmit);
2886         *rexmit = REXMIT_LOST;
2887 }
2888 
2889 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2890 {
2891         u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
2892         struct tcp_sock *tp = tcp_sk(sk);
2893 
2894         if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2895                 /* If the remote keeps returning delayed ACKs, eventually
2896                  * the min filter would pick it up and overestimate the
2897                  * prop. delay when it expires. Skip suspected delayed ACKs.
2898                  */
2899                 return;
2900         }
2901         minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2902                            rtt_us ? : jiffies_to_usecs(1));
2903 }
2904 
2905 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2906                                long seq_rtt_us, long sack_rtt_us,
2907                                long ca_rtt_us, struct rate_sample *rs)
2908 {
2909         const struct tcp_sock *tp = tcp_sk(sk);
2910 
2911         /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2912          * broken middle-boxes or peers may corrupt TS-ECR fields. But
2913          * Karn's algorithm forbids taking RTT if some retransmitted data
2914          * is acked (RFC6298).
2915          */
2916         if (seq_rtt_us < 0)
2917                 seq_rtt_us = sack_rtt_us;
2918 
2919         /* RTTM Rule: A TSecr value received in a segment is used to
2920          * update the averaged RTT measurement only if the segment
2921          * acknowledges some new data, i.e., only if it advances the
2922          * left edge of the send window.
2923          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2924          */
2925         if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2926             flag & FLAG_ACKED) {
2927                 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2928 
2929                 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
2930                         seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2931                         ca_rtt_us = seq_rtt_us;
2932                 }
2933         }
2934         rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2935         if (seq_rtt_us < 0)
2936                 return false;
2937 
2938         /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2939          * always taken together with ACK, SACK, or TS-opts. Any negative
2940          * values will be skipped with the seq_rtt_us < 0 check above.
2941          */
2942         tcp_update_rtt_min(sk, ca_rtt_us, flag);
2943         tcp_rtt_estimator(sk, seq_rtt_us);
2944         tcp_set_rto(sk);
2945 
2946         /* RFC6298: only reset backoff on valid RTT measurement. */
2947         inet_csk(sk)->icsk_backoff = 0;
2948         return true;
2949 }
2950 
2951 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2952 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2953 {
2954         struct rate_sample rs;
2955         long rtt_us = -1L;
2956 
2957         if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2958                 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2959 
2960         tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2961 }
2962 
2963 
2964 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2965 {
2966         const struct inet_connection_sock *icsk = inet_csk(sk);
2967 
2968         icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2969         tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2970 }
2971 
2972 /* Restart timer after forward progress on connection.
2973  * RFC2988 recommends to restart timer to now+rto.
2974  */
2975 void tcp_rearm_rto(struct sock *sk)
2976 {
2977         const struct inet_connection_sock *icsk = inet_csk(sk);
2978         struct tcp_sock *tp = tcp_sk(sk);
2979 
2980         /* If the retrans timer is currently being used by Fast Open
2981          * for SYN-ACK retrans purpose, stay put.
2982          */
2983         if (tp->fastopen_rsk)
2984                 return;
2985 
2986         if (!tp->packets_out) {
2987                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2988         } else {
2989                 u32 rto = inet_csk(sk)->icsk_rto;
2990                 /* Offset the time elapsed after installing regular RTO */
2991                 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2992                     icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2993                         s64 delta_us = tcp_rto_delta_us(sk);
2994                         /* delta_us may not be positive if the socket is locked
2995                          * when the retrans timer fires and is rescheduled.
2996                          */
2997                         rto = usecs_to_jiffies(max_t(int, delta_us, 1));
2998                 }
2999                 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3000                                      TCP_RTO_MAX, tcp_rtx_queue_head(sk));
3001         }
3002 }
3003 
3004 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3005 static void tcp_set_xmit_timer(struct sock *sk)
3006 {
3007         if (!tcp_schedule_loss_probe(sk, true))
3008                 tcp_rearm_rto(sk);
3009 }
3010 
3011 /* If we get here, the whole TSO packet has not been acked. */
3012 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3013 {
3014         struct tcp_sock *tp = tcp_sk(sk);
3015         u32 packets_acked;
3016 
3017         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3018 
3019         packets_acked = tcp_skb_pcount(skb);
3020         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3021                 return 0;
3022         packets_acked -= tcp_skb_pcount(skb);
3023 
3024         if (packets_acked) {
3025                 BUG_ON(tcp_skb_pcount(skb) == 0);
3026                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3027         }
3028 
3029         return packets_acked;
3030 }
3031 
3032 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3033                            u32 prior_snd_una)
3034 {
3035         const struct skb_shared_info *shinfo;
3036 
3037         /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3038         if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3039                 return;
3040 
3041         shinfo = skb_shinfo(skb);
3042         if (!before(shinfo->tskey, prior_snd_una) &&
3043             before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3044                 tcp_skb_tsorted_save(skb) {
3045                         __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3046                 } tcp_skb_tsorted_restore(skb);
3047         }
3048 }
3049 
3050 /* Remove acknowledged frames from the retransmission queue. If our packet
3051  * is before the ack sequence we can discard it as it's confirmed to have
3052  * arrived at the other end.
3053  */
3054 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3055                                u32 prior_snd_una,
3056                                struct tcp_sacktag_state *sack)
3057 {
3058         const struct inet_connection_sock *icsk = inet_csk(sk);
3059         u64 first_ackt, last_ackt;
3060         struct tcp_sock *tp = tcp_sk(sk);
3061         u32 prior_sacked = tp->sacked_out;
3062         u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3063         struct sk_buff *skb, *next;
3064         bool fully_acked = true;
3065         long sack_rtt_us = -1L;
3066         long seq_rtt_us = -1L;
3067         long ca_rtt_us = -1L;
3068         u32 pkts_acked = 0;
3069         u32 last_in_flight = 0;
3070         bool rtt_update;
3071         int flag = 0;
3072 
3073         first_ackt = 0;
3074 
3075         for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3076                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3077                 const u32 start_seq = scb->seq;
3078                 u8 sacked = scb->sacked;
3079                 u32 acked_pcount;
3080 
3081                 tcp_ack_tstamp(sk, skb, prior_snd_una);
3082 
3083                 /* Determine how many packets and what bytes were acked, tso and else */
3084                 if (after(scb->end_seq, tp->snd_una)) {
3085                         if (tcp_skb_pcount(skb) == 1 ||
3086                             !after(tp->snd_una, scb->seq))
3087                                 break;
3088 
3089                         acked_pcount = tcp_tso_acked(sk, skb);
3090                         if (!acked_pcount)
3091                                 break;
3092                         fully_acked = false;
3093                 } else {
3094                         acked_pcount = tcp_skb_pcount(skb);
3095                 }
3096 
3097                 if (unlikely(sacked & TCPCB_RETRANS)) {
3098                         if (sacked & TCPCB_SACKED_RETRANS)
3099                                 tp->retrans_out -= acked_pcount;
3100                         flag |= FLAG_RETRANS_DATA_ACKED;
3101                 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3102                         last_ackt = tcp_skb_timestamp_us(skb);
3103                         WARN_ON_ONCE(last_ackt == 0);
3104                         if (!first_ackt)
3105                                 first_ackt = last_ackt;
3106 
3107                         last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3108                         if (before(start_seq, reord))
3109                                 reord = start_seq;
3110                         if (!after(scb->end_seq, tp->high_seq))
3111                                 flag |= FLAG_ORIG_SACK_ACKED;
3112                 }
3113 
3114                 if (sacked & TCPCB_SACKED_ACKED) {
3115                         tp->sacked_out -= acked_pcount;
3116                 } else if (tcp_is_sack(tp)) {
3117                         tp->delivered += acked_pcount;
3118                         if (!tcp_skb_spurious_retrans(tp, skb))
3119                                 tcp_rack_advance(tp, sacked, scb->end_seq,
3120                                                  tcp_skb_timestamp_us(skb));
3121                 }
3122                 if (sacked & TCPCB_LOST)
3123                         tp->lost_out -= acked_pcount;
3124 
3125                 tp->packets_out -= acked_pcount;
3126                 pkts_acked += acked_pcount;
3127                 tcp_rate_skb_delivered(sk, skb, sack->rate);
3128 
3129                 /* Initial outgoing SYN's get put onto the write_queue
3130                  * just like anything else we transmit.  It is not
3131                  * true data, and if we misinform our callers that
3132                  * this ACK acks real data, we will erroneously exit
3133                  * connection startup slow start one packet too
3134                  * quickly.  This is severely frowned upon behavior.
3135                  */
3136                 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3137                         flag |= FLAG_DATA_ACKED;
3138                 } else {
3139                         flag |= FLAG_SYN_ACKED;
3140                         tp->retrans_stamp = 0;
3141                 }
3142 
3143                 if (!fully_acked)
3144                         break;
3145 
3146                 next = skb_rb_next(skb);
3147                 if (unlikely(skb == tp->retransmit_skb_hint))
3148                         tp->retransmit_skb_hint = NULL;
3149                 if (unlikely(skb == tp->lost_skb_hint))
3150                         tp->lost_skb_hint = NULL;
3151                 tcp_rtx_queue_unlink_and_free(skb, sk);
3152         }
3153 
3154         if (!skb)
3155                 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3156 
3157         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3158                 tp->snd_up = tp->snd_una;
3159 
3160         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3161                 flag |= FLAG_SACK_RENEGING;
3162 
3163         if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3164                 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3165                 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3166 
3167                 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3168                     last_in_flight && !prior_sacked && fully_acked &&
3169                     sack->rate->prior_delivered + 1 == tp->delivered &&
3170                     !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3171                         /* Conservatively mark a delayed ACK. It's typically
3172                          * from a lone runt packet over the round trip to
3173                          * a receiver w/o out-of-order or CE events.
3174                          */
3175                         flag |= FLAG_ACK_MAYBE_DELAYED;
3176                 }
3177         }
3178         if (sack->first_sackt) {
3179                 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3180                 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3181         }
3182         rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3183                                         ca_rtt_us, sack->rate);
3184 
3185         if (flag & FLAG_ACKED) {
3186                 flag |= FLAG_SET_XMIT_TIMER;  /* set TLP or RTO timer */
3187                 if (unlikely(icsk->icsk_mtup.probe_size &&
3188                              !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3189                         tcp_mtup_probe_success(sk);
3190                 }
3191 
3192                 if (tcp_is_reno(tp)) {
3193                         tcp_remove_reno_sacks(sk, pkts_acked);
3194 
3195                         /* If any of the cumulatively ACKed segments was
3196                          * retransmitted, non-SACK case cannot confirm that
3197                          * progress was due to original transmission due to
3198                          * lack of TCPCB_SACKED_ACKED bits even if some of
3199                          * the packets may have been never retransmitted.
3200                          */
3201                         if (flag & FLAG_RETRANS_DATA_ACKED)
3202                                 flag &= ~FLAG_ORIG_SACK_ACKED;
3203                 } else {
3204                         int delta;
3205 
3206                         /* Non-retransmitted hole got filled? That's reordering */
3207                         if (before(reord, prior_fack))
3208                                 tcp_check_sack_reordering(sk, reord, 0);
3209 
3210                         delta = prior_sacked - tp->sacked_out;
3211                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3212                 }
3213         } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3214                    sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3215                                                     tcp_skb_timestamp_us(skb))) {
3216                 /* Do not re-arm RTO if the sack RTT is measured from data sent
3217                  * after when the head was last (re)transmitted. Otherwise the
3218                  * timeout may continue to extend in loss recovery.
3219                  */
3220                 flag |= FLAG_SET_XMIT_TIMER;  /* set TLP or RTO timer */
3221         }
3222 
3223         if (icsk->icsk_ca_ops->pkts_acked) {
3224                 struct ack_sample sample = { .pkts_acked = pkts_acked,
3225                                              .rtt_us = sack->rate->rtt_us,
3226                                              .in_flight = last_in_flight };
3227 
3228                 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3229         }
3230 
3231 #if FASTRETRANS_DEBUG > 0
3232         WARN_ON((int)tp->sacked_out < 0);
3233         WARN_ON((int)tp->lost_out < 0);
3234         WARN_ON((int)tp->retrans_out < 0);
3235         if (!tp->packets_out && tcp_is_sack(tp)) {
3236                 icsk = inet_csk(sk);
3237                 if (tp->lost_out) {
3238                         pr_debug("Leak l=%u %d\n",
3239                                  tp->lost_out, icsk->icsk_ca_state);
3240                         tp->lost_out = 0;
3241                 }
3242                 if (tp->sacked_out) {
3243                         pr_debug("Leak s=%u %d\n",
3244                                  tp->sacked_out, icsk->icsk_ca_state);
3245                         tp->sacked_out = 0;
3246                 }
3247                 if (tp->retrans_out) {
3248                         pr_debug("Leak r=%u %d\n",
3249                                  tp->retrans_out, icsk->icsk_ca_state);
3250                         tp->retrans_out = 0;
3251                 }
3252         }
3253 #endif
3254         return flag;
3255 }
3256 
3257 static void tcp_ack_probe(struct sock *sk)
3258 {
3259         struct inet_connection_sock *icsk = inet_csk(sk);
3260         struct sk_buff *head = tcp_send_head(sk);
3261         const struct tcp_sock *tp = tcp_sk(sk);
3262 
3263         /* Was it a usable window open? */
3264         if (!head)
3265                 return;
3266         if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3267                 icsk->icsk_backoff = 0;
3268                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3269                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3270                  * This function is not for random using!
3271                  */
3272         } else {
3273                 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3274 
3275                 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3276                                      when, TCP_RTO_MAX, NULL);
3277         }
3278 }
3279 
3280 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3281 {
3282         return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3283                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3284 }
3285 
3286 /* Decide wheather to run the increase function of congestion control. */
3287 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3288 {
3289         /* If reordering is high then always grow cwnd whenever data is
3290          * delivered regardless of its ordering. Otherwise stay conservative
3291          * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3292          * new SACK or ECE mark may first advance cwnd here and later reduce
3293          * cwnd in tcp_fastretrans_alert() based on more states.
3294          */
3295         if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3296                 return flag & FLAG_FORWARD_PROGRESS;
3297 
3298         return flag & FLAG_DATA_ACKED;
3299 }
3300 
3301 /* The "ultimate" congestion control function that aims to replace the rigid
3302  * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3303  * It's called toward the end of processing an ACK with precise rate
3304  * information. All transmission or retransmission are delayed afterwards.
3305  */
3306 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3307                              int flag, const struct rate_sample *rs)
3308 {
3309         const struct inet_connection_sock *icsk = inet_csk(sk);
3310 
3311         if (icsk->icsk_ca_ops->cong_control) {
3312                 icsk->icsk_ca_ops->cong_control(sk, rs);
3313                 return;
3314         }
3315 
3316         if (tcp_in_cwnd_reduction(sk)) {
3317                 /* Reduce cwnd if state mandates */
3318                 tcp_cwnd_reduction(sk, acked_sacked, flag);
3319         } else if (tcp_may_raise_cwnd(sk, flag)) {
3320                 /* Advance cwnd if state allows */
3321                 tcp_cong_avoid(sk, ack, acked_sacked);
3322         }
3323         tcp_update_pacing_rate(sk);
3324 }
3325 
3326 /* Check that window update is acceptable.
3327  * The function assumes that snd_una<=ack<=snd_next.
3328  */
3329 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3330                                         const u32 ack, const u32 ack_seq,
3331                                         const u32 nwin)
3332 {
3333         return  after(ack, tp->snd_una) ||
3334                 after(ack_seq, tp->snd_wl1) ||
3335                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3336 }
3337 
3338 /* If we update tp->snd_una, also update tp->bytes_acked */
3339 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3340 {
3341         u32 delta = ack - tp->snd_una;
3342 
3343         sock_owned_by_me((struct sock *)tp);
3344         tp->bytes_acked += delta;
3345         tp->snd_una = ack;
3346 }
3347 
3348 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3349 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3350 {
3351         u32 delta = seq - tp->rcv_nxt;
3352 
3353         sock_owned_by_me((struct sock *)tp);
3354         tp->bytes_received += delta;
3355         tp->rcv_nxt = seq;
3356 }
3357 
3358 /* Update our send window.
3359  *
3360  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3361  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3362  */
3363 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3364                                  u32 ack_seq)
3365 {
3366         struct tcp_sock *tp = tcp_sk(sk);
3367         int flag = 0;
3368         u32 nwin = ntohs(tcp_hdr(skb)->window);
3369 
3370         if (likely(!tcp_hdr(skb)->syn))
3371                 nwin <<= tp->rx_opt.snd_wscale;
3372 
3373         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3374                 flag |= FLAG_WIN_UPDATE;
3375                 tcp_update_wl(tp, ack_seq);
3376 
3377                 if (tp->snd_wnd != nwin) {
3378                         tp->snd_wnd = nwin;
3379 
3380                         /* Note, it is the only place, where
3381                          * fast path is recovered for sending TCP.
3382                          */
3383                         tp->pred_flags = 0;
3384                         tcp_fast_path_check(sk);
3385 
3386                         if (!tcp_write_queue_empty(sk))
3387                                 tcp_slow_start_after_idle_check(sk);
3388 
3389                         if (nwin > tp->max_window) {
3390                                 tp->max_window = nwin;
3391                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3392                         }
3393                 }
3394         }
3395 
3396         tcp_snd_una_update(tp, ack);
3397 
3398         return flag;
3399 }
3400 
3401 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3402                                    u32 *last_oow_ack_time)
3403 {
3404         if (*last_oow_ack_time) {
3405                 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3406 
3407                 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3408                         NET_INC_STATS(net, mib_idx);
3409                         return true;    /* rate-limited: don't send yet! */
3410                 }
3411         }
3412 
3413         *last_oow_ack_time = tcp_jiffies32;
3414 
3415         return false;   /* not rate-limited: go ahead, send dupack now! */
3416 }
3417 
3418 /* Return true if we're currently rate-limiting out-of-window ACKs and
3419  * thus shouldn't send a dupack right now. We rate-limit dupacks in
3420  * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3421  * attacks that send repeated SYNs or ACKs for the same connection. To
3422  * do this, we do not send a duplicate SYNACK or ACK if the remote
3423  * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3424  */
3425 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3426                           int mib_idx, u32 *last_oow_ack_time)
3427 {
3428         /* Data packets without SYNs are not likely part of an ACK loop. */
3429         if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3430             !tcp_hdr(skb)->syn)
3431                 return false;
3432 
3433         return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3434 }
3435 
3436 /* RFC 5961 7 [ACK Throttling] */
3437 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3438 {
3439         /* unprotected vars, we dont care of overwrites */
3440         static u32 challenge_timestamp;
3441         static unsigned int challenge_count;
3442         struct tcp_sock *tp = tcp_sk(sk);
3443         struct net *net = sock_net(sk);
3444         u32 count, now;
3445 
3446         /* First check our per-socket dupack rate limit. */
3447         if (__tcp_oow_rate_limited(net,
3448                                    LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3449                                    &tp->last_oow_ack_time))
3450                 return;
3451 
3452         /* Then check host-wide RFC 5961 rate limit. */
3453         now = jiffies / HZ;
3454         if (now != challenge_timestamp) {
3455                 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3456                 u32 half = (ack_limit + 1) >> 1;
3457 
3458                 challenge_timestamp = now;
3459                 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3460         }
3461         count = READ_ONCE(challenge_count);
3462         if (count > 0) {
3463                 WRITE_ONCE(challenge_count, count - 1);
3464                 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3465                 tcp_send_ack(sk);
3466         }
3467 }
3468 
3469 static void tcp_store_ts_recent(struct tcp_sock *tp)
3470 {
3471         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3472         tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3473 }
3474 
3475 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3476 {
3477         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3478                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3479                  * extra check below makes sure this can only happen
3480                  * for pure ACK frames.  -DaveM
3481                  *
3482                  * Not only, also it occurs for expired timestamps.
3483                  */
3484 
3485                 if (tcp_paws_check(&tp->rx_opt, 0))
3486                         tcp_store_ts_recent(tp);
3487         }
3488 }
3489 
3490 /* This routine deals with acks during a TLP episode.
3491  * We mark the end of a TLP episode on receiving TLP dupack or when
3492  * ack is after tlp_high_seq.
3493  * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3494  */
3495 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3496 {
3497         struct tcp_sock *tp = tcp_sk(sk);
3498 
3499         if (before(ack, tp->tlp_high_seq))
3500                 return;
3501 
3502         if (flag & FLAG_DSACKING_ACK) {
3503                 /* This DSACK means original and TLP probe arrived; no loss */
3504                 tp->tlp_high_seq = 0;
3505         } else if (after(ack, tp->tlp_high_seq)) {
3506                 /* ACK advances: there was a loss, so reduce cwnd. Reset
3507                  * tlp_high_seq in tcp_init_cwnd_reduction()
3508                  */
3509                 tcp_init_cwnd_reduction(sk);
3510                 tcp_set_ca_state(sk, TCP_CA_CWR);
3511                 tcp_end_cwnd_reduction(sk);
3512                 tcp_try_keep_open(sk);
3513                 NET_INC_STATS(sock_net(sk),
3514                                 LINUX_MIB_TCPLOSSPROBERECOVERY);
3515         } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3516                              FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3517                 /* Pure dupack: original and TLP probe arrived; no loss */
3518                 tp->tlp_high_seq = 0;
3519         }
3520 }
3521 
3522 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3523 {
3524         const struct inet_connection_sock *icsk = inet_csk(sk);
3525 
3526         if (icsk->icsk_ca_ops->in_ack_event)
3527                 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3528 }
3529 
3530 /* Congestion control has updated the cwnd already. So if we're in
3531  * loss recovery then now we do any new sends (for FRTO) or
3532  * retransmits (for CA_Loss or CA_recovery) that make sense.
3533  */
3534 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3535 {
3536         struct tcp_sock *tp = tcp_sk(sk);
3537 
3538         if (rexmit == REXMIT_NONE)
3539                 return;
3540 
3541         if (unlikely(rexmit == 2)) {
3542                 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3543                                           TCP_NAGLE_OFF);
3544                 if (after(tp->snd_nxt, tp->high_seq))
3545                         return;
3546                 tp->frto = 0;
3547         }
3548         tcp_xmit_retransmit_queue(sk);
3549 }
3550 
3551 /* Returns the number of packets newly acked or sacked by the current ACK */
3552 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3553 {
3554         const struct net *net = sock_net(sk);
3555         struct tcp_sock *tp = tcp_sk(sk);
3556         u32 delivered;
3557 
3558         delivered = tp->delivered - prior_delivered;
3559         NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3560         if (flag & FLAG_ECE) {
3561                 tp->delivered_ce += delivered;
3562                 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3563         }
3564         return delivered;
3565 }
3566 
3567 /* This routine deals with incoming acks, but not outgoing ones. */
3568 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3569 {
3570         struct inet_connection_sock *icsk = inet_csk(sk);
3571         struct tcp_sock *tp = tcp_sk(sk);
3572         struct tcp_sacktag_state sack_state;
3573         struct rate_sample rs = { .prior_delivered = 0 };
3574         u32 prior_snd_una = tp->snd_una;
3575         bool is_sack_reneg = tp->is_sack_reneg;
3576         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3577         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3578         int num_dupack = 0;
3579         int prior_packets = tp->packets_out;
3580         u32 delivered = tp->delivered;
3581         u32 lost = tp->lost;
3582         int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3583         u32 prior_fack;
3584 
3585         sack_state.first_sackt = 0;
3586         sack_state.rate = &rs;
3587 
3588         /* We very likely will need to access rtx queue. */
3589         prefetch(sk->tcp_rtx_queue.rb_node);
3590 
3591         /* If the ack is older than previous acks
3592          * then we can probably ignore it.
3593          */
3594         if (before(ack, prior_snd_una)) {
3595                 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3596                 if (before(ack, prior_snd_una - tp->max_window)) {
3597                         if (!(flag & FLAG_NO_CHALLENGE_ACK))
3598                                 tcp_send_challenge_ack(sk, skb);
3599                         return -1;
3600                 }
3601                 goto old_ack;
3602         }
3603 
3604         /* If the ack includes data we haven't sent yet, discard
3605          * this segment (RFC793 Section 3.9).
3606          */
3607         if (after(ack, tp->snd_nxt))
3608                 return -1;
3609 
3610         if (after(ack, prior_snd_una)) {
3611                 flag |= FLAG_SND_UNA_ADVANCED;
3612                 icsk->icsk_retransmits = 0;
3613 
3614 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3615                 if (static_branch_unlikely(&clean_acked_data_enabled))
3616                         if (icsk->icsk_clean_acked)
3617                                 icsk->icsk_clean_acked(sk, ack);
3618 #endif
3619         }
3620 
3621         prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3622         rs.prior_in_flight = tcp_packets_in_flight(tp);
3623 
3624         /* ts_recent update must be made after we are sure that the packet
3625          * is in window.
3626          */
3627         if (flag & FLAG_UPDATE_TS_RECENT)
3628                 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3629 
3630         if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3631             FLAG_SND_UNA_ADVANCED) {
3632                 /* Window is constant, pure forward advance.
3633                  * No more checks are required.
3634                  * Note, we use the fact that SND.UNA>=SND.WL2.
3635                  */
3636                 tcp_update_wl(tp, ack_seq);
3637                 tcp_snd_una_update(tp, ack);
3638                 flag |= FLAG_WIN_UPDATE;
3639 
3640                 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3641 
3642                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3643         } else {
3644                 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3645 
3646                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3647                         flag |= FLAG_DATA;
3648                 else
3649                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3650 
3651                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3652 
3653                 if (TCP_SKB_CB(skb)->sacked)
3654                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3655                                                         &sack_state);
3656 
3657                 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3658                         flag |= FLAG_ECE;
3659                         ack_ev_flags |= CA_ACK_ECE;
3660                 }
3661 
3662                 if (flag & FLAG_WIN_UPDATE)
3663                         ack_ev_flags |= CA_ACK_WIN_UPDATE;
3664 
3665                 tcp_in_ack_event(sk, ack_ev_flags);
3666         }
3667 
3668         /* We passed data and got it acked, remove any soft error
3669          * log. Something worked...
3670          */
3671         sk->sk_err_soft = 0;
3672         icsk->icsk_probes_out = 0;
3673         tp->rcv_tstamp = tcp_jiffies32;
3674         if (!prior_packets)
3675                 goto no_queue;
3676 
3677         /* See if we can take anything off of the retransmit queue. */
3678         flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
3679 
3680         tcp_rack_update_reo_wnd(sk, &rs);
3681 
3682         if (tp->tlp_high_seq)
3683                 tcp_process_tlp_ack(sk, ack, flag);
3684         /* If needed, reset TLP/RTO timer; RACK may later override this. */
3685         if (flag & FLAG_SET_XMIT_TIMER)
3686                 tcp_set_xmit_timer(sk);
3687 
3688         if (tcp_ack_is_dubious(sk, flag)) {
3689                 if (!(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP))) {
3690                         num_dupack = 1;
3691                         /* Consider if pure acks were aggregated in tcp_add_backlog() */
3692                         if (!(flag & FLAG_DATA))
3693                                 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3694                 }
3695                 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3696                                       &rexmit);
3697         }
3698 
3699         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3700                 sk_dst_confirm(sk);
3701 
3702         delivered = tcp_newly_delivered(sk, delivered, flag);
3703         lost = tp->lost - lost;                 /* freshly marked lost */
3704         rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3705         tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3706         tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3707         tcp_xmit_recovery(sk, rexmit);
3708         return 1;
3709 
3710 no_queue:
3711         /* If data was DSACKed, see if we can undo a cwnd reduction. */
3712         if (flag & FLAG_DSACKING_ACK) {
3713                 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3714                                       &rexmit);
3715                 tcp_newly_delivered(sk, delivered, flag);
3716         }
3717         /* If this ack opens up a zero window, clear backoff.  It was
3718          * being used to time the probes, and is probably far higher than
3719          * it needs to be for normal retransmission.
3720          */
3721         tcp_ack_probe(sk);
3722 
3723         if (tp->tlp_high_seq)
3724                 tcp_process_tlp_ack(sk, ack, flag);
3725         return 1;
3726 
3727 old_ack:
3728         /* If data was SACKed, tag it and see if we should send more data.
3729          * If data was DSACKed, see if we can undo a cwnd reduction.
3730          */
3731         if (TCP_SKB_CB(skb)->sacked) {
3732                 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3733                                                 &sack_state);
3734                 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3735                                       &rexmit);
3736                 tcp_newly_delivered(sk, delivered, flag);
3737                 tcp_xmit_recovery(sk, rexmit);
3738         }
3739 
3740         return 0;
3741 }
3742 
3743 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3744                                       bool syn, struct tcp_fastopen_cookie *foc,
3745                                       bool exp_opt)
3746 {
3747         /* Valid only in SYN or SYN-ACK with an even length.  */
3748         if (!foc || !syn || len < 0 || (len & 1))
3749                 return;
3750 
3751         if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3752             len <= TCP_FASTOPEN_COOKIE_MAX)
3753                 memcpy(foc->val, cookie, len);
3754         else if (len != 0)
3755                 len = -1;
3756         foc->len = len;
3757         foc->exp = exp_opt;
3758 }
3759 
3760 static void smc_parse_options(const struct tcphdr *th,
3761                               struct tcp_options_received *opt_rx,
3762                               const unsigned char *ptr,
3763                               int opsize)
3764 {
3765 #if IS_ENABLED(CONFIG_SMC)
3766         if (static_branch_unlikely(&tcp_have_smc)) {
3767                 if (th->syn && !(opsize & 1) &&
3768                     opsize >= TCPOLEN_EXP_SMC_BASE &&
3769                     get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3770                         opt_rx->smc_ok = 1;
3771         }
3772 #endif
3773 }
3774 
3775 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3776  * But, this can also be called on packets in the established flow when
3777  * the fast version below fails.
3778  */
3779 void tcp_parse_options(const struct net *net,
3780                        const struct sk_buff *skb,
3781                        struct tcp_options_received *opt_rx, int estab,
3782                        struct tcp_fastopen_cookie *foc)
3783 {
3784         const unsigned char *ptr;
3785         const struct tcphdr *th = tcp_hdr(skb);
3786         int length = (th->doff * 4) - sizeof(struct tcphdr);
3787 
3788         ptr = (const unsigned char *)(th + 1);
3789         opt_rx->saw_tstamp = 0;
3790 
3791         while (length > 0) {
3792                 int opcode = *ptr++;
3793                 int opsize;
3794 
3795                 switch (opcode) {
3796                 case TCPOPT_EOL:
3797                         return;
3798                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3799                         length--;
3800                         continue;
3801                 default:
3802                         opsize = *ptr++;
3803                         if (opsize < 2) /* "silly options" */
3804                                 return;
3805                         if (opsize > length)
3806                                 return; /* don't parse partial options */
3807                         switch (opcode) {
3808                         case TCPOPT_MSS:
3809                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3810                                         u16 in_mss = get_unaligned_be16(ptr);
3811                                         if (in_mss) {
3812                                                 if (opt_rx->user_mss &&
3813                                                     opt_rx->user_mss < in_mss)
3814                                                         in_mss = opt_rx->user_mss;
3815                                                 opt_rx->mss_clamp = in_mss;
3816                                         }
3817                                 }
3818                                 break;
3819                         case TCPOPT_WINDOW:
3820                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3821                                     !estab && net->ipv4.sysctl_tcp_window_scaling) {
3822                                         __u8 snd_wscale = *(__u8 *)ptr;
3823                                         opt_rx->wscale_ok = 1;
3824                                         if (snd_wscale > TCP_MAX_WSCALE) {
3825                                                 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3826                                                                      __func__,
3827                                                                      snd_wscale,
3828                                                                      TCP_MAX_WSCALE);
3829                                                 snd_wscale = TCP_MAX_WSCALE;
3830                                         }
3831                                         opt_rx->snd_wscale = snd_wscale;
3832                                 }
3833                                 break;
3834                         case TCPOPT_TIMESTAMP:
3835                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3836                                     ((estab && opt_rx->tstamp_ok) ||
3837                                      (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3838                                         opt_rx->saw_tstamp = 1;
3839                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3840                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3841                                 }
3842                                 break;
3843                         case TCPOPT_SACK_PERM:
3844                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3845                                     !estab && net->ipv4.sysctl_tcp_sack) {
3846                                         opt_rx->sack_ok = TCP_SACK_SEEN;
3847                                         tcp_sack_reset(opt_rx);
3848                                 }
3849                                 break;
3850 
3851                         case TCPOPT_SACK:
3852                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3853                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3854                                    opt_rx->sack_ok) {
3855                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3856                                 }
3857                                 break;
3858 #ifdef CONFIG_TCP_MD5SIG
3859                         case TCPOPT_MD5SIG:
3860                                 /*
3861                                  * The MD5 Hash has already been
3862                                  * checked (see tcp_v{4,6}_do_rcv()).
3863                                  */
3864                                 break;
3865 #endif
3866                         case TCPOPT_FASTOPEN:
3867                                 tcp_parse_fastopen_option(
3868                                         opsize - TCPOLEN_FASTOPEN_BASE,
3869                                         ptr, th->syn, foc, false);
3870                                 break;
3871 
3872                         case TCPOPT_EXP:
3873                                 /* Fast Open option shares code 254 using a
3874                                  * 16 bits magic number.
3875                                  */
3876                                 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3877                                     get_unaligned_be16(ptr) ==
3878                                     TCPOPT_FASTOPEN_MAGIC)
3879                                         tcp_parse_fastopen_option(opsize -
3880                                                 TCPOLEN_EXP_FASTOPEN_BASE,
3881                                                 ptr + 2, th->syn, foc, true);
3882                                 else
3883                                         smc_parse_options(th, opt_rx, ptr,
3884                                                           opsize);
3885                                 break;
3886 
3887                         }
3888                         ptr += opsize-2;
3889                         length -= opsize;
3890                 }
3891         }
3892 }
3893 EXPORT_SYMBOL(tcp_parse_options);
3894 
3895 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3896 {
3897         const __be32 *ptr = (const __be32 *)(th + 1);
3898 
3899         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3900                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3901                 tp->rx_opt.saw_tstamp = 1;
3902                 ++ptr;
3903                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3904                 ++ptr;
3905                 if (*ptr)
3906                         tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3907                 else
3908                         tp->rx_opt.rcv_tsecr = 0;
3909                 return true;
3910         }
3911         return false;
3912 }
3913 
3914 /* Fast parse options. This hopes to only see timestamps.
3915  * If it is wrong it falls back on tcp_parse_options().
3916  */
3917 static bool tcp_fast_parse_options(const struct net *net,
3918                                    const struct sk_buff *skb,
3919                                    const struct tcphdr *th, struct tcp_sock *tp)
3920 {
3921         /* In the spirit of fast parsing, compare doff directly to constant
3922          * values.  Because equality is used, short doff can be ignored here.
3923          */
3924         if (th->doff == (sizeof(*th) / 4)) {
3925                 tp->rx_opt.saw_tstamp = 0;
3926                 return false;
3927         } else if (tp->rx_opt.tstamp_ok &&
3928                    th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3929                 if (tcp_parse_aligned_timestamp(tp, th))
3930                         return true;
3931         }
3932 
3933         tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
3934         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3935                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3936 
3937         return true;
3938 }
3939 
3940 #ifdef CONFIG_TCP_MD5SIG
3941 /*
3942  * Parse MD5 Signature option
3943  */
3944 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3945 {
3946         int length = (th->doff << 2) - sizeof(*th);
3947         const u8 *ptr = (const u8 *)(th + 1);
3948 
3949         /* If not enough data remaining, we can short cut */
3950         while (length >= TCPOLEN_MD5SIG) {
3951                 int opcode = *ptr++;
3952                 int opsize;
3953 
3954                 switch (opcode) {
3955                 case TCPOPT_EOL:
3956                         return NULL;
3957                 case TCPOPT_NOP:
3958                         length--;
3959                         continue;
3960                 default:
3961                         opsize = *ptr++;
3962                         if (opsize < 2 || opsize > length)
3963                                 return NULL;
3964                         if (opcode == TCPOPT_MD5SIG)
3965                                 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3966                 }
3967                 ptr += opsize - 2;
3968                 length -= opsize;
3969         }
3970         return NULL;
3971 }
3972 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3973 #endif
3974 
3975 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3976  *
3977  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3978  * it can pass through stack. So, the following predicate verifies that
3979  * this segment is not used for anything but congestion avoidance or
3980  * fast retransmit. Moreover, we even are able to eliminate most of such
3981  * second order effects, if we apply some small "replay" window (~RTO)
3982  * to timestamp space.
3983  *
3984  * All these measures still do not guarantee that we reject wrapped ACKs
3985  * on networks with high bandwidth, when sequence space is recycled fastly,
3986  * but it guarantees that such events will be very rare and do not affect
3987  * connection seriously. This doesn't look nice, but alas, PAWS is really
3988  * buggy extension.
3989  *
3990  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3991  * states that events when retransmit arrives after original data are rare.
3992  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3993  * the biggest problem on large power networks even with minor reordering.
3994  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3995  * up to bandwidth of 18Gigabit/sec. 8) ]
3996  */
3997 
3998 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3999 {
4000         const struct tcp_sock *tp = tcp_sk(sk);
4001         const struct tcphdr *th = tcp_hdr(skb);
4002         u32 seq = TCP_SKB_CB(skb)->seq;
4003         u32 ack = TCP_SKB_CB(skb)->ack_seq;
4004 
4005         return (/* 1. Pure ACK with correct sequence number. */
4006                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4007 
4008                 /* 2. ... and duplicate ACK. */
4009                 ack == tp->snd_una &&
4010 
4011                 /* 3. ... and does not update window. */
4012                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4013 
4014                 /* 4. ... and sits in replay window. */
4015                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4016 }
4017 
4018 static inline bool tcp_paws_discard(const struct sock *sk,
4019                                    const struct sk_buff *skb)
4020 {
4021         const struct tcp_sock *tp = tcp_sk(sk);
4022 
4023         return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4024                !tcp_disordered_ack(sk, skb);
4025 }
4026 
4027 /* Check segment sequence number for validity.
4028  *
4029  * Segment controls are considered valid, if the segment
4030  * fits to the window after truncation to the window. Acceptability
4031  * of data (and SYN, FIN, of course) is checked separately.
4032  * See tcp_data_queue(), for example.
4033  *
4034  * Also, controls (RST is main one) are accepted using RCV.WUP instead
4035  * of RCV.NXT. Peer still did not advance his SND.UNA when we
4036  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4037  * (borrowed from freebsd)
4038  */
4039 
4040 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4041 {
4042         return  !before(end_seq, tp->rcv_wup) &&
4043                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4044 }
4045 
4046 /* When we get a reset we do this. */
4047 void tcp_reset(struct sock *sk)
4048 {
4049         trace_tcp_receive_reset(sk);
4050 
4051         /* We want the right error as BSD sees it (and indeed as we do). */
4052         switch (sk->sk_state) {
4053         case TCP_SYN_SENT:
4054                 sk->sk_err = ECONNREFUSED;
4055                 break;
4056         case TCP_CLOSE_WAIT:
4057                 sk->sk_err = EPIPE;
4058                 break;
4059         case TCP_CLOSE:
4060                 return;
4061         default:
4062                 sk->sk_err = ECONNRESET;
4063         }
4064         /* This barrier is coupled with smp_rmb() in tcp_poll() */
4065         smp_wmb();
4066 
4067         tcp_write_queue_purge(sk);
4068         tcp_done(sk);
4069 
4070         if (!sock_flag(sk, SOCK_DEAD))
4071                 sk->sk_error_report(sk);
4072 }
4073 
4074 /*
4075  *      Process the FIN bit. This now behaves as it is supposed to work
4076  *      and the FIN takes effect when it is validly part of sequence
4077  *      space. Not before when we get holes.
4078  *
4079  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4080  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
4081  *      TIME-WAIT)
4082  *
4083  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
4084  *      close and we go into CLOSING (and later onto TIME-WAIT)
4085  *
4086  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4087  */
4088 void tcp_fin(struct sock *sk)
4089 {
4090         struct tcp_sock *tp = tcp_sk(sk);
4091 
4092         inet_csk_schedule_ack(sk);
4093 
4094         sk->sk_shutdown |= RCV_SHUTDOWN;
4095         sock_set_flag(sk, SOCK_DONE);
4096 
4097         switch (sk->sk_state) {
4098         case TCP_SYN_RECV:
4099         case TCP_ESTABLISHED:
4100                 /* Move to CLOSE_WAIT */
4101                 tcp_set_state(sk, TCP_CLOSE_WAIT);
4102                 inet_csk_enter_pingpong_mode(sk);
4103                 break;
4104 
4105         case TCP_CLOSE_WAIT:
4106         case TCP_CLOSING:
4107                 /* Received a retransmission of the FIN, do
4108                  * nothing.
4109                  */
4110                 break;
4111         case TCP_LAST_ACK:
4112                 /* RFC793: Remain in the LAST-ACK state. */
4113                 break;
4114 
4115         case TCP_FIN_WAIT1:
4116                 /* This case occurs when a simultaneous close
4117                  * happens, we must ack the received FIN and
4118                  * enter the CLOSING state.
4119                  */
4120                 tcp_send_ack(sk);
4121                 tcp_set_state(sk, TCP_CLOSING);
4122                 break;
4123         case TCP_FIN_WAIT2:
4124                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4125                 tcp_send_ack(sk);
4126                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4127                 break;
4128         default:
4129                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4130                  * cases we should never reach this piece of code.
4131                  */
4132                 pr_err("%s: Impossible, sk->sk_state=%d\n",
4133                        __func__, sk->sk_state);
4134                 break;
4135         }
4136 
4137         /* It _is_ possible, that we have something out-of-order _after_ FIN.
4138          * Probably, we should reset in this case. For now drop them.
4139          */
4140         skb_rbtree_purge(&tp->out_of_order_queue);
4141         if (tcp_is_sack(tp))
4142                 tcp_sack_reset(&tp->rx_opt);
4143         sk_mem_reclaim(sk);
4144 
4145         if (!sock_flag(sk, SOCK_DEAD)) {
4146                 sk->sk_state_change(sk);
4147 
4148                 /* Do not send POLL_HUP for half duplex close. */
4149                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4150                     sk->sk_state == TCP_CLOSE)
4151                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4152                 else
4153                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4154         }
4155 }
4156 
4157 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4158                                   u32 end_seq)
4159 {
4160         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4161                 if (before(seq, sp->start_seq))
4162                         sp->start_seq = seq;
4163                 if (after(end_seq, sp->end_seq))
4164                         sp->end_seq = end_seq;
4165                 return true;
4166         }
4167         return false;
4168 }
4169 
4170 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4171 {
4172         struct tcp_sock *tp = tcp_sk(sk);
4173 
4174         if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4175                 int mib_idx;
4176 
4177                 if (before(seq, tp->rcv_nxt))
4178                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4179                 else
4180                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4181 
4182                 NET_INC_STATS(sock_net(sk), mib_idx);
4183 
4184                 tp->rx_opt.dsack = 1;
4185                 tp->duplicate_sack[0].start_seq = seq;
4186                 tp->duplicate_sack[0].end_seq = end_seq;
4187         }
4188 }
4189 
4190 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4191 {
4192         struct tcp_sock *tp = tcp_sk(sk);
4193 
4194         if (!tp->rx_opt.dsack)
4195                 tcp_dsack_set(sk, seq, end_seq);
4196         else
4197                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4198 }
4199 
4200 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4201 {
4202         /* When the ACK path fails or drops most ACKs, the sender would
4203          * timeout and spuriously retransmit the same segment repeatedly.
4204          * The receiver remembers and reflects via DSACKs. Leverage the
4205          * DSACK state and change the txhash to re-route speculatively.
4206          */
4207         if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq)
4208                 sk_rethink_txhash(sk);
4209 }
4210 
4211 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4212 {
4213         struct tcp_sock *tp = tcp_sk(sk);
4214 
4215         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4216             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4217                 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4218                 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4219 
4220                 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4221                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4222 
4223                         tcp_rcv_spurious_retrans(sk, skb);
4224                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4225                                 end_seq = tp->rcv_nxt;
4226                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4227                 }
4228         }
4229 
4230         tcp_send_ack(sk);
4231 }
4232 
4233 /* These routines update the SACK block as out-of-order packets arrive or
4234  * in-order packets close up the sequence space.
4235  */
4236 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4237 {
4238         int this_sack;
4239         struct tcp_sack_block *sp = &tp->selective_acks[0];
4240         struct tcp_sack_block *swalk = sp + 1;
4241 
4242         /* See if the recent change to the first SACK eats into
4243          * or hits the sequence space of other SACK blocks, if so coalesce.
4244          */
4245         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4246                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4247                         int i;
4248 
4249                         /* Zap SWALK, by moving every further SACK up by one slot.
4250                          * Decrease num_sacks.
4251                          */
4252                         tp->rx_opt.num_sacks--;
4253                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4254                                 sp[i] = sp[i + 1];
4255                         continue;
4256                 }
4257                 this_sack++, swalk++;
4258         }
4259 }
4260 
4261 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4262 {
4263         struct tcp_sock *tp = tcp_sk(sk);
4264         struct tcp_sack_block *sp = &tp->selective_acks[0];
4265         int cur_sacks = tp->rx_opt.num_sacks;
4266         int this_sack;
4267 
4268         if (!cur_sacks)
4269                 goto new_sack;
4270 
4271         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4272                 if (tcp_sack_extend(sp, seq, end_seq)) {
4273                         /* Rotate this_sack to the first one. */
4274                         for (; this_sack > 0; this_sack--, sp--)
4275                                 swap(*sp, *(sp - 1));
4276                         if (cur_sacks > 1)
4277                                 tcp_sack_maybe_coalesce(tp);
4278                         return;
4279                 }
4280         }
4281 
4282         /* Could not find an adjacent existing SACK, build a new one,
4283          * put it at the front, and shift everyone else down.  We
4284          * always know there is at least one SACK present already here.
4285          *
4286          * If the sack array is full, forget about the last one.
4287          */
4288         if (this_sack >= TCP_NUM_SACKS) {
4289                 if (tp->compressed_ack > TCP_FASTRETRANS_THRESH)
4290                         tcp_send_ack(sk);
4291                 this_sack--;
4292                 tp->rx_opt.num_sacks--;
4293                 sp--;
4294         }
4295         for (; this_sack > 0; this_sack--, sp--)
4296                 *sp = *(sp - 1);
4297 
4298 new_sack:
4299         /* Build the new head SACK, and we're done. */
4300         sp->start_seq = seq;
4301         sp->end_seq = end_seq;
4302         tp->rx_opt.num_sacks++;
4303 }
4304 
4305 /* RCV.NXT advances, some SACKs should be eaten. */
4306 
4307 static void tcp_sack_remove(struct tcp_sock *tp)
4308 {
4309         struct tcp_sack_block *sp = &tp->selective_acks[0];
4310         int num_sacks = tp->rx_opt.num_sacks;
4311         int this_sack;
4312 
4313         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4314         if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4315                 tp->rx_opt.num_sacks = 0;
4316                 return;
4317         }
4318 
4319         for (this_sack = 0; this_sack < num_sacks;) {
4320                 /* Check if the start of the sack is covered by RCV.NXT. */
4321                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4322                         int i;
4323 
4324                         /* RCV.NXT must cover all the block! */
4325                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4326 
4327                         /* Zap this SACK, by moving forward any other SACKS. */
4328                         for (i = this_sack+1; i < num_sacks; i++)
4329                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4330                         num_sacks--;
4331                         continue;
4332                 }
4333                 this_sack++;
4334                 sp++;
4335         }
4336         tp->rx_opt.num_sacks = num_sacks;
4337 }
4338 
4339 /**
4340  * tcp_try_coalesce - try to merge skb to prior one
4341  * @sk: socket
4342  * @dest: destination queue
4343  * @to: prior buffer
4344  * @from: buffer to add in queue
4345  * @fragstolen: pointer to boolean
4346  *
4347  * Before queueing skb @from after @to, try to merge them
4348  * to reduce overall memory use and queue lengths, if cost is small.
4349  * Packets in ofo or receive queues can stay a long time.
4350  * Better try to coalesce them right now to avoid future collapses.
4351  * Returns true if caller should free @from instead of queueing it
4352  */
4353 static bool tcp_try_coalesce(struct sock *sk,
4354                              struct sk_buff *to,
4355                              struct sk_buff *from,
4356                              bool *fragstolen)
4357 {
4358         int delta;
4359 
4360         *fragstolen = false;
4361 
4362         /* Its possible this segment overlaps with prior segment in queue */
4363         if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4364                 return false;
4365 
4366 #ifdef CONFIG_TLS_DEVICE
4367         if (from->decrypted != to->decrypted)
4368                 return false;
4369 #endif
4370 
4371         if (!skb_try_coalesce(to, from, fragstolen, &delta))
4372                 return false;
4373 
4374         atomic_add(delta, &sk->sk_rmem_alloc);
4375         sk_mem_charge(sk, delta);
4376         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4377         TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4378         TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4379         TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4380 
4381         if (TCP_SKB_CB(from)->has_rxtstamp) {
4382                 TCP_SKB_CB(to)->has_rxtstamp = true;
4383                 to->tstamp = from->tstamp;
4384                 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4385         }
4386 
4387         return true;
4388 }
4389 
4390 static bool tcp_ooo_try_coalesce(struct sock *sk,
4391                              struct sk_buff *to,
4392                              struct sk_buff *from,
4393                              bool *fragstolen)
4394 {
4395         bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4396 
4397         /* In case tcp_drop() is called later, update to->gso_segs */
4398         if (res) {
4399                 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4400                                max_t(u16, 1, skb_shinfo(from)->gso_segs);
4401 
4402                 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4403         }
4404         return res;
4405 }
4406 
4407 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4408 {
4409         sk_drops_add(sk, skb);
4410         __kfree_skb(skb);
4411 }
4412 
4413 /* This one checks to see if we can put data from the
4414  * out_of_order queue into the receive_queue.
4415  */
4416 static void tcp_ofo_queue(struct sock *sk)
4417 {
4418         struct tcp_sock *tp = tcp_sk(sk);
4419         __u32 dsack_high = tp->rcv_nxt;
4420         bool fin, fragstolen, eaten;
4421         struct sk_buff *skb, *tail;
4422         struct rb_node *p;
4423 
4424         p = rb_first(&tp->out_of_order_queue);
4425         while (p) {
4426                 skb = rb_to_skb(p);
4427                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4428                         break;
4429 
4430                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4431                         __u32 dsack = dsack_high;
4432                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4433                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4434                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4435                 }
4436                 p = rb_next(p);
4437                 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4438 
4439                 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4440                         tcp_drop(sk, skb);
4441                         continue;
4442                 }
4443 
4444                 tail = skb_peek_tail(&sk->sk_receive_queue);
4445                 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4446                 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4447                 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4448                 if (!eaten)
4449                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4450                 else
4451                         kfree_skb_partial(skb, fragstolen);
4452 
4453                 if (unlikely(fin)) {
4454                         tcp_fin(sk);
4455                         /* tcp_fin() purges tp->out_of_order_queue,
4456                          * so we must end this loop right now.
4457                          */
4458                         break;
4459                 }
4460         }
4461 }
4462 
4463 static bool tcp_prune_ofo_queue(struct sock *sk);
4464 static int tcp_prune_queue(struct sock *sk);
4465 
4466 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4467                                  unsigned int size)
4468 {
4469         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4470             !sk_rmem_schedule(sk, skb, size)) {
4471 
4472                 if (tcp_prune_queue(sk) < 0)
4473                         return -1;
4474 
4475                 while (!sk_rmem_schedule(sk, skb, size)) {
4476                         if (!tcp_prune_ofo_queue(sk))
4477                                 return -1;
4478                 }
4479         }
4480         return 0;
4481 }
4482 
4483 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4484 {
4485         struct tcp_sock *tp = tcp_sk(sk);
4486         struct rb_node **p, *parent;
4487         struct sk_buff *skb1;
4488         u32 seq, end_seq;
4489         bool fragstolen;
4490 
4491         tcp_ecn_check_ce(sk, skb);
4492 
4493         if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4494                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4495                 tcp_drop(sk, skb);
4496                 return;
4497         }
4498 
4499         /* Disable header prediction. */
4500         tp->pred_flags = 0;
4501         inet_csk_schedule_ack(sk);
4502 
4503         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4504         seq = TCP_SKB_CB(skb)->seq;
4505         end_seq = TCP_SKB_CB(skb)->end_seq;
4506 
4507         p = &tp->out_of_order_queue.rb_node;
4508         if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4509                 /* Initial out of order segment, build 1 SACK. */
4510                 if (tcp_is_sack(tp)) {
4511                         tp->rx_opt.num_sacks = 1;
4512                         tp->selective_acks[0].start_seq = seq;
4513                         tp->selective_acks[0].end_seq = end_seq;
4514                 }
4515                 rb_link_node(&skb->rbnode, NULL, p);
4516                 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4517                 tp->ooo_last_skb = skb;
4518                 goto end;
4519         }
4520 
4521         /* In the typical case, we are adding an skb to the end of the list.
4522          * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4523          */
4524         if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4525                                  skb, &fragstolen)) {
4526 coalesce_done:
4527                 tcp_grow_window(sk, skb);
4528                 kfree_skb_partial(skb, fragstolen);
4529                 skb = NULL;
4530                 goto add_sack;
4531         }
4532         /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4533         if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4534                 parent = &tp->ooo_last_skb->rbnode;
4535                 p = &parent->rb_right;
4536                 goto insert;
4537         }
4538 
4539         /* Find place to insert this segment. Handle overlaps on the way. */
4540         parent = NULL;
4541         while (*p) {
4542                 parent = *p;
4543                 skb1 = rb_to_skb(parent);
4544                 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4545                         p = &parent->rb_left;
4546                         continue;
4547                 }
4548                 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4549                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4550                                 /* All the bits are present. Drop. */
4551                                 NET_INC_STATS(sock_net(sk),
4552                                               LINUX_MIB_TCPOFOMERGE);
4553                                 tcp_drop(sk, skb);
4554                                 skb = NULL;
4555                                 tcp_dsack_set(sk, seq, end_seq);
4556                                 goto add_sack;
4557                         }
4558                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4559                                 /* Partial overlap. */
4560                                 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4561                         } else {
4562                                 /* skb's seq == skb1's seq and skb covers skb1.
4563                                  * Replace skb1 with skb.
4564                                  */
4565                                 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4566                                                 &tp->out_of_order_queue);
4567                                 tcp_dsack_extend(sk,
4568                                                  TCP_SKB_CB(skb1)->seq,
4569                                                  TCP_SKB_CB(skb1)->end_seq);
4570                                 NET_INC_STATS(sock_net(sk),
4571                                               LINUX_MIB_TCPOFOMERGE);
4572                                 tcp_drop(sk, skb1);
4573                                 goto merge_right;
4574                         }
4575                 } else if (tcp_ooo_try_coalesce(sk, skb1,
4576                                                 skb, &fragstolen)) {
4577                         goto coalesce_done;
4578                 }
4579                 p = &parent->rb_right;
4580         }
4581 insert:
4582         /* Insert segment into RB tree. */
4583         rb_link_node(&skb->rbnode, parent, p);
4584         rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4585 
4586 merge_right:
4587         /* Remove other segments covered by skb. */
4588         while ((skb1 = skb_rb_next(skb)) != NULL) {
4589                 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4590                         break;
4591                 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4592                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4593                                          end_seq);
4594                         break;
4595                 }
4596                 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4597                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4598                                  TCP_SKB_CB(skb1)->end_seq);
4599                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4600                 tcp_drop(sk, skb1);
4601         }
4602         /* If there is no skb after us, we are the last_skb ! */
4603         if (!skb1)
4604                 tp->ooo_last_skb = skb;
4605 
4606 add_sack:
4607         if (tcp_is_sack(tp))
4608                 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4609 end:
4610         if (skb) {
4611                 tcp_grow_window(sk, skb);
4612                 skb_condense(skb);
4613                 skb_set_owner_r(skb, sk);
4614         }
4615 }
4616 
4617 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4618                                       bool *fragstolen)
4619 {
4620         int eaten;
4621         struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4622 
4623         eaten = (tail &&
4624                  tcp_try_coalesce(sk, tail,
4625                                   skb, fragstolen)) ? 1 : 0;
4626         tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4627         if (!eaten) {
4628                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4629                 skb_set_owner_r(skb, sk);
4630         }
4631         return eaten;
4632 }
4633 
4634 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4635 {
4636         struct sk_buff *skb;
4637         int err = -ENOMEM;
4638         int data_len = 0;
4639         bool fragstolen;
4640 
4641         if (size == 0)
4642                 return 0;
4643 
4644         if (size > PAGE_SIZE) {
4645                 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4646 
4647                 data_len = npages << PAGE_SHIFT;
4648                 size = data_len + (size & ~PAGE_MASK);
4649         }
4650         skb = alloc_skb_with_frags(size - data_len, data_len,
4651                                    PAGE_ALLOC_COSTLY_ORDER,
4652                                    &err, sk->sk_allocation);
4653         if (!skb)
4654                 goto err;
4655 
4656         skb_put(skb, size - data_len);
4657         skb->data_len = data_len;
4658         skb->len = size;
4659 
4660         if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4661                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4662                 goto err_free;
4663         }
4664 
4665         err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4666         if (err)
4667                 goto err_free;
4668 
4669         TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4670         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4671         TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4672 
4673         if (tcp_queue_rcv(sk, skb, &fragstolen)) {
4674                 WARN_ON_ONCE(fragstolen); /* should not happen */
4675                 __kfree_skb(skb);
4676         }
4677         return size;
4678 
4679 err_free:
4680         kfree_skb(skb);
4681 err:
4682         return err;
4683 
4684 }
4685 
4686 void tcp_data_ready(struct sock *sk)
4687 {
4688         const struct tcp_sock *tp = tcp_sk(sk);
4689         int avail = tp->rcv_nxt - tp->copied_seq;
4690 
4691         if (avail < sk->sk_rcvlowat && !sock_flag(sk, SOCK_DONE))
4692                 return;
4693 
4694         sk->sk_data_ready(sk);
4695 }
4696 
4697 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4698 {
4699         struct tcp_sock *tp = tcp_sk(sk);
4700         bool fragstolen;
4701         int eaten;
4702 
4703         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4704                 __kfree_skb(skb);
4705                 return;
4706         }
4707         skb_dst_drop(skb);
4708         __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4709 
4710         tcp_ecn_accept_cwr(sk, skb);
4711 
4712         tp->rx_opt.dsack = 0;
4713 
4714         /*  Queue data for delivery to the user.
4715          *  Packets in sequence go to the receive queue.
4716          *  Out of sequence packets to the out_of_order_queue.
4717          */
4718         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4719                 if (tcp_receive_window(tp) == 0) {
4720                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4721                         goto out_of_window;
4722                 }
4723 
4724                 /* Ok. In sequence. In window. */
4725 queue_and_out:
4726                 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4727                         sk_forced_mem_schedule(sk, skb->truesize);
4728                 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4729                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4730                         goto drop;
4731                 }
4732 
4733                 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
4734                 if (skb->len)
4735                         tcp_event_data_recv(sk, skb);
4736                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4737                         tcp_fin(sk);
4738 
4739                 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4740                         tcp_ofo_queue(sk);
4741 
4742                         /* RFC5681. 4.2. SHOULD send immediate ACK, when
4743                          * gap in queue is filled.
4744                          */
4745                         if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4746                                 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
4747                 }
4748 
4749                 if (tp->rx_opt.num_sacks)
4750                         tcp_sack_remove(tp);
4751 
4752                 tcp_fast_path_check(sk);
4753 
4754                 if (eaten > 0)
4755                         kfree_skb_partial(skb, fragstolen);
4756                 if (!sock_flag(sk, SOCK_DEAD))
4757                         tcp_data_ready(sk);
4758                 return;
4759         }
4760 
4761         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4762                 tcp_rcv_spurious_retrans(sk, skb);
4763                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4764                 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4765                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4766 
4767 out_of_window:
4768                 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4769                 inet_csk_schedule_ack(sk);
4770 drop:
4771                 tcp_drop(sk, skb);
4772                 return;
4773         }
4774 
4775         /* Out of window. F.e. zero window probe. */
4776         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4777                 goto out_of_window;
4778 
4779         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4780                 /* Partial packet, seq < rcv_next < end_seq */
4781                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4782 
4783                 /* If window is closed, drop tail of packet. But after
4784                  * remembering D-SACK for its head made in previous line.
4785                  */
4786                 if (!tcp_receive_window(tp)) {
4787                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4788                         goto out_of_window;
4789                 }
4790                 goto queue_and_out;
4791         }
4792 
4793         tcp_data_queue_ofo(sk, skb);
4794 }
4795 
4796 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4797 {
4798         if (list)
4799                 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4800 
4801         return skb_rb_next(skb);
4802 }
4803 
4804 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4805                                         struct sk_buff_head *list,
4806                                         struct rb_root *root)
4807 {
4808         struct sk_buff *next = tcp_skb_next(skb, list);
4809 
4810         if (list)
4811                 __skb_unlink(skb, list);
4812         else
4813                 rb_erase(&skb->rbnode, root);
4814 
4815         __kfree_skb(skb);
4816         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4817 
4818         return next;
4819 }
4820 
4821 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4822 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4823 {
4824         struct rb_node **p = &root->rb_node;
4825         struct rb_node *parent = NULL;
4826         struct sk_buff *skb1;
4827 
4828         while (*p) {
4829                 parent = *p;
4830                 skb1 = rb_to_skb(parent);
4831                 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4832                         p = &parent->rb_left;
4833                 else
4834                         p = &parent->rb_right;
4835         }
4836         rb_link_node(&skb->rbnode, parent, p);
4837         rb_insert_color(&skb->rbnode, root);
4838 }
4839 
4840 /* Collapse contiguous sequence of skbs head..tail with
4841  * sequence numbers start..end.
4842  *
4843  * If tail is NULL, this means until the end of the queue.
4844  *
4845  * Segments with FIN/SYN are not collapsed (only because this
4846  * simplifies code)
4847  */
4848 static void
4849 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4850              struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4851 {
4852         struct sk_buff *skb = head, *n;
4853         struct sk_buff_head tmp;
4854         bool end_of_skbs;
4855 
4856         /* First, check that queue is collapsible and find
4857          * the point where collapsing can be useful.
4858          */
4859 restart:
4860         for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4861                 n = tcp_skb_next(skb, list);
4862 
4863                 /* No new bits? It is possible on ofo queue. */
4864                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4865                         skb = tcp_collapse_one(sk, skb, list, root);
4866                         if (!skb)
4867                                 break;
4868                         goto restart;
4869                 }
4870 
4871                 /* The first skb to collapse is:
4872                  * - not SYN/FIN and
4873                  * - bloated or contains data before "start" or
4874                  *   overlaps to the next one.
4875                  */
4876                 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4877                     (tcp_win_from_space(sk, skb->truesize) > skb->len ||
4878                      before(TCP_SKB_CB(skb)->seq, start))) {
4879                         end_of_skbs = false;
4880                         break;
4881                 }
4882 
4883                 if (n && n != tail &&
4884                     TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4885                         end_of_skbs = false;
4886                         break;
4887                 }
4888 
4889                 /* Decided to skip this, advance start seq. */
4890                 start = TCP_SKB_CB(skb)->end_seq;
4891         }
4892         if (end_of_skbs ||
4893             (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4894                 return;
4895 
4896         __skb_queue_head_init(&tmp);
4897 
4898         while (before(start, end)) {
4899                 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4900                 struct sk_buff *nskb;
4901 
4902                 nskb = alloc_skb(copy, GFP_ATOMIC);
4903                 if (!nskb)
4904                         break;
4905 
4906                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4907 #ifdef CONFIG_TLS_DEVICE
4908                 nskb->decrypted = skb->decrypted;
4909 #endif
4910                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4911                 if (list)
4912                         __skb_queue_before(list, skb, nskb);
4913                 else
4914                         __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4915                 skb_set_owner_r(nskb, sk);
4916 
4917                 /* Copy data, releasing collapsed skbs. */
4918                 while (copy > 0) {
4919                         int offset = start - TCP_SKB_CB(skb)->seq;
4920                         int size = TCP_SKB_CB(skb)->end_seq - start;
4921 
4922                         BUG_ON(offset < 0);
4923                         if (size > 0) {
4924                                 size = min(copy, size);
4925                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4926                                         BUG();
4927                                 TCP_SKB_CB(nskb)->end_seq += size;
4928                                 copy -= size;
4929                                 start += size;
4930                         }
4931                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4932                                 skb = tcp_collapse_one(sk, skb, list, root);
4933                                 if (!skb ||
4934                                     skb == tail ||
4935                                     (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4936                                         goto end;
4937 #ifdef CONFIG_TLS_DEVICE
4938                                 if (skb->decrypted != nskb->decrypted)
4939                                         goto end;
4940 #endif
4941                         }
4942                 }
4943         }
4944 end:
4945         skb_queue_walk_safe(&tmp, skb, n)
4946                 tcp_rbtree_insert(root, skb);
4947 }
4948 
4949 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4950  * and tcp_collapse() them until all the queue is collapsed.
4951  */
4952 static void tcp_collapse_ofo_queue(struct sock *sk)
4953 {
4954         struct tcp_sock *tp = tcp_sk(sk);
4955         u32 range_truesize, sum_tiny = 0;
4956         struct sk_buff *skb, *head;
4957         u32 start, end;
4958 
4959         skb = skb_rb_first(&tp->out_of_order_queue);
4960 new_range:
4961         if (!skb) {
4962                 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
4963                 return;
4964         }
4965         start = TCP_SKB_CB(skb)->seq;
4966         end = TCP_SKB_CB(skb)->end_seq;
4967         range_truesize = skb->truesize;
4968 
4969         for (head = skb;;) {
4970                 skb = skb_rb_next(skb);
4971 
4972                 /* Range is terminated when we see a gap or when
4973                  * we are at the queue end.
4974                  */
4975                 if (!skb ||
4976                     after(TCP_SKB_CB(skb)->seq, end) ||
4977                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4978                         /* Do not attempt collapsing tiny skbs */
4979                         if (range_truesize != head->truesize ||
4980                             end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
4981                                 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4982                                              head, skb, start, end);
4983                         } else {
4984                                 sum_tiny += range_truesize;
4985                                 if (sum_tiny > sk->sk_rcvbuf >> 3)
4986                                         return;
4987                         }
4988                         goto new_range;
4989                 }
4990 
4991                 range_truesize += skb->truesize;
4992                 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4993                         start = TCP_SKB_CB(skb)->seq;
4994                 if (after(TCP_SKB_CB(skb)->end_seq, end))
4995                         end = TCP_SKB_CB(skb)->end_seq;
4996         }
4997 }
4998 
4999 /*
5000  * Clean the out-of-order queue to make room.
5001  * We drop high sequences packets to :
5002  * 1) Let a chance for holes to be filled.
5003  * 2) not add too big latencies if thousands of packets sit there.
5004  *    (But if application shrinks SO_RCVBUF, we could still end up
5005  *     freeing whole queue here)
5006  * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5007  *
5008  * Return true if queue has shrunk.
5009  */
5010 static bool tcp_prune_ofo_queue(struct sock *sk)
5011 {
5012         struct tcp_sock *tp = tcp_sk(sk);
5013         struct rb_node *node, *prev;
5014         int goal;
5015 
5016         if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5017                 return false;
5018 
5019         NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5020         goal = sk->sk_rcvbuf >> 3;
5021         node = &tp->ooo_last_skb->rbnode;
5022         do {
5023                 prev = rb_prev(node);
5024                 rb_erase(node, &tp->out_of_order_queue);
5025                 goal -= rb_to_skb(node)->truesize;
5026                 tcp_drop(sk, rb_to_skb(node));
5027                 if (!prev || goal <= 0) {
5028                         sk_mem_reclaim(sk);
5029                         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5030                             !tcp_under_memory_pressure(sk))
5031                                 break;
5032                         goal = sk->sk_rcvbuf >> 3;
5033                 }
5034                 node = prev;
5035         } while (node);
5036         tp->ooo_last_skb = rb_to_skb(prev);
5037 
5038         /* Reset SACK state.  A conforming SACK implementation will
5039          * do the same at a timeout based retransmit.  When a connection
5040          * is in a sad state like this, we care only about integrity
5041          * of the connection not performance.
5042          */
5043         if (tp->rx_opt.sack_ok)
5044                 tcp_sack_reset(&tp->rx_opt);
5045         return true;
5046 }
5047 
5048 /* Reduce allocated memory if we can, trying to get
5049  * the socket within its memory limits again.
5050  *
5051  * Return less than zero if we should start dropping frames
5052  * until the socket owning process reads some of the data
5053  * to stabilize the situation.
5054  */
5055 static int tcp_prune_queue(struct sock *sk)
5056 {
5057         struct tcp_sock *tp = tcp_sk(sk);
5058 
5059         NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5060 
5061         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5062                 tcp_clamp_window(sk);
5063         else if (tcp_under_memory_pressure(sk))
5064                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5065 
5066         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5067                 return 0;
5068 
5069         tcp_collapse_ofo_queue(sk);
5070         if (!skb_queue_empty(&sk->sk_receive_queue))
5071                 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5072                              skb_peek(&sk->sk_receive_queue),
5073                              NULL,
5074                              tp->copied_seq, tp->rcv_nxt);
5075         sk_mem_reclaim(sk);
5076 
5077         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5078                 return 0;
5079 
5080         /* Collapsing did not help, destructive actions follow.
5081          * This must not ever occur. */
5082 
5083         tcp_prune_ofo_queue(sk);
5084 
5085         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5086                 return 0;
5087 
5088         /* If we are really being abused, tell the caller to silently
5089          * drop receive data on the floor.  It will get retransmitted
5090          * and hopefully then we'll have sufficient space.
5091          */
5092         NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5093 
5094         /* Massive buffer overcommit. */
5095         tp->pred_flags = 0;
5096         return -1;
5097 }
5098 
5099 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5100 {
5101         const struct tcp_sock *tp = tcp_sk(sk);
5102 
5103         /* If the user specified a specific send buffer setting, do
5104          * not modify it.
5105          */
5106         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5107                 return false;
5108 
5109         /* If we are under global TCP memory pressure, do not expand.  */
5110         if (tcp_under_memory_pressure(sk))
5111                 return false;
5112 
5113         /* If we are under soft global TCP memory pressure, do not expand.  */
5114         if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5115                 return false;
5116 
5117         /* If we filled the congestion window, do not expand.  */
5118         if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5119                 return false;
5120 
5121         return true;
5122 }
5123 
5124 /* When incoming ACK allowed to free some skb from write_queue,
5125  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5126  * on the exit from tcp input handler.
5127  *
5128  * PROBLEM: sndbuf expansion does not work well with largesend.
5129  */
5130 static void tcp_new_space(struct sock *sk)
5131 {
5132         struct tcp_sock *tp = tcp_sk(sk);
5133 
5134         if (tcp_should_expand_sndbuf(sk)) {
5135                 tcp_sndbuf_expand(sk);
5136                 tp->snd_cwnd_stamp = tcp_jiffies32;
5137         }
5138 
5139         sk->sk_write_space(sk);
5140 }
5141 
5142 static void tcp_check_space(struct sock *sk)
5143 {
5144         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5145                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5146                 /* pairs with tcp_poll() */
5147                 smp_mb();
5148                 if (sk->sk_socket &&
5149                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5150                         tcp_new_space(sk);
5151                         if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5152                                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5153                 }
5154         }
5155 }
5156 
5157 static inline void tcp_data_snd_check(struct sock *sk)
5158 {
5159         tcp_push_pending_frames(sk);
5160         tcp_check_space(sk);
5161 }
5162 
5163 /*
5164  * Check if sending an ack is needed.
5165  */
5166 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5167 {
5168         struct tcp_sock *tp = tcp_sk(sk);
5169         unsigned long rtt, delay;
5170 
5171             /* More than one full frame received... */
5172         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5173              /* ... and right edge of window advances far enough.
5174               * (tcp_recvmsg() will send ACK otherwise).
5175               * If application uses SO_RCVLOWAT, we want send ack now if
5176               * we have not received enough bytes to satisfy the condition.
5177               */
5178             (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5179              __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5180             /* We ACK each frame or... */
5181             tcp_in_quickack_mode(sk) ||
5182             /* Protocol state mandates a one-time immediate ACK */
5183             inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5184 send_now:
5185                 tcp_send_ack(sk);
5186                 return;
5187         }
5188 
5189         if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5190                 tcp_send_delayed_ack(sk);
5191                 return;
5192         }
5193 
5194         if (!tcp_is_sack(tp) ||
5195             tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)
5196                 goto send_now;
5197 
5198         if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5199                 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5200                 if (tp->compressed_ack > TCP_FASTRETRANS_THRESH)
5201                         NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
5202                                       tp->compressed_ack - TCP_FASTRETRANS_THRESH);
5203                 tp->compressed_ack = 0;
5204         }
5205 
5206         if (++tp->compressed_ack <= TCP_FASTRETRANS_THRESH)
5207                 goto send_now;
5208 
5209         if (hrtimer_is_queued(&tp->compressed_ack_timer))
5210                 return;
5211 
5212         /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5213 
5214         rtt = tp->rcv_rtt_est.rtt_us;
5215         if (tp->srtt_us && tp->srtt_us < rtt)
5216                 rtt = tp->srtt_us;
5217 
5218         delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns,
5219                       rtt * (NSEC_PER_USEC >> 3)/20);
5220         sock_hold(sk);
5221         hrtimer_start(&tp->compressed_ack_timer, ns_to_ktime(delay),
5222                       HRTIMER_MODE_REL_PINNED_SOFT);
5223 }
5224 
5225 static inline void tcp_ack_snd_check(struct sock *sk)
5226 {
5227         if (!inet_csk_ack_scheduled(sk)) {
5228                 /* We sent a data segment already. */
5229                 return;
5230         }
5231         __tcp_ack_snd_check(sk, 1);
5232 }
5233 
5234 /*
5235  *      This routine is only called when we have urgent data
5236  *      signaled. Its the 'slow' part of tcp_urg. It could be
5237  *      moved inline now as tcp_urg is only called from one
5238  *      place. We handle URGent data wrong. We have to - as
5239  *      BSD still doesn't use the correction from RFC961.
5240  *      For 1003.1g we should support a new option TCP_STDURG to permit
5241  *      either form (or just set the sysctl tcp_stdurg).
5242  */
5243 
5244 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5245 {
5246         struct tcp_sock *tp = tcp_sk(sk);
5247         u32 ptr = ntohs(th->urg_ptr);
5248 
5249         if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5250                 ptr--;
5251         ptr += ntohl(th->seq);
5252 
5253         /* Ignore urgent data that we've already seen and read. */
5254         if (after(tp->copied_seq, ptr))
5255                 return;
5256 
5257         /* Do not replay urg ptr.
5258          *
5259          * NOTE: interesting situation not covered by specs.
5260          * Misbehaving sender may send urg ptr, pointing to segment,
5261          * which we already have in ofo queue. We are not able to fetch
5262          * such data and will stay in TCP_URG_NOTYET until will be eaten
5263          * by recvmsg(). Seems, we are not obliged to handle such wicked
5264          * situations. But it is worth to think about possibility of some
5265          * DoSes using some hypothetical application level deadlock.
5266          */
5267         if (before(ptr, tp->rcv_nxt))
5268                 return;
5269 
5270         /* Do we already have a newer (or duplicate) urgent pointer? */
5271         if (tp->urg_data && !after(ptr, tp->urg_seq))
5272                 return;
5273 
5274         /* Tell the world about our new urgent pointer. */
5275         sk_send_sigurg(sk);
5276 
5277         /* We may be adding urgent data when the last byte read was
5278          * urgent. To do this requires some care. We cannot just ignore
5279          * tp->copied_seq since we would read the last urgent byte again
5280          * as data, nor can we alter copied_seq until this data arrives
5281          * or we break the semantics of SIOCATMARK (and thus sockatmark())
5282          *
5283          * NOTE. Double Dutch. Rendering to plain English: author of comment
5284          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
5285          * and expect that both A and B disappear from stream. This is _wrong_.
5286          * Though this happens in BSD with high probability, this is occasional.
5287          * Any application relying on this is buggy. Note also, that fix "works"
5288          * only in this artificial test. Insert some normal data between A and B and we will
5289          * decline of BSD again. Verdict: it is better to remove to trap
5290          * buggy users.
5291          */
5292         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5293             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5294                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5295                 tp->copied_seq++;
5296                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5297                         __skb_unlink(skb, &sk->sk_receive_queue);
5298                         __kfree_skb(skb);
5299                 }
5300         }
5301 
5302         tp->urg_data = TCP_URG_NOTYET;
5303         tp->urg_seq = ptr;
5304 
5305         /* Disable header prediction. */
5306         tp->pred_flags = 0;
5307 }
5308 
5309 /* This is the 'fast' part of urgent handling. */
5310 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5311 {
5312         struct tcp_sock *tp = tcp_sk(sk);
5313 
5314         /* Check if we get a new urgent pointer - normally not. */
5315         if (th->urg)
5316                 tcp_check_urg(sk, th);
5317 
5318         /* Do we wait for any urgent data? - normally not... */
5319         if (tp->urg_data == TCP_URG_NOTYET) {
5320                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5321                           th->syn;
5322 
5323                 /* Is the urgent pointer pointing into this packet? */
5324                 if (ptr < skb->len) {
5325                         u8 tmp;
5326                         if (skb_copy_bits(skb, ptr, &tmp, 1))
5327                                 BUG();
5328                         tp->urg_data = TCP_URG_VALID | tmp;
5329                         if (!sock_flag(sk, SOCK_DEAD))
5330                                 sk->sk_data_ready(sk);
5331                 }
5332         }
5333 }
5334 
5335 /* Accept RST for rcv_nxt - 1 after a FIN.
5336  * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5337  * FIN is sent followed by a RST packet. The RST is sent with the same
5338  * sequence number as the FIN, and thus according to RFC 5961 a challenge
5339  * ACK should be sent. However, Mac OSX rate limits replies to challenge
5340  * ACKs on the closed socket. In addition middleboxes can drop either the
5341  * challenge ACK or a subsequent RST.
5342  */
5343 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5344 {
5345         struct tcp_sock *tp = tcp_sk(sk);
5346 
5347         return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5348                         (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5349                                                TCPF_CLOSING));
5350 }
5351 
5352 /* Does PAWS and seqno based validation of an incoming segment, flags will
5353  * play significant role here.
5354  */
5355 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5356                                   const struct tcphdr *th, int syn_inerr)
5357 {
5358         struct tcp_sock *tp = tcp_sk(sk);
5359         bool rst_seq_match = false;
5360 
5361         /* RFC1323: H1. Apply PAWS check first. */
5362         if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5363             tp->rx_opt.saw_tstamp &&
5364             tcp_paws_discard(sk, skb)) {
5365                 if (!th->rst) {
5366                         NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5367                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5368                                                   LINUX_MIB_TCPACKSKIPPEDPAWS,
5369                                                   &tp->last_oow_ack_time))
5370                                 tcp_send_dupack(sk, skb);
5371                         goto discard;
5372                 }
5373                 /* Reset is accepted even if it did not pass PAWS. */
5374         }
5375 
5376         /* Step 1: check sequence number */
5377         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5378                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5379                  * (RST) segments are validated by checking their SEQ-fields."
5380                  * And page 69: "If an incoming segment is not acceptable,
5381                  * an acknowledgment should be sent in reply (unless the RST
5382                  * bit is set, if so drop the segment and return)".
5383                  */
5384                 if (!th->rst) {
5385                         if (th->syn)
5386                                 goto syn_challenge;
5387                         if (!tcp_oow_rate_limited(sock_net(sk), skb,
5388                                                   LINUX_MIB_TCPACKSKIPPEDSEQ,
5389                                                   &tp->last_oow_ack_time))
5390                                 tcp_send_dupack(sk, skb);
5391                 } else if (tcp_reset_check(sk, skb)) {
5392                         tcp_reset(sk);
5393                 }
5394                 goto discard;
5395         }
5396 
5397         /* Step 2: check RST bit */
5398         if (th->rst) {
5399                 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5400                  * FIN and SACK too if available):
5401                  * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5402                  * the right-most SACK block,
5403                  * then
5404                  *     RESET the connection
5405                  * else
5406                  *     Send a challenge ACK
5407                  */
5408                 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5409                     tcp_reset_check(sk, skb)) {
5410                         rst_seq_match = true;
5411                 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5412                         struct tcp_sack_block *sp = &tp->selective_acks[0];
5413                         int max_sack = sp[0].end_seq;
5414                         int this_sack;
5415 
5416                         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5417                              ++this_sack) {
5418                                 max_sack = after(sp[this_sack].end_seq,
5419                                                  max_sack) ?
5420                                         sp[this_sack].end_seq : max_sack;
5421                         }
5422 
5423                         if (TCP_SKB_CB(skb)->seq == max_sack)
5424                                 rst_seq_match = true;
5425                 }
5426 
5427                 if (rst_seq_match)
5428                         tcp_reset(sk);
5429                 else {
5430                         /* Disable TFO if RST is out-of-order
5431                          * and no data has been received
5432                          * for current active TFO socket
5433                          */
5434                         if (tp->syn_fastopen && !tp->data_segs_in &&
5435                             sk->sk_state == TCP_ESTABLISHED)
5436                                 tcp_fastopen_active_disable(sk);
5437                         tcp_send_challenge_ack(sk, skb);
5438                 }
5439                 goto discard;
5440         }
5441 
5442         /* step 3: check security and precedence [ignored] */
5443 
5444         /* step 4: Check for a SYN
5445          * RFC 5961 4.2 : Send a challenge ack
5446          */
5447         if (th->syn) {
5448 syn_challenge:
5449                 if (syn_inerr)
5450                         TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5451                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5452                 tcp_send_challenge_ack(sk, skb);
5453                 goto discard;
5454         }
5455 
5456         return true;
5457 
5458 discard:
5459         tcp_drop(sk, skb);
5460         return false;
5461 }
5462 
5463 /*
5464  *      TCP receive function for the ESTABLISHED state.
5465  *
5466  *      It is split into a fast path and a slow path. The fast path is
5467  *      disabled when:
5468  *      - A zero window was announced from us - zero window probing
5469  *        is only handled properly in the slow path.
5470  *      - Out of order segments arrived.
5471  *      - Urgent data is expected.
5472  *      - There is no buffer space left
5473  *      - Unexpected TCP flags/window values/header lengths are received
5474  *        (detected by checking the TCP header against pred_flags)
5475  *      - Data is sent in both directions. Fast path only supports pure senders
5476  *        or pure receivers (this means either the sequence number or the ack
5477  *        value must stay constant)
5478  *      - Unexpected TCP option.
5479  *
5480  *      When these conditions are not satisfied it drops into a standard
5481  *      receive procedure patterned after RFC793 to handle all cases.
5482  *      The first three cases are guaranteed by proper pred_flags setting,
5483  *      the rest is checked inline. Fast processing is turned on in
5484  *      tcp_data_queue when everything is OK.
5485  */
5486 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5487 {
5488         const struct tcphdr *th = (const struct tcphdr *)skb->data;
5489         struct tcp_sock *tp = tcp_sk(sk);
5490         unsigned int len = skb->len;
5491 
5492         /* TCP congestion window tracking */
5493         trace_tcp_probe(sk, skb);
5494 
5495         tcp_mstamp_refresh(tp);
5496         if (unlikely(!sk->sk_rx_dst))
5497                 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5498         /*
5499          *      Header prediction.
5500          *      The code loosely follows the one in the famous
5501          *      "30 instruction TCP receive" Van Jacobson mail.
5502          *
5503          *      Van's trick is to deposit buffers into socket queue
5504          *      on a device interrupt, to call tcp_recv function
5505          *      on the receive process context and checksum and copy
5506          *      the buffer to user space. smart...
5507          *
5508          *      Our current scheme is not silly either but we take the
5509          *      extra cost of the net_bh soft interrupt processing...
5510          *      We do checksum and copy also but from device to kernel.
5511          */
5512 
5513         tp->rx_opt.saw_tstamp = 0;
5514 
5515         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5516          *      if header_prediction is to be made
5517          *      'S' will always be tp->tcp_header_len >> 2
5518          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5519          *  turn it off (when there are holes in the receive
5520          *       space for instance)
5521          *      PSH flag is ignored.
5522          */
5523 
5524         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5525             TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5526             !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5527                 int tcp_header_len = tp->tcp_header_len;
5528 
5529                 /* Timestamp header prediction: tcp_header_len
5530                  * is automatically equal to th->doff*4 due to pred_flags
5531                  * match.
5532                  */
5533 
5534                 /* Check timestamp */
5535                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5536                         /* No? Slow path! */
5537                         if (!tcp_parse_aligned_timestamp(tp, th))
5538                                 goto slow_path;
5539 
5540                         /* If PAWS failed, check it more carefully in slow path */
5541                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5542                                 goto slow_path;
5543 
5544                         /* DO NOT update ts_recent here, if checksum fails
5545                          * and timestamp was corrupted part, it will result
5546                          * in a hung connection since we will drop all
5547                          * future packets due to the PAWS test.
5548                          */
5549                 }
5550 
5551                 if (len <= tcp_header_len) {
5552                         /* Bulk data transfer: sender */
5553                         if (len == tcp_header_len) {
5554                                 /* Predicted packet is in window by definition.
5555                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5556                                  * Hence, check seq<=rcv_wup reduces to:
5557                                  */
5558                                 if (tcp_header_len ==
5559                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5560                                     tp->rcv_nxt == tp->rcv_wup)
5561                                         tcp_store_ts_recent(tp);
5562 
5563                                 /* We know that such packets are checksummed
5564                                  * on entry.
5565                                  */
5566                                 tcp_ack(sk, skb, 0);
5567                                 __kfree_skb(skb);
5568                                 tcp_data_snd_check(sk);
5569                                 /* When receiving pure ack in fast path, update
5570                                  * last ts ecr directly instead of calling
5571                                  * tcp_rcv_rtt_measure_ts()
5572                                  */
5573                                 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5574                                 return;
5575                         } else { /* Header too small */
5576                                 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5577                                 goto discard;
5578                         }
5579                 } else {
5580                         int eaten = 0;
5581                         bool fragstolen = false;
5582 
5583                         if (tcp_checksum_complete(skb))
5584                                 goto csum_error;
5585 
5586                         if ((int)skb->truesize > sk->sk_forward_alloc)
5587                                 goto step5;
5588 
5589                         /* Predicted packet is in window by definition.
5590                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5591                          * Hence, check seq<=rcv_wup reduces to:
5592                          */
5593                         if (tcp_header_len ==
5594                             (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5595                             tp->rcv_nxt == tp->rcv_wup)
5596                                 tcp_store_ts_recent(tp);
5597 
5598                         tcp_rcv_rtt_measure_ts(sk, skb);
5599 
5600                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5601 
5602                         /* Bulk data transfer: receiver */
5603                         __skb_pull(skb, tcp_header_len);
5604                         eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5605 
5606                         tcp_event_data_recv(sk, skb);
5607 
5608                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5609                                 /* Well, only one small jumplet in fast path... */
5610                                 tcp_ack(sk, skb, FLAG_DATA);
5611                                 tcp_data_snd_check(sk);
5612                                 if (!inet_csk_ack_scheduled(sk))
5613                                         goto no_ack;
5614                         }
5615 
5616                         __tcp_ack_snd_check(sk, 0);
5617 no_ack:
5618                         if (eaten)
5619                                 kfree_skb_partial(skb, fragstolen);
5620                         tcp_data_ready(sk);
5621                         return;
5622                 }
5623         }
5624 
5625 slow_path:
5626         if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5627                 goto csum_error;
5628 
5629         if (!th->ack && !th->rst && !th->syn)
5630                 goto discard;
5631 
5632         /*
5633          *      Standard slow path.
5634          */
5635 
5636         if (!tcp_validate_incoming(sk, skb, th, 1))
5637                 return;
5638 
5639 step5:
5640         if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5641                 goto discard;
5642 
5643         tcp_rcv_rtt_measure_ts(sk, skb);
5644 
5645         /* Process urgent data. */
5646         tcp_urg(sk, skb, th);
5647 
5648         /* step 7: process the segment text */
5649         tcp_data_queue(sk, skb);
5650 
5651         tcp_data_snd_check(sk);
5652         tcp_ack_snd_check(sk);
5653         return;
5654 
5655 csum_error:
5656         TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5657         TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5658 
5659 discard:
5660         tcp_drop(sk, skb);
5661 }
5662 EXPORT_SYMBOL(tcp_rcv_established);
5663 
5664 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5665 {
5666         struct tcp_sock *tp = tcp_sk(sk);
5667         struct inet_connection_sock *icsk = inet_csk(sk);
5668 
5669         tcp_set_state(sk, TCP_ESTABLISHED);
5670         icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5671 
5672         if (skb) {
5673                 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5674                 security_inet_conn_established(sk, skb);
5675                 sk_mark_napi_id(sk, skb);
5676         }
5677 
5678         tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5679 
5680         /* Prevent spurious tcp_cwnd_restart() on first data
5681          * packet.
5682          */
5683         tp->lsndtime = tcp_jiffies32;
5684 
5685         if (sock_flag(sk, SOCK_KEEPOPEN))
5686                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5687 
5688         if (!tp->rx_opt.snd_wscale)
5689                 __tcp_fast_path_on(tp, tp->snd_wnd);
5690         else
5691                 tp->pred_flags = 0;
5692 }
5693 
5694 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5695                                     struct tcp_fastopen_cookie *cookie)
5696 {
5697         struct tcp_sock *tp = tcp_sk(sk);
5698         struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5699         u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5700         bool syn_drop = false;
5701 
5702         if (mss == tp->rx_opt.user_mss) {
5703                 struct tcp_options_received opt;
5704 
5705                 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5706                 tcp_clear_options(&opt);
5707                 opt.user_mss = opt.mss_clamp = 0;
5708                 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5709                 mss = opt.mss_clamp;
5710         }
5711 
5712         if (!tp->syn_fastopen) {
5713                 /* Ignore an unsolicited cookie */
5714                 cookie->len = -1;
5715         } else if (tp->total_retrans) {
5716                 /* SYN timed out and the SYN-ACK neither has a cookie nor
5717                  * acknowledges data. Presumably the remote received only
5718                  * the retransmitted (regular) SYNs: either the original
5719                  * SYN-data or the corresponding SYN-ACK was dropped.
5720                  */
5721                 syn_drop = (cookie->len < 0 && data);
5722         } else if (cookie->len < 0 && !tp->syn_data) {
5723                 /* We requested a cookie but didn't get it. If we did not use
5724                  * the (old) exp opt format then try so next time (try_exp=1).
5725                  * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5726                  */
5727                 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5728         }
5729 
5730         tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5731 
5732         if (data) { /* Retransmit unacked data in SYN */
5733                 skb_rbtree_walk_from(data) {
5734                         if (__tcp_retransmit_skb(sk, data, 1))
5735                                 break;
5736                 }
5737                 tcp_rearm_rto(sk);
5738                 NET_INC_STATS(sock_net(sk),
5739                                 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5740                 return true;
5741         }
5742         tp->syn_data_acked = tp->syn_data;
5743         if (tp->syn_data_acked) {
5744                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5745                 /* SYN-data is counted as two separate packets in tcp_ack() */
5746                 if (tp->delivered > 1)
5747                         --tp->delivered;
5748         }
5749 
5750         tcp_fastopen_add_skb(sk, synack);
5751 
5752         return false;
5753 }
5754 
5755 static void smc_check_reset_syn(struct tcp_sock *tp)
5756 {
5757 #if IS_ENABLED(CONFIG_SMC)
5758         if (static_branch_unlikely(&tcp_have_smc)) {
5759                 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5760                         tp->syn_smc = 0;
5761         }
5762 #endif
5763 }
5764 
5765 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5766                                          const struct tcphdr *th)
5767 {
5768         struct inet_connection_sock *icsk = inet_csk(sk);
5769         struct tcp_sock *tp = tcp_sk(sk);
5770         struct tcp_fastopen_cookie foc = { .len = -1 };
5771         int saved_clamp = tp->rx_opt.mss_clamp;
5772         bool fastopen_fail;
5773 
5774         tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5775         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5776                 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5777 
5778         if (th->ack) {
5779                 /* rfc793:
5780                  * "If the state is SYN-SENT then
5781                  *    first check the ACK bit
5782                  *      If the ACK bit is set
5783                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5784                  *        a reset (unless the RST bit is set, if so drop
5785                  *        the segment and return)"
5786                  */
5787                 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5788                     after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5789                         goto reset_and_undo;
5790 
5791                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5792                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5793                              tcp_time_stamp(tp))) {
5794                         NET_INC_STATS(sock_net(sk),
5795                                         LINUX_MIB_PAWSACTIVEREJECTED);
5796                         goto reset_and_undo;
5797                 }
5798 
5799                 /* Now ACK is acceptable.
5800                  *
5801                  * "If the RST bit is set
5802                  *    If the ACK was acceptable then signal the user "error:
5803                  *    connection reset", drop the segment, enter CLOSED state,
5804                  *    delete TCB, and return."
5805                  */
5806 
5807                 if (th->rst) {
5808                         tcp_reset(sk);
5809                         goto discard;
5810                 }
5811 
5812                 /* rfc793:
5813                  *   "fifth, if neither of the SYN or RST bits is set then
5814                  *    drop the segment and return."
5815                  *
5816                  *    See note below!
5817                  *                                        --ANK(990513)
5818                  */
5819                 if (!th->syn)
5820                         goto discard_and_undo;
5821 
5822                 /* rfc793:
5823                  *   "If the SYN bit is on ...
5824                  *    are acceptable then ...
5825                  *    (our SYN has been ACKed), change the connection
5826                  *    state to ESTABLISHED..."
5827                  */
5828 
5829                 tcp_ecn_rcv_synack(tp, th);
5830 
5831                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5832                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5833 
5834                 /* Ok.. it's good. Set up sequence numbers and
5835                  * move to established.
5836                  */
5837                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5838                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5839 
5840                 /* RFC1323: The window in SYN & SYN/ACK segments is
5841                  * never scaled.
5842                  */
5843                 tp->snd_wnd = ntohs(th->window);
5844 
5845                 if (!tp->rx_opt.wscale_ok) {
5846                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5847                         tp->window_clamp = min(tp->window_clamp, 65535U);
5848                 }
5849 
5850                 if (tp->rx_opt.saw_tstamp) {
5851                         tp->rx_opt.tstamp_ok       = 1;
5852                         tp->tcp_header_len =
5853                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5854                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5855                         tcp_store_ts_recent(tp);
5856                 } else {
5857                         tp->tcp_header_len = sizeof(struct tcphdr);
5858                 }
5859 
5860                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5861                 tcp_initialize_rcv_mss(sk);
5862 
5863                 /* Remember, tcp_poll() does not lock socket!
5864                  * Change state from SYN-SENT only after copied_seq
5865                  * is initialized. */
5866                 tp->copied_seq = tp->rcv_nxt;
5867 
5868                 smc_check_reset_syn(tp);
5869 
5870                 smp_mb();
5871 
5872                 tcp_finish_connect(sk, skb);
5873 
5874                 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5875                                 tcp_rcv_fastopen_synack(sk, skb, &foc);
5876 
5877                 if (!sock_flag(sk, SOCK_DEAD)) {
5878                         sk->sk_state_change(sk);
5879                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5880                 }
5881                 if (fastopen_fail)
5882                         return -1;
5883                 if (sk->sk_write_pending ||
5884                     icsk->icsk_accept_queue.rskq_defer_accept ||
5885                     inet_csk_in_pingpong_mode(sk)) {
5886                         /* Save one ACK. Data will be ready after
5887                          * several ticks, if write_pending is set.
5888                          *
5889                          * It may be deleted, but with this feature tcpdumps
5890                          * look so _wonderfully_ clever, that I was not able
5891                          * to stand against the temptation 8)     --ANK
5892                          */
5893                         inet_csk_schedule_ack(sk);
5894                         tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5895                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5896                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5897 
5898 discard:
5899                         tcp_drop(sk, skb);
5900                         return 0;
5901                 } else {
5902                         tcp_send_ack(sk);
5903                 }
5904                 return -1;
5905         }
5906 
5907         /* No ACK in the segment */
5908 
5909         if (th->rst) {
5910                 /* rfc793:
5911                  * "If the RST bit is set
5912                  *
5913                  *      Otherwise (no ACK) drop the segment and return."
5914                  */
5915 
5916                 goto discard_and_undo;
5917         }
5918 
5919         /* PAWS check. */
5920         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5921             tcp_paws_reject(&tp->rx_opt, 0))
5922                 goto discard_and_undo;
5923 
5924         if (th->syn) {
5925                 /* We see SYN without ACK. It is attempt of
5926                  * simultaneous connect with crossed SYNs.
5927                  * Particularly, it can be connect to self.
5928                  */
5929                 tcp_set_state(sk, TCP_SYN_RECV);
5930 
5931                 if (tp->rx_opt.saw_tstamp) {
5932                         tp->rx_opt.tstamp_ok = 1;
5933                         tcp_store_ts_recent(tp);
5934                         tp->tcp_header_len =
5935                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5936                 } else {
5937                         tp->tcp_header_len = sizeof(struct tcphdr);
5938                 }
5939 
5940                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5941                 tp->copied_seq = tp->rcv_nxt;
5942                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5943 
5944                 /* RFC1323: The window in SYN & SYN/ACK segments is
5945                  * never scaled.
5946                  */
5947                 tp->snd_wnd    = ntohs(th->window);
5948                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5949                 tp->max_window = tp->snd_wnd;
5950