Version:  2.0.40 2.2.26 2.4.37 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5 4.6 4.7

Linux/net/ipv4/tcp_output.c

  1 /*
  2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
  3  *              operating system.  INET is implemented using the  BSD Socket
  4  *              interface as the means of communication with the user level.
  5  *
  6  *              Implementation of the Transmission Control Protocol(TCP).
  7  *
  8  * Authors:     Ross Biro
  9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
 10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
 11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
 12  *              Florian La Roche, <flla@stud.uni-sb.de>
 13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
 14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
 15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
 16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
 17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
 18  *              Jorge Cwik, <jorge@laser.satlink.net>
 19  */
 20 
 21 /*
 22  * Changes:     Pedro Roque     :       Retransmit queue handled by TCP.
 23  *                              :       Fragmentation on mtu decrease
 24  *                              :       Segment collapse on retransmit
 25  *                              :       AF independence
 26  *
 27  *              Linus Torvalds  :       send_delayed_ack
 28  *              David S. Miller :       Charge memory using the right skb
 29  *                                      during syn/ack processing.
 30  *              David S. Miller :       Output engine completely rewritten.
 31  *              Andrea Arcangeli:       SYNACK carry ts_recent in tsecr.
 32  *              Cacophonix Gaul :       draft-minshall-nagle-01
 33  *              J Hadi Salim    :       ECN support
 34  *
 35  */
 36 
 37 #define pr_fmt(fmt) "TCP: " fmt
 38 
 39 #include <net/tcp.h>
 40 
 41 #include <linux/compiler.h>
 42 #include <linux/gfp.h>
 43 #include <linux/module.h>
 44 
 45 /* People can turn this off for buggy TCP's found in printers etc. */
 46 int sysctl_tcp_retrans_collapse __read_mostly = 1;
 47 
 48 /* People can turn this on to work with those rare, broken TCPs that
 49  * interpret the window field as a signed quantity.
 50  */
 51 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
 52 
 53 /* Default TSQ limit of four TSO segments */
 54 int sysctl_tcp_limit_output_bytes __read_mostly = 262144;
 55 
 56 /* This limits the percentage of the congestion window which we
 57  * will allow a single TSO frame to consume.  Building TSO frames
 58  * which are too large can cause TCP streams to be bursty.
 59  */
 60 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
 61 
 62 /* By default, RFC2861 behavior.  */
 63 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
 64 
 65 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
 66                            int push_one, gfp_t gfp);
 67 
 68 /* Account for new data that has been sent to the network. */
 69 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
 70 {
 71         struct inet_connection_sock *icsk = inet_csk(sk);
 72         struct tcp_sock *tp = tcp_sk(sk);
 73         unsigned int prior_packets = tp->packets_out;
 74 
 75         tcp_advance_send_head(sk, skb);
 76         tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
 77 
 78         tp->packets_out += tcp_skb_pcount(skb);
 79         if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
 80             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
 81                 tcp_rearm_rto(sk);
 82         }
 83 
 84         NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
 85                       tcp_skb_pcount(skb));
 86 }
 87 
 88 /* SND.NXT, if window was not shrunk.
 89  * If window has been shrunk, what should we make? It is not clear at all.
 90  * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
 91  * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
 92  * invalid. OK, let's make this for now:
 93  */
 94 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
 95 {
 96         const struct tcp_sock *tp = tcp_sk(sk);
 97 
 98         if (!before(tcp_wnd_end(tp), tp->snd_nxt))
 99                 return tp->snd_nxt;
100         else
101                 return tcp_wnd_end(tp);
102 }
103 
104 /* Calculate mss to advertise in SYN segment.
105  * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
106  *
107  * 1. It is independent of path mtu.
108  * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
109  * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
110  *    attached devices, because some buggy hosts are confused by
111  *    large MSS.
112  * 4. We do not make 3, we advertise MSS, calculated from first
113  *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
114  *    This may be overridden via information stored in routing table.
115  * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
116  *    probably even Jumbo".
117  */
118 static __u16 tcp_advertise_mss(struct sock *sk)
119 {
120         struct tcp_sock *tp = tcp_sk(sk);
121         const struct dst_entry *dst = __sk_dst_get(sk);
122         int mss = tp->advmss;
123 
124         if (dst) {
125                 unsigned int metric = dst_metric_advmss(dst);
126 
127                 if (metric < mss) {
128                         mss = metric;
129                         tp->advmss = mss;
130                 }
131         }
132 
133         return (__u16)mss;
134 }
135 
136 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
137  * This is the first part of cwnd validation mechanism.
138  */
139 void tcp_cwnd_restart(struct sock *sk, s32 delta)
140 {
141         struct tcp_sock *tp = tcp_sk(sk);
142         u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
143         u32 cwnd = tp->snd_cwnd;
144 
145         tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
146 
147         tp->snd_ssthresh = tcp_current_ssthresh(sk);
148         restart_cwnd = min(restart_cwnd, cwnd);
149 
150         while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
151                 cwnd >>= 1;
152         tp->snd_cwnd = max(cwnd, restart_cwnd);
153         tp->snd_cwnd_stamp = tcp_time_stamp;
154         tp->snd_cwnd_used = 0;
155 }
156 
157 /* Congestion state accounting after a packet has been sent. */
158 static void tcp_event_data_sent(struct tcp_sock *tp,
159                                 struct sock *sk)
160 {
161         struct inet_connection_sock *icsk = inet_csk(sk);
162         const u32 now = tcp_time_stamp;
163 
164         if (tcp_packets_in_flight(tp) == 0)
165                 tcp_ca_event(sk, CA_EVENT_TX_START);
166 
167         tp->lsndtime = now;
168 
169         /* If it is a reply for ato after last received
170          * packet, enter pingpong mode.
171          */
172         if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
173                 icsk->icsk_ack.pingpong = 1;
174 }
175 
176 /* Account for an ACK we sent. */
177 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
178 {
179         tcp_dec_quickack_mode(sk, pkts);
180         inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
181 }
182 
183 
184 u32 tcp_default_init_rwnd(u32 mss)
185 {
186         /* Initial receive window should be twice of TCP_INIT_CWND to
187          * enable proper sending of new unsent data during fast recovery
188          * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
189          * limit when mss is larger than 1460.
190          */
191         u32 init_rwnd = TCP_INIT_CWND * 2;
192 
193         if (mss > 1460)
194                 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
195         return init_rwnd;
196 }
197 
198 /* Determine a window scaling and initial window to offer.
199  * Based on the assumption that the given amount of space
200  * will be offered. Store the results in the tp structure.
201  * NOTE: for smooth operation initial space offering should
202  * be a multiple of mss if possible. We assume here that mss >= 1.
203  * This MUST be enforced by all callers.
204  */
205 void tcp_select_initial_window(int __space, __u32 mss,
206                                __u32 *rcv_wnd, __u32 *window_clamp,
207                                int wscale_ok, __u8 *rcv_wscale,
208                                __u32 init_rcv_wnd)
209 {
210         unsigned int space = (__space < 0 ? 0 : __space);
211 
212         /* If no clamp set the clamp to the max possible scaled window */
213         if (*window_clamp == 0)
214                 (*window_clamp) = (65535 << 14);
215         space = min(*window_clamp, space);
216 
217         /* Quantize space offering to a multiple of mss if possible. */
218         if (space > mss)
219                 space = (space / mss) * mss;
220 
221         /* NOTE: offering an initial window larger than 32767
222          * will break some buggy TCP stacks. If the admin tells us
223          * it is likely we could be speaking with such a buggy stack
224          * we will truncate our initial window offering to 32K-1
225          * unless the remote has sent us a window scaling option,
226          * which we interpret as a sign the remote TCP is not
227          * misinterpreting the window field as a signed quantity.
228          */
229         if (sysctl_tcp_workaround_signed_windows)
230                 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
231         else
232                 (*rcv_wnd) = space;
233 
234         (*rcv_wscale) = 0;
235         if (wscale_ok) {
236                 /* Set window scaling on max possible window
237                  * See RFC1323 for an explanation of the limit to 14
238                  */
239                 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
240                 space = min_t(u32, space, *window_clamp);
241                 while (space > 65535 && (*rcv_wscale) < 14) {
242                         space >>= 1;
243                         (*rcv_wscale)++;
244                 }
245         }
246 
247         if (mss > (1 << *rcv_wscale)) {
248                 if (!init_rcv_wnd) /* Use default unless specified otherwise */
249                         init_rcv_wnd = tcp_default_init_rwnd(mss);
250                 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
251         }
252 
253         /* Set the clamp no higher than max representable value */
254         (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
255 }
256 EXPORT_SYMBOL(tcp_select_initial_window);
257 
258 /* Chose a new window to advertise, update state in tcp_sock for the
259  * socket, and return result with RFC1323 scaling applied.  The return
260  * value can be stuffed directly into th->window for an outgoing
261  * frame.
262  */
263 static u16 tcp_select_window(struct sock *sk)
264 {
265         struct tcp_sock *tp = tcp_sk(sk);
266         u32 old_win = tp->rcv_wnd;
267         u32 cur_win = tcp_receive_window(tp);
268         u32 new_win = __tcp_select_window(sk);
269 
270         /* Never shrink the offered window */
271         if (new_win < cur_win) {
272                 /* Danger Will Robinson!
273                  * Don't update rcv_wup/rcv_wnd here or else
274                  * we will not be able to advertise a zero
275                  * window in time.  --DaveM
276                  *
277                  * Relax Will Robinson.
278                  */
279                 if (new_win == 0)
280                         NET_INC_STATS(sock_net(sk),
281                                       LINUX_MIB_TCPWANTZEROWINDOWADV);
282                 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
283         }
284         tp->rcv_wnd = new_win;
285         tp->rcv_wup = tp->rcv_nxt;
286 
287         /* Make sure we do not exceed the maximum possible
288          * scaled window.
289          */
290         if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
291                 new_win = min(new_win, MAX_TCP_WINDOW);
292         else
293                 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
294 
295         /* RFC1323 scaling applied */
296         new_win >>= tp->rx_opt.rcv_wscale;
297 
298         /* If we advertise zero window, disable fast path. */
299         if (new_win == 0) {
300                 tp->pred_flags = 0;
301                 if (old_win)
302                         NET_INC_STATS(sock_net(sk),
303                                       LINUX_MIB_TCPTOZEROWINDOWADV);
304         } else if (old_win == 0) {
305                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
306         }
307 
308         return new_win;
309 }
310 
311 /* Packet ECN state for a SYN-ACK */
312 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
313 {
314         const struct tcp_sock *tp = tcp_sk(sk);
315 
316         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
317         if (!(tp->ecn_flags & TCP_ECN_OK))
318                 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
319         else if (tcp_ca_needs_ecn(sk))
320                 INET_ECN_xmit(sk);
321 }
322 
323 /* Packet ECN state for a SYN.  */
324 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
325 {
326         struct tcp_sock *tp = tcp_sk(sk);
327         bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
328                        tcp_ca_needs_ecn(sk);
329 
330         if (!use_ecn) {
331                 const struct dst_entry *dst = __sk_dst_get(sk);
332 
333                 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
334                         use_ecn = true;
335         }
336 
337         tp->ecn_flags = 0;
338 
339         if (use_ecn) {
340                 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
341                 tp->ecn_flags = TCP_ECN_OK;
342                 if (tcp_ca_needs_ecn(sk))
343                         INET_ECN_xmit(sk);
344         }
345 }
346 
347 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
348 {
349         if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
350                 /* tp->ecn_flags are cleared at a later point in time when
351                  * SYN ACK is ultimatively being received.
352                  */
353                 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
354 }
355 
356 static void
357 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
358 {
359         if (inet_rsk(req)->ecn_ok)
360                 th->ece = 1;
361 }
362 
363 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
364  * be sent.
365  */
366 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
367                          struct tcphdr *th, int tcp_header_len)
368 {
369         struct tcp_sock *tp = tcp_sk(sk);
370 
371         if (tp->ecn_flags & TCP_ECN_OK) {
372                 /* Not-retransmitted data segment: set ECT and inject CWR. */
373                 if (skb->len != tcp_header_len &&
374                     !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
375                         INET_ECN_xmit(sk);
376                         if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
377                                 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
378                                 th->cwr = 1;
379                                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
380                         }
381                 } else if (!tcp_ca_needs_ecn(sk)) {
382                         /* ACK or retransmitted segment: clear ECT|CE */
383                         INET_ECN_dontxmit(sk);
384                 }
385                 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
386                         th->ece = 1;
387         }
388 }
389 
390 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
391  * auto increment end seqno.
392  */
393 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
394 {
395         skb->ip_summed = CHECKSUM_PARTIAL;
396         skb->csum = 0;
397 
398         TCP_SKB_CB(skb)->tcp_flags = flags;
399         TCP_SKB_CB(skb)->sacked = 0;
400 
401         tcp_skb_pcount_set(skb, 1);
402 
403         TCP_SKB_CB(skb)->seq = seq;
404         if (flags & (TCPHDR_SYN | TCPHDR_FIN))
405                 seq++;
406         TCP_SKB_CB(skb)->end_seq = seq;
407 }
408 
409 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
410 {
411         return tp->snd_una != tp->snd_up;
412 }
413 
414 #define OPTION_SACK_ADVERTISE   (1 << 0)
415 #define OPTION_TS               (1 << 1)
416 #define OPTION_MD5              (1 << 2)
417 #define OPTION_WSCALE           (1 << 3)
418 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
419 
420 struct tcp_out_options {
421         u16 options;            /* bit field of OPTION_* */
422         u16 mss;                /* 0 to disable */
423         u8 ws;                  /* window scale, 0 to disable */
424         u8 num_sack_blocks;     /* number of SACK blocks to include */
425         u8 hash_size;           /* bytes in hash_location */
426         __u8 *hash_location;    /* temporary pointer, overloaded */
427         __u32 tsval, tsecr;     /* need to include OPTION_TS */
428         struct tcp_fastopen_cookie *fastopen_cookie;    /* Fast open cookie */
429 };
430 
431 /* Write previously computed TCP options to the packet.
432  *
433  * Beware: Something in the Internet is very sensitive to the ordering of
434  * TCP options, we learned this through the hard way, so be careful here.
435  * Luckily we can at least blame others for their non-compliance but from
436  * inter-operability perspective it seems that we're somewhat stuck with
437  * the ordering which we have been using if we want to keep working with
438  * those broken things (not that it currently hurts anybody as there isn't
439  * particular reason why the ordering would need to be changed).
440  *
441  * At least SACK_PERM as the first option is known to lead to a disaster
442  * (but it may well be that other scenarios fail similarly).
443  */
444 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
445                               struct tcp_out_options *opts)
446 {
447         u16 options = opts->options;    /* mungable copy */
448 
449         if (unlikely(OPTION_MD5 & options)) {
450                 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
451                                (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
452                 /* overload cookie hash location */
453                 opts->hash_location = (__u8 *)ptr;
454                 ptr += 4;
455         }
456 
457         if (unlikely(opts->mss)) {
458                 *ptr++ = htonl((TCPOPT_MSS << 24) |
459                                (TCPOLEN_MSS << 16) |
460                                opts->mss);
461         }
462 
463         if (likely(OPTION_TS & options)) {
464                 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
465                         *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
466                                        (TCPOLEN_SACK_PERM << 16) |
467                                        (TCPOPT_TIMESTAMP << 8) |
468                                        TCPOLEN_TIMESTAMP);
469                         options &= ~OPTION_SACK_ADVERTISE;
470                 } else {
471                         *ptr++ = htonl((TCPOPT_NOP << 24) |
472                                        (TCPOPT_NOP << 16) |
473                                        (TCPOPT_TIMESTAMP << 8) |
474                                        TCPOLEN_TIMESTAMP);
475                 }
476                 *ptr++ = htonl(opts->tsval);
477                 *ptr++ = htonl(opts->tsecr);
478         }
479 
480         if (unlikely(OPTION_SACK_ADVERTISE & options)) {
481                 *ptr++ = htonl((TCPOPT_NOP << 24) |
482                                (TCPOPT_NOP << 16) |
483                                (TCPOPT_SACK_PERM << 8) |
484                                TCPOLEN_SACK_PERM);
485         }
486 
487         if (unlikely(OPTION_WSCALE & options)) {
488                 *ptr++ = htonl((TCPOPT_NOP << 24) |
489                                (TCPOPT_WINDOW << 16) |
490                                (TCPOLEN_WINDOW << 8) |
491                                opts->ws);
492         }
493 
494         if (unlikely(opts->num_sack_blocks)) {
495                 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
496                         tp->duplicate_sack : tp->selective_acks;
497                 int this_sack;
498 
499                 *ptr++ = htonl((TCPOPT_NOP  << 24) |
500                                (TCPOPT_NOP  << 16) |
501                                (TCPOPT_SACK <<  8) |
502                                (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
503                                                      TCPOLEN_SACK_PERBLOCK)));
504 
505                 for (this_sack = 0; this_sack < opts->num_sack_blocks;
506                      ++this_sack) {
507                         *ptr++ = htonl(sp[this_sack].start_seq);
508                         *ptr++ = htonl(sp[this_sack].end_seq);
509                 }
510 
511                 tp->rx_opt.dsack = 0;
512         }
513 
514         if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
515                 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
516                 u8 *p = (u8 *)ptr;
517                 u32 len; /* Fast Open option length */
518 
519                 if (foc->exp) {
520                         len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
521                         *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
522                                      TCPOPT_FASTOPEN_MAGIC);
523                         p += TCPOLEN_EXP_FASTOPEN_BASE;
524                 } else {
525                         len = TCPOLEN_FASTOPEN_BASE + foc->len;
526                         *p++ = TCPOPT_FASTOPEN;
527                         *p++ = len;
528                 }
529 
530                 memcpy(p, foc->val, foc->len);
531                 if ((len & 3) == 2) {
532                         p[foc->len] = TCPOPT_NOP;
533                         p[foc->len + 1] = TCPOPT_NOP;
534                 }
535                 ptr += (len + 3) >> 2;
536         }
537 }
538 
539 /* Compute TCP options for SYN packets. This is not the final
540  * network wire format yet.
541  */
542 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
543                                 struct tcp_out_options *opts,
544                                 struct tcp_md5sig_key **md5)
545 {
546         struct tcp_sock *tp = tcp_sk(sk);
547         unsigned int remaining = MAX_TCP_OPTION_SPACE;
548         struct tcp_fastopen_request *fastopen = tp->fastopen_req;
549 
550 #ifdef CONFIG_TCP_MD5SIG
551         *md5 = tp->af_specific->md5_lookup(sk, sk);
552         if (*md5) {
553                 opts->options |= OPTION_MD5;
554                 remaining -= TCPOLEN_MD5SIG_ALIGNED;
555         }
556 #else
557         *md5 = NULL;
558 #endif
559 
560         /* We always get an MSS option.  The option bytes which will be seen in
561          * normal data packets should timestamps be used, must be in the MSS
562          * advertised.  But we subtract them from tp->mss_cache so that
563          * calculations in tcp_sendmsg are simpler etc.  So account for this
564          * fact here if necessary.  If we don't do this correctly, as a
565          * receiver we won't recognize data packets as being full sized when we
566          * should, and thus we won't abide by the delayed ACK rules correctly.
567          * SACKs don't matter, we never delay an ACK when we have any of those
568          * going out.  */
569         opts->mss = tcp_advertise_mss(sk);
570         remaining -= TCPOLEN_MSS_ALIGNED;
571 
572         if (likely(sysctl_tcp_timestamps && !*md5)) {
573                 opts->options |= OPTION_TS;
574                 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
575                 opts->tsecr = tp->rx_opt.ts_recent;
576                 remaining -= TCPOLEN_TSTAMP_ALIGNED;
577         }
578         if (likely(sysctl_tcp_window_scaling)) {
579                 opts->ws = tp->rx_opt.rcv_wscale;
580                 opts->options |= OPTION_WSCALE;
581                 remaining -= TCPOLEN_WSCALE_ALIGNED;
582         }
583         if (likely(sysctl_tcp_sack)) {
584                 opts->options |= OPTION_SACK_ADVERTISE;
585                 if (unlikely(!(OPTION_TS & opts->options)))
586                         remaining -= TCPOLEN_SACKPERM_ALIGNED;
587         }
588 
589         if (fastopen && fastopen->cookie.len >= 0) {
590                 u32 need = fastopen->cookie.len;
591 
592                 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
593                                                TCPOLEN_FASTOPEN_BASE;
594                 need = (need + 3) & ~3U;  /* Align to 32 bits */
595                 if (remaining >= need) {
596                         opts->options |= OPTION_FAST_OPEN_COOKIE;
597                         opts->fastopen_cookie = &fastopen->cookie;
598                         remaining -= need;
599                         tp->syn_fastopen = 1;
600                         tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
601                 }
602         }
603 
604         return MAX_TCP_OPTION_SPACE - remaining;
605 }
606 
607 /* Set up TCP options for SYN-ACKs. */
608 static unsigned int tcp_synack_options(struct request_sock *req,
609                                        unsigned int mss, struct sk_buff *skb,
610                                        struct tcp_out_options *opts,
611                                        const struct tcp_md5sig_key *md5,
612                                        struct tcp_fastopen_cookie *foc)
613 {
614         struct inet_request_sock *ireq = inet_rsk(req);
615         unsigned int remaining = MAX_TCP_OPTION_SPACE;
616 
617 #ifdef CONFIG_TCP_MD5SIG
618         if (md5) {
619                 opts->options |= OPTION_MD5;
620                 remaining -= TCPOLEN_MD5SIG_ALIGNED;
621 
622                 /* We can't fit any SACK blocks in a packet with MD5 + TS
623                  * options. There was discussion about disabling SACK
624                  * rather than TS in order to fit in better with old,
625                  * buggy kernels, but that was deemed to be unnecessary.
626                  */
627                 ireq->tstamp_ok &= !ireq->sack_ok;
628         }
629 #endif
630 
631         /* We always send an MSS option. */
632         opts->mss = mss;
633         remaining -= TCPOLEN_MSS_ALIGNED;
634 
635         if (likely(ireq->wscale_ok)) {
636                 opts->ws = ireq->rcv_wscale;
637                 opts->options |= OPTION_WSCALE;
638                 remaining -= TCPOLEN_WSCALE_ALIGNED;
639         }
640         if (likely(ireq->tstamp_ok)) {
641                 opts->options |= OPTION_TS;
642                 opts->tsval = tcp_skb_timestamp(skb);
643                 opts->tsecr = req->ts_recent;
644                 remaining -= TCPOLEN_TSTAMP_ALIGNED;
645         }
646         if (likely(ireq->sack_ok)) {
647                 opts->options |= OPTION_SACK_ADVERTISE;
648                 if (unlikely(!ireq->tstamp_ok))
649                         remaining -= TCPOLEN_SACKPERM_ALIGNED;
650         }
651         if (foc != NULL && foc->len >= 0) {
652                 u32 need = foc->len;
653 
654                 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
655                                    TCPOLEN_FASTOPEN_BASE;
656                 need = (need + 3) & ~3U;  /* Align to 32 bits */
657                 if (remaining >= need) {
658                         opts->options |= OPTION_FAST_OPEN_COOKIE;
659                         opts->fastopen_cookie = foc;
660                         remaining -= need;
661                 }
662         }
663 
664         return MAX_TCP_OPTION_SPACE - remaining;
665 }
666 
667 /* Compute TCP options for ESTABLISHED sockets. This is not the
668  * final wire format yet.
669  */
670 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
671                                         struct tcp_out_options *opts,
672                                         struct tcp_md5sig_key **md5)
673 {
674         struct tcp_sock *tp = tcp_sk(sk);
675         unsigned int size = 0;
676         unsigned int eff_sacks;
677 
678         opts->options = 0;
679 
680 #ifdef CONFIG_TCP_MD5SIG
681         *md5 = tp->af_specific->md5_lookup(sk, sk);
682         if (unlikely(*md5)) {
683                 opts->options |= OPTION_MD5;
684                 size += TCPOLEN_MD5SIG_ALIGNED;
685         }
686 #else
687         *md5 = NULL;
688 #endif
689 
690         if (likely(tp->rx_opt.tstamp_ok)) {
691                 opts->options |= OPTION_TS;
692                 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
693                 opts->tsecr = tp->rx_opt.ts_recent;
694                 size += TCPOLEN_TSTAMP_ALIGNED;
695         }
696 
697         eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
698         if (unlikely(eff_sacks)) {
699                 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
700                 opts->num_sack_blocks =
701                         min_t(unsigned int, eff_sacks,
702                               (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
703                               TCPOLEN_SACK_PERBLOCK);
704                 size += TCPOLEN_SACK_BASE_ALIGNED +
705                         opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
706         }
707 
708         return size;
709 }
710 
711 
712 /* TCP SMALL QUEUES (TSQ)
713  *
714  * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
715  * to reduce RTT and bufferbloat.
716  * We do this using a special skb destructor (tcp_wfree).
717  *
718  * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
719  * needs to be reallocated in a driver.
720  * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
721  *
722  * Since transmit from skb destructor is forbidden, we use a tasklet
723  * to process all sockets that eventually need to send more skbs.
724  * We use one tasklet per cpu, with its own queue of sockets.
725  */
726 struct tsq_tasklet {
727         struct tasklet_struct   tasklet;
728         struct list_head        head; /* queue of tcp sockets */
729 };
730 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
731 
732 static void tcp_tsq_handler(struct sock *sk)
733 {
734         if ((1 << sk->sk_state) &
735             (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
736              TCPF_CLOSE_WAIT  | TCPF_LAST_ACK))
737                 tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
738                                0, GFP_ATOMIC);
739 }
740 /*
741  * One tasklet per cpu tries to send more skbs.
742  * We run in tasklet context but need to disable irqs when
743  * transferring tsq->head because tcp_wfree() might
744  * interrupt us (non NAPI drivers)
745  */
746 static void tcp_tasklet_func(unsigned long data)
747 {
748         struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
749         LIST_HEAD(list);
750         unsigned long flags;
751         struct list_head *q, *n;
752         struct tcp_sock *tp;
753         struct sock *sk;
754 
755         local_irq_save(flags);
756         list_splice_init(&tsq->head, &list);
757         local_irq_restore(flags);
758 
759         list_for_each_safe(q, n, &list) {
760                 tp = list_entry(q, struct tcp_sock, tsq_node);
761                 list_del(&tp->tsq_node);
762 
763                 sk = (struct sock *)tp;
764                 bh_lock_sock(sk);
765 
766                 if (!sock_owned_by_user(sk)) {
767                         tcp_tsq_handler(sk);
768                 } else {
769                         /* defer the work to tcp_release_cb() */
770                         set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
771                 }
772                 bh_unlock_sock(sk);
773 
774                 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
775                 sk_free(sk);
776         }
777 }
778 
779 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) |           \
780                           (1UL << TCP_WRITE_TIMER_DEFERRED) |   \
781                           (1UL << TCP_DELACK_TIMER_DEFERRED) |  \
782                           (1UL << TCP_MTU_REDUCED_DEFERRED))
783 /**
784  * tcp_release_cb - tcp release_sock() callback
785  * @sk: socket
786  *
787  * called from release_sock() to perform protocol dependent
788  * actions before socket release.
789  */
790 void tcp_release_cb(struct sock *sk)
791 {
792         struct tcp_sock *tp = tcp_sk(sk);
793         unsigned long flags, nflags;
794 
795         /* perform an atomic operation only if at least one flag is set */
796         do {
797                 flags = tp->tsq_flags;
798                 if (!(flags & TCP_DEFERRED_ALL))
799                         return;
800                 nflags = flags & ~TCP_DEFERRED_ALL;
801         } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
802 
803         if (flags & (1UL << TCP_TSQ_DEFERRED))
804                 tcp_tsq_handler(sk);
805 
806         /* Here begins the tricky part :
807          * We are called from release_sock() with :
808          * 1) BH disabled
809          * 2) sk_lock.slock spinlock held
810          * 3) socket owned by us (sk->sk_lock.owned == 1)
811          *
812          * But following code is meant to be called from BH handlers,
813          * so we should keep BH disabled, but early release socket ownership
814          */
815         sock_release_ownership(sk);
816 
817         if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
818                 tcp_write_timer_handler(sk);
819                 __sock_put(sk);
820         }
821         if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
822                 tcp_delack_timer_handler(sk);
823                 __sock_put(sk);
824         }
825         if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
826                 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
827                 __sock_put(sk);
828         }
829 }
830 EXPORT_SYMBOL(tcp_release_cb);
831 
832 void __init tcp_tasklet_init(void)
833 {
834         int i;
835 
836         for_each_possible_cpu(i) {
837                 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
838 
839                 INIT_LIST_HEAD(&tsq->head);
840                 tasklet_init(&tsq->tasklet,
841                              tcp_tasklet_func,
842                              (unsigned long)tsq);
843         }
844 }
845 
846 /*
847  * Write buffer destructor automatically called from kfree_skb.
848  * We can't xmit new skbs from this context, as we might already
849  * hold qdisc lock.
850  */
851 void tcp_wfree(struct sk_buff *skb)
852 {
853         struct sock *sk = skb->sk;
854         struct tcp_sock *tp = tcp_sk(sk);
855         int wmem;
856 
857         /* Keep one reference on sk_wmem_alloc.
858          * Will be released by sk_free() from here or tcp_tasklet_func()
859          */
860         wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
861 
862         /* If this softirq is serviced by ksoftirqd, we are likely under stress.
863          * Wait until our queues (qdisc + devices) are drained.
864          * This gives :
865          * - less callbacks to tcp_write_xmit(), reducing stress (batches)
866          * - chance for incoming ACK (processed by another cpu maybe)
867          *   to migrate this flow (skb->ooo_okay will be eventually set)
868          */
869         if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
870                 goto out;
871 
872         if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
873             !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
874                 unsigned long flags;
875                 struct tsq_tasklet *tsq;
876 
877                 /* queue this socket to tasklet queue */
878                 local_irq_save(flags);
879                 tsq = this_cpu_ptr(&tsq_tasklet);
880                 list_add(&tp->tsq_node, &tsq->head);
881                 tasklet_schedule(&tsq->tasklet);
882                 local_irq_restore(flags);
883                 return;
884         }
885 out:
886         sk_free(sk);
887 }
888 
889 /* This routine actually transmits TCP packets queued in by
890  * tcp_do_sendmsg().  This is used by both the initial
891  * transmission and possible later retransmissions.
892  * All SKB's seen here are completely headerless.  It is our
893  * job to build the TCP header, and pass the packet down to
894  * IP so it can do the same plus pass the packet off to the
895  * device.
896  *
897  * We are working here with either a clone of the original
898  * SKB, or a fresh unique copy made by the retransmit engine.
899  */
900 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
901                             gfp_t gfp_mask)
902 {
903         const struct inet_connection_sock *icsk = inet_csk(sk);
904         struct inet_sock *inet;
905         struct tcp_sock *tp;
906         struct tcp_skb_cb *tcb;
907         struct tcp_out_options opts;
908         unsigned int tcp_options_size, tcp_header_size;
909         struct tcp_md5sig_key *md5;
910         struct tcphdr *th;
911         int err;
912 
913         BUG_ON(!skb || !tcp_skb_pcount(skb));
914 
915         if (clone_it) {
916                 skb_mstamp_get(&skb->skb_mstamp);
917 
918                 if (unlikely(skb_cloned(skb)))
919                         skb = pskb_copy(skb, gfp_mask);
920                 else
921                         skb = skb_clone(skb, gfp_mask);
922                 if (unlikely(!skb))
923                         return -ENOBUFS;
924         }
925 
926         inet = inet_sk(sk);
927         tp = tcp_sk(sk);
928         tcb = TCP_SKB_CB(skb);
929         memset(&opts, 0, sizeof(opts));
930 
931         if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
932                 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
933         else
934                 tcp_options_size = tcp_established_options(sk, skb, &opts,
935                                                            &md5);
936         tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
937 
938         /* if no packet is in qdisc/device queue, then allow XPS to select
939          * another queue. We can be called from tcp_tsq_handler()
940          * which holds one reference to sk_wmem_alloc.
941          *
942          * TODO: Ideally, in-flight pure ACK packets should not matter here.
943          * One way to get this would be to set skb->truesize = 2 on them.
944          */
945         skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
946 
947         skb_push(skb, tcp_header_size);
948         skb_reset_transport_header(skb);
949 
950         skb_orphan(skb);
951         skb->sk = sk;
952         skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
953         skb_set_hash_from_sk(skb, sk);
954         atomic_add(skb->truesize, &sk->sk_wmem_alloc);
955 
956         /* Build TCP header and checksum it. */
957         th = (struct tcphdr *)skb->data;
958         th->source              = inet->inet_sport;
959         th->dest                = inet->inet_dport;
960         th->seq                 = htonl(tcb->seq);
961         th->ack_seq             = htonl(tp->rcv_nxt);
962         *(((__be16 *)th) + 6)   = htons(((tcp_header_size >> 2) << 12) |
963                                         tcb->tcp_flags);
964 
965         th->check               = 0;
966         th->urg_ptr             = 0;
967 
968         /* The urg_mode check is necessary during a below snd_una win probe */
969         if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
970                 if (before(tp->snd_up, tcb->seq + 0x10000)) {
971                         th->urg_ptr = htons(tp->snd_up - tcb->seq);
972                         th->urg = 1;
973                 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
974                         th->urg_ptr = htons(0xFFFF);
975                         th->urg = 1;
976                 }
977         }
978 
979         tcp_options_write((__be32 *)(th + 1), tp, &opts);
980         skb_shinfo(skb)->gso_type = sk->sk_gso_type;
981         if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
982                 th->window      = htons(tcp_select_window(sk));
983                 tcp_ecn_send(sk, skb, th, tcp_header_size);
984         } else {
985                 /* RFC1323: The window in SYN & SYN/ACK segments
986                  * is never scaled.
987                  */
988                 th->window      = htons(min(tp->rcv_wnd, 65535U));
989         }
990 #ifdef CONFIG_TCP_MD5SIG
991         /* Calculate the MD5 hash, as we have all we need now */
992         if (md5) {
993                 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
994                 tp->af_specific->calc_md5_hash(opts.hash_location,
995                                                md5, sk, skb);
996         }
997 #endif
998 
999         icsk->icsk_af_ops->send_check(sk, skb);
1000 
1001         if (likely(tcb->tcp_flags & TCPHDR_ACK))
1002                 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1003 
1004         if (skb->len != tcp_header_size) {
1005                 tcp_event_data_sent(tp, sk);
1006                 tp->data_segs_out += tcp_skb_pcount(skb);
1007         }
1008 
1009         if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1010                 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1011                               tcp_skb_pcount(skb));
1012 
1013         tp->segs_out += tcp_skb_pcount(skb);
1014         /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1015         skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1016         skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1017 
1018         /* Our usage of tstamp should remain private */
1019         skb->tstamp.tv64 = 0;
1020 
1021         /* Cleanup our debris for IP stacks */
1022         memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1023                                sizeof(struct inet6_skb_parm)));
1024 
1025         err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1026 
1027         if (likely(err <= 0))
1028                 return err;
1029 
1030         tcp_enter_cwr(sk);
1031 
1032         return net_xmit_eval(err);
1033 }
1034 
1035 /* This routine just queues the buffer for sending.
1036  *
1037  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1038  * otherwise socket can stall.
1039  */
1040 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1041 {
1042         struct tcp_sock *tp = tcp_sk(sk);
1043 
1044         /* Advance write_seq and place onto the write_queue. */
1045         tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1046         __skb_header_release(skb);
1047         tcp_add_write_queue_tail(sk, skb);
1048         sk->sk_wmem_queued += skb->truesize;
1049         sk_mem_charge(sk, skb->truesize);
1050 }
1051 
1052 /* Initialize TSO segments for a packet. */
1053 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1054 {
1055         if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1056                 /* Avoid the costly divide in the normal
1057                  * non-TSO case.
1058                  */
1059                 tcp_skb_pcount_set(skb, 1);
1060                 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1061         } else {
1062                 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1063                 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1064         }
1065 }
1066 
1067 /* When a modification to fackets out becomes necessary, we need to check
1068  * skb is counted to fackets_out or not.
1069  */
1070 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1071                                    int decr)
1072 {
1073         struct tcp_sock *tp = tcp_sk(sk);
1074 
1075         if (!tp->sacked_out || tcp_is_reno(tp))
1076                 return;
1077 
1078         if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1079                 tp->fackets_out -= decr;
1080 }
1081 
1082 /* Pcount in the middle of the write queue got changed, we need to do various
1083  * tweaks to fix counters
1084  */
1085 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1086 {
1087         struct tcp_sock *tp = tcp_sk(sk);
1088 
1089         tp->packets_out -= decr;
1090 
1091         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1092                 tp->sacked_out -= decr;
1093         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1094                 tp->retrans_out -= decr;
1095         if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1096                 tp->lost_out -= decr;
1097 
1098         /* Reno case is special. Sigh... */
1099         if (tcp_is_reno(tp) && decr > 0)
1100                 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1101 
1102         tcp_adjust_fackets_out(sk, skb, decr);
1103 
1104         if (tp->lost_skb_hint &&
1105             before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1106             (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1107                 tp->lost_cnt_hint -= decr;
1108 
1109         tcp_verify_left_out(tp);
1110 }
1111 
1112 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1113 {
1114         return TCP_SKB_CB(skb)->txstamp_ack ||
1115                 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1116 }
1117 
1118 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1119 {
1120         struct skb_shared_info *shinfo = skb_shinfo(skb);
1121 
1122         if (unlikely(tcp_has_tx_tstamp(skb)) &&
1123             !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1124                 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1125                 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1126 
1127                 shinfo->tx_flags &= ~tsflags;
1128                 shinfo2->tx_flags |= tsflags;
1129                 swap(shinfo->tskey, shinfo2->tskey);
1130                 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1131                 TCP_SKB_CB(skb)->txstamp_ack = 0;
1132         }
1133 }
1134 
1135 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1136 {
1137         TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1138         TCP_SKB_CB(skb)->eor = 0;
1139 }
1140 
1141 /* Function to create two new TCP segments.  Shrinks the given segment
1142  * to the specified size and appends a new segment with the rest of the
1143  * packet to the list.  This won't be called frequently, I hope.
1144  * Remember, these are still headerless SKBs at this point.
1145  */
1146 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1147                  unsigned int mss_now, gfp_t gfp)
1148 {
1149         struct tcp_sock *tp = tcp_sk(sk);
1150         struct sk_buff *buff;
1151         int nsize, old_factor;
1152         int nlen;
1153         u8 flags;
1154 
1155         if (WARN_ON(len > skb->len))
1156                 return -EINVAL;
1157 
1158         nsize = skb_headlen(skb) - len;
1159         if (nsize < 0)
1160                 nsize = 0;
1161 
1162         if (skb_unclone(skb, gfp))
1163                 return -ENOMEM;
1164 
1165         /* Get a new skb... force flag on. */
1166         buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1167         if (!buff)
1168                 return -ENOMEM; /* We'll just try again later. */
1169 
1170         sk->sk_wmem_queued += buff->truesize;
1171         sk_mem_charge(sk, buff->truesize);
1172         nlen = skb->len - len - nsize;
1173         buff->truesize += nlen;
1174         skb->truesize -= nlen;
1175 
1176         /* Correct the sequence numbers. */
1177         TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1178         TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1179         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1180 
1181         /* PSH and FIN should only be set in the second packet. */
1182         flags = TCP_SKB_CB(skb)->tcp_flags;
1183         TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1184         TCP_SKB_CB(buff)->tcp_flags = flags;
1185         TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1186         tcp_skb_fragment_eor(skb, buff);
1187 
1188         if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1189                 /* Copy and checksum data tail into the new buffer. */
1190                 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1191                                                        skb_put(buff, nsize),
1192                                                        nsize, 0);
1193 
1194                 skb_trim(skb, len);
1195 
1196                 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1197         } else {
1198                 skb->ip_summed = CHECKSUM_PARTIAL;
1199                 skb_split(skb, buff, len);
1200         }
1201 
1202         buff->ip_summed = skb->ip_summed;
1203 
1204         buff->tstamp = skb->tstamp;
1205         tcp_fragment_tstamp(skb, buff);
1206 
1207         old_factor = tcp_skb_pcount(skb);
1208 
1209         /* Fix up tso_factor for both original and new SKB.  */
1210         tcp_set_skb_tso_segs(skb, mss_now);
1211         tcp_set_skb_tso_segs(buff, mss_now);
1212 
1213         /* If this packet has been sent out already, we must
1214          * adjust the various packet counters.
1215          */
1216         if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1217                 int diff = old_factor - tcp_skb_pcount(skb) -
1218                         tcp_skb_pcount(buff);
1219 
1220                 if (diff)
1221                         tcp_adjust_pcount(sk, skb, diff);
1222         }
1223 
1224         /* Link BUFF into the send queue. */
1225         __skb_header_release(buff);
1226         tcp_insert_write_queue_after(skb, buff, sk);
1227 
1228         return 0;
1229 }
1230 
1231 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1232  * eventually). The difference is that pulled data not copied, but
1233  * immediately discarded.
1234  */
1235 static void __pskb_trim_head(struct sk_buff *skb, int len)
1236 {
1237         struct skb_shared_info *shinfo;
1238         int i, k, eat;
1239 
1240         eat = min_t(int, len, skb_headlen(skb));
1241         if (eat) {
1242                 __skb_pull(skb, eat);
1243                 len -= eat;
1244                 if (!len)
1245                         return;
1246         }
1247         eat = len;
1248         k = 0;
1249         shinfo = skb_shinfo(skb);
1250         for (i = 0; i < shinfo->nr_frags; i++) {
1251                 int size = skb_frag_size(&shinfo->frags[i]);
1252 
1253                 if (size <= eat) {
1254                         skb_frag_unref(skb, i);
1255                         eat -= size;
1256                 } else {
1257                         shinfo->frags[k] = shinfo->frags[i];
1258                         if (eat) {
1259                                 shinfo->frags[k].page_offset += eat;
1260                                 skb_frag_size_sub(&shinfo->frags[k], eat);
1261                                 eat = 0;
1262                         }
1263                         k++;
1264                 }
1265         }
1266         shinfo->nr_frags = k;
1267 
1268         skb_reset_tail_pointer(skb);
1269         skb->data_len -= len;
1270         skb->len = skb->data_len;
1271 }
1272 
1273 /* Remove acked data from a packet in the transmit queue. */
1274 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1275 {
1276         if (skb_unclone(skb, GFP_ATOMIC))
1277                 return -ENOMEM;
1278 
1279         __pskb_trim_head(skb, len);
1280 
1281         TCP_SKB_CB(skb)->seq += len;
1282         skb->ip_summed = CHECKSUM_PARTIAL;
1283 
1284         skb->truesize        -= len;
1285         sk->sk_wmem_queued   -= len;
1286         sk_mem_uncharge(sk, len);
1287         sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1288 
1289         /* Any change of skb->len requires recalculation of tso factor. */
1290         if (tcp_skb_pcount(skb) > 1)
1291                 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1292 
1293         return 0;
1294 }
1295 
1296 /* Calculate MSS not accounting any TCP options.  */
1297 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1298 {
1299         const struct tcp_sock *tp = tcp_sk(sk);
1300         const struct inet_connection_sock *icsk = inet_csk(sk);
1301         int mss_now;
1302 
1303         /* Calculate base mss without TCP options:
1304            It is MMS_S - sizeof(tcphdr) of rfc1122
1305          */
1306         mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1307 
1308         /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1309         if (icsk->icsk_af_ops->net_frag_header_len) {
1310                 const struct dst_entry *dst = __sk_dst_get(sk);
1311 
1312                 if (dst && dst_allfrag(dst))
1313                         mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1314         }
1315 
1316         /* Clamp it (mss_clamp does not include tcp options) */
1317         if (mss_now > tp->rx_opt.mss_clamp)
1318                 mss_now = tp->rx_opt.mss_clamp;
1319 
1320         /* Now subtract optional transport overhead */
1321         mss_now -= icsk->icsk_ext_hdr_len;
1322 
1323         /* Then reserve room for full set of TCP options and 8 bytes of data */
1324         if (mss_now < 48)
1325                 mss_now = 48;
1326         return mss_now;
1327 }
1328 
1329 /* Calculate MSS. Not accounting for SACKs here.  */
1330 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1331 {
1332         /* Subtract TCP options size, not including SACKs */
1333         return __tcp_mtu_to_mss(sk, pmtu) -
1334                (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1335 }
1336 
1337 /* Inverse of above */
1338 int tcp_mss_to_mtu(struct sock *sk, int mss)
1339 {
1340         const struct tcp_sock *tp = tcp_sk(sk);
1341         const struct inet_connection_sock *icsk = inet_csk(sk);
1342         int mtu;
1343 
1344         mtu = mss +
1345               tp->tcp_header_len +
1346               icsk->icsk_ext_hdr_len +
1347               icsk->icsk_af_ops->net_header_len;
1348 
1349         /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1350         if (icsk->icsk_af_ops->net_frag_header_len) {
1351                 const struct dst_entry *dst = __sk_dst_get(sk);
1352 
1353                 if (dst && dst_allfrag(dst))
1354                         mtu += icsk->icsk_af_ops->net_frag_header_len;
1355         }
1356         return mtu;
1357 }
1358 
1359 /* MTU probing init per socket */
1360 void tcp_mtup_init(struct sock *sk)
1361 {
1362         struct tcp_sock *tp = tcp_sk(sk);
1363         struct inet_connection_sock *icsk = inet_csk(sk);
1364         struct net *net = sock_net(sk);
1365 
1366         icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1367         icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1368                                icsk->icsk_af_ops->net_header_len;
1369         icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1370         icsk->icsk_mtup.probe_size = 0;
1371         if (icsk->icsk_mtup.enabled)
1372                 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1373 }
1374 EXPORT_SYMBOL(tcp_mtup_init);
1375 
1376 /* This function synchronize snd mss to current pmtu/exthdr set.
1377 
1378    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1379    for TCP options, but includes only bare TCP header.
1380 
1381    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1382    It is minimum of user_mss and mss received with SYN.
1383    It also does not include TCP options.
1384 
1385    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1386 
1387    tp->mss_cache is current effective sending mss, including
1388    all tcp options except for SACKs. It is evaluated,
1389    taking into account current pmtu, but never exceeds
1390    tp->rx_opt.mss_clamp.
1391 
1392    NOTE1. rfc1122 clearly states that advertised MSS
1393    DOES NOT include either tcp or ip options.
1394 
1395    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1396    are READ ONLY outside this function.         --ANK (980731)
1397  */
1398 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1399 {
1400         struct tcp_sock *tp = tcp_sk(sk);
1401         struct inet_connection_sock *icsk = inet_csk(sk);
1402         int mss_now;
1403 
1404         if (icsk->icsk_mtup.search_high > pmtu)
1405                 icsk->icsk_mtup.search_high = pmtu;
1406 
1407         mss_now = tcp_mtu_to_mss(sk, pmtu);
1408         mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1409 
1410         /* And store cached results */
1411         icsk->icsk_pmtu_cookie = pmtu;
1412         if (icsk->icsk_mtup.enabled)
1413                 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1414         tp->mss_cache = mss_now;
1415 
1416         return mss_now;
1417 }
1418 EXPORT_SYMBOL(tcp_sync_mss);
1419 
1420 /* Compute the current effective MSS, taking SACKs and IP options,
1421  * and even PMTU discovery events into account.
1422  */
1423 unsigned int tcp_current_mss(struct sock *sk)
1424 {
1425         const struct tcp_sock *tp = tcp_sk(sk);
1426         const struct dst_entry *dst = __sk_dst_get(sk);
1427         u32 mss_now;
1428         unsigned int header_len;
1429         struct tcp_out_options opts;
1430         struct tcp_md5sig_key *md5;
1431 
1432         mss_now = tp->mss_cache;
1433 
1434         if (dst) {
1435                 u32 mtu = dst_mtu(dst);
1436                 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1437                         mss_now = tcp_sync_mss(sk, mtu);
1438         }
1439 
1440         header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1441                      sizeof(struct tcphdr);
1442         /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1443          * some common options. If this is an odd packet (because we have SACK
1444          * blocks etc) then our calculated header_len will be different, and
1445          * we have to adjust mss_now correspondingly */
1446         if (header_len != tp->tcp_header_len) {
1447                 int delta = (int) header_len - tp->tcp_header_len;
1448                 mss_now -= delta;
1449         }
1450 
1451         return mss_now;
1452 }
1453 
1454 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1455  * As additional protections, we do not touch cwnd in retransmission phases,
1456  * and if application hit its sndbuf limit recently.
1457  */
1458 static void tcp_cwnd_application_limited(struct sock *sk)
1459 {
1460         struct tcp_sock *tp = tcp_sk(sk);
1461 
1462         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1463             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1464                 /* Limited by application or receiver window. */
1465                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1466                 u32 win_used = max(tp->snd_cwnd_used, init_win);
1467                 if (win_used < tp->snd_cwnd) {
1468                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
1469                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1470                 }
1471                 tp->snd_cwnd_used = 0;
1472         }
1473         tp->snd_cwnd_stamp = tcp_time_stamp;
1474 }
1475 
1476 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1477 {
1478         struct tcp_sock *tp = tcp_sk(sk);
1479 
1480         /* Track the maximum number of outstanding packets in each
1481          * window, and remember whether we were cwnd-limited then.
1482          */
1483         if (!before(tp->snd_una, tp->max_packets_seq) ||
1484             tp->packets_out > tp->max_packets_out) {
1485                 tp->max_packets_out = tp->packets_out;
1486                 tp->max_packets_seq = tp->snd_nxt;
1487                 tp->is_cwnd_limited = is_cwnd_limited;
1488         }
1489 
1490         if (tcp_is_cwnd_limited(sk)) {
1491                 /* Network is feed fully. */
1492                 tp->snd_cwnd_used = 0;
1493                 tp->snd_cwnd_stamp = tcp_time_stamp;
1494         } else {
1495                 /* Network starves. */
1496                 if (tp->packets_out > tp->snd_cwnd_used)
1497                         tp->snd_cwnd_used = tp->packets_out;
1498 
1499                 if (sysctl_tcp_slow_start_after_idle &&
1500                     (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1501                         tcp_cwnd_application_limited(sk);
1502         }
1503 }
1504 
1505 /* Minshall's variant of the Nagle send check. */
1506 static bool tcp_minshall_check(const struct tcp_sock *tp)
1507 {
1508         return after(tp->snd_sml, tp->snd_una) &&
1509                 !after(tp->snd_sml, tp->snd_nxt);
1510 }
1511 
1512 /* Update snd_sml if this skb is under mss
1513  * Note that a TSO packet might end with a sub-mss segment
1514  * The test is really :
1515  * if ((skb->len % mss) != 0)
1516  *        tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1517  * But we can avoid doing the divide again given we already have
1518  *  skb_pcount = skb->len / mss_now
1519  */
1520 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1521                                 const struct sk_buff *skb)
1522 {
1523         if (skb->len < tcp_skb_pcount(skb) * mss_now)
1524                 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1525 }
1526 
1527 /* Return false, if packet can be sent now without violation Nagle's rules:
1528  * 1. It is full sized. (provided by caller in %partial bool)
1529  * 2. Or it contains FIN. (already checked by caller)
1530  * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1531  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1532  *    With Minshall's modification: all sent small packets are ACKed.
1533  */
1534 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1535                             int nonagle)
1536 {
1537         return partial &&
1538                 ((nonagle & TCP_NAGLE_CORK) ||
1539                  (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1540 }
1541 
1542 /* Return how many segs we'd like on a TSO packet,
1543  * to send one TSO packet per ms
1544  */
1545 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now)
1546 {
1547         u32 bytes, segs;
1548 
1549         bytes = min(sk->sk_pacing_rate >> 10,
1550                     sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1551 
1552         /* Goal is to send at least one packet per ms,
1553          * not one big TSO packet every 100 ms.
1554          * This preserves ACK clocking and is consistent
1555          * with tcp_tso_should_defer() heuristic.
1556          */
1557         segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs);
1558 
1559         return min_t(u32, segs, sk->sk_gso_max_segs);
1560 }
1561 
1562 /* Returns the portion of skb which can be sent right away */
1563 static unsigned int tcp_mss_split_point(const struct sock *sk,
1564                                         const struct sk_buff *skb,
1565                                         unsigned int mss_now,
1566                                         unsigned int max_segs,
1567                                         int nonagle)
1568 {
1569         const struct tcp_sock *tp = tcp_sk(sk);
1570         u32 partial, needed, window, max_len;
1571 
1572         window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1573         max_len = mss_now * max_segs;
1574 
1575         if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1576                 return max_len;
1577 
1578         needed = min(skb->len, window);
1579 
1580         if (max_len <= needed)
1581                 return max_len;
1582 
1583         partial = needed % mss_now;
1584         /* If last segment is not a full MSS, check if Nagle rules allow us
1585          * to include this last segment in this skb.
1586          * Otherwise, we'll split the skb at last MSS boundary
1587          */
1588         if (tcp_nagle_check(partial != 0, tp, nonagle))
1589                 return needed - partial;
1590 
1591         return needed;
1592 }
1593 
1594 /* Can at least one segment of SKB be sent right now, according to the
1595  * congestion window rules?  If so, return how many segments are allowed.
1596  */
1597 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1598                                          const struct sk_buff *skb)
1599 {
1600         u32 in_flight, cwnd, halfcwnd;
1601 
1602         /* Don't be strict about the congestion window for the final FIN.  */
1603         if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1604             tcp_skb_pcount(skb) == 1)
1605                 return 1;
1606 
1607         in_flight = tcp_packets_in_flight(tp);
1608         cwnd = tp->snd_cwnd;
1609         if (in_flight >= cwnd)
1610                 return 0;
1611 
1612         /* For better scheduling, ensure we have at least
1613          * 2 GSO packets in flight.
1614          */
1615         halfcwnd = max(cwnd >> 1, 1U);
1616         return min(halfcwnd, cwnd - in_flight);
1617 }
1618 
1619 /* Initialize TSO state of a skb.
1620  * This must be invoked the first time we consider transmitting
1621  * SKB onto the wire.
1622  */
1623 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1624 {
1625         int tso_segs = tcp_skb_pcount(skb);
1626 
1627         if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1628                 tcp_set_skb_tso_segs(skb, mss_now);
1629                 tso_segs = tcp_skb_pcount(skb);
1630         }
1631         return tso_segs;
1632 }
1633 
1634 
1635 /* Return true if the Nagle test allows this packet to be
1636  * sent now.
1637  */
1638 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1639                                   unsigned int cur_mss, int nonagle)
1640 {
1641         /* Nagle rule does not apply to frames, which sit in the middle of the
1642          * write_queue (they have no chances to get new data).
1643          *
1644          * This is implemented in the callers, where they modify the 'nonagle'
1645          * argument based upon the location of SKB in the send queue.
1646          */
1647         if (nonagle & TCP_NAGLE_PUSH)
1648                 return true;
1649 
1650         /* Don't use the nagle rule for urgent data (or for the final FIN). */
1651         if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1652                 return true;
1653 
1654         if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1655                 return true;
1656 
1657         return false;
1658 }
1659 
1660 /* Does at least the first segment of SKB fit into the send window? */
1661 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1662                              const struct sk_buff *skb,
1663                              unsigned int cur_mss)
1664 {
1665         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1666 
1667         if (skb->len > cur_mss)
1668                 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1669 
1670         return !after(end_seq, tcp_wnd_end(tp));
1671 }
1672 
1673 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1674  * should be put on the wire right now.  If so, it returns the number of
1675  * packets allowed by the congestion window.
1676  */
1677 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1678                                  unsigned int cur_mss, int nonagle)
1679 {
1680         const struct tcp_sock *tp = tcp_sk(sk);
1681         unsigned int cwnd_quota;
1682 
1683         tcp_init_tso_segs(skb, cur_mss);
1684 
1685         if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1686                 return 0;
1687 
1688         cwnd_quota = tcp_cwnd_test(tp, skb);
1689         if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1690                 cwnd_quota = 0;
1691 
1692         return cwnd_quota;
1693 }
1694 
1695 /* Test if sending is allowed right now. */
1696 bool tcp_may_send_now(struct sock *sk)
1697 {
1698         const struct tcp_sock *tp = tcp_sk(sk);
1699         struct sk_buff *skb = tcp_send_head(sk);
1700 
1701         return skb &&
1702                 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1703                              (tcp_skb_is_last(sk, skb) ?
1704                               tp->nonagle : TCP_NAGLE_PUSH));
1705 }
1706 
1707 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1708  * which is put after SKB on the list.  It is very much like
1709  * tcp_fragment() except that it may make several kinds of assumptions
1710  * in order to speed up the splitting operation.  In particular, we
1711  * know that all the data is in scatter-gather pages, and that the
1712  * packet has never been sent out before (and thus is not cloned).
1713  */
1714 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1715                         unsigned int mss_now, gfp_t gfp)
1716 {
1717         struct sk_buff *buff;
1718         int nlen = skb->len - len;
1719         u8 flags;
1720 
1721         /* All of a TSO frame must be composed of paged data.  */
1722         if (skb->len != skb->data_len)
1723                 return tcp_fragment(sk, skb, len, mss_now, gfp);
1724 
1725         buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1726         if (unlikely(!buff))
1727                 return -ENOMEM;
1728 
1729         sk->sk_wmem_queued += buff->truesize;
1730         sk_mem_charge(sk, buff->truesize);
1731         buff->truesize += nlen;
1732         skb->truesize -= nlen;
1733 
1734         /* Correct the sequence numbers. */
1735         TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1736         TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1737         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1738 
1739         /* PSH and FIN should only be set in the second packet. */
1740         flags = TCP_SKB_CB(skb)->tcp_flags;
1741         TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1742         TCP_SKB_CB(buff)->tcp_flags = flags;
1743 
1744         /* This packet was never sent out yet, so no SACK bits. */
1745         TCP_SKB_CB(buff)->sacked = 0;
1746 
1747         tcp_skb_fragment_eor(skb, buff);
1748 
1749         buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1750         skb_split(skb, buff, len);
1751         tcp_fragment_tstamp(skb, buff);
1752 
1753         /* Fix up tso_factor for both original and new SKB.  */
1754         tcp_set_skb_tso_segs(skb, mss_now);
1755         tcp_set_skb_tso_segs(buff, mss_now);
1756 
1757         /* Link BUFF into the send queue. */
1758         __skb_header_release(buff);
1759         tcp_insert_write_queue_after(skb, buff, sk);
1760 
1761         return 0;
1762 }
1763 
1764 /* Try to defer sending, if possible, in order to minimize the amount
1765  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1766  *
1767  * This algorithm is from John Heffner.
1768  */
1769 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1770                                  bool *is_cwnd_limited, u32 max_segs)
1771 {
1772         const struct inet_connection_sock *icsk = inet_csk(sk);
1773         u32 age, send_win, cong_win, limit, in_flight;
1774         struct tcp_sock *tp = tcp_sk(sk);
1775         struct skb_mstamp now;
1776         struct sk_buff *head;
1777         int win_divisor;
1778 
1779         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1780                 goto send_now;
1781 
1782         if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1783                 goto send_now;
1784 
1785         /* Avoid bursty behavior by allowing defer
1786          * only if the last write was recent.
1787          */
1788         if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1789                 goto send_now;
1790 
1791         in_flight = tcp_packets_in_flight(tp);
1792 
1793         BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1794 
1795         send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1796 
1797         /* From in_flight test above, we know that cwnd > in_flight.  */
1798         cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1799 
1800         limit = min(send_win, cong_win);
1801 
1802         /* If a full-sized TSO skb can be sent, do it. */
1803         if (limit >= max_segs * tp->mss_cache)
1804                 goto send_now;
1805 
1806         /* Middle in queue won't get any more data, full sendable already? */
1807         if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1808                 goto send_now;
1809 
1810         win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1811         if (win_divisor) {
1812                 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1813 
1814                 /* If at least some fraction of a window is available,
1815                  * just use it.
1816                  */
1817                 chunk /= win_divisor;
1818                 if (limit >= chunk)
1819                         goto send_now;
1820         } else {
1821                 /* Different approach, try not to defer past a single
1822                  * ACK.  Receiver should ACK every other full sized
1823                  * frame, so if we have space for more than 3 frames
1824                  * then send now.
1825                  */
1826                 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1827                         goto send_now;
1828         }
1829 
1830         head = tcp_write_queue_head(sk);
1831         skb_mstamp_get(&now);
1832         age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1833         /* If next ACK is likely to come too late (half srtt), do not defer */
1834         if (age < (tp->srtt_us >> 4))
1835                 goto send_now;
1836 
1837         /* Ok, it looks like it is advisable to defer. */
1838 
1839         if (cong_win < send_win && cong_win <= skb->len)
1840                 *is_cwnd_limited = true;
1841 
1842         return true;
1843 
1844 send_now:
1845         return false;
1846 }
1847 
1848 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1849 {
1850         struct inet_connection_sock *icsk = inet_csk(sk);
1851         struct tcp_sock *tp = tcp_sk(sk);
1852         struct net *net = sock_net(sk);
1853         u32 interval;
1854         s32 delta;
1855 
1856         interval = net->ipv4.sysctl_tcp_probe_interval;
1857         delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1858         if (unlikely(delta >= interval * HZ)) {
1859                 int mss = tcp_current_mss(sk);
1860 
1861                 /* Update current search range */
1862                 icsk->icsk_mtup.probe_size = 0;
1863                 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1864                         sizeof(struct tcphdr) +
1865                         icsk->icsk_af_ops->net_header_len;
1866                 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1867 
1868                 /* Update probe time stamp */
1869                 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1870         }
1871 }
1872 
1873 /* Create a new MTU probe if we are ready.
1874  * MTU probe is regularly attempting to increase the path MTU by
1875  * deliberately sending larger packets.  This discovers routing
1876  * changes resulting in larger path MTUs.
1877  *
1878  * Returns 0 if we should wait to probe (no cwnd available),
1879  *         1 if a probe was sent,
1880  *         -1 otherwise
1881  */
1882 static int tcp_mtu_probe(struct sock *sk)
1883 {
1884         struct tcp_sock *tp = tcp_sk(sk);
1885         struct inet_connection_sock *icsk = inet_csk(sk);
1886         struct sk_buff *skb, *nskb, *next;
1887         struct net *net = sock_net(sk);
1888         int len;
1889         int probe_size;
1890         int size_needed;
1891         int copy;
1892         int mss_now;
1893         int interval;
1894 
1895         /* Not currently probing/verifying,
1896          * not in recovery,
1897          * have enough cwnd, and
1898          * not SACKing (the variable headers throw things off) */
1899         if (!icsk->icsk_mtup.enabled ||
1900             icsk->icsk_mtup.probe_size ||
1901             inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1902             tp->snd_cwnd < 11 ||
1903             tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1904                 return -1;
1905 
1906         /* Use binary search for probe_size between tcp_mss_base,
1907          * and current mss_clamp. if (search_high - search_low)
1908          * smaller than a threshold, backoff from probing.
1909          */
1910         mss_now = tcp_current_mss(sk);
1911         probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1912                                     icsk->icsk_mtup.search_low) >> 1);
1913         size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1914         interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
1915         /* When misfortune happens, we are reprobing actively,
1916          * and then reprobe timer has expired. We stick with current
1917          * probing process by not resetting search range to its orignal.
1918          */
1919         if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
1920                 interval < net->ipv4.sysctl_tcp_probe_threshold) {
1921                 /* Check whether enough time has elaplased for
1922                  * another round of probing.
1923                  */
1924                 tcp_mtu_check_reprobe(sk);
1925                 return -1;
1926         }
1927 
1928         /* Have enough data in the send queue to probe? */
1929         if (tp->write_seq - tp->snd_nxt < size_needed)
1930                 return -1;
1931 
1932         if (tp->snd_wnd < size_needed)
1933                 return -1;
1934         if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1935                 return 0;
1936 
1937         /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1938         if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1939                 if (!tcp_packets_in_flight(tp))
1940                         return -1;
1941                 else
1942                         return 0;
1943         }
1944 
1945         /* We're allowed to probe.  Build it now. */
1946         nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
1947         if (!nskb)
1948                 return -1;
1949         sk->sk_wmem_queued += nskb->truesize;
1950         sk_mem_charge(sk, nskb->truesize);
1951 
1952         skb = tcp_send_head(sk);
1953 
1954         TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1955         TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1956         TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1957         TCP_SKB_CB(nskb)->sacked = 0;
1958         nskb->csum = 0;
1959         nskb->ip_summed = skb->ip_summed;
1960 
1961         tcp_insert_write_queue_before(nskb, skb, sk);
1962 
1963         len = 0;
1964         tcp_for_write_queue_from_safe(skb, next, sk) {
1965                 copy = min_t(int, skb->len, probe_size - len);
1966                 if (nskb->ip_summed)
1967                         skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1968                 else
1969                         nskb->csum = skb_copy_and_csum_bits(skb, 0,
1970                                                             skb_put(nskb, copy),
1971                                                             copy, nskb->csum);
1972 
1973                 if (skb->len <= copy) {
1974                         /* We've eaten all the data from this skb.
1975                          * Throw it away. */
1976                         TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1977                         tcp_unlink_write_queue(skb, sk);
1978                         sk_wmem_free_skb(sk, skb);
1979                 } else {
1980                         TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1981                                                    ~(TCPHDR_FIN|TCPHDR_PSH);
1982                         if (!skb_shinfo(skb)->nr_frags) {
1983                                 skb_pull(skb, copy);
1984                                 if (skb->ip_summed != CHECKSUM_PARTIAL)
1985                                         skb->csum = csum_partial(skb->data,
1986                                                                  skb->len, 0);
1987                         } else {
1988                                 __pskb_trim_head(skb, copy);
1989                                 tcp_set_skb_tso_segs(skb, mss_now);
1990                         }
1991                         TCP_SKB_CB(skb)->seq += copy;
1992                 }
1993 
1994                 len += copy;
1995 
1996                 if (len >= probe_size)
1997                         break;
1998         }
1999         tcp_init_tso_segs(nskb, nskb->len);
2000 
2001         /* We're ready to send.  If this fails, the probe will
2002          * be resegmented into mss-sized pieces by tcp_write_xmit().
2003          */
2004         if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2005                 /* Decrement cwnd here because we are sending
2006                  * effectively two packets. */
2007                 tp->snd_cwnd--;
2008                 tcp_event_new_data_sent(sk, nskb);
2009 
2010                 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2011                 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2012                 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2013 
2014                 return 1;
2015         }
2016 
2017         return -1;
2018 }
2019 
2020 /* This routine writes packets to the network.  It advances the
2021  * send_head.  This happens as incoming acks open up the remote
2022  * window for us.
2023  *
2024  * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2025  * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2026  * account rare use of URG, this is not a big flaw.
2027  *
2028  * Send at most one packet when push_one > 0. Temporarily ignore
2029  * cwnd limit to force at most one packet out when push_one == 2.
2030 
2031  * Returns true, if no segments are in flight and we have queued segments,
2032  * but cannot send anything now because of SWS or another problem.
2033  */
2034 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2035                            int push_one, gfp_t gfp)
2036 {
2037         struct tcp_sock *tp = tcp_sk(sk);
2038         struct sk_buff *skb;
2039         unsigned int tso_segs, sent_pkts;
2040         int cwnd_quota;
2041         int result;
2042         bool is_cwnd_limited = false;
2043         u32 max_segs;
2044 
2045         sent_pkts = 0;
2046 
2047         if (!push_one) {
2048                 /* Do MTU probing. */
2049                 result = tcp_mtu_probe(sk);
2050                 if (!result) {
2051                         return false;
2052                 } else if (result > 0) {
2053                         sent_pkts = 1;
2054                 }
2055         }
2056 
2057         max_segs = tcp_tso_autosize(sk, mss_now);
2058         while ((skb = tcp_send_head(sk))) {
2059                 unsigned int limit;
2060 
2061                 tso_segs = tcp_init_tso_segs(skb, mss_now);
2062                 BUG_ON(!tso_segs);
2063 
2064                 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2065                         /* "skb_mstamp" is used as a start point for the retransmit timer */
2066                         skb_mstamp_get(&skb->skb_mstamp);
2067                         goto repair; /* Skip network transmission */
2068                 }
2069 
2070                 cwnd_quota = tcp_cwnd_test(tp, skb);
2071                 if (!cwnd_quota) {
2072                         if (push_one == 2)
2073                                 /* Force out a loss probe pkt. */
2074                                 cwnd_quota = 1;
2075                         else
2076                                 break;
2077                 }
2078 
2079                 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2080                         break;
2081 
2082                 if (tso_segs == 1) {
2083                         if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2084                                                      (tcp_skb_is_last(sk, skb) ?
2085                                                       nonagle : TCP_NAGLE_PUSH))))
2086                                 break;
2087                 } else {
2088                         if (!push_one &&
2089                             tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2090                                                  max_segs))
2091                                 break;
2092                 }
2093 
2094                 limit = mss_now;
2095                 if (tso_segs > 1 && !tcp_urg_mode(tp))
2096                         limit = tcp_mss_split_point(sk, skb, mss_now,
2097                                                     min_t(unsigned int,
2098                                                           cwnd_quota,
2099                                                           max_segs),
2100                                                     nonagle);
2101 
2102                 if (skb->len > limit &&
2103                     unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2104                         break;
2105 
2106                 /* TCP Small Queues :
2107                  * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2108                  * This allows for :
2109                  *  - better RTT estimation and ACK scheduling
2110                  *  - faster recovery
2111                  *  - high rates
2112                  * Alas, some drivers / subsystems require a fair amount
2113                  * of queued bytes to ensure line rate.
2114                  * One example is wifi aggregation (802.11 AMPDU)
2115                  */
2116                 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2117                 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2118 
2119                 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2120                         set_bit(TSQ_THROTTLED, &tp->tsq_flags);
2121                         /* It is possible TX completion already happened
2122                          * before we set TSQ_THROTTLED, so we must
2123                          * test again the condition.
2124                          */
2125                         smp_mb__after_atomic();
2126                         if (atomic_read(&sk->sk_wmem_alloc) > limit)
2127                                 break;
2128                 }
2129 
2130                 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2131                         break;
2132 
2133 repair:
2134                 /* Advance the send_head.  This one is sent out.
2135                  * This call will increment packets_out.
2136                  */
2137                 tcp_event_new_data_sent(sk, skb);
2138 
2139                 tcp_minshall_update(tp, mss_now, skb);
2140                 sent_pkts += tcp_skb_pcount(skb);
2141 
2142                 if (push_one)
2143                         break;
2144         }
2145 
2146         if (likely(sent_pkts)) {
2147                 if (tcp_in_cwnd_reduction(sk))
2148                         tp->prr_out += sent_pkts;
2149 
2150                 /* Send one loss probe per tail loss episode. */
2151                 if (push_one != 2)
2152                         tcp_schedule_loss_probe(sk);
2153                 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2154                 tcp_cwnd_validate(sk, is_cwnd_limited);
2155                 return false;
2156         }
2157         return !tp->packets_out && tcp_send_head(sk);
2158 }
2159 
2160 bool tcp_schedule_loss_probe(struct sock *sk)
2161 {
2162         struct inet_connection_sock *icsk = inet_csk(sk);
2163         struct tcp_sock *tp = tcp_sk(sk);
2164         u32 timeout, tlp_time_stamp, rto_time_stamp;
2165         u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2166 
2167         if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2168                 return false;
2169         /* No consecutive loss probes. */
2170         if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2171                 tcp_rearm_rto(sk);
2172                 return false;
2173         }
2174         /* Don't do any loss probe on a Fast Open connection before 3WHS
2175          * finishes.
2176          */
2177         if (tp->fastopen_rsk)
2178                 return false;
2179 
2180         /* TLP is only scheduled when next timer event is RTO. */
2181         if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2182                 return false;
2183 
2184         /* Schedule a loss probe in 2*RTT for SACK capable connections
2185          * in Open state, that are either limited by cwnd or application.
2186          */
2187         if (sysctl_tcp_early_retrans < 3 || !tp->packets_out ||
2188             !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2189                 return false;
2190 
2191         if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2192              tcp_send_head(sk))
2193                 return false;
2194 
2195         /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2196          * for delayed ack when there's one outstanding packet. If no RTT
2197          * sample is available then probe after TCP_TIMEOUT_INIT.
2198          */
2199         timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
2200         if (tp->packets_out == 1)
2201                 timeout = max_t(u32, timeout,
2202                                 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2203         timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2204 
2205         /* If RTO is shorter, just schedule TLP in its place. */
2206         tlp_time_stamp = tcp_time_stamp + timeout;
2207         rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2208         if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2209                 s32 delta = rto_time_stamp - tcp_time_stamp;
2210                 if (delta > 0)
2211                         timeout = delta;
2212         }
2213 
2214         inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2215                                   TCP_RTO_MAX);
2216         return true;
2217 }
2218 
2219 /* Thanks to skb fast clones, we can detect if a prior transmit of
2220  * a packet is still in a qdisc or driver queue.
2221  * In this case, there is very little point doing a retransmit !
2222  */
2223 static bool skb_still_in_host_queue(const struct sock *sk,
2224                                     const struct sk_buff *skb)
2225 {
2226         if (unlikely(skb_fclone_busy(sk, skb))) {
2227                 NET_INC_STATS(sock_net(sk),
2228                               LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2229                 return true;
2230         }
2231         return false;
2232 }
2233 
2234 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2235  * retransmit the last segment.
2236  */
2237 void tcp_send_loss_probe(struct sock *sk)
2238 {
2239         struct tcp_sock *tp = tcp_sk(sk);
2240         struct sk_buff *skb;
2241         int pcount;
2242         int mss = tcp_current_mss(sk);
2243 
2244         skb = tcp_send_head(sk);
2245         if (skb) {
2246                 if (tcp_snd_wnd_test(tp, skb, mss)) {
2247                         pcount = tp->packets_out;
2248                         tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2249                         if (tp->packets_out > pcount)
2250                                 goto probe_sent;
2251                         goto rearm_timer;
2252                 }
2253                 skb = tcp_write_queue_prev(sk, skb);
2254         } else {
2255                 skb = tcp_write_queue_tail(sk);
2256         }
2257 
2258         /* At most one outstanding TLP retransmission. */
2259         if (tp->tlp_high_seq)
2260                 goto rearm_timer;
2261 
2262         /* Retransmit last segment. */
2263         if (WARN_ON(!skb))
2264                 goto rearm_timer;
2265 
2266         if (skb_still_in_host_queue(sk, skb))
2267                 goto rearm_timer;
2268 
2269         pcount = tcp_skb_pcount(skb);
2270         if (WARN_ON(!pcount))
2271                 goto rearm_timer;
2272 
2273         if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2274                 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2275                                           GFP_ATOMIC)))
2276                         goto rearm_timer;
2277                 skb = tcp_write_queue_next(sk, skb);
2278         }
2279 
2280         if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2281                 goto rearm_timer;
2282 
2283         if (__tcp_retransmit_skb(sk, skb, 1))
2284                 goto rearm_timer;
2285 
2286         /* Record snd_nxt for loss detection. */
2287         tp->tlp_high_seq = tp->snd_nxt;
2288 
2289 probe_sent:
2290         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2291         /* Reset s.t. tcp_rearm_rto will restart timer from now */
2292         inet_csk(sk)->icsk_pending = 0;
2293 rearm_timer:
2294         tcp_rearm_rto(sk);
2295 }
2296 
2297 /* Push out any pending frames which were held back due to
2298  * TCP_CORK or attempt at coalescing tiny packets.
2299  * The socket must be locked by the caller.
2300  */
2301 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2302                                int nonagle)
2303 {
2304         /* If we are closed, the bytes will have to remain here.
2305          * In time closedown will finish, we empty the write queue and
2306          * all will be happy.
2307          */
2308         if (unlikely(sk->sk_state == TCP_CLOSE))
2309                 return;
2310 
2311         if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2312                            sk_gfp_mask(sk, GFP_ATOMIC)))
2313                 tcp_check_probe_timer(sk);
2314 }
2315 
2316 /* Send _single_ skb sitting at the send head. This function requires
2317  * true push pending frames to setup probe timer etc.
2318  */
2319 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2320 {
2321         struct sk_buff *skb = tcp_send_head(sk);
2322 
2323         BUG_ON(!skb || skb->len < mss_now);
2324 
2325         tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2326 }
2327 
2328 /* This function returns the amount that we can raise the
2329  * usable window based on the following constraints
2330  *
2331  * 1. The window can never be shrunk once it is offered (RFC 793)
2332  * 2. We limit memory per socket
2333  *
2334  * RFC 1122:
2335  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2336  *  RECV.NEXT + RCV.WIN fixed until:
2337  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2338  *
2339  * i.e. don't raise the right edge of the window until you can raise
2340  * it at least MSS bytes.
2341  *
2342  * Unfortunately, the recommended algorithm breaks header prediction,
2343  * since header prediction assumes th->window stays fixed.
2344  *
2345  * Strictly speaking, keeping th->window fixed violates the receiver
2346  * side SWS prevention criteria. The problem is that under this rule
2347  * a stream of single byte packets will cause the right side of the
2348  * window to always advance by a single byte.
2349  *
2350  * Of course, if the sender implements sender side SWS prevention
2351  * then this will not be a problem.
2352  *
2353  * BSD seems to make the following compromise:
2354  *
2355  *      If the free space is less than the 1/4 of the maximum
2356  *      space available and the free space is less than 1/2 mss,
2357  *      then set the window to 0.
2358  *      [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2359  *      Otherwise, just prevent the window from shrinking
2360  *      and from being larger than the largest representable value.
2361  *
2362  * This prevents incremental opening of the window in the regime
2363  * where TCP is limited by the speed of the reader side taking
2364  * data out of the TCP receive queue. It does nothing about
2365  * those cases where the window is constrained on the sender side
2366  * because the pipeline is full.
2367  *
2368  * BSD also seems to "accidentally" limit itself to windows that are a
2369  * multiple of MSS, at least until the free space gets quite small.
2370  * This would appear to be a side effect of the mbuf implementation.
2371  * Combining these two algorithms results in the observed behavior
2372  * of having a fixed window size at almost all times.
2373  *
2374  * Below we obtain similar behavior by forcing the offered window to
2375  * a multiple of the mss when it is feasible to do so.
2376  *
2377  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2378  * Regular options like TIMESTAMP are taken into account.
2379  */
2380 u32 __tcp_select_window(struct sock *sk)
2381 {
2382         struct inet_connection_sock *icsk = inet_csk(sk);
2383         struct tcp_sock *tp = tcp_sk(sk);
2384         /* MSS for the peer's data.  Previous versions used mss_clamp
2385          * here.  I don't know if the value based on our guesses
2386          * of peer's MSS is better for the performance.  It's more correct
2387          * but may be worse for the performance because of rcv_mss
2388          * fluctuations.  --SAW  1998/11/1
2389          */
2390         int mss = icsk->icsk_ack.rcv_mss;
2391         int free_space = tcp_space(sk);
2392         int allowed_space = tcp_full_space(sk);
2393         int full_space = min_t(int, tp->window_clamp, allowed_space);
2394         int window;
2395 
2396         if (mss > full_space)
2397                 mss = full_space;
2398 
2399         if (free_space < (full_space >> 1)) {
2400                 icsk->icsk_ack.quick = 0;
2401 
2402                 if (tcp_under_memory_pressure(sk))
2403                         tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2404                                                4U * tp->advmss);
2405 
2406                 /* free_space might become our new window, make sure we don't
2407                  * increase it due to wscale.
2408                  */
2409                 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2410 
2411                 /* if free space is less than mss estimate, or is below 1/16th
2412                  * of the maximum allowed, try to move to zero-window, else
2413                  * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2414                  * new incoming data is dropped due to memory limits.
2415                  * With large window, mss test triggers way too late in order
2416                  * to announce zero window in time before rmem limit kicks in.
2417                  */
2418                 if (free_space < (allowed_space >> 4) || free_space < mss)
2419                         return 0;
2420         }
2421 
2422         if (free_space > tp->rcv_ssthresh)
2423                 free_space = tp->rcv_ssthresh;
2424 
2425         /* Don't do rounding if we are using window scaling, since the
2426          * scaled window will not line up with the MSS boundary anyway.
2427          */
2428         window = tp->rcv_wnd;
2429         if (tp->rx_opt.rcv_wscale) {
2430                 window = free_space;
2431 
2432                 /* Advertise enough space so that it won't get scaled away.
2433                  * Import case: prevent zero window announcement if
2434                  * 1<<rcv_wscale > mss.
2435                  */
2436                 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2437                         window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2438                                   << tp->rx_opt.rcv_wscale);
2439         } else {
2440                 /* Get the largest window that is a nice multiple of mss.
2441                  * Window clamp already applied above.
2442                  * If our current window offering is within 1 mss of the
2443                  * free space we just keep it. This prevents the divide
2444                  * and multiply from happening most of the time.
2445                  * We also don't do any window rounding when the free space
2446                  * is too small.
2447                  */
2448                 if (window <= free_space - mss || window > free_space)
2449                         window = (free_space / mss) * mss;
2450                 else if (mss == full_space &&
2451                          free_space > window + (full_space >> 1))
2452                         window = free_space;
2453         }
2454 
2455         return window;
2456 }
2457 
2458 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2459                              const struct sk_buff *next_skb)
2460 {
2461         if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2462                 const struct skb_shared_info *next_shinfo =
2463                         skb_shinfo(next_skb);
2464                 struct skb_shared_info *shinfo = skb_shinfo(skb);
2465 
2466                 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2467                 shinfo->tskey = next_shinfo->tskey;
2468                 TCP_SKB_CB(skb)->txstamp_ack |=
2469                         TCP_SKB_CB(next_skb)->txstamp_ack;
2470         }
2471 }
2472 
2473 /* Collapses two adjacent SKB's during retransmission. */
2474 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2475 {
2476         struct tcp_sock *tp = tcp_sk(sk);
2477         struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2478         int skb_size, next_skb_size;
2479 
2480         skb_size = skb->len;
2481         next_skb_size = next_skb->len;
2482 
2483         BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2484 
2485         tcp_highest_sack_combine(sk, next_skb, skb);
2486 
2487         tcp_unlink_write_queue(next_skb, sk);
2488 
2489         skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2490                                   next_skb_size);
2491 
2492         if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2493                 skb->ip_summed = CHECKSUM_PARTIAL;
2494 
2495         if (skb->ip_summed != CHECKSUM_PARTIAL)
2496                 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2497 
2498         /* Update sequence range on original skb. */
2499         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2500 
2501         /* Merge over control information. This moves PSH/FIN etc. over */
2502         TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2503 
2504         /* All done, get rid of second SKB and account for it so
2505          * packet counting does not break.
2506          */
2507         TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2508         TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2509 
2510         /* changed transmit queue under us so clear hints */
2511         tcp_clear_retrans_hints_partial(tp);
2512         if (next_skb == tp->retransmit_skb_hint)
2513                 tp->retransmit_skb_hint = skb;
2514 
2515         tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2516 
2517         tcp_skb_collapse_tstamp(skb, next_skb);
2518 
2519         sk_wmem_free_skb(sk, next_skb);
2520 }
2521 
2522 /* Check if coalescing SKBs is legal. */
2523 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2524 {
2525         if (tcp_skb_pcount(skb) > 1)
2526                 return false;
2527         /* TODO: SACK collapsing could be used to remove this condition */
2528         if (skb_shinfo(skb)->nr_frags != 0)
2529                 return false;
2530         if (skb_cloned(skb))
2531                 return false;
2532         if (skb == tcp_send_head(sk))
2533                 return false;
2534         /* Some heurestics for collapsing over SACK'd could be invented */
2535         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2536                 return false;
2537 
2538         return true;
2539 }
2540 
2541 /* Collapse packets in the retransmit queue to make to create
2542  * less packets on the wire. This is only done on retransmission.
2543  */
2544 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2545                                      int space)
2546 {
2547         struct tcp_sock *tp = tcp_sk(sk);
2548         struct sk_buff *skb = to, *tmp;
2549         bool first = true;
2550 
2551         if (!sysctl_tcp_retrans_collapse)
2552                 return;
2553         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2554                 return;
2555 
2556         tcp_for_write_queue_from_safe(skb, tmp, sk) {
2557                 if (!tcp_can_collapse(sk, skb))
2558                         break;
2559 
2560                 if (!tcp_skb_can_collapse_to(to))
2561                         break;
2562 
2563                 space -= skb->len;
2564 
2565                 if (first) {
2566                         first = false;
2567                         continue;
2568                 }
2569 
2570                 if (space < 0)
2571                         break;
2572                 /* Punt if not enough space exists in the first SKB for
2573                  * the data in the second
2574                  */
2575                 if (skb->len > skb_availroom(to))
2576                         break;
2577 
2578                 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2579                         break;
2580 
2581                 tcp_collapse_retrans(sk, to);
2582         }
2583 }
2584 
2585 /* This retransmits one SKB.  Policy decisions and retransmit queue
2586  * state updates are done by the caller.  Returns non-zero if an
2587  * error occurred which prevented the send.
2588  */
2589 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2590 {
2591         struct inet_connection_sock *icsk = inet_csk(sk);
2592         struct tcp_sock *tp = tcp_sk(sk);
2593         unsigned int cur_mss;
2594         int diff, len, err;
2595 
2596 
2597         /* Inconclusive MTU probe */
2598         if (icsk->icsk_mtup.probe_size)
2599                 icsk->icsk_mtup.probe_size = 0;
2600 
2601         /* Do not sent more than we queued. 1/4 is reserved for possible
2602          * copying overhead: fragmentation, tunneling, mangling etc.
2603          */
2604         if (atomic_read(&sk->sk_wmem_alloc) >
2605             min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2606                 return -EAGAIN;
2607 
2608         if (skb_still_in_host_queue(sk, skb))
2609                 return -EBUSY;
2610 
2611         if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2612                 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2613                         BUG();
2614                 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2615                         return -ENOMEM;
2616         }
2617 
2618         if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2619                 return -EHOSTUNREACH; /* Routing failure or similar. */
2620 
2621         cur_mss = tcp_current_mss(sk);
2622 
2623         /* If receiver has shrunk his window, and skb is out of
2624          * new window, do not retransmit it. The exception is the
2625          * case, when window is shrunk to zero. In this case
2626          * our retransmit serves as a zero window probe.
2627          */
2628         if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2629             TCP_SKB_CB(skb)->seq != tp->snd_una)
2630                 return -EAGAIN;
2631 
2632         len = cur_mss * segs;
2633         if (skb->len > len) {
2634                 if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC))
2635                         return -ENOMEM; /* We'll try again later. */
2636         } else {
2637                 if (skb_unclone(skb, GFP_ATOMIC))
2638                         return -ENOMEM;
2639 
2640                 diff = tcp_skb_pcount(skb);
2641                 tcp_set_skb_tso_segs(skb, cur_mss);
2642                 diff -= tcp_skb_pcount(skb);
2643                 if (diff)
2644                         tcp_adjust_pcount(sk, skb, diff);
2645                 if (skb->len < cur_mss)
2646                         tcp_retrans_try_collapse(sk, skb, cur_mss);
2647         }
2648 
2649         /* RFC3168, section 6.1.1.1. ECN fallback */
2650         if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2651                 tcp_ecn_clear_syn(sk, skb);
2652 
2653         /* make sure skb->data is aligned on arches that require it
2654          * and check if ack-trimming & collapsing extended the headroom
2655          * beyond what csum_start can cover.
2656          */
2657         if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2658                      skb_headroom(skb) >= 0xFFFF)) {
2659                 struct sk_buff *nskb;
2660 
2661                 skb_mstamp_get(&skb->skb_mstamp);
2662                 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2663                 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2664                              -ENOBUFS;
2665         } else {
2666                 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2667         }
2668 
2669         if (likely(!err)) {
2670                 segs = tcp_skb_pcount(skb);
2671 
2672                 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2673                 /* Update global TCP statistics. */
2674                 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2675                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2676                         __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2677                 tp->total_retrans += segs;
2678         }
2679         return err;
2680 }
2681 
2682 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2683 {
2684         struct tcp_sock *tp = tcp_sk(sk);
2685         int err = __tcp_retransmit_skb(sk, skb, segs);
2686 
2687         if (err == 0) {
2688 #if FASTRETRANS_DEBUG > 0
2689                 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2690                         net_dbg_ratelimited("retrans_out leaked\n");
2691                 }
2692 #endif
2693                 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2694                 tp->retrans_out += tcp_skb_pcount(skb);
2695 
2696                 /* Save stamp of the first retransmit. */
2697                 if (!tp->retrans_stamp)
2698                         tp->retrans_stamp = tcp_skb_timestamp(skb);
2699 
2700         } else if (err != -EBUSY) {
2701                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2702         }
2703 
2704         if (tp->undo_retrans < 0)
2705                 tp->undo_retrans = 0;
2706         tp->undo_retrans += tcp_skb_pcount(skb);
2707         return err;
2708 }
2709 
2710 /* Check if we forward retransmits are possible in the current
2711  * window/congestion state.
2712  */
2713 static bool tcp_can_forward_retransmit(struct sock *sk)
2714 {
2715         const struct inet_connection_sock *icsk = inet_csk(sk);
2716         const struct tcp_sock *tp = tcp_sk(sk);
2717 
2718         /* Forward retransmissions are possible only during Recovery. */
2719         if (icsk->icsk_ca_state != TCP_CA_Recovery)
2720                 return false;
2721 
2722         /* No forward retransmissions in Reno are possible. */
2723         if (tcp_is_reno(tp))
2724                 return false;
2725 
2726         /* Yeah, we have to make difficult choice between forward transmission
2727          * and retransmission... Both ways have their merits...
2728          *
2729          * For now we do not retransmit anything, while we have some new
2730          * segments to send. In the other cases, follow rule 3 for
2731          * NextSeg() specified in RFC3517.
2732          */
2733 
2734         if (tcp_may_send_now(sk))
2735                 return false;
2736 
2737         return true;
2738 }
2739 
2740 /* This gets called after a retransmit timeout, and the initially
2741  * retransmitted data is acknowledged.  It tries to continue
2742  * resending the rest of the retransmit queue, until either
2743  * we've sent it all or the congestion window limit is reached.
2744  * If doing SACK, the first ACK which comes back for a timeout
2745  * based retransmit packet might feed us FACK information again.
2746  * If so, we use it to avoid unnecessarily retransmissions.
2747  */
2748 void tcp_xmit_retransmit_queue(struct sock *sk)
2749 {
2750         const struct inet_connection_sock *icsk = inet_csk(sk);
2751         struct tcp_sock *tp = tcp_sk(sk);
2752         struct sk_buff *skb;
2753         struct sk_buff *hole = NULL;
2754         u32 max_segs, last_lost;
2755         int mib_idx;
2756         int fwd_rexmitting = 0;
2757 
2758         if (!tp->packets_out)
2759                 return;
2760 
2761         if (!tp->lost_out)
2762                 tp->retransmit_high = tp->snd_una;
2763 
2764         if (tp->retransmit_skb_hint) {
2765                 skb = tp->retransmit_skb_hint;
2766                 last_lost = TCP_SKB_CB(skb)->end_seq;
2767                 if (after(last_lost, tp->retransmit_high))
2768                         last_lost = tp->retransmit_high;
2769         } else {
2770                 skb = tcp_write_queue_head(sk);
2771                 last_lost = tp->snd_una;
2772         }
2773 
2774         max_segs = tcp_tso_autosize(sk, tcp_current_mss(sk));
2775         tcp_for_write_queue_from(skb, sk) {
2776                 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2777                 int segs;
2778 
2779                 if (skb == tcp_send_head(sk))
2780                         break;
2781                 /* we could do better than to assign each time */
2782                 if (!hole)
2783                         tp->retransmit_skb_hint = skb;
2784 
2785                 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2786                 if (segs <= 0)
2787                         return;
2788                 /* In case tcp_shift_skb_data() have aggregated large skbs,
2789                  * we need to make sure not sending too bigs TSO packets
2790                  */
2791                 segs = min_t(int, segs, max_segs);
2792 
2793                 if (fwd_rexmitting) {
2794 begin_fwd:
2795                         if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2796                                 break;
2797                         mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2798 
2799                 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2800                         tp->retransmit_high = last_lost;
2801                         if (!tcp_can_forward_retransmit(sk))
2802                                 break;
2803                         /* Backtrack if necessary to non-L'ed skb */
2804                         if (hole) {
2805                                 skb = hole;
2806                                 hole = NULL;
2807                         }
2808                         fwd_rexmitting = 1;
2809                         goto begin_fwd;
2810 
2811                 } else if (!(sacked & TCPCB_LOST)) {
2812                         if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2813                                 hole = skb;
2814                         continue;
2815 
2816                 } else {
2817                         last_lost = TCP_SKB_CB(skb)->end_seq;
2818                         if (icsk->icsk_ca_state != TCP_CA_Loss)
2819                                 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2820                         else
2821                                 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2822                 }
2823 
2824                 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2825                         continue;
2826 
2827                 if (tcp_retransmit_skb(sk, skb, segs))
2828                         return;
2829 
2830                 NET_INC_STATS(sock_net(sk), mib_idx);
2831 
2832                 if (tcp_in_cwnd_reduction(sk))
2833                         tp->prr_out += tcp_skb_pcount(skb);
2834 
2835                 if (skb == tcp_write_queue_head(sk))
2836                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2837                                                   inet_csk(sk)->icsk_rto,
2838                                                   TCP_RTO_MAX);
2839         }
2840 }
2841 
2842 /* We allow to exceed memory limits for FIN packets to expedite
2843  * connection tear down and (memory) recovery.
2844  * Otherwise tcp_send_fin() could be tempted to either delay FIN
2845  * or even be forced to close flow without any FIN.
2846  * In general, we want to allow one skb per socket to avoid hangs
2847  * with edge trigger epoll()
2848  */
2849 void sk_forced_mem_schedule(struct sock *sk, int size)
2850 {
2851         int amt;
2852 
2853         if (size <= sk->sk_forward_alloc)
2854                 return;
2855         amt = sk_mem_pages(size);
2856         sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2857         sk_memory_allocated_add(sk, amt);
2858 
2859         if (mem_cgroup_sockets_enabled && sk->sk_memcg)
2860                 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
2861 }
2862 
2863 /* Send a FIN. The caller locks the socket for us.
2864  * We should try to send a FIN packet really hard, but eventually give up.
2865  */
2866 void tcp_send_fin(struct sock *sk)
2867 {
2868         struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
2869         struct tcp_sock *tp = tcp_sk(sk);
2870 
2871         /* Optimization, tack on the FIN if we have one skb in write queue and
2872          * this skb was not yet sent, or we are under memory pressure.
2873          * Note: in the latter case, FIN packet will be sent after a timeout,
2874          * as TCP stack thinks it has already been transmitted.
2875          */
2876         if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
2877 coalesce:
2878                 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2879                 TCP_SKB_CB(tskb)->end_seq++;
2880                 tp->write_seq++;
2881                 if (!tcp_send_head(sk)) {
2882                         /* This means tskb was already sent.
2883                          * Pretend we included the FIN on previous transmit.
2884                          * We need to set tp->snd_nxt to the value it would have
2885                          * if FIN had been sent. This is because retransmit path
2886                          * does not change tp->snd_nxt.
2887                          */
2888                         tp->snd_nxt++;
2889                         return;
2890                 }
2891         } else {
2892                 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2893                 if (unlikely(!skb)) {
2894                         if (tskb)
2895                                 goto coalesce;
2896                         return;
2897                 }
2898                 skb_reserve(skb, MAX_TCP_HEADER);
2899                 sk_forced_mem_schedule(sk, skb->truesize);
2900                 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2901                 tcp_init_nondata_skb(skb, tp->write_seq,
2902                                      TCPHDR_ACK | TCPHDR_FIN);
2903                 tcp_queue_skb(sk, skb);
2904         }
2905         __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
2906 }
2907 
2908 /* We get here when a process closes a file descriptor (either due to
2909  * an explicit close() or as a byproduct of exit()'ing) and there
2910  * was unread data in the receive queue.  This behavior is recommended
2911  * by RFC 2525, section 2.17.  -DaveM
2912  */
2913 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2914 {
2915         struct sk_buff *skb;
2916 
2917         /* NOTE: No TCP options attached and we never retransmit this. */
2918         skb = alloc_skb(MAX_TCP_HEADER, priority);
2919         if (!skb) {
2920                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2921                 return;
2922         }
2923 
2924         /* Reserve space for headers and prepare control bits. */
2925         skb_reserve(skb, MAX_TCP_HEADER);
2926         tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2927                              TCPHDR_ACK | TCPHDR_RST);
2928         skb_mstamp_get(&skb->skb_mstamp);
2929         /* Send it off. */
2930         if (tcp_transmit_skb(sk, skb, 0, priority))
2931                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2932 
2933         TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2934 }
2935 
2936 /* Send a crossed SYN-ACK during socket establishment.
2937  * WARNING: This routine must only be called when we have already sent
2938  * a SYN packet that crossed the incoming SYN that caused this routine
2939  * to get called. If this assumption fails then the initial rcv_wnd
2940  * and rcv_wscale values will not be correct.
2941  */
2942 int tcp_send_synack(struct sock *sk)
2943 {
2944         struct sk_buff *skb;
2945 
2946         skb = tcp_write_queue_head(sk);
2947         if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2948                 pr_debug("%s: wrong queue state\n", __func__);
2949                 return -EFAULT;
2950         }
2951         if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2952                 if (skb_cloned(skb)) {
2953                         struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2954                         if (!nskb)
2955                                 return -ENOMEM;
2956                         tcp_unlink_write_queue(skb, sk);
2957                         __skb_header_release(nskb);
2958                         __tcp_add_write_queue_head(sk, nskb);
2959                         sk_wmem_free_skb(sk, skb);
2960                         sk->sk_wmem_queued += nskb->truesize;
2961                         sk_mem_charge(sk, nskb->truesize);
2962                         skb = nskb;
2963                 }
2964 
2965                 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2966                 tcp_ecn_send_synack(sk, skb);
2967         }
2968         return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2969 }
2970 
2971 /**
2972  * tcp_make_synack - Prepare a SYN-ACK.
2973  * sk: listener socket
2974  * dst: dst entry attached to the SYNACK
2975  * req: request_sock pointer
2976  *
2977  * Allocate one skb and build a SYNACK packet.
2978  * @dst is consumed : Caller should not use it again.
2979  */
2980 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
2981                                 struct request_sock *req,
2982                                 struct tcp_fastopen_cookie *foc,
2983                                 enum tcp_synack_type synack_type)
2984 {
2985         struct inet_request_sock *ireq = inet_rsk(req);
2986         const struct tcp_sock *tp = tcp_sk(sk);
2987         struct tcp_md5sig_key *md5 = NULL;
2988         struct tcp_out_options opts;
2989         struct sk_buff *skb;
2990         int tcp_header_size;
2991         struct tcphdr *th;
2992         u16 user_mss;
2993         int mss;
2994 
2995         skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2996         if (unlikely(!skb)) {
2997                 dst_release(dst);
2998                 return NULL;
2999         }
3000         /* Reserve space for headers. */
3001         skb_reserve(skb, MAX_TCP_HEADER);
3002 
3003         switch (synack_type) {
3004         case TCP_SYNACK_NORMAL:
3005                 skb_set_owner_w(skb, req_to_sk(req));
3006                 break;
3007         case TCP_SYNACK_COOKIE:
3008                 /* Under synflood, we do not attach skb to a socket,
3009                  * to avoid false sharing.
3010                  */
3011                 break;
3012         case TCP_SYNACK_FASTOPEN:
3013                 /* sk is a const pointer, because we want to express multiple
3014                  * cpu might call us concurrently.
3015                  * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3016                  */
3017                 skb_set_owner_w(skb, (struct sock *)sk);
3018                 break;
3019         }
3020         skb_dst_set(skb, dst);
3021 
3022         mss = dst_metric_advmss(dst);
3023         user_mss = READ_ONCE(tp->rx_opt.user_mss);
3024         if (user_mss && user_mss < mss)
3025                 mss = user_mss;
3026 
3027         memset(&opts, 0, sizeof(opts));
3028 #ifdef CONFIG_SYN_COOKIES
3029         if (unlikely(req->cookie_ts))
3030                 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
3031         else
3032 #endif
3033         skb_mstamp_get(&skb->skb_mstamp);
3034 
3035 #ifdef CONFIG_TCP_MD5SIG
3036         rcu_read_lock();
3037         md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3038 #endif
3039         skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3040         tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
3041                           sizeof(*th);
3042 
3043         skb_push(skb, tcp_header_size);
3044         skb_reset_transport_header(skb);
3045 
3046         th = (struct tcphdr *)skb->data;
3047         memset(th, 0, sizeof(struct tcphdr));
3048         th->syn = 1;
3049         th->ack = 1;
3050         tcp_ecn_make_synack(req, th);
3051         th->source = htons(ireq->ir_num);
3052         th->dest = ireq->ir_rmt_port;
3053         /* Setting of flags are superfluous here for callers (and ECE is
3054          * not even correctly set)
3055          */
3056         tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
3057                              TCPHDR_SYN | TCPHDR_ACK);
3058 
3059         th->seq = htonl(TCP_SKB_CB(skb)->seq);
3060         /* XXX data is queued and acked as is. No buffer/window check */
3061         th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3062 
3063         /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3064         th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3065         tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3066         th->doff = (tcp_header_size >> 2);
3067         __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3068 
3069 #ifdef CONFIG_TCP_MD5SIG
3070         /* Okay, we have all we need - do the md5 hash if needed */
3071         if (md5)
3072                 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3073                                                md5, req_to_sk(req), skb);
3074         rcu_read_unlock();
3075 #endif
3076 
3077         /* Do not fool tcpdump (if any), clean our debris */
3078         skb->tstamp.tv64 = 0;
3079         return skb;
3080 }
3081 EXPORT_SYMBOL(tcp_make_synack);
3082 
3083 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3084 {
3085         struct inet_connection_sock *icsk = inet_csk(sk);
3086         const struct tcp_congestion_ops *ca;
3087         u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3088 
3089         if (ca_key == TCP_CA_UNSPEC)
3090                 return;
3091 
3092         rcu_read_lock();
3093         ca = tcp_ca_find_key(ca_key);
3094         if (likely(ca && try_module_get(ca->owner))) {
3095                 module_put(icsk->icsk_ca_ops->owner);
3096                 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3097                 icsk->icsk_ca_ops = ca;
3098         }
3099         rcu_read_unlock();
3100 }
3101 
3102 /* Do all connect socket setups that can be done AF independent. */
3103 static void tcp_connect_init(struct sock *sk)
3104 {
3105         const struct dst_entry *dst = __sk_dst_get(sk);
3106         struct tcp_sock *tp = tcp_sk(sk);
3107         __u8 rcv_wscale;
3108 
3109         /* We'll fix this up when we get a response from the other end.
3110          * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3111          */
3112         tp->tcp_header_len = sizeof(struct tcphdr) +
3113                 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3114 
3115 #ifdef CONFIG_TCP_MD5SIG
3116         if (tp->af_specific->md5_lookup(sk, sk))
3117                 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3118 #endif
3119 
3120         /* If user gave his TCP_MAXSEG, record it to clamp */
3121         if (tp->rx_opt.user_mss)
3122                 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3123         tp->max_window = 0;
3124         tcp_mtup_init(sk);
3125         tcp_sync_mss(sk, dst_mtu(dst));
3126 
3127         tcp_ca_dst_init(sk, dst);
3128 
3129         if (!tp->window_clamp)
3130                 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3131         tp->advmss = dst_metric_advmss(dst);
3132         if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3133                 tp->advmss = tp->rx_opt.user_mss;
3134 
3135         tcp_initialize_rcv_mss(sk);
3136 
3137         /* limit the window selection if the user enforce a smaller rx buffer */
3138         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3139             (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3140                 tp->window_clamp = tcp_full_space(sk);
3141 
3142         tcp_select_initial_window(tcp_full_space(sk),
3143                                   tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3144                                   &tp->rcv_wnd,
3145                                   &tp->window_clamp,
3146                                   sysctl_tcp_window_scaling,
3147                                   &rcv_wscale,
3148                                   dst_metric(dst, RTAX_INITRWND));
3149 
3150         tp->rx_opt.rcv_wscale = rcv_wscale;
3151         tp->rcv_ssthresh = tp->rcv_wnd;
3152 
3153         sk->sk_err = 0;
3154         sock_reset_flag(sk, SOCK_DONE);
3155         tp->snd_wnd = 0;
3156         tcp_init_wl(tp, 0);
3157         tp->snd_una = tp->write_seq;
3158         tp->snd_sml = tp->write_seq;
3159         tp->snd_up = tp->write_seq;
3160         tp->snd_nxt = tp->write_seq;
3161 
3162         if (likely(!tp->repair))
3163                 tp->rcv_nxt = 0;
3164         else
3165                 tp->rcv_tstamp = tcp_time_stamp;
3166         tp->rcv_wup = tp->rcv_nxt;
3167         tp->copied_seq = tp->rcv_nxt;
3168 
3169         inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3170         inet_csk(sk)->icsk_retransmits = 0;
3171         tcp_clear_retrans(tp);
3172 }
3173 
3174 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3175 {
3176         struct tcp_sock *tp = tcp_sk(sk);
3177         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3178 
3179         tcb->end_seq += skb->len;
3180         __skb_header_release(skb);
3181         __tcp_add_write_queue_tail(sk, skb);
3182         sk->sk_wmem_queued += skb->truesize;
3183         sk_mem_charge(sk, skb->truesize);
3184         tp->write_seq = tcb->end_seq;
3185         tp->packets_out += tcp_skb_pcount(skb);
3186 }
3187 
3188 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3189  * queue a data-only packet after the regular SYN, such that regular SYNs
3190  * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3191  * only the SYN sequence, the data are retransmitted in the first ACK.
3192  * If cookie is not cached or other error occurs, falls back to send a
3193  * regular SYN with Fast Open cookie request option.
3194  */
3195 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3196 {
3197         struct tcp_sock *tp = tcp_sk(sk);
3198         struct tcp_fastopen_request *fo = tp->fastopen_req;
3199         int syn_loss = 0, space, err = 0;
3200         unsigned long last_syn_loss = 0;
3201         struct sk_buff *syn_data;
3202 
3203         tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
3204         tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3205                                &syn_loss, &last_syn_loss);
3206         /* Recurring FO SYN losses: revert to regular handshake temporarily */
3207         if (syn_loss > 1 &&
3208             time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3209                 fo->cookie.len = -1;
3210                 goto fallback;
3211         }
3212 
3213         if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3214                 fo->cookie.len = -1;
3215         else if (fo->cookie.len <= 0)
3216                 goto fallback;
3217 
3218         /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3219          * user-MSS. Reserve maximum option space for middleboxes that add
3220          * private TCP options. The cost is reduced data space in SYN :(
3221          */
3222         if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3223                 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3224         space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3225                 MAX_TCP_OPTION_SPACE;
3226 
3227         space = min_t(size_t, space, fo->size);
3228 
3229         /* limit to order-0 allocations */
3230         space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3231 
3232         syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3233         if (!syn_data)
3234                 goto fallback;
3235         syn_data->ip_summed = CHECKSUM_PARTIAL;
3236         memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3237         if (space) {
3238                 int copied = copy_from_iter(skb_put(syn_data, space), space,
3239                                             &fo->data->msg_iter);
3240                 if (unlikely(!copied)) {
3241                         kfree_skb(syn_data);
3242                         goto fallback;
3243                 }
3244                 if (copied != space) {
3245                         skb_trim(syn_data, copied);
3246                         space = copied;
3247                 }
3248         }
3249         /* No more data pending in inet_wait_for_connect() */
3250         if (space == fo->size)
3251                 fo->data = NULL;
3252         fo->copied = space;
3253 
3254         tcp_connect_queue_skb(sk, syn_data);
3255 
3256         err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3257 
3258         syn->skb_mstamp = syn_data->skb_mstamp;
3259 
3260         /* Now full SYN+DATA was cloned and sent (or not),
3261          * remove the SYN from the original skb (syn_data)
3262          * we keep in write queue in case of a retransmit, as we
3263          * also have the SYN packet (with no data) in the same queue.
3264          */
3265         TCP_SKB_CB(syn_data)->seq++;
3266         TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3267         if (!err) {
3268                 tp->syn_data = (fo->copied > 0);
3269                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3270                 goto done;
3271         }
3272 
3273 fallback:
3274         /* Send a regular SYN with Fast Open cookie request option */
3275         if (fo->cookie.len > 0)
3276                 fo->cookie.len = 0;
3277         err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3278         if (err)
3279                 tp->syn_fastopen = 0;
3280 done:
3281         fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
3282         return err;
3283 }
3284 
3285 /* Build a SYN and send it off. */
3286 int tcp_connect(struct sock *sk)
3287 {
3288         struct tcp_sock *tp = tcp_sk(sk);
3289         struct sk_buff *buff;
3290         int err;
3291 
3292         tcp_connect_init(sk);
3293 
3294         if (unlikely(tp->repair)) {
3295                 tcp_finish_connect(sk, NULL);
3296                 return 0;
3297         }
3298 
3299         buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3300         if (unlikely(!buff))
3301                 return -ENOBUFS;
3302 
3303         tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3304         tp->retrans_stamp = tcp_time_stamp;
3305         tcp_connect_queue_skb(sk, buff);
3306         tcp_ecn_send_syn(sk, buff);
3307 
3308         /* Send off SYN; include data in Fast Open. */
3309         err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3310               tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3311         if (err == -ECONNREFUSED)
3312                 return err;
3313 
3314         /* We change tp->snd_nxt after the tcp_transmit_skb() call
3315          * in order to make this packet get counted in tcpOutSegs.
3316          */
3317         tp->snd_nxt = tp->write_seq;
3318         tp->pushed_seq = tp->write_seq;
3319         TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3320 
3321         /* Timer for repeating the SYN until an answer. */
3322         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3323                                   inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3324         return 0;
3325 }
3326 EXPORT_SYMBOL(tcp_connect);
3327 
3328 /* Send out a delayed ack, the caller does the policy checking
3329  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
3330  * for details.
3331  */
3332 void tcp_send_delayed_ack(struct sock *sk)
3333 {
3334         struct inet_connection_sock *icsk = inet_csk(sk);
3335         int ato = icsk->icsk_ack.ato;
3336         unsigned long timeout;
3337 
3338         tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3339 
3340         if (ato > TCP_DELACK_MIN) {
3341                 const struct tcp_sock *tp = tcp_sk(sk);
3342                 int max_ato = HZ / 2;
3343 
3344                 if (icsk->icsk_ack.pingpong ||
3345                     (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3346                         max_ato = TCP_DELACK_MAX;
3347 
3348                 /* Slow path, intersegment interval is "high". */
3349 
3350                 /* If some rtt estimate is known, use it to bound delayed ack.
3351                  * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3352                  * directly.
3353                  */
3354                 if (tp->srtt_us) {
3355                         int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3356                                         TCP_DELACK_MIN);
3357 
3358                         if (rtt < max_ato)
3359                                 max_ato = rtt;
3360                 }
3361 
3362                 ato = min(ato, max_ato);
3363         }
3364 
3365         /* Stay within the limit we were given */
3366         timeout = jiffies + ato;
3367 
3368         /* Use new timeout only if there wasn't a older one earlier. */
3369         if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3370                 /* If delack timer was blocked or is about to expire,
3371                  * send ACK now.
3372                  */
3373                 if (icsk->icsk_ack.blocked ||
3374                     time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3375                         tcp_send_ack(sk);
3376                         return;
3377                 }
3378 
3379                 if (!time_before(timeout, icsk->icsk_ack.timeout))
3380                         timeout = icsk->icsk_ack.timeout;
3381         }
3382         icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3383         icsk->icsk_ack.timeout = timeout;
3384         sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3385 }
3386 
3387 /* This routine sends an ack and also updates the window. */
3388 void tcp_send_ack(struct sock *sk)
3389 {
3390         struct sk_buff *buff;
3391 
3392         /* If we have been reset, we may not send again. */
3393         if (sk->sk_state == TCP_CLOSE)
3394                 return;
3395 
3396         tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3397 
3398         /* We are not putting this on the write queue, so
3399          * tcp_transmit_skb() will set the ownership to this
3400          * sock.
3401          */
3402         buff = alloc_skb(MAX_TCP_HEADER,
3403                          sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3404         if (unlikely(!buff)) {
3405                 inet_csk_schedule_ack(sk);
3406                 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3407                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3408                                           TCP_DELACK_MAX, TCP_RTO_MAX);
3409                 return;
3410         }
3411 
3412         /* Reserve space for headers and prepare control bits. */
3413         skb_reserve(buff, MAX_TCP_HEADER);
3414         tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3415 
3416         /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3417          * too much.
3418          * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3419          * We also avoid tcp_wfree() overhead (cache line miss accessing
3420          * tp->tsq_flags) by using regular sock_wfree()
3421          */
3422         skb_set_tcp_pure_ack(buff);
3423 
3424         /* Send it off, this clears delayed acks for us. */
3425         skb_mstamp_get(&buff->skb_mstamp);
3426         tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3427 }
3428 EXPORT_SYMBOL_GPL(tcp_send_ack);
3429 
3430 /* This routine sends a packet with an out of date sequence
3431  * number. It assumes the other end will try to ack it.
3432  *
3433  * Question: what should we make while urgent mode?
3434  * 4.4BSD forces sending single byte of data. We cannot send
3435  * out of window data, because we have SND.NXT==SND.MAX...
3436  *
3437  * Current solution: to send TWO zero-length segments in urgent mode:
3438  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3439  * out-of-date with SND.UNA-1 to probe window.
3440  */
3441 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3442 {
3443         struct tcp_sock *tp = tcp_sk(sk);
3444         struct sk_buff *skb;
3445 
3446         /* We don't queue it, tcp_transmit_skb() sets ownership. */
3447         skb = alloc_skb(MAX_TCP_HEADER,
3448                         sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3449         if (!skb)
3450                 return -1;
3451 
3452         /* Reserve space for headers and set control bits. */
3453         skb_reserve(skb, MAX_TCP_HEADER);
3454         /* Use a previous sequence.  This should cause the other
3455          * end to send an ack.  Don't queue or clone SKB, just
3456          * send it.
3457          */
3458         tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3459         skb_mstamp_get(&skb->skb_mstamp);
3460         NET_INC_STATS(sock_net(sk), mib);
3461         return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3462 }
3463 
3464 void tcp_send_window_probe(struct sock *sk)
3465 {
3466         if (sk->sk_state == TCP_ESTABLISHED) {
3467                 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3468                 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3469         }
3470 }
3471 
3472 /* Initiate keepalive or window probe from timer. */
3473 int tcp_write_wakeup(struct sock *sk, int mib)
3474 {
3475         struct tcp_sock *tp = tcp_sk(sk);
3476         struct sk_buff *skb;
3477 
3478         if (sk->sk_state == TCP_CLOSE)
3479                 return -1;
3480 
3481         skb = tcp_send_head(sk);
3482         if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3483                 int err;
3484                 unsigned int mss = tcp_current_mss(sk);
3485                 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3486 
3487                 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3488                         tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3489 
3490                 /* We are probing the opening of a window
3491                  * but the window size is != 0
3492                  * must have been a result SWS avoidance ( sender )
3493                  */
3494                 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3495                     skb->len > mss) {
3496                         seg_size = min(seg_size, mss);
3497                         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3498                         if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3499                                 return -1;
3500                 } else if (!tcp_skb_pcount(skb))
3501                         tcp_set_skb_tso_segs(skb, mss);
3502 
3503                 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3504                 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3505                 if (!err)
3506                         tcp_event_new_data_sent(sk, skb);
3507                 return err;
3508         } else {
3509                 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3510                         tcp_xmit_probe_skb(sk, 1, mib);
3511                 return tcp_xmit_probe_skb(sk, 0, mib);
3512         }
3513 }
3514 
3515 /* A window probe timeout has occurred.  If window is not closed send
3516  * a partial packet else a zero probe.
3517  */
3518 void tcp_send_probe0(struct sock *sk)
3519 {
3520         struct inet_connection_sock *icsk = inet_csk(sk);
3521         struct tcp_sock *tp = tcp_sk(sk);
3522         struct net *net = sock_net(sk);
3523         unsigned long probe_max;
3524         int err;
3525 
3526         err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3527 
3528         if (tp->packets_out || !tcp_send_head(sk)) {
3529                 /* Cancel probe timer, if it is not required. */
3530                 icsk->icsk_probes_out = 0;
3531                 icsk->icsk_backoff = 0;
3532                 return;
3533         }
3534 
3535         if (err <= 0) {
3536                 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3537                         icsk->icsk_backoff++;
3538                 icsk->icsk_probes_out++;
3539                 probe_max = TCP_RTO_MAX;
3540         } else {
3541                 /* If packet was not sent due to local congestion,
3542                  * do not backoff and do not remember icsk_probes_out.
3543                  * Let local senders to fight for local resources.
3544                  *
3545                  * Use accumulated backoff yet.
3546                  */
3547                 if (!icsk->icsk_probes_out)
3548                         icsk->icsk_probes_out = 1;
3549                 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3550         }
3551         inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3552                                   tcp_probe0_when(sk, probe_max),
3553                                   TCP_RTO_MAX);
3554 }
3555 
3556 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3557 {
3558         const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3559         struct flowi fl;
3560         int res;
3561 
3562         tcp_rsk(req)->txhash = net_tx_rndhash();
3563         res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3564         if (!res) {
3565                 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3566                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3567         }
3568         return res;
3569 }
3570 EXPORT_SYMBOL(tcp_rtx_synack);
3571 

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us