Version:  2.0.40 2.2.26 2.4.37 3.9 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

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                                 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                                 tcp_hdr(skb)->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                         tcp_hdr(skb)->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 = tcp_hdr(skb);
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         if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
966                 /* RFC1323: The window in SYN & SYN/ACK segments
967                  * is never scaled.
968                  */
969                 th->window      = htons(min(tp->rcv_wnd, 65535U));
970         } else {
971                 th->window      = htons(tcp_select_window(sk));
972         }
973         th->check               = 0;
974         th->urg_ptr             = 0;
975 
976         /* The urg_mode check is necessary during a below snd_una win probe */
977         if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
978                 if (before(tp->snd_up, tcb->seq + 0x10000)) {
979                         th->urg_ptr = htons(tp->snd_up - tcb->seq);
980                         th->urg = 1;
981                 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
982                         th->urg_ptr = htons(0xFFFF);
983                         th->urg = 1;
984                 }
985         }
986 
987         tcp_options_write((__be32 *)(th + 1), tp, &opts);
988         skb_shinfo(skb)->gso_type = sk->sk_gso_type;
989         if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
990                 tcp_ecn_send(sk, skb, tcp_header_size);
991 
992 #ifdef CONFIG_TCP_MD5SIG
993         /* Calculate the MD5 hash, as we have all we need now */
994         if (md5) {
995                 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
996                 tp->af_specific->calc_md5_hash(opts.hash_location,
997                                                md5, sk, skb);
998         }
999 #endif
1000 
1001         icsk->icsk_af_ops->send_check(sk, skb);
1002 
1003         if (likely(tcb->tcp_flags & TCPHDR_ACK))
1004                 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1005 
1006         if (skb->len != tcp_header_size) {
1007                 tcp_event_data_sent(tp, sk);
1008                 tp->data_segs_out += tcp_skb_pcount(skb);
1009         }
1010 
1011         if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1012                 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1013                               tcp_skb_pcount(skb));
1014 
1015         tp->segs_out += tcp_skb_pcount(skb);
1016         /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1017         skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1018         skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1019 
1020         /* Our usage of tstamp should remain private */
1021         skb->tstamp.tv64 = 0;
1022 
1023         /* Cleanup our debris for IP stacks */
1024         memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1025                                sizeof(struct inet6_skb_parm)));
1026 
1027         err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1028 
1029         if (likely(err <= 0))
1030                 return err;
1031 
1032         tcp_enter_cwr(sk);
1033 
1034         return net_xmit_eval(err);
1035 }
1036 
1037 /* This routine just queues the buffer for sending.
1038  *
1039  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1040  * otherwise socket can stall.
1041  */
1042 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1043 {
1044         struct tcp_sock *tp = tcp_sk(sk);
1045 
1046         /* Advance write_seq and place onto the write_queue. */
1047         tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1048         __skb_header_release(skb);
1049         tcp_add_write_queue_tail(sk, skb);
1050         sk->sk_wmem_queued += skb->truesize;
1051         sk_mem_charge(sk, skb->truesize);
1052 }
1053 
1054 /* Initialize TSO segments for a packet. */
1055 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1056 {
1057         if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1058                 /* Avoid the costly divide in the normal
1059                  * non-TSO case.
1060                  */
1061                 tcp_skb_pcount_set(skb, 1);
1062                 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1063         } else {
1064                 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1065                 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1066         }
1067 }
1068 
1069 /* When a modification to fackets out becomes necessary, we need to check
1070  * skb is counted to fackets_out or not.
1071  */
1072 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1073                                    int decr)
1074 {
1075         struct tcp_sock *tp = tcp_sk(sk);
1076 
1077         if (!tp->sacked_out || tcp_is_reno(tp))
1078                 return;
1079 
1080         if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1081                 tp->fackets_out -= decr;
1082 }
1083 
1084 /* Pcount in the middle of the write queue got changed, we need to do various
1085  * tweaks to fix counters
1086  */
1087 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1088 {
1089         struct tcp_sock *tp = tcp_sk(sk);
1090 
1091         tp->packets_out -= decr;
1092 
1093         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1094                 tp->sacked_out -= decr;
1095         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1096                 tp->retrans_out -= decr;
1097         if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1098                 tp->lost_out -= decr;
1099 
1100         /* Reno case is special. Sigh... */
1101         if (tcp_is_reno(tp) && decr > 0)
1102                 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1103 
1104         tcp_adjust_fackets_out(sk, skb, decr);
1105 
1106         if (tp->lost_skb_hint &&
1107             before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1108             (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1109                 tp->lost_cnt_hint -= decr;
1110 
1111         tcp_verify_left_out(tp);
1112 }
1113 
1114 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1115 {
1116         struct skb_shared_info *shinfo = skb_shinfo(skb);
1117 
1118         if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) &&
1119             !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1120                 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1121                 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1122 
1123                 shinfo->tx_flags &= ~tsflags;
1124                 shinfo2->tx_flags |= tsflags;
1125                 swap(shinfo->tskey, shinfo2->tskey);
1126         }
1127 }
1128 
1129 /* Function to create two new TCP segments.  Shrinks the given segment
1130  * to the specified size and appends a new segment with the rest of the
1131  * packet to the list.  This won't be called frequently, I hope.
1132  * Remember, these are still headerless SKBs at this point.
1133  */
1134 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1135                  unsigned int mss_now, gfp_t gfp)
1136 {
1137         struct tcp_sock *tp = tcp_sk(sk);
1138         struct sk_buff *buff;
1139         int nsize, old_factor;
1140         int nlen;
1141         u8 flags;
1142 
1143         if (WARN_ON(len > skb->len))
1144                 return -EINVAL;
1145 
1146         nsize = skb_headlen(skb) - len;
1147         if (nsize < 0)
1148                 nsize = 0;
1149 
1150         if (skb_unclone(skb, gfp))
1151                 return -ENOMEM;
1152 
1153         /* Get a new skb... force flag on. */
1154         buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1155         if (!buff)
1156                 return -ENOMEM; /* We'll just try again later. */
1157 
1158         sk->sk_wmem_queued += buff->truesize;
1159         sk_mem_charge(sk, buff->truesize);
1160         nlen = skb->len - len - nsize;
1161         buff->truesize += nlen;
1162         skb->truesize -= nlen;
1163 
1164         /* Correct the sequence numbers. */
1165         TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1166         TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1167         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1168 
1169         /* PSH and FIN should only be set in the second packet. */
1170         flags = TCP_SKB_CB(skb)->tcp_flags;
1171         TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1172         TCP_SKB_CB(buff)->tcp_flags = flags;
1173         TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1174 
1175         if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1176                 /* Copy and checksum data tail into the new buffer. */
1177                 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1178                                                        skb_put(buff, nsize),
1179                                                        nsize, 0);
1180 
1181                 skb_trim(skb, len);
1182 
1183                 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1184         } else {
1185                 skb->ip_summed = CHECKSUM_PARTIAL;
1186                 skb_split(skb, buff, len);
1187         }
1188 
1189         buff->ip_summed = skb->ip_summed;
1190 
1191         buff->tstamp = skb->tstamp;
1192         tcp_fragment_tstamp(skb, buff);
1193 
1194         old_factor = tcp_skb_pcount(skb);
1195 
1196         /* Fix up tso_factor for both original and new SKB.  */
1197         tcp_set_skb_tso_segs(skb, mss_now);
1198         tcp_set_skb_tso_segs(buff, mss_now);
1199 
1200         /* If this packet has been sent out already, we must
1201          * adjust the various packet counters.
1202          */
1203         if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1204                 int diff = old_factor - tcp_skb_pcount(skb) -
1205                         tcp_skb_pcount(buff);
1206 
1207                 if (diff)
1208                         tcp_adjust_pcount(sk, skb, diff);
1209         }
1210 
1211         /* Link BUFF into the send queue. */
1212         __skb_header_release(buff);
1213         tcp_insert_write_queue_after(skb, buff, sk);
1214 
1215         return 0;
1216 }
1217 
1218 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1219  * eventually). The difference is that pulled data not copied, but
1220  * immediately discarded.
1221  */
1222 static void __pskb_trim_head(struct sk_buff *skb, int len)
1223 {
1224         struct skb_shared_info *shinfo;
1225         int i, k, eat;
1226 
1227         eat = min_t(int, len, skb_headlen(skb));
1228         if (eat) {
1229                 __skb_pull(skb, eat);
1230                 len -= eat;
1231                 if (!len)
1232                         return;
1233         }
1234         eat = len;
1235         k = 0;
1236         shinfo = skb_shinfo(skb);
1237         for (i = 0; i < shinfo->nr_frags; i++) {
1238                 int size = skb_frag_size(&shinfo->frags[i]);
1239 
1240                 if (size <= eat) {
1241                         skb_frag_unref(skb, i);
1242                         eat -= size;
1243                 } else {
1244                         shinfo->frags[k] = shinfo->frags[i];
1245                         if (eat) {
1246                                 shinfo->frags[k].page_offset += eat;
1247                                 skb_frag_size_sub(&shinfo->frags[k], eat);
1248                                 eat = 0;
1249                         }
1250                         k++;
1251                 }
1252         }
1253         shinfo->nr_frags = k;
1254 
1255         skb_reset_tail_pointer(skb);
1256         skb->data_len -= len;
1257         skb->len = skb->data_len;
1258 }
1259 
1260 /* Remove acked data from a packet in the transmit queue. */
1261 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1262 {
1263         if (skb_unclone(skb, GFP_ATOMIC))
1264                 return -ENOMEM;
1265 
1266         __pskb_trim_head(skb, len);
1267 
1268         TCP_SKB_CB(skb)->seq += len;
1269         skb->ip_summed = CHECKSUM_PARTIAL;
1270 
1271         skb->truesize        -= len;
1272         sk->sk_wmem_queued   -= len;
1273         sk_mem_uncharge(sk, len);
1274         sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1275 
1276         /* Any change of skb->len requires recalculation of tso factor. */
1277         if (tcp_skb_pcount(skb) > 1)
1278                 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1279 
1280         return 0;
1281 }
1282 
1283 /* Calculate MSS not accounting any TCP options.  */
1284 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1285 {
1286         const struct tcp_sock *tp = tcp_sk(sk);
1287         const struct inet_connection_sock *icsk = inet_csk(sk);
1288         int mss_now;
1289 
1290         /* Calculate base mss without TCP options:
1291            It is MMS_S - sizeof(tcphdr) of rfc1122
1292          */
1293         mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1294 
1295         /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1296         if (icsk->icsk_af_ops->net_frag_header_len) {
1297                 const struct dst_entry *dst = __sk_dst_get(sk);
1298 
1299                 if (dst && dst_allfrag(dst))
1300                         mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1301         }
1302 
1303         /* Clamp it (mss_clamp does not include tcp options) */
1304         if (mss_now > tp->rx_opt.mss_clamp)
1305                 mss_now = tp->rx_opt.mss_clamp;
1306 
1307         /* Now subtract optional transport overhead */
1308         mss_now -= icsk->icsk_ext_hdr_len;
1309 
1310         /* Then reserve room for full set of TCP options and 8 bytes of data */
1311         if (mss_now < 48)
1312                 mss_now = 48;
1313         return mss_now;
1314 }
1315 
1316 /* Calculate MSS. Not accounting for SACKs here.  */
1317 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1318 {
1319         /* Subtract TCP options size, not including SACKs */
1320         return __tcp_mtu_to_mss(sk, pmtu) -
1321                (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1322 }
1323 
1324 /* Inverse of above */
1325 int tcp_mss_to_mtu(struct sock *sk, int mss)
1326 {
1327         const struct tcp_sock *tp = tcp_sk(sk);
1328         const struct inet_connection_sock *icsk = inet_csk(sk);
1329         int mtu;
1330 
1331         mtu = mss +
1332               tp->tcp_header_len +
1333               icsk->icsk_ext_hdr_len +
1334               icsk->icsk_af_ops->net_header_len;
1335 
1336         /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1337         if (icsk->icsk_af_ops->net_frag_header_len) {
1338                 const struct dst_entry *dst = __sk_dst_get(sk);
1339 
1340                 if (dst && dst_allfrag(dst))
1341                         mtu += icsk->icsk_af_ops->net_frag_header_len;
1342         }
1343         return mtu;
1344 }
1345 
1346 /* MTU probing init per socket */
1347 void tcp_mtup_init(struct sock *sk)
1348 {
1349         struct tcp_sock *tp = tcp_sk(sk);
1350         struct inet_connection_sock *icsk = inet_csk(sk);
1351         struct net *net = sock_net(sk);
1352 
1353         icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1354         icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1355                                icsk->icsk_af_ops->net_header_len;
1356         icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1357         icsk->icsk_mtup.probe_size = 0;
1358         if (icsk->icsk_mtup.enabled)
1359                 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1360 }
1361 EXPORT_SYMBOL(tcp_mtup_init);
1362 
1363 /* This function synchronize snd mss to current pmtu/exthdr set.
1364 
1365    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1366    for TCP options, but includes only bare TCP header.
1367 
1368    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1369    It is minimum of user_mss and mss received with SYN.
1370    It also does not include TCP options.
1371 
1372    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1373 
1374    tp->mss_cache is current effective sending mss, including
1375    all tcp options except for SACKs. It is evaluated,
1376    taking into account current pmtu, but never exceeds
1377    tp->rx_opt.mss_clamp.
1378 
1379    NOTE1. rfc1122 clearly states that advertised MSS
1380    DOES NOT include either tcp or ip options.
1381 
1382    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1383    are READ ONLY outside this function.         --ANK (980731)
1384  */
1385 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1386 {
1387         struct tcp_sock *tp = tcp_sk(sk);
1388         struct inet_connection_sock *icsk = inet_csk(sk);
1389         int mss_now;
1390 
1391         if (icsk->icsk_mtup.search_high > pmtu)
1392                 icsk->icsk_mtup.search_high = pmtu;
1393 
1394         mss_now = tcp_mtu_to_mss(sk, pmtu);
1395         mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1396 
1397         /* And store cached results */
1398         icsk->icsk_pmtu_cookie = pmtu;
1399         if (icsk->icsk_mtup.enabled)
1400                 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1401         tp->mss_cache = mss_now;
1402 
1403         return mss_now;
1404 }
1405 EXPORT_SYMBOL(tcp_sync_mss);
1406 
1407 /* Compute the current effective MSS, taking SACKs and IP options,
1408  * and even PMTU discovery events into account.
1409  */
1410 unsigned int tcp_current_mss(struct sock *sk)
1411 {
1412         const struct tcp_sock *tp = tcp_sk(sk);
1413         const struct dst_entry *dst = __sk_dst_get(sk);
1414         u32 mss_now;
1415         unsigned int header_len;
1416         struct tcp_out_options opts;
1417         struct tcp_md5sig_key *md5;
1418 
1419         mss_now = tp->mss_cache;
1420 
1421         if (dst) {
1422                 u32 mtu = dst_mtu(dst);
1423                 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1424                         mss_now = tcp_sync_mss(sk, mtu);
1425         }
1426 
1427         header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1428                      sizeof(struct tcphdr);
1429         /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1430          * some common options. If this is an odd packet (because we have SACK
1431          * blocks etc) then our calculated header_len will be different, and
1432          * we have to adjust mss_now correspondingly */
1433         if (header_len != tp->tcp_header_len) {
1434                 int delta = (int) header_len - tp->tcp_header_len;
1435                 mss_now -= delta;
1436         }
1437 
1438         return mss_now;
1439 }
1440 
1441 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1442  * As additional protections, we do not touch cwnd in retransmission phases,
1443  * and if application hit its sndbuf limit recently.
1444  */
1445 static void tcp_cwnd_application_limited(struct sock *sk)
1446 {
1447         struct tcp_sock *tp = tcp_sk(sk);
1448 
1449         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1450             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1451                 /* Limited by application or receiver window. */
1452                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1453                 u32 win_used = max(tp->snd_cwnd_used, init_win);
1454                 if (win_used < tp->snd_cwnd) {
1455                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
1456                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1457                 }
1458                 tp->snd_cwnd_used = 0;
1459         }
1460         tp->snd_cwnd_stamp = tcp_time_stamp;
1461 }
1462 
1463 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1464 {
1465         struct tcp_sock *tp = tcp_sk(sk);
1466 
1467         /* Track the maximum number of outstanding packets in each
1468          * window, and remember whether we were cwnd-limited then.
1469          */
1470         if (!before(tp->snd_una, tp->max_packets_seq) ||
1471             tp->packets_out > tp->max_packets_out) {
1472                 tp->max_packets_out = tp->packets_out;
1473                 tp->max_packets_seq = tp->snd_nxt;
1474                 tp->is_cwnd_limited = is_cwnd_limited;
1475         }
1476 
1477         if (tcp_is_cwnd_limited(sk)) {
1478                 /* Network is feed fully. */
1479                 tp->snd_cwnd_used = 0;
1480                 tp->snd_cwnd_stamp = tcp_time_stamp;
1481         } else {
1482                 /* Network starves. */
1483                 if (tp->packets_out > tp->snd_cwnd_used)
1484                         tp->snd_cwnd_used = tp->packets_out;
1485 
1486                 if (sysctl_tcp_slow_start_after_idle &&
1487                     (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1488                         tcp_cwnd_application_limited(sk);
1489         }
1490 }
1491 
1492 /* Minshall's variant of the Nagle send check. */
1493 static bool tcp_minshall_check(const struct tcp_sock *tp)
1494 {
1495         return after(tp->snd_sml, tp->snd_una) &&
1496                 !after(tp->snd_sml, tp->snd_nxt);
1497 }
1498 
1499 /* Update snd_sml if this skb is under mss
1500  * Note that a TSO packet might end with a sub-mss segment
1501  * The test is really :
1502  * if ((skb->len % mss) != 0)
1503  *        tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1504  * But we can avoid doing the divide again given we already have
1505  *  skb_pcount = skb->len / mss_now
1506  */
1507 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1508                                 const struct sk_buff *skb)
1509 {
1510         if (skb->len < tcp_skb_pcount(skb) * mss_now)
1511                 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1512 }
1513 
1514 /* Return false, if packet can be sent now without violation Nagle's rules:
1515  * 1. It is full sized. (provided by caller in %partial bool)
1516  * 2. Or it contains FIN. (already checked by caller)
1517  * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1518  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1519  *    With Minshall's modification: all sent small packets are ACKed.
1520  */
1521 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1522                             int nonagle)
1523 {
1524         return partial &&
1525                 ((nonagle & TCP_NAGLE_CORK) ||
1526                  (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1527 }
1528 
1529 /* Return how many segs we'd like on a TSO packet,
1530  * to send one TSO packet per ms
1531  */
1532 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now)
1533 {
1534         u32 bytes, segs;
1535 
1536         bytes = min(sk->sk_pacing_rate >> 10,
1537                     sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1538 
1539         /* Goal is to send at least one packet per ms,
1540          * not one big TSO packet every 100 ms.
1541          * This preserves ACK clocking and is consistent
1542          * with tcp_tso_should_defer() heuristic.
1543          */
1544         segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs);
1545 
1546         return min_t(u32, segs, sk->sk_gso_max_segs);
1547 }
1548 
1549 /* Returns the portion of skb which can be sent right away */
1550 static unsigned int tcp_mss_split_point(const struct sock *sk,
1551                                         const struct sk_buff *skb,
1552                                         unsigned int mss_now,
1553                                         unsigned int max_segs,
1554                                         int nonagle)
1555 {
1556         const struct tcp_sock *tp = tcp_sk(sk);
1557         u32 partial, needed, window, max_len;
1558 
1559         window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1560         max_len = mss_now * max_segs;
1561 
1562         if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1563                 return max_len;
1564 
1565         needed = min(skb->len, window);
1566 
1567         if (max_len <= needed)
1568                 return max_len;
1569 
1570         partial = needed % mss_now;
1571         /* If last segment is not a full MSS, check if Nagle rules allow us
1572          * to include this last segment in this skb.
1573          * Otherwise, we'll split the skb at last MSS boundary
1574          */
1575         if (tcp_nagle_check(partial != 0, tp, nonagle))
1576                 return needed - partial;
1577 
1578         return needed;
1579 }
1580 
1581 /* Can at least one segment of SKB be sent right now, according to the
1582  * congestion window rules?  If so, return how many segments are allowed.
1583  */
1584 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1585                                          const struct sk_buff *skb)
1586 {
1587         u32 in_flight, cwnd, halfcwnd;
1588 
1589         /* Don't be strict about the congestion window for the final FIN.  */
1590         if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1591             tcp_skb_pcount(skb) == 1)
1592                 return 1;
1593 
1594         in_flight = tcp_packets_in_flight(tp);
1595         cwnd = tp->snd_cwnd;
1596         if (in_flight >= cwnd)
1597                 return 0;
1598 
1599         /* For better scheduling, ensure we have at least
1600          * 2 GSO packets in flight.
1601          */
1602         halfcwnd = max(cwnd >> 1, 1U);
1603         return min(halfcwnd, cwnd - in_flight);
1604 }
1605 
1606 /* Initialize TSO state of a skb.
1607  * This must be invoked the first time we consider transmitting
1608  * SKB onto the wire.
1609  */
1610 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1611 {
1612         int tso_segs = tcp_skb_pcount(skb);
1613 
1614         if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1615                 tcp_set_skb_tso_segs(skb, mss_now);
1616                 tso_segs = tcp_skb_pcount(skb);
1617         }
1618         return tso_segs;
1619 }
1620 
1621 
1622 /* Return true if the Nagle test allows this packet to be
1623  * sent now.
1624  */
1625 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1626                                   unsigned int cur_mss, int nonagle)
1627 {
1628         /* Nagle rule does not apply to frames, which sit in the middle of the
1629          * write_queue (they have no chances to get new data).
1630          *
1631          * This is implemented in the callers, where they modify the 'nonagle'
1632          * argument based upon the location of SKB in the send queue.
1633          */
1634         if (nonagle & TCP_NAGLE_PUSH)
1635                 return true;
1636 
1637         /* Don't use the nagle rule for urgent data (or for the final FIN). */
1638         if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1639                 return true;
1640 
1641         if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1642                 return true;
1643 
1644         return false;
1645 }
1646 
1647 /* Does at least the first segment of SKB fit into the send window? */
1648 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1649                              const struct sk_buff *skb,
1650                              unsigned int cur_mss)
1651 {
1652         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1653 
1654         if (skb->len > cur_mss)
1655                 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1656 
1657         return !after(end_seq, tcp_wnd_end(tp));
1658 }
1659 
1660 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1661  * should be put on the wire right now.  If so, it returns the number of
1662  * packets allowed by the congestion window.
1663  */
1664 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1665                                  unsigned int cur_mss, int nonagle)
1666 {
1667         const struct tcp_sock *tp = tcp_sk(sk);
1668         unsigned int cwnd_quota;
1669 
1670         tcp_init_tso_segs(skb, cur_mss);
1671 
1672         if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1673                 return 0;
1674 
1675         cwnd_quota = tcp_cwnd_test(tp, skb);
1676         if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1677                 cwnd_quota = 0;
1678 
1679         return cwnd_quota;
1680 }
1681 
1682 /* Test if sending is allowed right now. */
1683 bool tcp_may_send_now(struct sock *sk)
1684 {
1685         const struct tcp_sock *tp = tcp_sk(sk);
1686         struct sk_buff *skb = tcp_send_head(sk);
1687 
1688         return skb &&
1689                 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1690                              (tcp_skb_is_last(sk, skb) ?
1691                               tp->nonagle : TCP_NAGLE_PUSH));
1692 }
1693 
1694 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1695  * which is put after SKB on the list.  It is very much like
1696  * tcp_fragment() except that it may make several kinds of assumptions
1697  * in order to speed up the splitting operation.  In particular, we
1698  * know that all the data is in scatter-gather pages, and that the
1699  * packet has never been sent out before (and thus is not cloned).
1700  */
1701 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1702                         unsigned int mss_now, gfp_t gfp)
1703 {
1704         struct sk_buff *buff;
1705         int nlen = skb->len - len;
1706         u8 flags;
1707 
1708         /* All of a TSO frame must be composed of paged data.  */
1709         if (skb->len != skb->data_len)
1710                 return tcp_fragment(sk, skb, len, mss_now, gfp);
1711 
1712         buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1713         if (unlikely(!buff))
1714                 return -ENOMEM;
1715 
1716         sk->sk_wmem_queued += buff->truesize;
1717         sk_mem_charge(sk, buff->truesize);
1718         buff->truesize += nlen;
1719         skb->truesize -= nlen;
1720 
1721         /* Correct the sequence numbers. */
1722         TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1723         TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1724         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1725 
1726         /* PSH and FIN should only be set in the second packet. */
1727         flags = TCP_SKB_CB(skb)->tcp_flags;
1728         TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1729         TCP_SKB_CB(buff)->tcp_flags = flags;
1730 
1731         /* This packet was never sent out yet, so no SACK bits. */
1732         TCP_SKB_CB(buff)->sacked = 0;
1733 
1734         buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1735         skb_split(skb, buff, len);
1736         tcp_fragment_tstamp(skb, buff);
1737 
1738         /* Fix up tso_factor for both original and new SKB.  */
1739         tcp_set_skb_tso_segs(skb, mss_now);
1740         tcp_set_skb_tso_segs(buff, mss_now);
1741 
1742         /* Link BUFF into the send queue. */
1743         __skb_header_release(buff);
1744         tcp_insert_write_queue_after(skb, buff, sk);
1745 
1746         return 0;
1747 }
1748 
1749 /* Try to defer sending, if possible, in order to minimize the amount
1750  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1751  *
1752  * This algorithm is from John Heffner.
1753  */
1754 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1755                                  bool *is_cwnd_limited, u32 max_segs)
1756 {
1757         const struct inet_connection_sock *icsk = inet_csk(sk);
1758         u32 age, send_win, cong_win, limit, in_flight;
1759         struct tcp_sock *tp = tcp_sk(sk);
1760         struct skb_mstamp now;
1761         struct sk_buff *head;
1762         int win_divisor;
1763 
1764         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1765                 goto send_now;
1766 
1767         if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1768                 goto send_now;
1769 
1770         /* Avoid bursty behavior by allowing defer
1771          * only if the last write was recent.
1772          */
1773         if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1774                 goto send_now;
1775 
1776         in_flight = tcp_packets_in_flight(tp);
1777 
1778         BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1779 
1780         send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1781 
1782         /* From in_flight test above, we know that cwnd > in_flight.  */
1783         cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1784 
1785         limit = min(send_win, cong_win);
1786 
1787         /* If a full-sized TSO skb can be sent, do it. */
1788         if (limit >= max_segs * tp->mss_cache)
1789                 goto send_now;
1790 
1791         /* Middle in queue won't get any more data, full sendable already? */
1792         if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1793                 goto send_now;
1794 
1795         win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1796         if (win_divisor) {
1797                 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1798 
1799                 /* If at least some fraction of a window is available,
1800                  * just use it.
1801                  */
1802                 chunk /= win_divisor;
1803                 if (limit >= chunk)
1804                         goto send_now;
1805         } else {
1806                 /* Different approach, try not to defer past a single
1807                  * ACK.  Receiver should ACK every other full sized
1808                  * frame, so if we have space for more than 3 frames
1809                  * then send now.
1810                  */
1811                 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1812                         goto send_now;
1813         }
1814 
1815         head = tcp_write_queue_head(sk);
1816         skb_mstamp_get(&now);
1817         age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1818         /* If next ACK is likely to come too late (half srtt), do not defer */
1819         if (age < (tp->srtt_us >> 4))
1820                 goto send_now;
1821 
1822         /* Ok, it looks like it is advisable to defer. */
1823 
1824         if (cong_win < send_win && cong_win <= skb->len)
1825                 *is_cwnd_limited = true;
1826 
1827         return true;
1828 
1829 send_now:
1830         return false;
1831 }
1832 
1833 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1834 {
1835         struct inet_connection_sock *icsk = inet_csk(sk);
1836         struct tcp_sock *tp = tcp_sk(sk);
1837         struct net *net = sock_net(sk);
1838         u32 interval;
1839         s32 delta;
1840 
1841         interval = net->ipv4.sysctl_tcp_probe_interval;
1842         delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1843         if (unlikely(delta >= interval * HZ)) {
1844                 int mss = tcp_current_mss(sk);
1845 
1846                 /* Update current search range */
1847                 icsk->icsk_mtup.probe_size = 0;
1848                 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1849                         sizeof(struct tcphdr) +
1850                         icsk->icsk_af_ops->net_header_len;
1851                 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1852 
1853                 /* Update probe time stamp */
1854                 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1855         }
1856 }
1857 
1858 /* Create a new MTU probe if we are ready.
1859  * MTU probe is regularly attempting to increase the path MTU by
1860  * deliberately sending larger packets.  This discovers routing
1861  * changes resulting in larger path MTUs.
1862  *
1863  * Returns 0 if we should wait to probe (no cwnd available),
1864  *         1 if a probe was sent,
1865  *         -1 otherwise
1866  */
1867 static int tcp_mtu_probe(struct sock *sk)
1868 {
1869         struct tcp_sock *tp = tcp_sk(sk);
1870         struct inet_connection_sock *icsk = inet_csk(sk);
1871         struct sk_buff *skb, *nskb, *next;
1872         struct net *net = sock_net(sk);
1873         int len;
1874         int probe_size;
1875         int size_needed;
1876         int copy;
1877         int mss_now;
1878         int interval;
1879 
1880         /* Not currently probing/verifying,
1881          * not in recovery,
1882          * have enough cwnd, and
1883          * not SACKing (the variable headers throw things off) */
1884         if (!icsk->icsk_mtup.enabled ||
1885             icsk->icsk_mtup.probe_size ||
1886             inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1887             tp->snd_cwnd < 11 ||
1888             tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1889                 return -1;
1890 
1891         /* Use binary search for probe_size between tcp_mss_base,
1892          * and current mss_clamp. if (search_high - search_low)
1893          * smaller than a threshold, backoff from probing.
1894          */
1895         mss_now = tcp_current_mss(sk);
1896         probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1897                                     icsk->icsk_mtup.search_low) >> 1);
1898         size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1899         interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
1900         /* When misfortune happens, we are reprobing actively,
1901          * and then reprobe timer has expired. We stick with current
1902          * probing process by not resetting search range to its orignal.
1903          */
1904         if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
1905                 interval < net->ipv4.sysctl_tcp_probe_threshold) {
1906                 /* Check whether enough time has elaplased for
1907                  * another round of probing.
1908                  */
1909                 tcp_mtu_check_reprobe(sk);
1910                 return -1;
1911         }
1912 
1913         /* Have enough data in the send queue to probe? */
1914         if (tp->write_seq - tp->snd_nxt < size_needed)
1915                 return -1;
1916 
1917         if (tp->snd_wnd < size_needed)
1918                 return -1;
1919         if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1920                 return 0;
1921 
1922         /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1923         if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1924                 if (!tcp_packets_in_flight(tp))
1925                         return -1;
1926                 else
1927                         return 0;
1928         }
1929 
1930         /* We're allowed to probe.  Build it now. */
1931         nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
1932         if (!nskb)
1933                 return -1;
1934         sk->sk_wmem_queued += nskb->truesize;
1935         sk_mem_charge(sk, nskb->truesize);
1936 
1937         skb = tcp_send_head(sk);
1938 
1939         TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1940         TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1941         TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1942         TCP_SKB_CB(nskb)->sacked = 0;
1943         nskb->csum = 0;
1944         nskb->ip_summed = skb->ip_summed;
1945 
1946         tcp_insert_write_queue_before(nskb, skb, sk);
1947 
1948         len = 0;
1949         tcp_for_write_queue_from_safe(skb, next, sk) {
1950                 copy = min_t(int, skb->len, probe_size - len);
1951                 if (nskb->ip_summed)
1952                         skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1953                 else
1954                         nskb->csum = skb_copy_and_csum_bits(skb, 0,
1955                                                             skb_put(nskb, copy),
1956                                                             copy, nskb->csum);
1957 
1958                 if (skb->len <= copy) {
1959                         /* We've eaten all the data from this skb.
1960                          * Throw it away. */
1961                         TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1962                         tcp_unlink_write_queue(skb, sk);
1963                         sk_wmem_free_skb(sk, skb);
1964                 } else {
1965                         TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1966                                                    ~(TCPHDR_FIN|TCPHDR_PSH);
1967                         if (!skb_shinfo(skb)->nr_frags) {
1968                                 skb_pull(skb, copy);
1969                                 if (skb->ip_summed != CHECKSUM_PARTIAL)
1970                                         skb->csum = csum_partial(skb->data,
1971                                                                  skb->len, 0);
1972                         } else {
1973                                 __pskb_trim_head(skb, copy);
1974                                 tcp_set_skb_tso_segs(skb, mss_now);
1975                         }
1976                         TCP_SKB_CB(skb)->seq += copy;
1977                 }
1978 
1979                 len += copy;
1980 
1981                 if (len >= probe_size)
1982                         break;
1983         }
1984         tcp_init_tso_segs(nskb, nskb->len);
1985 
1986         /* We're ready to send.  If this fails, the probe will
1987          * be resegmented into mss-sized pieces by tcp_write_xmit().
1988          */
1989         if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1990                 /* Decrement cwnd here because we are sending
1991                  * effectively two packets. */
1992                 tp->snd_cwnd--;
1993                 tcp_event_new_data_sent(sk, nskb);
1994 
1995                 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1996                 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1997                 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1998 
1999                 return 1;
2000         }
2001 
2002         return -1;
2003 }
2004 
2005 /* This routine writes packets to the network.  It advances the
2006  * send_head.  This happens as incoming acks open up the remote
2007  * window for us.
2008  *
2009  * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2010  * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2011  * account rare use of URG, this is not a big flaw.
2012  *
2013  * Send at most one packet when push_one > 0. Temporarily ignore
2014  * cwnd limit to force at most one packet out when push_one == 2.
2015 
2016  * Returns true, if no segments are in flight and we have queued segments,
2017  * but cannot send anything now because of SWS or another problem.
2018  */
2019 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2020                            int push_one, gfp_t gfp)
2021 {
2022         struct tcp_sock *tp = tcp_sk(sk);
2023         struct sk_buff *skb;
2024         unsigned int tso_segs, sent_pkts;
2025         int cwnd_quota;
2026         int result;
2027         bool is_cwnd_limited = false;
2028         u32 max_segs;
2029 
2030         sent_pkts = 0;
2031 
2032         if (!push_one) {
2033                 /* Do MTU probing. */
2034                 result = tcp_mtu_probe(sk);
2035                 if (!result) {
2036                         return false;
2037                 } else if (result > 0) {
2038                         sent_pkts = 1;
2039                 }
2040         }
2041 
2042         max_segs = tcp_tso_autosize(sk, mss_now);
2043         while ((skb = tcp_send_head(sk))) {
2044                 unsigned int limit;
2045 
2046                 tso_segs = tcp_init_tso_segs(skb, mss_now);
2047                 BUG_ON(!tso_segs);
2048 
2049                 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2050                         /* "skb_mstamp" is used as a start point for the retransmit timer */
2051                         skb_mstamp_get(&skb->skb_mstamp);
2052                         goto repair; /* Skip network transmission */
2053                 }
2054 
2055                 cwnd_quota = tcp_cwnd_test(tp, skb);
2056                 if (!cwnd_quota) {
2057                         if (push_one == 2)
2058                                 /* Force out a loss probe pkt. */
2059                                 cwnd_quota = 1;
2060                         else
2061                                 break;
2062                 }
2063 
2064                 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2065                         break;
2066 
2067                 if (tso_segs == 1) {
2068                         if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2069                                                      (tcp_skb_is_last(sk, skb) ?
2070                                                       nonagle : TCP_NAGLE_PUSH))))
2071                                 break;
2072                 } else {
2073                         if (!push_one &&
2074                             tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2075                                                  max_segs))
2076                                 break;
2077                 }
2078 
2079                 limit = mss_now;
2080                 if (tso_segs > 1 && !tcp_urg_mode(tp))
2081                         limit = tcp_mss_split_point(sk, skb, mss_now,
2082                                                     min_t(unsigned int,
2083                                                           cwnd_quota,
2084                                                           max_segs),
2085                                                     nonagle);
2086 
2087                 if (skb->len > limit &&
2088                     unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2089                         break;
2090 
2091                 /* TCP Small Queues :
2092                  * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2093                  * This allows for :
2094                  *  - better RTT estimation and ACK scheduling
2095                  *  - faster recovery
2096                  *  - high rates
2097                  * Alas, some drivers / subsystems require a fair amount
2098                  * of queued bytes to ensure line rate.
2099                  * One example is wifi aggregation (802.11 AMPDU)
2100                  */
2101                 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2102                 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2103 
2104                 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2105                         set_bit(TSQ_THROTTLED, &tp->tsq_flags);
2106                         /* It is possible TX completion already happened
2107                          * before we set TSQ_THROTTLED, so we must
2108                          * test again the condition.
2109                          */
2110                         smp_mb__after_atomic();
2111                         if (atomic_read(&sk->sk_wmem_alloc) > limit)
2112                                 break;
2113                 }
2114 
2115                 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2116                         break;
2117 
2118 repair:
2119                 /* Advance the send_head.  This one is sent out.
2120                  * This call will increment packets_out.
2121                  */
2122                 tcp_event_new_data_sent(sk, skb);
2123 
2124                 tcp_minshall_update(tp, mss_now, skb);
2125                 sent_pkts += tcp_skb_pcount(skb);
2126 
2127                 if (push_one)
2128                         break;
2129         }
2130 
2131         if (likely(sent_pkts)) {
2132                 if (tcp_in_cwnd_reduction(sk))
2133                         tp->prr_out += sent_pkts;
2134 
2135                 /* Send one loss probe per tail loss episode. */
2136                 if (push_one != 2)
2137                         tcp_schedule_loss_probe(sk);
2138                 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2139                 tcp_cwnd_validate(sk, is_cwnd_limited);
2140                 return false;
2141         }
2142         return !tp->packets_out && tcp_send_head(sk);
2143 }
2144 
2145 bool tcp_schedule_loss_probe(struct sock *sk)
2146 {
2147         struct inet_connection_sock *icsk = inet_csk(sk);
2148         struct tcp_sock *tp = tcp_sk(sk);
2149         u32 timeout, tlp_time_stamp, rto_time_stamp;
2150         u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2151 
2152         if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2153                 return false;
2154         /* No consecutive loss probes. */
2155         if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2156                 tcp_rearm_rto(sk);
2157                 return false;
2158         }
2159         /* Don't do any loss probe on a Fast Open connection before 3WHS
2160          * finishes.
2161          */
2162         if (tp->fastopen_rsk)
2163                 return false;
2164 
2165         /* TLP is only scheduled when next timer event is RTO. */
2166         if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2167                 return false;
2168 
2169         /* Schedule a loss probe in 2*RTT for SACK capable connections
2170          * in Open state, that are either limited by cwnd or application.
2171          */
2172         if (sysctl_tcp_early_retrans < 3 || !tp->packets_out ||
2173             !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2174                 return false;
2175 
2176         if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2177              tcp_send_head(sk))
2178                 return false;
2179 
2180         /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2181          * for delayed ack when there's one outstanding packet. If no RTT
2182          * sample is available then probe after TCP_TIMEOUT_INIT.
2183          */
2184         timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
2185         if (tp->packets_out == 1)
2186                 timeout = max_t(u32, timeout,
2187                                 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2188         timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2189 
2190         /* If RTO is shorter, just schedule TLP in its place. */
2191         tlp_time_stamp = tcp_time_stamp + timeout;
2192         rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2193         if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2194                 s32 delta = rto_time_stamp - tcp_time_stamp;
2195                 if (delta > 0)
2196                         timeout = delta;
2197         }
2198 
2199         inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2200                                   TCP_RTO_MAX);
2201         return true;
2202 }
2203 
2204 /* Thanks to skb fast clones, we can detect if a prior transmit of
2205  * a packet is still in a qdisc or driver queue.
2206  * In this case, there is very little point doing a retransmit !
2207  * Note: This is called from BH context only.
2208  */
2209 static bool skb_still_in_host_queue(const struct sock *sk,
2210                                     const struct sk_buff *skb)
2211 {
2212         if (unlikely(skb_fclone_busy(sk, skb))) {
2213                 NET_INC_STATS_BH(sock_net(sk),
2214                                  LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2215                 return true;
2216         }
2217         return false;
2218 }
2219 
2220 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2221  * retransmit the last segment.
2222  */
2223 void tcp_send_loss_probe(struct sock *sk)
2224 {
2225         struct tcp_sock *tp = tcp_sk(sk);
2226         struct sk_buff *skb;
2227         int pcount;
2228         int mss = tcp_current_mss(sk);
2229 
2230         skb = tcp_send_head(sk);
2231         if (skb) {
2232                 if (tcp_snd_wnd_test(tp, skb, mss)) {
2233                         pcount = tp->packets_out;
2234                         tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2235                         if (tp->packets_out > pcount)
2236                                 goto probe_sent;
2237                         goto rearm_timer;
2238                 }
2239                 skb = tcp_write_queue_prev(sk, skb);
2240         } else {
2241                 skb = tcp_write_queue_tail(sk);
2242         }
2243 
2244         /* At most one outstanding TLP retransmission. */
2245         if (tp->tlp_high_seq)
2246                 goto rearm_timer;
2247 
2248         /* Retransmit last segment. */
2249         if (WARN_ON(!skb))
2250                 goto rearm_timer;
2251 
2252         if (skb_still_in_host_queue(sk, skb))
2253                 goto rearm_timer;
2254 
2255         pcount = tcp_skb_pcount(skb);
2256         if (WARN_ON(!pcount))
2257                 goto rearm_timer;
2258 
2259         if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2260                 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2261                                           GFP_ATOMIC)))
2262                         goto rearm_timer;
2263                 skb = tcp_write_queue_next(sk, skb);
2264         }
2265 
2266         if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2267                 goto rearm_timer;
2268 
2269         if (__tcp_retransmit_skb(sk, skb))
2270                 goto rearm_timer;
2271 
2272         /* Record snd_nxt for loss detection. */
2273         tp->tlp_high_seq = tp->snd_nxt;
2274 
2275 probe_sent:
2276         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2277         /* Reset s.t. tcp_rearm_rto will restart timer from now */
2278         inet_csk(sk)->icsk_pending = 0;
2279 rearm_timer:
2280         tcp_rearm_rto(sk);
2281 }
2282 
2283 /* Push out any pending frames which were held back due to
2284  * TCP_CORK or attempt at coalescing tiny packets.
2285  * The socket must be locked by the caller.
2286  */
2287 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2288                                int nonagle)
2289 {
2290         /* If we are closed, the bytes will have to remain here.
2291          * In time closedown will finish, we empty the write queue and
2292          * all will be happy.
2293          */
2294         if (unlikely(sk->sk_state == TCP_CLOSE))
2295                 return;
2296 
2297         if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2298                            sk_gfp_mask(sk, GFP_ATOMIC)))
2299                 tcp_check_probe_timer(sk);
2300 }
2301 
2302 /* Send _single_ skb sitting at the send head. This function requires
2303  * true push pending frames to setup probe timer etc.
2304  */
2305 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2306 {
2307         struct sk_buff *skb = tcp_send_head(sk);
2308 
2309         BUG_ON(!skb || skb->len < mss_now);
2310 
2311         tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2312 }
2313 
2314 /* This function returns the amount that we can raise the
2315  * usable window based on the following constraints
2316  *
2317  * 1. The window can never be shrunk once it is offered (RFC 793)
2318  * 2. We limit memory per socket
2319  *
2320  * RFC 1122:
2321  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2322  *  RECV.NEXT + RCV.WIN fixed until:
2323  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2324  *
2325  * i.e. don't raise the right edge of the window until you can raise
2326  * it at least MSS bytes.
2327  *
2328  * Unfortunately, the recommended algorithm breaks header prediction,
2329  * since header prediction assumes th->window stays fixed.
2330  *
2331  * Strictly speaking, keeping th->window fixed violates the receiver
2332  * side SWS prevention criteria. The problem is that under this rule
2333  * a stream of single byte packets will cause the right side of the
2334  * window to always advance by a single byte.
2335  *
2336  * Of course, if the sender implements sender side SWS prevention
2337  * then this will not be a problem.
2338  *
2339  * BSD seems to make the following compromise:
2340  *
2341  *      If the free space is less than the 1/4 of the maximum
2342  *      space available and the free space is less than 1/2 mss,
2343  *      then set the window to 0.
2344  *      [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2345  *      Otherwise, just prevent the window from shrinking
2346  *      and from being larger than the largest representable value.
2347  *
2348  * This prevents incremental opening of the window in the regime
2349  * where TCP is limited by the speed of the reader side taking
2350  * data out of the TCP receive queue. It does nothing about
2351  * those cases where the window is constrained on the sender side
2352  * because the pipeline is full.
2353  *
2354  * BSD also seems to "accidentally" limit itself to windows that are a
2355  * multiple of MSS, at least until the free space gets quite small.
2356  * This would appear to be a side effect of the mbuf implementation.
2357  * Combining these two algorithms results in the observed behavior
2358  * of having a fixed window size at almost all times.
2359  *
2360  * Below we obtain similar behavior by forcing the offered window to
2361  * a multiple of the mss when it is feasible to do so.
2362  *
2363  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2364  * Regular options like TIMESTAMP are taken into account.
2365  */
2366 u32 __tcp_select_window(struct sock *sk)
2367 {
2368         struct inet_connection_sock *icsk = inet_csk(sk);
2369         struct tcp_sock *tp = tcp_sk(sk);
2370         /* MSS for the peer's data.  Previous versions used mss_clamp
2371          * here.  I don't know if the value based on our guesses
2372          * of peer's MSS is better for the performance.  It's more correct
2373          * but may be worse for the performance because of rcv_mss
2374          * fluctuations.  --SAW  1998/11/1
2375          */
2376         int mss = icsk->icsk_ack.rcv_mss;
2377         int free_space = tcp_space(sk);
2378         int allowed_space = tcp_full_space(sk);
2379         int full_space = min_t(int, tp->window_clamp, allowed_space);
2380         int window;
2381 
2382         if (mss > full_space)
2383                 mss = full_space;
2384 
2385         if (free_space < (full_space >> 1)) {
2386                 icsk->icsk_ack.quick = 0;
2387 
2388                 if (tcp_under_memory_pressure(sk))
2389                         tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2390                                                4U * tp->advmss);
2391 
2392                 /* free_space might become our new window, make sure we don't
2393                  * increase it due to wscale.
2394                  */
2395                 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2396 
2397                 /* if free space is less than mss estimate, or is below 1/16th
2398                  * of the maximum allowed, try to move to zero-window, else
2399                  * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2400                  * new incoming data is dropped due to memory limits.
2401                  * With large window, mss test triggers way too late in order
2402                  * to announce zero window in time before rmem limit kicks in.
2403                  */
2404                 if (free_space < (allowed_space >> 4) || free_space < mss)
2405                         return 0;
2406         }
2407 
2408         if (free_space > tp->rcv_ssthresh)
2409                 free_space = tp->rcv_ssthresh;
2410 
2411         /* Don't do rounding if we are using window scaling, since the
2412          * scaled window will not line up with the MSS boundary anyway.
2413          */
2414         window = tp->rcv_wnd;
2415         if (tp->rx_opt.rcv_wscale) {
2416                 window = free_space;
2417 
2418                 /* Advertise enough space so that it won't get scaled away.
2419                  * Import case: prevent zero window announcement if
2420                  * 1<<rcv_wscale > mss.
2421                  */
2422                 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2423                         window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2424                                   << tp->rx_opt.rcv_wscale);
2425         } else {
2426                 /* Get the largest window that is a nice multiple of mss.
2427                  * Window clamp already applied above.
2428                  * If our current window offering is within 1 mss of the
2429                  * free space we just keep it. This prevents the divide
2430                  * and multiply from happening most of the time.
2431                  * We also don't do any window rounding when the free space
2432                  * is too small.
2433                  */
2434                 if (window <= free_space - mss || window > free_space)
2435                         window = (free_space / mss) * mss;
2436                 else if (mss == full_space &&
2437                          free_space > window + (full_space >> 1))
2438                         window = free_space;
2439         }
2440 
2441         return window;
2442 }
2443 
2444 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2445                              const struct sk_buff *next_skb)
2446 {
2447         const struct skb_shared_info *next_shinfo = skb_shinfo(next_skb);
2448         u8 tsflags = next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2449 
2450         if (unlikely(tsflags)) {
2451                 struct skb_shared_info *shinfo = skb_shinfo(skb);
2452 
2453                 shinfo->tx_flags |= tsflags;
2454                 shinfo->tskey = next_shinfo->tskey;
2455         }
2456 }
2457 
2458 /* Collapses two adjacent SKB's during retransmission. */
2459 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2460 {
2461         struct tcp_sock *tp = tcp_sk(sk);
2462         struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2463         int skb_size, next_skb_size;
2464 
2465         skb_size = skb->len;
2466         next_skb_size = next_skb->len;
2467 
2468         BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2469 
2470         tcp_highest_sack_combine(sk, next_skb, skb);
2471 
2472         tcp_unlink_write_queue(next_skb, sk);
2473 
2474         skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2475                                   next_skb_size);
2476 
2477         if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2478                 skb->ip_summed = CHECKSUM_PARTIAL;
2479 
2480         if (skb->ip_summed != CHECKSUM_PARTIAL)
2481                 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2482 
2483         /* Update sequence range on original skb. */
2484         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2485 
2486         /* Merge over control information. This moves PSH/FIN etc. over */
2487         TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2488 
2489         /* All done, get rid of second SKB and account for it so
2490          * packet counting does not break.
2491          */
2492         TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2493 
2494         /* changed transmit queue under us so clear hints */
2495         tcp_clear_retrans_hints_partial(tp);
2496         if (next_skb == tp->retransmit_skb_hint)
2497                 tp->retransmit_skb_hint = skb;
2498 
2499         tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2500 
2501         tcp_skb_collapse_tstamp(skb, next_skb);
2502 
2503         sk_wmem_free_skb(sk, next_skb);
2504 }
2505 
2506 /* Check if coalescing SKBs is legal. */
2507 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2508 {
2509         if (tcp_skb_pcount(skb) > 1)
2510                 return false;
2511         /* TODO: SACK collapsing could be used to remove this condition */
2512         if (skb_shinfo(skb)->nr_frags != 0)
2513                 return false;
2514         if (skb_cloned(skb))
2515                 return false;
2516         if (skb == tcp_send_head(sk))
2517                 return false;
2518         /* Some heurestics for collapsing over SACK'd could be invented */
2519         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2520                 return false;
2521 
2522         return true;
2523 }
2524 
2525 /* Collapse packets in the retransmit queue to make to create
2526  * less packets on the wire. This is only done on retransmission.
2527  */
2528 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2529                                      int space)
2530 {
2531         struct tcp_sock *tp = tcp_sk(sk);
2532         struct sk_buff *skb = to, *tmp;
2533         bool first = true;
2534 
2535         if (!sysctl_tcp_retrans_collapse)
2536                 return;
2537         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2538                 return;
2539 
2540         tcp_for_write_queue_from_safe(skb, tmp, sk) {
2541                 if (!tcp_can_collapse(sk, skb))
2542                         break;
2543 
2544                 space -= skb->len;
2545 
2546                 if (first) {
2547                         first = false;
2548                         continue;
2549                 }
2550 
2551                 if (space < 0)
2552                         break;
2553                 /* Punt if not enough space exists in the first SKB for
2554                  * the data in the second
2555                  */
2556                 if (skb->len > skb_availroom(to))
2557                         break;
2558 
2559                 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2560                         break;
2561 
2562                 tcp_collapse_retrans(sk, to);
2563         }
2564 }
2565 
2566 /* This retransmits one SKB.  Policy decisions and retransmit queue
2567  * state updates are done by the caller.  Returns non-zero if an
2568  * error occurred which prevented the send.
2569  */
2570 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2571 {
2572         struct tcp_sock *tp = tcp_sk(sk);
2573         struct inet_connection_sock *icsk = inet_csk(sk);
2574         unsigned int cur_mss;
2575         int err;
2576 
2577         /* Inconslusive MTU probe */
2578         if (icsk->icsk_mtup.probe_size) {
2579                 icsk->icsk_mtup.probe_size = 0;
2580         }
2581 
2582         /* Do not sent more than we queued. 1/4 is reserved for possible
2583          * copying overhead: fragmentation, tunneling, mangling etc.
2584          */
2585         if (atomic_read(&sk->sk_wmem_alloc) >
2586             min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2587                 return -EAGAIN;
2588 
2589         if (skb_still_in_host_queue(sk, skb))
2590                 return -EBUSY;
2591 
2592         if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2593                 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2594                         BUG();
2595                 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2596                         return -ENOMEM;
2597         }
2598 
2599         if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2600                 return -EHOSTUNREACH; /* Routing failure or similar. */
2601 
2602         cur_mss = tcp_current_mss(sk);
2603 
2604         /* If receiver has shrunk his window, and skb is out of
2605          * new window, do not retransmit it. The exception is the
2606          * case, when window is shrunk to zero. In this case
2607          * our retransmit serves as a zero window probe.
2608          */
2609         if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2610             TCP_SKB_CB(skb)->seq != tp->snd_una)
2611                 return -EAGAIN;
2612 
2613         if (skb->len > cur_mss) {
2614                 if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC))
2615                         return -ENOMEM; /* We'll try again later. */
2616         } else {
2617                 int oldpcount = tcp_skb_pcount(skb);
2618 
2619                 if (unlikely(oldpcount > 1)) {
2620                         if (skb_unclone(skb, GFP_ATOMIC))
2621                                 return -ENOMEM;
2622                         tcp_init_tso_segs(skb, cur_mss);
2623                         tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2624                 }
2625         }
2626 
2627         /* RFC3168, section 6.1.1.1. ECN fallback */
2628         if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2629                 tcp_ecn_clear_syn(sk, skb);
2630 
2631         tcp_retrans_try_collapse(sk, skb, cur_mss);
2632 
2633         /* Make a copy, if the first transmission SKB clone we made
2634          * is still in somebody's hands, else make a clone.
2635          */
2636 
2637         /* make sure skb->data is aligned on arches that require it
2638          * and check if ack-trimming & collapsing extended the headroom
2639          * beyond what csum_start can cover.
2640          */
2641         if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2642                      skb_headroom(skb) >= 0xFFFF)) {
2643                 struct sk_buff *nskb;
2644 
2645                 skb_mstamp_get(&skb->skb_mstamp);
2646                 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2647                 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2648                              -ENOBUFS;
2649         } else {
2650                 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2651         }
2652 
2653         if (likely(!err)) {
2654                 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2655                 /* Update global TCP statistics. */
2656                 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2657                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2658                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2659                 tp->total_retrans++;
2660         }
2661         return err;
2662 }
2663 
2664 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2665 {
2666         struct tcp_sock *tp = tcp_sk(sk);
2667         int err = __tcp_retransmit_skb(sk, skb);
2668 
2669         if (err == 0) {
2670 #if FASTRETRANS_DEBUG > 0
2671                 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2672                         net_dbg_ratelimited("retrans_out leaked\n");
2673                 }
2674 #endif
2675                 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2676                 tp->retrans_out += tcp_skb_pcount(skb);
2677 
2678                 /* Save stamp of the first retransmit. */
2679                 if (!tp->retrans_stamp)
2680                         tp->retrans_stamp = tcp_skb_timestamp(skb);
2681 
2682         } else if (err != -EBUSY) {
2683                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2684         }
2685 
2686         if (tp->undo_retrans < 0)
2687                 tp->undo_retrans = 0;
2688         tp->undo_retrans += tcp_skb_pcount(skb);
2689         return err;
2690 }
2691 
2692 /* Check if we forward retransmits are possible in the current
2693  * window/congestion state.
2694  */
2695 static bool tcp_can_forward_retransmit(struct sock *sk)
2696 {
2697         const struct inet_connection_sock *icsk = inet_csk(sk);
2698         const struct tcp_sock *tp = tcp_sk(sk);
2699 
2700         /* Forward retransmissions are possible only during Recovery. */
2701         if (icsk->icsk_ca_state != TCP_CA_Recovery)
2702                 return false;
2703 
2704         /* No forward retransmissions in Reno are possible. */
2705         if (tcp_is_reno(tp))
2706                 return false;
2707 
2708         /* Yeah, we have to make difficult choice between forward transmission
2709          * and retransmission... Both ways have their merits...
2710          *
2711          * For now we do not retransmit anything, while we have some new
2712          * segments to send. In the other cases, follow rule 3 for
2713          * NextSeg() specified in RFC3517.
2714          */
2715 
2716         if (tcp_may_send_now(sk))
2717                 return false;
2718 
2719         return true;
2720 }
2721 
2722 /* This gets called after a retransmit timeout, and the initially
2723  * retransmitted data is acknowledged.  It tries to continue
2724  * resending the rest of the retransmit queue, until either
2725  * we've sent it all or the congestion window limit is reached.
2726  * If doing SACK, the first ACK which comes back for a timeout
2727  * based retransmit packet might feed us FACK information again.
2728  * If so, we use it to avoid unnecessarily retransmissions.
2729  */
2730 void tcp_xmit_retransmit_queue(struct sock *sk)
2731 {
2732         const struct inet_connection_sock *icsk = inet_csk(sk);
2733         struct tcp_sock *tp = tcp_sk(sk);
2734         struct sk_buff *skb;
2735         struct sk_buff *hole = NULL;
2736         u32 last_lost;
2737         int mib_idx;
2738         int fwd_rexmitting = 0;
2739 
2740         if (!tp->packets_out)
2741                 return;
2742 
2743         if (!tp->lost_out)
2744                 tp->retransmit_high = tp->snd_una;
2745 
2746         if (tp->retransmit_skb_hint) {
2747                 skb = tp->retransmit_skb_hint;
2748                 last_lost = TCP_SKB_CB(skb)->end_seq;
2749                 if (after(last_lost, tp->retransmit_high))
2750                         last_lost = tp->retransmit_high;
2751         } else {
2752                 skb = tcp_write_queue_head(sk);
2753                 last_lost = tp->snd_una;
2754         }
2755 
2756         tcp_for_write_queue_from(skb, sk) {
2757                 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2758 
2759                 if (skb == tcp_send_head(sk))
2760                         break;
2761                 /* we could do better than to assign each time */
2762                 if (!hole)
2763                         tp->retransmit_skb_hint = skb;
2764 
2765                 /* Assume this retransmit will generate
2766                  * only one packet for congestion window
2767                  * calculation purposes.  This works because
2768                  * tcp_retransmit_skb() will chop up the
2769                  * packet to be MSS sized and all the
2770                  * packet counting works out.
2771                  */
2772                 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2773                         return;
2774 
2775                 if (fwd_rexmitting) {
2776 begin_fwd:
2777                         if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2778                                 break;
2779                         mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2780 
2781                 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2782                         tp->retransmit_high = last_lost;
2783                         if (!tcp_can_forward_retransmit(sk))
2784                                 break;
2785                         /* Backtrack if necessary to non-L'ed skb */
2786                         if (hole) {
2787                                 skb = hole;
2788                                 hole = NULL;
2789                         }
2790                         fwd_rexmitting = 1;
2791                         goto begin_fwd;
2792 
2793                 } else if (!(sacked & TCPCB_LOST)) {
2794                         if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2795                                 hole = skb;
2796                         continue;
2797 
2798                 } else {
2799                         last_lost = TCP_SKB_CB(skb)->end_seq;
2800                         if (icsk->icsk_ca_state != TCP_CA_Loss)
2801                                 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2802                         else
2803                                 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2804                 }
2805 
2806                 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2807                         continue;
2808 
2809                 if (tcp_retransmit_skb(sk, skb))
2810                         return;
2811 
2812                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2813 
2814                 if (tcp_in_cwnd_reduction(sk))
2815                         tp->prr_out += tcp_skb_pcount(skb);
2816 
2817                 if (skb == tcp_write_queue_head(sk))
2818                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2819                                                   inet_csk(sk)->icsk_rto,
2820                                                   TCP_RTO_MAX);
2821         }
2822 }
2823 
2824 /* We allow to exceed memory limits for FIN packets to expedite
2825  * connection tear down and (memory) recovery.
2826  * Otherwise tcp_send_fin() could be tempted to either delay FIN
2827  * or even be forced to close flow without any FIN.
2828  * In general, we want to allow one skb per socket to avoid hangs
2829  * with edge trigger epoll()
2830  */
2831 void sk_forced_mem_schedule(struct sock *sk, int size)
2832 {
2833         int amt;
2834 
2835         if (size <= sk->sk_forward_alloc)
2836                 return;
2837         amt = sk_mem_pages(size);
2838         sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2839         sk_memory_allocated_add(sk, amt);
2840 
2841         if (mem_cgroup_sockets_enabled && sk->sk_memcg)
2842                 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
2843 }
2844 
2845 /* Send a FIN. The caller locks the socket for us.
2846  * We should try to send a FIN packet really hard, but eventually give up.
2847  */
2848 void tcp_send_fin(struct sock *sk)
2849 {
2850         struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
2851         struct tcp_sock *tp = tcp_sk(sk);
2852 
2853         /* Optimization, tack on the FIN if we have one skb in write queue and
2854          * this skb was not yet sent, or we are under memory pressure.
2855          * Note: in the latter case, FIN packet will be sent after a timeout,
2856          * as TCP stack thinks it has already been transmitted.
2857          */
2858         if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
2859 coalesce:
2860                 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2861                 TCP_SKB_CB(tskb)->end_seq++;
2862                 tp->write_seq++;
2863                 if (!tcp_send_head(sk)) {
2864                         /* This means tskb was already sent.
2865                          * Pretend we included the FIN on previous transmit.
2866                          * We need to set tp->snd_nxt to the value it would have
2867                          * if FIN had been sent. This is because retransmit path
2868                          * does not change tp->snd_nxt.
2869                          */
2870                         tp->snd_nxt++;
2871                         return;
2872                 }
2873         } else {
2874                 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2875                 if (unlikely(!skb)) {
2876                         if (tskb)
2877                                 goto coalesce;
2878                         return;
2879                 }
2880                 skb_reserve(skb, MAX_TCP_HEADER);
2881                 sk_forced_mem_schedule(sk, skb->truesize);
2882                 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2883                 tcp_init_nondata_skb(skb, tp->write_seq,
2884                                      TCPHDR_ACK | TCPHDR_FIN);
2885                 tcp_queue_skb(sk, skb);
2886         }
2887         __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
2888 }
2889 
2890 /* We get here when a process closes a file descriptor (either due to
2891  * an explicit close() or as a byproduct of exit()'ing) and there
2892  * was unread data in the receive queue.  This behavior is recommended
2893  * by RFC 2525, section 2.17.  -DaveM
2894  */
2895 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2896 {
2897         struct sk_buff *skb;
2898 
2899         /* NOTE: No TCP options attached and we never retransmit this. */
2900         skb = alloc_skb(MAX_TCP_HEADER, priority);
2901         if (!skb) {
2902                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2903                 return;
2904         }
2905 
2906         /* Reserve space for headers and prepare control bits. */
2907         skb_reserve(skb, MAX_TCP_HEADER);
2908         tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2909                              TCPHDR_ACK | TCPHDR_RST);
2910         skb_mstamp_get(&skb->skb_mstamp);
2911         /* Send it off. */
2912         if (tcp_transmit_skb(sk, skb, 0, priority))
2913                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2914 
2915         TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2916 }
2917 
2918 /* Send a crossed SYN-ACK during socket establishment.
2919  * WARNING: This routine must only be called when we have already sent
2920  * a SYN packet that crossed the incoming SYN that caused this routine
2921  * to get called. If this assumption fails then the initial rcv_wnd
2922  * and rcv_wscale values will not be correct.
2923  */
2924 int tcp_send_synack(struct sock *sk)
2925 {
2926         struct sk_buff *skb;
2927 
2928         skb = tcp_write_queue_head(sk);
2929         if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2930                 pr_debug("%s: wrong queue state\n", __func__);
2931                 return -EFAULT;
2932         }
2933         if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2934                 if (skb_cloned(skb)) {
2935                         struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2936                         if (!nskb)
2937                                 return -ENOMEM;
2938                         tcp_unlink_write_queue(skb, sk);
2939                         __skb_header_release(nskb);
2940                         __tcp_add_write_queue_head(sk, nskb);
2941                         sk_wmem_free_skb(sk, skb);
2942                         sk->sk_wmem_queued += nskb->truesize;
2943                         sk_mem_charge(sk, nskb->truesize);
2944                         skb = nskb;
2945                 }
2946 
2947                 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2948                 tcp_ecn_send_synack(sk, skb);
2949         }
2950         return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2951 }
2952 
2953 /**
2954  * tcp_make_synack - Prepare a SYN-ACK.
2955  * sk: listener socket
2956  * dst: dst entry attached to the SYNACK
2957  * req: request_sock pointer
2958  *
2959  * Allocate one skb and build a SYNACK packet.
2960  * @dst is consumed : Caller should not use it again.
2961  */
2962 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
2963                                 struct request_sock *req,
2964                                 struct tcp_fastopen_cookie *foc,
2965                                 bool attach_req)
2966 {
2967         struct inet_request_sock *ireq = inet_rsk(req);
2968         const struct tcp_sock *tp = tcp_sk(sk);
2969         struct tcp_md5sig_key *md5 = NULL;
2970         struct tcp_out_options opts;
2971         struct sk_buff *skb;
2972         int tcp_header_size;
2973         struct tcphdr *th;
2974         u16 user_mss;
2975         int mss;
2976 
2977         skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2978         if (unlikely(!skb)) {
2979                 dst_release(dst);
2980                 return NULL;
2981         }
2982         /* Reserve space for headers. */
2983         skb_reserve(skb, MAX_TCP_HEADER);
2984 
2985         if (attach_req) {
2986                 skb_set_owner_w(skb, req_to_sk(req));
2987         } else {
2988                 /* sk is a const pointer, because we want to express multiple
2989                  * cpu might call us concurrently.
2990                  * sk->sk_wmem_alloc in an atomic, we can promote to rw.
2991                  */
2992                 skb_set_owner_w(skb, (struct sock *)sk);
2993         }
2994         skb_dst_set(skb, dst);
2995 
2996         mss = dst_metric_advmss(dst);
2997         user_mss = READ_ONCE(tp->rx_opt.user_mss);
2998         if (user_mss && user_mss < mss)
2999                 mss = user_mss;
3000 
3001         memset(&opts, 0, sizeof(opts));
3002 #ifdef CONFIG_SYN_COOKIES
3003         if (unlikely(req->cookie_ts))
3004                 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
3005         else
3006 #endif
3007         skb_mstamp_get(&skb->skb_mstamp);
3008 
3009 #ifdef CONFIG_TCP_MD5SIG
3010         rcu_read_lock();
3011         md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3012 #endif
3013         skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3014         tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
3015                           sizeof(*th);
3016 
3017         skb_push(skb, tcp_header_size);
3018         skb_reset_transport_header(skb);
3019 
3020         th = tcp_hdr(skb);
3021         memset(th, 0, sizeof(struct tcphdr));
3022         th->syn = 1;
3023         th->ack = 1;
3024         tcp_ecn_make_synack(req, th);
3025         th->source = htons(ireq->ir_num);
3026         th->dest = ireq->ir_rmt_port;
3027         /* Setting of flags are superfluous here for callers (and ECE is
3028          * not even correctly set)
3029          */
3030         tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
3031                              TCPHDR_SYN | TCPHDR_ACK);
3032 
3033         th->seq = htonl(TCP_SKB_CB(skb)->seq);
3034         /* XXX data is queued and acked as is. No buffer/window check */
3035         th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3036 
3037         /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3038         th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3039         tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3040         th->doff = (tcp_header_size >> 2);
3041         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
3042 
3043 #ifdef CONFIG_TCP_MD5SIG
3044         /* Okay, we have all we need - do the md5 hash if needed */
3045         if (md5)
3046                 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3047                                                md5, req_to_sk(req), skb);
3048         rcu_read_unlock();
3049 #endif
3050 
3051         /* Do not fool tcpdump (if any), clean our debris */
3052         skb->tstamp.tv64 = 0;
3053         return skb;
3054 }
3055 EXPORT_SYMBOL(tcp_make_synack);
3056 
3057 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3058 {
3059         struct inet_connection_sock *icsk = inet_csk(sk);
3060         const struct tcp_congestion_ops *ca;
3061         u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3062 
3063         if (ca_key == TCP_CA_UNSPEC)
3064                 return;
3065 
3066         rcu_read_lock();
3067         ca = tcp_ca_find_key(ca_key);
3068         if (likely(ca && try_module_get(ca->owner))) {
3069                 module_put(icsk->icsk_ca_ops->owner);
3070                 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3071                 icsk->icsk_ca_ops = ca;
3072         }
3073         rcu_read_unlock();
3074 }
3075 
3076 /* Do all connect socket setups that can be done AF independent. */
3077 static void tcp_connect_init(struct sock *sk)
3078 {
3079         const struct dst_entry *dst = __sk_dst_get(sk);
3080         struct tcp_sock *tp = tcp_sk(sk);
3081         __u8 rcv_wscale;
3082 
3083         /* We'll fix this up when we get a response from the other end.
3084          * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3085          */
3086         tp->tcp_header_len = sizeof(struct tcphdr) +
3087                 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3088 
3089 #ifdef CONFIG_TCP_MD5SIG
3090         if (tp->af_specific->md5_lookup(sk, sk))
3091                 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3092 #endif
3093 
3094         /* If user gave his TCP_MAXSEG, record it to clamp */
3095         if (tp->rx_opt.user_mss)
3096                 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3097         tp->max_window = 0;
3098         tcp_mtup_init(sk);
3099         tcp_sync_mss(sk, dst_mtu(dst));
3100 
3101         tcp_ca_dst_init(sk, dst);
3102 
3103         if (!tp->window_clamp)
3104                 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3105         tp->advmss = dst_metric_advmss(dst);
3106         if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3107                 tp->advmss = tp->rx_opt.user_mss;
3108 
3109         tcp_initialize_rcv_mss(sk);
3110 
3111         /* limit the window selection if the user enforce a smaller rx buffer */
3112         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3113             (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3114                 tp->window_clamp = tcp_full_space(sk);
3115 
3116         tcp_select_initial_window(tcp_full_space(sk),
3117                                   tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3118                                   &tp->rcv_wnd,
3119                                   &tp->window_clamp,
3120                                   sysctl_tcp_window_scaling,
3121                                   &rcv_wscale,
3122                                   dst_metric(dst, RTAX_INITRWND));
3123 
3124         tp->rx_opt.rcv_wscale = rcv_wscale;
3125         tp->rcv_ssthresh = tp->rcv_wnd;
3126 
3127         sk->sk_err = 0;
3128         sock_reset_flag(sk, SOCK_DONE);
3129         tp->snd_wnd = 0;
3130         tcp_init_wl(tp, 0);
3131         tp->snd_una = tp->write_seq;
3132         tp->snd_sml = tp->write_seq;
3133         tp->snd_up = tp->write_seq;
3134         tp->snd_nxt = tp->write_seq;
3135 
3136         if (likely(!tp->repair))
3137                 tp->rcv_nxt = 0;
3138         else
3139                 tp->rcv_tstamp = tcp_time_stamp;
3140         tp->rcv_wup = tp->rcv_nxt;
3141         tp->copied_seq = tp->rcv_nxt;
3142 
3143         inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3144         inet_csk(sk)->icsk_retransmits = 0;
3145         tcp_clear_retrans(tp);
3146 }
3147 
3148 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3149 {
3150         struct tcp_sock *tp = tcp_sk(sk);
3151         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3152 
3153         tcb->end_seq += skb->len;
3154         __skb_header_release(skb);
3155         __tcp_add_write_queue_tail(sk, skb);
3156         sk->sk_wmem_queued += skb->truesize;
3157         sk_mem_charge(sk, skb->truesize);
3158         tp->write_seq = tcb->end_seq;
3159         tp->packets_out += tcp_skb_pcount(skb);
3160 }
3161 
3162 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3163  * queue a data-only packet after the regular SYN, such that regular SYNs
3164  * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3165  * only the SYN sequence, the data are retransmitted in the first ACK.
3166  * If cookie is not cached or other error occurs, falls back to send a
3167  * regular SYN with Fast Open cookie request option.
3168  */
3169 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3170 {
3171         struct tcp_sock *tp = tcp_sk(sk);
3172         struct tcp_fastopen_request *fo = tp->fastopen_req;
3173         int syn_loss = 0, space, err = 0;
3174         unsigned long last_syn_loss = 0;
3175         struct sk_buff *syn_data;
3176 
3177         tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
3178         tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3179                                &syn_loss, &last_syn_loss);
3180         /* Recurring FO SYN losses: revert to regular handshake temporarily */
3181         if (syn_loss > 1 &&
3182             time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3183                 fo->cookie.len = -1;
3184                 goto fallback;
3185         }
3186 
3187         if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3188                 fo->cookie.len = -1;
3189         else if (fo->cookie.len <= 0)
3190                 goto fallback;
3191 
3192         /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3193          * user-MSS. Reserve maximum option space for middleboxes that add
3194          * private TCP options. The cost is reduced data space in SYN :(
3195          */
3196         if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3197                 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3198         space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3199                 MAX_TCP_OPTION_SPACE;
3200 
3201         space = min_t(size_t, space, fo->size);
3202 
3203         /* limit to order-0 allocations */
3204         space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3205 
3206         syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3207         if (!syn_data)
3208                 goto fallback;
3209         syn_data->ip_summed = CHECKSUM_PARTIAL;
3210         memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3211         if (space) {
3212                 int copied = copy_from_iter(skb_put(syn_data, space), space,
3213                                             &fo->data->msg_iter);
3214                 if (unlikely(!copied)) {
3215                         kfree_skb(syn_data);
3216                         goto fallback;
3217                 }
3218                 if (copied != space) {
3219                         skb_trim(syn_data, copied);
3220                         space = copied;
3221                 }
3222         }
3223         /* No more data pending in inet_wait_for_connect() */
3224         if (space == fo->size)
3225                 fo->data = NULL;
3226         fo->copied = space;
3227 
3228         tcp_connect_queue_skb(sk, syn_data);
3229 
3230         err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3231 
3232         syn->skb_mstamp = syn_data->skb_mstamp;
3233 
3234         /* Now full SYN+DATA was cloned and sent (or not),
3235          * remove the SYN from the original skb (syn_data)
3236          * we keep in write queue in case of a retransmit, as we
3237          * also have the SYN packet (with no data) in the same queue.
3238          */
3239         TCP_SKB_CB(syn_data)->seq++;
3240         TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3241         if (!err) {
3242                 tp->syn_data = (fo->copied > 0);
3243                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3244                 goto done;
3245         }
3246 
3247 fallback:
3248         /* Send a regular SYN with Fast Open cookie request option */
3249         if (fo->cookie.len > 0)
3250                 fo->cookie.len = 0;
3251         err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3252         if (err)
3253                 tp->syn_fastopen = 0;
3254 done:
3255         fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
3256         return err;
3257 }
3258 
3259 /* Build a SYN and send it off. */
3260 int tcp_connect(struct sock *sk)
3261 {
3262         struct tcp_sock *tp = tcp_sk(sk);
3263         struct sk_buff *buff;
3264         int err;
3265 
3266         tcp_connect_init(sk);
3267 
3268         if (unlikely(tp->repair)) {
3269                 tcp_finish_connect(sk, NULL);
3270                 return 0;
3271         }
3272 
3273         buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3274         if (unlikely(!buff))
3275                 return -ENOBUFS;
3276 
3277         tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3278         tp->retrans_stamp = tcp_time_stamp;
3279         tcp_connect_queue_skb(sk, buff);
3280         tcp_ecn_send_syn(sk, buff);
3281 
3282         /* Send off SYN; include data in Fast Open. */
3283         err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3284               tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3285         if (err == -ECONNREFUSED)
3286                 return err;
3287 
3288         /* We change tp->snd_nxt after the tcp_transmit_skb() call
3289          * in order to make this packet get counted in tcpOutSegs.
3290          */
3291         tp->snd_nxt = tp->write_seq;
3292         tp->pushed_seq = tp->write_seq;
3293         TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3294 
3295         /* Timer for repeating the SYN until an answer. */
3296         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3297                                   inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3298         return 0;
3299 }
3300 EXPORT_SYMBOL(tcp_connect);
3301 
3302 /* Send out a delayed ack, the caller does the policy checking
3303  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
3304  * for details.
3305  */
3306 void tcp_send_delayed_ack(struct sock *sk)
3307 {
3308         struct inet_connection_sock *icsk = inet_csk(sk);
3309         int ato = icsk->icsk_ack.ato;
3310         unsigned long timeout;
3311 
3312         tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3313 
3314         if (ato > TCP_DELACK_MIN) {
3315                 const struct tcp_sock *tp = tcp_sk(sk);
3316                 int max_ato = HZ / 2;
3317 
3318                 if (icsk->icsk_ack.pingpong ||
3319                     (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3320                         max_ato = TCP_DELACK_MAX;
3321 
3322                 /* Slow path, intersegment interval is "high". */
3323 
3324                 /* If some rtt estimate is known, use it to bound delayed ack.
3325                  * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3326                  * directly.
3327                  */
3328                 if (tp->srtt_us) {
3329                         int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3330                                         TCP_DELACK_MIN);
3331 
3332                         if (rtt < max_ato)
3333                                 max_ato = rtt;
3334                 }
3335 
3336                 ato = min(ato, max_ato);
3337         }
3338 
3339         /* Stay within the limit we were given */
3340         timeout = jiffies + ato;
3341 
3342         /* Use new timeout only if there wasn't a older one earlier. */
3343         if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3344                 /* If delack timer was blocked or is about to expire,
3345                  * send ACK now.
3346                  */
3347                 if (icsk->icsk_ack.blocked ||
3348                     time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3349                         tcp_send_ack(sk);
3350                         return;
3351                 }
3352 
3353                 if (!time_before(timeout, icsk->icsk_ack.timeout))
3354                         timeout = icsk->icsk_ack.timeout;
3355         }
3356         icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3357         icsk->icsk_ack.timeout = timeout;
3358         sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3359 }
3360 
3361 /* This routine sends an ack and also updates the window. */
3362 void tcp_send_ack(struct sock *sk)
3363 {
3364         struct sk_buff *buff;
3365 
3366         /* If we have been reset, we may not send again. */
3367         if (sk->sk_state == TCP_CLOSE)
3368                 return;
3369 
3370         tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3371 
3372         /* We are not putting this on the write queue, so
3373          * tcp_transmit_skb() will set the ownership to this
3374          * sock.
3375          */
3376         buff = alloc_skb(MAX_TCP_HEADER,
3377                          sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3378         if (unlikely(!buff)) {
3379                 inet_csk_schedule_ack(sk);
3380                 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3381                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3382                                           TCP_DELACK_MAX, TCP_RTO_MAX);
3383                 return;
3384         }
3385 
3386         /* Reserve space for headers and prepare control bits. */
3387         skb_reserve(buff, MAX_TCP_HEADER);
3388         tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3389 
3390         /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3391          * too much.
3392          * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3393          * We also avoid tcp_wfree() overhead (cache line miss accessing
3394          * tp->tsq_flags) by using regular sock_wfree()
3395          */
3396         skb_set_tcp_pure_ack(buff);
3397 
3398         /* Send it off, this clears delayed acks for us. */
3399         skb_mstamp_get(&buff->skb_mstamp);
3400         tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3401 }
3402 EXPORT_SYMBOL_GPL(tcp_send_ack);
3403 
3404 /* This routine sends a packet with an out of date sequence
3405  * number. It assumes the other end will try to ack it.
3406  *
3407  * Question: what should we make while urgent mode?
3408  * 4.4BSD forces sending single byte of data. We cannot send
3409  * out of window data, because we have SND.NXT==SND.MAX...
3410  *
3411  * Current solution: to send TWO zero-length segments in urgent mode:
3412  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3413  * out-of-date with SND.UNA-1 to probe window.
3414  */
3415 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3416 {
3417         struct tcp_sock *tp = tcp_sk(sk);
3418         struct sk_buff *skb;
3419 
3420         /* We don't queue it, tcp_transmit_skb() sets ownership. */
3421         skb = alloc_skb(MAX_TCP_HEADER,
3422                         sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3423         if (!skb)
3424                 return -1;
3425 
3426         /* Reserve space for headers and set control bits. */
3427         skb_reserve(skb, MAX_TCP_HEADER);
3428         /* Use a previous sequence.  This should cause the other
3429          * end to send an ack.  Don't queue or clone SKB, just
3430          * send it.
3431          */
3432         tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3433         skb_mstamp_get(&skb->skb_mstamp);
3434         NET_INC_STATS(sock_net(sk), mib);
3435         return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3436 }
3437 
3438 void tcp_send_window_probe(struct sock *sk)
3439 {
3440         if (sk->sk_state == TCP_ESTABLISHED) {
3441                 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3442                 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3443         }
3444 }
3445 
3446 /* Initiate keepalive or window probe from timer. */
3447 int tcp_write_wakeup(struct sock *sk, int mib)
3448 {
3449         struct tcp_sock *tp = tcp_sk(sk);
3450         struct sk_buff *skb;
3451 
3452         if (sk->sk_state == TCP_CLOSE)
3453                 return -1;
3454 
3455         skb = tcp_send_head(sk);
3456         if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3457                 int err;
3458                 unsigned int mss = tcp_current_mss(sk);
3459                 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3460 
3461                 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3462                         tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3463 
3464                 /* We are probing the opening of a window
3465                  * but the window size is != 0
3466                  * must have been a result SWS avoidance ( sender )
3467                  */
3468                 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3469                     skb->len > mss) {
3470                         seg_size = min(seg_size, mss);
3471                         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3472                         if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3473                                 return -1;
3474                 } else if (!tcp_skb_pcount(skb))
3475                         tcp_set_skb_tso_segs(skb, mss);
3476 
3477                 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3478                 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3479                 if (!err)
3480                         tcp_event_new_data_sent(sk, skb);
3481                 return err;
3482         } else {
3483                 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3484                         tcp_xmit_probe_skb(sk, 1, mib);
3485                 return tcp_xmit_probe_skb(sk, 0, mib);
3486         }
3487 }
3488 
3489 /* A window probe timeout has occurred.  If window is not closed send
3490  * a partial packet else a zero probe.
3491  */
3492 void tcp_send_probe0(struct sock *sk)
3493 {
3494         struct inet_connection_sock *icsk = inet_csk(sk);
3495         struct tcp_sock *tp = tcp_sk(sk);
3496         struct net *net = sock_net(sk);
3497         unsigned long probe_max;
3498         int err;
3499 
3500         err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3501 
3502         if (tp->packets_out || !tcp_send_head(sk)) {
3503                 /* Cancel probe timer, if it is not required. */
3504                 icsk->icsk_probes_out = 0;
3505                 icsk->icsk_backoff = 0;
3506                 return;
3507         }
3508 
3509         if (err <= 0) {
3510                 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3511                         icsk->icsk_backoff++;
3512                 icsk->icsk_probes_out++;
3513                 probe_max = TCP_RTO_MAX;
3514         } else {
3515                 /* If packet was not sent due to local congestion,
3516                  * do not backoff and do not remember icsk_probes_out.
3517                  * Let local senders to fight for local resources.
3518                  *
3519                  * Use accumulated backoff yet.
3520                  */
3521                 if (!icsk->icsk_probes_out)
3522                         icsk->icsk_probes_out = 1;
3523                 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3524         }
3525         inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3526                                   tcp_probe0_when(sk, probe_max),
3527                                   TCP_RTO_MAX);
3528 }
3529 
3530 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3531 {
3532         const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3533         struct flowi fl;
3534         int res;
3535 
3536         tcp_rsk(req)->txhash = net_tx_rndhash();
3537         res = af_ops->send_synack(sk, NULL, &fl, req, NULL, true);
3538         if (!res) {
3539                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS);
3540                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3541         }
3542         return res;
3543 }
3544 EXPORT_SYMBOL(tcp_rtx_synack);
3545 

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