Version:  2.0.40 2.2.26 2.4.37 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8

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, space, sysctl_tcp_rmem[2]);
240                 space = max_t(u32, space, sysctl_rmem_max);
241                 space = min_t(u32, space, *window_clamp);
242                 while (space > 65535 && (*rcv_wscale) < 14) {
243                         space >>= 1;
244                         (*rcv_wscale)++;
245                 }
246         }
247 
248         if (mss > (1 << *rcv_wscale)) {
249                 if (!init_rcv_wnd) /* Use default unless specified otherwise */
250                         init_rcv_wnd = tcp_default_init_rwnd(mss);
251                 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
252         }
253 
254         /* Set the clamp no higher than max representable value */
255         (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
256 }
257 EXPORT_SYMBOL(tcp_select_initial_window);
258 
259 /* Chose a new window to advertise, update state in tcp_sock for the
260  * socket, and return result with RFC1323 scaling applied.  The return
261  * value can be stuffed directly into th->window for an outgoing
262  * frame.
263  */
264 static u16 tcp_select_window(struct sock *sk)
265 {
266         struct tcp_sock *tp = tcp_sk(sk);
267         u32 old_win = tp->rcv_wnd;
268         u32 cur_win = tcp_receive_window(tp);
269         u32 new_win = __tcp_select_window(sk);
270 
271         /* Never shrink the offered window */
272         if (new_win < cur_win) {
273                 /* Danger Will Robinson!
274                  * Don't update rcv_wup/rcv_wnd here or else
275                  * we will not be able to advertise a zero
276                  * window in time.  --DaveM
277                  *
278                  * Relax Will Robinson.
279                  */
280                 if (new_win == 0)
281                         NET_INC_STATS(sock_net(sk),
282                                       LINUX_MIB_TCPWANTZEROWINDOWADV);
283                 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
284         }
285         tp->rcv_wnd = new_win;
286         tp->rcv_wup = tp->rcv_nxt;
287 
288         /* Make sure we do not exceed the maximum possible
289          * scaled window.
290          */
291         if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
292                 new_win = min(new_win, MAX_TCP_WINDOW);
293         else
294                 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
295 
296         /* RFC1323 scaling applied */
297         new_win >>= tp->rx_opt.rcv_wscale;
298 
299         /* If we advertise zero window, disable fast path. */
300         if (new_win == 0) {
301                 tp->pred_flags = 0;
302                 if (old_win)
303                         NET_INC_STATS(sock_net(sk),
304                                       LINUX_MIB_TCPTOZEROWINDOWADV);
305         } else if (old_win == 0) {
306                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
307         }
308 
309         return new_win;
310 }
311 
312 /* Packet ECN state for a SYN-ACK */
313 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
314 {
315         const struct tcp_sock *tp = tcp_sk(sk);
316 
317         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
318         if (!(tp->ecn_flags & TCP_ECN_OK))
319                 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
320         else if (tcp_ca_needs_ecn(sk))
321                 INET_ECN_xmit(sk);
322 }
323 
324 /* Packet ECN state for a SYN.  */
325 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
326 {
327         struct tcp_sock *tp = tcp_sk(sk);
328         bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
329                        tcp_ca_needs_ecn(sk);
330 
331         if (!use_ecn) {
332                 const struct dst_entry *dst = __sk_dst_get(sk);
333 
334                 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
335                         use_ecn = true;
336         }
337 
338         tp->ecn_flags = 0;
339 
340         if (use_ecn) {
341                 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
342                 tp->ecn_flags = TCP_ECN_OK;
343                 if (tcp_ca_needs_ecn(sk))
344                         INET_ECN_xmit(sk);
345         }
346 }
347 
348 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
349 {
350         if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
351                 /* tp->ecn_flags are cleared at a later point in time when
352                  * SYN ACK is ultimatively being received.
353                  */
354                 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
355 }
356 
357 static void
358 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
359 {
360         if (inet_rsk(req)->ecn_ok)
361                 th->ece = 1;
362 }
363 
364 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
365  * be sent.
366  */
367 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
368                          struct tcphdr *th, int tcp_header_len)
369 {
370         struct tcp_sock *tp = tcp_sk(sk);
371 
372         if (tp->ecn_flags & TCP_ECN_OK) {
373                 /* Not-retransmitted data segment: set ECT and inject CWR. */
374                 if (skb->len != tcp_header_len &&
375                     !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
376                         INET_ECN_xmit(sk);
377                         if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
378                                 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
379                                 th->cwr = 1;
380                                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
381                         }
382                 } else if (!tcp_ca_needs_ecn(sk)) {
383                         /* ACK or retransmitted segment: clear ECT|CE */
384                         INET_ECN_dontxmit(sk);
385                 }
386                 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
387                         th->ece = 1;
388         }
389 }
390 
391 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
392  * auto increment end seqno.
393  */
394 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
395 {
396         skb->ip_summed = CHECKSUM_PARTIAL;
397         skb->csum = 0;
398 
399         TCP_SKB_CB(skb)->tcp_flags = flags;
400         TCP_SKB_CB(skb)->sacked = 0;
401 
402         tcp_skb_pcount_set(skb, 1);
403 
404         TCP_SKB_CB(skb)->seq = seq;
405         if (flags & (TCPHDR_SYN | TCPHDR_FIN))
406                 seq++;
407         TCP_SKB_CB(skb)->end_seq = seq;
408 }
409 
410 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
411 {
412         return tp->snd_una != tp->snd_up;
413 }
414 
415 #define OPTION_SACK_ADVERTISE   (1 << 0)
416 #define OPTION_TS               (1 << 1)
417 #define OPTION_MD5              (1 << 2)
418 #define OPTION_WSCALE           (1 << 3)
419 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
420 
421 struct tcp_out_options {
422         u16 options;            /* bit field of OPTION_* */
423         u16 mss;                /* 0 to disable */
424         u8 ws;                  /* window scale, 0 to disable */
425         u8 num_sack_blocks;     /* number of SACK blocks to include */
426         u8 hash_size;           /* bytes in hash_location */
427         __u8 *hash_location;    /* temporary pointer, overloaded */
428         __u32 tsval, tsecr;     /* need to include OPTION_TS */
429         struct tcp_fastopen_cookie *fastopen_cookie;    /* Fast open cookie */
430 };
431 
432 /* Write previously computed TCP options to the packet.
433  *
434  * Beware: Something in the Internet is very sensitive to the ordering of
435  * TCP options, we learned this through the hard way, so be careful here.
436  * Luckily we can at least blame others for their non-compliance but from
437  * inter-operability perspective it seems that we're somewhat stuck with
438  * the ordering which we have been using if we want to keep working with
439  * those broken things (not that it currently hurts anybody as there isn't
440  * particular reason why the ordering would need to be changed).
441  *
442  * At least SACK_PERM as the first option is known to lead to a disaster
443  * (but it may well be that other scenarios fail similarly).
444  */
445 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
446                               struct tcp_out_options *opts)
447 {
448         u16 options = opts->options;    /* mungable copy */
449 
450         if (unlikely(OPTION_MD5 & options)) {
451                 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
452                                (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
453                 /* overload cookie hash location */
454                 opts->hash_location = (__u8 *)ptr;
455                 ptr += 4;
456         }
457 
458         if (unlikely(opts->mss)) {
459                 *ptr++ = htonl((TCPOPT_MSS << 24) |
460                                (TCPOLEN_MSS << 16) |
461                                opts->mss);
462         }
463 
464         if (likely(OPTION_TS & options)) {
465                 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
466                         *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
467                                        (TCPOLEN_SACK_PERM << 16) |
468                                        (TCPOPT_TIMESTAMP << 8) |
469                                        TCPOLEN_TIMESTAMP);
470                         options &= ~OPTION_SACK_ADVERTISE;
471                 } else {
472                         *ptr++ = htonl((TCPOPT_NOP << 24) |
473                                        (TCPOPT_NOP << 16) |
474                                        (TCPOPT_TIMESTAMP << 8) |
475                                        TCPOLEN_TIMESTAMP);
476                 }
477                 *ptr++ = htonl(opts->tsval);
478                 *ptr++ = htonl(opts->tsecr);
479         }
480 
481         if (unlikely(OPTION_SACK_ADVERTISE & options)) {
482                 *ptr++ = htonl((TCPOPT_NOP << 24) |
483                                (TCPOPT_NOP << 16) |
484                                (TCPOPT_SACK_PERM << 8) |
485                                TCPOLEN_SACK_PERM);
486         }
487 
488         if (unlikely(OPTION_WSCALE & options)) {
489                 *ptr++ = htonl((TCPOPT_NOP << 24) |
490                                (TCPOPT_WINDOW << 16) |
491                                (TCPOLEN_WINDOW << 8) |
492                                opts->ws);
493         }
494 
495         if (unlikely(opts->num_sack_blocks)) {
496                 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
497                         tp->duplicate_sack : tp->selective_acks;
498                 int this_sack;
499 
500                 *ptr++ = htonl((TCPOPT_NOP  << 24) |
501                                (TCPOPT_NOP  << 16) |
502                                (TCPOPT_SACK <<  8) |
503                                (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
504                                                      TCPOLEN_SACK_PERBLOCK)));
505 
506                 for (this_sack = 0; this_sack < opts->num_sack_blocks;
507                      ++this_sack) {
508                         *ptr++ = htonl(sp[this_sack].start_seq);
509                         *ptr++ = htonl(sp[this_sack].end_seq);
510                 }
511 
512                 tp->rx_opt.dsack = 0;
513         }
514 
515         if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
516                 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
517                 u8 *p = (u8 *)ptr;
518                 u32 len; /* Fast Open option length */
519 
520                 if (foc->exp) {
521                         len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
522                         *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
523                                      TCPOPT_FASTOPEN_MAGIC);
524                         p += TCPOLEN_EXP_FASTOPEN_BASE;
525                 } else {
526                         len = TCPOLEN_FASTOPEN_BASE + foc->len;
527                         *p++ = TCPOPT_FASTOPEN;
528                         *p++ = len;
529                 }
530 
531                 memcpy(p, foc->val, foc->len);
532                 if ((len & 3) == 2) {
533                         p[foc->len] = TCPOPT_NOP;
534                         p[foc->len + 1] = TCPOPT_NOP;
535                 }
536                 ptr += (len + 3) >> 2;
537         }
538 }
539 
540 /* Compute TCP options for SYN packets. This is not the final
541  * network wire format yet.
542  */
543 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
544                                 struct tcp_out_options *opts,
545                                 struct tcp_md5sig_key **md5)
546 {
547         struct tcp_sock *tp = tcp_sk(sk);
548         unsigned int remaining = MAX_TCP_OPTION_SPACE;
549         struct tcp_fastopen_request *fastopen = tp->fastopen_req;
550 
551 #ifdef CONFIG_TCP_MD5SIG
552         *md5 = tp->af_specific->md5_lookup(sk, sk);
553         if (*md5) {
554                 opts->options |= OPTION_MD5;
555                 remaining -= TCPOLEN_MD5SIG_ALIGNED;
556         }
557 #else
558         *md5 = NULL;
559 #endif
560 
561         /* We always get an MSS option.  The option bytes which will be seen in
562          * normal data packets should timestamps be used, must be in the MSS
563          * advertised.  But we subtract them from tp->mss_cache so that
564          * calculations in tcp_sendmsg are simpler etc.  So account for this
565          * fact here if necessary.  If we don't do this correctly, as a
566          * receiver we won't recognize data packets as being full sized when we
567          * should, and thus we won't abide by the delayed ACK rules correctly.
568          * SACKs don't matter, we never delay an ACK when we have any of those
569          * going out.  */
570         opts->mss = tcp_advertise_mss(sk);
571         remaining -= TCPOLEN_MSS_ALIGNED;
572 
573         if (likely(sysctl_tcp_timestamps && !*md5)) {
574                 opts->options |= OPTION_TS;
575                 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
576                 opts->tsecr = tp->rx_opt.ts_recent;
577                 remaining -= TCPOLEN_TSTAMP_ALIGNED;
578         }
579         if (likely(sysctl_tcp_window_scaling)) {
580                 opts->ws = tp->rx_opt.rcv_wscale;
581                 opts->options |= OPTION_WSCALE;
582                 remaining -= TCPOLEN_WSCALE_ALIGNED;
583         }
584         if (likely(sysctl_tcp_sack)) {
585                 opts->options |= OPTION_SACK_ADVERTISE;
586                 if (unlikely(!(OPTION_TS & opts->options)))
587                         remaining -= TCPOLEN_SACKPERM_ALIGNED;
588         }
589 
590         if (fastopen && fastopen->cookie.len >= 0) {
591                 u32 need = fastopen->cookie.len;
592 
593                 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
594                                                TCPOLEN_FASTOPEN_BASE;
595                 need = (need + 3) & ~3U;  /* Align to 32 bits */
596                 if (remaining >= need) {
597                         opts->options |= OPTION_FAST_OPEN_COOKIE;
598                         opts->fastopen_cookie = &fastopen->cookie;
599                         remaining -= need;
600                         tp->syn_fastopen = 1;
601                         tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
602                 }
603         }
604 
605         return MAX_TCP_OPTION_SPACE - remaining;
606 }
607 
608 /* Set up TCP options for SYN-ACKs. */
609 static unsigned int tcp_synack_options(struct request_sock *req,
610                                        unsigned int mss, struct sk_buff *skb,
611                                        struct tcp_out_options *opts,
612                                        const struct tcp_md5sig_key *md5,
613                                        struct tcp_fastopen_cookie *foc)
614 {
615         struct inet_request_sock *ireq = inet_rsk(req);
616         unsigned int remaining = MAX_TCP_OPTION_SPACE;
617 
618 #ifdef CONFIG_TCP_MD5SIG
619         if (md5) {
620                 opts->options |= OPTION_MD5;
621                 remaining -= TCPOLEN_MD5SIG_ALIGNED;
622 
623                 /* We can't fit any SACK blocks in a packet with MD5 + TS
624                  * options. There was discussion about disabling SACK
625                  * rather than TS in order to fit in better with old,
626                  * buggy kernels, but that was deemed to be unnecessary.
627                  */
628                 ireq->tstamp_ok &= !ireq->sack_ok;
629         }
630 #endif
631 
632         /* We always send an MSS option. */
633         opts->mss = mss;
634         remaining -= TCPOLEN_MSS_ALIGNED;
635 
636         if (likely(ireq->wscale_ok)) {
637                 opts->ws = ireq->rcv_wscale;
638                 opts->options |= OPTION_WSCALE;
639                 remaining -= TCPOLEN_WSCALE_ALIGNED;
640         }
641         if (likely(ireq->tstamp_ok)) {
642                 opts->options |= OPTION_TS;
643                 opts->tsval = tcp_skb_timestamp(skb);
644                 opts->tsecr = req->ts_recent;
645                 remaining -= TCPOLEN_TSTAMP_ALIGNED;
646         }
647         if (likely(ireq->sack_ok)) {
648                 opts->options |= OPTION_SACK_ADVERTISE;
649                 if (unlikely(!ireq->tstamp_ok))
650                         remaining -= TCPOLEN_SACKPERM_ALIGNED;
651         }
652         if (foc != NULL && foc->len >= 0) {
653                 u32 need = foc->len;
654 
655                 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
656                                    TCPOLEN_FASTOPEN_BASE;
657                 need = (need + 3) & ~3U;  /* Align to 32 bits */
658                 if (remaining >= need) {
659                         opts->options |= OPTION_FAST_OPEN_COOKIE;
660                         opts->fastopen_cookie = foc;
661                         remaining -= need;
662                 }
663         }
664 
665         return MAX_TCP_OPTION_SPACE - remaining;
666 }
667 
668 /* Compute TCP options for ESTABLISHED sockets. This is not the
669  * final wire format yet.
670  */
671 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
672                                         struct tcp_out_options *opts,
673                                         struct tcp_md5sig_key **md5)
674 {
675         struct tcp_sock *tp = tcp_sk(sk);
676         unsigned int size = 0;
677         unsigned int eff_sacks;
678 
679         opts->options = 0;
680 
681 #ifdef CONFIG_TCP_MD5SIG
682         *md5 = tp->af_specific->md5_lookup(sk, sk);
683         if (unlikely(*md5)) {
684                 opts->options |= OPTION_MD5;
685                 size += TCPOLEN_MD5SIG_ALIGNED;
686         }
687 #else
688         *md5 = NULL;
689 #endif
690 
691         if (likely(tp->rx_opt.tstamp_ok)) {
692                 opts->options |= OPTION_TS;
693                 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
694                 opts->tsecr = tp->rx_opt.ts_recent;
695                 size += TCPOLEN_TSTAMP_ALIGNED;
696         }
697 
698         eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
699         if (unlikely(eff_sacks)) {
700                 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
701                 opts->num_sack_blocks =
702                         min_t(unsigned int, eff_sacks,
703                               (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
704                               TCPOLEN_SACK_PERBLOCK);
705                 size += TCPOLEN_SACK_BASE_ALIGNED +
706                         opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
707         }
708 
709         return size;
710 }
711 
712 
713 /* TCP SMALL QUEUES (TSQ)
714  *
715  * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
716  * to reduce RTT and bufferbloat.
717  * We do this using a special skb destructor (tcp_wfree).
718  *
719  * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
720  * needs to be reallocated in a driver.
721  * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
722  *
723  * Since transmit from skb destructor is forbidden, we use a tasklet
724  * to process all sockets that eventually need to send more skbs.
725  * We use one tasklet per cpu, with its own queue of sockets.
726  */
727 struct tsq_tasklet {
728         struct tasklet_struct   tasklet;
729         struct list_head        head; /* queue of tcp sockets */
730 };
731 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
732 
733 static void tcp_tsq_handler(struct sock *sk)
734 {
735         if ((1 << sk->sk_state) &
736             (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
737              TCPF_CLOSE_WAIT  | TCPF_LAST_ACK))
738                 tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
739                                0, GFP_ATOMIC);
740 }
741 /*
742  * One tasklet per cpu tries to send more skbs.
743  * We run in tasklet context but need to disable irqs when
744  * transferring tsq->head because tcp_wfree() might
745  * interrupt us (non NAPI drivers)
746  */
747 static void tcp_tasklet_func(unsigned long data)
748 {
749         struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
750         LIST_HEAD(list);
751         unsigned long flags;
752         struct list_head *q, *n;
753         struct tcp_sock *tp;
754         struct sock *sk;
755 
756         local_irq_save(flags);
757         list_splice_init(&tsq->head, &list);
758         local_irq_restore(flags);
759 
760         list_for_each_safe(q, n, &list) {
761                 tp = list_entry(q, struct tcp_sock, tsq_node);
762                 list_del(&tp->tsq_node);
763 
764                 sk = (struct sock *)tp;
765                 bh_lock_sock(sk);
766 
767                 if (!sock_owned_by_user(sk)) {
768                         tcp_tsq_handler(sk);
769                 } else {
770                         /* defer the work to tcp_release_cb() */
771                         set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
772                 }
773                 bh_unlock_sock(sk);
774 
775                 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
776                 sk_free(sk);
777         }
778 }
779 
780 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) |           \
781                           (1UL << TCP_WRITE_TIMER_DEFERRED) |   \
782                           (1UL << TCP_DELACK_TIMER_DEFERRED) |  \
783                           (1UL << TCP_MTU_REDUCED_DEFERRED))
784 /**
785  * tcp_release_cb - tcp release_sock() callback
786  * @sk: socket
787  *
788  * called from release_sock() to perform protocol dependent
789  * actions before socket release.
790  */
791 void tcp_release_cb(struct sock *sk)
792 {
793         struct tcp_sock *tp = tcp_sk(sk);
794         unsigned long flags, nflags;
795 
796         /* perform an atomic operation only if at least one flag is set */
797         do {
798                 flags = tp->tsq_flags;
799                 if (!(flags & TCP_DEFERRED_ALL))
800                         return;
801                 nflags = flags & ~TCP_DEFERRED_ALL;
802         } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
803 
804         if (flags & (1UL << TCP_TSQ_DEFERRED))
805                 tcp_tsq_handler(sk);
806 
807         /* Here begins the tricky part :
808          * We are called from release_sock() with :
809          * 1) BH disabled
810          * 2) sk_lock.slock spinlock held
811          * 3) socket owned by us (sk->sk_lock.owned == 1)
812          *
813          * But following code is meant to be called from BH handlers,
814          * so we should keep BH disabled, but early release socket ownership
815          */
816         sock_release_ownership(sk);
817 
818         if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
819                 tcp_write_timer_handler(sk);
820                 __sock_put(sk);
821         }
822         if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
823                 tcp_delack_timer_handler(sk);
824                 __sock_put(sk);
825         }
826         if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
827                 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
828                 __sock_put(sk);
829         }
830 }
831 EXPORT_SYMBOL(tcp_release_cb);
832 
833 void __init tcp_tasklet_init(void)
834 {
835         int i;
836 
837         for_each_possible_cpu(i) {
838                 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
839 
840                 INIT_LIST_HEAD(&tsq->head);
841                 tasklet_init(&tsq->tasklet,
842                              tcp_tasklet_func,
843                              (unsigned long)tsq);
844         }
845 }
846 
847 /*
848  * Write buffer destructor automatically called from kfree_skb.
849  * We can't xmit new skbs from this context, as we might already
850  * hold qdisc lock.
851  */
852 void tcp_wfree(struct sk_buff *skb)
853 {
854         struct sock *sk = skb->sk;
855         struct tcp_sock *tp = tcp_sk(sk);
856         int wmem;
857 
858         /* Keep one reference on sk_wmem_alloc.
859          * Will be released by sk_free() from here or tcp_tasklet_func()
860          */
861         wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
862 
863         /* If this softirq is serviced by ksoftirqd, we are likely under stress.
864          * Wait until our queues (qdisc + devices) are drained.
865          * This gives :
866          * - less callbacks to tcp_write_xmit(), reducing stress (batches)
867          * - chance for incoming ACK (processed by another cpu maybe)
868          *   to migrate this flow (skb->ooo_okay will be eventually set)
869          */
870         if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
871                 goto out;
872 
873         if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
874             !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
875                 unsigned long flags;
876                 struct tsq_tasklet *tsq;
877 
878                 /* queue this socket to tasklet queue */
879                 local_irq_save(flags);
880                 tsq = this_cpu_ptr(&tsq_tasklet);
881                 list_add(&tp->tsq_node, &tsq->head);
882                 tasklet_schedule(&tsq->tasklet);
883                 local_irq_restore(flags);
884                 return;
885         }
886 out:
887         sk_free(sk);
888 }
889 
890 /* This routine actually transmits TCP packets queued in by
891  * tcp_do_sendmsg().  This is used by both the initial
892  * transmission and possible later retransmissions.
893  * All SKB's seen here are completely headerless.  It is our
894  * job to build the TCP header, and pass the packet down to
895  * IP so it can do the same plus pass the packet off to the
896  * device.
897  *
898  * We are working here with either a clone of the original
899  * SKB, or a fresh unique copy made by the retransmit engine.
900  */
901 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
902                             gfp_t gfp_mask)
903 {
904         const struct inet_connection_sock *icsk = inet_csk(sk);
905         struct inet_sock *inet;
906         struct tcp_sock *tp;
907         struct tcp_skb_cb *tcb;
908         struct tcp_out_options opts;
909         unsigned int tcp_options_size, tcp_header_size;
910         struct tcp_md5sig_key *md5;
911         struct tcphdr *th;
912         int err;
913 
914         BUG_ON(!skb || !tcp_skb_pcount(skb));
915         tp = tcp_sk(sk);
916 
917         if (clone_it) {
918                 skb_mstamp_get(&skb->skb_mstamp);
919                 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
920                         - tp->snd_una;
921 
922                 if (unlikely(skb_cloned(skb)))
923                         skb = pskb_copy(skb, gfp_mask);
924                 else
925                         skb = skb_clone(skb, gfp_mask);
926                 if (unlikely(!skb))
927                         return -ENOBUFS;
928         }
929 
930         inet = inet_sk(sk);
931         tcb = TCP_SKB_CB(skb);
932         memset(&opts, 0, sizeof(opts));
933 
934         if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
935                 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
936         else
937                 tcp_options_size = tcp_established_options(sk, skb, &opts,
938                                                            &md5);
939         tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
940 
941         /* if no packet is in qdisc/device queue, then allow XPS to select
942          * another queue. We can be called from tcp_tsq_handler()
943          * which holds one reference to sk_wmem_alloc.
944          *
945          * TODO: Ideally, in-flight pure ACK packets should not matter here.
946          * One way to get this would be to set skb->truesize = 2 on them.
947          */
948         skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
949 
950         skb_push(skb, tcp_header_size);
951         skb_reset_transport_header(skb);
952 
953         skb_orphan(skb);
954         skb->sk = sk;
955         skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
956         skb_set_hash_from_sk(skb, sk);
957         atomic_add(skb->truesize, &sk->sk_wmem_alloc);
958 
959         /* Build TCP header and checksum it. */
960         th = (struct tcphdr *)skb->data;
961         th->source              = inet->inet_sport;
962         th->dest                = inet->inet_dport;
963         th->seq                 = htonl(tcb->seq);
964         th->ack_seq             = htonl(tp->rcv_nxt);
965         *(((__be16 *)th) + 6)   = htons(((tcp_header_size >> 2) << 12) |
966                                         tcb->tcp_flags);
967 
968         th->check               = 0;
969         th->urg_ptr             = 0;
970 
971         /* The urg_mode check is necessary during a below snd_una win probe */
972         if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
973                 if (before(tp->snd_up, tcb->seq + 0x10000)) {
974                         th->urg_ptr = htons(tp->snd_up - tcb->seq);
975                         th->urg = 1;
976                 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
977                         th->urg_ptr = htons(0xFFFF);
978                         th->urg = 1;
979                 }
980         }
981 
982         tcp_options_write((__be32 *)(th + 1), tp, &opts);
983         skb_shinfo(skb)->gso_type = sk->sk_gso_type;
984         if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
985                 th->window      = htons(tcp_select_window(sk));
986                 tcp_ecn_send(sk, skb, th, tcp_header_size);
987         } else {
988                 /* RFC1323: The window in SYN & SYN/ACK segments
989                  * is never scaled.
990                  */
991                 th->window      = htons(min(tp->rcv_wnd, 65535U));
992         }
993 #ifdef CONFIG_TCP_MD5SIG
994         /* Calculate the MD5 hash, as we have all we need now */
995         if (md5) {
996                 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
997                 tp->af_specific->calc_md5_hash(opts.hash_location,
998                                                md5, sk, skb);
999         }
1000 #endif
1001 
1002         icsk->icsk_af_ops->send_check(sk, skb);
1003 
1004         if (likely(tcb->tcp_flags & TCPHDR_ACK))
1005                 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1006 
1007         if (skb->len != tcp_header_size) {
1008                 tcp_event_data_sent(tp, sk);
1009                 tp->data_segs_out += tcp_skb_pcount(skb);
1010         }
1011 
1012         if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1013                 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1014                               tcp_skb_pcount(skb));
1015 
1016         tp->segs_out += tcp_skb_pcount(skb);
1017         /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1018         skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1019         skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1020 
1021         /* Our usage of tstamp should remain private */
1022         skb->tstamp.tv64 = 0;
1023 
1024         /* Cleanup our debris for IP stacks */
1025         memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1026                                sizeof(struct inet6_skb_parm)));
1027 
1028         err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1029 
1030         if (likely(err <= 0))
1031                 return err;
1032 
1033         tcp_enter_cwr(sk);
1034 
1035         return net_xmit_eval(err);
1036 }
1037 
1038 /* This routine just queues the buffer for sending.
1039  *
1040  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1041  * otherwise socket can stall.
1042  */
1043 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1044 {
1045         struct tcp_sock *tp = tcp_sk(sk);
1046 
1047         /* Advance write_seq and place onto the write_queue. */
1048         tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1049         __skb_header_release(skb);
1050         tcp_add_write_queue_tail(sk, skb);
1051         sk->sk_wmem_queued += skb->truesize;
1052         sk_mem_charge(sk, skb->truesize);
1053 }
1054 
1055 /* Initialize TSO segments for a packet. */
1056 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1057 {
1058         if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1059                 /* Avoid the costly divide in the normal
1060                  * non-TSO case.
1061                  */
1062                 tcp_skb_pcount_set(skb, 1);
1063                 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1064         } else {
1065                 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1066                 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1067         }
1068 }
1069 
1070 /* When a modification to fackets out becomes necessary, we need to check
1071  * skb is counted to fackets_out or not.
1072  */
1073 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1074                                    int decr)
1075 {
1076         struct tcp_sock *tp = tcp_sk(sk);
1077 
1078         if (!tp->sacked_out || tcp_is_reno(tp))
1079                 return;
1080 
1081         if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1082                 tp->fackets_out -= decr;
1083 }
1084 
1085 /* Pcount in the middle of the write queue got changed, we need to do various
1086  * tweaks to fix counters
1087  */
1088 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1089 {
1090         struct tcp_sock *tp = tcp_sk(sk);
1091 
1092         tp->packets_out -= decr;
1093 
1094         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1095                 tp->sacked_out -= decr;
1096         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1097                 tp->retrans_out -= decr;
1098         if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1099                 tp->lost_out -= decr;
1100 
1101         /* Reno case is special. Sigh... */
1102         if (tcp_is_reno(tp) && decr > 0)
1103                 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1104 
1105         tcp_adjust_fackets_out(sk, skb, decr);
1106 
1107         if (tp->lost_skb_hint &&
1108             before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1109             (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1110                 tp->lost_cnt_hint -= decr;
1111 
1112         tcp_verify_left_out(tp);
1113 }
1114 
1115 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1116 {
1117         return TCP_SKB_CB(skb)->txstamp_ack ||
1118                 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1119 }
1120 
1121 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1122 {
1123         struct skb_shared_info *shinfo = skb_shinfo(skb);
1124 
1125         if (unlikely(tcp_has_tx_tstamp(skb)) &&
1126             !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1127                 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1128                 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1129 
1130                 shinfo->tx_flags &= ~tsflags;
1131                 shinfo2->tx_flags |= tsflags;
1132                 swap(shinfo->tskey, shinfo2->tskey);
1133                 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1134                 TCP_SKB_CB(skb)->txstamp_ack = 0;
1135         }
1136 }
1137 
1138 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1139 {
1140         TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1141         TCP_SKB_CB(skb)->eor = 0;
1142 }
1143 
1144 /* Function to create two new TCP segments.  Shrinks the given segment
1145  * to the specified size and appends a new segment with the rest of the
1146  * packet to the list.  This won't be called frequently, I hope.
1147  * Remember, these are still headerless SKBs at this point.
1148  */
1149 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1150                  unsigned int mss_now, gfp_t gfp)
1151 {
1152         struct tcp_sock *tp = tcp_sk(sk);
1153         struct sk_buff *buff;
1154         int nsize, old_factor;
1155         int nlen;
1156         u8 flags;
1157 
1158         if (WARN_ON(len > skb->len))
1159                 return -EINVAL;
1160 
1161         nsize = skb_headlen(skb) - len;
1162         if (nsize < 0)
1163                 nsize = 0;
1164 
1165         if (skb_unclone(skb, gfp))
1166                 return -ENOMEM;
1167 
1168         /* Get a new skb... force flag on. */
1169         buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1170         if (!buff)
1171                 return -ENOMEM; /* We'll just try again later. */
1172 
1173         sk->sk_wmem_queued += buff->truesize;
1174         sk_mem_charge(sk, buff->truesize);
1175         nlen = skb->len - len - nsize;
1176         buff->truesize += nlen;
1177         skb->truesize -= nlen;
1178 
1179         /* Correct the sequence numbers. */
1180         TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1181         TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1182         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1183 
1184         /* PSH and FIN should only be set in the second packet. */
1185         flags = TCP_SKB_CB(skb)->tcp_flags;
1186         TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1187         TCP_SKB_CB(buff)->tcp_flags = flags;
1188         TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1189         tcp_skb_fragment_eor(skb, buff);
1190 
1191         if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1192                 /* Copy and checksum data tail into the new buffer. */
1193                 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1194                                                        skb_put(buff, nsize),
1195                                                        nsize, 0);
1196 
1197                 skb_trim(skb, len);
1198 
1199                 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1200         } else {
1201                 skb->ip_summed = CHECKSUM_PARTIAL;
1202                 skb_split(skb, buff, len);
1203         }
1204 
1205         buff->ip_summed = skb->ip_summed;
1206 
1207         buff->tstamp = skb->tstamp;
1208         tcp_fragment_tstamp(skb, buff);
1209 
1210         old_factor = tcp_skb_pcount(skb);
1211 
1212         /* Fix up tso_factor for both original and new SKB.  */
1213         tcp_set_skb_tso_segs(skb, mss_now);
1214         tcp_set_skb_tso_segs(buff, mss_now);
1215 
1216         /* If this packet has been sent out already, we must
1217          * adjust the various packet counters.
1218          */
1219         if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1220                 int diff = old_factor - tcp_skb_pcount(skb) -
1221                         tcp_skb_pcount(buff);
1222 
1223                 if (diff)
1224                         tcp_adjust_pcount(sk, skb, diff);
1225         }
1226 
1227         /* Link BUFF into the send queue. */
1228         __skb_header_release(buff);
1229         tcp_insert_write_queue_after(skb, buff, sk);
1230 
1231         return 0;
1232 }
1233 
1234 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1235  * eventually). The difference is that pulled data not copied, but
1236  * immediately discarded.
1237  */
1238 static void __pskb_trim_head(struct sk_buff *skb, int len)
1239 {
1240         struct skb_shared_info *shinfo;
1241         int i, k, eat;
1242 
1243         eat = min_t(int, len, skb_headlen(skb));
1244         if (eat) {
1245                 __skb_pull(skb, eat);
1246                 len -= eat;
1247                 if (!len)
1248                         return;
1249         }
1250         eat = len;
1251         k = 0;
1252         shinfo = skb_shinfo(skb);
1253         for (i = 0; i < shinfo->nr_frags; i++) {
1254                 int size = skb_frag_size(&shinfo->frags[i]);
1255 
1256                 if (size <= eat) {
1257                         skb_frag_unref(skb, i);
1258                         eat -= size;
1259                 } else {
1260                         shinfo->frags[k] = shinfo->frags[i];
1261                         if (eat) {
1262                                 shinfo->frags[k].page_offset += eat;
1263                                 skb_frag_size_sub(&shinfo->frags[k], eat);
1264                                 eat = 0;
1265                         }
1266                         k++;
1267                 }
1268         }
1269         shinfo->nr_frags = k;
1270 
1271         skb_reset_tail_pointer(skb);
1272         skb->data_len -= len;
1273         skb->len = skb->data_len;
1274 }
1275 
1276 /* Remove acked data from a packet in the transmit queue. */
1277 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1278 {
1279         if (skb_unclone(skb, GFP_ATOMIC))
1280                 return -ENOMEM;
1281 
1282         __pskb_trim_head(skb, len);
1283 
1284         TCP_SKB_CB(skb)->seq += len;
1285         skb->ip_summed = CHECKSUM_PARTIAL;
1286 
1287         skb->truesize        -= len;
1288         sk->sk_wmem_queued   -= len;
1289         sk_mem_uncharge(sk, len);
1290         sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1291 
1292         /* Any change of skb->len requires recalculation of tso factor. */
1293         if (tcp_skb_pcount(skb) > 1)
1294                 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1295 
1296         return 0;
1297 }
1298 
1299 /* Calculate MSS not accounting any TCP options.  */
1300 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1301 {
1302         const struct tcp_sock *tp = tcp_sk(sk);
1303         const struct inet_connection_sock *icsk = inet_csk(sk);
1304         int mss_now;
1305 
1306         /* Calculate base mss without TCP options:
1307            It is MMS_S - sizeof(tcphdr) of rfc1122
1308          */
1309         mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1310 
1311         /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1312         if (icsk->icsk_af_ops->net_frag_header_len) {
1313                 const struct dst_entry *dst = __sk_dst_get(sk);
1314 
1315                 if (dst && dst_allfrag(dst))
1316                         mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1317         }
1318 
1319         /* Clamp it (mss_clamp does not include tcp options) */
1320         if (mss_now > tp->rx_opt.mss_clamp)
1321                 mss_now = tp->rx_opt.mss_clamp;
1322 
1323         /* Now subtract optional transport overhead */
1324         mss_now -= icsk->icsk_ext_hdr_len;
1325 
1326         /* Then reserve room for full set of TCP options and 8 bytes of data */
1327         if (mss_now < 48)
1328                 mss_now = 48;
1329         return mss_now;
1330 }
1331 
1332 /* Calculate MSS. Not accounting for SACKs here.  */
1333 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1334 {
1335         /* Subtract TCP options size, not including SACKs */
1336         return __tcp_mtu_to_mss(sk, pmtu) -
1337                (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1338 }
1339 
1340 /* Inverse of above */
1341 int tcp_mss_to_mtu(struct sock *sk, int mss)
1342 {
1343         const struct tcp_sock *tp = tcp_sk(sk);
1344         const struct inet_connection_sock *icsk = inet_csk(sk);
1345         int mtu;
1346 
1347         mtu = mss +
1348               tp->tcp_header_len +
1349               icsk->icsk_ext_hdr_len +
1350               icsk->icsk_af_ops->net_header_len;
1351 
1352         /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1353         if (icsk->icsk_af_ops->net_frag_header_len) {
1354                 const struct dst_entry *dst = __sk_dst_get(sk);
1355 
1356                 if (dst && dst_allfrag(dst))
1357                         mtu += icsk->icsk_af_ops->net_frag_header_len;
1358         }
1359         return mtu;
1360 }
1361 
1362 /* MTU probing init per socket */
1363 void tcp_mtup_init(struct sock *sk)
1364 {
1365         struct tcp_sock *tp = tcp_sk(sk);
1366         struct inet_connection_sock *icsk = inet_csk(sk);
1367         struct net *net = sock_net(sk);
1368 
1369         icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1370         icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1371                                icsk->icsk_af_ops->net_header_len;
1372         icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1373         icsk->icsk_mtup.probe_size = 0;
1374         if (icsk->icsk_mtup.enabled)
1375                 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1376 }
1377 EXPORT_SYMBOL(tcp_mtup_init);
1378 
1379 /* This function synchronize snd mss to current pmtu/exthdr set.
1380 
1381    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1382    for TCP options, but includes only bare TCP header.
1383 
1384    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1385    It is minimum of user_mss and mss received with SYN.
1386    It also does not include TCP options.
1387 
1388    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1389 
1390    tp->mss_cache is current effective sending mss, including
1391    all tcp options except for SACKs. It is evaluated,
1392    taking into account current pmtu, but never exceeds
1393    tp->rx_opt.mss_clamp.
1394 
1395    NOTE1. rfc1122 clearly states that advertised MSS
1396    DOES NOT include either tcp or ip options.
1397 
1398    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1399    are READ ONLY outside this function.         --ANK (980731)
1400  */
1401 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1402 {
1403         struct tcp_sock *tp = tcp_sk(sk);
1404         struct inet_connection_sock *icsk = inet_csk(sk);
1405         int mss_now;
1406 
1407         if (icsk->icsk_mtup.search_high > pmtu)
1408                 icsk->icsk_mtup.search_high = pmtu;
1409 
1410         mss_now = tcp_mtu_to_mss(sk, pmtu);
1411         mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1412 
1413         /* And store cached results */
1414         icsk->icsk_pmtu_cookie = pmtu;
1415         if (icsk->icsk_mtup.enabled)
1416                 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1417         tp->mss_cache = mss_now;
1418 
1419         return mss_now;
1420 }
1421 EXPORT_SYMBOL(tcp_sync_mss);
1422 
1423 /* Compute the current effective MSS, taking SACKs and IP options,
1424  * and even PMTU discovery events into account.
1425  */
1426 unsigned int tcp_current_mss(struct sock *sk)
1427 {
1428         const struct tcp_sock *tp = tcp_sk(sk);
1429         const struct dst_entry *dst = __sk_dst_get(sk);
1430         u32 mss_now;
1431         unsigned int header_len;
1432         struct tcp_out_options opts;
1433         struct tcp_md5sig_key *md5;
1434 
1435         mss_now = tp->mss_cache;
1436 
1437         if (dst) {
1438                 u32 mtu = dst_mtu(dst);
1439                 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1440                         mss_now = tcp_sync_mss(sk, mtu);
1441         }
1442 
1443         header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1444                      sizeof(struct tcphdr);
1445         /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1446          * some common options. If this is an odd packet (because we have SACK
1447          * blocks etc) then our calculated header_len will be different, and
1448          * we have to adjust mss_now correspondingly */
1449         if (header_len != tp->tcp_header_len) {
1450                 int delta = (int) header_len - tp->tcp_header_len;
1451                 mss_now -= delta;
1452         }
1453 
1454         return mss_now;
1455 }
1456 
1457 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1458  * As additional protections, we do not touch cwnd in retransmission phases,
1459  * and if application hit its sndbuf limit recently.
1460  */
1461 static void tcp_cwnd_application_limited(struct sock *sk)
1462 {
1463         struct tcp_sock *tp = tcp_sk(sk);
1464 
1465         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1466             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1467                 /* Limited by application or receiver window. */
1468                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1469                 u32 win_used = max(tp->snd_cwnd_used, init_win);
1470                 if (win_used < tp->snd_cwnd) {
1471                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
1472                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1473                 }
1474                 tp->snd_cwnd_used = 0;
1475         }
1476         tp->snd_cwnd_stamp = tcp_time_stamp;
1477 }
1478 
1479 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1480 {
1481         struct tcp_sock *tp = tcp_sk(sk);
1482 
1483         /* Track the maximum number of outstanding packets in each
1484          * window, and remember whether we were cwnd-limited then.
1485          */
1486         if (!before(tp->snd_una, tp->max_packets_seq) ||
1487             tp->packets_out > tp->max_packets_out) {
1488                 tp->max_packets_out = tp->packets_out;
1489                 tp->max_packets_seq = tp->snd_nxt;
1490                 tp->is_cwnd_limited = is_cwnd_limited;
1491         }
1492 
1493         if (tcp_is_cwnd_limited(sk)) {
1494                 /* Network is feed fully. */
1495                 tp->snd_cwnd_used = 0;
1496                 tp->snd_cwnd_stamp = tcp_time_stamp;
1497         } else {
1498                 /* Network starves. */
1499                 if (tp->packets_out > tp->snd_cwnd_used)
1500                         tp->snd_cwnd_used = tp->packets_out;
1501 
1502                 if (sysctl_tcp_slow_start_after_idle &&
1503                     (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1504                         tcp_cwnd_application_limited(sk);
1505         }
1506 }
1507 
1508 /* Minshall's variant of the Nagle send check. */
1509 static bool tcp_minshall_check(const struct tcp_sock *tp)
1510 {
1511         return after(tp->snd_sml, tp->snd_una) &&
1512                 !after(tp->snd_sml, tp->snd_nxt);
1513 }
1514 
1515 /* Update snd_sml if this skb is under mss
1516  * Note that a TSO packet might end with a sub-mss segment
1517  * The test is really :
1518  * if ((skb->len % mss) != 0)
1519  *        tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1520  * But we can avoid doing the divide again given we already have
1521  *  skb_pcount = skb->len / mss_now
1522  */
1523 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1524                                 const struct sk_buff *skb)
1525 {
1526         if (skb->len < tcp_skb_pcount(skb) * mss_now)
1527                 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1528 }
1529 
1530 /* Return false, if packet can be sent now without violation Nagle's rules:
1531  * 1. It is full sized. (provided by caller in %partial bool)
1532  * 2. Or it contains FIN. (already checked by caller)
1533  * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1534  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1535  *    With Minshall's modification: all sent small packets are ACKed.
1536  */
1537 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1538                             int nonagle)
1539 {
1540         return partial &&
1541                 ((nonagle & TCP_NAGLE_CORK) ||
1542                  (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1543 }
1544 
1545 /* Return how many segs we'd like on a TSO packet,
1546  * to send one TSO packet per ms
1547  */
1548 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now)
1549 {
1550         u32 bytes, segs;
1551 
1552         bytes = min(sk->sk_pacing_rate >> 10,
1553                     sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1554 
1555         /* Goal is to send at least one packet per ms,
1556          * not one big TSO packet every 100 ms.
1557          * This preserves ACK clocking and is consistent
1558          * with tcp_tso_should_defer() heuristic.
1559          */
1560         segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs);
1561 
1562         return min_t(u32, segs, sk->sk_gso_max_segs);
1563 }
1564 
1565 /* Returns the portion of skb which can be sent right away */
1566 static unsigned int tcp_mss_split_point(const struct sock *sk,
1567                                         const struct sk_buff *skb,
1568                                         unsigned int mss_now,
1569                                         unsigned int max_segs,
1570                                         int nonagle)
1571 {
1572         const struct tcp_sock *tp = tcp_sk(sk);
1573         u32 partial, needed, window, max_len;
1574 
1575         window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1576         max_len = mss_now * max_segs;
1577 
1578         if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1579                 return max_len;
1580 
1581         needed = min(skb->len, window);
1582 
1583         if (max_len <= needed)
1584                 return max_len;
1585 
1586         partial = needed % mss_now;
1587         /* If last segment is not a full MSS, check if Nagle rules allow us
1588          * to include this last segment in this skb.
1589          * Otherwise, we'll split the skb at last MSS boundary
1590          */
1591         if (tcp_nagle_check(partial != 0, tp, nonagle))
1592                 return needed - partial;
1593 
1594         return needed;
1595 }
1596 
1597 /* Can at least one segment of SKB be sent right now, according to the
1598  * congestion window rules?  If so, return how many segments are allowed.
1599  */
1600 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1601                                          const struct sk_buff *skb)
1602 {
1603         u32 in_flight, cwnd, halfcwnd;
1604 
1605         /* Don't be strict about the congestion window for the final FIN.  */
1606         if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1607             tcp_skb_pcount(skb) == 1)
1608                 return 1;
1609 
1610         in_flight = tcp_packets_in_flight(tp);
1611         cwnd = tp->snd_cwnd;
1612         if (in_flight >= cwnd)
1613                 return 0;
1614 
1615         /* For better scheduling, ensure we have at least
1616          * 2 GSO packets in flight.
1617          */
1618         halfcwnd = max(cwnd >> 1, 1U);
1619         return min(halfcwnd, cwnd - in_flight);
1620 }
1621 
1622 /* Initialize TSO state of a skb.
1623  * This must be invoked the first time we consider transmitting
1624  * SKB onto the wire.
1625  */
1626 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1627 {
1628         int tso_segs = tcp_skb_pcount(skb);
1629 
1630         if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1631                 tcp_set_skb_tso_segs(skb, mss_now);
1632                 tso_segs = tcp_skb_pcount(skb);
1633         }
1634         return tso_segs;
1635 }
1636 
1637 
1638 /* Return true if the Nagle test allows this packet to be
1639  * sent now.
1640  */
1641 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1642                                   unsigned int cur_mss, int nonagle)
1643 {
1644         /* Nagle rule does not apply to frames, which sit in the middle of the
1645          * write_queue (they have no chances to get new data).
1646          *
1647          * This is implemented in the callers, where they modify the 'nonagle'
1648          * argument based upon the location of SKB in the send queue.
1649          */
1650         if (nonagle & TCP_NAGLE_PUSH)
1651                 return true;
1652 
1653         /* Don't use the nagle rule for urgent data (or for the final FIN). */
1654         if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1655                 return true;
1656 
1657         if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1658                 return true;
1659 
1660         return false;
1661 }
1662 
1663 /* Does at least the first segment of SKB fit into the send window? */
1664 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1665                              const struct sk_buff *skb,
1666                              unsigned int cur_mss)
1667 {
1668         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1669 
1670         if (skb->len > cur_mss)
1671                 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1672 
1673         return !after(end_seq, tcp_wnd_end(tp));
1674 }
1675 
1676 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1677  * should be put on the wire right now.  If so, it returns the number of
1678  * packets allowed by the congestion window.
1679  */
1680 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1681                                  unsigned int cur_mss, int nonagle)
1682 {
1683         const struct tcp_sock *tp = tcp_sk(sk);
1684         unsigned int cwnd_quota;
1685 
1686         tcp_init_tso_segs(skb, cur_mss);
1687 
1688         if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1689                 return 0;
1690 
1691         cwnd_quota = tcp_cwnd_test(tp, skb);
1692         if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1693                 cwnd_quota = 0;
1694 
1695         return cwnd_quota;
1696 }
1697 
1698 /* Test if sending is allowed right now. */
1699 bool tcp_may_send_now(struct sock *sk)
1700 {
1701         const struct tcp_sock *tp = tcp_sk(sk);
1702         struct sk_buff *skb = tcp_send_head(sk);
1703 
1704         return skb &&
1705                 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1706                              (tcp_skb_is_last(sk, skb) ?
1707                               tp->nonagle : TCP_NAGLE_PUSH));
1708 }
1709 
1710 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1711  * which is put after SKB on the list.  It is very much like
1712  * tcp_fragment() except that it may make several kinds of assumptions
1713  * in order to speed up the splitting operation.  In particular, we
1714  * know that all the data is in scatter-gather pages, and that the
1715  * packet has never been sent out before (and thus is not cloned).
1716  */
1717 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1718                         unsigned int mss_now, gfp_t gfp)
1719 {
1720         struct sk_buff *buff;
1721         int nlen = skb->len - len;
1722         u8 flags;
1723 
1724         /* All of a TSO frame must be composed of paged data.  */
1725         if (skb->len != skb->data_len)
1726                 return tcp_fragment(sk, skb, len, mss_now, gfp);
1727 
1728         buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1729         if (unlikely(!buff))
1730                 return -ENOMEM;
1731 
1732         sk->sk_wmem_queued += buff->truesize;
1733         sk_mem_charge(sk, buff->truesize);
1734         buff->truesize += nlen;
1735         skb->truesize -= nlen;
1736 
1737         /* Correct the sequence numbers. */
1738         TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1739         TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1740         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1741 
1742         /* PSH and FIN should only be set in the second packet. */
1743         flags = TCP_SKB_CB(skb)->tcp_flags;
1744         TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1745         TCP_SKB_CB(buff)->tcp_flags = flags;
1746 
1747         /* This packet was never sent out yet, so no SACK bits. */
1748         TCP_SKB_CB(buff)->sacked = 0;
1749 
1750         tcp_skb_fragment_eor(skb, buff);
1751 
1752         buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1753         skb_split(skb, buff, len);
1754         tcp_fragment_tstamp(skb, buff);
1755 
1756         /* Fix up tso_factor for both original and new SKB.  */
1757         tcp_set_skb_tso_segs(skb, mss_now);
1758         tcp_set_skb_tso_segs(buff, mss_now);
1759 
1760         /* Link BUFF into the send queue. */
1761         __skb_header_release(buff);
1762         tcp_insert_write_queue_after(skb, buff, sk);
1763 
1764         return 0;
1765 }
1766 
1767 /* Try to defer sending, if possible, in order to minimize the amount
1768  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1769  *
1770  * This algorithm is from John Heffner.
1771  */
1772 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1773                                  bool *is_cwnd_limited, u32 max_segs)
1774 {
1775         const struct inet_connection_sock *icsk = inet_csk(sk);
1776         u32 age, send_win, cong_win, limit, in_flight;
1777         struct tcp_sock *tp = tcp_sk(sk);
1778         struct skb_mstamp now;
1779         struct sk_buff *head;
1780         int win_divisor;
1781 
1782         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1783                 goto send_now;
1784 
1785         if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1786                 goto send_now;
1787 
1788         /* Avoid bursty behavior by allowing defer
1789          * only if the last write was recent.
1790          */
1791         if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1792                 goto send_now;
1793 
1794         in_flight = tcp_packets_in_flight(tp);
1795 
1796         BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1797 
1798         send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1799 
1800         /* From in_flight test above, we know that cwnd > in_flight.  */
1801         cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1802 
1803         limit = min(send_win, cong_win);
1804 
1805         /* If a full-sized TSO skb can be sent, do it. */
1806         if (limit >= max_segs * tp->mss_cache)
1807                 goto send_now;
1808 
1809         /* Middle in queue won't get any more data, full sendable already? */
1810         if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1811                 goto send_now;
1812 
1813         win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1814         if (win_divisor) {
1815                 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1816 
1817                 /* If at least some fraction of a window is available,
1818                  * just use it.
1819                  */
1820                 chunk /= win_divisor;
1821                 if (limit >= chunk)
1822                         goto send_now;
1823         } else {
1824                 /* Different approach, try not to defer past a single
1825                  * ACK.  Receiver should ACK every other full sized
1826                  * frame, so if we have space for more than 3 frames
1827                  * then send now.
1828                  */
1829                 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1830                         goto send_now;
1831         }
1832 
1833         head = tcp_write_queue_head(sk);
1834         skb_mstamp_get(&now);
1835         age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1836         /* If next ACK is likely to come too late (half srtt), do not defer */
1837         if (age < (tp->srtt_us >> 4))
1838                 goto send_now;
1839 
1840         /* Ok, it looks like it is advisable to defer. */
1841 
1842         if (cong_win < send_win && cong_win <= skb->len)
1843                 *is_cwnd_limited = true;
1844 
1845         return true;
1846 
1847 send_now:
1848         return false;
1849 }
1850 
1851 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1852 {
1853         struct inet_connection_sock *icsk = inet_csk(sk);
1854         struct tcp_sock *tp = tcp_sk(sk);
1855         struct net *net = sock_net(sk);
1856         u32 interval;
1857         s32 delta;
1858 
1859         interval = net->ipv4.sysctl_tcp_probe_interval;
1860         delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1861         if (unlikely(delta >= interval * HZ)) {
1862                 int mss = tcp_current_mss(sk);
1863 
1864                 /* Update current search range */
1865                 icsk->icsk_mtup.probe_size = 0;
1866                 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1867                         sizeof(struct tcphdr) +
1868                         icsk->icsk_af_ops->net_header_len;
1869                 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1870 
1871                 /* Update probe time stamp */
1872                 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1873         }
1874 }
1875 
1876 /* Create a new MTU probe if we are ready.
1877  * MTU probe is regularly attempting to increase the path MTU by
1878  * deliberately sending larger packets.  This discovers routing
1879  * changes resulting in larger path MTUs.
1880  *
1881  * Returns 0 if we should wait to probe (no cwnd available),
1882  *         1 if a probe was sent,
1883  *         -1 otherwise
1884  */
1885 static int tcp_mtu_probe(struct sock *sk)
1886 {
1887         struct tcp_sock *tp = tcp_sk(sk);
1888         struct inet_connection_sock *icsk = inet_csk(sk);
1889         struct sk_buff *skb, *nskb, *next;
1890         struct net *net = sock_net(sk);
1891         int len;
1892         int probe_size;
1893         int size_needed;
1894         int copy;
1895         int mss_now;
1896         int interval;
1897 
1898         /* Not currently probing/verifying,
1899          * not in recovery,
1900          * have enough cwnd, and
1901          * not SACKing (the variable headers throw things off) */
1902         if (!icsk->icsk_mtup.enabled ||
1903             icsk->icsk_mtup.probe_size ||
1904             inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1905             tp->snd_cwnd < 11 ||
1906             tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1907                 return -1;
1908 
1909         /* Use binary search for probe_size between tcp_mss_base,
1910          * and current mss_clamp. if (search_high - search_low)
1911          * smaller than a threshold, backoff from probing.
1912          */
1913         mss_now = tcp_current_mss(sk);
1914         probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1915                                     icsk->icsk_mtup.search_low) >> 1);
1916         size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1917         interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
1918         /* When misfortune happens, we are reprobing actively,
1919          * and then reprobe timer has expired. We stick with current
1920          * probing process by not resetting search range to its orignal.
1921          */
1922         if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
1923                 interval < net->ipv4.sysctl_tcp_probe_threshold) {
1924                 /* Check whether enough time has elaplased for
1925                  * another round of probing.
1926                  */
1927                 tcp_mtu_check_reprobe(sk);
1928                 return -1;
1929         }
1930 
1931         /* Have enough data in the send queue to probe? */
1932         if (tp->write_seq - tp->snd_nxt < size_needed)
1933                 return -1;
1934 
1935         if (tp->snd_wnd < size_needed)
1936                 return -1;
1937         if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1938                 return 0;
1939 
1940         /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1941         if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1942                 if (!tcp_packets_in_flight(tp))
1943                         return -1;
1944                 else
1945                         return 0;
1946         }
1947 
1948         /* We're allowed to probe.  Build it now. */
1949         nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
1950         if (!nskb)
1951                 return -1;
1952         sk->sk_wmem_queued += nskb->truesize;
1953         sk_mem_charge(sk, nskb->truesize);
1954 
1955         skb = tcp_send_head(sk);
1956 
1957         TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1958         TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1959         TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1960         TCP_SKB_CB(nskb)->sacked = 0;
1961         nskb->csum = 0;
1962         nskb->ip_summed = skb->ip_summed;
1963 
1964         tcp_insert_write_queue_before(nskb, skb, sk);
1965 
1966         len = 0;
1967         tcp_for_write_queue_from_safe(skb, next, sk) {
1968                 copy = min_t(int, skb->len, probe_size - len);
1969                 if (nskb->ip_summed) {
1970                         skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1971                 } else {
1972                         __wsum csum = skb_copy_and_csum_bits(skb, 0,
1973                                                              skb_put(nskb, copy),
1974                                                              copy, 0);
1975                         nskb->csum = csum_block_add(nskb->csum, csum, len);
1976                 }
1977 
1978                 if (skb->len <= copy) {
1979                         /* We've eaten all the data from this skb.
1980                          * Throw it away. */
1981                         TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1982                         tcp_unlink_write_queue(skb, sk);
1983                         sk_wmem_free_skb(sk, skb);
1984                 } else {
1985                         TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1986                                                    ~(TCPHDR_FIN|TCPHDR_PSH);
1987                         if (!skb_shinfo(skb)->nr_frags) {
1988                                 skb_pull(skb, copy);
1989                                 if (skb->ip_summed != CHECKSUM_PARTIAL)
1990                                         skb->csum = csum_partial(skb->data,
1991                                                                  skb->len, 0);
1992                         } else {
1993                                 __pskb_trim_head(skb, copy);
1994                                 tcp_set_skb_tso_segs(skb, mss_now);
1995                         }
1996                         TCP_SKB_CB(skb)->seq += copy;
1997                 }
1998 
1999                 len += copy;
2000 
2001                 if (len >= probe_size)
2002                         break;
2003         }
2004         tcp_init_tso_segs(nskb, nskb->len);
2005 
2006         /* We're ready to send.  If this fails, the probe will
2007          * be resegmented into mss-sized pieces by tcp_write_xmit().
2008          */
2009         if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2010                 /* Decrement cwnd here because we are sending
2011                  * effectively two packets. */
2012                 tp->snd_cwnd--;
2013                 tcp_event_new_data_sent(sk, nskb);
2014 
2015                 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2016                 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2017                 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2018 
2019                 return 1;
2020         }
2021 
2022         return -1;
2023 }
2024 
2025 /* This routine writes packets to the network.  It advances the
2026  * send_head.  This happens as incoming acks open up the remote
2027  * window for us.
2028  *
2029  * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2030  * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2031  * account rare use of URG, this is not a big flaw.
2032  *
2033  * Send at most one packet when push_one > 0. Temporarily ignore
2034  * cwnd limit to force at most one packet out when push_one == 2.
2035 
2036  * Returns true, if no segments are in flight and we have queued segments,
2037  * but cannot send anything now because of SWS or another problem.
2038  */
2039 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2040                            int push_one, gfp_t gfp)
2041 {
2042         struct tcp_sock *tp = tcp_sk(sk);
2043         struct sk_buff *skb;
2044         unsigned int tso_segs, sent_pkts;
2045         int cwnd_quota;
2046         int result;
2047         bool is_cwnd_limited = false;
2048         u32 max_segs;
2049 
2050         sent_pkts = 0;
2051 
2052         if (!push_one) {
2053                 /* Do MTU probing. */
2054                 result = tcp_mtu_probe(sk);
2055                 if (!result) {
2056                         return false;
2057                 } else if (result > 0) {
2058                         sent_pkts = 1;
2059                 }
2060         }
2061 
2062         max_segs = tcp_tso_autosize(sk, mss_now);
2063         while ((skb = tcp_send_head(sk))) {
2064                 unsigned int limit;
2065 
2066                 tso_segs = tcp_init_tso_segs(skb, mss_now);
2067                 BUG_ON(!tso_segs);
2068 
2069                 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2070                         /* "skb_mstamp" is used as a start point for the retransmit timer */
2071                         skb_mstamp_get(&skb->skb_mstamp);
2072                         goto repair; /* Skip network transmission */
2073                 }
2074 
2075                 cwnd_quota = tcp_cwnd_test(tp, skb);
2076                 if (!cwnd_quota) {
2077                         if (push_one == 2)
2078                                 /* Force out a loss probe pkt. */
2079                                 cwnd_quota = 1;
2080                         else
2081                                 break;
2082                 }
2083 
2084                 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2085                         break;
2086 
2087                 if (tso_segs == 1) {
2088                         if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2089                                                      (tcp_skb_is_last(sk, skb) ?
2090                                                       nonagle : TCP_NAGLE_PUSH))))
2091                                 break;
2092                 } else {
2093                         if (!push_one &&
2094                             tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2095                                                  max_segs))
2096                                 break;
2097                 }
2098 
2099                 limit = mss_now;
2100                 if (tso_segs > 1 && !tcp_urg_mode(tp))
2101                         limit = tcp_mss_split_point(sk, skb, mss_now,
2102                                                     min_t(unsigned int,
2103                                                           cwnd_quota,
2104                                                           max_segs),
2105                                                     nonagle);
2106 
2107                 if (skb->len > limit &&
2108                     unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2109                         break;
2110 
2111                 /* TCP Small Queues :
2112                  * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2113                  * This allows for :
2114                  *  - better RTT estimation and ACK scheduling
2115                  *  - faster recovery
2116                  *  - high rates
2117                  * Alas, some drivers / subsystems require a fair amount
2118                  * of queued bytes to ensure line rate.
2119                  * One example is wifi aggregation (802.11 AMPDU)
2120                  */
2121                 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2122                 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2123 
2124                 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2125                         set_bit(TSQ_THROTTLED, &tp->tsq_flags);
2126                         /* It is possible TX completion already happened
2127                          * before we set TSQ_THROTTLED, so we must
2128                          * test again the condition.
2129                          */
2130                         smp_mb__after_atomic();
2131                         if (atomic_read(&sk->sk_wmem_alloc) > limit)
2132                                 break;
2133                 }
2134 
2135                 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2136                         break;
2137 
2138 repair:
2139                 /* Advance the send_head.  This one is sent out.
2140                  * This call will increment packets_out.
2141                  */
2142                 tcp_event_new_data_sent(sk, skb);
2143 
2144                 tcp_minshall_update(tp, mss_now, skb);
2145                 sent_pkts += tcp_skb_pcount(skb);
2146 
2147                 if (push_one)
2148                         break;
2149         }
2150 
2151         if (likely(sent_pkts)) {
2152                 if (tcp_in_cwnd_reduction(sk))
2153                         tp->prr_out += sent_pkts;
2154 
2155                 /* Send one loss probe per tail loss episode. */
2156                 if (push_one != 2)
2157                         tcp_schedule_loss_probe(sk);
2158                 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2159                 tcp_cwnd_validate(sk, is_cwnd_limited);
2160                 return false;
2161         }
2162         return !tp->packets_out && tcp_send_head(sk);
2163 }
2164 
2165 bool tcp_schedule_loss_probe(struct sock *sk)
2166 {
2167         struct inet_connection_sock *icsk = inet_csk(sk);
2168         struct tcp_sock *tp = tcp_sk(sk);
2169         u32 timeout, tlp_time_stamp, rto_time_stamp;
2170         u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2171 
2172         if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2173                 return false;
2174         /* No consecutive loss probes. */
2175         if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2176                 tcp_rearm_rto(sk);
2177                 return false;
2178         }
2179         /* Don't do any loss probe on a Fast Open connection before 3WHS
2180          * finishes.
2181          */
2182         if (tp->fastopen_rsk)
2183                 return false;
2184 
2185         /* TLP is only scheduled when next timer event is RTO. */
2186         if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2187                 return false;
2188 
2189         /* Schedule a loss probe in 2*RTT for SACK capable connections
2190          * in Open state, that are either limited by cwnd or application.
2191          */
2192         if (sysctl_tcp_early_retrans < 3 || !tp->packets_out ||
2193             !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2194                 return false;
2195 
2196         if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2197              tcp_send_head(sk))
2198                 return false;
2199 
2200         /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2201          * for delayed ack when there's one outstanding packet. If no RTT
2202          * sample is available then probe after TCP_TIMEOUT_INIT.
2203          */
2204         timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
2205         if (tp->packets_out == 1)
2206                 timeout = max_t(u32, timeout,
2207                                 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2208         timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2209 
2210         /* If RTO is shorter, just schedule TLP in its place. */
2211         tlp_time_stamp = tcp_time_stamp + timeout;
2212         rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2213         if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2214                 s32 delta = rto_time_stamp - tcp_time_stamp;
2215                 if (delta > 0)
2216                         timeout = delta;
2217         }
2218 
2219         inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2220                                   TCP_RTO_MAX);
2221         return true;
2222 }
2223 
2224 /* Thanks to skb fast clones, we can detect if a prior transmit of
2225  * a packet is still in a qdisc or driver queue.
2226  * In this case, there is very little point doing a retransmit !
2227  */
2228 static bool skb_still_in_host_queue(const struct sock *sk,
2229                                     const struct sk_buff *skb)
2230 {
2231         if (unlikely(skb_fclone_busy(sk, skb))) {
2232                 NET_INC_STATS(sock_net(sk),
2233                               LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2234                 return true;
2235         }
2236         return false;
2237 }
2238 
2239 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2240  * retransmit the last segment.
2241  */
2242 void tcp_send_loss_probe(struct sock *sk)
2243 {
2244         struct tcp_sock *tp = tcp_sk(sk);
2245         struct sk_buff *skb;
2246         int pcount;
2247         int mss = tcp_current_mss(sk);
2248 
2249         skb = tcp_send_head(sk);
2250         if (skb) {
2251                 if (tcp_snd_wnd_test(tp, skb, mss)) {
2252                         pcount = tp->packets_out;
2253                         tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2254                         if (tp->packets_out > pcount)
2255                                 goto probe_sent;
2256                         goto rearm_timer;
2257                 }
2258                 skb = tcp_write_queue_prev(sk, skb);
2259         } else {
2260                 skb = tcp_write_queue_tail(sk);
2261         }
2262 
2263         /* At most one outstanding TLP retransmission. */
2264         if (tp->tlp_high_seq)
2265                 goto rearm_timer;
2266 
2267         /* Retransmit last segment. */
2268         if (WARN_ON(!skb))
2269                 goto rearm_timer;
2270 
2271         if (skb_still_in_host_queue(sk, skb))
2272                 goto rearm_timer;
2273 
2274         pcount = tcp_skb_pcount(skb);
2275         if (WARN_ON(!pcount))
2276                 goto rearm_timer;
2277 
2278         if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2279                 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2280                                           GFP_ATOMIC)))
2281                         goto rearm_timer;
2282                 skb = tcp_write_queue_next(sk, skb);
2283         }
2284 
2285         if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2286                 goto rearm_timer;
2287 
2288         if (__tcp_retransmit_skb(sk, skb, 1))
2289                 goto rearm_timer;
2290 
2291         /* Record snd_nxt for loss detection. */
2292         tp->tlp_high_seq = tp->snd_nxt;
2293 
2294 probe_sent:
2295         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2296         /* Reset s.t. tcp_rearm_rto will restart timer from now */
2297         inet_csk(sk)->icsk_pending = 0;
2298 rearm_timer:
2299         tcp_rearm_rto(sk);
2300 }
2301 
2302 /* Push out any pending frames which were held back due to
2303  * TCP_CORK or attempt at coalescing tiny packets.
2304  * The socket must be locked by the caller.
2305  */
2306 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2307                                int nonagle)
2308 {
2309         /* If we are closed, the bytes will have to remain here.
2310          * In time closedown will finish, we empty the write queue and
2311          * all will be happy.
2312          */
2313         if (unlikely(sk->sk_state == TCP_CLOSE))
2314                 return;
2315 
2316         if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2317                            sk_gfp_mask(sk, GFP_ATOMIC)))
2318                 tcp_check_probe_timer(sk);
2319 }
2320 
2321 /* Send _single_ skb sitting at the send head. This function requires
2322  * true push pending frames to setup probe timer etc.
2323  */
2324 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2325 {
2326         struct sk_buff *skb = tcp_send_head(sk);
2327 
2328         BUG_ON(!skb || skb->len < mss_now);
2329 
2330         tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2331 }
2332 
2333 /* This function returns the amount that we can raise the
2334  * usable window based on the following constraints
2335  *
2336  * 1. The window can never be shrunk once it is offered (RFC 793)
2337  * 2. We limit memory per socket
2338  *
2339  * RFC 1122:
2340  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2341  *  RECV.NEXT + RCV.WIN fixed until:
2342  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2343  *
2344  * i.e. don't raise the right edge of the window until you can raise
2345  * it at least MSS bytes.
2346  *
2347  * Unfortunately, the recommended algorithm breaks header prediction,
2348  * since header prediction assumes th->window stays fixed.
2349  *
2350  * Strictly speaking, keeping th->window fixed violates the receiver
2351  * side SWS prevention criteria. The problem is that under this rule
2352  * a stream of single byte packets will cause the right side of the
2353  * window to always advance by a single byte.
2354  *
2355  * Of course, if the sender implements sender side SWS prevention
2356  * then this will not be a problem.
2357  *
2358  * BSD seems to make the following compromise:
2359  *
2360  *      If the free space is less than the 1/4 of the maximum
2361  *      space available and the free space is less than 1/2 mss,
2362  *      then set the window to 0.
2363  *      [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2364  *      Otherwise, just prevent the window from shrinking
2365  *      and from being larger than the largest representable value.
2366  *
2367  * This prevents incremental opening of the window in the regime
2368  * where TCP is limited by the speed of the reader side taking
2369  * data out of the TCP receive queue. It does nothing about
2370  * those cases where the window is constrained on the sender side
2371  * because the pipeline is full.
2372  *
2373  * BSD also seems to "accidentally" limit itself to windows that are a
2374  * multiple of MSS, at least until the free space gets quite small.
2375  * This would appear to be a side effect of the mbuf implementation.
2376  * Combining these two algorithms results in the observed behavior
2377  * of having a fixed window size at almost all times.
2378  *
2379  * Below we obtain similar behavior by forcing the offered window to
2380  * a multiple of the mss when it is feasible to do so.
2381  *
2382  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2383  * Regular options like TIMESTAMP are taken into account.
2384  */
2385 u32 __tcp_select_window(struct sock *sk)
2386 {
2387         struct inet_connection_sock *icsk = inet_csk(sk);
2388         struct tcp_sock *tp = tcp_sk(sk);
2389         /* MSS for the peer's data.  Previous versions used mss_clamp
2390          * here.  I don't know if the value based on our guesses
2391          * of peer's MSS is better for the performance.  It's more correct
2392          * but may be worse for the performance because of rcv_mss
2393          * fluctuations.  --SAW  1998/11/1
2394          */
2395         int mss = icsk->icsk_ack.rcv_mss;
2396         int free_space = tcp_space(sk);
2397         int allowed_space = tcp_full_space(sk);
2398         int full_space = min_t(int, tp->window_clamp, allowed_space);
2399         int window;
2400 
2401         if (mss > full_space)
2402                 mss = full_space;
2403 
2404         if (free_space < (full_space >> 1)) {
2405                 icsk->icsk_ack.quick = 0;
2406 
2407                 if (tcp_under_memory_pressure(sk))
2408                         tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2409                                                4U * tp->advmss);
2410 
2411                 /* free_space might become our new window, make sure we don't
2412                  * increase it due to wscale.
2413                  */
2414                 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2415 
2416                 /* if free space is less than mss estimate, or is below 1/16th
2417                  * of the maximum allowed, try to move to zero-window, else
2418                  * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2419                  * new incoming data is dropped due to memory limits.
2420                  * With large window, mss test triggers way too late in order
2421                  * to announce zero window in time before rmem limit kicks in.
2422                  */
2423                 if (free_space < (allowed_space >> 4) || free_space < mss)
2424                         return 0;
2425         }
2426 
2427         if (free_space > tp->rcv_ssthresh)
2428                 free_space = tp->rcv_ssthresh;
2429 
2430         /* Don't do rounding if we are using window scaling, since the
2431          * scaled window will not line up with the MSS boundary anyway.
2432          */
2433         window = tp->rcv_wnd;
2434         if (tp->rx_opt.rcv_wscale) {
2435                 window = free_space;
2436 
2437                 /* Advertise enough space so that it won't get scaled away.
2438                  * Import case: prevent zero window announcement if
2439                  * 1<<rcv_wscale > mss.
2440                  */
2441                 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2442                         window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2443                                   << tp->rx_opt.rcv_wscale);
2444         } else {
2445                 /* Get the largest window that is a nice multiple of mss.
2446                  * Window clamp already applied above.
2447                  * If our current window offering is within 1 mss of the
2448                  * free space we just keep it. This prevents the divide
2449                  * and multiply from happening most of the time.
2450                  * We also don't do any window rounding when the free space
2451                  * is too small.
2452                  */
2453                 if (window <= free_space - mss || window > free_space)
2454                         window = (free_space / mss) * mss;
2455                 else if (mss == full_space &&
2456                          free_space > window + (full_space >> 1))
2457                         window = free_space;
2458         }
2459 
2460         return window;
2461 }
2462 
2463 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2464                              const struct sk_buff *next_skb)
2465 {
2466         if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2467                 const struct skb_shared_info *next_shinfo =
2468                         skb_shinfo(next_skb);
2469                 struct skb_shared_info *shinfo = skb_shinfo(skb);
2470 
2471                 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2472                 shinfo->tskey = next_shinfo->tskey;
2473                 TCP_SKB_CB(skb)->txstamp_ack |=
2474                         TCP_SKB_CB(next_skb)->txstamp_ack;
2475         }
2476 }
2477 
2478 /* Collapses two adjacent SKB's during retransmission. */
2479 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2480 {
2481         struct tcp_sock *tp = tcp_sk(sk);
2482         struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2483         int skb_size, next_skb_size;
2484 
2485         skb_size = skb->len;
2486         next_skb_size = next_skb->len;
2487 
2488         BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2489 
2490         tcp_highest_sack_combine(sk, next_skb, skb);
2491 
2492         tcp_unlink_write_queue(next_skb, sk);
2493 
2494         skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2495                                   next_skb_size);
2496 
2497         if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2498                 skb->ip_summed = CHECKSUM_PARTIAL;
2499 
2500         if (skb->ip_summed != CHECKSUM_PARTIAL)
2501                 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2502 
2503         /* Update sequence range on original skb. */
2504         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2505 
2506         /* Merge over control information. This moves PSH/FIN etc. over */
2507         TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2508 
2509         /* All done, get rid of second SKB and account for it so
2510          * packet counting does not break.
2511          */
2512         TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2513         TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2514 
2515         /* changed transmit queue under us so clear hints */
2516         tcp_clear_retrans_hints_partial(tp);
2517         if (next_skb == tp->retransmit_skb_hint)
2518                 tp->retransmit_skb_hint = skb;
2519 
2520         tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2521 
2522         tcp_skb_collapse_tstamp(skb, next_skb);
2523 
2524         sk_wmem_free_skb(sk, next_skb);
2525 }
2526 
2527 /* Check if coalescing SKBs is legal. */
2528 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2529 {
2530         if (tcp_skb_pcount(skb) > 1)
2531                 return false;
2532         /* TODO: SACK collapsing could be used to remove this condition */
2533         if (skb_shinfo(skb)->nr_frags != 0)
2534                 return false;
2535         if (skb_cloned(skb))
2536                 return false;
2537         if (skb == tcp_send_head(sk))
2538                 return false;
2539         /* Some heurestics for collapsing over SACK'd could be invented */
2540         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2541                 return false;
2542 
2543         return true;
2544 }
2545 
2546 /* Collapse packets in the retransmit queue to make to create
2547  * less packets on the wire. This is only done on retransmission.
2548  */
2549 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2550                                      int space)
2551 {
2552         struct tcp_sock *tp = tcp_sk(sk);
2553         struct sk_buff *skb = to, *tmp;
2554         bool first = true;
2555 
2556         if (!sysctl_tcp_retrans_collapse)
2557                 return;
2558         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2559                 return;
2560 
2561         tcp_for_write_queue_from_safe(skb, tmp, sk) {
2562                 if (!tcp_can_collapse(sk, skb))
2563                         break;
2564 
2565                 if (!tcp_skb_can_collapse_to(to))
2566                         break;
2567 
2568                 space -= skb->len;
2569 
2570                 if (first) {
2571                         first = false;
2572                         continue;
2573                 }
2574 
2575                 if (space < 0)
2576                         break;
2577                 /* Punt if not enough space exists in the first SKB for
2578                  * the data in the second
2579                  */
2580                 if (skb->len > skb_availroom(to))
2581                         break;
2582 
2583                 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2584                         break;
2585 
2586                 tcp_collapse_retrans(sk, to);
2587         }
2588 }
2589 
2590 /* This retransmits one SKB.  Policy decisions and retransmit queue
2591  * state updates are done by the caller.  Returns non-zero if an
2592  * error occurred which prevented the send.
2593  */
2594 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2595 {
2596         struct inet_connection_sock *icsk = inet_csk(sk);
2597         struct tcp_sock *tp = tcp_sk(sk);
2598         unsigned int cur_mss;
2599         int diff, len, err;
2600 
2601 
2602         /* Inconclusive MTU probe */
2603         if (icsk->icsk_mtup.probe_size)
2604                 icsk->icsk_mtup.probe_size = 0;
2605 
2606         /* Do not sent more than we queued. 1/4 is reserved for possible
2607          * copying overhead: fragmentation, tunneling, mangling etc.
2608          */
2609         if (atomic_read(&sk->sk_wmem_alloc) >
2610             min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2611                   sk->sk_sndbuf))
2612                 return -EAGAIN;
2613 
2614         if (skb_still_in_host_queue(sk, skb))
2615                 return -EBUSY;
2616 
2617         if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2618                 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2619                         BUG();
2620                 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2621                         return -ENOMEM;
2622         }
2623 
2624         if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2625                 return -EHOSTUNREACH; /* Routing failure or similar. */
2626 
2627         cur_mss = tcp_current_mss(sk);
2628 
2629         /* If receiver has shrunk his window, and skb is out of
2630          * new window, do not retransmit it. The exception is the
2631          * case, when window is shrunk to zero. In this case
2632          * our retransmit serves as a zero window probe.
2633          */
2634         if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2635             TCP_SKB_CB(skb)->seq != tp->snd_una)
2636                 return -EAGAIN;
2637 
2638         len = cur_mss * segs;
2639         if (skb->len > len) {
2640                 if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC))
2641                         return -ENOMEM; /* We'll try again later. */
2642         } else {
2643                 if (skb_unclone(skb, GFP_ATOMIC))
2644                         return -ENOMEM;
2645 
2646                 diff = tcp_skb_pcount(skb);
2647                 tcp_set_skb_tso_segs(skb, cur_mss);
2648                 diff -= tcp_skb_pcount(skb);
2649                 if (diff)
2650                         tcp_adjust_pcount(sk, skb, diff);
2651                 if (skb->len < cur_mss)
2652                         tcp_retrans_try_collapse(sk, skb, cur_mss);
2653         }
2654 
2655         /* RFC3168, section 6.1.1.1. ECN fallback */
2656         if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2657                 tcp_ecn_clear_syn(sk, skb);
2658 
2659         /* make sure skb->data is aligned on arches that require it
2660          * and check if ack-trimming & collapsing extended the headroom
2661          * beyond what csum_start can cover.
2662          */
2663         if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2664                      skb_headroom(skb) >= 0xFFFF)) {
2665                 struct sk_buff *nskb;
2666 
2667                 skb_mstamp_get(&skb->skb_mstamp);
2668                 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2669                 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2670                              -ENOBUFS;
2671         } else {
2672                 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2673         }
2674 
2675         if (likely(!err)) {
2676                 segs = tcp_skb_pcount(skb);
2677 
2678                 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2679                 /* Update global TCP statistics. */
2680                 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2681                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2682                         __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2683                 tp->total_retrans += segs;
2684         }
2685         return err;
2686 }
2687 
2688 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2689 {
2690         struct tcp_sock *tp = tcp_sk(sk);
2691         int err = __tcp_retransmit_skb(sk, skb, segs);
2692 
2693         if (err == 0) {
2694 #if FASTRETRANS_DEBUG > 0
2695                 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2696                         net_dbg_ratelimited("retrans_out leaked\n");
2697                 }
2698 #endif
2699                 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2700                 tp->retrans_out += tcp_skb_pcount(skb);
2701 
2702                 /* Save stamp of the first retransmit. */
2703                 if (!tp->retrans_stamp)
2704                         tp->retrans_stamp = tcp_skb_timestamp(skb);
2705 
2706         } else if (err != -EBUSY) {
2707                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2708         }
2709 
2710         if (tp->undo_retrans < 0)
2711                 tp->undo_retrans = 0;
2712         tp->undo_retrans += tcp_skb_pcount(skb);
2713         return err;
2714 }
2715 
2716 /* Check if we forward retransmits are possible in the current
2717  * window/congestion state.
2718  */
2719 static bool tcp_can_forward_retransmit(struct sock *sk)
2720 {
2721         const struct inet_connection_sock *icsk = inet_csk(sk);
2722         const struct tcp_sock *tp = tcp_sk(sk);
2723 
2724         /* Forward retransmissions are possible only during Recovery. */
2725         if (icsk->icsk_ca_state != TCP_CA_Recovery)
2726                 return false;
2727 
2728         /* No forward retransmissions in Reno are possible. */
2729         if (tcp_is_reno(tp))
2730                 return false;
2731 
2732         /* Yeah, we have to make difficult choice between forward transmission
2733          * and retransmission... Both ways have their merits...
2734          *
2735          * For now we do not retransmit anything, while we have some new
2736          * segments to send. In the other cases, follow rule 3 for
2737          * NextSeg() specified in RFC3517.
2738          */
2739 
2740         if (tcp_may_send_now(sk))
2741                 return false;
2742 
2743         return true;
2744 }
2745 
2746 /* This gets called after a retransmit timeout, and the initially
2747  * retransmitted data is acknowledged.  It tries to continue
2748  * resending the rest of the retransmit queue, until either
2749  * we've sent it all or the congestion window limit is reached.
2750  * If doing SACK, the first ACK which comes back for a timeout
2751  * based retransmit packet might feed us FACK information again.
2752  * If so, we use it to avoid unnecessarily retransmissions.
2753  */
2754 void tcp_xmit_retransmit_queue(struct sock *sk)
2755 {
2756         const struct inet_connection_sock *icsk = inet_csk(sk);
2757         struct tcp_sock *tp = tcp_sk(sk);
2758         struct sk_buff *skb;
2759         struct sk_buff *hole = NULL;
2760         u32 max_segs, last_lost;
2761         int mib_idx;
2762         int fwd_rexmitting = 0;
2763 
2764         if (!tp->packets_out)
2765                 return;
2766 
2767         if (!tp->lost_out)
2768                 tp->retransmit_high = tp->snd_una;
2769 
2770         if (tp->retransmit_skb_hint) {
2771                 skb = tp->retransmit_skb_hint;
2772                 last_lost = TCP_SKB_CB(skb)->end_seq;
2773                 if (after(last_lost, tp->retransmit_high))
2774                         last_lost = tp->retransmit_high;
2775         } else {
2776                 skb = tcp_write_queue_head(sk);
2777                 last_lost = tp->snd_una;
2778         }
2779 
2780         max_segs = tcp_tso_autosize(sk, tcp_current_mss(sk));
2781         tcp_for_write_queue_from(skb, sk) {
2782                 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2783                 int segs;
2784 
2785                 if (skb == tcp_send_head(sk))
2786                         break;
2787                 /* we could do better than to assign each time */
2788                 if (!hole)
2789                         tp->retransmit_skb_hint = skb;
2790 
2791                 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2792                 if (segs <= 0)
2793                         return;
2794                 /* In case tcp_shift_skb_data() have aggregated large skbs,
2795                  * we need to make sure not sending too bigs TSO packets
2796                  */
2797                 segs = min_t(int, segs, max_segs);
2798 
2799                 if (fwd_rexmitting) {
2800 begin_fwd:
2801                         if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2802                                 break;
2803                         mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2804 
2805                 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2806                         tp->retransmit_high = last_lost;
2807                         if (!tcp_can_forward_retransmit(sk))
2808                                 break;
2809                         /* Backtrack if necessary to non-L'ed skb */
2810                         if (hole) {
2811                                 skb = hole;
2812                                 hole = NULL;
2813                         }
2814                         fwd_rexmitting = 1;
2815                         goto begin_fwd;
2816 
2817                 } else if (!(sacked & TCPCB_LOST)) {
2818                         if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2819                                 hole = skb;
2820                         continue;
2821 
2822                 } else {
2823                         last_lost = TCP_SKB_CB(skb)->end_seq;
2824                         if (icsk->icsk_ca_state != TCP_CA_Loss)
2825                                 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2826                         else
2827                                 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2828                 }
2829 
2830                 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2831                         continue;
2832 
2833                 if (tcp_retransmit_skb(sk, skb, segs))
2834                         return;
2835 
2836                 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
2837 
2838                 if (tcp_in_cwnd_reduction(sk))
2839                         tp->prr_out += tcp_skb_pcount(skb);
2840 
2841                 if (skb == tcp_write_queue_head(sk))
2842                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2843                                                   inet_csk(sk)->icsk_rto,
2844                                                   TCP_RTO_MAX);
2845         }
2846 }
2847 
2848 /* We allow to exceed memory limits for FIN packets to expedite
2849  * connection tear down and (memory) recovery.
2850  * Otherwise tcp_send_fin() could be tempted to either delay FIN
2851  * or even be forced to close flow without any FIN.
2852  * In general, we want to allow one skb per socket to avoid hangs
2853  * with edge trigger epoll()
2854  */
2855 void sk_forced_mem_schedule(struct sock *sk, int size)
2856 {
2857         int amt;
2858 
2859         if (size <= sk->sk_forward_alloc)
2860                 return;
2861         amt = sk_mem_pages(size);
2862         sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2863         sk_memory_allocated_add(sk, amt);
2864 
2865         if (mem_cgroup_sockets_enabled && sk->sk_memcg)
2866                 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
2867 }
2868 
2869 /* Send a FIN. The caller locks the socket for us.
2870  * We should try to send a FIN packet really hard, but eventually give up.
2871  */
2872 void tcp_send_fin(struct sock *sk)
2873 {
2874         struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
2875         struct tcp_sock *tp = tcp_sk(sk);
2876 
2877         /* Optimization, tack on the FIN if we have one skb in write queue and
2878          * this skb was not yet sent, or we are under memory pressure.
2879          * Note: in the latter case, FIN packet will be sent after a timeout,
2880          * as TCP stack thinks it has already been transmitted.
2881          */
2882         if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
2883 coalesce:
2884                 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2885                 TCP_SKB_CB(tskb)->end_seq++;
2886                 tp->write_seq++;
2887                 if (!tcp_send_head(sk)) {
2888                         /* This means tskb was already sent.
2889                          * Pretend we included the FIN on previous transmit.
2890                          * We need to set tp->snd_nxt to the value it would have
2891                          * if FIN had been sent. This is because retransmit path
2892                          * does not change tp->snd_nxt.
2893                          */
2894                         tp->snd_nxt++;
2895                         return;
2896                 }
2897         } else {
2898                 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2899                 if (unlikely(!skb)) {
2900                         if (tskb)
2901                                 goto coalesce;
2902                         return;
2903                 }
2904                 skb_reserve(skb, MAX_TCP_HEADER);
2905                 sk_forced_mem_schedule(sk, skb->truesize);
2906                 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2907                 tcp_init_nondata_skb(skb, tp->write_seq,
2908                                      TCPHDR_ACK | TCPHDR_FIN);
2909                 tcp_queue_skb(sk, skb);
2910         }
2911         __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
2912 }
2913 
2914 /* We get here when a process closes a file descriptor (either due to
2915  * an explicit close() or as a byproduct of exit()'ing) and there
2916  * was unread data in the receive queue.  This behavior is recommended
2917  * by RFC 2525, section 2.17.  -DaveM
2918  */
2919 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2920 {
2921         struct sk_buff *skb;
2922 
2923         /* NOTE: No TCP options attached and we never retransmit this. */
2924         skb = alloc_skb(MAX_TCP_HEADER, priority);
2925         if (!skb) {
2926                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2927                 return;
2928         }
2929 
2930         /* Reserve space for headers and prepare control bits. */
2931         skb_reserve(skb, MAX_TCP_HEADER);
2932         tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2933                              TCPHDR_ACK | TCPHDR_RST);
2934         skb_mstamp_get(&skb->skb_mstamp);
2935         /* Send it off. */
2936         if (tcp_transmit_skb(sk, skb, 0, priority))
2937                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2938 
2939         TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2940 }
2941 
2942 /* Send a crossed SYN-ACK during socket establishment.
2943  * WARNING: This routine must only be called when we have already sent
2944  * a SYN packet that crossed the incoming SYN that caused this routine
2945  * to get called. If this assumption fails then the initial rcv_wnd
2946  * and rcv_wscale values will not be correct.
2947  */
2948 int tcp_send_synack(struct sock *sk)
2949 {
2950         struct sk_buff *skb;
2951 
2952         skb = tcp_write_queue_head(sk);
2953         if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2954                 pr_debug("%s: wrong queue state\n", __func__);
2955                 return -EFAULT;
2956         }
2957         if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2958                 if (skb_cloned(skb)) {
2959                         struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2960                         if (!nskb)
2961                                 return -ENOMEM;
2962                         tcp_unlink_write_queue(skb, sk);
2963                         __skb_header_release(nskb);
2964                         __tcp_add_write_queue_head(sk, nskb);
2965                         sk_wmem_free_skb(sk, skb);
2966                         sk->sk_wmem_queued += nskb->truesize;
2967                         sk_mem_charge(sk, nskb->truesize);
2968                         skb = nskb;
2969                 }
2970 
2971                 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2972                 tcp_ecn_send_synack(sk, skb);
2973         }
2974         return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2975 }
2976 
2977 /**
2978  * tcp_make_synack - Prepare a SYN-ACK.
2979  * sk: listener socket
2980  * dst: dst entry attached to the SYNACK
2981  * req: request_sock pointer
2982  *
2983  * Allocate one skb and build a SYNACK packet.
2984  * @dst is consumed : Caller should not use it again.
2985  */
2986 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
2987                                 struct request_sock *req,
2988                                 struct tcp_fastopen_cookie *foc,
2989                                 enum tcp_synack_type synack_type)
2990 {
2991         struct inet_request_sock *ireq = inet_rsk(req);
2992         const struct tcp_sock *tp = tcp_sk(sk);
2993         struct tcp_md5sig_key *md5 = NULL;
2994         struct tcp_out_options opts;
2995         struct sk_buff *skb;
2996         int tcp_header_size;
2997         struct tcphdr *th;
2998         u16 user_mss;
2999         int mss;
3000 
3001         skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3002         if (unlikely(!skb)) {
3003                 dst_release(dst);
3004                 return NULL;
3005         }
3006         /* Reserve space for headers. */
3007         skb_reserve(skb, MAX_TCP_HEADER);
3008 
3009         switch (synack_type) {
3010         case TCP_SYNACK_NORMAL:
3011                 skb_set_owner_w(skb, req_to_sk(req));
3012                 break;
3013         case TCP_SYNACK_COOKIE:
3014                 /* Under synflood, we do not attach skb to a socket,
3015                  * to avoid false sharing.
3016                  */
3017                 break;
3018         case TCP_SYNACK_FASTOPEN:
3019                 /* sk is a const pointer, because we want to express multiple
3020                  * cpu might call us concurrently.
3021                  * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3022                  */
3023                 skb_set_owner_w(skb, (struct sock *)sk);
3024                 break;
3025         }
3026         skb_dst_set(skb, dst);
3027 
3028         mss = dst_metric_advmss(dst);
3029         user_mss = READ_ONCE(tp->rx_opt.user_mss);
3030         if (user_mss && user_mss < mss)
3031                 mss = user_mss;
3032 
3033         memset(&opts, 0, sizeof(opts));
3034 #ifdef CONFIG_SYN_COOKIES
3035         if (unlikely(req->cookie_ts))
3036                 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
3037         else
3038 #endif
3039         skb_mstamp_get(&skb->skb_mstamp);
3040 
3041 #ifdef CONFIG_TCP_MD5SIG
3042         rcu_read_lock();
3043         md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3044 #endif
3045         skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3046         tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
3047                           sizeof(*th);
3048 
3049         skb_push(skb, tcp_header_size);
3050         skb_reset_transport_header(skb);
3051 
3052         th = (struct tcphdr *)skb->data;
3053         memset(th, 0, sizeof(struct tcphdr));
3054         th->syn = 1;
3055         th->ack = 1;
3056         tcp_ecn_make_synack(req, th);
3057         th->source = htons(ireq->ir_num);
3058         th->dest = ireq->ir_rmt_port;
3059         /* Setting of flags are superfluous here for callers (and ECE is
3060          * not even correctly set)
3061          */
3062         tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
3063                              TCPHDR_SYN | TCPHDR_ACK);
3064 
3065         th->seq = htonl(TCP_SKB_CB(skb)->seq);
3066         /* XXX data is queued and acked as is. No buffer/window check */
3067         th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3068 
3069         /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3070         th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3071         tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3072         th->doff = (tcp_header_size >> 2);
3073         __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3074 
3075 #ifdef CONFIG_TCP_MD5SIG
3076         /* Okay, we have all we need - do the md5 hash if needed */
3077         if (md5)
3078                 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3079                                                md5, req_to_sk(req), skb);
3080         rcu_read_unlock();
3081 #endif
3082 
3083         /* Do not fool tcpdump (if any), clean our debris */
3084         skb->tstamp.tv64 = 0;
3085         return skb;
3086 }
3087 EXPORT_SYMBOL(tcp_make_synack);
3088 
3089 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3090 {
3091         struct inet_connection_sock *icsk = inet_csk(sk);
3092         const struct tcp_congestion_ops *ca;
3093         u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3094 
3095         if (ca_key == TCP_CA_UNSPEC)
3096                 return;
3097 
3098         rcu_read_lock();
3099         ca = tcp_ca_find_key(ca_key);
3100         if (likely(ca && try_module_get(ca->owner))) {
3101                 module_put(icsk->icsk_ca_ops->owner);
3102                 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3103                 icsk->icsk_ca_ops = ca;
3104         }
3105         rcu_read_unlock();
3106 }
3107 
3108 /* Do all connect socket setups that can be done AF independent. */
3109 static void tcp_connect_init(struct sock *sk)
3110 {
3111         const struct dst_entry *dst = __sk_dst_get(sk);
3112         struct tcp_sock *tp = tcp_sk(sk);
3113         __u8 rcv_wscale;
3114 
3115         /* We'll fix this up when we get a response from the other end.
3116          * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3117          */
3118         tp->tcp_header_len = sizeof(struct tcphdr) +
3119                 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3120 
3121 #ifdef CONFIG_TCP_MD5SIG
3122         if (tp->af_specific->md5_lookup(sk, sk))
3123                 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3124 #endif
3125 
3126         /* If user gave his TCP_MAXSEG, record it to clamp */
3127         if (tp->rx_opt.user_mss)
3128                 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3129         tp->max_window = 0;
3130         tcp_mtup_init(sk);
3131         tcp_sync_mss(sk, dst_mtu(dst));
3132 
3133         tcp_ca_dst_init(sk, dst);
3134 
3135         if (!tp->window_clamp)
3136                 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3137         tp->advmss = dst_metric_advmss(dst);
3138         if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3139                 tp->advmss = tp->rx_opt.user_mss;
3140 
3141         tcp_initialize_rcv_mss(sk);
3142 
3143         /* limit the window selection if the user enforce a smaller rx buffer */
3144         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3145             (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3146                 tp->window_clamp = tcp_full_space(sk);
3147 
3148         tcp_select_initial_window(tcp_full_space(sk),
3149                                   tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3150                                   &tp->rcv_wnd,
3151                                   &tp->window_clamp,
3152                                   sysctl_tcp_window_scaling,
3153                                   &rcv_wscale,
3154                                   dst_metric(dst, RTAX_INITRWND));
3155 
3156         tp->rx_opt.rcv_wscale = rcv_wscale;
3157         tp->rcv_ssthresh = tp->rcv_wnd;
3158 
3159         sk->sk_err = 0;
3160         sock_reset_flag(sk, SOCK_DONE);
3161         tp->snd_wnd = 0;
3162         tcp_init_wl(tp, 0);
3163         tp->snd_una = tp->write_seq;
3164         tp->snd_sml = tp->write_seq;
3165         tp->snd_up = tp->write_seq;
3166         tp->snd_nxt = tp->write_seq;
3167 
3168         if (likely(!tp->repair))
3169                 tp->rcv_nxt = 0;
3170         else
3171                 tp->rcv_tstamp = tcp_time_stamp;
3172         tp->rcv_wup = tp->rcv_nxt;
3173         tp->copied_seq = tp->rcv_nxt;
3174 
3175         inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3176         inet_csk(sk)->icsk_retransmits = 0;
3177         tcp_clear_retrans(tp);
3178 }
3179 
3180 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3181 {
3182         struct tcp_sock *tp = tcp_sk(sk);
3183         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3184 
3185         tcb->end_seq += skb->len;
3186         __skb_header_release(skb);
3187         __tcp_add_write_queue_tail(sk, skb);
3188         sk->sk_wmem_queued += skb->truesize;
3189         sk_mem_charge(sk, skb->truesize);
3190         tp->write_seq = tcb->end_seq;
3191         tp->packets_out += tcp_skb_pcount(skb);
3192 }
3193 
3194 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3195  * queue a data-only packet after the regular SYN, such that regular SYNs
3196  * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3197  * only the SYN sequence, the data are retransmitted in the first ACK.
3198  * If cookie is not cached or other error occurs, falls back to send a
3199  * regular SYN with Fast Open cookie request option.
3200  */
3201 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3202 {
3203         struct tcp_sock *tp = tcp_sk(sk);
3204         struct tcp_fastopen_request *fo = tp->fastopen_req;
3205         int syn_loss = 0, space, err = 0;
3206         unsigned long last_syn_loss = 0;
3207         struct sk_buff *syn_data;
3208 
3209         tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
3210         tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3211                                &syn_loss, &last_syn_loss);
3212         /* Recurring FO SYN losses: revert to regular handshake temporarily */
3213         if (syn_loss > 1 &&
3214             time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3215                 fo->cookie.len = -1;
3216                 goto fallback;
3217         }
3218 
3219         if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3220                 fo->cookie.len = -1;
3221         else if (fo->cookie.len <= 0)
3222                 goto fallback;
3223 
3224         /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3225          * user-MSS. Reserve maximum option space for middleboxes that add
3226          * private TCP options. The cost is reduced data space in SYN :(
3227          */
3228         if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3229                 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3230         space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3231                 MAX_TCP_OPTION_SPACE;
3232 
3233         space = min_t(size_t, space, fo->size);
3234 
3235         /* limit to order-0 allocations */
3236         space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3237 
3238         syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3239         if (!syn_data)
3240                 goto fallback;
3241         syn_data->ip_summed = CHECKSUM_PARTIAL;
3242         memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3243         if (space) {
3244                 int copied = copy_from_iter(skb_put(syn_data, space), space,
3245                                             &fo->data->msg_iter);
3246                 if (unlikely(!copied)) {
3247                         kfree_skb(syn_data);
3248                         goto fallback;
3249                 }
3250                 if (copied != space) {
3251                         skb_trim(syn_data, copied);
3252                         space = copied;
3253                 }
3254         }
3255         /* No more data pending in inet_wait_for_connect() */
3256         if (space == fo->size)
3257                 fo->data = NULL;
3258         fo->copied = space;
3259 
3260         tcp_connect_queue_skb(sk, syn_data);
3261 
3262         err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3263 
3264         syn->skb_mstamp = syn_data->skb_mstamp;
3265 
3266         /* Now full SYN+DATA was cloned and sent (or not),
3267          * remove the SYN from the original skb (syn_data)
3268          * we keep in write queue in case of a retransmit, as we
3269          * also have the SYN packet (with no data) in the same queue.
3270          */
3271         TCP_SKB_CB(syn_data)->seq++;
3272         TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3273         if (!err) {
3274                 tp->syn_data = (fo->copied > 0);
3275                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3276                 goto done;
3277         }
3278 
3279 fallback:
3280         /* Send a regular SYN with Fast Open cookie request option */
3281         if (fo->cookie.len > 0)
3282                 fo->cookie.len = 0;
3283         err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3284         if (err)
3285                 tp->syn_fastopen = 0;
3286 done:
3287         fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
3288         return err;
3289 }
3290 
3291 /* Build a SYN and send it off. */
3292 int tcp_connect(struct sock *sk)
3293 {
3294         struct tcp_sock *tp = tcp_sk(sk);
3295         struct sk_buff *buff;
3296         int err;
3297 
3298         tcp_connect_init(sk);
3299 
3300         if (unlikely(tp->repair)) {
3301                 tcp_finish_connect(sk, NULL);
3302                 return 0;
3303         }
3304 
3305         buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3306         if (unlikely(!buff))
3307                 return -ENOBUFS;
3308 
3309         tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3310         tp->retrans_stamp = tcp_time_stamp;
3311         tcp_connect_queue_skb(sk, buff);
3312         tcp_ecn_send_syn(sk, buff);
3313 
3314         /* Send off SYN; include data in Fast Open. */
3315         err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3316               tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3317         if (err == -ECONNREFUSED)
3318                 return err;
3319 
3320         /* We change tp->snd_nxt after the tcp_transmit_skb() call
3321          * in order to make this packet get counted in tcpOutSegs.
3322          */
3323         tp->snd_nxt = tp->write_seq;
3324         tp->pushed_seq = tp->write_seq;
3325         TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3326 
3327         /* Timer for repeating the SYN until an answer. */
3328         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3329                                   inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3330         return 0;
3331 }
3332 EXPORT_SYMBOL(tcp_connect);
3333 
3334 /* Send out a delayed ack, the caller does the policy checking
3335  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
3336  * for details.
3337  */
3338 void tcp_send_delayed_ack(struct sock *sk)
3339 {
3340         struct inet_connection_sock *icsk = inet_csk(sk);
3341         int ato = icsk->icsk_ack.ato;
3342         unsigned long timeout;
3343 
3344         tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3345 
3346         if (ato > TCP_DELACK_MIN) {
3347                 const struct tcp_sock *tp = tcp_sk(sk);
3348                 int max_ato = HZ / 2;
3349 
3350                 if (icsk->icsk_ack.pingpong ||
3351                     (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3352                         max_ato = TCP_DELACK_MAX;
3353 
3354                 /* Slow path, intersegment interval is "high". */
3355 
3356                 /* If some rtt estimate is known, use it to bound delayed ack.
3357                  * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3358                  * directly.
3359                  */
3360                 if (tp->srtt_us) {
3361                         int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3362                                         TCP_DELACK_MIN);
3363 
3364                         if (rtt < max_ato)
3365                                 max_ato = rtt;
3366                 }
3367 
3368                 ato = min(ato, max_ato);
3369         }
3370 
3371         /* Stay within the limit we were given */
3372         timeout = jiffies + ato;
3373 
3374         /* Use new timeout only if there wasn't a older one earlier. */
3375         if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3376                 /* If delack timer was blocked or is about to expire,
3377                  * send ACK now.
3378                  */
3379                 if (icsk->icsk_ack.blocked ||
3380                     time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3381                         tcp_send_ack(sk);
3382                         return;
3383                 }
3384 
3385                 if (!time_before(timeout, icsk->icsk_ack.timeout))
3386                         timeout = icsk->icsk_ack.timeout;
3387         }
3388         icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3389         icsk->icsk_ack.timeout = timeout;
3390         sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3391 }
3392 
3393 /* This routine sends an ack and also updates the window. */
3394 void tcp_send_ack(struct sock *sk)
3395 {
3396         struct sk_buff *buff;
3397 
3398         /* If we have been reset, we may not send again. */
3399         if (sk->sk_state == TCP_CLOSE)
3400                 return;
3401 
3402         tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3403 
3404         /* We are not putting this on the write queue, so
3405          * tcp_transmit_skb() will set the ownership to this
3406          * sock.
3407          */
3408         buff = alloc_skb(MAX_TCP_HEADER,
3409                          sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3410         if (unlikely(!buff)) {
3411                 inet_csk_schedule_ack(sk);
3412                 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3413                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3414                                           TCP_DELACK_MAX, TCP_RTO_MAX);
3415                 return;
3416         }
3417 
3418         /* Reserve space for headers and prepare control bits. */
3419         skb_reserve(buff, MAX_TCP_HEADER);
3420         tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3421 
3422         /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3423          * too much.
3424          * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3425          * We also avoid tcp_wfree() overhead (cache line miss accessing
3426          * tp->tsq_flags) by using regular sock_wfree()
3427          */
3428         skb_set_tcp_pure_ack(buff);
3429 
3430         /* Send it off, this clears delayed acks for us. */
3431         skb_mstamp_get(&buff->skb_mstamp);
3432         tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3433 }
3434 EXPORT_SYMBOL_GPL(tcp_send_ack);
3435 
3436 /* This routine sends a packet with an out of date sequence
3437  * number. It assumes the other end will try to ack it.
3438  *
3439  * Question: what should we make while urgent mode?
3440  * 4.4BSD forces sending single byte of data. We cannot send
3441  * out of window data, because we have SND.NXT==SND.MAX...
3442  *
3443  * Current solution: to send TWO zero-length segments in urgent mode:
3444  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3445  * out-of-date with SND.UNA-1 to probe window.
3446  */
3447 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3448 {
3449         struct tcp_sock *tp = tcp_sk(sk);
3450         struct sk_buff *skb;
3451 
3452         /* We don't queue it, tcp_transmit_skb() sets ownership. */
3453         skb = alloc_skb(MAX_TCP_HEADER,
3454                         sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3455         if (!skb)
3456                 return -1;
3457 
3458         /* Reserve space for headers and set control bits. */
3459         skb_reserve(skb, MAX_TCP_HEADER);
3460         /* Use a previous sequence.  This should cause the other
3461          * end to send an ack.  Don't queue or clone SKB, just
3462          * send it.
3463          */
3464         tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3465         skb_mstamp_get(&skb->skb_mstamp);
3466         NET_INC_STATS(sock_net(sk), mib);
3467         return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3468 }
3469 
3470 void tcp_send_window_probe(struct sock *sk)
3471 {
3472         if (sk->sk_state == TCP_ESTABLISHED) {
3473                 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3474                 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3475         }
3476 }
3477 
3478 /* Initiate keepalive or window probe from timer. */
3479 int tcp_write_wakeup(struct sock *sk, int mib)
3480 {
3481         struct tcp_sock *tp = tcp_sk(sk);
3482         struct sk_buff *skb;
3483 
3484         if (sk->sk_state == TCP_CLOSE)
3485                 return -1;
3486 
3487         skb = tcp_send_head(sk);
3488         if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3489                 int err;
3490                 unsigned int mss = tcp_current_mss(sk);
3491                 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3492 
3493                 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3494                         tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3495 
3496                 /* We are probing the opening of a window
3497                  * but the window size is != 0
3498                  * must have been a result SWS avoidance ( sender )
3499                  */
3500                 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3501                     skb->len > mss) {
3502                         seg_size = min(seg_size, mss);
3503                         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3504                         if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3505                                 return -1;
3506                 } else if (!tcp_skb_pcount(skb))
3507                         tcp_set_skb_tso_segs(skb, mss);
3508 
3509                 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3510                 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3511                 if (!err)
3512                         tcp_event_new_data_sent(sk, skb);
3513                 return err;
3514         } else {
3515                 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3516                         tcp_xmit_probe_skb(sk, 1, mib);
3517                 return tcp_xmit_probe_skb(sk, 0, mib);
3518         }
3519 }
3520 
3521 /* A window probe timeout has occurred.  If window is not closed send
3522  * a partial packet else a zero probe.
3523  */
3524 void tcp_send_probe0(struct sock *sk)
3525 {
3526         struct inet_connection_sock *icsk = inet_csk(sk);
3527         struct tcp_sock *tp = tcp_sk(sk);
3528         struct net *net = sock_net(sk);
3529         unsigned long probe_max;
3530         int err;
3531 
3532         err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3533 
3534         if (tp->packets_out || !tcp_send_head(sk)) {
3535                 /* Cancel probe timer, if it is not required. */
3536                 icsk->icsk_probes_out = 0;
3537                 icsk->icsk_backoff = 0;
3538                 return;
3539         }
3540 
3541         if (err <= 0) {
3542                 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3543                         icsk->icsk_backoff++;
3544                 icsk->icsk_probes_out++;
3545                 probe_max = TCP_RTO_MAX;
3546         } else {
3547                 /* If packet was not sent due to local congestion,
3548                  * do not backoff and do not remember icsk_probes_out.
3549                  * Let local senders to fight for local resources.
3550                  *
3551                  * Use accumulated backoff yet.
3552                  */
3553                 if (!icsk->icsk_probes_out)
3554                         icsk->icsk_probes_out = 1;
3555                 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3556         }
3557         inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3558                                   tcp_probe0_when(sk, probe_max),
3559                                   TCP_RTO_MAX);
3560 }
3561 
3562 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3563 {
3564         const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3565         struct flowi fl;
3566         int res;
3567 
3568         tcp_rsk(req)->txhash = net_tx_rndhash();
3569         res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3570         if (!res) {
3571                 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3572                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3573                 if (unlikely(tcp_passive_fastopen(sk)))
3574                         tcp_sk(sk)->total_retrans++;
3575         }
3576         return res;
3577 }
3578 EXPORT_SYMBOL(tcp_rtx_synack);
3579 

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